bump to 3.3-alpha1

. Missing } and unprotected min/max calls and definitions.

.hgeol

@@ -1,6 +1,9 @@
 [patterns]
+*.sh = LF
+*.MINPACK = CRLF
 scripts/*.in = LF
 debug/msvc/*.dat = CRLF
+debug/msvc/*.natvis = CRLF
 unsupported/test/mpreal/*.* = CRLF
 ** = native
 

.hgignore

@@ -30,3 +30,5 @@ log
 patch
 a
 a.*
+lapack/testing
+lapack/reference

CMakeLists.txt

@@ -1,6 +1,6 @@
 project(Eigen)
 
-cmake_minimum_required(VERSION 2.8.2)
+cmake_minimum_required(VERSION 2.8.4)
 
 # guard against in-source builds
 
@@ -108,7 +108,8 @@ endif()
 set(EIGEN_TEST_MAX_SIZE "320" CACHE STRING "Maximal matrix/vector size, default is 320")
 
 macro(ei_add_cxx_compiler_flag FLAG)
-  string(REGEX REPLACE "-" "" SFLAG ${FLAG})
+  string(REGEX REPLACE "-" "" SFLAG1 ${FLAG})
+  string(REGEX REPLACE "\\+" "p" SFLAG ${SFLAG1})
   check_cxx_compiler_flag(${FLAG} COMPILER_SUPPORT_${SFLAG})
   if(COMPILER_SUPPORT_${SFLAG})
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${FLAG}")
@@ -118,7 +119,7 @@ endmacro(ei_add_cxx_compiler_flag)
 if(NOT MSVC)
   # We assume that other compilers are partly compatible with GNUCC
   
-  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fexceptions")
+#  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fexceptions")
   set(CMAKE_CXX_FLAGS_DEBUG "-g3")
   set(CMAKE_CXX_FLAGS_RELEASE "-g0 -O2")
   
@@ -142,6 +143,15 @@ if(NOT MSVC)
   ei_add_cxx_compiler_flag("-Wpointer-arith")
   ei_add_cxx_compiler_flag("-Wwrite-strings")
   ei_add_cxx_compiler_flag("-Wformat-security")
+  ei_add_cxx_compiler_flag("-Wshorten-64-to-32")
+  ei_add_cxx_compiler_flag("-Wenum-conversion")
+  ei_add_cxx_compiler_flag("-Wc++11-extensions")
+  
+  # -Wshadow is insanely too strict with gcc, hopefully it will become usable with gcc 6
+  # if(NOT CMAKE_COMPILER_IS_GNUCXX OR (CMAKE_CXX_COMPILER_VERSION VERSION_GREATER "5.0.0"))
+  if(NOT CMAKE_COMPILER_IS_GNUCXX)
+    ei_add_cxx_compiler_flag("-Wshadow")
+  endif()
   
   ei_add_cxx_compiler_flag("-Wno-psabi")
   ei_add_cxx_compiler_flag("-Wno-variadic-macros")
@@ -153,6 +163,7 @@ if(NOT MSVC)
   ei_add_cxx_compiler_flag("-wd981")                    # disable ICC's "operands are evaluated in unspecified order" remark
   ei_add_cxx_compiler_flag("-wd2304")                   # disbale ICC's "warning #2304: non-explicit constructor with single argument may cause implicit type conversion" produced by -Wnon-virtual-dtor
   
+  
   # The -ansi flag must be added last, otherwise it is also used as a linker flag by check_cxx_compiler_flag making it fails
   # Moreover we should not set both -strict-ansi and -ansi
   check_cxx_compiler_flag("-strict-ansi" COMPILER_SUPPORT_STRICTANSI)
@@ -163,6 +174,11 @@ if(NOT MSVC)
   else()
     ei_add_cxx_compiler_flag("-ansi")
   endif()
+
+  if(ANDROID_NDK)
+    ei_add_cxx_compiler_flag("-pie")
+    ei_add_cxx_compiler_flag("-fPIE")
+  endif()
   
   set(CMAKE_REQUIRED_FLAGS "")
 
@@ -196,18 +212,49 @@ if(NOT MSVC)
     message(STATUS "Enabling SSE4.2 in tests/examples")
   endif()
 
+  option(EIGEN_TEST_AVX "Enable/Disable AVX in tests/examples" OFF)
+  if(EIGEN_TEST_AVX)
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mavx")
+    message(STATUS "Enabling AVX in tests/examples")
+  endif()
+
+  option(EIGEN_TEST_FMA "Enable/Disable FMA in tests/examples" OFF)
+  if(EIGEN_TEST_FMA AND NOT EIGEN_TEST_NEON)
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mfma")
+    message(STATUS "Enabling FMA in tests/examples")
+  endif()
+
   option(EIGEN_TEST_ALTIVEC "Enable/Disable AltiVec in tests/examples" OFF)
   if(EIGEN_TEST_ALTIVEC)
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -maltivec -mabi=altivec")
     message(STATUS "Enabling AltiVec in tests/examples")
   endif()
 
+  option(EIGEN_TEST_VSX "Enable/Disable VSX in tests/examples" OFF)
+  if(EIGEN_TEST_VSX)
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -m64 -mvsx")
+    message(STATUS "Enabling VSX in tests/examples")
+  endif()
+
   option(EIGEN_TEST_NEON "Enable/Disable Neon in tests/examples" OFF)
   if(EIGEN_TEST_NEON)
-    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mfpu=neon -mcpu=cortex-a"8)
+    if(EIGEN_TEST_FMA)
+      set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mfpu=neon-vfpv4")
+    else()
+      set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mfpu=neon")
+    endif()
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mfloat-abi=softfp")
     message(STATUS "Enabling NEON in tests/examples")
   endif()
 
+  option(EIGEN_TEST_NEON64 "Enable/Disable Neon in tests/examples" OFF)
+  if(EIGEN_TEST_NEON64)
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS}")
+    message(STATUS "Enabling NEON in tests/examples")
+  endif()
+
+
+
   check_cxx_compiler_flag("-fopenmp" COMPILER_SUPPORT_OPENMP)
   if(COMPILER_SUPPORT_OPENMP)
     option(EIGEN_TEST_OPENMP "Enable/Disable OpenMP in tests/examples" OFF)
@@ -284,7 +331,13 @@ if(EIGEN_TEST_NO_EXPLICIT_ALIGNMENT)
   message(STATUS "Disabling alignment in tests/examples")
 endif()
 
-option(EIGEN_TEST_C++0x "Enables all C++0x features." OFF)
+option(EIGEN_TEST_NO_EXCEPTIONS "Disables C++ exceptions" OFF)
+if(EIGEN_TEST_NO_EXCEPTIONS)
+  ei_add_cxx_compiler_flag("-fno-exceptions")
+  message(STATUS "Disabling exceptions in tests/examples")
+endif()
+
+option(EIGEN_TEST_CXX11 "Enable testing with C++11 and C++11 features (e.g. Tensor module)." OFF)
 
 include_directories(${CMAKE_CURRENT_SOURCE_DIR} ${CMAKE_CURRENT_BINARY_DIR})
 
@@ -411,6 +464,7 @@ if(cmake_generator_tolower MATCHES "makefile")
   message(STATUS "make check    | Build and run the unit-tests. Read this page:")
   message(STATUS "              |   http://eigen.tuxfamily.org/index.php?title=Tests")
   message(STATUS "make blas     | Build BLAS library (not the same thing as Eigen)")
+  message(STATUS "make uninstall| Removes files installed by make install")
   message(STATUS "--------------+--------------------------------------------------------------")
 else()
   message(STATUS "To build/run the unit tests, read this page:")
@@ -418,3 +472,35 @@ else()
 endif()
 
 message(STATUS "")
+
+set ( EIGEN_CONFIG_CMAKE_PATH
+      lib${LIB_SUFFIX}/cmake/eigen3
+      CACHE PATH "The directory where the CMake files are installed"
+    )
+if ( NOT IS_ABSOLUTE EIGEN_CONFIG_CMAKE_PATH )
+  set ( EIGEN_CONFIG_CMAKE_PATH ${CMAKE_INSTALL_PREFIX}/${EIGEN_CONFIG_CMAKE_PATH} )
+endif ()
+
+set ( EIGEN_USE_FILE ${EIGEN_CONFIG_CMAKE_PATH}/UseEigen3.cmake )
+set ( EIGEN_VERSION_STRING ${EIGEN_VERSION_NUMBER} )
+set ( EIGEN_VERSION_MAJOR  ${EIGEN_WORLD_VERSION} )
+set ( EIGEN_VERSION_MINOR  ${EIGEN_MAJOR_VERSION} )
+set ( EIGEN_VERSION_PATCH  ${EIGEN_MINOR_VERSION} )
+set ( EIGEN_DEFINITIONS "")
+set ( EIGEN_INCLUDE_DIR ${INCLUDE_INSTALL_DIR} )
+set ( EIGEN_INCLUDE_DIRS ${EIGEN_INCLUDE_DIR} )
+set ( EIGEN_ROOT_DIR ${CMAKE_INSTALL_PREFIX} )
+
+configure_file ( ${CMAKE_CURRENT_SOURCE_DIR}/cmake/Eigen3Config.cmake.in
+                 ${CMAKE_CURRENT_BINARY_DIR}/Eigen3Config.cmake
+                 @ONLY ESCAPE_QUOTES
+               )
+
+install ( FILES ${CMAKE_CURRENT_SOURCE_DIR}/cmake/UseEigen3.cmake
+                ${CMAKE_CURRENT_BINARY_DIR}/Eigen3Config.cmake
+          DESTINATION ${EIGEN_CONFIG_CMAKE_PATH}
+        )
+
+# Add uninstall target
+add_custom_target ( uninstall
+    COMMAND ${CMAKE_COMMAND} -P ${CMAKE_CURRENT_SOURCE_DIR}/cmake/EigenUninstall.cmake)

Eigen/Array

@@ -1,11 +0,0 @@
-#ifndef EIGEN_ARRAY_MODULE_H
-#define EIGEN_ARRAY_MODULE_H
-
-// include Core first to handle Eigen2 support macros
-#include "Core"
-
-#ifndef EIGEN2_SUPPORT
-  #error The Eigen/Array header does no longer exist in Eigen3. All that functionality has moved to Eigen/Core.
-#endif
-
-#endif // EIGEN_ARRAY_MODULE_H

Eigen/Cholesky

@@ -10,16 +10,17 @@
   *
   *
   * This module provides two variants of the Cholesky decomposition for selfadjoint (hermitian) matrices.
-  * Those decompositions are accessible via the following MatrixBase methods:
-  *  - MatrixBase::llt(),
+  * Those decompositions are also accessible via the following methods:
+  *  - MatrixBase::llt()
   *  - MatrixBase::ldlt()
+  *  - SelfAdjointView::llt()
+  *  - SelfAdjointView::ldlt()
   *
   * \code
   * #include <Eigen/Cholesky>
   * \endcode
   */
 
-#include "src/misc/Solve.h"
 #include "src/Cholesky/LLT.h"
 #include "src/Cholesky/LDLT.h"
 #ifdef EIGEN_USE_LAPACKE

Eigen/CholmodSupport

@@ -33,12 +33,8 @@ extern "C" {
   *
   */
 
-#include "src/misc/Solve.h"
-#include "src/misc/SparseSolve.h"
-
 #include "src/CholmodSupport/CholmodSupport.h"
 
-
 #include "src/Core/util/ReenableStupidWarnings.h"
 
 #endif // EIGEN_CHOLMODSUPPORT_MODULE_H

Eigen/Core

@@ -14,6 +14,48 @@
 // first thing Eigen does: stop the compiler from committing suicide
 #include "src/Core/util/DisableStupidWarnings.h"
 
+// Handle NVCC/CUDA
+#ifdef __CUDACC__
+  // Do not try asserts on CUDA!
+  #ifndef EIGEN_NO_DEBUG
+  #define EIGEN_NO_DEBUG
+  #endif
+
+  #ifdef EIGEN_INTERNAL_DEBUGGING
+  #undef EIGEN_INTERNAL_DEBUGGING
+  #endif
+
+  // Do not try to vectorize on CUDA!
+  #ifndef EIGEN_DONT_VECTORIZE
+  #define EIGEN_DONT_VECTORIZE
+  #endif
+
+  #ifdef EIGEN_EXCEPTIONS
+  #undef EIGEN_EXCEPTIONS
+  #endif
+  
+  // All functions callable from CUDA code must be qualified with __device__
+  #define EIGEN_DEVICE_FUNC __host__ __device__
+  
+#else
+  #define EIGEN_DEVICE_FUNC
+  
+#endif
+
+#if defined(__CUDA_ARCH__)
+  #define EIGEN_USING_STD_MATH(FUNC) using ::FUNC;
+#else
+  #define EIGEN_USING_STD_MATH(FUNC) using std::FUNC;
+#endif
+
+#if (defined(_CPPUNWIND) || defined(__EXCEPTIONS)) && !defined(__CUDA_ARCH__) && !defined(EIGEN_EXCEPTIONS)
+  #define EIGEN_EXCEPTIONS
+#endif
+
+#ifdef EIGEN_EXCEPTIONS
+  #include <new>
+#endif
+
 // then include this file where all our macros are defined. It's really important to do it first because
 // it's where we do all the alignment settings (platform detection and honoring the user's will if he
 // defined e.g. EIGEN_DONT_ALIGN) so it needs to be done before we do anything with vectorization.
@@ -21,7 +63,7 @@
 
 // Disable the ipa-cp-clone optimization flag with MinGW 6.x or newer (enabled by default with -O3)
 // See http://eigen.tuxfamily.org/bz/show_bug.cgi?id=556 for details.
-#if defined(__MINGW32__) && EIGEN_GNUC_AT_LEAST(4,6)
+#if EIGEN_COMP_MINGW && EIGEN_GNUC_AT_LEAST(4,6)
   #pragma GCC optimize ("-fno-ipa-cp-clone")
 #endif
 
@@ -31,26 +73,26 @@
 // and inclusion of their respective header files
 #include "src/Core/util/MKL_support.h"
 
-// if alignment is disabled, then disable vectorization. Note: EIGEN_ALIGN is the proper check, it takes into
-// account both the user's will (EIGEN_DONT_ALIGN) and our own platform checks
-#if !EIGEN_ALIGN
+// if alignment is disabled, then disable vectorization. Note: EIGEN_MAX_ALIGN_BYTES is the proper check, it takes into
+// account both the user's will (EIGEN_MAX_ALIGN_BYTES,EIGEN_DONT_ALIGN) and our own platform checks
+#if EIGEN_MAX_ALIGN_BYTES==0
   #ifndef EIGEN_DONT_VECTORIZE
     #define EIGEN_DONT_VECTORIZE
   #endif
 #endif
 
-#ifdef _MSC_VER
+#if EIGEN_COMP_MSVC
   #include <malloc.h> // for _aligned_malloc -- need it regardless of whether vectorization is enabled
-  #if (_MSC_VER >= 1500) // 2008 or later
+  #if (EIGEN_COMP_MSVC >= 1500) // 2008 or later
     // Remember that usage of defined() in a #define is undefined by the standard.
     // a user reported that in 64-bit mode, MSVC doesn't care to define _M_IX86_FP.
-    #if (defined(_M_IX86_FP) && (_M_IX86_FP >= 2)) || defined(_M_X64)
+    #if (defined(_M_IX86_FP) && (_M_IX86_FP >= 2)) || EIGEN_ARCH_x86_64
       #define EIGEN_SSE2_ON_MSVC_2008_OR_LATER
     #endif
   #endif
 #else
   // Remember that usage of defined() in a #define is undefined by the standard
-  #if (defined __SSE2__) && ( (!defined __GNUC__) || (defined __INTEL_COMPILER) || EIGEN_GNUC_AT_LEAST(4,2) )
+  #if (defined __SSE2__) && ( (!EIGEN_COMP_GNUC) || EIGEN_COMP_ICC || EIGEN_GNUC_AT_LEAST(4,2) )
     #define EIGEN_SSE2_ON_NON_MSVC_BUT_NOT_OLD_GCC
   #endif
 #endif
@@ -82,6 +124,19 @@
     #ifdef __SSE4_2__
       #define EIGEN_VECTORIZE_SSE4_2
     #endif
+    #ifdef __AVX__
+      #define EIGEN_VECTORIZE_AVX
+      #define EIGEN_VECTORIZE_SSE3
+      #define EIGEN_VECTORIZE_SSSE3
+      #define EIGEN_VECTORIZE_SSE4_1
+      #define EIGEN_VECTORIZE_SSE4_2
+    #endif
+    #ifdef __AVX2__
+      #define EIGEN_VECTORIZE_AVX2
+    #endif
+    #ifdef __FMA__
+      #define EIGEN_VECTORIZE_FMA
+    #endif
 
     // include files
 
@@ -95,7 +150,7 @@
     extern "C" {
       // In theory we should only include immintrin.h and not the other *mmintrin.h header files directly.
       // Doing so triggers some issues with ICC. However old gcc versions seems to not have this file, thus:
-      #ifdef __INTEL_COMPILER
+      #if EIGEN_COMP_ICC >= 1110
         #include <immintrin.h>
       #else
         #include <emmintrin.h>
@@ -112,8 +167,20 @@
         #ifdef EIGEN_VECTORIZE_SSE4_2
         #include <nmmintrin.h>
         #endif
+        #ifdef EIGEN_VECTORIZE_AVX
+        #include <immintrin.h>
+        #endif
       #endif
     } // end extern "C"
+  #elif defined __VSX__
+    #define EIGEN_VECTORIZE
+    #define EIGEN_VECTORIZE_VSX
+    #include <altivec.h>
+    // We need to #undef all these ugly tokens defined in <altivec.h>
+    // => use __vector instead of vector
+    #undef bool
+    #undef vector
+    #undef pixel
   #elif defined __ALTIVEC__
     #define EIGEN_VECTORIZE
     #define EIGEN_VECTORIZE_ALTIVEC
@@ -123,13 +190,18 @@
     #undef bool
     #undef vector
     #undef pixel
-  #elif defined  __ARM_NEON__
+  #elif (defined  __ARM_NEON) || (defined __ARM_NEON__)
     #define EIGEN_VECTORIZE
     #define EIGEN_VECTORIZE_NEON
     #include <arm_neon.h>
   #endif
 #endif
 
+#if defined __CUDACC__
+  #define EIGEN_VECTORIZE_CUDA
+  #include <vector_types.h>
+#endif
+
 #if (defined _OPENMP) && (!defined EIGEN_DONT_PARALLELIZE)
   #define EIGEN_HAS_OPENMP
 #endif
@@ -139,7 +211,7 @@
 #endif
 
 // MSVC for windows mobile does not have the errno.h file
-#if !(defined(_MSC_VER) && defined(_WIN32_WCE)) && !defined(__ARMCC_VERSION)
+#if !(EIGEN_COMP_MSVC && EIGEN_OS_WINCE) && !EIGEN_COMP_ARM
 #define EIGEN_HAS_ERRNO
 #endif
 
@@ -165,23 +237,17 @@
 #endif
 
 // required for __cpuid, needs to be included after cmath
-#if defined(_MSC_VER) && (defined(_M_IX86)||defined(_M_X64))
+#if EIGEN_COMP_MSVC && EIGEN_ARCH_i386_OR_x86_64 && !EIGEN_OS_WINCE
   #include <intrin.h>
 #endif
 
-#if defined(_CPPUNWIND) || defined(__EXCEPTIONS)
-  #define EIGEN_EXCEPTIONS
-#endif
-
-#ifdef EIGEN_EXCEPTIONS
-  #include <new>
-#endif
-
 /** \brief Namespace containing all symbols from the %Eigen library. */
 namespace Eigen {
 
 inline static const char *SimdInstructionSetsInUse(void) {
-#if defined(EIGEN_VECTORIZE_SSE4_2)
+#if defined(EIGEN_VECTORIZE_AVX)
+  return "AVX SSE, SSE2, SSE3, SSSE3, SSE4.1, SSE4.2";
+#elif defined(EIGEN_VECTORIZE_SSE4_2)
   return "SSE, SSE2, SSE3, SSSE3, SSE4.1, SSE4.2";
 #elif defined(EIGEN_VECTORIZE_SSE4_1)
   return "SSE, SSE2, SSE3, SSSE3, SSE4.1";
@@ -193,6 +259,8 @@ inline static const char *SimdInstructionSetsInUse(void) {
   return "SSE, SSE2";
 #elif defined(EIGEN_VECTORIZE_ALTIVEC)
   return "AltiVec";
+#elif defined(EIGEN_VECTORIZE_VSX)
+  return "VSX";
 #elif defined(EIGEN_VECTORIZE_NEON)
   return "ARM NEON";
 #else
@@ -202,34 +270,9 @@ inline static const char *SimdInstructionSetsInUse(void) {
 
 } // end namespace Eigen
 
-#define STAGE10_FULL_EIGEN2_API             10
-#define STAGE20_RESOLVE_API_CONFLICTS       20
-#define STAGE30_FULL_EIGEN3_API             30
-#define STAGE40_FULL_EIGEN3_STRICTNESS      40
-#define STAGE99_NO_EIGEN2_SUPPORT           99
-
-#if   defined EIGEN2_SUPPORT_STAGE40_FULL_EIGEN3_STRICTNESS
-  #define EIGEN2_SUPPORT
-  #define EIGEN2_SUPPORT_STAGE STAGE40_FULL_EIGEN3_STRICTNESS
-#elif defined EIGEN2_SUPPORT_STAGE30_FULL_EIGEN3_API
-  #define EIGEN2_SUPPORT
-  #define EIGEN2_SUPPORT_STAGE STAGE30_FULL_EIGEN3_API
-#elif defined EIGEN2_SUPPORT_STAGE20_RESOLVE_API_CONFLICTS
-  #define EIGEN2_SUPPORT
-  #define EIGEN2_SUPPORT_STAGE STAGE20_RESOLVE_API_CONFLICTS
-#elif defined EIGEN2_SUPPORT_STAGE10_FULL_EIGEN2_API
-  #define EIGEN2_SUPPORT
-  #define EIGEN2_SUPPORT_STAGE STAGE10_FULL_EIGEN2_API
-#elif defined EIGEN2_SUPPORT
-  // default to stage 3, that's what it's always meant
-  #define EIGEN2_SUPPORT_STAGE30_FULL_EIGEN3_API
-  #define EIGEN2_SUPPORT_STAGE STAGE30_FULL_EIGEN3_API
-#else
-  #define EIGEN2_SUPPORT_STAGE STAGE99_NO_EIGEN2_SUPPORT
-#endif
-
-#ifdef EIGEN2_SUPPORT
-#undef minor
+#if defined EIGEN2_SUPPORT_STAGE40_FULL_EIGEN3_STRICTNESS || defined EIGEN2_SUPPORT_STAGE30_FULL_EIGEN3_API || defined EIGEN2_SUPPORT_STAGE20_RESOLVE_API_CONFLICTS || defined EIGEN2_SUPPORT_STAGE10_FULL_EIGEN2_API || defined EIGEN2_SUPPORT
+// This will generate an error message:
+#error Eigen2-support is only available up to version 3.2. Please go to "http://eigen.tuxfamily.org/index.php?title=Eigen2" for further information
 #endif
 
 // we use size_t frequently and we'll never remember to prepend it with std:: everytime just to
@@ -249,8 +292,8 @@ using std::ptrdiff_t;
   */
 
 #include "src/Core/util/Constants.h"
-#include "src/Core/util/ForwardDeclarations.h"
 #include "src/Core/util/Meta.h"
+#include "src/Core/util/ForwardDeclarations.h"
 #include "src/Core/util/StaticAssert.h"
 #include "src/Core/util/XprHelper.h"
 #include "src/Core/util/Memory.h"
@@ -259,35 +302,64 @@ using std::ptrdiff_t;
 #include "src/Core/MathFunctions.h"
 #include "src/Core/GenericPacketMath.h"
 
-#if defined EIGEN_VECTORIZE_SSE
+#if defined EIGEN_VECTORIZE_AVX
+  // Use AVX for floats and doubles, SSE for integers
+  #include "src/Core/arch/SSE/PacketMath.h"
+  #include "src/Core/arch/SSE/Complex.h"
+  #include "src/Core/arch/SSE/MathFunctions.h"
+  #include "src/Core/arch/AVX/PacketMath.h"
+  #include "src/Core/arch/AVX/MathFunctions.h"
+  #include "src/Core/arch/AVX/Complex.h"
+  #include "src/Core/arch/AVX/TypeCasting.h"
+#elif defined EIGEN_VECTORIZE_SSE
   #include "src/Core/arch/SSE/PacketMath.h"
   #include "src/Core/arch/SSE/MathFunctions.h"
   #include "src/Core/arch/SSE/Complex.h"
-#elif defined EIGEN_VECTORIZE_ALTIVEC
+  #include "src/Core/arch/SSE/TypeCasting.h"
+#elif defined(EIGEN_VECTORIZE_ALTIVEC) || defined(EIGEN_VECTORIZE_VSX)
   #include "src/Core/arch/AltiVec/PacketMath.h"
+  #include "src/Core/arch/AltiVec/MathFunctions.h"
   #include "src/Core/arch/AltiVec/Complex.h"
 #elif defined EIGEN_VECTORIZE_NEON
   #include "src/Core/arch/NEON/PacketMath.h"
+  #include "src/Core/arch/NEON/MathFunctions.h"
   #include "src/Core/arch/NEON/Complex.h"
 #endif
 
+#if defined EIGEN_VECTORIZE_CUDA
+  #include "src/Core/arch/CUDA/PacketMath.h"
+  #include "src/Core/arch/CUDA/MathFunctions.h"
+#endif
+
 #include "src/Core/arch/Default/Settings.h"
 
-#include "src/Core/Functors.h"
+#include "src/Core/functors/BinaryFunctors.h"
+#include "src/Core/functors/UnaryFunctors.h"
+#include "src/Core/functors/NullaryFunctors.h"
+#include "src/Core/functors/StlFunctors.h"
+#include "src/Core/functors/AssignmentFunctors.h"
+
 #include "src/Core/DenseCoeffsBase.h"
 #include "src/Core/DenseBase.h"
 #include "src/Core/MatrixBase.h"
 #include "src/Core/EigenBase.h"
 
+#include "src/Core/Product.h"
+#include "src/Core/CoreEvaluators.h"
+#include "src/Core/AssignEvaluator.h"
+
 #ifndef EIGEN_PARSED_BY_DOXYGEN // work around Doxygen bug triggered by Assign.h r814874
                                 // at least confirmed with Doxygen 1.5.5 and 1.5.6
   #include "src/Core/Assign.h"
 #endif
 
+#include "src/Core/ArrayBase.h"
 #include "src/Core/util/BlasUtil.h"
 #include "src/Core/DenseStorage.h"
 #include "src/Core/NestByValue.h"
-#include "src/Core/ForceAlignedAccess.h"
+
+// #include "src/Core/ForceAlignedAccess.h"
+
 #include "src/Core/ReturnByValue.h"
 #include "src/Core/NoAlias.h"
 #include "src/Core/PlainObjectBase.h"
@@ -300,12 +372,12 @@ using std::ptrdiff_t;
 #include "src/Core/SelfCwiseBinaryOp.h"
 #include "src/Core/Dot.h"
 #include "src/Core/StableNorm.h"
-#include "src/Core/MapBase.h"
 #include "src/Core/Stride.h"
+#include "src/Core/MapBase.h"
 #include "src/Core/Map.h"
+#include "src/Core/Ref.h"
 #include "src/Core/Block.h"
 #include "src/Core/VectorBlock.h"
-#include "src/Core/Ref.h"
 #include "src/Core/Transpose.h"
 #include "src/Core/DiagonalMatrix.h"
 #include "src/Core/Diagonal.h"
@@ -318,14 +390,14 @@ using std::ptrdiff_t;
 #include "src/Core/IO.h"
 #include "src/Core/Swap.h"
 #include "src/Core/CommaInitializer.h"
-#include "src/Core/Flagged.h"
-#include "src/Core/ProductBase.h"
 #include "src/Core/GeneralProduct.h"
+#include "src/Core/Solve.h"
+#include "src/Core/Inverse.h"
 #include "src/Core/TriangularMatrix.h"
 #include "src/Core/SelfAdjointView.h"
 #include "src/Core/products/GeneralBlockPanelKernel.h"
 #include "src/Core/products/Parallelizer.h"
-#include "src/Core/products/CoeffBasedProduct.h"
+#include "src/Core/ProductEvaluators.h"
 #include "src/Core/products/GeneralMatrixVector.h"
 #include "src/Core/products/GeneralMatrixMatrix.h"
 #include "src/Core/SolveTriangular.h"
@@ -347,16 +419,8 @@ using std::ptrdiff_t;
 #include "src/Core/Random.h"
 #include "src/Core/Replicate.h"
 #include "src/Core/Reverse.h"
-#include "src/Core/ArrayBase.h"
 #include "src/Core/ArrayWrapper.h"
 
-#ifdef EIGEN_ENABLE_EVALUATORS
-#include "src/Core/Product.h"
-#include "src/Core/CoreEvaluators.h"
-#include "src/Core/AssignEvaluator.h"
-#include "src/Core/ProductEvaluators.h"
-#endif
-
 #ifdef EIGEN_USE_BLAS
 #include "src/Core/products/GeneralMatrixMatrix_MKL.h"
 #include "src/Core/products/GeneralMatrixVector_MKL.h"
@@ -376,8 +440,4 @@ using std::ptrdiff_t;
 
 #include "src/Core/util/ReenableStupidWarnings.h"
 
-#ifdef EIGEN2_SUPPORT
-#include "Eigen2Support"
-#endif
-
 #endif // EIGEN_CORE_H

Eigen/Eigen

@@ -1,2 +1,2 @@
 #include "Dense"
-//#include "Sparse"
+#include "Sparse"

Eigen/Eigen2Support

@@ -1,82 +0,0 @@
-// This file is part of Eigen, a lightweight C++ template library
-// for linear algebra.
-//
-// Copyright (C) 2009 Gael Guennebaud <gael.guennebaud@inria.fr>
-//
-// This Source Code Form is subject to the terms of the Mozilla
-// Public License v. 2.0. If a copy of the MPL was not distributed
-// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
-
-#ifndef EIGEN2SUPPORT_H
-#define EIGEN2SUPPORT_H
-
-#if (!defined(EIGEN2_SUPPORT)) || (!defined(EIGEN_CORE_H))
-#error Eigen2 support must be enabled by defining EIGEN2_SUPPORT before including any Eigen header
-#endif
-
-#include "src/Core/util/DisableStupidWarnings.h"
-
-/** \ingroup Support_modules
-  * \defgroup Eigen2Support_Module Eigen2 support module
-  * This module provides a couple of deprecated functions improving the compatibility with Eigen2.
-  *
-  * To use it, define EIGEN2_SUPPORT before including any Eigen header
-  * \code
-  * #define EIGEN2_SUPPORT
-  * \endcode
-  *
-  */
-
-#include "src/Eigen2Support/Macros.h"
-#include "src/Eigen2Support/Memory.h"
-#include "src/Eigen2Support/Meta.h"
-#include "src/Eigen2Support/Lazy.h"
-#include "src/Eigen2Support/Cwise.h"
-#include "src/Eigen2Support/CwiseOperators.h"
-#include "src/Eigen2Support/TriangularSolver.h"
-#include "src/Eigen2Support/Block.h"
-#include "src/Eigen2Support/VectorBlock.h"
-#include "src/Eigen2Support/Minor.h"
-#include "src/Eigen2Support/MathFunctions.h"
-
-
-#include "src/Core/util/ReenableStupidWarnings.h"
-
-// Eigen2 used to include iostream
-#include<iostream>
-
-#define EIGEN_USING_MATRIX_TYPEDEFS_FOR_TYPE_AND_SIZE(TypeSuffix, SizeSuffix) \
-using Eigen::Matrix##SizeSuffix##TypeSuffix; \
-using Eigen::Vector##SizeSuffix##TypeSuffix; \
-using Eigen::RowVector##SizeSuffix##TypeSuffix;
-
-#define EIGEN_USING_MATRIX_TYPEDEFS_FOR_TYPE(TypeSuffix) \
-EIGEN_USING_MATRIX_TYPEDEFS_FOR_TYPE_AND_SIZE(TypeSuffix, 2) \
-EIGEN_USING_MATRIX_TYPEDEFS_FOR_TYPE_AND_SIZE(TypeSuffix, 3) \
-EIGEN_USING_MATRIX_TYPEDEFS_FOR_TYPE_AND_SIZE(TypeSuffix, 4) \
-EIGEN_USING_MATRIX_TYPEDEFS_FOR_TYPE_AND_SIZE(TypeSuffix, X) \
-
-#define EIGEN_USING_MATRIX_TYPEDEFS \
-EIGEN_USING_MATRIX_TYPEDEFS_FOR_TYPE(i) \
-EIGEN_USING_MATRIX_TYPEDEFS_FOR_TYPE(f) \
-EIGEN_USING_MATRIX_TYPEDEFS_FOR_TYPE(d) \
-EIGEN_USING_MATRIX_TYPEDEFS_FOR_TYPE(cf) \
-EIGEN_USING_MATRIX_TYPEDEFS_FOR_TYPE(cd)
-
-#define USING_PART_OF_NAMESPACE_EIGEN \
-EIGEN_USING_MATRIX_TYPEDEFS \
-using Eigen::Matrix; \
-using Eigen::MatrixBase; \
-using Eigen::ei_random; \
-using Eigen::ei_real; \
-using Eigen::ei_imag; \
-using Eigen::ei_conj; \
-using Eigen::ei_abs; \
-using Eigen::ei_abs2; \
-using Eigen::ei_sqrt; \
-using Eigen::ei_exp; \
-using Eigen::ei_log; \
-using Eigen::ei_sin; \
-using Eigen::ei_cos;
-
-#endif // EIGEN2SUPPORT_H

Eigen/Geometry

@@ -9,10 +9,6 @@
 #include "LU"
 #include <limits>
 
-#ifndef M_PI
-#define M_PI 3.14159265358979323846
-#endif
-
 /** \defgroup Geometry_Module Geometry module
   *
   *
@@ -33,27 +29,23 @@
 #include "src/Geometry/OrthoMethods.h"
 #include "src/Geometry/EulerAngles.h"
 
-#if EIGEN2_SUPPORT_STAGE > STAGE20_RESOLVE_API_CONFLICTS
-  #include "src/Geometry/Homogeneous.h"
-  #include "src/Geometry/RotationBase.h"
-  #include "src/Geometry/Rotation2D.h"
-  #include "src/Geometry/Quaternion.h"
-  #include "src/Geometry/AngleAxis.h"
-  #include "src/Geometry/Transform.h"
-  #include "src/Geometry/Translation.h"
-  #include "src/Geometry/Scaling.h"
-  #include "src/Geometry/Hyperplane.h"
-  #include "src/Geometry/ParametrizedLine.h"
-  #include "src/Geometry/AlignedBox.h"
-  #include "src/Geometry/Umeyama.h"
+#include "src/Geometry/Homogeneous.h"
+#include "src/Geometry/RotationBase.h"
+#include "src/Geometry/Rotation2D.h"
+#include "src/Geometry/Quaternion.h"
+#include "src/Geometry/AngleAxis.h"
+#include "src/Geometry/Transform.h"
+#include "src/Geometry/Translation.h"
+#include "src/Geometry/Scaling.h"
+#include "src/Geometry/Hyperplane.h"
+#include "src/Geometry/ParametrizedLine.h"
+#include "src/Geometry/AlignedBox.h"
+#include "src/Geometry/Umeyama.h"
 
-  #if defined EIGEN_VECTORIZE_SSE
-    #include "src/Geometry/arch/Geometry_SSE.h"
-  #endif
-#endif
-
-#ifdef EIGEN2_SUPPORT
-#include "src/Eigen2Support/Geometry/All.h"
+// Use the SSE optimized version whenever possible. At the moment the
+// SSE version doesn't compile when AVX is enabled
+#if defined EIGEN_VECTORIZE_SSE && !defined EIGEN_VECTORIZE_AVX
+#include "src/Geometry/arch/Geometry_SSE.h"
 #endif
 
 #include "src/Core/util/ReenableStupidWarnings.h"

Eigen/IterativeLinearSolvers

@@ -12,26 +12,26 @@
   * This module currently provides iterative methods to solve problems of the form \c A \c x = \c b, where \c A is a squared matrix, usually very large and sparse.
   * Those solvers are accessible via the following classes:
   *  - ConjugateGradient for selfadjoint (hermitian) matrices,
+  *  - LeastSquaresConjugateGradient for rectangular least-square problems,
   *  - BiCGSTAB for general square matrices.
   *
   * These iterative solvers are associated with some preconditioners:
   *  - IdentityPreconditioner - not really useful
-  *  - DiagonalPreconditioner - also called JAcobi preconditioner, work very well on diagonal dominant matrices.
-  *  - IncompleteILUT - incomplete LU factorization with dual thresholding
+  *  - DiagonalPreconditioner - also called Jacobi preconditioner, work very well on diagonal dominant matrices.
+  *  - IncompleteLUT - incomplete LU factorization with dual thresholding
   *
   * Such problems can also be solved using the direct sparse decomposition modules: SparseCholesky, CholmodSupport, UmfPackSupport, SuperLUSupport.
   *
-  * \code
-  * #include <Eigen/IterativeLinearSolvers>
-  * \endcode
+    \code
+    #include <Eigen/IterativeLinearSolvers>
+    \endcode
   */
 
-#include "src/misc/Solve.h"
-#include "src/misc/SparseSolve.h"
-
+#include "src/IterativeLinearSolvers/SolveWithGuess.h"
 #include "src/IterativeLinearSolvers/IterativeSolverBase.h"
 #include "src/IterativeLinearSolvers/BasicPreconditioners.h"
 #include "src/IterativeLinearSolvers/ConjugateGradient.h"
+#include "src/IterativeLinearSolvers/LeastSquareConjugateGradient.h"
 #include "src/IterativeLinearSolvers/BiCGSTAB.h"
 #include "src/IterativeLinearSolvers/IncompleteLUT.h"
 

Eigen/LU

@@ -16,7 +16,6 @@
   * \endcode
   */
 
-#include "src/misc/Solve.h"
 #include "src/misc/Kernel.h"
 #include "src/misc/Image.h"
 #include "src/LU/FullPivLU.h"
@@ -25,16 +24,14 @@
 #include "src/LU/PartialPivLU_MKL.h"
 #endif
 #include "src/LU/Determinant.h"
-#include "src/LU/Inverse.h"
+#include "src/LU/InverseImpl.h"
 
-#if defined EIGEN_VECTORIZE_SSE
+// Use the SSE optimized version whenever possible. At the moment the
+// SSE version doesn't compile when AVX is enabled
+#if defined EIGEN_VECTORIZE_SSE && !defined EIGEN_VECTORIZE_AVX
   #include "src/LU/arch/Inverse_SSE.h"
 #endif
 
-#ifdef EIGEN2_SUPPORT
-  #include "src/Eigen2Support/LU.h"
-#endif
-
 #include "src/Core/util/ReenableStupidWarnings.h"
 
 #endif // EIGEN_LU_MODULE_H

Eigen/LeastSquares

@@ -1,32 +0,0 @@
-#ifndef EIGEN_REGRESSION_MODULE_H
-#define EIGEN_REGRESSION_MODULE_H
-
-#ifndef EIGEN2_SUPPORT
-#error LeastSquares is only available in Eigen2 support mode (define EIGEN2_SUPPORT)
-#endif
-
-// exclude from normal eigen3-only documentation
-#ifdef EIGEN2_SUPPORT
-
-#include "Core"
-
-#include "src/Core/util/DisableStupidWarnings.h"
-
-#include "Eigenvalues"
-#include "Geometry"
-
-/** \defgroup LeastSquares_Module LeastSquares module
-  * This module provides linear regression and related features.
-  *
-  * \code
-  * #include <Eigen/LeastSquares>
-  * \endcode
-  */
-
-#include "src/Eigen2Support/LeastSquares.h"
-
-#include "src/Core/util/ReenableStupidWarnings.h"
-
-#endif // EIGEN2_SUPPORT
-
-#endif // EIGEN_REGRESSION_MODULE_H

Eigen/PaStiXSupport

@@ -35,12 +35,8 @@ extern "C" {
   *
   */
 
-#include "src/misc/Solve.h"
-#include "src/misc/SparseSolve.h"
-
 #include "src/PaStiXSupport/PaStiXSupport.h"
 
-
 #include "src/Core/util/ReenableStupidWarnings.h"
 
 #endif // EIGEN_PASTIXSUPPORT_MODULE_H

Eigen/QR

@@ -15,14 +15,15 @@
   *
   * This module provides various QR decompositions
   * This module also provides some MatrixBase methods, including:
-  *  - MatrixBase::qr(),
+  *  - MatrixBase::householderQr()
+  *  - MatrixBase::colPivHouseholderQr()
+  *  - MatrixBase::fullPivHouseholderQr()
   *
   * \code
   * #include <Eigen/QR>
   * \endcode
   */
 
-#include "src/misc/Solve.h"
 #include "src/QR/HouseholderQR.h"
 #include "src/QR/FullPivHouseholderQR.h"
 #include "src/QR/ColPivHouseholderQR.h"
@@ -31,15 +32,7 @@
 #include "src/QR/ColPivHouseholderQR_MKL.h"
 #endif
 
-#ifdef EIGEN2_SUPPORT
-#include "src/Eigen2Support/QR.h"
-#endif
-
 #include "src/Core/util/ReenableStupidWarnings.h"
 
-#ifdef EIGEN2_SUPPORT
-#include "Eigenvalues"
-#endif
-
 #endif // EIGEN_QR_MODULE_H
 /* vim: set filetype=cpp et sw=2 ts=2 ai: */

Eigen/SPQRSupport

@@ -21,8 +21,6 @@
   *
   */
 
-#include "src/misc/Solve.h"
-#include "src/misc/SparseSolve.h"
 #include "src/CholmodSupport/CholmodSupport.h"
 #include "src/SPQRSupport/SuiteSparseQRSupport.h"
 

Eigen/SVD

@@ -12,24 +12,25 @@
   *
   *
   * This module provides SVD decomposition for matrices (both real and complex).
-  * This decomposition is accessible via the following MatrixBase method:
+  * Two decomposition algorithms are provided:
+  *  - JacobiSVD implementing two-sided Jacobi iterations is numerically very accurate, fast for small matrices, but very slow for larger ones.
+  *  - BDCSVD implementing a recursive divide & conquer strategy on top of an upper-bidiagonalization which remains fast for large problems.
+  * These decompositions are accessible via the respective classes and following MatrixBase methods:
   *  - MatrixBase::jacobiSvd()
+  *  - MatrixBase::bdcSvd()
   *
   * \code
   * #include <Eigen/SVD>
   * \endcode
   */
 
-#include "src/misc/Solve.h"
+#include "src/SVD/UpperBidiagonalization.h"
+#include "src/SVD/SVDBase.h"
 #include "src/SVD/JacobiSVD.h"
+#include "src/SVD/BDCSVD.h"
 #if defined(EIGEN_USE_LAPACKE) && !defined(EIGEN_USE_LAPACKE_STRICT)
 #include "src/SVD/JacobiSVD_MKL.h"
 #endif
-#include "src/SVD/UpperBidiagonalization.h"
-
-#ifdef EIGEN2_SUPPORT
-#include "src/Eigen2Support/SVD.h"
-#endif
 
 #include "src/Core/util/ReenableStupidWarnings.h"
 

Eigen/Sparse

@@ -11,9 +11,9 @@
   * - \ref SparseQR_Module
   * - \ref IterativeLinearSolvers_Module
   *
-  * \code
-  * #include <Eigen/Sparse>
-  * \endcode
+    \code
+    #include <Eigen/Sparse>
+    \endcode
   */
 
 #include "SparseCore"

Eigen/SparseCholesky

@@ -34,8 +34,6 @@
 #error The SparseCholesky module has nothing to offer in MPL2 only mode
 #endif
 
-#include "src/misc/Solve.h"
-#include "src/misc/SparseSolve.h"
 #include "src/SparseCholesky/SimplicialCholesky.h"
 
 #ifndef EIGEN_MPL2_ONLY

Eigen/SparseCore

@@ -26,37 +26,35 @@
   * This module depends on: Core.
   */
 
-namespace Eigen {
-
-/** The type used to identify a general sparse storage. */
-struct Sparse {};
-
-}
-
 #include "src/SparseCore/SparseUtil.h"
 #include "src/SparseCore/SparseMatrixBase.h"
+#include "src/SparseCore/SparseAssign.h"
 #include "src/SparseCore/CompressedStorage.h"
 #include "src/SparseCore/AmbiVector.h"
+#include "src/SparseCore/SparseCompressedBase.h"
 #include "src/SparseCore/SparseMatrix.h"
+#include "src/SparseCore/SparseMap.h"
 #include "src/SparseCore/MappedSparseMatrix.h"
 #include "src/SparseCore/SparseVector.h"
-#include "src/SparseCore/SparseBlock.h"
-#include "src/SparseCore/SparseTranspose.h"
+#include "src/SparseCore/SparseRef.h"
 #include "src/SparseCore/SparseCwiseUnaryOp.h"
 #include "src/SparseCore/SparseCwiseBinaryOp.h"
+#include "src/SparseCore/SparseTranspose.h"
+#include "src/SparseCore/SparseBlock.h"
 #include "src/SparseCore/SparseDot.h"
-#include "src/SparseCore/SparsePermutation.h"
 #include "src/SparseCore/SparseRedux.h"
-#include "src/SparseCore/SparseFuzzy.h"
+#include "src/SparseCore/SparseView.h"
+#include "src/SparseCore/SparseDiagonalProduct.h"
 #include "src/SparseCore/ConservativeSparseSparseProduct.h"
 #include "src/SparseCore/SparseSparseProductWithPruning.h"
 #include "src/SparseCore/SparseProduct.h"
 #include "src/SparseCore/SparseDenseProduct.h"
-#include "src/SparseCore/SparseDiagonalProduct.h"
-#include "src/SparseCore/SparseTriangularView.h"
 #include "src/SparseCore/SparseSelfAdjointView.h"
+#include "src/SparseCore/SparseTriangularView.h"
 #include "src/SparseCore/TriangularSolver.h"
-#include "src/SparseCore/SparseView.h"
+#include "src/SparseCore/SparsePermutation.h"
+#include "src/SparseCore/SparseFuzzy.h"
+#include "src/SparseCore/SparseSolverBase.h"
 
 #include "src/Core/util/ReenableStupidWarnings.h"
 

Eigen/SparseLU

@@ -20,9 +20,6 @@
   * Please, see the documentation of the SparseLU class for more details.
   */
 
-#include "src/misc/Solve.h"
-#include "src/misc/SparseSolve.h"
-
 // Ordering interface
 #include "OrderingMethods"
 

Eigen/SparseQR

@@ -21,9 +21,6 @@
   * 
   */
 
-#include "src/misc/Solve.h"
-#include "src/misc/SparseSolve.h"
-
 #include "OrderingMethods"
 #include "src/SparseCore/SparseColEtree.h"
 #include "src/SparseQR/SparseQR.h"

Eigen/StdDeque

@@ -14,7 +14,7 @@
 #include "Core"
 #include <deque>
 
-#if (defined(_MSC_VER) && defined(_WIN64)) /* MSVC auto aligns in 64 bit builds */
+#if EIGEN_COMP_MSVC && EIGEN_OS_WIN64 /* MSVC auto aligns in 64 bit builds */
 
 #define EIGEN_DEFINE_STL_DEQUE_SPECIALIZATION(...)
 

Eigen/StdList

@@ -13,7 +13,7 @@
 #include "Core"
 #include <list>
 
-#if (defined(_MSC_VER) && defined(_WIN64)) /* MSVC auto aligns in 64 bit builds */    
+#if EIGEN_COMP_MSVC && EIGEN_OS_WIN64 /* MSVC auto aligns in 64 bit builds */    
 
 #define EIGEN_DEFINE_STL_LIST_SPECIALIZATION(...)
 

Eigen/StdVector

@@ -14,7 +14,7 @@
 #include "Core"
 #include <vector>
 
-#if (defined(_MSC_VER) && defined(_WIN64)) /* MSVC auto aligns in 64 bit builds */
+#if EIGEN_COMP_MSVC && EIGEN_OS_WIN64 /* MSVC auto aligns in 64 bit builds */
 
 #define EIGEN_DEFINE_STL_VECTOR_SPECIALIZATION(...)
 

Eigen/SuperLUSupport

@@ -48,12 +48,8 @@ namespace Eigen { struct SluMatrix; }
   *
   */
 
-#include "src/misc/Solve.h"
-#include "src/misc/SparseSolve.h"
-
 #include "src/SuperLUSupport/SuperLUSupport.h"
 
-
 #include "src/Core/util/ReenableStupidWarnings.h"
 
 #endif // EIGEN_SUPERLUSUPPORT_MODULE_H

Eigen/UmfPackSupport

@@ -26,9 +26,6 @@ extern "C" {
   *
   */
 
-#include "src/misc/Solve.h"
-#include "src/misc/SparseSolve.h"
-
 #include "src/UmfPackSupport/UmfPackSupport.h"
 
 #include "src/Core/util/ReenableStupidWarnings.h"

Eigen/src/Cholesky/LDLT.h

@@ -16,7 +16,10 @@
 namespace Eigen { 
 
 namespace internal {
-template<typename MatrixType, int UpLo> struct LDLT_Traits;
+  template<typename MatrixType, int UpLo> struct LDLT_Traits;
+
+  // PositiveSemiDef means positive semi-definite and non-zero; same for NegativeSemiDef
+  enum SignMatrix { PositiveSemiDef, NegativeSemiDef, ZeroSign, Indefinite };
 }
 
 /** \ingroup Cholesky_Module
@@ -40,7 +43,7 @@ template<typename MatrixType, int UpLo> struct LDLT_Traits;
   * Remember that Cholesky decompositions are not rank-revealing. Also, do not use a Cholesky
   * decomposition to determine whether a system of equations has a solution.
   *
-  * \sa MatrixBase::ldlt(), class LLT
+  * \sa MatrixBase::ldlt(), SelfAdjointView::ldlt(), class LLT
   */
 template<typename _MatrixType, int _UpLo> class LDLT
 {
@@ -56,7 +59,8 @@ template<typename _MatrixType, int _UpLo> class LDLT
     };
     typedef typename MatrixType::Scalar Scalar;
     typedef typename NumTraits<typename MatrixType::Scalar>::Real RealScalar;
-    typedef typename MatrixType::Index Index;
+    typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3
+    typedef typename MatrixType::StorageIndex StorageIndex;
     typedef Matrix<Scalar, RowsAtCompileTime, 1, Options, MaxRowsAtCompileTime, 1> TmpMatrixType;
 
     typedef Transpositions<RowsAtCompileTime, MaxRowsAtCompileTime> TranspositionType;
@@ -69,7 +73,12 @@ template<typename _MatrixType, int _UpLo> class LDLT
       * The default constructor is useful in cases in which the user intends to
       * perform decompositions via LDLT::compute(const MatrixType&).
       */
-    LDLT() : m_matrix(), m_transpositions(), m_isInitialized(false) {}
+    LDLT() 
+      : m_matrix(), 
+        m_transpositions(), 
+        m_sign(internal::ZeroSign),
+        m_isInitialized(false) 
+    {}
 
     /** \brief Default Constructor with memory preallocation
       *
@@ -77,10 +86,11 @@ template<typename _MatrixType, int _UpLo> class LDLT
       * according to the specified problem \a size.
       * \sa LDLT()
       */
-    LDLT(Index size)
+    explicit LDLT(Index size)
       : m_matrix(size, size),
         m_transpositions(size),
         m_temporary(size),
+        m_sign(internal::ZeroSign),
         m_isInitialized(false)
     {}
 
@@ -89,10 +99,11 @@ template<typename _MatrixType, int _UpLo> class LDLT
       * This calculates the decomposition for the input \a matrix.
       * \sa LDLT(Index size)
       */
-    LDLT(const MatrixType& matrix)
+    explicit LDLT(const MatrixType& matrix)
       : m_matrix(matrix.rows(), matrix.cols()),
         m_transpositions(matrix.rows()),
         m_temporary(matrix.rows()),
+        m_sign(internal::ZeroSign),
         m_isInitialized(false)
     {
       compute(matrix);
@@ -139,21 +150,14 @@ template<typename _MatrixType, int _UpLo> class LDLT
     inline bool isPositive() const
     {
       eigen_assert(m_isInitialized && "LDLT is not initialized.");
-      return m_sign == 1;
+      return m_sign == internal::PositiveSemiDef || m_sign == internal::ZeroSign;
     }
-    
-    #ifdef EIGEN2_SUPPORT
-    inline bool isPositiveDefinite() const
-    {
-      return isPositive();
-    }
-    #endif
 
     /** \returns true if the matrix is negative (semidefinite) */
     inline bool isNegative(void) const
     {
       eigen_assert(m_isInitialized && "LDLT is not initialized.");
-      return m_sign == -1;
+      return m_sign == internal::NegativeSemiDef || m_sign == internal::ZeroSign;
     }
 
     /** \returns a solution x of \f$ A x = b \f$ using the current decomposition of A.
@@ -169,27 +173,18 @@ template<typename _MatrixType, int _UpLo> class LDLT
       * least-square solution of \f$ D y_3 = y_2 \f$ is computed. This does not mean that this function
       * computes the least-square solution of \f$ A x = b \f$ is \f$ A \f$ is singular.
       *
-      * \sa MatrixBase::ldlt()
+      * \sa MatrixBase::ldlt(), SelfAdjointView::ldlt()
       */
     template<typename Rhs>
-    inline const internal::solve_retval<LDLT, Rhs>
+    inline const Solve<LDLT, Rhs>
     solve(const MatrixBase<Rhs>& b) const
     {
       eigen_assert(m_isInitialized && "LDLT is not initialized.");
       eigen_assert(m_matrix.rows()==b.rows()
                 && "LDLT::solve(): invalid number of rows of the right hand side matrix b");
-      return internal::solve_retval<LDLT, Rhs>(*this, b.derived());
+      return Solve<LDLT, Rhs>(*this, b.derived());
     }
 
-    #ifdef EIGEN2_SUPPORT
-    template<typename OtherDerived, typename ResultType>
-    bool solve(const MatrixBase<OtherDerived>& b, ResultType *result) const
-    {
-      *result = this->solve(b);
-      return true;
-    }
-    #endif
-
     template<typename Derived>
     bool solveInPlace(MatrixBase<Derived> &bAndX) const;
 
@@ -223,8 +218,19 @@ template<typename _MatrixType, int _UpLo> class LDLT
       eigen_assert(m_isInitialized && "LDLT is not initialized.");
       return Success;
     }
+    
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    template<typename RhsType, typename DstType>
+    EIGEN_DEVICE_FUNC
+    void _solve_impl(const RhsType &rhs, DstType &dst) const;
+    #endif
 
   protected:
+    
+    static void check_template_parameters()
+    {
+      EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar);
+    }
 
     /** \internal
       * Used to compute and store the Cholesky decomposition A = L D L^* = U^* D U.
@@ -235,7 +241,7 @@ template<typename _MatrixType, int _UpLo> class LDLT
     MatrixType m_matrix;
     TranspositionType m_transpositions;
     TmpMatrixType m_temporary;
-    int m_sign;
+    internal::SignMatrix m_sign;
     bool m_isInitialized;
 };
 
@@ -246,49 +252,32 @@ template<int UpLo> struct ldlt_inplace;
 template<> struct ldlt_inplace<Lower>
 {
   template<typename MatrixType, typename TranspositionType, typename Workspace>
-  static bool unblocked(MatrixType& mat, TranspositionType& transpositions, Workspace& temp, int* sign=0)
+  static bool unblocked(MatrixType& mat, TranspositionType& transpositions, Workspace& temp, SignMatrix& sign)
   {
     using std::abs;
     typedef typename MatrixType::Scalar Scalar;
     typedef typename MatrixType::RealScalar RealScalar;
-    typedef typename MatrixType::Index Index;
+    typedef typename TranspositionType::StorageIndex IndexType;
     eigen_assert(mat.rows()==mat.cols());
     const Index size = mat.rows();
 
     if (size <= 1)
     {
       transpositions.setIdentity();
-      if(sign)
-        *sign = numext::real(mat.coeff(0,0))>0 ? 1:-1;
+      if (numext::real(mat.coeff(0,0)) > 0) sign = PositiveSemiDef;
+      else if (numext::real(mat.coeff(0,0)) < 0) sign = NegativeSemiDef;
+      else sign = ZeroSign;
       return true;
     }
 
-    RealScalar cutoff(0), biggest_in_corner;
-
     for (Index k = 0; k < size; ++k)
     {
       // Find largest diagonal element
       Index index_of_biggest_in_corner;
-      biggest_in_corner = mat.diagonal().tail(size-k).cwiseAbs().maxCoeff(&index_of_biggest_in_corner);
+      mat.diagonal().tail(size-k).cwiseAbs().maxCoeff(&index_of_biggest_in_corner);
       index_of_biggest_in_corner += k;
 
-      if(k == 0)
-      {
-        // The biggest overall is the point of reference to which further diagonals
-        // are compared; if any diagonal is negligible compared
-        // to the largest overall, the algorithm bails.
-        cutoff = abs(NumTraits<Scalar>::epsilon() * biggest_in_corner);
-      }
-
-      // Finish early if the matrix is not full rank.
-      if(biggest_in_corner < cutoff)
-      {
-        for(Index i = k; i < size; i++) transpositions.coeffRef(i) = i;
-        if(sign) *sign = 0;
-        break;
-      }
-
-      transpositions.coeffRef(k) = index_of_biggest_in_corner;
+      transpositions.coeffRef(k) = IndexType(index_of_biggest_in_corner);
       if(k != index_of_biggest_in_corner)
       {
         // apply the transposition while taking care to consider only
@@ -297,7 +286,7 @@ template<> struct ldlt_inplace<Lower>
         mat.row(k).head(k).swap(mat.row(index_of_biggest_in_corner).head(k));
         mat.col(k).tail(s).swap(mat.col(index_of_biggest_in_corner).tail(s));
         std::swap(mat.coeffRef(k,k),mat.coeffRef(index_of_biggest_in_corner,index_of_biggest_in_corner));
-        for(int i=k+1;i<index_of_biggest_in_corner;++i)
+        for(Index i=k+1;i<index_of_biggest_in_corner;++i)
         {
           Scalar tmp = mat.coeffRef(i,k);
           mat.coeffRef(i,k) = numext::conj(mat.coeffRef(index_of_biggest_in_corner,i));
@@ -318,22 +307,27 @@ template<> struct ldlt_inplace<Lower>
 
       if(k>0)
       {
-        temp.head(k) = mat.diagonal().head(k).asDiagonal() * A10.adjoint();
+        temp.head(k) = mat.diagonal().real().head(k).asDiagonal() * A10.adjoint();
         mat.coeffRef(k,k) -= (A10 * temp.head(k)).value();
         if(rs>0)
           A21.noalias() -= A20 * temp.head(k);
       }
-      if((rs>0) && (abs(mat.coeffRef(k,k)) > cutoff))
-        A21 /= mat.coeffRef(k,k);
       
-      if(sign)
-      {
-        // LDLT is not guaranteed to work for indefinite matrices, but let's try to get the sign right
-        int newSign = numext::real(mat.diagonal().coeff(index_of_biggest_in_corner)) > 0;
-        if(k == 0)
-          *sign = newSign;
-        else if(*sign != newSign)
-          *sign = 0;
+      // In some previous versions of Eigen (e.g., 3.2.1), the scaling was omitted if the pivot
+      // was smaller than the cutoff value. However, since LDLT is not rank-revealing
+      // we should only make sure that we do not introduce INF or NaN values.
+      // Remark that LAPACK also uses 0 as the cutoff value.
+      RealScalar realAkk = numext::real(mat.coeffRef(k,k));
+      if((rs>0) && (abs(realAkk) > RealScalar(0)))
+        A21 /= realAkk;
+
+      if (sign == PositiveSemiDef) {
+        if (realAkk < 0) sign = Indefinite;
+      } else if (sign == NegativeSemiDef) {
+        if (realAkk > 0) sign = Indefinite;
+      } else if (sign == ZeroSign) {
+        if (realAkk > 0) sign = PositiveSemiDef;
+        else if (realAkk < 0) sign = NegativeSemiDef;
       }
     }
 
@@ -353,7 +347,6 @@ template<> struct ldlt_inplace<Lower>
     using numext::isfinite;
     typedef typename MatrixType::Scalar Scalar;
     typedef typename MatrixType::RealScalar RealScalar;
-    typedef typename MatrixType::Index Index;
 
     const Index size = mat.rows();
     eigen_assert(mat.cols() == size && w.size()==size);
@@ -399,7 +392,7 @@ template<> struct ldlt_inplace<Lower>
 template<> struct ldlt_inplace<Upper>
 {
   template<typename MatrixType, typename TranspositionType, typename Workspace>
-  static EIGEN_STRONG_INLINE bool unblocked(MatrixType& mat, TranspositionType& transpositions, Workspace& temp, int* sign=0)
+  static EIGEN_STRONG_INLINE bool unblocked(MatrixType& mat, TranspositionType& transpositions, Workspace& temp, SignMatrix& sign)
   {
     Transpose<MatrixType> matt(mat);
     return ldlt_inplace<Lower>::unblocked(matt, transpositions, temp, sign);
@@ -417,16 +410,16 @@ template<typename MatrixType> struct LDLT_Traits<MatrixType,Lower>
 {
   typedef const TriangularView<const MatrixType, UnitLower> MatrixL;
   typedef const TriangularView<const typename MatrixType::AdjointReturnType, UnitUpper> MatrixU;
-  static inline MatrixL getL(const MatrixType& m) { return m; }
-  static inline MatrixU getU(const MatrixType& m) { return m.adjoint(); }
+  static inline MatrixL getL(const MatrixType& m) { return MatrixL(m); }
+  static inline MatrixU getU(const MatrixType& m) { return MatrixU(m.adjoint()); }
 };
 
 template<typename MatrixType> struct LDLT_Traits<MatrixType,Upper>
 {
   typedef const TriangularView<const typename MatrixType::AdjointReturnType, UnitLower> MatrixL;
   typedef const TriangularView<const MatrixType, UnitUpper> MatrixU;
-  static inline MatrixL getL(const MatrixType& m) { return m.adjoint(); }
-  static inline MatrixU getU(const MatrixType& m) { return m; }
+  static inline MatrixL getL(const MatrixType& m) { return MatrixL(m.adjoint()); }
+  static inline MatrixU getU(const MatrixType& m) { return MatrixU(m); }
 };
 
 } // end namespace internal
@@ -436,6 +429,8 @@ template<typename MatrixType> struct LDLT_Traits<MatrixType,Upper>
 template<typename MatrixType, int _UpLo>
 LDLT<MatrixType,_UpLo>& LDLT<MatrixType,_UpLo>::compute(const MatrixType& a)
 {
+  check_template_parameters();
+  
   eigen_assert(a.rows()==a.cols());
   const Index size = a.rows();
 
@@ -444,8 +439,9 @@ LDLT<MatrixType,_UpLo>& LDLT<MatrixType,_UpLo>::compute(const MatrixType& a)
   m_transpositions.resize(size);
   m_isInitialized = false;
   m_temporary.resize(size);
+  m_sign = internal::ZeroSign;
 
-  internal::ldlt_inplace<UpLo>::unblocked(m_matrix, m_transpositions, m_temporary, &m_sign);
+  internal::ldlt_inplace<UpLo>::unblocked(m_matrix, m_transpositions, m_temporary, m_sign);
 
   m_isInitialized = true;
   return *this;
@@ -458,8 +454,9 @@ LDLT<MatrixType,_UpLo>& LDLT<MatrixType,_UpLo>::compute(const MatrixType& a)
   */
 template<typename MatrixType, int _UpLo>
 template<typename Derived>
-LDLT<MatrixType,_UpLo>& LDLT<MatrixType,_UpLo>::rankUpdate(const MatrixBase<Derived>& w, const typename NumTraits<typename MatrixType::Scalar>::Real& sigma)
+LDLT<MatrixType,_UpLo>& LDLT<MatrixType,_UpLo>::rankUpdate(const MatrixBase<Derived>& w, const typename LDLT<MatrixType,_UpLo>::RealScalar& sigma)
 {
+  typedef typename TranspositionType::StorageIndex IndexType;
   const Index size = w.rows();
   if (m_isInitialized)
   {
@@ -471,9 +468,9 @@ LDLT<MatrixType,_UpLo>& LDLT<MatrixType,_UpLo>::rankUpdate(const MatrixBase<Deri
     m_matrix.setZero();
     m_transpositions.resize(size);
     for (Index i = 0; i < size; i++)
-      m_transpositions.coeffRef(i) = i;
+      m_transpositions.coeffRef(i) = IndexType(i);
     m_temporary.resize(size);
-    m_sign = sigma>=0 ? 1 : -1;
+    m_sign = sigma>=0 ? internal::PositiveSemiDef : internal::NegativeSemiDef;
     m_isInitialized = true;
   }
 
@@ -482,48 +479,45 @@ LDLT<MatrixType,_UpLo>& LDLT<MatrixType,_UpLo>::rankUpdate(const MatrixBase<Deri
   return *this;
 }
 
-namespace internal {
-template<typename _MatrixType, int _UpLo, typename Rhs>
-struct solve_retval<LDLT<_MatrixType,_UpLo>, Rhs>
-  : solve_retval_base<LDLT<_MatrixType,_UpLo>, Rhs>
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+template<typename _MatrixType, int _UpLo>
+template<typename RhsType, typename DstType>
+void LDLT<_MatrixType,_UpLo>::_solve_impl(const RhsType &rhs, DstType &dst) const
 {
-  typedef LDLT<_MatrixType,_UpLo> LDLTType;
-  EIGEN_MAKE_SOLVE_HELPERS(LDLTType,Rhs)
+  eigen_assert(rhs.rows() == rows());
+  // dst = P b
+  dst = m_transpositions * rhs;
 
-  template<typename Dest> void evalTo(Dest& dst) const
+  // dst = L^-1 (P b)
+  matrixL().solveInPlace(dst);
+
+  // dst = D^-1 (L^-1 P b)
+  // more precisely, use pseudo-inverse of D (see bug 241)
+  using std::abs;
+  const typename Diagonal<const MatrixType>::RealReturnType vecD(vectorD());
+  // In some previous versions, tolerance was set to the max of 1/highest and the maximal diagonal entry * epsilon
+  // as motivated by LAPACK's xGELSS:
+  // RealScalar tolerance = numext::maxi(vecD.array().abs().maxCoeff() * NumTraits<RealScalar>::epsilon(),RealScalar(1) / NumTraits<RealScalar>::highest());
+  // However, LDLT is not rank revealing, and so adjusting the tolerance wrt to the highest
+  // diagonal element is not well justified and leads to numerical issues in some cases.
+  // Moreover, Lapack's xSYTRS routines use 0 for the tolerance.
+  RealScalar tolerance = RealScalar(1) / NumTraits<RealScalar>::highest();
+  
+  for (Index i = 0; i < vecD.size(); ++i)
   {
-    eigen_assert(rhs().rows() == dec().matrixLDLT().rows());
-    // dst = P b
-    dst = dec().transpositionsP() * rhs();
-
-    // dst = L^-1 (P b)
-    dec().matrixL().solveInPlace(dst);
-
-    // dst = D^-1 (L^-1 P b)
-    // more precisely, use pseudo-inverse of D (see bug 241)
-    using std::abs;
-    using std::max;
-    typedef typename LDLTType::MatrixType MatrixType;
-    typedef typename LDLTType::Scalar Scalar;
-    typedef typename LDLTType::RealScalar RealScalar;
-    const Diagonal<const MatrixType> vectorD = dec().vectorD();
-    RealScalar tolerance = (max)(vectorD.array().abs().maxCoeff() * NumTraits<Scalar>::epsilon(),
-				 RealScalar(1) / NumTraits<RealScalar>::highest()); // motivated by LAPACK's xGELSS
-    for (Index i = 0; i < vectorD.size(); ++i) {
-      if(abs(vectorD(i)) > tolerance)
-	dst.row(i) /= vectorD(i);
-      else
-	dst.row(i).setZero();
-    }
-
-    // dst = L^-T (D^-1 L^-1 P b)
-    dec().matrixU().solveInPlace(dst);
-
-    // dst = P^-1 (L^-T D^-1 L^-1 P b) = A^-1 b
-    dst = dec().transpositionsP().transpose() * dst;
+    if(abs(vecD(i)) > tolerance)
+      dst.row(i) /= vecD(i);
+    else
+      dst.row(i).setZero();
   }
-};
+
+  // dst = L^-T (D^-1 L^-1 P b)
+  matrixU().solveInPlace(dst);
+
+  // dst = P^-1 (L^-T D^-1 L^-1 P b) = A^-1 b
+  dst = m_transpositions.transpose() * dst;
 }
+#endif
 
 /** \internal use x = ldlt_object.solve(x);
   *
@@ -566,7 +560,7 @@ MatrixType LDLT<MatrixType,_UpLo>::reconstructedMatrix() const
   // L^* P
   res = matrixU() * res;
   // D(L^*P)
-  res = vectorD().asDiagonal() * res;
+  res = vectorD().real().asDiagonal() * res;
   // L(DL^*P)
   res = matrixL() * res;
   // P^T (LDL^*P)
@@ -575,8 +569,10 @@ MatrixType LDLT<MatrixType,_UpLo>::reconstructedMatrix() const
   return res;
 }
 
+#ifndef __CUDACC__
 /** \cholesky_module
   * \returns the Cholesky decomposition with full pivoting without square root of \c *this
+  * \sa MatrixBase::ldlt()
   */
 template<typename MatrixType, unsigned int UpLo>
 inline const LDLT<typename SelfAdjointView<MatrixType, UpLo>::PlainObject, UpLo>
@@ -587,6 +583,7 @@ SelfAdjointView<MatrixType, UpLo>::ldlt() const
 
 /** \cholesky_module
   * \returns the Cholesky decomposition with full pivoting without square root of \c *this
+  * \sa SelfAdjointView::ldlt()
   */
 template<typename Derived>
 inline const LDLT<typename MatrixBase<Derived>::PlainObject>
@@ -594,6 +591,7 @@ MatrixBase<Derived>::ldlt() const
 {
   return LDLT<PlainObject>(derived());
 }
+#endif // __CUDACC__
 
 } // end namespace Eigen
 

Eigen/src/Cholesky/LLT.h

@@ -41,7 +41,7 @@ template<typename MatrixType, int UpLo> struct LLT_Traits;
   * Example: \include LLT_example.cpp
   * Output: \verbinclude LLT_example.out
   *    
-  * \sa MatrixBase::llt(), class LDLT
+  * \sa MatrixBase::llt(), SelfAdjointView::llt(), class LDLT
   */
  /* HEY THIS DOX IS DISABLED BECAUSE THERE's A BUG EITHER HERE OR IN LDLT ABOUT THAT (OR BOTH)
   * Note that during the decomposition, only the upper triangular part of A is considered. Therefore,
@@ -59,7 +59,8 @@ template<typename _MatrixType, int _UpLo> class LLT
     };
     typedef typename MatrixType::Scalar Scalar;
     typedef typename NumTraits<typename MatrixType::Scalar>::Real RealScalar;
-    typedef typename MatrixType::Index Index;
+    typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3
+    typedef typename MatrixType::StorageIndex StorageIndex;
 
     enum {
       PacketSize = internal::packet_traits<Scalar>::size,
@@ -83,10 +84,10 @@ template<typename _MatrixType, int _UpLo> class LLT
       * according to the specified problem \a size.
       * \sa LLT()
       */
-    LLT(Index size) : m_matrix(size, size),
+    explicit LLT(Index size) : m_matrix(size, size),
                     m_isInitialized(false) {}
 
-    LLT(const MatrixType& matrix)
+    explicit LLT(const MatrixType& matrix)
       : m_matrix(matrix.rows(), matrix.cols()),
         m_isInitialized(false)
     {
@@ -115,29 +116,18 @@ template<typename _MatrixType, int _UpLo> class LLT
       * Example: \include LLT_solve.cpp
       * Output: \verbinclude LLT_solve.out
       *
-      * \sa solveInPlace(), MatrixBase::llt()
+      * \sa solveInPlace(), MatrixBase::llt(), SelfAdjointView::llt()
       */
     template<typename Rhs>
-    inline const internal::solve_retval<LLT, Rhs>
+    inline const Solve<LLT, Rhs>
     solve(const MatrixBase<Rhs>& b) const
     {
       eigen_assert(m_isInitialized && "LLT is not initialized.");
       eigen_assert(m_matrix.rows()==b.rows()
                 && "LLT::solve(): invalid number of rows of the right hand side matrix b");
-      return internal::solve_retval<LLT, Rhs>(*this, b.derived());
+      return Solve<LLT, Rhs>(*this, b.derived());
     }
 
-    #ifdef EIGEN2_SUPPORT
-    template<typename OtherDerived, typename ResultType>
-    bool solve(const MatrixBase<OtherDerived>& b, ResultType *result) const
-    {
-      *result = this->solve(b);
-      return true;
-    }
-    
-    bool isPositiveDefinite() const { return true; }
-    #endif
-
     template<typename Derived>
     void solveInPlace(MatrixBase<Derived> &bAndX) const;
 
@@ -172,8 +162,20 @@ template<typename _MatrixType, int _UpLo> class LLT
 
     template<typename VectorType>
     LLT rankUpdate(const VectorType& vec, const RealScalar& sigma = 1);
+    
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    template<typename RhsType, typename DstType>
+    EIGEN_DEVICE_FUNC
+    void _solve_impl(const RhsType &rhs, DstType &dst) const;
+    #endif
 
   protected:
+    
+    static void check_template_parameters()
+    {
+      EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar);
+    }
+    
     /** \internal
       * Used to compute and store L
       * The strict upper part is not used and even not initialized.
@@ -188,12 +190,11 @@ namespace internal {
 template<typename Scalar, int UpLo> struct llt_inplace;
 
 template<typename MatrixType, typename VectorType>
-static typename MatrixType::Index llt_rank_update_lower(MatrixType& mat, const VectorType& vec, const typename MatrixType::RealScalar& sigma)
+static Index llt_rank_update_lower(MatrixType& mat, const VectorType& vec, const typename MatrixType::RealScalar& sigma)
 {
   using std::sqrt;
   typedef typename MatrixType::Scalar Scalar;
   typedef typename MatrixType::RealScalar RealScalar;
-  typedef typename MatrixType::Index Index;
   typedef typename MatrixType::ColXpr ColXpr;
   typedef typename internal::remove_all<ColXpr>::type ColXprCleaned;
   typedef typename ColXprCleaned::SegmentReturnType ColXprSegment;
@@ -262,10 +263,9 @@ template<typename Scalar> struct llt_inplace<Scalar, Lower>
 {
   typedef typename NumTraits<Scalar>::Real RealScalar;
   template<typename MatrixType>
-  static typename MatrixType::Index unblocked(MatrixType& mat)
+  static Index unblocked(MatrixType& mat)
   {
     using std::sqrt;
-    typedef typename MatrixType::Index Index;
     
     eigen_assert(mat.rows()==mat.cols());
     const Index size = mat.rows();
@@ -289,9 +289,8 @@ template<typename Scalar> struct llt_inplace<Scalar, Lower>
   }
 
   template<typename MatrixType>
-  static typename MatrixType::Index blocked(MatrixType& m)
+  static Index blocked(MatrixType& m)
   {
-    typedef typename MatrixType::Index Index;
     eigen_assert(m.rows()==m.cols());
     Index size = m.rows();
     if(size<32)
@@ -322,7 +321,7 @@ template<typename Scalar> struct llt_inplace<Scalar, Lower>
   }
 
   template<typename MatrixType, typename VectorType>
-  static typename MatrixType::Index rankUpdate(MatrixType& mat, const VectorType& vec, const RealScalar& sigma)
+  static Index rankUpdate(MatrixType& mat, const VectorType& vec, const RealScalar& sigma)
   {
     return Eigen::internal::llt_rank_update_lower(mat, vec, sigma);
   }
@@ -333,19 +332,19 @@ template<typename Scalar> struct llt_inplace<Scalar, Upper>
   typedef typename NumTraits<Scalar>::Real RealScalar;
 
   template<typename MatrixType>
-  static EIGEN_STRONG_INLINE typename MatrixType::Index unblocked(MatrixType& mat)
+  static EIGEN_STRONG_INLINE Index unblocked(MatrixType& mat)
   {
     Transpose<MatrixType> matt(mat);
     return llt_inplace<Scalar, Lower>::unblocked(matt);
   }
   template<typename MatrixType>
-  static EIGEN_STRONG_INLINE typename MatrixType::Index blocked(MatrixType& mat)
+  static EIGEN_STRONG_INLINE Index blocked(MatrixType& mat)
   {
     Transpose<MatrixType> matt(mat);
     return llt_inplace<Scalar, Lower>::blocked(matt);
   }
   template<typename MatrixType, typename VectorType>
-  static typename MatrixType::Index rankUpdate(MatrixType& mat, const VectorType& vec, const RealScalar& sigma)
+  static Index rankUpdate(MatrixType& mat, const VectorType& vec, const RealScalar& sigma)
   {
     Transpose<MatrixType> matt(mat);
     return llt_inplace<Scalar, Lower>::rankUpdate(matt, vec.conjugate(), sigma);
@@ -356,8 +355,8 @@ template<typename MatrixType> struct LLT_Traits<MatrixType,Lower>
 {
   typedef const TriangularView<const MatrixType, Lower> MatrixL;
   typedef const TriangularView<const typename MatrixType::AdjointReturnType, Upper> MatrixU;
-  static inline MatrixL getL(const MatrixType& m) { return m; }
-  static inline MatrixU getU(const MatrixType& m) { return m.adjoint(); }
+  static inline MatrixL getL(const MatrixType& m) { return MatrixL(m); }
+  static inline MatrixU getU(const MatrixType& m) { return MatrixU(m.adjoint()); }
   static bool inplace_decomposition(MatrixType& m)
   { return llt_inplace<typename MatrixType::Scalar, Lower>::blocked(m)==-1; }
 };
@@ -366,8 +365,8 @@ template<typename MatrixType> struct LLT_Traits<MatrixType,Upper>
 {
   typedef const TriangularView<const typename MatrixType::AdjointReturnType, Lower> MatrixL;
   typedef const TriangularView<const MatrixType, Upper> MatrixU;
-  static inline MatrixL getL(const MatrixType& m) { return m.adjoint(); }
-  static inline MatrixU getU(const MatrixType& m) { return m; }
+  static inline MatrixL getL(const MatrixType& m) { return MatrixL(m.adjoint()); }
+  static inline MatrixU getU(const MatrixType& m) { return MatrixU(m); }
   static bool inplace_decomposition(MatrixType& m)
   { return llt_inplace<typename MatrixType::Scalar, Upper>::blocked(m)==-1; }
 };
@@ -384,6 +383,8 @@ template<typename MatrixType> struct LLT_Traits<MatrixType,Upper>
 template<typename MatrixType, int _UpLo>
 LLT<MatrixType,_UpLo>& LLT<MatrixType,_UpLo>::compute(const MatrixType& a)
 {
+  check_template_parameters();
+  
   eigen_assert(a.rows()==a.cols());
   const Index size = a.rows();
   m_matrix.resize(size, size);
@@ -415,22 +416,16 @@ LLT<_MatrixType,_UpLo> LLT<_MatrixType,_UpLo>::rankUpdate(const VectorType& v, c
 
   return *this;
 }
-    
-namespace internal {
-template<typename _MatrixType, int UpLo, typename Rhs>
-struct solve_retval<LLT<_MatrixType, UpLo>, Rhs>
-  : solve_retval_base<LLT<_MatrixType, UpLo>, Rhs>
+ 
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+template<typename _MatrixType,int _UpLo>
+template<typename RhsType, typename DstType>
+void LLT<_MatrixType,_UpLo>::_solve_impl(const RhsType &rhs, DstType &dst) const
 {
-  typedef LLT<_MatrixType,UpLo> LLTType;
-  EIGEN_MAKE_SOLVE_HELPERS(LLTType,Rhs)
-
-  template<typename Dest> void evalTo(Dest& dst) const
-  {
-    dst = rhs();
-    dec().solveInPlace(dst);
-  }
-};
+  dst = rhs;
+  solveInPlace(dst);
 }
+#endif
 
 /** \internal use x = llt_object.solve(x);
   * 
@@ -465,8 +460,10 @@ MatrixType LLT<MatrixType,_UpLo>::reconstructedMatrix() const
   return matrixL() * matrixL().adjoint().toDenseMatrix();
 }
 
+#ifndef __CUDACC__
 /** \cholesky_module
   * \returns the LLT decomposition of \c *this
+  * \sa SelfAdjointView::llt()
   */
 template<typename Derived>
 inline const LLT<typename MatrixBase<Derived>::PlainObject>
@@ -477,6 +474,7 @@ MatrixBase<Derived>::llt() const
 
 /** \cholesky_module
   * \returns the LLT decomposition of \c *this
+  * \sa SelfAdjointView::llt()
   */
 template<typename MatrixType, unsigned int UpLo>
 inline const LLT<typename SelfAdjointView<MatrixType, UpLo>::PlainObject, UpLo>
@@ -484,7 +482,8 @@ SelfAdjointView<MatrixType, UpLo>::llt() const
 {
   return LLT<PlainObject,UpLo>(m_matrix);
 }
-
+#endif // __CUDACC__
+  
 } // end namespace Eigen
 
 #endif // EIGEN_LLT_H

Eigen/src/Cholesky/LLT_MKL.h

@@ -46,7 +46,7 @@ template<typename Scalar> struct mkl_llt;
 template<> struct mkl_llt<EIGTYPE> \
 { \
   template<typename MatrixType> \
-  static inline typename MatrixType::Index potrf(MatrixType& m, char uplo) \
+  static inline Index potrf(MatrixType& m, char uplo) \
   { \
     lapack_int matrix_order; \
     lapack_int size, lda, info, StorageOrder; \
@@ -60,30 +60,30 @@ template<> struct mkl_llt<EIGTYPE> \
     lda = m.outerStride(); \
 \
     info = LAPACKE_##MKLPREFIX##potrf( matrix_order, uplo, size, (MKLTYPE*)a, lda ); \
-    info = (info==0) ? Success : NumericalIssue; \
+    info = (info==0) ? -1 : info>0 ? info-1 : size; \
     return info; \
   } \
 }; \
 template<> struct llt_inplace<EIGTYPE, Lower> \
 { \
   template<typename MatrixType> \
-  static typename MatrixType::Index blocked(MatrixType& m) \
+  static Index blocked(MatrixType& m) \
   { \
     return mkl_llt<EIGTYPE>::potrf(m, 'L'); \
   } \
   template<typename MatrixType, typename VectorType> \
-  static typename MatrixType::Index rankUpdate(MatrixType& mat, const VectorType& vec, const typename MatrixType::RealScalar& sigma) \
+  static Index rankUpdate(MatrixType& mat, const VectorType& vec, const typename MatrixType::RealScalar& sigma) \
   { return Eigen::internal::llt_rank_update_lower(mat, vec, sigma); } \
 }; \
 template<> struct llt_inplace<EIGTYPE, Upper> \
 { \
   template<typename MatrixType> \
-  static typename MatrixType::Index blocked(MatrixType& m) \
+  static Index blocked(MatrixType& m) \
   { \
     return mkl_llt<EIGTYPE>::potrf(m, 'U'); \
   } \
   template<typename MatrixType, typename VectorType> \
-  static typename MatrixType::Index rankUpdate(MatrixType& mat, const VectorType& vec, const typename MatrixType::RealScalar& sigma) \
+  static Index rankUpdate(MatrixType& mat, const VectorType& vec, const typename MatrixType::RealScalar& sigma) \
   { \
     Transpose<MatrixType> matt(mat); \
     return llt_inplace<EIGTYPE, Lower>::rankUpdate(matt, vec.conjugate(), sigma); \

Eigen/src/CholmodSupport/CholmodSupport.h

@@ -48,8 +48,8 @@ void cholmod_configure_matrix(CholmodType& mat)
 /** Wraps the Eigen sparse matrix \a mat into a Cholmod sparse matrix object.
   * Note that the data are shared.
   */
-template<typename _Scalar, int _Options, typename _Index>
-cholmod_sparse viewAsCholmod(SparseMatrix<_Scalar,_Options,_Index>& mat)
+template<typename _Scalar, int _Options, typename _StorageIndex>
+cholmod_sparse viewAsCholmod(SparseMatrix<_Scalar,_Options,_StorageIndex>& mat)
 {
   cholmod_sparse res;
   res.nzmax   = mat.nonZeros();
@@ -58,10 +58,12 @@ cholmod_sparse viewAsCholmod(SparseMatrix<_Scalar,_Options,_Index>& mat)
   res.p       = mat.outerIndexPtr();
   res.i       = mat.innerIndexPtr();
   res.x       = mat.valuePtr();
+  res.z       = 0;
   res.sorted  = 1;
   if(mat.isCompressed())
   {
     res.packed  = 1;
+    res.nz = 0;
   }
   else
   {
@@ -72,11 +74,11 @@ cholmod_sparse viewAsCholmod(SparseMatrix<_Scalar,_Options,_Index>& mat)
   res.dtype   = 0;
   res.stype   = -1;
   
-  if (internal::is_same<_Index,int>::value)
+  if (internal::is_same<_StorageIndex,int>::value)
   {
     res.itype = CHOLMOD_INT;
   }
-  else if (internal::is_same<_Index,UF_long>::value)
+  else if (internal::is_same<_StorageIndex,UF_long>::value)
   {
     res.itype = CHOLMOD_LONG;
   }
@@ -103,7 +105,7 @@ const cholmod_sparse viewAsCholmod(const SparseMatrix<_Scalar,_Options,_Index>&
 /** Returns a view of the Eigen sparse matrix \a mat as Cholmod sparse matrix.
   * The data are not copied but shared. */
 template<typename _Scalar, int _Options, typename _Index, unsigned int UpLo>
-cholmod_sparse viewAsCholmod(const SparseSelfAdjointView<SparseMatrix<_Scalar,_Options,_Index>, UpLo>& mat)
+cholmod_sparse viewAsCholmod(const SparseSelfAdjointView<const SparseMatrix<_Scalar,_Options,_Index>, UpLo>& mat)
 {
   cholmod_sparse res = viewAsCholmod(mat.matrix().const_cast_derived());
   
@@ -136,12 +138,12 @@ cholmod_dense viewAsCholmod(MatrixBase<Derived>& mat)
 
 /** Returns a view of the Cholmod sparse matrix \a cm as an Eigen sparse matrix.
   * The data are not copied but shared. */
-template<typename Scalar, int Flags, typename Index>
-MappedSparseMatrix<Scalar,Flags,Index> viewAsEigen(cholmod_sparse& cm)
+template<typename Scalar, int Flags, typename StorageIndex>
+MappedSparseMatrix<Scalar,Flags,StorageIndex> viewAsEigen(cholmod_sparse& cm)
 {
-  return MappedSparseMatrix<Scalar,Flags,Index>
-         (cm.nrow, cm.ncol, static_cast<Index*>(cm.p)[cm.ncol],
-          static_cast<Index*>(cm.p), static_cast<Index*>(cm.i),static_cast<Scalar*>(cm.x) );
+  return MappedSparseMatrix<Scalar,Flags,StorageIndex>
+         (cm.nrow, cm.ncol, static_cast<StorageIndex*>(cm.p)[cm.ncol],
+          static_cast<StorageIndex*>(cm.p), static_cast<StorageIndex*>(cm.i),static_cast<Scalar*>(cm.x) );
 }
 
 enum CholmodMode {
@@ -155,26 +157,31 @@ enum CholmodMode {
   * \sa class CholmodSupernodalLLT, class CholmodSimplicialLDLT, class CholmodSimplicialLLT
   */
 template<typename _MatrixType, int _UpLo, typename Derived>
-class CholmodBase : internal::noncopyable
+class CholmodBase : public SparseSolverBase<Derived>
 {
+  protected:
+    typedef SparseSolverBase<Derived> Base;
+    using Base::derived;
+    using Base::m_isInitialized;
   public:
     typedef _MatrixType MatrixType;
     enum { UpLo = _UpLo };
     typedef typename MatrixType::Scalar Scalar;
     typedef typename MatrixType::RealScalar RealScalar;
     typedef MatrixType CholMatrixType;
-    typedef typename MatrixType::Index Index;
+    typedef typename MatrixType::StorageIndex StorageIndex;
 
   public:
 
     CholmodBase()
-      : m_cholmodFactor(0), m_info(Success), m_isInitialized(false)
+      : m_cholmodFactor(0), m_info(Success)
     {
+      m_shiftOffset[0] = m_shiftOffset[1] = RealScalar(0.0);
       cholmod_start(&m_cholmod);
     }
 
-    CholmodBase(const MatrixType& matrix)
-      : m_cholmodFactor(0), m_info(Success), m_isInitialized(false)
+    explicit CholmodBase(const MatrixType& matrix)
+      : m_cholmodFactor(0), m_info(Success)
     {
       m_shiftOffset[0] = m_shiftOffset[1] = RealScalar(0.0);
       cholmod_start(&m_cholmod);
@@ -188,11 +195,8 @@ class CholmodBase : internal::noncopyable
       cholmod_finish(&m_cholmod);
     }
     
-    inline Index cols() const { return m_cholmodFactor->n; }
-    inline Index rows() const { return m_cholmodFactor->n; }
-    
-    Derived& derived() { return *static_cast<Derived*>(this); }
-    const Derived& derived() const { return *static_cast<const Derived*>(this); }
+    inline StorageIndex cols() const { return internal::convert_index<StorageIndex, Index>(m_cholmodFactor->n); }
+    inline StorageIndex rows() const { return internal::convert_index<StorageIndex, Index>(m_cholmodFactor->n); }
     
     /** \brief Reports whether previous computation was successful.
       *
@@ -213,35 +217,7 @@ class CholmodBase : internal::noncopyable
       return derived();
     }
     
-    /** \returns the solution x of \f$ A x = b \f$ using the current decomposition of A.
-      *
-      * \sa compute()
-      */
-    template<typename Rhs>
-    inline const internal::solve_retval<CholmodBase, Rhs>
-    solve(const MatrixBase<Rhs>& b) const
-    {
-      eigen_assert(m_isInitialized && "LLT is not initialized.");
-      eigen_assert(rows()==b.rows()
-                && "CholmodDecomposition::solve(): invalid number of rows of the right hand side matrix b");
-      return internal::solve_retval<CholmodBase, Rhs>(*this, b.derived());
-    }
-    
-    /** \returns the solution x of \f$ A x = b \f$ using the current decomposition of A.
-      *
-      * \sa compute()
-      */
-    template<typename Rhs>
-    inline const internal::sparse_solve_retval<CholmodBase, Rhs>
-    solve(const SparseMatrixBase<Rhs>& b) const
-    {
-      eigen_assert(m_isInitialized && "LLT is not initialized.");
-      eigen_assert(rows()==b.rows()
-                && "CholmodDecomposition::solve(): invalid number of rows of the right hand side matrix b");
-      return internal::sparse_solve_retval<CholmodBase, Rhs>(*this, b.derived());
-    }
-    
-    /** Performs a symbolic decomposition on the sparcity of \a matrix.
+    /** Performs a symbolic decomposition on the sparsity pattern of \a matrix.
       *
       * This function is particularly useful when solving for several problems having the same structure.
       * 
@@ -265,7 +241,7 @@ class CholmodBase : internal::noncopyable
     
     /** Performs a numeric decomposition of \a matrix
       *
-      * The given matrix must has the same sparcity than the matrix on which the symbolic decomposition has been performed.
+      * The given matrix must have the same sparsity pattern as the matrix on which the symbolic decomposition has been performed.
       *
       * \sa analyzePattern()
       */
@@ -287,7 +263,7 @@ class CholmodBase : internal::noncopyable
     #ifndef EIGEN_PARSED_BY_DOXYGEN
     /** \internal */
     template<typename Rhs,typename Dest>
-    void _solve(const MatrixBase<Rhs> &b, MatrixBase<Dest> &dest) const
+    void _solve_impl(const MatrixBase<Rhs> &b, MatrixBase<Dest> &dest) const
     {
       eigen_assert(m_factorizationIsOk && "The decomposition is not in a valid state for solving, you must first call either compute() or symbolic()/numeric()");
       const Index size = m_cholmodFactor->n;
@@ -301,15 +277,16 @@ class CholmodBase : internal::noncopyable
       if(!x_cd)
       {
         this->m_info = NumericalIssue;
+        return;
       }
-      // TODO optimize this copy by swapping when possible (be carreful with alignment, etc.)
+      // TODO optimize this copy by swapping when possible (be careful with alignment, etc.)
       dest = Matrix<Scalar,Dest::RowsAtCompileTime,Dest::ColsAtCompileTime>::Map(reinterpret_cast<Scalar*>(x_cd->x),b.rows(),b.cols());
       cholmod_free_dense(&x_cd, &m_cholmod);
     }
     
     /** \internal */
     template<typename RhsScalar, int RhsOptions, typename RhsIndex, typename DestScalar, int DestOptions, typename DestIndex>
-    void _solve(const SparseMatrix<RhsScalar,RhsOptions,RhsIndex> &b, SparseMatrix<DestScalar,DestOptions,DestIndex> &dest) const
+    void _solve_impl(const SparseMatrix<RhsScalar,RhsOptions,RhsIndex> &b, SparseMatrix<DestScalar,DestOptions,DestIndex> &dest) const
     {
       eigen_assert(m_factorizationIsOk && "The decomposition is not in a valid state for solving, you must first call either compute() or symbolic()/numeric()");
       const Index size = m_cholmodFactor->n;
@@ -322,8 +299,9 @@ class CholmodBase : internal::noncopyable
       if(!x_cs)
       {
         this->m_info = NumericalIssue;
+        return;
       }
-      // TODO optimize this copy by swapping when possible (be carreful with alignment, etc.)
+      // TODO optimize this copy by swapping when possible (be careful with alignment, etc.)
       dest = viewAsEigen<DestScalar,DestOptions,DestIndex>(*x_cs);
       cholmod_free_sparse(&x_cs, &m_cholmod);
     }
@@ -354,7 +332,6 @@ class CholmodBase : internal::noncopyable
     cholmod_factor* m_cholmodFactor;
     RealScalar m_shiftOffset[2];
     mutable ComputationInfo m_info;
-    bool m_isInitialized;
     int m_factorizationIsOk;
     int m_analysisIsOk;
 };
@@ -365,8 +342,8 @@ class CholmodBase : internal::noncopyable
   *
   * This class allows to solve for A.X = B sparse linear problems via a simplicial LL^T Cholesky factorization
   * using the Cholmod library.
-  * This simplicial variant is equivalent to Eigen's built-in SimplicialLLT class. Thefore, it has little practical interest.
-  * The sparse matrix A must be selfajoint and positive definite. The vectors or matrices
+  * This simplicial variant is equivalent to Eigen's built-in SimplicialLLT class. Therefore, it has little practical interest.
+  * The sparse matrix A must be selfadjoint and positive definite. The vectors or matrices
   * X and B can be either dense or sparse.
   *
   * \tparam _MatrixType the type of the sparse matrix A, it must be a SparseMatrix<>
@@ -392,7 +369,7 @@ class CholmodSimplicialLLT : public CholmodBase<_MatrixType, _UpLo, CholmodSimpl
     CholmodSimplicialLLT(const MatrixType& matrix) : Base()
     {
       init();
-      compute(matrix);
+      this->compute(matrix);
     }
 
     ~CholmodSimplicialLLT() {}
@@ -412,8 +389,8 @@ class CholmodSimplicialLLT : public CholmodBase<_MatrixType, _UpLo, CholmodSimpl
   *
   * This class allows to solve for A.X = B sparse linear problems via a simplicial LDL^T Cholesky factorization
   * using the Cholmod library.
-  * This simplicial variant is equivalent to Eigen's built-in SimplicialLDLT class. Thefore, it has little practical interest.
-  * The sparse matrix A must be selfajoint and positive definite. The vectors or matrices
+  * This simplicial variant is equivalent to Eigen's built-in SimplicialLDLT class. Therefore, it has little practical interest.
+  * The sparse matrix A must be selfadjoint and positive definite. The vectors or matrices
   * X and B can be either dense or sparse.
   *
   * \tparam _MatrixType the type of the sparse matrix A, it must be a SparseMatrix<>
@@ -439,7 +416,7 @@ class CholmodSimplicialLDLT : public CholmodBase<_MatrixType, _UpLo, CholmodSimp
     CholmodSimplicialLDLT(const MatrixType& matrix) : Base()
     {
       init();
-      compute(matrix);
+      this->compute(matrix);
     }
 
     ~CholmodSimplicialLDLT() {}
@@ -458,7 +435,7 @@ class CholmodSimplicialLDLT : public CholmodBase<_MatrixType, _UpLo, CholmodSimp
   * This class allows to solve for A.X = B sparse linear problems via a supernodal LL^T Cholesky factorization
   * using the Cholmod library.
   * This supernodal variant performs best on dense enough problems, e.g., 3D FEM, or very high order 2D FEM.
-  * The sparse matrix A must be selfajoint and positive definite. The vectors or matrices
+  * The sparse matrix A must be selfadjoint and positive definite. The vectors or matrices
   * X and B can be either dense or sparse.
   *
   * \tparam _MatrixType the type of the sparse matrix A, it must be a SparseMatrix<>
@@ -484,7 +461,7 @@ class CholmodSupernodalLLT : public CholmodBase<_MatrixType, _UpLo, CholmodSuper
     CholmodSupernodalLLT(const MatrixType& matrix) : Base()
     {
       init();
-      compute(matrix);
+      this->compute(matrix);
     }
 
     ~CholmodSupernodalLLT() {}
@@ -501,7 +478,7 @@ class CholmodSupernodalLLT : public CholmodBase<_MatrixType, _UpLo, CholmodSuper
   * \brief A general Cholesky factorization and solver based on Cholmod
   *
   * This class allows to solve for A.X = B sparse linear problems via a LL^T or LDL^T Cholesky factorization
-  * using the Cholmod library. The sparse matrix A must be selfajoint and positive definite. The vectors or matrices
+  * using the Cholmod library. The sparse matrix A must be selfadjoint and positive definite. The vectors or matrices
   * X and B can be either dense or sparse.
   *
   * This variant permits to change the underlying Cholesky method at runtime.
@@ -531,7 +508,7 @@ class CholmodDecomposition : public CholmodBase<_MatrixType, _UpLo, CholmodDecom
     CholmodDecomposition(const MatrixType& matrix) : Base()
     {
       init();
-      compute(matrix);
+      this->compute(matrix);
     }
 
     ~CholmodDecomposition() {}
@@ -569,36 +546,6 @@ class CholmodDecomposition : public CholmodBase<_MatrixType, _UpLo, CholmodDecom
     }
 };
 
-namespace internal {
-  
-template<typename _MatrixType, int _UpLo, typename Derived, typename Rhs>
-struct solve_retval<CholmodBase<_MatrixType,_UpLo,Derived>, Rhs>
-  : solve_retval_base<CholmodBase<_MatrixType,_UpLo,Derived>, Rhs>
-{
-  typedef CholmodBase<_MatrixType,_UpLo,Derived> Dec;
-  EIGEN_MAKE_SOLVE_HELPERS(Dec,Rhs)
-
-  template<typename Dest> void evalTo(Dest& dst) const
-  {
-    dec()._solve(rhs(),dst);
-  }
-};
-
-template<typename _MatrixType, int _UpLo, typename Derived, typename Rhs>
-struct sparse_solve_retval<CholmodBase<_MatrixType,_UpLo,Derived>, Rhs>
-  : sparse_solve_retval_base<CholmodBase<_MatrixType,_UpLo,Derived>, Rhs>
-{
-  typedef CholmodBase<_MatrixType,_UpLo,Derived> Dec;
-  EIGEN_MAKE_SPARSE_SOLVE_HELPERS(Dec,Rhs)
-
-  template<typename Dest> void evalTo(Dest& dst) const
-  {
-    dec()._solve(rhs(),dst);
-  }
-};
-
-} // end namespace internal
-
 } // end namespace Eigen
 
 #endif // EIGEN_CHOLMODSUPPORT_H

Eigen/src/Core/Array.h

@@ -24,6 +24,9 @@ namespace Eigen {
   * API for the %Matrix class provides easy access to linear-algebra
   * operations.
   *
+  * See documentation of class Matrix for detailed information on the template parameters
+  * storage layout.
+  * 
   * This class can be extended with the help of the plugin mechanism described on the page
   * \ref TopicCustomizingEigen by defining the preprocessor symbol \c EIGEN_ARRAY_PLUGIN.
   *
@@ -69,11 +72,27 @@ class Array
       * the usage of 'using'. This should be done only for operator=.
       */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Array& operator=(const EigenBase<OtherDerived> &other)
     {
       return Base::operator=(other);
     }
 
+    /** Set all the entries to \a value.
+      * \sa DenseBase::setConstant(), DenseBase::fill()
+      */
+    /* This overload is needed because the usage of
+      *   using Base::operator=;
+      * fails on MSVC. Since the code below is working with GCC and MSVC, we skipped
+      * the usage of 'using'. This should be done only for operator=.
+      */
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Array& operator=(const Scalar &value)
+    {
+      Base::setConstant(value);
+      return *this;
+    }
+
     /** Copies the value of the expression \a other into \c *this with automatic resizing.
       *
       * *this might be resized to match the dimensions of \a other. If *this was a null matrix (not already initialized),
@@ -84,7 +103,8 @@ class Array
       * remain row-vectors and vectors remain vectors.
       */
     template<typename OtherDerived>
-    EIGEN_STRONG_INLINE Array& operator=(const ArrayBase<OtherDerived>& other)
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Array& operator=(const DenseBase<OtherDerived>& other)
     {
       return Base::_set(other);
     }
@@ -92,11 +112,12 @@ class Array
     /** This is a special case of the templated operator=. Its purpose is to
       * prevent a default operator= from hiding the templated operator=.
       */
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Array& operator=(const Array& other)
     {
       return Base::_set(other);
     }
-
+    
     /** Default constructor.
       *
       * For fixed-size matrices, does nothing.
@@ -107,6 +128,7 @@ class Array
       *
       * \sa resize(Index,Index)
       */
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Array() : Base()
     {
       Base::_check_template_params();
@@ -116,6 +138,7 @@ class Array
 #ifndef EIGEN_PARSED_BY_DOXYGEN
     // FIXME is it still needed ??
     /** \internal */
+    EIGEN_DEVICE_FUNC
     Array(internal::constructor_without_unaligned_array_assert)
       : Base(internal::constructor_without_unaligned_array_assert())
     {
@@ -124,41 +147,64 @@ class Array
     }
 #endif
 
-    /** Constructs a vector or row-vector with given dimension. \only_for_vectors
-      *
-      * Note that this is only useful for dynamic-size vectors. For fixed-size vectors,
-      * it is redundant to pass the dimension here, so it makes more sense to use the default
-      * constructor Matrix() instead.
-      */
-    EIGEN_STRONG_INLINE explicit Array(Index dim)
-      : Base(dim, RowsAtCompileTime == 1 ? 1 : dim, ColsAtCompileTime == 1 ? 1 : dim)
+#ifdef EIGEN_HAVE_RVALUE_REFERENCES
+    EIGEN_DEVICE_FUNC
+    Array(Array&& other)
+      : Base(std::move(other))
     {
       Base::_check_template_params();
-      EIGEN_STATIC_ASSERT_VECTOR_ONLY(Array)
-      eigen_assert(dim >= 0);
-      eigen_assert(SizeAtCompileTime == Dynamic || SizeAtCompileTime == dim);
-      EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED
+      if (RowsAtCompileTime!=Dynamic && ColsAtCompileTime!=Dynamic)
+        Base::_set_noalias(other);
     }
+    EIGEN_DEVICE_FUNC
+    Array& operator=(Array&& other)
+    {
+      other.swap(*this);
+      return *this;
+    }
+#endif
 
     #ifndef EIGEN_PARSED_BY_DOXYGEN
+    template<typename T>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE explicit Array(const T& x)
+    {
+      Base::_check_template_params();
+      Base::template _init1<T>(x);
+    }
+
     template<typename T0, typename T1>
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Array(const T0& val0, const T1& val1)
     {
       Base::_check_template_params();
       this->template _init2<T0,T1>(val0, val1);
     }
     #else
-    /** constructs an uninitialized matrix with \a rows rows and \a cols columns.
+    /** \brief Constructs a fixed-sized array initialized with coefficients starting at \a data */
+    EIGEN_DEVICE_FUNC explicit Array(const Scalar *data);
+    /** Constructs a vector or row-vector with given dimension. \only_for_vectors
       *
-      * This is useful for dynamic-size matrices. For fixed-size matrices,
+      * Note that this is only useful for dynamic-size vectors. For fixed-size vectors,
+      * it is redundant to pass the dimension here, so it makes more sense to use the default
+      * constructor Array() instead.
+      */
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE explicit Array(Index dim);
+    /** constructs an initialized 1x1 Array with the given coefficient */
+    Array(const Scalar& value);
+    /** constructs an uninitialized array with \a rows rows and \a cols columns.
+      *
+      * This is useful for dynamic-size arrays. For fixed-size arrays,
       * it is redundant to pass these parameters, so one should use the default constructor
-      * Matrix() instead. */
+      * Array() instead. */
     Array(Index rows, Index cols);
     /** constructs an initialized 2D vector with given coefficients */
     Array(const Scalar& val0, const Scalar& val1);
     #endif
 
     /** constructs an initialized 3D vector with given coefficients */
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Array(const Scalar& val0, const Scalar& val1, const Scalar& val2)
     {
       Base::_check_template_params();
@@ -168,6 +214,7 @@ class Array
       m_storage.data()[2] = val2;
     }
     /** constructs an initialized 4D vector with given coefficients */
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Array(const Scalar& val0, const Scalar& val1, const Scalar& val2, const Scalar& val3)
     {
       Base::_check_template_params();
@@ -178,51 +225,21 @@ class Array
       m_storage.data()[3] = val3;
     }
 
-    explicit Array(const Scalar *data);
-
-    /** Constructor copying the value of the expression \a other */
-    template<typename OtherDerived>
-    EIGEN_STRONG_INLINE Array(const ArrayBase<OtherDerived>& other)
-             : Base(other.rows() * other.cols(), other.rows(), other.cols())
-    {
-      Base::_check_template_params();
-      Base::_set_noalias(other);
-    }
     /** Copy constructor */
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Array(const Array& other)
-            : Base(other.rows() * other.cols(), other.rows(), other.cols())
-    {
-      Base::_check_template_params();
-      Base::_set_noalias(other);
-    }
-    /** Copy constructor with in-place evaluation */
-    template<typename OtherDerived>
-    EIGEN_STRONG_INLINE Array(const ReturnByValue<OtherDerived>& other)
-    {
-      Base::_check_template_params();
-      Base::resize(other.rows(), other.cols());
-      other.evalTo(*this);
-    }
+            : Base(other)
+    { }
 
     /** \sa MatrixBase::operator=(const EigenBase<OtherDerived>&) */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Array(const EigenBase<OtherDerived> &other)
-      : Base(other.derived().rows() * other.derived().cols(), other.derived().rows(), other.derived().cols())
-    {
-      Base::_check_template_params();
-      Base::resize(other.rows(), other.cols());
-      *this = other;
-    }
+      : Base(other.derived())
+    { }
 
-    /** Override MatrixBase::swap() since for dynamic-sized matrices of same type it is enough to swap the
-      * data pointers.
-      */
-    template<typename OtherDerived>
-    void swap(ArrayBase<OtherDerived> const & other)
-    { this->_swap(other.derived()); }
-
-    inline Index innerStride() const { return 1; }
-    inline Index outerStride() const { return this->innerSize(); }
+    EIGEN_DEVICE_FUNC inline Index innerStride() const { return 1; }
+    EIGEN_DEVICE_FUNC inline Index outerStride() const { return this->innerSize(); }
 
     #ifdef EIGEN_ARRAY_PLUGIN
     #include EIGEN_ARRAY_PLUGIN

Eigen/src/Core/ArrayBase.h

@@ -50,7 +50,6 @@ template<typename Derived> class ArrayBase
                 typename NumTraits<typename internal::traits<Derived>::Scalar>::Real>::operator*;
 
     typedef typename internal::traits<Derived>::StorageKind StorageKind;
-    typedef typename internal::traits<Derived>::Index Index;
     typedef typename internal::traits<Derived>::Scalar Scalar;
     typedef typename internal::packet_traits<Scalar>::type PacketScalar;
     typedef typename NumTraits<Scalar>::Real RealScalar;
@@ -64,8 +63,7 @@ template<typename Derived> class ArrayBase
     using Base::MaxSizeAtCompileTime;
     using Base::IsVectorAtCompileTime;
     using Base::Flags;
-    using Base::CoeffReadCost;
-
+    
     using Base::derived;
     using Base::const_cast_derived;
     using Base::rows;
@@ -85,22 +83,10 @@ template<typename Derived> class ArrayBase
 #endif // not EIGEN_PARSED_BY_DOXYGEN
 
 #ifndef EIGEN_PARSED_BY_DOXYGEN
-    /** \internal the plain matrix type corresponding to this expression. Note that is not necessarily
-      * exactly the return type of eval(): in the case of plain matrices, the return type of eval() is a const
-      * reference to a matrix, not a matrix! It is however guaranteed that the return type of eval() is either
-      * PlainObject or const PlainObject&.
-      */
-    typedef Array<typename internal::traits<Derived>::Scalar,
-                internal::traits<Derived>::RowsAtCompileTime,
-                internal::traits<Derived>::ColsAtCompileTime,
-                AutoAlign | (internal::traits<Derived>::Flags&RowMajorBit ? RowMajor : ColMajor),
-                internal::traits<Derived>::MaxRowsAtCompileTime,
-                internal::traits<Derived>::MaxColsAtCompileTime
-          > PlainObject;
-
+    typedef typename Base::PlainObject PlainObject;
 
     /** \internal Represents a matrix with all coefficients equal to one another*/
-    typedef CwiseNullaryOp<internal::scalar_constant_op<Scalar>,Derived> ConstantReturnType;
+    typedef CwiseNullaryOp<internal::scalar_constant_op<Scalar>,PlainObject> ConstantReturnType;
 #endif // not EIGEN_PARSED_BY_DOXYGEN
 
 #define EIGEN_CURRENT_STORAGE_BASE_CLASS Eigen::ArrayBase
@@ -118,40 +104,57 @@ template<typename Derived> class ArrayBase
     /** Special case of the template operator=, in order to prevent the compiler
       * from generating a default operator= (issue hit with g++ 4.1)
       */
+    EIGEN_DEVICE_FUNC
     Derived& operator=(const ArrayBase& other)
     {
-      return internal::assign_selector<Derived,Derived>::run(derived(), other.derived());
+      internal::call_assignment(derived(), other.derived());
+      return derived();
     }
+    
+    /** Set all the entries to \a value.
+      * \sa DenseBase::setConstant(), DenseBase::fill() */
+    EIGEN_DEVICE_FUNC
+    Derived& operator=(const Scalar &value)
+    { Base::setConstant(value); return derived(); }
 
-    Derived& operator+=(const Scalar& scalar)
-    { return *this = derived() + scalar; }
-    Derived& operator-=(const Scalar& scalar)
-    { return *this = derived() - scalar; }
+    EIGEN_DEVICE_FUNC
+    Derived& operator+=(const Scalar& scalar);
+    EIGEN_DEVICE_FUNC
+    Derived& operator-=(const Scalar& scalar);
 
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     Derived& operator+=(const ArrayBase<OtherDerived>& other);
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     Derived& operator-=(const ArrayBase<OtherDerived>& other);
 
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     Derived& operator*=(const ArrayBase<OtherDerived>& other);
 
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     Derived& operator/=(const ArrayBase<OtherDerived>& other);
 
   public:
+    EIGEN_DEVICE_FUNC
     ArrayBase<Derived>& array() { return *this; }
+    EIGEN_DEVICE_FUNC
     const ArrayBase<Derived>& array() const { return *this; }
 
     /** \returns an \link Eigen::MatrixBase Matrix \endlink expression of this array
       * \sa MatrixBase::array() */
-    MatrixWrapper<Derived> matrix() { return derived(); }
-    const MatrixWrapper<const Derived> matrix() const { return derived(); }
+    EIGEN_DEVICE_FUNC
+    MatrixWrapper<Derived> matrix() { return MatrixWrapper<Derived>(derived()); }
+    EIGEN_DEVICE_FUNC
+    const MatrixWrapper<const Derived> matrix() const { return MatrixWrapper<const Derived>(derived()); }
 
 //     template<typename Dest>
 //     inline void evalTo(Dest& dst) const { dst = matrix(); }
 
   protected:
+    EIGEN_DEVICE_FUNC
     ArrayBase() : Base() {}
 
   private:
@@ -176,8 +179,7 @@ template<typename OtherDerived>
 EIGEN_STRONG_INLINE Derived &
 ArrayBase<Derived>::operator-=(const ArrayBase<OtherDerived> &other)
 {
-  SelfCwiseBinaryOp<internal::scalar_difference_op<Scalar>, Derived, OtherDerived> tmp(derived());
-  tmp = other.derived();
+  call_assignment(derived(), other.derived(), internal::sub_assign_op<Scalar>());
   return derived();
 }
 
@@ -190,8 +192,7 @@ template<typename OtherDerived>
 EIGEN_STRONG_INLINE Derived &
 ArrayBase<Derived>::operator+=(const ArrayBase<OtherDerived>& other)
 {
-  SelfCwiseBinaryOp<internal::scalar_sum_op<Scalar>, Derived, OtherDerived> tmp(derived());
-  tmp = other.derived();
+  call_assignment(derived(), other.derived(), internal::add_assign_op<Scalar>());
   return derived();
 }
 
@@ -204,8 +205,7 @@ template<typename OtherDerived>
 EIGEN_STRONG_INLINE Derived &
 ArrayBase<Derived>::operator*=(const ArrayBase<OtherDerived>& other)
 {
-  SelfCwiseBinaryOp<internal::scalar_product_op<Scalar>, Derived, OtherDerived> tmp(derived());
-  tmp = other.derived();
+  call_assignment(derived(), other.derived(), internal::mul_assign_op<Scalar,typename OtherDerived::Scalar>());
   return derived();
 }
 
@@ -218,8 +218,7 @@ template<typename OtherDerived>
 EIGEN_STRONG_INLINE Derived &
 ArrayBase<Derived>::operator/=(const ArrayBase<OtherDerived>& other)
 {
-  SelfCwiseBinaryOp<internal::scalar_quotient_op<Scalar>, Derived, OtherDerived> tmp(derived());
-  tmp = other.derived();
+  call_assignment(derived(), other.derived(), internal::div_assign_op<Scalar>());
   return derived();
 }
 

Eigen/src/Core/ArrayWrapper.h

@@ -29,6 +29,11 @@ struct traits<ArrayWrapper<ExpressionType> >
   : public traits<typename remove_all<typename ExpressionType::Nested>::type >
 {
   typedef ArrayXpr XprKind;
+  // Let's remove NestByRefBit
+  enum {
+    Flags0 = traits<typename remove_all<typename ExpressionType::Nested>::type >::Flags,
+    Flags = Flags0 & ~NestByRefBit
+  };
 };
 }
 
@@ -39,6 +44,7 @@ class ArrayWrapper : public ArrayBase<ArrayWrapper<ExpressionType> >
     typedef ArrayBase<ArrayWrapper> Base;
     EIGEN_DENSE_PUBLIC_INTERFACE(ArrayWrapper)
     EIGEN_INHERIT_ASSIGNMENT_OPERATORS(ArrayWrapper)
+    typedef typename internal::remove_all<ExpressionType>::type NestedExpression;
 
     typedef typename internal::conditional<
                        internal::is_lvalue<ExpressionType>::value,
@@ -46,43 +52,56 @@ class ArrayWrapper : public ArrayBase<ArrayWrapper<ExpressionType> >
                        const Scalar
                      >::type ScalarWithConstIfNotLvalue;
 
-    typedef typename internal::nested<ExpressionType>::type NestedExpressionType;
+    typedef typename internal::ref_selector<ExpressionType>::type NestedExpressionType;
 
-    inline ArrayWrapper(ExpressionType& matrix) : m_expression(matrix) {}
+    EIGEN_DEVICE_FUNC
+    explicit EIGEN_STRONG_INLINE ArrayWrapper(ExpressionType& matrix) : m_expression(matrix) {}
 
+    EIGEN_DEVICE_FUNC
     inline Index rows() const { return m_expression.rows(); }
+    EIGEN_DEVICE_FUNC
     inline Index cols() const { return m_expression.cols(); }
+    EIGEN_DEVICE_FUNC
     inline Index outerStride() const { return m_expression.outerStride(); }
+    EIGEN_DEVICE_FUNC
     inline Index innerStride() const { return m_expression.innerStride(); }
 
+    EIGEN_DEVICE_FUNC
     inline ScalarWithConstIfNotLvalue* data() { return m_expression.const_cast_derived().data(); }
+    EIGEN_DEVICE_FUNC
     inline const Scalar* data() const { return m_expression.data(); }
 
+    EIGEN_DEVICE_FUNC
     inline CoeffReturnType coeff(Index rowId, Index colId) const
     {
       return m_expression.coeff(rowId, colId);
     }
 
+    EIGEN_DEVICE_FUNC
     inline Scalar& coeffRef(Index rowId, Index colId)
     {
       return m_expression.const_cast_derived().coeffRef(rowId, colId);
     }
 
+    EIGEN_DEVICE_FUNC
     inline const Scalar& coeffRef(Index rowId, Index colId) const
     {
       return m_expression.const_cast_derived().coeffRef(rowId, colId);
     }
 
+    EIGEN_DEVICE_FUNC
     inline CoeffReturnType coeff(Index index) const
     {
       return m_expression.coeff(index);
     }
 
+    EIGEN_DEVICE_FUNC
     inline Scalar& coeffRef(Index index)
     {
       return m_expression.const_cast_derived().coeffRef(index);
     }
 
+    EIGEN_DEVICE_FUNC
     inline const Scalar& coeffRef(Index index) const
     {
       return m_expression.const_cast_derived().coeffRef(index);
@@ -113,9 +132,11 @@ class ArrayWrapper : public ArrayBase<ArrayWrapper<ExpressionType> >
     }
 
     template<typename Dest>
+    EIGEN_DEVICE_FUNC
     inline void evalTo(Dest& dst) const { dst = m_expression; }
 
     const typename internal::remove_all<NestedExpressionType>::type& 
+    EIGEN_DEVICE_FUNC
     nestedExpression() const 
     {
       return m_expression;
@@ -123,10 +144,12 @@ class ArrayWrapper : public ArrayBase<ArrayWrapper<ExpressionType> >
 
     /** Forwards the resizing request to the nested expression
       * \sa DenseBase::resize(Index)  */
+    EIGEN_DEVICE_FUNC
     void resize(Index newSize) { m_expression.const_cast_derived().resize(newSize); }
     /** Forwards the resizing request to the nested expression
       * \sa DenseBase::resize(Index,Index)*/
-    void resize(Index nbRows, Index nbCols) { m_expression.const_cast_derived().resize(nbRows,nbCols); }
+    EIGEN_DEVICE_FUNC
+    void resize(Index rows, Index cols) { m_expression.const_cast_derived().resize(rows,cols); }
 
   protected:
     NestedExpressionType m_expression;
@@ -149,6 +172,11 @@ struct traits<MatrixWrapper<ExpressionType> >
  : public traits<typename remove_all<typename ExpressionType::Nested>::type >
 {
   typedef MatrixXpr XprKind;
+  // Let's remove NestByRefBit
+  enum {
+    Flags0 = traits<typename remove_all<typename ExpressionType::Nested>::type >::Flags,
+    Flags = Flags0 & ~NestByRefBit
+  };
 };
 }
 
@@ -159,6 +187,7 @@ class MatrixWrapper : public MatrixBase<MatrixWrapper<ExpressionType> >
     typedef MatrixBase<MatrixWrapper<ExpressionType> > Base;
     EIGEN_DENSE_PUBLIC_INTERFACE(MatrixWrapper)
     EIGEN_INHERIT_ASSIGNMENT_OPERATORS(MatrixWrapper)
+    typedef typename internal::remove_all<ExpressionType>::type NestedExpression;
 
     typedef typename internal::conditional<
                        internal::is_lvalue<ExpressionType>::value,
@@ -166,43 +195,56 @@ class MatrixWrapper : public MatrixBase<MatrixWrapper<ExpressionType> >
                        const Scalar
                      >::type ScalarWithConstIfNotLvalue;
 
-    typedef typename internal::nested<ExpressionType>::type NestedExpressionType;
+    typedef typename internal::ref_selector<ExpressionType>::type NestedExpressionType;
 
-    inline MatrixWrapper(ExpressionType& a_matrix) : m_expression(a_matrix) {}
+    EIGEN_DEVICE_FUNC
+    explicit inline MatrixWrapper(ExpressionType& matrix) : m_expression(matrix) {}
 
+    EIGEN_DEVICE_FUNC
     inline Index rows() const { return m_expression.rows(); }
+    EIGEN_DEVICE_FUNC
     inline Index cols() const { return m_expression.cols(); }
+    EIGEN_DEVICE_FUNC
     inline Index outerStride() const { return m_expression.outerStride(); }
+    EIGEN_DEVICE_FUNC
     inline Index innerStride() const { return m_expression.innerStride(); }
 
+    EIGEN_DEVICE_FUNC
     inline ScalarWithConstIfNotLvalue* data() { return m_expression.const_cast_derived().data(); }
+    EIGEN_DEVICE_FUNC
     inline const Scalar* data() const { return m_expression.data(); }
 
+    EIGEN_DEVICE_FUNC
     inline CoeffReturnType coeff(Index rowId, Index colId) const
     {
       return m_expression.coeff(rowId, colId);
     }
 
+    EIGEN_DEVICE_FUNC
     inline Scalar& coeffRef(Index rowId, Index colId)
     {
       return m_expression.const_cast_derived().coeffRef(rowId, colId);
     }
 
+    EIGEN_DEVICE_FUNC
     inline const Scalar& coeffRef(Index rowId, Index colId) const
     {
       return m_expression.derived().coeffRef(rowId, colId);
     }
 
+    EIGEN_DEVICE_FUNC
     inline CoeffReturnType coeff(Index index) const
     {
       return m_expression.coeff(index);
     }
 
+    EIGEN_DEVICE_FUNC
     inline Scalar& coeffRef(Index index)
     {
       return m_expression.const_cast_derived().coeffRef(index);
     }
 
+    EIGEN_DEVICE_FUNC
     inline const Scalar& coeffRef(Index index) const
     {
       return m_expression.const_cast_derived().coeffRef(index);
@@ -232,6 +274,7 @@ class MatrixWrapper : public MatrixBase<MatrixWrapper<ExpressionType> >
       m_expression.const_cast_derived().template writePacket<LoadMode>(index, val);
     }
 
+    EIGEN_DEVICE_FUNC
     const typename internal::remove_all<NestedExpressionType>::type& 
     nestedExpression() const 
     {
@@ -240,10 +283,12 @@ class MatrixWrapper : public MatrixBase<MatrixWrapper<ExpressionType> >
 
     /** Forwards the resizing request to the nested expression
       * \sa DenseBase::resize(Index)  */
+    EIGEN_DEVICE_FUNC
     void resize(Index newSize) { m_expression.const_cast_derived().resize(newSize); }
     /** Forwards the resizing request to the nested expression
       * \sa DenseBase::resize(Index,Index)*/
-    void resize(Index nbRows, Index nbCols) { m_expression.const_cast_derived().resize(nbRows,nbCols); }
+    EIGEN_DEVICE_FUNC
+    void resize(Index rows, Index cols) { m_expression.const_cast_derived().resize(rows,cols); }
 
   protected:
     NestedExpressionType m_expression;

Eigen/src/Core/Assign.h

@@ -14,471 +14,6 @@
 
 namespace Eigen {
 
-namespace internal {
-
-/***************************************************************************
-* Part 1 : the logic deciding a strategy for traversal and unrolling       *
-***************************************************************************/
-
-template <typename Derived, typename OtherDerived>
-struct assign_traits
-{
-public:
-  enum {
-    DstIsAligned = Derived::Flags & AlignedBit,
-    DstHasDirectAccess = Derived::Flags & DirectAccessBit,
-    SrcIsAligned = OtherDerived::Flags & AlignedBit,
-    JointAlignment = bool(DstIsAligned) && bool(SrcIsAligned) ? Aligned : Unaligned
-  };
-
-private:
-  enum {
-    InnerSize = int(Derived::IsVectorAtCompileTime) ? int(Derived::SizeAtCompileTime)
-              : int(Derived::Flags)&RowMajorBit ? int(Derived::ColsAtCompileTime)
-              : int(Derived::RowsAtCompileTime),
-    InnerMaxSize = int(Derived::IsVectorAtCompileTime) ? int(Derived::MaxSizeAtCompileTime)
-              : int(Derived::Flags)&RowMajorBit ? int(Derived::MaxColsAtCompileTime)
-              : int(Derived::MaxRowsAtCompileTime),
-    MaxSizeAtCompileTime = Derived::SizeAtCompileTime,
-    PacketSize = packet_traits<typename Derived::Scalar>::size
-  };
-
-  enum {
-    StorageOrdersAgree = (int(Derived::IsRowMajor) == int(OtherDerived::IsRowMajor)),
-    MightVectorize = StorageOrdersAgree
-                  && (int(Derived::Flags) & int(OtherDerived::Flags) & ActualPacketAccessBit),
-    MayInnerVectorize  = MightVectorize && int(InnerSize)!=Dynamic && int(InnerSize)%int(PacketSize)==0
-                       && int(DstIsAligned) && int(SrcIsAligned),
-    MayLinearize = StorageOrdersAgree && (int(Derived::Flags) & int(OtherDerived::Flags) & LinearAccessBit),
-    MayLinearVectorize = MightVectorize && MayLinearize && DstHasDirectAccess
-                       && (DstIsAligned || MaxSizeAtCompileTime == Dynamic),
-      /* If the destination isn't aligned, we have to do runtime checks and we don't unroll,
-         so it's only good for large enough sizes. */
-    MaySliceVectorize  = MightVectorize && DstHasDirectAccess
-                       && (int(InnerMaxSize)==Dynamic || int(InnerMaxSize)>=3*PacketSize)
-      /* slice vectorization can be slow, so we only want it if the slices are big, which is
-         indicated by InnerMaxSize rather than InnerSize, think of the case of a dynamic block
-         in a fixed-size matrix */
-  };
-
-public:
-  enum {
-    Traversal = int(MayInnerVectorize)  ? int(InnerVectorizedTraversal)
-              : int(MayLinearVectorize) ? int(LinearVectorizedTraversal)
-              : int(MaySliceVectorize)  ? int(SliceVectorizedTraversal)
-              : int(MayLinearize)       ? int(LinearTraversal)
-                                        : int(DefaultTraversal),
-    Vectorized = int(Traversal) == InnerVectorizedTraversal
-              || int(Traversal) == LinearVectorizedTraversal
-              || int(Traversal) == SliceVectorizedTraversal
-  };
-
-private:
-  enum {
-    UnrollingLimit      = EIGEN_UNROLLING_LIMIT * (Vectorized ? int(PacketSize) : 1),
-    MayUnrollCompletely = int(Derived::SizeAtCompileTime) != Dynamic
-                       && int(OtherDerived::CoeffReadCost) != Dynamic
-                       && int(Derived::SizeAtCompileTime) * int(OtherDerived::CoeffReadCost) <= int(UnrollingLimit),
-    MayUnrollInner      = int(InnerSize) != Dynamic
-                       && int(OtherDerived::CoeffReadCost) != Dynamic
-                       && int(InnerSize) * int(OtherDerived::CoeffReadCost) <= int(UnrollingLimit)
-  };
-
-public:
-  enum {
-    Unrolling = (int(Traversal) == int(InnerVectorizedTraversal) || int(Traversal) == int(DefaultTraversal))
-                ? (
-                    int(MayUnrollCompletely) ? int(CompleteUnrolling)
-                  : int(MayUnrollInner)      ? int(InnerUnrolling)
-                                             : int(NoUnrolling)
-                  )
-              : int(Traversal) == int(LinearVectorizedTraversal)
-                ? ( bool(MayUnrollCompletely) && bool(DstIsAligned) ? int(CompleteUnrolling) : int(NoUnrolling) )
-              : int(Traversal) == int(LinearTraversal)
-                ? ( bool(MayUnrollCompletely) ? int(CompleteUnrolling) : int(NoUnrolling) )
-              : int(NoUnrolling)
-  };
-
-#ifdef EIGEN_DEBUG_ASSIGN
-  static void debug()
-  {
-    EIGEN_DEBUG_VAR(DstIsAligned)
-    EIGEN_DEBUG_VAR(SrcIsAligned)
-    EIGEN_DEBUG_VAR(JointAlignment)
-    EIGEN_DEBUG_VAR(InnerSize)
-    EIGEN_DEBUG_VAR(InnerMaxSize)
-    EIGEN_DEBUG_VAR(PacketSize)
-    EIGEN_DEBUG_VAR(StorageOrdersAgree)
-    EIGEN_DEBUG_VAR(MightVectorize)
-    EIGEN_DEBUG_VAR(MayLinearize)
-    EIGEN_DEBUG_VAR(MayInnerVectorize)
-    EIGEN_DEBUG_VAR(MayLinearVectorize)
-    EIGEN_DEBUG_VAR(MaySliceVectorize)
-    EIGEN_DEBUG_VAR(Traversal)
-    EIGEN_DEBUG_VAR(UnrollingLimit)
-    EIGEN_DEBUG_VAR(MayUnrollCompletely)
-    EIGEN_DEBUG_VAR(MayUnrollInner)
-    EIGEN_DEBUG_VAR(Unrolling)
-  }
-#endif
-};
-
-/***************************************************************************
-* Part 2 : meta-unrollers
-***************************************************************************/
-
-/************************
-*** Default traversal ***
-************************/
-
-template<typename Derived1, typename Derived2, int Index, int Stop>
-struct assign_DefaultTraversal_CompleteUnrolling
-{
-  enum {
-    outer = Index / Derived1::InnerSizeAtCompileTime,
-    inner = Index % Derived1::InnerSizeAtCompileTime
-  };
-
-  static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src)
-  {
-    dst.copyCoeffByOuterInner(outer, inner, src);
-    assign_DefaultTraversal_CompleteUnrolling<Derived1, Derived2, Index+1, Stop>::run(dst, src);
-  }
-};
-
-template<typename Derived1, typename Derived2, int Stop>
-struct assign_DefaultTraversal_CompleteUnrolling<Derived1, Derived2, Stop, Stop>
-{
-  static EIGEN_STRONG_INLINE void run(Derived1 &, const Derived2 &) {}
-};
-
-template<typename Derived1, typename Derived2, int Index, int Stop>
-struct assign_DefaultTraversal_InnerUnrolling
-{
-  static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src, typename Derived1::Index outer)
-  {
-    dst.copyCoeffByOuterInner(outer, Index, src);
-    assign_DefaultTraversal_InnerUnrolling<Derived1, Derived2, Index+1, Stop>::run(dst, src, outer);
-  }
-};
-
-template<typename Derived1, typename Derived2, int Stop>
-struct assign_DefaultTraversal_InnerUnrolling<Derived1, Derived2, Stop, Stop>
-{
-  static EIGEN_STRONG_INLINE void run(Derived1 &, const Derived2 &, typename Derived1::Index) {}
-};
-
-/***********************
-*** Linear traversal ***
-***********************/
-
-template<typename Derived1, typename Derived2, int Index, int Stop>
-struct assign_LinearTraversal_CompleteUnrolling
-{
-  static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src)
-  {
-    dst.copyCoeff(Index, src);
-    assign_LinearTraversal_CompleteUnrolling<Derived1, Derived2, Index+1, Stop>::run(dst, src);
-  }
-};
-
-template<typename Derived1, typename Derived2, int Stop>
-struct assign_LinearTraversal_CompleteUnrolling<Derived1, Derived2, Stop, Stop>
-{
-  static EIGEN_STRONG_INLINE void run(Derived1 &, const Derived2 &) {}
-};
-
-/**************************
-*** Inner vectorization ***
-**************************/
-
-template<typename Derived1, typename Derived2, int Index, int Stop>
-struct assign_innervec_CompleteUnrolling
-{
-  enum {
-    outer = Index / Derived1::InnerSizeAtCompileTime,
-    inner = Index % Derived1::InnerSizeAtCompileTime,
-    JointAlignment = assign_traits<Derived1,Derived2>::JointAlignment
-  };
-
-  static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src)
-  {
-    dst.template copyPacketByOuterInner<Derived2, Aligned, JointAlignment>(outer, inner, src);
-    assign_innervec_CompleteUnrolling<Derived1, Derived2,
-      Index+packet_traits<typename Derived1::Scalar>::size, Stop>::run(dst, src);
-  }
-};
-
-template<typename Derived1, typename Derived2, int Stop>
-struct assign_innervec_CompleteUnrolling<Derived1, Derived2, Stop, Stop>
-{
-  static EIGEN_STRONG_INLINE void run(Derived1 &, const Derived2 &) {}
-};
-
-template<typename Derived1, typename Derived2, int Index, int Stop>
-struct assign_innervec_InnerUnrolling
-{
-  static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src, typename Derived1::Index outer)
-  {
-    dst.template copyPacketByOuterInner<Derived2, Aligned, Aligned>(outer, Index, src);
-    assign_innervec_InnerUnrolling<Derived1, Derived2,
-      Index+packet_traits<typename Derived1::Scalar>::size, Stop>::run(dst, src, outer);
-  }
-};
-
-template<typename Derived1, typename Derived2, int Stop>
-struct assign_innervec_InnerUnrolling<Derived1, Derived2, Stop, Stop>
-{
-  static EIGEN_STRONG_INLINE void run(Derived1 &, const Derived2 &, typename Derived1::Index) {}
-};
-
-/***************************************************************************
-* Part 3 : implementation of all cases
-***************************************************************************/
-
-template<typename Derived1, typename Derived2,
-         int Traversal = assign_traits<Derived1, Derived2>::Traversal,
-         int Unrolling = assign_traits<Derived1, Derived2>::Unrolling,
-         int Version = Specialized>
-struct assign_impl;
-
-/************************
-*** Default traversal ***
-************************/
-
-template<typename Derived1, typename Derived2, int Unrolling, int Version>
-struct assign_impl<Derived1, Derived2, InvalidTraversal, Unrolling, Version>
-{
-  static inline void run(Derived1 &, const Derived2 &) { }
-};
-
-template<typename Derived1, typename Derived2, int Version>
-struct assign_impl<Derived1, Derived2, DefaultTraversal, NoUnrolling, Version>
-{
-  typedef typename Derived1::Index Index;
-  static inline void run(Derived1 &dst, const Derived2 &src)
-  {
-    const Index innerSize = dst.innerSize();
-    const Index outerSize = dst.outerSize();
-    for(Index outer = 0; outer < outerSize; ++outer)
-      for(Index inner = 0; inner < innerSize; ++inner)
-        dst.copyCoeffByOuterInner(outer, inner, src);
-  }
-};
-
-template<typename Derived1, typename Derived2, int Version>
-struct assign_impl<Derived1, Derived2, DefaultTraversal, CompleteUnrolling, Version>
-{
-  static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src)
-  {
-    assign_DefaultTraversal_CompleteUnrolling<Derived1, Derived2, 0, Derived1::SizeAtCompileTime>
-      ::run(dst, src);
-  }
-};
-
-template<typename Derived1, typename Derived2, int Version>
-struct assign_impl<Derived1, Derived2, DefaultTraversal, InnerUnrolling, Version>
-{
-  typedef typename Derived1::Index Index;
-  static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src)
-  {
-    const Index outerSize = dst.outerSize();
-    for(Index outer = 0; outer < outerSize; ++outer)
-      assign_DefaultTraversal_InnerUnrolling<Derived1, Derived2, 0, Derived1::InnerSizeAtCompileTime>
-        ::run(dst, src, outer);
-  }
-};
-
-/***********************
-*** Linear traversal ***
-***********************/
-
-template<typename Derived1, typename Derived2, int Version>
-struct assign_impl<Derived1, Derived2, LinearTraversal, NoUnrolling, Version>
-{
-  typedef typename Derived1::Index Index;
-  static inline void run(Derived1 &dst, const Derived2 &src)
-  {
-    const Index size = dst.size();
-    for(Index i = 0; i < size; ++i)
-      dst.copyCoeff(i, src);
-  }
-};
-
-template<typename Derived1, typename Derived2, int Version>
-struct assign_impl<Derived1, Derived2, LinearTraversal, CompleteUnrolling, Version>
-{
-  static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src)
-  {
-    assign_LinearTraversal_CompleteUnrolling<Derived1, Derived2, 0, Derived1::SizeAtCompileTime>
-      ::run(dst, src);
-  }
-};
-
-/**************************
-*** Inner vectorization ***
-**************************/
-
-template<typename Derived1, typename Derived2, int Version>
-struct assign_impl<Derived1, Derived2, InnerVectorizedTraversal, NoUnrolling, Version>
-{
-  typedef typename Derived1::Index Index;
-  static inline void run(Derived1 &dst, const Derived2 &src)
-  {
-    const Index innerSize = dst.innerSize();
-    const Index outerSize = dst.outerSize();
-    const Index packetSize = packet_traits<typename Derived1::Scalar>::size;
-    for(Index outer = 0; outer < outerSize; ++outer)
-      for(Index inner = 0; inner < innerSize; inner+=packetSize)
-        dst.template copyPacketByOuterInner<Derived2, Aligned, Aligned>(outer, inner, src);
-  }
-};
-
-template<typename Derived1, typename Derived2, int Version>
-struct assign_impl<Derived1, Derived2, InnerVectorizedTraversal, CompleteUnrolling, Version>
-{
-  static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src)
-  {
-    assign_innervec_CompleteUnrolling<Derived1, Derived2, 0, Derived1::SizeAtCompileTime>
-      ::run(dst, src);
-  }
-};
-
-template<typename Derived1, typename Derived2, int Version>
-struct assign_impl<Derived1, Derived2, InnerVectorizedTraversal, InnerUnrolling, Version>
-{
-  typedef typename Derived1::Index Index;
-  static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src)
-  {
-    const Index outerSize = dst.outerSize();
-    for(Index outer = 0; outer < outerSize; ++outer)
-      assign_innervec_InnerUnrolling<Derived1, Derived2, 0, Derived1::InnerSizeAtCompileTime>
-        ::run(dst, src, outer);
-  }
-};
-
-/***************************
-*** Linear vectorization ***
-***************************/
-
-template <bool IsAligned = false>
-struct unaligned_assign_impl
-{
-  template <typename Derived, typename OtherDerived>
-  static EIGEN_STRONG_INLINE void run(const Derived&, OtherDerived&, typename Derived::Index, typename Derived::Index) {}
-};
-
-template <>
-struct unaligned_assign_impl<false>
-{
-  // MSVC must not inline this functions. If it does, it fails to optimize the
-  // packet access path.
-#ifdef _MSC_VER
-  template <typename Derived, typename OtherDerived>
-  static EIGEN_DONT_INLINE void run(const Derived& src, OtherDerived& dst, typename Derived::Index start, typename Derived::Index end)
-#else
-  template <typename Derived, typename OtherDerived>
-  static EIGEN_STRONG_INLINE void run(const Derived& src, OtherDerived& dst, typename Derived::Index start, typename Derived::Index end)
-#endif
-  {
-    for (typename Derived::Index index = start; index < end; ++index)
-      dst.copyCoeff(index, src);
-  }
-};
-
-template<typename Derived1, typename Derived2, int Version>
-struct assign_impl<Derived1, Derived2, LinearVectorizedTraversal, NoUnrolling, Version>
-{
-  typedef typename Derived1::Index Index;
-  static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src)
-  {
-    const Index size = dst.size();
-    typedef packet_traits<typename Derived1::Scalar> PacketTraits;
-    enum {
-      packetSize = PacketTraits::size,
-      dstAlignment = PacketTraits::AlignedOnScalar ? Aligned : int(assign_traits<Derived1,Derived2>::DstIsAligned) ,
-      srcAlignment = assign_traits<Derived1,Derived2>::JointAlignment
-    };
-    const Index alignedStart = assign_traits<Derived1,Derived2>::DstIsAligned ? 0
-                             : internal::first_aligned(&dst.coeffRef(0), size);
-    const Index alignedEnd = alignedStart + ((size-alignedStart)/packetSize)*packetSize;
-
-    unaligned_assign_impl<assign_traits<Derived1,Derived2>::DstIsAligned!=0>::run(src,dst,0,alignedStart);
-
-    for(Index index = alignedStart; index < alignedEnd; index += packetSize)
-    {
-      dst.template copyPacket<Derived2, dstAlignment, srcAlignment>(index, src);
-    }
-
-    unaligned_assign_impl<>::run(src,dst,alignedEnd,size);
-  }
-};
-
-template<typename Derived1, typename Derived2, int Version>
-struct assign_impl<Derived1, Derived2, LinearVectorizedTraversal, CompleteUnrolling, Version>
-{
-  typedef typename Derived1::Index Index;
-  static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src)
-  {
-    enum { size = Derived1::SizeAtCompileTime,
-           packetSize = packet_traits<typename Derived1::Scalar>::size,
-           alignedSize = (size/packetSize)*packetSize };
-
-    assign_innervec_CompleteUnrolling<Derived1, Derived2, 0, alignedSize>::run(dst, src);
-    assign_DefaultTraversal_CompleteUnrolling<Derived1, Derived2, alignedSize, size>::run(dst, src);
-  }
-};
-
-/**************************
-*** Slice vectorization ***
-***************************/
-
-template<typename Derived1, typename Derived2, int Version>
-struct assign_impl<Derived1, Derived2, SliceVectorizedTraversal, NoUnrolling, Version>
-{
-  typedef typename Derived1::Index Index;
-  static inline void run(Derived1 &dst, const Derived2 &src)
-  {
-    typedef packet_traits<typename Derived1::Scalar> PacketTraits;
-    enum {
-      packetSize = PacketTraits::size,
-      alignable = PacketTraits::AlignedOnScalar,
-      dstAlignment = alignable ? Aligned : int(assign_traits<Derived1,Derived2>::DstIsAligned) ,
-      srcAlignment = assign_traits<Derived1,Derived2>::JointAlignment
-    };
-    const Index packetAlignedMask = packetSize - 1;
-    const Index innerSize = dst.innerSize();
-    const Index outerSize = dst.outerSize();
-    const Index alignedStep = alignable ? (packetSize - dst.outerStride() % packetSize) & packetAlignedMask : 0;
-    Index alignedStart = ((!alignable) || assign_traits<Derived1,Derived2>::DstIsAligned) ? 0
-                       : internal::first_aligned(&dst.coeffRef(0,0), innerSize);
-
-    for(Index outer = 0; outer < outerSize; ++outer)
-    {
-      const Index alignedEnd = alignedStart + ((innerSize-alignedStart) & ~packetAlignedMask);
-      // do the non-vectorizable part of the assignment
-      for(Index inner = 0; inner<alignedStart ; ++inner)
-        dst.copyCoeffByOuterInner(outer, inner, src);
-
-      // do the vectorizable part of the assignment
-      for(Index inner = alignedStart; inner<alignedEnd; inner+=packetSize)
-        dst.template copyPacketByOuterInner<Derived2, dstAlignment, Unaligned>(outer, inner, src);
-
-      // do the non-vectorizable part of the assignment
-      for(Index inner = alignedEnd; inner<innerSize ; ++inner)
-        dst.copyCoeffByOuterInner(outer, inner, src);
-
-      alignedStart = std::min<Index>((alignedStart+alignedStep)%packetSize, innerSize);
-    }
-  }
-};
-
-} // end namespace internal
-
-/***************************************************************************
-* Part 4 : implementation of DenseBase methods
-***************************************************************************/
-
 template<typename Derived>
 template<typename OtherDerived>
 EIGEN_STRONG_INLINE Derived& DenseBase<Derived>
@@ -492,90 +27,62 @@ EIGEN_STRONG_INLINE Derived& DenseBase<Derived>
   EIGEN_STATIC_ASSERT_SAME_MATRIX_SIZE(Derived,OtherDerived)
   EIGEN_STATIC_ASSERT(SameType,YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY)
 
-#ifdef EIGEN_DEBUG_ASSIGN
-  internal::assign_traits<Derived, OtherDerived>::debug();
-#endif
   eigen_assert(rows() == other.rows() && cols() == other.cols());
-  internal::assign_impl<Derived, OtherDerived, int(SameType) ? int(internal::assign_traits<Derived, OtherDerived>::Traversal)
-                                                       : int(InvalidTraversal)>::run(derived(),other.derived());
-#ifndef EIGEN_NO_DEBUG
-  checkTransposeAliasing(other.derived());
-#endif
+  internal::call_assignment_no_alias(derived(),other.derived());
+  
   return derived();
 }
 
-namespace internal {
-
-template<typename Derived, typename OtherDerived,
-         bool EvalBeforeAssigning = (int(internal::traits<OtherDerived>::Flags) & EvalBeforeAssigningBit) != 0,
-         bool NeedToTranspose = ((int(Derived::RowsAtCompileTime) == 1 && int(OtherDerived::ColsAtCompileTime) == 1)
-                              |   // FIXME | instead of || to please GCC 4.4.0 stupid warning "suggest parentheses around &&".
-                                  // revert to || as soon as not needed anymore.
-                                  (int(Derived::ColsAtCompileTime) == 1 && int(OtherDerived::RowsAtCompileTime) == 1))
-                              && int(Derived::SizeAtCompileTime) != 1>
-struct assign_selector;
-
-template<typename Derived, typename OtherDerived>
-struct assign_selector<Derived,OtherDerived,false,false> {
-  static EIGEN_STRONG_INLINE Derived& run(Derived& dst, const OtherDerived& other) { return dst.lazyAssign(other.derived()); }
-  template<typename ActualDerived, typename ActualOtherDerived>
-  static EIGEN_STRONG_INLINE Derived& evalTo(ActualDerived& dst, const ActualOtherDerived& other) { other.evalTo(dst); return dst; }
-};
-template<typename Derived, typename OtherDerived>
-struct assign_selector<Derived,OtherDerived,true,false> {
-  static EIGEN_STRONG_INLINE Derived& run(Derived& dst, const OtherDerived& other) { return dst.lazyAssign(other.eval()); }
-};
-template<typename Derived, typename OtherDerived>
-struct assign_selector<Derived,OtherDerived,false,true> {
-  static EIGEN_STRONG_INLINE Derived& run(Derived& dst, const OtherDerived& other) { return dst.lazyAssign(other.transpose()); }
-  template<typename ActualDerived, typename ActualOtherDerived>
-  static EIGEN_STRONG_INLINE Derived& evalTo(ActualDerived& dst, const ActualOtherDerived& other) { Transpose<ActualDerived> dstTrans(dst); other.evalTo(dstTrans); return dst; }
-};
-template<typename Derived, typename OtherDerived>
-struct assign_selector<Derived,OtherDerived,true,true> {
-  static EIGEN_STRONG_INLINE Derived& run(Derived& dst, const OtherDerived& other) { return dst.lazyAssign(other.transpose().eval()); }
-};
-
-} // end namespace internal
-
 template<typename Derived>
 template<typename OtherDerived>
+EIGEN_DEVICE_FUNC
 EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::operator=(const DenseBase<OtherDerived>& other)
 {
-  return internal::assign_selector<Derived,OtherDerived>::run(derived(), other.derived());
+  internal::call_assignment(derived(), other.derived());
+  return derived();
 }
 
 template<typename Derived>
+EIGEN_DEVICE_FUNC
 EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::operator=(const DenseBase& other)
 {
-  return internal::assign_selector<Derived,Derived>::run(derived(), other.derived());
+  internal::call_assignment(derived(), other.derived());
+  return derived();
 }
 
 template<typename Derived>
+EIGEN_DEVICE_FUNC
 EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::operator=(const MatrixBase& other)
 {
-  return internal::assign_selector<Derived,Derived>::run(derived(), other.derived());
+  internal::call_assignment(derived(), other.derived());
+  return derived();
 }
 
 template<typename Derived>
 template <typename OtherDerived>
+EIGEN_DEVICE_FUNC
 EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::operator=(const DenseBase<OtherDerived>& other)
 {
-  return internal::assign_selector<Derived,OtherDerived>::run(derived(), other.derived());
+  internal::call_assignment(derived(), other.derived());
+  return derived();
 }
 
 template<typename Derived>
 template <typename OtherDerived>
+EIGEN_DEVICE_FUNC
 EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::operator=(const EigenBase<OtherDerived>& other)
 {
-  return internal::assign_selector<Derived,OtherDerived,false>::evalTo(derived(), other.derived());
+  internal::call_assignment(derived(), other.derived());
+  return derived();
 }
 
 template<typename Derived>
 template<typename OtherDerived>
+EIGEN_DEVICE_FUNC
 EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::operator=(const ReturnByValue<OtherDerived>& other)
 {
-  return internal::assign_selector<Derived,OtherDerived,false>::evalTo(derived(), other.derived());
+  other.derived().evalTo(derived());
+  return derived();
 }
 
 } // end namespace Eigen

Eigen/src/Core/Assign_MKL.h

@@ -1,6 +1,7 @@
 /*
  Copyright (c) 2011, Intel Corporation. All rights reserved.
-
+ Copyright (C) 2015 Gael Guennebaud <gael.guennebaud@inria.fr>
+ 
  Redistribution and use in source and binary forms, with or without modification,
  are permitted provided that the following conditions are met:
 
@@ -37,17 +38,13 @@ namespace Eigen {
 
 namespace internal {
 
-template<typename Op> struct vml_call
-{ enum { IsSupported = 0 }; };
-
-template<typename Dst, typename Src, typename UnaryOp>
+template<typename Dst, typename Src>
 class vml_assign_traits
 {
   private:
     enum {
       DstHasDirectAccess = Dst::Flags & DirectAccessBit,
       SrcHasDirectAccess = Src::Flags & DirectAccessBit,
-
       StorageOrdersAgree = (int(Dst::IsRowMajor) == int(Src::IsRowMajor)),
       InnerSize = int(Dst::IsVectorAtCompileTime) ? int(Dst::SizeAtCompileTime)
                 : int(Dst::Flags)&RowMajorBit ? int(Dst::ColsAtCompileTime)
@@ -57,165 +54,118 @@ class vml_assign_traits
                     : int(Dst::MaxRowsAtCompileTime),
       MaxSizeAtCompileTime = Dst::SizeAtCompileTime,
 
-      MightEnableVml =  vml_call<UnaryOp>::IsSupported && StorageOrdersAgree && DstHasDirectAccess && SrcHasDirectAccess
-                     && Src::InnerStrideAtCompileTime==1 && Dst::InnerStrideAtCompileTime==1,
+      MightEnableVml = StorageOrdersAgree && DstHasDirectAccess && SrcHasDirectAccess && Src::InnerStrideAtCompileTime==1 && Dst::InnerStrideAtCompileTime==1,
       MightLinearize = MightEnableVml && (int(Dst::Flags) & int(Src::Flags) & LinearAccessBit),
       VmlSize = MightLinearize ? MaxSizeAtCompileTime : InnerMaxSize,
-      LargeEnough = VmlSize==Dynamic || VmlSize>=EIGEN_MKL_VML_THRESHOLD,
-      MayEnableVml = MightEnableVml && LargeEnough,
-      MayLinearize = MayEnableVml && MightLinearize
+      LargeEnough = VmlSize==Dynamic || VmlSize>=EIGEN_MKL_VML_THRESHOLD
     };
   public:
     enum {
-      Traversal = MayLinearize ? LinearVectorizedTraversal
-                : MayEnableVml ? InnerVectorizedTraversal
-                : DefaultTraversal
+      EnableVml = MightEnableVml && LargeEnough,
+      Traversal = MightLinearize ? LinearTraversal : DefaultTraversal
     };
 };
 
-template<typename Derived1, typename Derived2, typename UnaryOp, int Traversal, int Unrolling,
-         int VmlTraversal = vml_assign_traits<Derived1, Derived2, UnaryOp>::Traversal >
-struct vml_assign_impl
-  : assign_impl<Derived1, Eigen::CwiseUnaryOp<UnaryOp, Derived2>,Traversal,Unrolling,BuiltIn>
-{
-};
-
-template<typename Derived1, typename Derived2, typename UnaryOp, int Traversal, int Unrolling>
-struct vml_assign_impl<Derived1, Derived2, UnaryOp, Traversal, Unrolling, InnerVectorizedTraversal>
-{
-  typedef typename Derived1::Scalar Scalar;
-  typedef typename Derived1::Index Index;
-  static inline void run(Derived1& dst, const CwiseUnaryOp<UnaryOp, Derived2>& src)
-  {
-    // in case we want to (or have to) skip VML at runtime we can call:
-    // assign_impl<Derived1,Eigen::CwiseUnaryOp<UnaryOp, Derived2>,Traversal,Unrolling,BuiltIn>::run(dst,src);
-    const Index innerSize = dst.innerSize();
-    const Index outerSize = dst.outerSize();
-    for(Index outer = 0; outer < outerSize; ++outer) {
-      const Scalar *src_ptr = src.IsRowMajor ?  &(src.nestedExpression().coeffRef(outer,0)) :
-                                                &(src.nestedExpression().coeffRef(0, outer));
-      Scalar *dst_ptr = dst.IsRowMajor ? &(dst.coeffRef(outer,0)) : &(dst.coeffRef(0, outer));
-      vml_call<UnaryOp>::run(src.functor(), innerSize, src_ptr, dst_ptr );
-    }
-  }
-};
-
-template<typename Derived1, typename Derived2, typename UnaryOp, int Traversal, int Unrolling>
-struct vml_assign_impl<Derived1, Derived2, UnaryOp, Traversal, Unrolling, LinearVectorizedTraversal>
-{
-  static inline void run(Derived1& dst, const CwiseUnaryOp<UnaryOp, Derived2>& src)
-  {
-    // in case we want to (or have to) skip VML at runtime we can call:
-    // assign_impl<Derived1,Eigen::CwiseUnaryOp<UnaryOp, Derived2>,Traversal,Unrolling,BuiltIn>::run(dst,src);
-    vml_call<UnaryOp>::run(src.functor(), dst.size(), src.nestedExpression().data(), dst.data() );
-  }
-};
-
-// Macroses
-
-#define EIGEN_MKL_VML_SPECIALIZE_ASSIGN(TRAVERSAL,UNROLLING) \
-  template<typename Derived1, typename Derived2, typename UnaryOp> \
-  struct assign_impl<Derived1, Eigen::CwiseUnaryOp<UnaryOp, Derived2>, TRAVERSAL, UNROLLING, Specialized>  {  \
-    static inline void run(Derived1 &dst, const Eigen::CwiseUnaryOp<UnaryOp, Derived2> &src) { \
-      vml_assign_impl<Derived1,Derived2,UnaryOp,TRAVERSAL,UNROLLING>::run(dst, src); \
-    } \
-  };
-
-EIGEN_MKL_VML_SPECIALIZE_ASSIGN(DefaultTraversal,NoUnrolling)
-EIGEN_MKL_VML_SPECIALIZE_ASSIGN(DefaultTraversal,CompleteUnrolling)
-EIGEN_MKL_VML_SPECIALIZE_ASSIGN(DefaultTraversal,InnerUnrolling)
-EIGEN_MKL_VML_SPECIALIZE_ASSIGN(LinearTraversal,NoUnrolling)
-EIGEN_MKL_VML_SPECIALIZE_ASSIGN(LinearTraversal,CompleteUnrolling)
-EIGEN_MKL_VML_SPECIALIZE_ASSIGN(InnerVectorizedTraversal,NoUnrolling)
-EIGEN_MKL_VML_SPECIALIZE_ASSIGN(InnerVectorizedTraversal,CompleteUnrolling)
-EIGEN_MKL_VML_SPECIALIZE_ASSIGN(InnerVectorizedTraversal,InnerUnrolling)
-EIGEN_MKL_VML_SPECIALIZE_ASSIGN(LinearVectorizedTraversal,CompleteUnrolling)
-EIGEN_MKL_VML_SPECIALIZE_ASSIGN(LinearVectorizedTraversal,NoUnrolling)
-EIGEN_MKL_VML_SPECIALIZE_ASSIGN(SliceVectorizedTraversal,NoUnrolling)
-
-
+#define EIGEN_PP_EXPAND(ARG) ARG
 #if !defined (EIGEN_FAST_MATH) || (EIGEN_FAST_MATH != 1)
-#define  EIGEN_MKL_VML_MODE VML_HA
+#define EIGEN_VMLMODE_EXPAND_LA , VML_HA
 #else
-#define  EIGEN_MKL_VML_MODE VML_LA
+#define EIGEN_VMLMODE_EXPAND_LA , VML_LA
 #endif
 
-#define EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, VMLOP, EIGENTYPE, VMLTYPE)     \
-  template<> struct vml_call< scalar_##EIGENOP##_op<EIGENTYPE> > {               \
-    enum { IsSupported = 1 };                                                    \
-    static inline void run( const scalar_##EIGENOP##_op<EIGENTYPE>& /*func*/,        \
-                            int size, const EIGENTYPE* src, EIGENTYPE* dst) {    \
-      VMLOP(size, (const VMLTYPE*)src, (VMLTYPE*)dst);                           \
-    }                                                                            \
+#define EIGEN_VMLMODE_EXPAND__ 
+
+#define EIGEN_VMLMODE_PREFIX_LA vm
+#define EIGEN_VMLMODE_PREFIX__  v
+#define EIGEN_VMLMODE_PREFIX(VMLMODE) EIGEN_CAT(EIGEN_VMLMODE_PREFIX_,VMLMODE)
+
+#define EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, VMLOP, EIGENTYPE, VMLTYPE, VMLMODE)                                           \
+  template< typename DstXprType, typename SrcXprNested>                                                                         \
+  struct Assignment<DstXprType, CwiseUnaryOp<scalar_##EIGENOP##_op<EIGENTYPE>, SrcXprNested>, assign_op<EIGENTYPE>,             \
+                   Dense2Dense, typename enable_if<vml_assign_traits<DstXprType,SrcXprNested>::EnableVml,EIGENTYPE>::type> {    \
+    typedef CwiseUnaryOp<scalar_##EIGENOP##_op<EIGENTYPE>, SrcXprNested> SrcXprType;                                            \
+    static void run(DstXprType &dst, const SrcXprType &src, const assign_op<EIGENTYPE> &/*func*/) {                             \
+      eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols());                                                       \
+      if(vml_assign_traits<DstXprType,SrcXprNested>::Traversal==LinearTraversal) {                                              \
+        VMLOP(dst.size(), (const VMLTYPE*)src.nestedExpression().data(),                                                        \
+              (VMLTYPE*)dst.data() EIGEN_PP_EXPAND(EIGEN_VMLMODE_EXPAND_##VMLMODE) );                                           \
+      } else {                                                                                                                  \
+        const Index outerSize = dst.outerSize();                                                                                \
+        for(Index outer = 0; outer < outerSize; ++outer) {                                                                      \
+          const EIGENTYPE *src_ptr = src.IsRowMajor ? &(src.nestedExpression().coeffRef(outer,0)) :                             \
+                                                      &(src.nestedExpression().coeffRef(0, outer));                             \
+          EIGENTYPE *dst_ptr = dst.IsRowMajor ? &(dst.coeffRef(outer,0)) : &(dst.coeffRef(0, outer));                           \
+          VMLOP( dst.innerSize(), (const VMLTYPE*)src_ptr,                                                                      \
+                (VMLTYPE*)dst_ptr EIGEN_PP_EXPAND(EIGEN_VMLMODE_EXPAND_##VMLMODE));                                             \
+        }                                                                                                                       \
+      }                                                                                                                         \
+    }                                                                                                                           \
+  };                                                                                                                            \
+
+
+#define EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(EIGENOP, VMLOP, VMLMODE)                                                         \
+  EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, EIGEN_CAT(EIGEN_VMLMODE_PREFIX(VMLMODE),s##VMLOP), float, float, VMLMODE)           \
+  EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, EIGEN_CAT(EIGEN_VMLMODE_PREFIX(VMLMODE),d##VMLOP), double, double, VMLMODE)
+
+#define EIGEN_MKL_VML_DECLARE_UNARY_CALLS_CPLX(EIGENOP, VMLOP, VMLMODE)                                                         \
+  EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, EIGEN_CAT(EIGEN_VMLMODE_PREFIX(VMLMODE),c##VMLOP), scomplex, MKL_Complex8, VMLMODE) \
+  EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, EIGEN_CAT(EIGEN_VMLMODE_PREFIX(VMLMODE),z##VMLOP), dcomplex, MKL_Complex16, VMLMODE)
+  
+#define EIGEN_MKL_VML_DECLARE_UNARY_CALLS(EIGENOP, VMLOP, VMLMODE)                                                              \
+  EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(EIGENOP, VMLOP, VMLMODE)                                                               \
+  EIGEN_MKL_VML_DECLARE_UNARY_CALLS_CPLX(EIGENOP, VMLOP, VMLMODE)
+
+  
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(sin,   Sin,   LA)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(asin,  Asin,  LA)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(sinh,  Sinh,  LA)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(cos,   Cos,   LA)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(acos,  Acos,  LA)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(cosh,  Cosh,  LA)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(tan,   Tan,   LA)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(atan,  Atan,  LA)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(tanh,  Tanh,  LA)
+// EIGEN_MKL_VML_DECLARE_UNARY_CALLS(abs,   Abs,    _)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(exp,   Exp,   LA)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(log,   Ln,    LA)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(log10, Log10, LA)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(sqrt,  Sqrt,  _)
+
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(square, Sqr,   _)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS_CPLX(arg, Arg,      _)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(round, Round,  _)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(floor, Floor,  _)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(ceil,  Ceil,   _)
+
+#define EIGEN_MKL_VML_DECLARE_POW_CALL(EIGENOP, VMLOP, EIGENTYPE, VMLTYPE, VMLMODE)                                           \
+  template< typename DstXprType, typename SrcXprNested>                                                                       \
+  struct Assignment<DstXprType, CwiseUnaryOp<scalar_##EIGENOP##_op<EIGENTYPE>, SrcXprNested>, assign_op<EIGENTYPE>,           \
+                   Dense2Dense, typename enable_if<vml_assign_traits<DstXprType,SrcXprNested>::EnableVml,EIGENTYPE>::type> {  \
+    typedef CwiseUnaryOp<scalar_##EIGENOP##_op<EIGENTYPE>, SrcXprNested> SrcXprType;                                          \
+    static void run(DstXprType &dst, const SrcXprType &src, const assign_op<EIGENTYPE> &/*func*/) {                           \
+      eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols());                                                     \
+      VMLTYPE exponent = reinterpret_cast<const VMLTYPE&>(src.functor().m_exponent);                                          \
+      if(vml_assign_traits<DstXprType,SrcXprNested>::Traversal==LinearTraversal)                                              \
+      {                                                                                                                       \
+        VMLOP( dst.size(), (const VMLTYPE*)src.nestedExpression().data(), exponent,                                           \
+              (VMLTYPE*)dst.data() EIGEN_PP_EXPAND(EIGEN_VMLMODE_EXPAND_##VMLMODE) );                                         \
+      } else {                                                                                                                \
+        const Index outerSize = dst.outerSize();                                                                              \
+        for(Index outer = 0; outer < outerSize; ++outer) {                                                                    \
+          const EIGENTYPE *src_ptr = src.IsRowMajor ? &(src.nestedExpression().coeffRef(outer,0)) :                           \
+                                                      &(src.nestedExpression().coeffRef(0, outer));                           \
+          EIGENTYPE *dst_ptr = dst.IsRowMajor ? &(dst.coeffRef(outer,0)) : &(dst.coeffRef(0, outer));                         \
+          VMLOP( dst.innerSize(), (const VMLTYPE*)src_ptr, exponent,                                                          \
+                 (VMLTYPE*)dst_ptr EIGEN_PP_EXPAND(EIGEN_VMLMODE_EXPAND_##VMLMODE));                                          \
+        }                                                                                                                     \
+      }                                                                                                                       \
+    }                                                                                                                         \
   };
-
-#define EIGEN_MKL_VML_DECLARE_UNARY_CALL_LA(EIGENOP, VMLOP, EIGENTYPE, VMLTYPE)  \
-  template<> struct vml_call< scalar_##EIGENOP##_op<EIGENTYPE> > {               \
-    enum { IsSupported = 1 };                                                    \
-    static inline void run( const scalar_##EIGENOP##_op<EIGENTYPE>& /*func*/,        \
-                            int size, const EIGENTYPE* src, EIGENTYPE* dst) {    \
-      MKL_INT64 vmlMode = EIGEN_MKL_VML_MODE;                                    \
-      VMLOP(size, (const VMLTYPE*)src, (VMLTYPE*)dst, vmlMode);                  \
-    }                                                                            \
-  };
-
-#define EIGEN_MKL_VML_DECLARE_POW_CALL(EIGENOP, VMLOP, EIGENTYPE, VMLTYPE)       \
-  template<> struct vml_call< scalar_##EIGENOP##_op<EIGENTYPE> > {               \
-    enum { IsSupported = 1 };                                                    \
-    static inline void run( const scalar_##EIGENOP##_op<EIGENTYPE>& func,        \
-                          int size, const EIGENTYPE* src, EIGENTYPE* dst) {      \
-      EIGENTYPE exponent = func.m_exponent;                                      \
-      MKL_INT64 vmlMode = EIGEN_MKL_VML_MODE;                                    \
-      VMLOP(&size, (const VMLTYPE*)src, (const VMLTYPE*)&exponent,               \
-                        (VMLTYPE*)dst, &vmlMode);                                \
-    }                                                                            \
-  };
-
-#define EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(EIGENOP, VMLOP)                   \
-  EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, vs##VMLOP, float, float)             \
-  EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, vd##VMLOP, double, double)
-
-#define EIGEN_MKL_VML_DECLARE_UNARY_CALLS_COMPLEX(EIGENOP, VMLOP)                \
-  EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, vc##VMLOP, scomplex, MKL_Complex8)   \
-  EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, vz##VMLOP, dcomplex, MKL_Complex16)
-
-#define EIGEN_MKL_VML_DECLARE_UNARY_CALLS(EIGENOP, VMLOP)                        \
-  EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(EIGENOP, VMLOP)                         \
-  EIGEN_MKL_VML_DECLARE_UNARY_CALLS_COMPLEX(EIGENOP, VMLOP)
-
-
-#define EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL_LA(EIGENOP, VMLOP)                \
-  EIGEN_MKL_VML_DECLARE_UNARY_CALL_LA(EIGENOP, vms##VMLOP, float, float)         \
-  EIGEN_MKL_VML_DECLARE_UNARY_CALL_LA(EIGENOP, vmd##VMLOP, double, double)
-
-#define EIGEN_MKL_VML_DECLARE_UNARY_CALLS_COMPLEX_LA(EIGENOP, VMLOP)             \
-  EIGEN_MKL_VML_DECLARE_UNARY_CALL_LA(EIGENOP, vmc##VMLOP, scomplex, MKL_Complex8)  \
-  EIGEN_MKL_VML_DECLARE_UNARY_CALL_LA(EIGENOP, vmz##VMLOP, dcomplex, MKL_Complex16)
-
-#define EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(EIGENOP, VMLOP)                     \
-  EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL_LA(EIGENOP, VMLOP)                      \
-  EIGEN_MKL_VML_DECLARE_UNARY_CALLS_COMPLEX_LA(EIGENOP, VMLOP)
-
-
-EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(sin,  Sin)
-EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(asin, Asin)
-EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(cos,  Cos)
-EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(acos, Acos)
-EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(tan,  Tan)
-//EIGEN_MKL_VML_DECLARE_UNARY_CALLS(abs,  Abs)
-EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(exp,  Exp)
-EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(log,  Ln)
-EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(sqrt, Sqrt)
-
-EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(square, Sqr)
-
-// The vm*powx functions are not avaibale in the windows version of MKL.
-#ifndef _WIN32
-EIGEN_MKL_VML_DECLARE_POW_CALL(pow, vmspowx_, float, float)
-EIGEN_MKL_VML_DECLARE_POW_CALL(pow, vmdpowx_, double, double)
-EIGEN_MKL_VML_DECLARE_POW_CALL(pow, vmcpowx_, scomplex, MKL_Complex8)
-EIGEN_MKL_VML_DECLARE_POW_CALL(pow, vmzpowx_, dcomplex, MKL_Complex16)
-#endif
+  
+EIGEN_MKL_VML_DECLARE_POW_CALL(pow, vmsPowx, float,    float,         LA)
+EIGEN_MKL_VML_DECLARE_POW_CALL(pow, vmdPowx, double,   double,        LA)
+EIGEN_MKL_VML_DECLARE_POW_CALL(pow, vmcPowx, scomplex, MKL_Complex8,  LA)
+EIGEN_MKL_VML_DECLARE_POW_CALL(pow, vmzPowx, dcomplex, MKL_Complex16, LA)
 
 } // end namespace internal
 

Eigen/src/Core/BandMatrix.h

@@ -32,7 +32,7 @@ class BandMatrixBase : public EigenBase<Derived>
     };
     typedef typename internal::traits<Derived>::Scalar Scalar;
     typedef Matrix<Scalar,RowsAtCompileTime,ColsAtCompileTime> DenseMatrixType;
-    typedef typename DenseMatrixType::Index Index;
+    typedef typename DenseMatrixType::StorageIndex StorageIndex;
     typedef typename internal::traits<Derived>::CoefficientsType CoefficientsType;
     typedef EigenBase<Derived> Base;
 
@@ -179,7 +179,7 @@ struct traits<BandMatrix<_Scalar,_Rows,_Cols,_Supers,_Subs,_Options> >
 {
   typedef _Scalar Scalar;
   typedef Dense StorageKind;
-  typedef DenseIndex Index;
+  typedef Eigen::Index StorageIndex;
   enum {
     CoeffReadCost = NumTraits<Scalar>::ReadCost,
     RowsAtCompileTime = _Rows,
@@ -201,10 +201,10 @@ class BandMatrix : public BandMatrixBase<BandMatrix<_Scalar,Rows,Cols,Supers,Sub
   public:
 
     typedef typename internal::traits<BandMatrix>::Scalar Scalar;
-    typedef typename internal::traits<BandMatrix>::Index Index;
+    typedef typename internal::traits<BandMatrix>::StorageIndex StorageIndex;
     typedef typename internal::traits<BandMatrix>::CoefficientsType CoefficientsType;
 
-    inline BandMatrix(Index rows=Rows, Index cols=Cols, Index supers=Supers, Index subs=Subs)
+    explicit inline BandMatrix(Index rows=Rows, Index cols=Cols, Index supers=Supers, Index subs=Subs)
       : m_coeffs(1+supers+subs,cols),
         m_rows(rows), m_supers(supers), m_subs(subs)
     {
@@ -241,7 +241,7 @@ struct traits<BandMatrixWrapper<_CoefficientsType,_Rows,_Cols,_Supers,_Subs,_Opt
 {
   typedef typename _CoefficientsType::Scalar Scalar;
   typedef typename _CoefficientsType::StorageKind StorageKind;
-  typedef typename _CoefficientsType::Index Index;
+  typedef typename _CoefficientsType::StorageIndex StorageIndex;
   enum {
     CoeffReadCost = internal::traits<_CoefficientsType>::CoeffReadCost,
     RowsAtCompileTime = _Rows,
@@ -264,9 +264,9 @@ class BandMatrixWrapper : public BandMatrixBase<BandMatrixWrapper<_CoefficientsT
 
     typedef typename internal::traits<BandMatrixWrapper>::Scalar Scalar;
     typedef typename internal::traits<BandMatrixWrapper>::CoefficientsType CoefficientsType;
-    typedef typename internal::traits<BandMatrixWrapper>::Index Index;
+    typedef typename internal::traits<BandMatrixWrapper>::StorageIndex StorageIndex;
 
-    inline BandMatrixWrapper(const CoefficientsType& coeffs, Index rows=_Rows, Index cols=_Cols, Index supers=_Supers, Index subs=_Subs)
+    explicit inline BandMatrixWrapper(const CoefficientsType& coeffs, Index rows=_Rows, Index cols=_Cols, Index supers=_Supers, Index subs=_Subs)
       : m_coeffs(coeffs),
         m_rows(rows), m_supers(supers), m_subs(subs)
     {
@@ -312,9 +312,9 @@ template<typename Scalar, int Size, int Options>
 class TridiagonalMatrix : public BandMatrix<Scalar,Size,Size,Options&SelfAdjoint?0:1,1,Options|RowMajor>
 {
     typedef BandMatrix<Scalar,Size,Size,Options&SelfAdjoint?0:1,1,Options|RowMajor> Base;
-    typedef typename Base::Index Index;
+    typedef typename Base::StorageIndex StorageIndex;
   public:
-    TridiagonalMatrix(Index size = Size) : Base(size,size,Options&SelfAdjoint?0:1,1) {}
+    explicit TridiagonalMatrix(Index size = Size) : Base(size,size,Options&SelfAdjoint?0:1,1) {}
 
     inline typename Base::template DiagonalIntReturnType<1>::Type super()
     { return Base::template diagonal<1>(); }
@@ -327,6 +327,25 @@ class TridiagonalMatrix : public BandMatrix<Scalar,Size,Size,Options&SelfAdjoint
   protected:
 };
 
+
+struct BandShape {};
+
+template<typename _Scalar, int _Rows, int _Cols, int _Supers, int _Subs, int _Options>
+struct evaluator_traits<BandMatrix<_Scalar,_Rows,_Cols,_Supers,_Subs,_Options> >
+  : public evaluator_traits_base<BandMatrix<_Scalar,_Rows,_Cols,_Supers,_Subs,_Options> >
+{
+  typedef BandShape Shape;
+};
+
+template<typename _CoefficientsType,int _Rows, int _Cols, int _Supers, int _Subs,int _Options>
+struct evaluator_traits<BandMatrixWrapper<_CoefficientsType,_Rows,_Cols,_Supers,_Subs,_Options> >
+  : public evaluator_traits_base<BandMatrixWrapper<_CoefficientsType,_Rows,_Cols,_Supers,_Subs,_Options> >
+{
+  typedef BandShape Shape;
+};
+
+template<> struct AssignmentKind<DenseShape,BandShape> { typedef EigenBase2EigenBase Kind; };
+
 } // end namespace internal
 
 } // end namespace Eigen

Eigen/src/Core/Block.h

@@ -21,6 +21,9 @@ namespace Eigen {
   * \param XprType the type of the expression in which we are taking a block
   * \param BlockRows the number of rows of the block we are taking at compile time (optional)
   * \param BlockCols the number of columns of the block we are taking at compile time (optional)
+  * \param InnerPanel is true, if the block maps to a set of rows of a row major matrix or
+  *        to set of columns of a column major matrix (optional). The parameter allows to determine
+  *        at compile time whether aligned access is possible on the block expression.
   *
   * This class represents an expression of either a fixed-size or dynamic-size block. It is the return
   * type of DenseBase::block(Index,Index,Index,Index) and DenseBase::block<int,int>(Index,Index) and
@@ -52,7 +55,7 @@ struct traits<Block<XprType, BlockRows, BlockCols, InnerPanel> > : traits<XprTyp
   typedef typename traits<XprType>::Scalar Scalar;
   typedef typename traits<XprType>::StorageKind StorageKind;
   typedef typename traits<XprType>::XprKind XprKind;
-  typedef typename nested<XprType>::type XprTypeNested;
+  typedef typename ref_selector<XprType>::type XprTypeNested;
   typedef typename remove_reference<XprTypeNested>::type _XprTypeNested;
   enum{
     MatrixRows = traits<XprType>::RowsAtCompileTime,
@@ -65,6 +68,7 @@ struct traits<Block<XprType, BlockRows, BlockCols, InnerPanel> > : traits<XprTyp
     MaxColsAtCompileTime = BlockCols==0 ? 0
                          : ColsAtCompileTime != Dynamic ? int(ColsAtCompileTime)
                          : int(traits<XprType>::MaxColsAtCompileTime),
+
     XprTypeIsRowMajor = (int(traits<XprType>::Flags)&RowMajorBit) != 0,
     IsRowMajor = (MaxRowsAtCompileTime==1&&MaxColsAtCompileTime!=1) ? 1
                : (MaxColsAtCompileTime==1&&MaxRowsAtCompileTime!=1) ? 0
@@ -77,18 +81,16 @@ struct traits<Block<XprType, BlockRows, BlockCols, InnerPanel> > : traits<XprTyp
     OuterStrideAtCompileTime = HasSameStorageOrderAsXprType
                              ? int(outer_stride_at_compile_time<XprType>::ret)
                              : int(inner_stride_at_compile_time<XprType>::ret),
-    MaskPacketAccessBit = (InnerSize == Dynamic || (InnerSize % packet_traits<Scalar>::size) == 0)
-                       && (InnerStrideAtCompileTime == 1)
-                        ? PacketAccessBit : 0,
-    MaskAlignedBit = (InnerPanel && (OuterStrideAtCompileTime!=Dynamic) && (((OuterStrideAtCompileTime * int(sizeof(Scalar))) % 16) == 0)) ? AlignedBit : 0,
-    FlagsLinearAccessBit = (RowsAtCompileTime == 1 || ColsAtCompileTime == 1) ? LinearAccessBit : 0,
+
+    // FIXME, this traits is rather specialized for dense object and it needs to be cleaned further
     FlagsLvalueBit = is_lvalue<XprType>::value ? LvalueBit : 0,
     FlagsRowMajorBit = IsRowMajor ? RowMajorBit : 0,
-    Flags0 = traits<XprType>::Flags & ( (HereditaryBits & ~RowMajorBit) |
-                                        DirectAccessBit |
-                                        MaskPacketAccessBit |
-                                        MaskAlignedBit),
-    Flags = Flags0 | FlagsLinearAccessBit | FlagsLvalueBit | FlagsRowMajorBit
+    Flags = (traits<XprType>::Flags & (DirectAccessBit | (InnerPanel?CompressedAccessBit:0))) | FlagsLvalueBit | FlagsRowMajorBit,
+    // FIXME DirectAccessBit should not be handled by expressions
+    // 
+    // Alignment is needed by MapBase's assertions
+    // We can sefely set it to false here. Internal alignment errors will be detected by an eigen_internal_assert in the respective evaluator
+    Alignment = 0
   };
 };
 
@@ -108,9 +110,12 @@ template<typename XprType, int BlockRows, int BlockCols, bool InnerPanel> class
     typedef Impl Base;
     EIGEN_GENERIC_PUBLIC_INTERFACE(Block)
     EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Block)
+    
+    typedef typename internal::remove_all<XprType>::type NestedExpression;
   
     /** Column or Row constructor
       */
+    EIGEN_DEVICE_FUNC
     inline Block(XprType& xpr, Index i) : Impl(xpr,i)
     {
       eigen_assert( (i>=0) && (
@@ -120,25 +125,27 @@ template<typename XprType, int BlockRows, int BlockCols, bool InnerPanel> class
 
     /** Fixed-size constructor
       */
-    inline Block(XprType& xpr, Index a_startRow, Index a_startCol)
-      : Impl(xpr, a_startRow, a_startCol)
+    EIGEN_DEVICE_FUNC
+    inline Block(XprType& xpr, Index startRow, Index startCol)
+      : Impl(xpr, startRow, startCol)
     {
       EIGEN_STATIC_ASSERT(RowsAtCompileTime!=Dynamic && ColsAtCompileTime!=Dynamic,THIS_METHOD_IS_ONLY_FOR_FIXED_SIZE)
-      eigen_assert(a_startRow >= 0 && BlockRows >= 1 && a_startRow + BlockRows <= xpr.rows()
-             && a_startCol >= 0 && BlockCols >= 1 && a_startCol + BlockCols <= xpr.cols());
+      eigen_assert(startRow >= 0 && BlockRows >= 1 && startRow + BlockRows <= xpr.rows()
+             && startCol >= 0 && BlockCols >= 1 && startCol + BlockCols <= xpr.cols());
     }
 
     /** Dynamic-size constructor
       */
+    EIGEN_DEVICE_FUNC
     inline Block(XprType& xpr,
-          Index a_startRow, Index a_startCol,
+          Index startRow, Index startCol,
           Index blockRows, Index blockCols)
-      : Impl(xpr, a_startRow, a_startCol, blockRows, blockCols)
+      : Impl(xpr, startRow, startCol, blockRows, blockCols)
     {
       eigen_assert((RowsAtCompileTime==Dynamic || RowsAtCompileTime==blockRows)
           && (ColsAtCompileTime==Dynamic || ColsAtCompileTime==blockCols));
-      eigen_assert(a_startRow >= 0 && blockRows >= 0 && a_startRow  <= xpr.rows() - blockRows
-          && a_startCol >= 0 && blockCols >= 0 && a_startCol <= xpr.cols() - blockCols);
+      eigen_assert(startRow >= 0 && blockRows >= 0 && startRow  <= xpr.rows() - blockRows
+          && startCol >= 0 && blockCols >= 0 && startCol <= xpr.cols() - blockCols);
     }
 };
          
@@ -149,14 +156,15 @@ class BlockImpl<XprType, BlockRows, BlockCols, InnerPanel, Dense>
   : public internal::BlockImpl_dense<XprType, BlockRows, BlockCols, InnerPanel>
 {
     typedef internal::BlockImpl_dense<XprType, BlockRows, BlockCols, InnerPanel> Impl;
-    typedef typename XprType::Index Index;
+    typedef typename XprType::StorageIndex StorageIndex;
   public:
     typedef Impl Base;
     EIGEN_INHERIT_ASSIGNMENT_OPERATORS(BlockImpl)
-    inline BlockImpl(XprType& xpr, Index i) : Impl(xpr,i) {}
-    inline BlockImpl(XprType& xpr, Index a_startRow, Index a_startCol) : Impl(xpr, a_startRow, a_startCol) {}
-    inline BlockImpl(XprType& xpr, Index a_startRow, Index a_startCol, Index blockRows, Index blockCols)
-      : Impl(xpr, a_startRow, a_startCol, blockRows, blockCols) {}
+    EIGEN_DEVICE_FUNC inline BlockImpl(XprType& xpr, Index i) : Impl(xpr,i) {}
+    EIGEN_DEVICE_FUNC inline BlockImpl(XprType& xpr, Index startRow, Index startCol) : Impl(xpr, startRow, startCol) {}
+    EIGEN_DEVICE_FUNC
+    inline BlockImpl(XprType& xpr, Index startRow, Index startCol, Index blockRows, Index blockCols)
+      : Impl(xpr, startRow, startCol, blockRows, blockCols) {}
 };
 
 namespace internal {
@@ -172,10 +180,11 @@ template<typename XprType, int BlockRows, int BlockCols, bool InnerPanel, bool H
     EIGEN_DENSE_PUBLIC_INTERFACE(BlockType)
     EIGEN_INHERIT_ASSIGNMENT_OPERATORS(BlockImpl_dense)
 
-    class InnerIterator;
+    // class InnerIterator; // FIXME apparently never used
 
     /** Column or Row constructor
       */
+    EIGEN_DEVICE_FUNC
     inline BlockImpl_dense(XprType& xpr, Index i)
       : m_xpr(xpr),
         // It is a row if and only if BlockRows==1 and BlockCols==XprType::ColsAtCompileTime,
@@ -190,23 +199,26 @@ template<typename XprType, int BlockRows, int BlockCols, bool InnerPanel, bool H
 
     /** Fixed-size constructor
       */
-    inline BlockImpl_dense(XprType& xpr, Index a_startRow, Index a_startCol)
-      : m_xpr(xpr), m_startRow(a_startRow), m_startCol(a_startCol),
+    EIGEN_DEVICE_FUNC
+    inline BlockImpl_dense(XprType& xpr, Index startRow, Index startCol)
+      : m_xpr(xpr), m_startRow(startRow), m_startCol(startCol),
                     m_blockRows(BlockRows), m_blockCols(BlockCols)
     {}
 
     /** Dynamic-size constructor
       */
+    EIGEN_DEVICE_FUNC
     inline BlockImpl_dense(XprType& xpr,
-          Index a_startRow, Index a_startCol,
+          Index startRow, Index startCol,
           Index blockRows, Index blockCols)
-      : m_xpr(xpr), m_startRow(a_startRow), m_startCol(a_startCol),
+      : m_xpr(xpr), m_startRow(startRow), m_startCol(startCol),
                     m_blockRows(blockRows), m_blockCols(blockCols)
     {}
 
-    inline Index rows() const { return m_blockRows.value(); }
-    inline Index cols() const { return m_blockCols.value(); }
+    EIGEN_DEVICE_FUNC inline Index rows() const { return m_blockRows.value(); }
+    EIGEN_DEVICE_FUNC inline Index cols() const { return m_blockCols.value(); }
 
+    EIGEN_DEVICE_FUNC
     inline Scalar& coeffRef(Index rowId, Index colId)
     {
       EIGEN_STATIC_ASSERT_LVALUE(XprType)
@@ -214,17 +226,20 @@ template<typename XprType, int BlockRows, int BlockCols, bool InnerPanel, bool H
                .coeffRef(rowId + m_startRow.value(), colId + m_startCol.value());
     }
 
+    EIGEN_DEVICE_FUNC
     inline const Scalar& coeffRef(Index rowId, Index colId) const
     {
       return m_xpr.derived()
                .coeffRef(rowId + m_startRow.value(), colId + m_startCol.value());
     }
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE const CoeffReturnType coeff(Index rowId, Index colId) const
     {
       return m_xpr.coeff(rowId + m_startRow.value(), colId + m_startCol.value());
     }
 
+    EIGEN_DEVICE_FUNC
     inline Scalar& coeffRef(Index index)
     {
       EIGEN_STATIC_ASSERT_LVALUE(XprType)
@@ -233,6 +248,7 @@ template<typename XprType, int BlockRows, int BlockCols, bool InnerPanel, bool H
                        m_startCol.value() + (RowsAtCompileTime == 1 ? index : 0));
     }
 
+    EIGEN_DEVICE_FUNC
     inline const Scalar& coeffRef(Index index) const
     {
       return m_xpr.const_cast_derived()
@@ -240,6 +256,7 @@ template<typename XprType, int BlockRows, int BlockCols, bool InnerPanel, bool H
                        m_startCol.value() + (RowsAtCompileTime == 1 ? index : 0));
     }
 
+    EIGEN_DEVICE_FUNC
     inline const CoeffReturnType coeff(Index index) const
     {
       return m_xpr
@@ -279,22 +296,25 @@ template<typename XprType, int BlockRows, int BlockCols, bool InnerPanel, bool H
 
     #ifdef EIGEN_PARSED_BY_DOXYGEN
     /** \sa MapBase::data() */
-    inline const Scalar* data() const;
-    inline Index innerStride() const;
-    inline Index outerStride() const;
+    EIGEN_DEVICE_FUNC inline const Scalar* data() const;
+    EIGEN_DEVICE_FUNC inline Index innerStride() const;
+    EIGEN_DEVICE_FUNC inline Index outerStride() const;
     #endif
 
-    const typename internal::remove_all<typename XprType::Nested>::type& nestedExpression() const 
+    EIGEN_DEVICE_FUNC
+    const typename internal::remove_all<typename XprType::Nested>::type& nestedExpression() const
     { 
       return m_xpr; 
     }
       
-    Index startRow() const 
+    EIGEN_DEVICE_FUNC
+    StorageIndex startRow() const
     { 
       return m_startRow.value(); 
     }
       
-    Index startCol() const 
+    EIGEN_DEVICE_FUNC
+    StorageIndex startCol() const
     { 
       return m_startCol.value(); 
     }
@@ -302,10 +322,10 @@ template<typename XprType, int BlockRows, int BlockCols, bool InnerPanel, bool H
   protected:
 
     const typename XprType::Nested m_xpr;
-    const internal::variable_if_dynamic<Index, XprType::RowsAtCompileTime == 1 ? 0 : Dynamic> m_startRow;
-    const internal::variable_if_dynamic<Index, XprType::ColsAtCompileTime == 1 ? 0 : Dynamic> m_startCol;
-    const internal::variable_if_dynamic<Index, RowsAtCompileTime> m_blockRows;
-    const internal::variable_if_dynamic<Index, ColsAtCompileTime> m_blockCols;
+    const internal::variable_if_dynamic<StorageIndex, XprType::RowsAtCompileTime == 1 ? 0 : Dynamic> m_startRow;
+    const internal::variable_if_dynamic<StorageIndex, XprType::ColsAtCompileTime == 1 ? 0 : Dynamic> m_startCol;
+    const internal::variable_if_dynamic<StorageIndex, RowsAtCompileTime> m_blockRows;
+    const internal::variable_if_dynamic<StorageIndex, ColsAtCompileTime> m_blockCols;
 };
 
 /** \internal Internal implementation of dense Blocks in the direct access case.*/
@@ -314,6 +334,9 @@ class BlockImpl_dense<XprType,BlockRows,BlockCols, InnerPanel,true>
   : public MapBase<Block<XprType, BlockRows, BlockCols, InnerPanel> >
 {
     typedef Block<XprType, BlockRows, BlockCols, InnerPanel> BlockType;
+    enum {
+      XprTypeIsRowMajor = (int(traits<XprType>::Flags)&RowMajorBit) != 0
+    };
   public:
 
     typedef MapBase<BlockType> Base;
@@ -322,10 +345,10 @@ class BlockImpl_dense<XprType,BlockRows,BlockCols, InnerPanel,true>
 
     /** Column or Row constructor
       */
+    EIGEN_DEVICE_FUNC
     inline BlockImpl_dense(XprType& xpr, Index i)
-      : Base(internal::const_cast_ptr(&xpr.coeffRef(
-              (BlockRows==1) && (BlockCols==XprType::ColsAtCompileTime) ? i : 0,
-              (BlockRows==XprType::RowsAtCompileTime) && (BlockCols==1) ? i : 0)),
+      : Base(xpr.data() + i * (    ((BlockRows==1) && (BlockCols==XprType::ColsAtCompileTime) && (!XprTypeIsRowMajor)) 
+                                || ((BlockRows==XprType::RowsAtCompileTime) && (BlockCols==1) && ( XprTypeIsRowMajor)) ? xpr.innerStride() : xpr.outerStride()),
              BlockRows==1 ? 1 : xpr.rows(),
              BlockCols==1 ? 1 : xpr.cols()),
         m_xpr(xpr)
@@ -335,29 +358,34 @@ class BlockImpl_dense<XprType,BlockRows,BlockCols, InnerPanel,true>
 
     /** Fixed-size constructor
       */
+    EIGEN_DEVICE_FUNC
     inline BlockImpl_dense(XprType& xpr, Index startRow, Index startCol)
-      : Base(internal::const_cast_ptr(&xpr.coeffRef(startRow,startCol))), m_xpr(xpr)
+      : Base(xpr.data()+xpr.innerStride()*(XprTypeIsRowMajor?startCol:startRow) + xpr.outerStride()*(XprTypeIsRowMajor?startRow:startCol)),
+        m_xpr(xpr)
     {
       init();
     }
 
     /** Dynamic-size constructor
       */
+    EIGEN_DEVICE_FUNC
     inline BlockImpl_dense(XprType& xpr,
           Index startRow, Index startCol,
           Index blockRows, Index blockCols)
-      : Base(internal::const_cast_ptr(&xpr.coeffRef(startRow,startCol)), blockRows, blockCols),
+      : Base(xpr.data()+xpr.innerStride()*(XprTypeIsRowMajor?startCol:startRow) + xpr.outerStride()*(XprTypeIsRowMajor?startRow:startCol), blockRows, blockCols),
         m_xpr(xpr)
     {
       init();
     }
 
-    const typename internal::remove_all<typename XprType::Nested>::type& nestedExpression() const 
+    EIGEN_DEVICE_FUNC
+    const typename internal::remove_all<typename XprType::Nested>::type& nestedExpression() const
     { 
       return m_xpr; 
     }
       
     /** \sa MapBase::innerStride() */
+    EIGEN_DEVICE_FUNC
     inline Index innerStride() const
     {
       return internal::traits<BlockType>::HasSameStorageOrderAsXprType
@@ -366,6 +394,7 @@ class BlockImpl_dense<XprType,BlockRows,BlockCols, InnerPanel,true>
     }
 
     /** \sa MapBase::outerStride() */
+    EIGEN_DEVICE_FUNC
     inline Index outerStride() const
     {
       return m_outerStride;
@@ -379,6 +408,7 @@ class BlockImpl_dense<XprType,BlockRows,BlockCols, InnerPanel,true>
 
     #ifndef EIGEN_PARSED_BY_DOXYGEN
     /** \internal used by allowAligned() */
+    EIGEN_DEVICE_FUNC
     inline BlockImpl_dense(XprType& xpr, const Scalar* data, Index blockRows, Index blockCols)
       : Base(data, blockRows, blockCols), m_xpr(xpr)
     {
@@ -387,6 +417,7 @@ class BlockImpl_dense<XprType,BlockRows,BlockCols, InnerPanel,true>
     #endif
 
   protected:
+    EIGEN_DEVICE_FUNC
     void init()
     {
       m_outerStride = internal::traits<BlockType>::HasSameStorageOrderAsXprType

Eigen/src/Core/BooleanRedux.h

@@ -17,9 +17,10 @@ namespace internal {
 template<typename Derived, int UnrollCount>
 struct all_unroller
 {
+  typedef typename Derived::ExpressionTraits Traits;
   enum {
-    col = (UnrollCount-1) / Derived::RowsAtCompileTime,
-    row = (UnrollCount-1) % Derived::RowsAtCompileTime
+    col = (UnrollCount-1) / Traits::RowsAtCompileTime,
+    row = (UnrollCount-1) % Traits::RowsAtCompileTime
   };
 
   static inline bool run(const Derived &mat)
@@ -29,9 +30,9 @@ struct all_unroller
 };
 
 template<typename Derived>
-struct all_unroller<Derived, 1>
+struct all_unroller<Derived, 0>
 {
-  static inline bool run(const Derived &mat) { return mat.coeff(0, 0); }
+  static inline bool run(const Derived &/*mat*/) { return true; }
 };
 
 template<typename Derived>
@@ -43,11 +44,12 @@ struct all_unroller<Derived, Dynamic>
 template<typename Derived, int UnrollCount>
 struct any_unroller
 {
+  typedef typename Derived::ExpressionTraits Traits;
   enum {
-    col = (UnrollCount-1) / Derived::RowsAtCompileTime,
-    row = (UnrollCount-1) % Derived::RowsAtCompileTime
+    col = (UnrollCount-1) / Traits::RowsAtCompileTime,
+    row = (UnrollCount-1) % Traits::RowsAtCompileTime
   };
-
+  
   static inline bool run(const Derived &mat)
   {
     return any_unroller<Derived, UnrollCount-1>::run(mat) || mat.coeff(row, col);
@@ -55,9 +57,9 @@ struct any_unroller
 };
 
 template<typename Derived>
-struct any_unroller<Derived, 1>
+struct any_unroller<Derived, 0>
 {
-  static inline bool run(const Derived &mat) { return mat.coeff(0, 0); }
+  static inline bool run(const Derived & /*mat*/) { return false; }
 };
 
 template<typename Derived>
@@ -78,19 +80,21 @@ struct any_unroller<Derived, Dynamic>
 template<typename Derived>
 inline bool DenseBase<Derived>::all() const
 {
+  typedef internal::evaluator<Derived> Evaluator;
   enum {
     unroll = SizeAtCompileTime != Dynamic
-          && CoeffReadCost != Dynamic
+          && Evaluator::CoeffReadCost != Dynamic
           && NumTraits<Scalar>::AddCost != Dynamic
-          && SizeAtCompileTime * (CoeffReadCost + NumTraits<Scalar>::AddCost) <= EIGEN_UNROLLING_LIMIT
+          && SizeAtCompileTime * (Evaluator::CoeffReadCost + NumTraits<Scalar>::AddCost) <= EIGEN_UNROLLING_LIMIT
   };
+  Evaluator evaluator(derived());
   if(unroll)
-    return internal::all_unroller<Derived, unroll ? int(SizeAtCompileTime) : Dynamic>::run(derived());
+    return internal::all_unroller<Evaluator, unroll ? int(SizeAtCompileTime) : Dynamic>::run(evaluator);
   else
   {
     for(Index j = 0; j < cols(); ++j)
       for(Index i = 0; i < rows(); ++i)
-        if (!coeff(i, j)) return false;
+        if (!evaluator.coeff(i, j)) return false;
     return true;
   }
 }
@@ -102,19 +106,21 @@ inline bool DenseBase<Derived>::all() const
 template<typename Derived>
 inline bool DenseBase<Derived>::any() const
 {
+  typedef internal::evaluator<Derived> Evaluator;
   enum {
     unroll = SizeAtCompileTime != Dynamic
-          && CoeffReadCost != Dynamic
+          && Evaluator::CoeffReadCost != Dynamic
           && NumTraits<Scalar>::AddCost != Dynamic
-          && SizeAtCompileTime * (CoeffReadCost + NumTraits<Scalar>::AddCost) <= EIGEN_UNROLLING_LIMIT
+          && SizeAtCompileTime * (Evaluator::CoeffReadCost + NumTraits<Scalar>::AddCost) <= EIGEN_UNROLLING_LIMIT
   };
+  Evaluator evaluator(derived());
   if(unroll)
-    return internal::any_unroller<Derived, unroll ? int(SizeAtCompileTime) : Dynamic>::run(derived());
+    return internal::any_unroller<Evaluator, unroll ? int(SizeAtCompileTime) : Dynamic>::run(evaluator);
   else
   {
     for(Index j = 0; j < cols(); ++j)
       for(Index i = 0; i < rows(); ++i)
-        if (coeff(i, j)) return true;
+        if (evaluator.coeff(i, j)) return true;
     return false;
   }
 }
@@ -124,14 +130,14 @@ inline bool DenseBase<Derived>::any() const
   * \sa all(), any()
   */
 template<typename Derived>
-inline typename DenseBase<Derived>::Index DenseBase<Derived>::count() const
+inline Eigen::Index DenseBase<Derived>::count() const
 {
   return derived().template cast<bool>().template cast<Index>().sum();
 }
 
 /** \returns true is \c *this contains at least one Not A Number (NaN).
   *
-  * \sa isFinite()
+  * \sa allFinite()
   */
 template<typename Derived>
 inline bool DenseBase<Derived>::hasNaN() const
@@ -144,7 +150,7 @@ inline bool DenseBase<Derived>::hasNaN() const
   * \sa hasNaN()
   */
 template<typename Derived>
-inline bool DenseBase<Derived>::isFinite() const
+inline bool DenseBase<Derived>::allFinite() const
 {
   return !((derived()-derived()).hasNaN());
 }

Eigen/src/Core/CMakeLists.txt

@@ -8,3 +8,4 @@ INSTALL(FILES
 ADD_SUBDIRECTORY(products)
 ADD_SUBDIRECTORY(util)
 ADD_SUBDIRECTORY(arch)
+ADD_SUBDIRECTORY(functors)

Eigen/src/Core/CommaInitializer.h

@@ -28,8 +28,8 @@ template<typename XprType>
 struct CommaInitializer
 {
   typedef typename XprType::Scalar Scalar;
-  typedef typename XprType::Index Index;
 
+  EIGEN_DEVICE_FUNC
   inline CommaInitializer(XprType& xpr, const Scalar& s)
     : m_xpr(xpr), m_row(0), m_col(1), m_currentBlockRows(1)
   {
@@ -37,13 +37,27 @@ struct CommaInitializer
   }
 
   template<typename OtherDerived>
+  EIGEN_DEVICE_FUNC
   inline CommaInitializer(XprType& xpr, const DenseBase<OtherDerived>& other)
     : m_xpr(xpr), m_row(0), m_col(other.cols()), m_currentBlockRows(other.rows())
   {
     m_xpr.block(0, 0, other.rows(), other.cols()) = other;
   }
 
+  /* Copy/Move constructor which transfers ownership. This is crucial in 
+   * absence of return value optimization to avoid assertions during destruction. */
+  // FIXME in C++11 mode this could be replaced by a proper RValue constructor
+  EIGEN_DEVICE_FUNC
+  inline CommaInitializer(const CommaInitializer& o)
+  : m_xpr(o.m_xpr), m_row(o.m_row), m_col(o.m_col), m_currentBlockRows(o.m_currentBlockRows) {
+    // Mark original object as finished. In absence of R-value references we need to const_cast:
+    const_cast<CommaInitializer&>(o).m_row = m_xpr.rows();
+    const_cast<CommaInitializer&>(o).m_col = m_xpr.cols();
+    const_cast<CommaInitializer&>(o).m_currentBlockRows = 0;
+  }
+
   /* inserts a scalar value in the target matrix */
+  EIGEN_DEVICE_FUNC
   CommaInitializer& operator,(const Scalar& s)
   {
     if (m_col==m_xpr.cols())
@@ -63,6 +77,7 @@ struct CommaInitializer
 
   /* inserts a matrix expression in the target matrix */
   template<typename OtherDerived>
+  EIGEN_DEVICE_FUNC
   CommaInitializer& operator,(const DenseBase<OtherDerived>& other)
   {
     if(other.cols()==0 || other.rows()==0)
@@ -88,7 +103,11 @@ struct CommaInitializer
     return *this;
   }
 
+  EIGEN_DEVICE_FUNC
   inline ~CommaInitializer()
+#if defined VERIFY_RAISES_ASSERT && (!defined EIGEN_NO_ASSERTION_CHECKING) && defined EIGEN_EXCEPTIONS
+  throw(Eigen::eigen_assert_exception)
+#endif
   {
     eigen_assert((m_row+m_currentBlockRows) == m_xpr.rows()
          && m_col == m_xpr.cols()
@@ -102,9 +121,10 @@ struct CommaInitializer
     * quaternion.fromRotationMatrix((Matrix3f() << axis0, axis1, axis2).finished());
     * \endcode
     */
+  EIGEN_DEVICE_FUNC
   inline XprType& finished() { return m_xpr; }
 
-  XprType& m_xpr;   // target expression
+  XprType& m_xpr;           // target expression
   Index m_row;              // current row id
   Index m_col;              // current col id
   Index m_currentBlockRows; // current block height

Eigen/src/Core/CoreIterators.h

@@ -1,7 +1,7 @@
 // This file is part of Eigen, a lightweight C++ template library
 // for linear algebra.
 //
-// Copyright (C) 2008-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2008-2014 Gael Guennebaud <gael.guennebaud@inria.fr>
 //
 // This Source Code Form is subject to the terms of the Mozilla
 // Public License v. 2.0. If a copy of the MPL was not distributed
@@ -15,47 +15,113 @@ namespace Eigen {
 /* This file contains the respective InnerIterator definition of the expressions defined in Eigen/Core
  */
 
-/** \ingroup SparseCore_Module
-  * \class InnerIterator
-  * \brief An InnerIterator allows to loop over the element of a sparse (or dense) matrix or expression
-  *
-  * todo
+namespace internal {
+
+template<typename XprType, typename EvaluatorKind>
+class inner_iterator_selector;
+
+}
+
+/** \class InnerIterator
+  * \brief An InnerIterator allows to loop over the element of any matrix expression.
+  * 
+  * \warning To be used with care because an evaluator is constructed every time an InnerIterator iterator is constructed.
+  * 
+  * TODO: add a usage example
   */
-
-// generic version for dense matrix and expressions
-template<typename Derived> class DenseBase<Derived>::InnerIterator
+template<typename XprType>
+class InnerIterator
 {
-  protected:
-    typedef typename Derived::Scalar Scalar;
-    typedef typename Derived::Index Index;
-
-    enum { IsRowMajor = (Derived::Flags&RowMajorBit)==RowMajorBit };
-  public:
-    EIGEN_STRONG_INLINE InnerIterator(const Derived& expr, Index outer)
-      : m_expression(expr), m_inner(0), m_outer(outer), m_end(expr.innerSize())
-    {}
-
-    EIGEN_STRONG_INLINE Scalar value() const
-    {
-      return (IsRowMajor) ? m_expression.coeff(m_outer, m_inner)
-                          : m_expression.coeff(m_inner, m_outer);
-    }
-
-    EIGEN_STRONG_INLINE InnerIterator& operator++() { m_inner++; return *this; }
-
-    EIGEN_STRONG_INLINE Index index() const { return m_inner; }
-    inline Index row() const { return IsRowMajor ? m_outer : index(); }
-    inline Index col() const { return IsRowMajor ? index() : m_outer; }
-
-    EIGEN_STRONG_INLINE operator bool() const { return m_inner < m_end && m_inner>=0; }
-
-  protected:
-    const Derived& m_expression;
-    Index m_inner;
-    const Index m_outer;
-    const Index m_end;
+protected:
+  typedef internal::inner_iterator_selector<XprType, typename internal::evaluator_traits<XprType>::Kind> IteratorType;
+  typedef internal::evaluator<XprType> EvaluatorType;
+  typedef typename internal::traits<XprType>::Scalar Scalar;
+public:
+  /** Construct an iterator over the \a outerId -th row or column of \a xpr */
+  InnerIterator(const XprType &xpr, const Index &outerId)
+    : m_eval(xpr), m_iter(m_eval, outerId, xpr.innerSize())
+  {}
+  
+  /// \returns the value of the current coefficient.
+  EIGEN_STRONG_INLINE Scalar value() const          { return m_iter.value(); }
+  /** Increment the iterator \c *this to the next non-zero coefficient.
+    * Explicit zeros are not skipped over. To skip explicit zeros, see class SparseView
+    */
+  EIGEN_STRONG_INLINE InnerIterator& operator++()   { m_iter.operator++(); return *this; }
+  /// \returns the column or row index of the current coefficient.
+  EIGEN_STRONG_INLINE Index index() const           { return m_iter.index(); }
+  /// \returns the row index of the current coefficient.
+  EIGEN_STRONG_INLINE Index row() const             { return m_iter.row(); }
+  /// \returns the column index of the current coefficient.
+  EIGEN_STRONG_INLINE Index col() const             { return m_iter.col(); }
+  /// \returns \c true if the iterator \c *this still references a valid coefficient.
+  EIGEN_STRONG_INLINE operator bool() const         { return m_iter; }
+  
+protected:
+  EvaluatorType m_eval;
+  IteratorType m_iter;
+private:
+  // If you get here, then you're not using the right InnerIterator type, e.g.:
+  //   SparseMatrix<double,RowMajor> A;
+  //   SparseMatrix<double>::InnerIterator it(A,0);
+  template<typename T> InnerIterator(const EigenBase<T>&,Index outer);
 };
 
+namespace internal {
+
+// Generic inner iterator implementation for dense objects
+template<typename XprType>
+class inner_iterator_selector<XprType, IndexBased>
+{
+protected:
+  typedef evaluator<XprType> EvaluatorType;
+  typedef typename traits<XprType>::Scalar Scalar;
+  enum { IsRowMajor = (XprType::Flags&RowMajorBit)==RowMajorBit };
+  
+public:
+  EIGEN_STRONG_INLINE inner_iterator_selector(const EvaluatorType &eval, const Index &outerId, const Index &innerSize)
+    : m_eval(eval), m_inner(0), m_outer(outerId), m_end(innerSize)
+  {}
+
+  EIGEN_STRONG_INLINE Scalar value() const
+  {
+    return (IsRowMajor) ? m_eval.coeff(m_outer, m_inner)
+                        : m_eval.coeff(m_inner, m_outer);
+  }
+
+  EIGEN_STRONG_INLINE inner_iterator_selector& operator++() { m_inner++; return *this; }
+
+  EIGEN_STRONG_INLINE Index index() const { return m_inner; }
+  inline Index row() const { return IsRowMajor ? m_outer : index(); }
+  inline Index col() const { return IsRowMajor ? index() : m_outer; }
+
+  EIGEN_STRONG_INLINE operator bool() const { return m_inner < m_end && m_inner>=0; }
+
+protected:
+  const EvaluatorType& m_eval;
+  Index m_inner;
+  const Index m_outer;
+  const Index m_end;
+};
+
+// For iterator-based evaluator, inner-iterator is already implemented as
+// evaluator<>::InnerIterator
+template<typename XprType>
+class inner_iterator_selector<XprType, IteratorBased>
+ : public evaluator<XprType>::InnerIterator
+{
+protected:
+  typedef typename evaluator<XprType>::InnerIterator Base;
+  typedef evaluator<XprType> EvaluatorType;
+  
+public:
+  EIGEN_STRONG_INLINE inner_iterator_selector(const EvaluatorType &eval, const Index &outerId, const Index &/*innerSize*/)
+    : Base(eval, outerId)
+  {}  
+};
+
+} // end namespace internal
+
 } // end namespace Eigen
 
 #endif // EIGEN_COREITERATORS_H

Eigen/src/Core/CwiseBinaryOp.h

@@ -1,7 +1,7 @@
 // This file is part of Eigen, a lightweight C++ template library
 // for linear algebra.
 //
-// Copyright (C) 2008-2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2008-2014 Gael Guennebaud <gael.guennebaud@inria.fr>
 // Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
 //
 // This Source Code Form is subject to the terms of the Mozilla
@@ -56,72 +56,51 @@ struct traits<CwiseBinaryOp<BinaryOp, Lhs, Rhs> >
                        typename Rhs::Scalar
                      )
                    >::type Scalar;
-  typedef typename promote_storage_type<typename traits<Lhs>::StorageKind,
-                                           typename traits<Rhs>::StorageKind>::ret StorageKind;
-  typedef typename promote_index_type<typename traits<Lhs>::Index,
-                                         typename traits<Rhs>::Index>::type Index;
+  typedef typename cwise_promote_storage_type<typename traits<Lhs>::StorageKind,
+                                              typename traits<Rhs>::StorageKind,
+                                              BinaryOp>::ret StorageKind;
+  typedef typename promote_index_type<typename traits<Lhs>::StorageIndex,
+                                      typename traits<Rhs>::StorageIndex>::type StorageIndex;
   typedef typename Lhs::Nested LhsNested;
   typedef typename Rhs::Nested RhsNested;
   typedef typename remove_reference<LhsNested>::type _LhsNested;
   typedef typename remove_reference<RhsNested>::type _RhsNested;
   enum {
-    LhsCoeffReadCost = _LhsNested::CoeffReadCost,
-    RhsCoeffReadCost = _RhsNested::CoeffReadCost,
-    LhsFlags = _LhsNested::Flags,
-    RhsFlags = _RhsNested::Flags,
-    SameType = is_same<typename _LhsNested::Scalar,typename _RhsNested::Scalar>::value,
-    StorageOrdersAgree = (int(Lhs::Flags)&RowMajorBit)==(int(Rhs::Flags)&RowMajorBit),
-    Flags0 = (int(LhsFlags) | int(RhsFlags)) & (
-        HereditaryBits
-      | (int(LhsFlags) & int(RhsFlags) &
-           ( AlignedBit
-           | (StorageOrdersAgree ? LinearAccessBit : 0)
-           | (functor_traits<BinaryOp>::PacketAccess && StorageOrdersAgree && SameType ? PacketAccessBit : 0)
-           )
-        )
-     ),
-    Flags = (Flags0 & ~RowMajorBit) | (LhsFlags & RowMajorBit),
-    CoeffReadCost = LhsCoeffReadCost + RhsCoeffReadCost + functor_traits<BinaryOp>::Cost
+    Flags = _LhsNested::Flags & RowMajorBit
   };
 };
 } // end namespace internal
 
-// we require Lhs and Rhs to have the same scalar type. Currently there is no example of a binary functor
-// that would take two operands of different types. If there were such an example, then this check should be
-// moved to the BinaryOp functors, on a per-case basis. This would however require a change in the BinaryOp functors, as
-// currently they take only one typename Scalar template parameter.
-// It is tempting to always allow mixing different types but remember that this is often impossible in the vectorized paths.
-// So allowing mixing different types gives very unexpected errors when enabling vectorization, when the user tries to
-// add together a float matrix and a double matrix.
-#define EIGEN_CHECK_BINARY_COMPATIBILIY(BINOP,LHS,RHS) \
-  EIGEN_STATIC_ASSERT((internal::functor_is_product_like<BINOP>::ret \
-                        ? int(internal::scalar_product_traits<LHS, RHS>::Defined) \
-                        : int(internal::is_same<LHS, RHS>::value)), \
-    YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY)
-
 template<typename BinaryOp, typename Lhs, typename Rhs, typename StorageKind>
 class CwiseBinaryOpImpl;
 
-template<typename BinaryOp, typename Lhs, typename Rhs>
-class CwiseBinaryOp : internal::no_assignment_operator,
+template<typename BinaryOp, typename LhsType, typename RhsType>
+class CwiseBinaryOp : 
   public CwiseBinaryOpImpl<
-          BinaryOp, Lhs, Rhs,
-          typename internal::promote_storage_type<typename internal::traits<Lhs>::StorageKind,
-                                           typename internal::traits<Rhs>::StorageKind>::ret>
+          BinaryOp, LhsType, RhsType,
+          typename internal::cwise_promote_storage_type<typename internal::traits<LhsType>::StorageKind,
+                                                        typename internal::traits<RhsType>::StorageKind,
+                                                        BinaryOp>::ret>,
+  internal::no_assignment_operator
 {
   public:
+    
+    typedef typename internal::remove_all<LhsType>::type Lhs;
+    typedef typename internal::remove_all<RhsType>::type Rhs;
 
     typedef typename CwiseBinaryOpImpl<
-        BinaryOp, Lhs, Rhs,
-        typename internal::promote_storage_type<typename internal::traits<Lhs>::StorageKind,
-                                         typename internal::traits<Rhs>::StorageKind>::ret>::Base Base;
+        BinaryOp, LhsType, RhsType,
+        typename internal::cwise_promote_storage_type<typename internal::traits<LhsType>::StorageKind,
+                                                      typename internal::traits<Rhs>::StorageKind,
+                                                      BinaryOp>::ret>::Base Base;
     EIGEN_GENERIC_PUBLIC_INTERFACE(CwiseBinaryOp)
 
-    typedef typename internal::nested<Lhs>::type LhsNested;
-    typedef typename internal::nested<Rhs>::type RhsNested;
+    typedef typename internal::ref_selector<LhsType>::type LhsNested;
+    typedef typename internal::ref_selector<RhsType>::type RhsNested;
     typedef typename internal::remove_reference<LhsNested>::type _LhsNested;
     typedef typename internal::remove_reference<RhsNested>::type _RhsNested;
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE CwiseBinaryOp(const Lhs& aLhs, const Rhs& aRhs, const BinaryOp& func = BinaryOp())
       : m_lhs(aLhs), m_rhs(aRhs), m_functor(func)
     {
@@ -131,6 +110,7 @@ class CwiseBinaryOp : internal::no_assignment_operator,
       eigen_assert(aLhs.rows() == aRhs.rows() && aLhs.cols() == aRhs.cols());
     }
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Index rows() const {
       // return the fixed size type if available to enable compile time optimizations
       if (internal::traits<typename internal::remove_all<LhsNested>::type>::RowsAtCompileTime==Dynamic)
@@ -138,6 +118,7 @@ class CwiseBinaryOp : internal::no_assignment_operator,
       else
         return m_lhs.rows();
     }
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Index cols() const {
       // return the fixed size type if available to enable compile time optimizations
       if (internal::traits<typename internal::remove_all<LhsNested>::type>::ColsAtCompileTime==Dynamic)
@@ -147,10 +128,13 @@ class CwiseBinaryOp : internal::no_assignment_operator,
     }
 
     /** \returns the left hand side nested expression */
+    EIGEN_DEVICE_FUNC
     const _LhsNested& lhs() const { return m_lhs; }
     /** \returns the right hand side nested expression */
+    EIGEN_DEVICE_FUNC
     const _RhsNested& rhs() const { return m_rhs; }
     /** \returns the functor representing the binary operation */
+    EIGEN_DEVICE_FUNC
     const BinaryOp& functor() const { return m_functor; }
 
   protected:
@@ -159,41 +143,13 @@ class CwiseBinaryOp : internal::no_assignment_operator,
     const BinaryOp m_functor;
 };
 
-template<typename BinaryOp, typename Lhs, typename Rhs>
-class CwiseBinaryOpImpl<BinaryOp, Lhs, Rhs, Dense>
-  : public internal::dense_xpr_base<CwiseBinaryOp<BinaryOp, Lhs, Rhs> >::type
+// Generic API dispatcher
+template<typename BinaryOp, typename Lhs, typename Rhs, typename StorageKind>
+class CwiseBinaryOpImpl
+  : public internal::generic_xpr_base<CwiseBinaryOp<BinaryOp, Lhs, Rhs> >::type
 {
-    typedef CwiseBinaryOp<BinaryOp, Lhs, Rhs> Derived;
-  public:
-
-    typedef typename internal::dense_xpr_base<CwiseBinaryOp<BinaryOp, Lhs, Rhs> >::type Base;
-    EIGEN_DENSE_PUBLIC_INTERFACE( Derived )
-
-    EIGEN_STRONG_INLINE const Scalar coeff(Index rowId, Index colId) const
-    {
-      return derived().functor()(derived().lhs().coeff(rowId, colId),
-                                 derived().rhs().coeff(rowId, colId));
-    }
-
-    template<int LoadMode>
-    EIGEN_STRONG_INLINE PacketScalar packet(Index rowId, Index colId) const
-    {
-      return derived().functor().packetOp(derived().lhs().template packet<LoadMode>(rowId, colId),
-                                          derived().rhs().template packet<LoadMode>(rowId, colId));
-    }
-
-    EIGEN_STRONG_INLINE const Scalar coeff(Index index) const
-    {
-      return derived().functor()(derived().lhs().coeff(index),
-                                 derived().rhs().coeff(index));
-    }
-
-    template<int LoadMode>
-    EIGEN_STRONG_INLINE PacketScalar packet(Index index) const
-    {
-      return derived().functor().packetOp(derived().lhs().template packet<LoadMode>(index),
-                                          derived().rhs().template packet<LoadMode>(index));
-    }
+public:
+  typedef typename internal::generic_xpr_base<CwiseBinaryOp<BinaryOp, Lhs, Rhs> >::type Base;
 };
 
 /** replaces \c *this by \c *this - \a other.
@@ -205,8 +161,7 @@ template<typename OtherDerived>
 EIGEN_STRONG_INLINE Derived &
 MatrixBase<Derived>::operator-=(const MatrixBase<OtherDerived> &other)
 {
-  SelfCwiseBinaryOp<internal::scalar_difference_op<Scalar>, Derived, OtherDerived> tmp(derived());
-  tmp = other.derived();
+  call_assignment(derived(), other.derived(), internal::sub_assign_op<Scalar>());
   return derived();
 }
 
@@ -219,11 +174,11 @@ template<typename OtherDerived>
 EIGEN_STRONG_INLINE Derived &
 MatrixBase<Derived>::operator+=(const MatrixBase<OtherDerived>& other)
 {
-  SelfCwiseBinaryOp<internal::scalar_sum_op<Scalar>, Derived, OtherDerived> tmp(derived());
-  tmp = other.derived();
+  call_assignment(derived(), other.derived(), internal::add_assign_op<Scalar>());
   return derived();
 }
 
 } // end namespace Eigen
 
 #endif // EIGEN_CWISE_BINARY_OP_H
+

Eigen/src/Core/CwiseNullaryOp.h

@@ -35,37 +35,35 @@ template<typename NullaryOp, typename PlainObjectType>
 struct traits<CwiseNullaryOp<NullaryOp, PlainObjectType> > : traits<PlainObjectType>
 {
   enum {
-    Flags = (traits<PlainObjectType>::Flags
-      & (  HereditaryBits
-         | (functor_has_linear_access<NullaryOp>::ret ? LinearAccessBit : 0)
-         | (functor_traits<NullaryOp>::PacketAccess ? PacketAccessBit : 0)))
-      | (functor_traits<NullaryOp>::IsRepeatable ? 0 : EvalBeforeNestingBit),
-    CoeffReadCost = functor_traits<NullaryOp>::Cost
+    Flags = traits<PlainObjectType>::Flags & RowMajorBit
   };
 };
 }
 
 template<typename NullaryOp, typename PlainObjectType>
-class CwiseNullaryOp : internal::no_assignment_operator,
-  public internal::dense_xpr_base< CwiseNullaryOp<NullaryOp, PlainObjectType> >::type
+class CwiseNullaryOp : public internal::dense_xpr_base< CwiseNullaryOp<NullaryOp, PlainObjectType> >::type, internal::no_assignment_operator
 {
   public:
 
     typedef typename internal::dense_xpr_base<CwiseNullaryOp>::type Base;
     EIGEN_DENSE_PUBLIC_INTERFACE(CwiseNullaryOp)
 
-    CwiseNullaryOp(Index nbRows, Index nbCols, const NullaryOp& func = NullaryOp())
-      : m_rows(nbRows), m_cols(nbCols), m_functor(func)
+    EIGEN_DEVICE_FUNC
+    CwiseNullaryOp(Index rows, Index cols, const NullaryOp& func = NullaryOp())
+      : m_rows(rows), m_cols(cols), m_functor(func)
     {
-      eigen_assert(nbRows >= 0
-            && (RowsAtCompileTime == Dynamic || RowsAtCompileTime == nbRows)
-            &&  nbCols >= 0
-            && (ColsAtCompileTime == Dynamic || ColsAtCompileTime == nbCols));
+      eigen_assert(rows >= 0
+            && (RowsAtCompileTime == Dynamic || RowsAtCompileTime == rows)
+            &&  cols >= 0
+            && (ColsAtCompileTime == Dynamic || ColsAtCompileTime == cols));
     }
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Index rows() const { return m_rows.value(); }
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Index cols() const { return m_cols.value(); }
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE const Scalar coeff(Index rowId, Index colId) const
     {
       return m_functor(rowId, colId);
@@ -77,6 +75,7 @@ class CwiseNullaryOp : internal::no_assignment_operator,
       return m_functor.packetOp(rowId, colId);
     }
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE const Scalar coeff(Index index) const
     {
       return m_functor(index);
@@ -89,6 +88,7 @@ class CwiseNullaryOp : internal::no_assignment_operator,
     }
 
     /** \returns the functor representing the nullary operation */
+    EIGEN_DEVICE_FUNC
     const NullaryOp& functor() const { return m_functor; }
 
   protected:
@@ -113,10 +113,10 @@ class CwiseNullaryOp : internal::no_assignment_operator,
   */
 template<typename Derived>
 template<typename CustomNullaryOp>
-EIGEN_STRONG_INLINE const CwiseNullaryOp<CustomNullaryOp, Derived>
+EIGEN_STRONG_INLINE const CwiseNullaryOp<CustomNullaryOp, typename DenseBase<Derived>::PlainObject>
 DenseBase<Derived>::NullaryExpr(Index rows, Index cols, const CustomNullaryOp& func)
 {
-  return CwiseNullaryOp<CustomNullaryOp, Derived>(rows, cols, func);
+  return CwiseNullaryOp<CustomNullaryOp, PlainObject>(rows, cols, func);
 }
 
 /** \returns an expression of a matrix defined by a custom functor \a func
@@ -132,16 +132,19 @@ DenseBase<Derived>::NullaryExpr(Index rows, Index cols, const CustomNullaryOp& f
   *
   * The template parameter \a CustomNullaryOp is the type of the functor.
   *
+  * Here is an example with C++11 random generators: \include random_cpp11.cpp
+  * Output: \verbinclude random_cpp11.out
+  * 
   * \sa class CwiseNullaryOp
   */
 template<typename Derived>
 template<typename CustomNullaryOp>
-EIGEN_STRONG_INLINE const CwiseNullaryOp<CustomNullaryOp, Derived>
+EIGEN_STRONG_INLINE const CwiseNullaryOp<CustomNullaryOp, typename DenseBase<Derived>::PlainObject>
 DenseBase<Derived>::NullaryExpr(Index size, const CustomNullaryOp& func)
 {
   EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
-  if(RowsAtCompileTime == 1) return CwiseNullaryOp<CustomNullaryOp, Derived>(1, size, func);
-  else return CwiseNullaryOp<CustomNullaryOp, Derived>(size, 1, func);
+  if(RowsAtCompileTime == 1) return CwiseNullaryOp<CustomNullaryOp, PlainObject>(1, size, func);
+  else return CwiseNullaryOp<CustomNullaryOp, PlainObject>(size, 1, func);
 }
 
 /** \returns an expression of a matrix defined by a custom functor \a func
@@ -155,19 +158,19 @@ DenseBase<Derived>::NullaryExpr(Index size, const CustomNullaryOp& func)
   */
 template<typename Derived>
 template<typename CustomNullaryOp>
-EIGEN_STRONG_INLINE const CwiseNullaryOp<CustomNullaryOp, Derived>
+EIGEN_STRONG_INLINE const CwiseNullaryOp<CustomNullaryOp, typename DenseBase<Derived>::PlainObject>
 DenseBase<Derived>::NullaryExpr(const CustomNullaryOp& func)
 {
-  return CwiseNullaryOp<CustomNullaryOp, Derived>(RowsAtCompileTime, ColsAtCompileTime, func);
+  return CwiseNullaryOp<CustomNullaryOp, PlainObject>(RowsAtCompileTime, ColsAtCompileTime, func);
 }
 
 /** \returns an expression of a constant matrix of value \a value
   *
-  * The parameters \a nbRows and \a nbCols are the number of rows and of columns of
+  * The parameters \a rows and \a cols are the number of rows and of columns of
   * the returned matrix. Must be compatible with this DenseBase type.
   *
   * This variant is meant to be used for dynamic-size matrix types. For fixed-size types,
-  * it is redundant to pass \a nbRows and \a nbCols as arguments, so Zero() should be used
+  * it is redundant to pass \a rows and \a cols as arguments, so Zero() should be used
   * instead.
   *
   * The template parameter \a CustomNullaryOp is the type of the functor.
@@ -176,9 +179,9 @@ DenseBase<Derived>::NullaryExpr(const CustomNullaryOp& func)
   */
 template<typename Derived>
 EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
-DenseBase<Derived>::Constant(Index nbRows, Index nbCols, const Scalar& value)
+DenseBase<Derived>::Constant(Index rows, Index cols, const Scalar& value)
 {
-  return DenseBase<Derived>::NullaryExpr(nbRows, nbCols, internal::scalar_constant_op<Scalar>(value));
+  return DenseBase<Derived>::NullaryExpr(rows, cols, internal::scalar_constant_op<Scalar>(value));
 }
 
 /** \returns an expression of a constant matrix of value \a value
@@ -242,7 +245,7 @@ EIGEN_STRONG_INLINE const typename DenseBase<Derived>::SequentialLinSpacedReturn
 DenseBase<Derived>::LinSpaced(Sequential_t, Index size, const Scalar& low, const Scalar& high)
 {
   EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
-  return DenseBase<Derived>::NullaryExpr(size, internal::linspaced_op<Scalar,false>(low,high,size));
+  return DenseBase<Derived>::NullaryExpr(size, internal::linspaced_op<Scalar,PacketScalar,false>(low,high,size));
 }
 
 /**
@@ -255,7 +258,7 @@ DenseBase<Derived>::LinSpaced(Sequential_t, const Scalar& low, const Scalar& hig
 {
   EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
   EIGEN_STATIC_ASSERT_FIXED_SIZE(Derived)
-  return DenseBase<Derived>::NullaryExpr(Derived::SizeAtCompileTime, internal::linspaced_op<Scalar,false>(low,high,Derived::SizeAtCompileTime));
+  return DenseBase<Derived>::NullaryExpr(Derived::SizeAtCompileTime, internal::linspaced_op<Scalar,PacketScalar,false>(low,high,Derived::SizeAtCompileTime));
 }
 
 /**
@@ -276,7 +279,7 @@ EIGEN_STRONG_INLINE const typename DenseBase<Derived>::RandomAccessLinSpacedRetu
 DenseBase<Derived>::LinSpaced(Index size, const Scalar& low, const Scalar& high)
 {
   EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
-  return DenseBase<Derived>::NullaryExpr(size, internal::linspaced_op<Scalar,true>(low,high,size));
+  return DenseBase<Derived>::NullaryExpr(size, internal::linspaced_op<Scalar,PacketScalar,true>(low,high,size));
 }
 
 /**
@@ -289,7 +292,7 @@ DenseBase<Derived>::LinSpaced(const Scalar& low, const Scalar& high)
 {
   EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
   EIGEN_STATIC_ASSERT_FIXED_SIZE(Derived)
-  return DenseBase<Derived>::NullaryExpr(Derived::SizeAtCompileTime, internal::linspaced_op<Scalar,true>(low,high,Derived::SizeAtCompileTime));
+  return DenseBase<Derived>::NullaryExpr(Derived::SizeAtCompileTime, internal::linspaced_op<Scalar,PacketScalar,true>(low,high,Derived::SizeAtCompileTime));
 }
 
 /** \returns true if all coefficients in this matrix are approximately equal to \a val, to within precision \a prec */
@@ -297,9 +300,10 @@ template<typename Derived>
 bool DenseBase<Derived>::isApproxToConstant
 (const Scalar& val, const RealScalar& prec) const
 {
+  typename internal::nested_eval<Derived,1>::type self(derived());
   for(Index j = 0; j < cols(); ++j)
     for(Index i = 0; i < rows(); ++i)
-      if(!internal::isApprox(this->coeff(i, j), val, prec))
+      if(!internal::isApprox(self.coeff(i, j), val, prec))
         return false;
   return true;
 }
@@ -353,8 +357,8 @@ PlainObjectBase<Derived>::setConstant(Index size, const Scalar& val)
 
 /** Resizes to the given size, and sets all coefficients in this expression to the given \a value.
   *
-  * \param nbRows the new number of rows
-  * \param nbCols the new number of columns
+  * \param rows the new number of rows
+  * \param cols the new number of columns
   * \param val the value to which all coefficients are set
   *
   * Example: \include Matrix_setConstant_int_int.cpp
@@ -364,9 +368,9 @@ PlainObjectBase<Derived>::setConstant(Index size, const Scalar& val)
   */
 template<typename Derived>
 EIGEN_STRONG_INLINE Derived&
-PlainObjectBase<Derived>::setConstant(Index nbRows, Index nbCols, const Scalar& val)
+PlainObjectBase<Derived>::setConstant(Index rows, Index cols, const Scalar& val)
 {
-  resize(nbRows, nbCols);
+  resize(rows, cols);
   return setConstant(val);
 }
 
@@ -387,7 +391,7 @@ template<typename Derived>
 EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::setLinSpaced(Index newSize, const Scalar& low, const Scalar& high)
 {
   EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
-  return derived() = Derived::NullaryExpr(newSize, internal::linspaced_op<Scalar,false>(low,high,newSize));
+  return derived() = Derived::NullaryExpr(newSize, internal::linspaced_op<Scalar,PacketScalar,false>(low,high,newSize));
 }
 
 /**
@@ -425,9 +429,9 @@ EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::setLinSpaced(const Scalar& low,
   */
 template<typename Derived>
 EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
-DenseBase<Derived>::Zero(Index nbRows, Index nbCols)
+DenseBase<Derived>::Zero(Index rows, Index cols)
 {
-  return Constant(nbRows, nbCols, Scalar(0));
+  return Constant(rows, cols, Scalar(0));
 }
 
 /** \returns an expression of a zero vector.
@@ -481,9 +485,10 @@ DenseBase<Derived>::Zero()
 template<typename Derived>
 bool DenseBase<Derived>::isZero(const RealScalar& prec) const
 {
+  typename internal::nested_eval<Derived,1>::type self(derived());
   for(Index j = 0; j < cols(); ++j)
     for(Index i = 0; i < rows(); ++i)
-      if(!internal::isMuchSmallerThan(this->coeff(i, j), static_cast<Scalar>(1), prec))
+      if(!internal::isMuchSmallerThan(self.coeff(i, j), static_cast<Scalar>(1), prec))
         return false;
   return true;
 }
@@ -520,8 +525,8 @@ PlainObjectBase<Derived>::setZero(Index newSize)
 
 /** Resizes to the given size, and sets all coefficients in this expression to zero.
   *
-  * \param nbRows the new number of rows
-  * \param nbCols the new number of columns
+  * \param rows the new number of rows
+  * \param cols the new number of columns
   *
   * Example: \include Matrix_setZero_int_int.cpp
   * Output: \verbinclude Matrix_setZero_int_int.out
@@ -530,9 +535,9 @@ PlainObjectBase<Derived>::setZero(Index newSize)
   */
 template<typename Derived>
 EIGEN_STRONG_INLINE Derived&
-PlainObjectBase<Derived>::setZero(Index nbRows, Index nbCols)
+PlainObjectBase<Derived>::setZero(Index rows, Index cols)
 {
-  resize(nbRows, nbCols);
+  resize(rows, cols);
   return setConstant(Scalar(0));
 }
 
@@ -540,7 +545,7 @@ PlainObjectBase<Derived>::setZero(Index nbRows, Index nbCols)
 
 /** \returns an expression of a matrix where all coefficients equal one.
   *
-  * The parameters \a nbRows and \a nbCols are the number of rows and of columns of
+  * The parameters \a rows and \a cols are the number of rows and of columns of
   * the returned matrix. Must be compatible with this MatrixBase type.
   *
   * This variant is meant to be used for dynamic-size matrix types. For fixed-size types,
@@ -554,9 +559,9 @@ PlainObjectBase<Derived>::setZero(Index nbRows, Index nbCols)
   */
 template<typename Derived>
 EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
-DenseBase<Derived>::Ones(Index nbRows, Index nbCols)
+DenseBase<Derived>::Ones(Index rows, Index cols)
 {
-  return Constant(nbRows, nbCols, Scalar(1));
+  return Constant(rows, cols, Scalar(1));
 }
 
 /** \returns an expression of a vector where all coefficients equal one.
@@ -646,8 +651,8 @@ PlainObjectBase<Derived>::setOnes(Index newSize)
 
 /** Resizes to the given size, and sets all coefficients in this expression to one.
   *
-  * \param nbRows the new number of rows
-  * \param nbCols the new number of columns
+  * \param rows the new number of rows
+  * \param cols the new number of columns
   *
   * Example: \include Matrix_setOnes_int_int.cpp
   * Output: \verbinclude Matrix_setOnes_int_int.out
@@ -656,9 +661,9 @@ PlainObjectBase<Derived>::setOnes(Index newSize)
   */
 template<typename Derived>
 EIGEN_STRONG_INLINE Derived&
-PlainObjectBase<Derived>::setOnes(Index nbRows, Index nbCols)
+PlainObjectBase<Derived>::setOnes(Index rows, Index cols)
 {
-  resize(nbRows, nbCols);
+  resize(rows, cols);
   return setConstant(Scalar(1));
 }
 
@@ -666,7 +671,7 @@ PlainObjectBase<Derived>::setOnes(Index nbRows, Index nbCols)
 
 /** \returns an expression of the identity matrix (not necessarily square).
   *
-  * The parameters \a nbRows and \a nbCols are the number of rows and of columns of
+  * The parameters \a rows and \a cols are the number of rows and of columns of
   * the returned matrix. Must be compatible with this MatrixBase type.
   *
   * This variant is meant to be used for dynamic-size matrix types. For fixed-size types,
@@ -680,9 +685,9 @@ PlainObjectBase<Derived>::setOnes(Index nbRows, Index nbCols)
   */
 template<typename Derived>
 EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::IdentityReturnType
-MatrixBase<Derived>::Identity(Index nbRows, Index nbCols)
+MatrixBase<Derived>::Identity(Index rows, Index cols)
 {
-  return DenseBase<Derived>::NullaryExpr(nbRows, nbCols, internal::scalar_identity_op<Scalar>());
+  return DenseBase<Derived>::NullaryExpr(rows, cols, internal::scalar_identity_op<Scalar>());
 }
 
 /** \returns an expression of the identity matrix (not necessarily square).
@@ -716,18 +721,19 @@ template<typename Derived>
 bool MatrixBase<Derived>::isIdentity
 (const RealScalar& prec) const
 {
+  typename internal::nested_eval<Derived,1>::type self(derived());
   for(Index j = 0; j < cols(); ++j)
   {
     for(Index i = 0; i < rows(); ++i)
     {
       if(i == j)
       {
-        if(!internal::isApprox(this->coeff(i, j), static_cast<Scalar>(1), prec))
+        if(!internal::isApprox(self.coeff(i, j), static_cast<Scalar>(1), prec))
           return false;
       }
       else
       {
-        if(!internal::isMuchSmallerThan(this->coeff(i, j), static_cast<RealScalar>(1), prec))
+        if(!internal::isMuchSmallerThan(self.coeff(i, j), static_cast<RealScalar>(1), prec))
           return false;
       }
     }
@@ -740,6 +746,7 @@ namespace internal {
 template<typename Derived, bool Big = (Derived::SizeAtCompileTime>=16)>
 struct setIdentity_impl
 {
+  EIGEN_DEVICE_FUNC
   static EIGEN_STRONG_INLINE Derived& run(Derived& m)
   {
     return m = Derived::Identity(m.rows(), m.cols());
@@ -749,7 +756,7 @@ struct setIdentity_impl
 template<typename Derived>
 struct setIdentity_impl<Derived, true>
 {
-  typedef typename Derived::Index Index;
+  EIGEN_DEVICE_FUNC
   static EIGEN_STRONG_INLINE Derived& run(Derived& m)
   {
     m.setZero();
@@ -776,8 +783,8 @@ EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::setIdentity()
 
 /** \brief Resizes to the given size, and writes the identity expression (not necessarily square) into *this.
   *
-  * \param nbRows the new number of rows
-  * \param nbCols the new number of columns
+  * \param rows the new number of rows
+  * \param cols the new number of columns
   *
   * Example: \include Matrix_setIdentity_int_int.cpp
   * Output: \verbinclude Matrix_setIdentity_int_int.out
@@ -785,9 +792,9 @@ EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::setIdentity()
   * \sa MatrixBase::setIdentity(), class CwiseNullaryOp, MatrixBase::Identity()
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::setIdentity(Index nbRows, Index nbCols)
+EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::setIdentity(Index rows, Index cols)
 {
-  derived().resize(nbRows, nbCols);
+  derived().resize(rows, cols);
   return setIdentity();
 }
 

Eigen/src/Core/CwiseUnaryOp.h

@@ -1,7 +1,7 @@
 // This file is part of Eigen, a lightweight C++ template library
 // for linear algebra.
 //
-// Copyright (C) 2008-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2008-2014 Gael Guennebaud <gael.guennebaud@inria.fr>
 // Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
 //
 // This Source Code Form is subject to the terms of the Mozilla
@@ -44,10 +44,7 @@ struct traits<CwiseUnaryOp<UnaryOp, XprType> >
   typedef typename XprType::Nested XprTypeNested;
   typedef typename remove_reference<XprTypeNested>::type _XprTypeNested;
   enum {
-    Flags = _XprTypeNested::Flags & (
-      HereditaryBits | LinearAccessBit | AlignedBit
-      | (functor_traits<UnaryOp>::PacketAccess ? PacketAccessBit : 0)),
-    CoeffReadCost = _XprTypeNested::CoeffReadCost + functor_traits<UnaryOp>::Cost
+    Flags = _XprTypeNested::Flags & RowMajorBit 
   };
 };
 }
@@ -56,28 +53,34 @@ template<typename UnaryOp, typename XprType, typename StorageKind>
 class CwiseUnaryOpImpl;
 
 template<typename UnaryOp, typename XprType>
-class CwiseUnaryOp : internal::no_assignment_operator,
-  public CwiseUnaryOpImpl<UnaryOp, XprType, typename internal::traits<XprType>::StorageKind>
+class CwiseUnaryOp : public CwiseUnaryOpImpl<UnaryOp, XprType, typename internal::traits<XprType>::StorageKind>, internal::no_assignment_operator
 {
   public:
 
     typedef typename CwiseUnaryOpImpl<UnaryOp, XprType,typename internal::traits<XprType>::StorageKind>::Base Base;
     EIGEN_GENERIC_PUBLIC_INTERFACE(CwiseUnaryOp)
+    typedef typename internal::remove_all<XprType>::type NestedExpression;
 
-    inline CwiseUnaryOp(const XprType& xpr, const UnaryOp& func = UnaryOp())
+    EIGEN_DEVICE_FUNC
+    explicit inline CwiseUnaryOp(const XprType& xpr, const UnaryOp& func = UnaryOp())
       : m_xpr(xpr), m_functor(func) {}
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Index rows() const { return m_xpr.rows(); }
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Index cols() const { return m_xpr.cols(); }
 
     /** \returns the functor representing the unary operation */
+    EIGEN_DEVICE_FUNC
     const UnaryOp& functor() const { return m_functor; }
 
     /** \returns the nested expression */
+    EIGEN_DEVICE_FUNC
     const typename internal::remove_all<typename XprType::Nested>::type&
     nestedExpression() const { return m_xpr; }
 
     /** \returns the nested expression */
+    EIGEN_DEVICE_FUNC
     typename internal::remove_all<typename XprType::Nested>::type&
     nestedExpression() { return m_xpr.const_cast_derived(); }
 
@@ -86,39 +89,13 @@ class CwiseUnaryOp : internal::no_assignment_operator,
     const UnaryOp m_functor;
 };
 
-// This is the generic implementation for dense storage.
-// It can be used for any expression types implementing the dense concept.
-template<typename UnaryOp, typename XprType>
-class CwiseUnaryOpImpl<UnaryOp,XprType,Dense>
-  : public internal::dense_xpr_base<CwiseUnaryOp<UnaryOp, XprType> >::type
+// Generic API dispatcher
+template<typename UnaryOp, typename XprType, typename StorageKind>
+class CwiseUnaryOpImpl
+  : public internal::generic_xpr_base<CwiseUnaryOp<UnaryOp, XprType> >::type
 {
-  public:
-
-    typedef CwiseUnaryOp<UnaryOp, XprType> Derived;
-    typedef typename internal::dense_xpr_base<CwiseUnaryOp<UnaryOp, XprType> >::type Base;
-    EIGEN_DENSE_PUBLIC_INTERFACE(Derived)
-
-    EIGEN_STRONG_INLINE const Scalar coeff(Index rowId, Index colId) const
-    {
-      return derived().functor()(derived().nestedExpression().coeff(rowId, colId));
-    }
-
-    template<int LoadMode>
-    EIGEN_STRONG_INLINE PacketScalar packet(Index rowId, Index colId) const
-    {
-      return derived().functor().packetOp(derived().nestedExpression().template packet<LoadMode>(rowId, colId));
-    }
-
-    EIGEN_STRONG_INLINE const Scalar coeff(Index index) const
-    {
-      return derived().functor()(derived().nestedExpression().coeff(index));
-    }
-
-    template<int LoadMode>
-    EIGEN_STRONG_INLINE PacketScalar packet(Index index) const
-    {
-      return derived().functor().packetOp(derived().nestedExpression().template packet<LoadMode>(index));
-    }
+public:
+  typedef typename internal::generic_xpr_base<CwiseUnaryOp<UnaryOp, XprType> >::type Base;
 };
 
 } // end namespace Eigen

Eigen/src/Core/CwiseUnaryView.h

@@ -37,8 +37,8 @@ struct traits<CwiseUnaryView<ViewOp, MatrixType> >
   typedef typename MatrixType::Nested MatrixTypeNested;
   typedef typename remove_all<MatrixTypeNested>::type _MatrixTypeNested;
   enum {
-    Flags = (traits<_MatrixTypeNested>::Flags & (HereditaryBits | LvalueBit | LinearAccessBit | DirectAccessBit)),
-    CoeffReadCost = traits<_MatrixTypeNested>::CoeffReadCost + functor_traits<ViewOp>::Cost,
+    FlagsLvalueBit = is_lvalue<MatrixType>::value ? LvalueBit : 0,
+    Flags = traits<_MatrixTypeNested>::Flags & (RowMajorBit | FlagsLvalueBit | DirectAccessBit), // FIXME DirectAccessBit should not be handled by expressions
     MatrixTypeInnerStride =  inner_stride_at_compile_time<MatrixType>::ret,
     // need to cast the sizeof's from size_t to int explicitly, otherwise:
     // "error: no integral type can represent all of the enumerator values
@@ -62,8 +62,9 @@ class CwiseUnaryView : public CwiseUnaryViewImpl<ViewOp, MatrixType, typename in
 
     typedef typename CwiseUnaryViewImpl<ViewOp, MatrixType,typename internal::traits<MatrixType>::StorageKind>::Base Base;
     EIGEN_GENERIC_PUBLIC_INTERFACE(CwiseUnaryView)
+    typedef typename internal::remove_all<MatrixType>::type NestedExpression;
 
-    inline CwiseUnaryView(const MatrixType& mat, const ViewOp& func = ViewOp())
+    explicit inline CwiseUnaryView(MatrixType& mat, const ViewOp& func = ViewOp())
       : m_matrix(mat), m_functor(func) {}
 
     EIGEN_INHERIT_ASSIGNMENT_OPERATORS(CwiseUnaryView)
@@ -83,11 +84,19 @@ class CwiseUnaryView : public CwiseUnaryViewImpl<ViewOp, MatrixType, typename in
     nestedExpression() { return m_matrix.const_cast_derived(); }
 
   protected:
-    // FIXME changed from MatrixType::Nested because of a weird compilation error with sun CC
-    typename internal::nested<MatrixType>::type m_matrix;
+    typename internal::ref_selector<MatrixType>::type m_matrix;
     ViewOp m_functor;
 };
 
+// Generic API dispatcher
+template<typename ViewOp, typename XprType, typename StorageKind>
+class CwiseUnaryViewImpl
+  : public internal::generic_xpr_base<CwiseUnaryView<ViewOp, XprType> >::type
+{
+public:
+  typedef typename internal::generic_xpr_base<CwiseUnaryView<ViewOp, XprType> >::type Base;
+};
+
 template<typename ViewOp, typename MatrixType>
 class CwiseUnaryViewImpl<ViewOp,MatrixType,Dense>
   : public internal::dense_xpr_base< CwiseUnaryView<ViewOp, MatrixType> >::type
@@ -100,38 +109,18 @@ class CwiseUnaryViewImpl<ViewOp,MatrixType,Dense>
     EIGEN_DENSE_PUBLIC_INTERFACE(Derived)
     EIGEN_INHERIT_ASSIGNMENT_OPERATORS(CwiseUnaryViewImpl)
     
-    inline Scalar* data() { return &coeffRef(0); }
-    inline const Scalar* data() const { return &coeff(0); }
+    EIGEN_DEVICE_FUNC inline Scalar* data() { return &(this->coeffRef(0)); }
+    EIGEN_DEVICE_FUNC inline const Scalar* data() const { return &(this->coeff(0)); }
 
-    inline Index innerStride() const
+    EIGEN_DEVICE_FUNC inline Index innerStride() const
     {
       return derived().nestedExpression().innerStride() * sizeof(typename internal::traits<MatrixType>::Scalar) / sizeof(Scalar);
     }
 
-    inline Index outerStride() const
+    EIGEN_DEVICE_FUNC inline Index outerStride() const
     {
       return derived().nestedExpression().outerStride() * sizeof(typename internal::traits<MatrixType>::Scalar) / sizeof(Scalar);
     }
-
-    EIGEN_STRONG_INLINE CoeffReturnType coeff(Index row, Index col) const
-    {
-      return derived().functor()(derived().nestedExpression().coeff(row, col));
-    }
-
-    EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
-    {
-      return derived().functor()(derived().nestedExpression().coeff(index));
-    }
-
-    EIGEN_STRONG_INLINE Scalar& coeffRef(Index row, Index col)
-    {
-      return derived().functor()(const_cast_derived().nestedExpression().coeffRef(row, col));
-    }
-
-    EIGEN_STRONG_INLINE Scalar& coeffRef(Index index)
-    {
-      return derived().functor()(const_cast_derived().nestedExpression().coeffRef(index));
-    }
 };
 
 } // end namespace Eigen

Eigen/src/Core/DenseBase.h

@@ -49,22 +49,37 @@ template<typename Derived> class DenseBase
   public:
     using internal::special_scalar_op_base<Derived,typename internal::traits<Derived>::Scalar,
                 typename NumTraits<typename internal::traits<Derived>::Scalar>::Real>::operator*;
+    using internal::special_scalar_op_base<Derived,typename internal::traits<Derived>::Scalar,
+                typename NumTraits<typename internal::traits<Derived>::Scalar>::Real>::operator/;
 
-    class InnerIterator;
+
+    /** Inner iterator type to iterate over the coefficients of a row or column.
+      * \sa class InnerIterator
+      */
+    typedef Eigen::InnerIterator<Derived> InnerIterator;
 
     typedef typename internal::traits<Derived>::StorageKind StorageKind;
 
-    /** \brief The type of indices 
-      * \details To change this, \c \#define the preprocessor symbol \c EIGEN_DEFAULT_DENSE_INDEX_TYPE.
-      * \sa \ref TopicPreprocessorDirectives.
-      */
-    typedef typename internal::traits<Derived>::Index Index; 
+    /**
+      * \brief The type used to store indices
+      * \details This typedef is relevant for types that store multiple indices such as
+      *          PermutationMatrix or Transpositions, otherwise it defaults to Eigen::Index
+      * \sa \ref TopicPreprocessorDirectives, Eigen::Index, SparseMatrixBase.
+     */
+    typedef typename internal::traits<Derived>::StorageIndex StorageIndex;
 
+    /** The numeric type of the expression' coefficients, e.g. float, double, int or std::complex<float>, etc. */
     typedef typename internal::traits<Derived>::Scalar Scalar;
-    typedef typename internal::packet_traits<Scalar>::type PacketScalar;
+    
+    /** The numeric type of the expression' coefficients, e.g. float, double, int or std::complex<float>, etc.
+      *
+      * It is an alias for the Scalar type */
+    typedef Scalar value_type;
+    
     typedef typename NumTraits<Scalar>::Real RealScalar;
 
-    typedef DenseCoeffsBase<Derived> Base;
+    typedef internal::special_scalar_op_base<Derived,typename internal::traits<Derived>::Scalar,
+                      typename NumTraits<typename internal::traits<Derived>::Scalar>::Real> Base;
     using Base::derived;
     using Base::const_cast_derived;
     using Base::rows;
@@ -74,16 +89,6 @@ template<typename Derived> class DenseBase
     using Base::colIndexByOuterInner;
     using Base::coeff;
     using Base::coeffByOuterInner;
-    using Base::packet;
-    using Base::packetByOuterInner;
-    using Base::writePacket;
-    using Base::writePacketByOuterInner;
-    using Base::coeffRef;
-    using Base::coeffRefByOuterInner;
-    using Base::copyCoeff;
-    using Base::copyCoeffByOuterInner;
-    using Base::copyPacket;
-    using Base::copyPacketByOuterInner;
     using Base::operator();
     using Base::operator[];
     using Base::x;
@@ -169,19 +174,46 @@ template<typename Derived> class DenseBase
       InnerSizeAtCompileTime = int(IsVectorAtCompileTime) ? int(SizeAtCompileTime)
                              : int(IsRowMajor) ? int(ColsAtCompileTime) : int(RowsAtCompileTime),
 
-      CoeffReadCost = internal::traits<Derived>::CoeffReadCost,
-        /**< This is a rough measure of how expensive it is to read one coefficient from
-          * this expression.
-          */
-
       InnerStrideAtCompileTime = internal::inner_stride_at_compile_time<Derived>::ret,
       OuterStrideAtCompileTime = internal::outer_stride_at_compile_time<Derived>::ret
     };
+    
+    typedef typename internal::find_best_packet<Scalar,SizeAtCompileTime>::type PacketScalar;
 
-    enum { ThisConstantIsPrivateInPlainObjectBase };
+    enum { IsPlainObjectBase = 0 };
+    
+    /** The plain matrix type corresponding to this expression.
+      * \sa PlainObject */
+    typedef Matrix<typename internal::traits<Derived>::Scalar,
+                internal::traits<Derived>::RowsAtCompileTime,
+                internal::traits<Derived>::ColsAtCompileTime,
+                AutoAlign | (internal::traits<Derived>::Flags&RowMajorBit ? RowMajor : ColMajor),
+                internal::traits<Derived>::MaxRowsAtCompileTime,
+                internal::traits<Derived>::MaxColsAtCompileTime
+          > PlainMatrix;
+    
+    /** The plain array type corresponding to this expression.
+      * \sa PlainObject */
+    typedef Array<typename internal::traits<Derived>::Scalar,
+                internal::traits<Derived>::RowsAtCompileTime,
+                internal::traits<Derived>::ColsAtCompileTime,
+                AutoAlign | (internal::traits<Derived>::Flags&RowMajorBit ? RowMajor : ColMajor),
+                internal::traits<Derived>::MaxRowsAtCompileTime,
+                internal::traits<Derived>::MaxColsAtCompileTime
+          > PlainArray;
+
+    /** \brief The plain matrix or array type corresponding to this expression.
+      *
+      * This is not necessarily exactly the return type of eval(). In the case of plain matrices,
+      * the return type of eval() is a const reference to a matrix, not a matrix! It is however guaranteed
+      * that the return type of eval() is either PlainObject or const PlainObject&.
+      */
+    typedef typename internal::conditional<internal::is_same<typename internal::traits<Derived>::XprKind,MatrixXpr >::value,
+                                 PlainMatrix, PlainArray>::type PlainObject;
 
     /** \returns the number of nonzero coefficients which is in practice the number
       * of stored coefficients. */
+    EIGEN_DEVICE_FUNC
     inline Index nonZeros() const { return size(); }
     /** \returns true if either the number of rows or the number of columns is equal to 1.
       * In other words, this function returns
@@ -193,6 +225,7 @@ template<typename Derived> class DenseBase
       * \note For a vector, this returns just 1. For a matrix (non-vector), this is the major dimension
       * with respect to the \ref TopicStorageOrders "storage order", i.e., the number of columns for a
       * column-major matrix, and the number of rows for a row-major matrix. */
+    EIGEN_DEVICE_FUNC
     Index outerSize() const
     {
       return IsVectorAtCompileTime ? 1
@@ -204,6 +237,7 @@ template<typename Derived> class DenseBase
       * \note For a vector, this is just the size. For a matrix (non-vector), this is the minor dimension
       * with respect to the \ref TopicStorageOrders "storage order", i.e., the number of rows for a 
       * column-major matrix, and the number of columns for a row-major matrix. */
+    EIGEN_DEVICE_FUNC
     Index innerSize() const
     {
       return IsVectorAtCompileTime ? this->size()
@@ -214,6 +248,7 @@ template<typename Derived> class DenseBase
       * Matrix::resize() and Array::resize(). The present method only asserts that the new size equals the old size, and does
       * nothing else.
       */
+    EIGEN_DEVICE_FUNC
     void resize(Index newSize)
     {
       EIGEN_ONLY_USED_FOR_DEBUG(newSize);
@@ -224,22 +259,22 @@ template<typename Derived> class DenseBase
       * Matrix::resize() and Array::resize(). The present method only asserts that the new size equals the old size, and does
       * nothing else.
       */
-    void resize(Index nbRows, Index nbCols)
+    EIGEN_DEVICE_FUNC
+    void resize(Index rows, Index cols)
     {
-      EIGEN_ONLY_USED_FOR_DEBUG(nbRows);
-      EIGEN_ONLY_USED_FOR_DEBUG(nbCols);
-      eigen_assert(nbRows == this->rows() && nbCols == this->cols()
+      EIGEN_ONLY_USED_FOR_DEBUG(rows);
+      EIGEN_ONLY_USED_FOR_DEBUG(cols);
+      eigen_assert(rows == this->rows() && cols == this->cols()
                 && "DenseBase::resize() does not actually allow to resize.");
     }
 
 #ifndef EIGEN_PARSED_BY_DOXYGEN
-
     /** \internal Represents a matrix with all coefficients equal to one another*/
-    typedef CwiseNullaryOp<internal::scalar_constant_op<Scalar>,Derived> ConstantReturnType;
+    typedef CwiseNullaryOp<internal::scalar_constant_op<Scalar>,PlainObject> ConstantReturnType;
     /** \internal Represents a vector with linearly spaced coefficients that allows sequential access only. */
-    typedef CwiseNullaryOp<internal::linspaced_op<Scalar,false>,Derived> SequentialLinSpacedReturnType;
+    typedef CwiseNullaryOp<internal::linspaced_op<Scalar,PacketScalar,false>,PlainObject> SequentialLinSpacedReturnType;
     /** \internal Represents a vector with linearly spaced coefficients that allows random access. */
-    typedef CwiseNullaryOp<internal::linspaced_op<Scalar,true>,Derived> RandomAccessLinSpacedReturnType;
+    typedef CwiseNullaryOp<internal::linspaced_op<Scalar,PacketScalar,true>,PlainObject> RandomAccessLinSpacedReturnType;
     /** \internal the return type of MatrixBase::eigenvalues() */
     typedef Matrix<typename NumTraits<typename internal::traits<Derived>::Scalar>::Real, internal::traits<Derived>::ColsAtCompileTime, 1> EigenvaluesReturnType;
 
@@ -247,110 +282,122 @@ template<typename Derived> class DenseBase
 
     /** Copies \a other into *this. \returns a reference to *this. */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     Derived& operator=(const DenseBase<OtherDerived>& other);
 
     /** Special case of the template operator=, in order to prevent the compiler
       * from generating a default operator= (issue hit with g++ 4.1)
       */
+    EIGEN_DEVICE_FUNC
     Derived& operator=(const DenseBase& other);
 
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     Derived& operator=(const EigenBase<OtherDerived> &other);
 
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     Derived& operator+=(const EigenBase<OtherDerived> &other);
 
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     Derived& operator-=(const EigenBase<OtherDerived> &other);
 
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     Derived& operator=(const ReturnByValue<OtherDerived>& func);
 
-#ifndef EIGEN_PARSED_BY_DOXYGEN
-    /** Copies \a other into *this without evaluating other. \returns a reference to *this. */
+    /** \ínternal
+      * Copies \a other into *this without evaluating other. \returns a reference to *this.
+      * \deprecated */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     Derived& lazyAssign(const DenseBase<OtherDerived>& other);
-#endif // not EIGEN_PARSED_BY_DOXYGEN
 
+    EIGEN_DEVICE_FUNC
     CommaInitializer<Derived> operator<< (const Scalar& s);
 
+    /** \deprecated it now returns \c *this */
     template<unsigned int Added,unsigned int Removed>
-    const Flagged<Derived, Added, Removed> flagged() const;
+    EIGEN_DEPRECATED
+    const Derived& flagged() const
+    { return derived(); }
 
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     CommaInitializer<Derived> operator<< (const DenseBase<OtherDerived>& other);
 
-    Eigen::Transpose<Derived> transpose();
-	typedef typename internal::add_const<Transpose<const Derived> >::type ConstTransposeReturnType;
+    typedef Transpose<Derived> TransposeReturnType;
+    EIGEN_DEVICE_FUNC
+    TransposeReturnType transpose();
+    typedef typename internal::add_const<Transpose<const Derived> >::type ConstTransposeReturnType;
+    EIGEN_DEVICE_FUNC
     ConstTransposeReturnType transpose() const;
+    EIGEN_DEVICE_FUNC
     void transposeInPlace();
-#ifndef EIGEN_NO_DEBUG
-  protected:
-    template<typename OtherDerived>
-    void checkTransposeAliasing(const OtherDerived& other) const;
-  public:
-#endif
 
-
-    static const ConstantReturnType
+    EIGEN_DEVICE_FUNC static const ConstantReturnType
     Constant(Index rows, Index cols, const Scalar& value);
-    static const ConstantReturnType
+    EIGEN_DEVICE_FUNC static const ConstantReturnType
     Constant(Index size, const Scalar& value);
-    static const ConstantReturnType
+    EIGEN_DEVICE_FUNC static const ConstantReturnType
     Constant(const Scalar& value);
 
-    static const SequentialLinSpacedReturnType
+    EIGEN_DEVICE_FUNC static const SequentialLinSpacedReturnType
     LinSpaced(Sequential_t, Index size, const Scalar& low, const Scalar& high);
-    static const RandomAccessLinSpacedReturnType
+    EIGEN_DEVICE_FUNC static const RandomAccessLinSpacedReturnType
     LinSpaced(Index size, const Scalar& low, const Scalar& high);
-    static const SequentialLinSpacedReturnType
+    EIGEN_DEVICE_FUNC static const SequentialLinSpacedReturnType
     LinSpaced(Sequential_t, const Scalar& low, const Scalar& high);
-    static const RandomAccessLinSpacedReturnType
+    EIGEN_DEVICE_FUNC static const RandomAccessLinSpacedReturnType
     LinSpaced(const Scalar& low, const Scalar& high);
 
-    template<typename CustomNullaryOp>
-    static const CwiseNullaryOp<CustomNullaryOp, Derived>
+    template<typename CustomNullaryOp> EIGEN_DEVICE_FUNC
+    static const CwiseNullaryOp<CustomNullaryOp, PlainObject>
     NullaryExpr(Index rows, Index cols, const CustomNullaryOp& func);
-    template<typename CustomNullaryOp>
-    static const CwiseNullaryOp<CustomNullaryOp, Derived>
+    template<typename CustomNullaryOp> EIGEN_DEVICE_FUNC
+    static const CwiseNullaryOp<CustomNullaryOp, PlainObject>
     NullaryExpr(Index size, const CustomNullaryOp& func);
-    template<typename CustomNullaryOp>
-    static const CwiseNullaryOp<CustomNullaryOp, Derived>
+    template<typename CustomNullaryOp> EIGEN_DEVICE_FUNC
+    static const CwiseNullaryOp<CustomNullaryOp, PlainObject>
     NullaryExpr(const CustomNullaryOp& func);
 
-    static const ConstantReturnType Zero(Index rows, Index cols);
-    static const ConstantReturnType Zero(Index size);
-    static const ConstantReturnType Zero();
-    static const ConstantReturnType Ones(Index rows, Index cols);
-    static const ConstantReturnType Ones(Index size);
-    static const ConstantReturnType Ones();
+    EIGEN_DEVICE_FUNC static const ConstantReturnType Zero(Index rows, Index cols);
+    EIGEN_DEVICE_FUNC static const ConstantReturnType Zero(Index size);
+    EIGEN_DEVICE_FUNC static const ConstantReturnType Zero();
+    EIGEN_DEVICE_FUNC static const ConstantReturnType Ones(Index rows, Index cols);
+    EIGEN_DEVICE_FUNC static const ConstantReturnType Ones(Index size);
+    EIGEN_DEVICE_FUNC static const ConstantReturnType Ones();
 
-    void fill(const Scalar& value);
-    Derived& setConstant(const Scalar& value);
-    Derived& setLinSpaced(Index size, const Scalar& low, const Scalar& high);
-    Derived& setLinSpaced(const Scalar& low, const Scalar& high);
-    Derived& setZero();
-    Derived& setOnes();
-    Derived& setRandom();
+    EIGEN_DEVICE_FUNC void fill(const Scalar& value);
+    EIGEN_DEVICE_FUNC Derived& setConstant(const Scalar& value);
+    EIGEN_DEVICE_FUNC Derived& setLinSpaced(Index size, const Scalar& low, const Scalar& high);
+    EIGEN_DEVICE_FUNC Derived& setLinSpaced(const Scalar& low, const Scalar& high);
+    EIGEN_DEVICE_FUNC Derived& setZero();
+    EIGEN_DEVICE_FUNC Derived& setOnes();
+    EIGEN_DEVICE_FUNC Derived& setRandom();
 
-    template<typename OtherDerived>
+    template<typename OtherDerived> EIGEN_DEVICE_FUNC
     bool isApprox(const DenseBase<OtherDerived>& other,
                   const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const;
+    EIGEN_DEVICE_FUNC 
     bool isMuchSmallerThan(const RealScalar& other,
                            const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const;
-    template<typename OtherDerived>
+    template<typename OtherDerived> EIGEN_DEVICE_FUNC
     bool isMuchSmallerThan(const DenseBase<OtherDerived>& other,
                            const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const;
 
-    bool isApproxToConstant(const Scalar& value, const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const;
-    bool isConstant(const Scalar& value, const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const;
-    bool isZero(const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const;
-    bool isOnes(const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const;
+    EIGEN_DEVICE_FUNC bool isApproxToConstant(const Scalar& value, const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const;
+    EIGEN_DEVICE_FUNC bool isConstant(const Scalar& value, const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const;
+    EIGEN_DEVICE_FUNC bool isZero(const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const;
+    EIGEN_DEVICE_FUNC bool isOnes(const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const;
     
     inline bool hasNaN() const;
-    inline bool isFinite() const;
+    inline bool allFinite() const;
 
+    EIGEN_DEVICE_FUNC
     inline Derived& operator*=(const Scalar& other);
+    EIGEN_DEVICE_FUNC
     inline Derived& operator/=(const Scalar& other);
 
     typedef typename internal::add_const_on_value_type<typename internal::eval<Derived>::type>::type EvalReturnType;
@@ -358,7 +405,10 @@ template<typename Derived> class DenseBase
       *
       * Notice that in the case of a plain matrix or vector (not an expression) this function just returns
       * a const reference, in order to avoid a useless copy.
+      * 
+      * \warning Be carefull with eval() and the auto C++ keyword, as detailed in this \link TopicPitfalls_auto_keyword page \endlink.
       */
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE EvalReturnType eval() const
     {
       // Even though MSVC does not honor strong inlining when the return type
@@ -366,61 +416,68 @@ template<typename Derived> class DenseBase
       // size types on MSVC.
       return typename internal::eval<Derived>::type(derived());
     }
-
+    
     /** swaps *this with the expression \a other.
       *
       */
     template<typename OtherDerived>
-    void swap(const DenseBase<OtherDerived>& other,
-              int = OtherDerived::ThisConstantIsPrivateInPlainObjectBase)
+    EIGEN_DEVICE_FUNC
+    void swap(const DenseBase<OtherDerived>& other)
     {
-      SwapWrapper<Derived>(derived()).lazyAssign(other.derived());
+      EIGEN_STATIC_ASSERT(!OtherDerived::IsPlainObjectBase,THIS_EXPRESSION_IS_NOT_A_LVALUE__IT_IS_READ_ONLY);
+      eigen_assert(rows()==other.rows() && cols()==other.cols());
+      call_assignment(derived(), other.const_cast_derived(), internal::swap_assign_op<Scalar>());
     }
 
     /** swaps *this with the matrix or array \a other.
       *
       */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     void swap(PlainObjectBase<OtherDerived>& other)
     {
-      SwapWrapper<Derived>(derived()).lazyAssign(other.derived());
+      eigen_assert(rows()==other.rows() && cols()==other.cols());
+      call_assignment(derived(), other.derived(), internal::swap_assign_op<Scalar>());
     }
 
+    EIGEN_DEVICE_FUNC inline const NestByValue<Derived> nestByValue() const;
+    EIGEN_DEVICE_FUNC inline const ForceAlignedAccess<Derived> forceAlignedAccess() const;
+    EIGEN_DEVICE_FUNC inline ForceAlignedAccess<Derived> forceAlignedAccess();
+    template<bool Enable> EIGEN_DEVICE_FUNC
+    inline const typename internal::conditional<Enable,ForceAlignedAccess<Derived>,Derived&>::type forceAlignedAccessIf() const;
+    template<bool Enable> EIGEN_DEVICE_FUNC
+    inline typename internal::conditional<Enable,ForceAlignedAccess<Derived>,Derived&>::type forceAlignedAccessIf();
 
-    inline const NestByValue<Derived> nestByValue() const;
-    inline const ForceAlignedAccess<Derived> forceAlignedAccess() const;
-    inline ForceAlignedAccess<Derived> forceAlignedAccess();
-    template<bool Enable> inline const typename internal::conditional<Enable,ForceAlignedAccess<Derived>,Derived&>::type forceAlignedAccessIf() const;
-    template<bool Enable> inline typename internal::conditional<Enable,ForceAlignedAccess<Derived>,Derived&>::type forceAlignedAccessIf();
+    EIGEN_DEVICE_FUNC Scalar sum() const;
+    EIGEN_DEVICE_FUNC Scalar mean() const;
+    EIGEN_DEVICE_FUNC Scalar trace() const;
 
-    Scalar sum() const;
-    Scalar mean() const;
-    Scalar trace() const;
+    EIGEN_DEVICE_FUNC Scalar prod() const;
 
-    Scalar prod() const;
+    EIGEN_DEVICE_FUNC typename internal::traits<Derived>::Scalar minCoeff() const;
+    EIGEN_DEVICE_FUNC typename internal::traits<Derived>::Scalar maxCoeff() const;
 
-    typename internal::traits<Derived>::Scalar minCoeff() const;
-    typename internal::traits<Derived>::Scalar maxCoeff() const;
-
-    template<typename IndexType>
+    template<typename IndexType> EIGEN_DEVICE_FUNC
     typename internal::traits<Derived>::Scalar minCoeff(IndexType* row, IndexType* col) const;
-    template<typename IndexType>
+    template<typename IndexType> EIGEN_DEVICE_FUNC
     typename internal::traits<Derived>::Scalar maxCoeff(IndexType* row, IndexType* col) const;
-    template<typename IndexType>
+    template<typename IndexType> EIGEN_DEVICE_FUNC
     typename internal::traits<Derived>::Scalar minCoeff(IndexType* index) const;
-    template<typename IndexType>
+    template<typename IndexType> EIGEN_DEVICE_FUNC
     typename internal::traits<Derived>::Scalar maxCoeff(IndexType* index) const;
 
     template<typename BinaryOp>
-    typename internal::result_of<BinaryOp(typename internal::traits<Derived>::Scalar)>::type
-    redux(const BinaryOp& func) const;
+    EIGEN_DEVICE_FUNC
+    Scalar redux(const BinaryOp& func) const;
 
     template<typename Visitor>
+    EIGEN_DEVICE_FUNC
     void visit(Visitor& func) const;
 
     inline const WithFormat<Derived> format(const IOFormat& fmt) const;
 
     /** \returns the unique coefficient of a 1x1 expression */
+    EIGEN_DEVICE_FUNC
     CoeffReturnType value() const
     {
       EIGEN_STATIC_ASSERT_SIZE_1x1(Derived)
@@ -428,8 +485,8 @@ template<typename Derived> class DenseBase
       return derived().coeff(0,0);
     }
 
-    bool all(void) const;
-    bool any(void) const;
+    bool all() const;
+    bool any() const;
     Index count() const;
 
     typedef VectorwiseOp<Derived, Horizontal> RowwiseReturnType;
@@ -437,14 +494,35 @@ template<typename Derived> class DenseBase
     typedef VectorwiseOp<Derived, Vertical> ColwiseReturnType;
     typedef const VectorwiseOp<const Derived, Vertical> ConstColwiseReturnType;
 
-    ConstRowwiseReturnType rowwise() const;
-    RowwiseReturnType rowwise();
-    ConstColwiseReturnType colwise() const;
-    ColwiseReturnType colwise();
+    /** \returns a VectorwiseOp wrapper of *this providing additional partial reduction operations
+    *
+    * Example: \include MatrixBase_rowwise.cpp
+    * Output: \verbinclude MatrixBase_rowwise.out
+    *
+    * \sa colwise(), class VectorwiseOp, \ref TutorialReductionsVisitorsBroadcasting
+    */
+    //Code moved here due to a CUDA compiler bug
+    EIGEN_DEVICE_FUNC inline ConstRowwiseReturnType rowwise() const {
+      return ConstRowwiseReturnType(derived());
+    }
+    EIGEN_DEVICE_FUNC RowwiseReturnType rowwise();
 
-    static const CwiseNullaryOp<internal::scalar_random_op<Scalar>,Derived> Random(Index rows, Index cols);
-    static const CwiseNullaryOp<internal::scalar_random_op<Scalar>,Derived> Random(Index size);
-    static const CwiseNullaryOp<internal::scalar_random_op<Scalar>,Derived> Random();
+    /** \returns a VectorwiseOp wrapper of *this providing additional partial reduction operations
+    *
+    * Example: \include MatrixBase_colwise.cpp
+    * Output: \verbinclude MatrixBase_colwise.out
+    *
+    * \sa rowwise(), class VectorwiseOp, \ref TutorialReductionsVisitorsBroadcasting
+    */
+    EIGEN_DEVICE_FUNC inline ConstColwiseReturnType colwise() const {
+      return ConstColwiseReturnType(derived());
+    }
+    EIGEN_DEVICE_FUNC ColwiseReturnType colwise();
+
+    typedef CwiseNullaryOp<internal::scalar_random_op<Scalar>,PlainObject> RandomReturnType;
+    static const RandomReturnType Random(Index rows, Index cols);
+    static const RandomReturnType Random(Index size);
+    static const RandomReturnType Random();
 
     template<typename ThenDerived,typename ElseDerived>
     const Select<Derived,ThenDerived,ElseDerived>
@@ -462,14 +540,33 @@ template<typename Derived> class DenseBase
     template<int p> RealScalar lpNorm() const;
 
     template<int RowFactor, int ColFactor>
+    EIGEN_DEVICE_FUNC
     const Replicate<Derived,RowFactor,ColFactor> replicate() const;
-    const Replicate<Derived,Dynamic,Dynamic> replicate(Index rowFacor,Index colFactor) const;
+    /**
+    * \return an expression of the replication of \c *this
+    *
+    * Example: \include MatrixBase_replicate_int_int.cpp
+    * Output: \verbinclude MatrixBase_replicate_int_int.out
+    *
+    * \sa VectorwiseOp::replicate(), DenseBase::replicate<int,int>(), class Replicate
+    */
+    //Code moved here due to a CUDA compiler bug
+    EIGEN_DEVICE_FUNC
+    const Replicate<Derived, Dynamic, Dynamic> replicate(Index rowFactor, Index colFactor) const
+    {
+      return Replicate<Derived, Dynamic, Dynamic>(derived(), rowFactor, colFactor);
+    }
 
     typedef Reverse<Derived, BothDirections> ReverseReturnType;
     typedef const Reverse<const Derived, BothDirections> ConstReverseReturnType;
-    ReverseReturnType reverse();
-    ConstReverseReturnType reverse() const;
-    void reverseInPlace();
+    EIGEN_DEVICE_FUNC ReverseReturnType reverse();
+    /** This is the const version of reverse(). */
+    //Code moved here due to a CUDA compiler bug
+    EIGEN_DEVICE_FUNC ConstReverseReturnType reverse() const
+    {
+      return ConstReverseReturnType(derived());
+    }
+    EIGEN_DEVICE_FUNC void reverseInPlace();
 
 #define EIGEN_CURRENT_STORAGE_BASE_CLASS Eigen::DenseBase
 #   include "../plugins/BlockMethods.h"
@@ -478,27 +575,18 @@ template<typename Derived> class DenseBase
 #   endif
 #undef EIGEN_CURRENT_STORAGE_BASE_CLASS
 
-#ifdef EIGEN2_SUPPORT
-
-    Block<Derived> corner(CornerType type, Index cRows, Index cCols);
-    const Block<Derived> corner(CornerType type, Index cRows, Index cCols) const;
-    template<int CRows, int CCols>
-    Block<Derived, CRows, CCols> corner(CornerType type);
-    template<int CRows, int CCols>
-    const Block<Derived, CRows, CCols> corner(CornerType type) const;
-
-#endif // EIGEN2_SUPPORT
-
 
     // disable the use of evalTo for dense objects with a nice compilation error
-    template<typename Dest> inline void evalTo(Dest& ) const
+    template<typename Dest>
+    EIGEN_DEVICE_FUNC
+    inline void evalTo(Dest& ) const
     {
       EIGEN_STATIC_ASSERT((internal::is_same<Dest,void>::value),THE_EVAL_EVALTO_FUNCTION_SHOULD_NEVER_BE_CALLED_FOR_DENSE_OBJECTS);
     }
 
   protected:
     /** Default constructor. Do nothing. */
-    DenseBase()
+    EIGEN_DEVICE_FUNC DenseBase()
     {
       /* Just checks for self-consistency of the flags.
        * Only do it when debugging Eigen, as this borders on paranoiac and could slow compilation down
@@ -511,9 +599,9 @@ template<typename Derived> class DenseBase
     }
 
   private:
-    explicit DenseBase(int);
-    DenseBase(int,int);
-    template<typename OtherDerived> explicit DenseBase(const DenseBase<OtherDerived>&);
+    EIGEN_DEVICE_FUNC explicit DenseBase(int);
+    EIGEN_DEVICE_FUNC DenseBase(int,int);
+    template<typename OtherDerived> EIGEN_DEVICE_FUNC explicit DenseBase(const DenseBase<OtherDerived>&);
 };
 
 } // end namespace Eigen

Eigen/src/Core/DenseCoeffsBase.h

@@ -35,7 +35,6 @@ class DenseCoeffsBase<Derived,ReadOnlyAccessors> : public EigenBase<Derived>
 {
   public:
     typedef typename internal::traits<Derived>::StorageKind StorageKind;
-    typedef typename internal::traits<Derived>::Index Index;
     typedef typename internal::traits<Derived>::Scalar Scalar;
     typedef typename internal::packet_traits<Scalar>::type PacketScalar;
 
@@ -61,6 +60,7 @@ class DenseCoeffsBase<Derived,ReadOnlyAccessors> : public EigenBase<Derived>
     using Base::size;
     using Base::derived;
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Index rowIndexByOuterInner(Index outer, Index inner) const
     {
       return int(Derived::RowsAtCompileTime) == 1 ? 0
@@ -69,6 +69,7 @@ class DenseCoeffsBase<Derived,ReadOnlyAccessors> : public EigenBase<Derived>
           : inner;
     }
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Index colIndexByOuterInner(Index outer, Index inner) const
     {
       return int(Derived::ColsAtCompileTime) == 1 ? 0
@@ -91,13 +92,15 @@ class DenseCoeffsBase<Derived,ReadOnlyAccessors> : public EigenBase<Derived>
       *
       * \sa operator()(Index,Index) const, coeffRef(Index,Index), coeff(Index) const
       */
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE CoeffReturnType coeff(Index row, Index col) const
     {
       eigen_internal_assert(row >= 0 && row < rows()
-                        && col >= 0 && col < cols());
-      return derived().coeff(row, col);
+                         && col >= 0 && col < cols());
+      return internal::evaluator<Derived>(derived()).coeff(row,col);
     }
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE CoeffReturnType coeffByOuterInner(Index outer, Index inner) const
     {
       return coeff(rowIndexByOuterInner(outer, inner),
@@ -108,11 +111,12 @@ class DenseCoeffsBase<Derived,ReadOnlyAccessors> : public EigenBase<Derived>
       *
       * \sa operator()(Index,Index), operator[](Index)
       */
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE CoeffReturnType operator()(Index row, Index col) const
     {
       eigen_assert(row >= 0 && row < rows()
           && col >= 0 && col < cols());
-      return derived().coeff(row, col);
+      return coeff(row, col);
     }
 
     /** Short version: don't use this function, use
@@ -130,11 +134,12 @@ class DenseCoeffsBase<Derived,ReadOnlyAccessors> : public EigenBase<Derived>
       * \sa operator[](Index) const, coeffRef(Index), coeff(Index,Index) const
       */
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE CoeffReturnType
     coeff(Index index) const
     {
       eigen_internal_assert(index >= 0 && index < size());
-      return derived().coeff(index);
+      return internal::evaluator<Derived>(derived()).coeff(index);
     }
 
 
@@ -146,15 +151,14 @@ class DenseCoeffsBase<Derived,ReadOnlyAccessors> : public EigenBase<Derived>
       * z() const, w() const
       */
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE CoeffReturnType
     operator[](Index index) const
     {
-      #ifndef EIGEN2_SUPPORT
       EIGEN_STATIC_ASSERT(Derived::IsVectorAtCompileTime,
                           THE_BRACKET_OPERATOR_IS_ONLY_FOR_VECTORS__USE_THE_PARENTHESIS_OPERATOR_INSTEAD)
-      #endif
       eigen_assert(index >= 0 && index < size());
-      return derived().coeff(index);
+      return coeff(index);
     }
 
     /** \returns the coefficient at given index.
@@ -167,30 +171,35 @@ class DenseCoeffsBase<Derived,ReadOnlyAccessors> : public EigenBase<Derived>
       * z() const, w() const
       */
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE CoeffReturnType
     operator()(Index index) const
     {
       eigen_assert(index >= 0 && index < size());
-      return derived().coeff(index);
+      return coeff(index);
     }
 
     /** equivalent to operator[](0).  */
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE CoeffReturnType
     x() const { return (*this)[0]; }
 
     /** equivalent to operator[](1).  */
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE CoeffReturnType
     y() const { return (*this)[1]; }
 
     /** equivalent to operator[](2).  */
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE CoeffReturnType
     z() const { return (*this)[2]; }
 
     /** equivalent to operator[](3).  */
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE CoeffReturnType
     w() const { return (*this)[3]; }
 
@@ -207,9 +216,9 @@ class DenseCoeffsBase<Derived,ReadOnlyAccessors> : public EigenBase<Derived>
     template<int LoadMode>
     EIGEN_STRONG_INLINE PacketReturnType packet(Index row, Index col) const
     {
-      eigen_internal_assert(row >= 0 && row < rows()
-                      && col >= 0 && col < cols());
-      return derived().template packet<LoadMode>(row,col);
+      typedef typename internal::packet_traits<Scalar>::type DefaultPacketType;
+      eigen_internal_assert(row >= 0 && row < rows() && col >= 0 && col < cols());
+      return internal::evaluator<Derived>(derived()).template packet<LoadMode,DefaultPacketType>(row,col);
     }
 
 
@@ -234,8 +243,9 @@ class DenseCoeffsBase<Derived,ReadOnlyAccessors> : public EigenBase<Derived>
     template<int LoadMode>
     EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const
     {
+      typedef typename internal::packet_traits<Scalar>::type DefaultPacketType;
       eigen_internal_assert(index >= 0 && index < size());
-      return derived().template packet<LoadMode>(index);
+      return internal::evaluator<Derived>(derived()).template packet<LoadMode,DefaultPacketType>(index);
     }
 
   protected:
@@ -278,7 +288,6 @@ class DenseCoeffsBase<Derived, WriteAccessors> : public DenseCoeffsBase<Derived,
     typedef DenseCoeffsBase<Derived, ReadOnlyAccessors> Base;
 
     typedef typename internal::traits<Derived>::StorageKind StorageKind;
-    typedef typename internal::traits<Derived>::Index Index;
     typedef typename internal::traits<Derived>::Scalar Scalar;
     typedef typename internal::packet_traits<Scalar>::type PacketScalar;
     typedef typename NumTraits<Scalar>::Real RealScalar;
@@ -311,13 +320,15 @@ class DenseCoeffsBase<Derived, WriteAccessors> : public DenseCoeffsBase<Derived,
       *
       * \sa operator()(Index,Index), coeff(Index, Index) const, coeffRef(Index)
       */
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Scalar& coeffRef(Index row, Index col)
     {
       eigen_internal_assert(row >= 0 && row < rows()
-                        && col >= 0 && col < cols());
-      return derived().coeffRef(row, col);
+                         && col >= 0 && col < cols());
+      return internal::evaluator<Derived>(derived()).coeffRef(row,col);
     }
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Scalar&
     coeffRefByOuterInner(Index outer, Index inner)
     {
@@ -330,12 +341,13 @@ class DenseCoeffsBase<Derived, WriteAccessors> : public DenseCoeffsBase<Derived,
       * \sa operator[](Index)
       */
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Scalar&
     operator()(Index row, Index col)
     {
       eigen_assert(row >= 0 && row < rows()
           && col >= 0 && col < cols());
-      return derived().coeffRef(row, col);
+      return coeffRef(row, col);
     }
 
 
@@ -354,11 +366,12 @@ class DenseCoeffsBase<Derived, WriteAccessors> : public DenseCoeffsBase<Derived,
       * \sa operator[](Index), coeff(Index) const, coeffRef(Index,Index)
       */
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Scalar&
     coeffRef(Index index)
     {
       eigen_internal_assert(index >= 0 && index < size());
-      return derived().coeffRef(index);
+      return internal::evaluator<Derived>(derived()).coeffRef(index);
     }
 
     /** \returns a reference to the coefficient at given index.
@@ -368,15 +381,14 @@ class DenseCoeffsBase<Derived, WriteAccessors> : public DenseCoeffsBase<Derived,
       * \sa operator[](Index) const, operator()(Index,Index), x(), y(), z(), w()
       */
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Scalar&
     operator[](Index index)
     {
-      #ifndef EIGEN2_SUPPORT
       EIGEN_STATIC_ASSERT(Derived::IsVectorAtCompileTime,
                           THE_BRACKET_OPERATOR_IS_ONLY_FOR_VECTORS__USE_THE_PARENTHESIS_OPERATOR_INSTEAD)
-      #endif
       eigen_assert(index >= 0 && index < size());
-      return derived().coeffRef(index);
+      return coeffRef(index);
     }
 
     /** \returns a reference to the coefficient at given index.
@@ -388,167 +400,37 @@ class DenseCoeffsBase<Derived, WriteAccessors> : public DenseCoeffsBase<Derived,
       * \sa operator[](Index) const, operator()(Index,Index), x(), y(), z(), w()
       */
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Scalar&
     operator()(Index index)
     {
       eigen_assert(index >= 0 && index < size());
-      return derived().coeffRef(index);
+      return coeffRef(index);
     }
 
     /** equivalent to operator[](0).  */
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Scalar&
     x() { return (*this)[0]; }
 
     /** equivalent to operator[](1).  */
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Scalar&
     y() { return (*this)[1]; }
 
     /** equivalent to operator[](2).  */
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Scalar&
     z() { return (*this)[2]; }
 
     /** equivalent to operator[](3).  */
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Scalar&
     w() { return (*this)[3]; }
-
-    /** \internal
-      * Stores the given packet of coefficients, at the given row and column of this expression. It is your responsibility
-      * to ensure that a packet really starts there. This method is only available on expressions having the
-      * PacketAccessBit.
-      *
-      * The \a LoadMode parameter may have the value \a #Aligned or \a #Unaligned. Its effect is to select
-      * the appropriate vectorization instruction. Aligned access is faster, but is only possible for packets
-      * starting at an address which is a multiple of the packet size.
-      */
-
-    template<int StoreMode>
-    EIGEN_STRONG_INLINE void writePacket
-    (Index row, Index col, const typename internal::packet_traits<Scalar>::type& val)
-    {
-      eigen_internal_assert(row >= 0 && row < rows()
-                        && col >= 0 && col < cols());
-      derived().template writePacket<StoreMode>(row,col,val);
-    }
-
-
-    /** \internal */
-    template<int StoreMode>
-    EIGEN_STRONG_INLINE void writePacketByOuterInner
-    (Index outer, Index inner, const typename internal::packet_traits<Scalar>::type& val)
-    {
-      writePacket<StoreMode>(rowIndexByOuterInner(outer, inner),
-                            colIndexByOuterInner(outer, inner),
-                            val);
-    }
-
-    /** \internal
-      * Stores the given packet of coefficients, at the given index in this expression. It is your responsibility
-      * to ensure that a packet really starts there. This method is only available on expressions having the
-      * PacketAccessBit and the LinearAccessBit.
-      *
-      * The \a LoadMode parameter may have the value \a Aligned or \a Unaligned. Its effect is to select
-      * the appropriate vectorization instruction. Aligned access is faster, but is only possible for packets
-      * starting at an address which is a multiple of the packet size.
-      */
-    template<int StoreMode>
-    EIGEN_STRONG_INLINE void writePacket
-    (Index index, const typename internal::packet_traits<Scalar>::type& val)
-    {
-      eigen_internal_assert(index >= 0 && index < size());
-      derived().template writePacket<StoreMode>(index,val);
-    }
-
-#ifndef EIGEN_PARSED_BY_DOXYGEN
-
-    /** \internal Copies the coefficient at position (row,col) of other into *this.
-      *
-      * This method is overridden in SwapWrapper, allowing swap() assignments to share 99% of their code
-      * with usual assignments.
-      *
-      * Outside of this internal usage, this method has probably no usefulness. It is hidden in the public API dox.
-      */
-
-    template<typename OtherDerived>
-    EIGEN_STRONG_INLINE void copyCoeff(Index row, Index col, const DenseBase<OtherDerived>& other)
-    {
-      eigen_internal_assert(row >= 0 && row < rows()
-                        && col >= 0 && col < cols());
-      derived().coeffRef(row, col) = other.derived().coeff(row, col);
-    }
-
-    /** \internal Copies the coefficient at the given index of other into *this.
-      *
-      * This method is overridden in SwapWrapper, allowing swap() assignments to share 99% of their code
-      * with usual assignments.
-      *
-      * Outside of this internal usage, this method has probably no usefulness. It is hidden in the public API dox.
-      */
-
-    template<typename OtherDerived>
-    EIGEN_STRONG_INLINE void copyCoeff(Index index, const DenseBase<OtherDerived>& other)
-    {
-      eigen_internal_assert(index >= 0 && index < size());
-      derived().coeffRef(index) = other.derived().coeff(index);
-    }
-
-
-    template<typename OtherDerived>
-    EIGEN_STRONG_INLINE void copyCoeffByOuterInner(Index outer, Index inner, const DenseBase<OtherDerived>& other)
-    {
-      const Index row = rowIndexByOuterInner(outer,inner);
-      const Index col = colIndexByOuterInner(outer,inner);
-      // derived() is important here: copyCoeff() may be reimplemented in Derived!
-      derived().copyCoeff(row, col, other);
-    }
-
-    /** \internal Copies the packet at position (row,col) of other into *this.
-      *
-      * This method is overridden in SwapWrapper, allowing swap() assignments to share 99% of their code
-      * with usual assignments.
-      *
-      * Outside of this internal usage, this method has probably no usefulness. It is hidden in the public API dox.
-      */
-
-    template<typename OtherDerived, int StoreMode, int LoadMode>
-    EIGEN_STRONG_INLINE void copyPacket(Index row, Index col, const DenseBase<OtherDerived>& other)
-    {
-      eigen_internal_assert(row >= 0 && row < rows()
-                        && col >= 0 && col < cols());
-      derived().template writePacket<StoreMode>(row, col,
-        other.derived().template packet<LoadMode>(row, col));
-    }
-
-    /** \internal Copies the packet at the given index of other into *this.
-      *
-      * This method is overridden in SwapWrapper, allowing swap() assignments to share 99% of their code
-      * with usual assignments.
-      *
-      * Outside of this internal usage, this method has probably no usefulness. It is hidden in the public API dox.
-      */
-
-    template<typename OtherDerived, int StoreMode, int LoadMode>
-    EIGEN_STRONG_INLINE void copyPacket(Index index, const DenseBase<OtherDerived>& other)
-    {
-      eigen_internal_assert(index >= 0 && index < size());
-      derived().template writePacket<StoreMode>(index,
-        other.derived().template packet<LoadMode>(index));
-    }
-
-    /** \internal */
-    template<typename OtherDerived, int StoreMode, int LoadMode>
-    EIGEN_STRONG_INLINE void copyPacketByOuterInner(Index outer, Index inner, const DenseBase<OtherDerived>& other)
-    {
-      const Index row = rowIndexByOuterInner(outer,inner);
-      const Index col = colIndexByOuterInner(outer,inner);
-      // derived() is important here: copyCoeff() may be reimplemented in Derived!
-      derived().template copyPacket< OtherDerived, StoreMode, LoadMode>(row, col, other);
-    }
-#endif
-
 };
 
 /** \brief Base class providing direct read-only coefficient access to matrices and arrays.
@@ -568,7 +450,6 @@ class DenseCoeffsBase<Derived, DirectAccessors> : public DenseCoeffsBase<Derived
   public:
 
     typedef DenseCoeffsBase<Derived, ReadOnlyAccessors> Base;
-    typedef typename internal::traits<Derived>::Index Index;
     typedef typename internal::traits<Derived>::Scalar Scalar;
     typedef typename NumTraits<Scalar>::Real RealScalar;
 
@@ -581,6 +462,7 @@ class DenseCoeffsBase<Derived, DirectAccessors> : public DenseCoeffsBase<Derived
       *
       * \sa outerStride(), rowStride(), colStride()
       */
+    EIGEN_DEVICE_FUNC
     inline Index innerStride() const
     {
       return derived().innerStride();
@@ -591,6 +473,7 @@ class DenseCoeffsBase<Derived, DirectAccessors> : public DenseCoeffsBase<Derived
       *
       * \sa innerStride(), rowStride(), colStride()
       */
+    EIGEN_DEVICE_FUNC
     inline Index outerStride() const
     {
       return derived().outerStride();
@@ -606,6 +489,7 @@ class DenseCoeffsBase<Derived, DirectAccessors> : public DenseCoeffsBase<Derived
       *
       * \sa innerStride(), outerStride(), colStride()
       */
+    EIGEN_DEVICE_FUNC
     inline Index rowStride() const
     {
       return Derived::IsRowMajor ? outerStride() : innerStride();
@@ -615,6 +499,7 @@ class DenseCoeffsBase<Derived, DirectAccessors> : public DenseCoeffsBase<Derived
       *
       * \sa innerStride(), outerStride(), rowStride()
       */
+    EIGEN_DEVICE_FUNC
     inline Index colStride() const
     {
       return Derived::IsRowMajor ? innerStride() : outerStride();
@@ -639,7 +524,6 @@ class DenseCoeffsBase<Derived, DirectWriteAccessors>
   public:
 
     typedef DenseCoeffsBase<Derived, WriteAccessors> Base;
-    typedef typename internal::traits<Derived>::Index Index;
     typedef typename internal::traits<Derived>::Scalar Scalar;
     typedef typename NumTraits<Scalar>::Real RealScalar;
 
@@ -652,6 +536,7 @@ class DenseCoeffsBase<Derived, DirectWriteAccessors>
       *
       * \sa outerStride(), rowStride(), colStride()
       */
+    EIGEN_DEVICE_FUNC
     inline Index innerStride() const
     {
       return derived().innerStride();
@@ -662,6 +547,7 @@ class DenseCoeffsBase<Derived, DirectWriteAccessors>
       *
       * \sa innerStride(), rowStride(), colStride()
       */
+    EIGEN_DEVICE_FUNC
     inline Index outerStride() const
     {
       return derived().outerStride();
@@ -677,6 +563,7 @@ class DenseCoeffsBase<Derived, DirectWriteAccessors>
       *
       * \sa innerStride(), outerStride(), colStride()
       */
+    EIGEN_DEVICE_FUNC
     inline Index rowStride() const
     {
       return Derived::IsRowMajor ? outerStride() : innerStride();
@@ -686,6 +573,7 @@ class DenseCoeffsBase<Derived, DirectWriteAccessors>
       *
       * \sa innerStride(), outerStride(), rowStride()
       */
+    EIGEN_DEVICE_FUNC
     inline Index colStride() const
     {
       return Derived::IsRowMajor ? innerStride() : outerStride();
@@ -694,33 +582,42 @@ class DenseCoeffsBase<Derived, DirectWriteAccessors>
 
 namespace internal {
 
-template<typename Derived, bool JustReturnZero>
+template<int Alignment, typename Derived, bool JustReturnZero>
 struct first_aligned_impl
 {
-  static inline typename Derived::Index run(const Derived&)
+  static inline Index run(const Derived&)
   { return 0; }
 };
 
-template<typename Derived>
-struct first_aligned_impl<Derived, false>
+template<int Alignment, typename Derived>
+struct first_aligned_impl<Alignment, Derived, false>
 {
-  static inline typename Derived::Index run(const Derived& m)
+  static inline Index run(const Derived& m)
   {
-    return internal::first_aligned(&m.const_cast_derived().coeffRef(0,0), m.size());
+    return internal::first_aligned<Alignment>(&m.const_cast_derived().coeffRef(0,0), m.size());
   }
 };
 
-/** \internal \returns the index of the first element of the array that is well aligned for vectorization.
+/** \internal \returns the index of the first element of the array stored by \a m that is properly aligned with respect to \a Alignment for vectorization.
+  *
+  * \tparam Alignment requested alignment in Bytes.
   *
   * There is also the variant first_aligned(const Scalar*, Integer) defined in Memory.h. See it for more
   * documentation.
   */
-template<typename Derived>
-static inline typename Derived::Index first_aligned(const Derived& m)
+template<int Alignment, typename Derived>
+static inline Index first_aligned(const DenseBase<Derived>& m)
 {
-  return first_aligned_impl
-          <Derived, (Derived::Flags & AlignedBit) || !(Derived::Flags & DirectAccessBit)>
-          ::run(m);
+  enum { ReturnZero = (int(evaluator<Derived>::Alignment) >= Alignment) || !(Derived::Flags & DirectAccessBit) };
+  return first_aligned_impl<Alignment, Derived, ReturnZero>::run(m.derived());
+}
+
+template<typename Derived>
+static inline Index first_default_aligned(const DenseBase<Derived>& m)
+{
+  typedef typename Derived::Scalar Scalar;
+  typedef typename packet_traits<Scalar>::type DefaultPacketType;
+  return first_aligned<unpacket_traits<DefaultPacketType>::alignment>(m);
 }
 
 template<typename Derived, bool HasDirectAccess = has_direct_access<Derived>::ret>

Eigen/src/Core/DenseStorage.h

@@ -3,7 +3,7 @@
 //
 // Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr>
 // Copyright (C) 2006-2009 Benoit Jacob <jacob.benoit.1@gmail.com>
-// Copyright (C) 2010 Hauke Heibel <hauke.heibel@gmail.com>
+// Copyright (C) 2010-2013 Hauke Heibel <hauke.heibel@gmail.com>
 //
 // This Source Code Form is subject to the terms of the Mozilla
 // Public License v. 2.0. If a copy of the MPL was not distributed
@@ -24,26 +24,37 @@ namespace internal {
 
 struct constructor_without_unaligned_array_assert {};
 
+template<typename T, int Size>
+EIGEN_DEVICE_FUNC
+void check_static_allocation_size()
+{
+  // if EIGEN_STACK_ALLOCATION_LIMIT is defined to 0, then no limit
+  #if EIGEN_STACK_ALLOCATION_LIMIT
+  EIGEN_STATIC_ASSERT(Size * sizeof(T) <= EIGEN_STACK_ALLOCATION_LIMIT, OBJECT_ALLOCATED_ON_STACK_IS_TOO_BIG);
+  #endif
+}
+
 /** \internal
   * Static array. If the MatrixOrArrayOptions require auto-alignment, the array will be automatically aligned:
   * to 16 bytes boundary if the total size is a multiple of 16 bytes.
   */
 template <typename T, int Size, int MatrixOrArrayOptions,
           int Alignment = (MatrixOrArrayOptions&DontAlign) ? 0
-                        : (((Size*sizeof(T))%16)==0) ? 16
-                        : 0 >
+                        : compute_default_alignment<T,Size>::value >
 struct plain_array
 {
   T array[Size];
 
-  plain_array() 
+  EIGEN_DEVICE_FUNC
+  plain_array()
   { 
-    EIGEN_STATIC_ASSERT(Size * sizeof(T) <= 128 * 128 * 8, OBJECT_ALLOCATED_ON_STACK_IS_TOO_BIG);
+    check_static_allocation_size<T,Size>();
   }
 
-  plain_array(constructor_without_unaligned_array_assert) 
+  EIGEN_DEVICE_FUNC
+  plain_array(constructor_without_unaligned_array_assert)
   { 
-    EIGEN_STATIC_ASSERT(Size * sizeof(T) <= 128 * 128 * 8, OBJECT_ALLOCATED_ON_STACK_IS_TOO_BIG);
+    check_static_allocation_size<T,Size>();
   }
 };
 
@@ -56,41 +67,100 @@ struct plain_array
   template<typename PtrType>
   EIGEN_ALWAYS_INLINE PtrType eigen_unaligned_array_assert_workaround_gcc47(PtrType array) { return array; }
   #define EIGEN_MAKE_UNALIGNED_ARRAY_ASSERT(sizemask) \
-    eigen_assert((reinterpret_cast<size_t>(eigen_unaligned_array_assert_workaround_gcc47(array)) & sizemask) == 0 \
+    eigen_assert((reinterpret_cast<size_t>(eigen_unaligned_array_assert_workaround_gcc47(array)) & (sizemask)) == 0 \
               && "this assertion is explained here: " \
               "http://eigen.tuxfamily.org/dox-devel/group__TopicUnalignedArrayAssert.html" \
               " **** READ THIS WEB PAGE !!! ****");
 #else
   #define EIGEN_MAKE_UNALIGNED_ARRAY_ASSERT(sizemask) \
-    eigen_assert((reinterpret_cast<size_t>(array) & sizemask) == 0 \
+    eigen_assert((reinterpret_cast<size_t>(array) & (sizemask)) == 0 \
               && "this assertion is explained here: " \
               "http://eigen.tuxfamily.org/dox-devel/group__TopicUnalignedArrayAssert.html" \
               " **** READ THIS WEB PAGE !!! ****");
 #endif
 
 template <typename T, int Size, int MatrixOrArrayOptions>
-struct plain_array<T, Size, MatrixOrArrayOptions, 16>
+struct plain_array<T, Size, MatrixOrArrayOptions, 8>
 {
-  EIGEN_USER_ALIGN16 T array[Size];
+  EIGEN_ALIGN_TO_BOUNDARY(8) T array[Size];
 
+  EIGEN_DEVICE_FUNC
   plain_array() 
-  { 
-    EIGEN_MAKE_UNALIGNED_ARRAY_ASSERT(0xf);
-    EIGEN_STATIC_ASSERT(Size * sizeof(T) <= 128 * 128 * 8, OBJECT_ALLOCATED_ON_STACK_IS_TOO_BIG);
+  {
+    EIGEN_MAKE_UNALIGNED_ARRAY_ASSERT(7);
+    check_static_allocation_size<T,Size>();
   }
 
+  EIGEN_DEVICE_FUNC
   plain_array(constructor_without_unaligned_array_assert) 
   { 
-    EIGEN_STATIC_ASSERT(Size * sizeof(T) <= 128 * 128 * 8, OBJECT_ALLOCATED_ON_STACK_IS_TOO_BIG);
+    check_static_allocation_size<T,Size>();
+  }
+};
+
+template <typename T, int Size, int MatrixOrArrayOptions>
+struct plain_array<T, Size, MatrixOrArrayOptions, 16>
+{
+  EIGEN_ALIGN_TO_BOUNDARY(16) T array[Size];
+
+  EIGEN_DEVICE_FUNC
+  plain_array() 
+  { 
+    EIGEN_MAKE_UNALIGNED_ARRAY_ASSERT(15);
+    check_static_allocation_size<T,Size>();
+  }
+
+  EIGEN_DEVICE_FUNC
+  plain_array(constructor_without_unaligned_array_assert) 
+  { 
+    check_static_allocation_size<T,Size>();
+  }
+};
+
+template <typename T, int Size, int MatrixOrArrayOptions>
+struct plain_array<T, Size, MatrixOrArrayOptions, 32>
+{
+  EIGEN_ALIGN_TO_BOUNDARY(32) T array[Size];
+
+  EIGEN_DEVICE_FUNC
+  plain_array() 
+  {
+    EIGEN_MAKE_UNALIGNED_ARRAY_ASSERT(31);
+    check_static_allocation_size<T,Size>();
+  }
+
+  EIGEN_DEVICE_FUNC
+  plain_array(constructor_without_unaligned_array_assert) 
+  { 
+    check_static_allocation_size<T,Size>();
+  }
+};
+
+template <typename T, int Size, int MatrixOrArrayOptions>
+struct plain_array<T, Size, MatrixOrArrayOptions, 64>
+{
+  EIGEN_ALIGN_TO_BOUNDARY(64) T array[Size];
+
+  EIGEN_DEVICE_FUNC
+  plain_array() 
+  { 
+    EIGEN_MAKE_UNALIGNED_ARRAY_ASSERT(63);
+    check_static_allocation_size<T,Size>();
+  }
+
+  EIGEN_DEVICE_FUNC
+  plain_array(constructor_without_unaligned_array_assert) 
+  { 
+    check_static_allocation_size<T,Size>();
   }
 };
 
 template <typename T, int MatrixOrArrayOptions, int Alignment>
 struct plain_array<T, 0, MatrixOrArrayOptions, Alignment>
 {
-  EIGEN_USER_ALIGN16 T array[1];
-  plain_array() {}
-  plain_array(constructor_without_unaligned_array_assert) {}
+  T array[1];
+  EIGEN_DEVICE_FUNC plain_array() {}
+  EIGEN_DEVICE_FUNC plain_array(constructor_without_unaligned_array_assert) {}
 };
 
 } // end namespace internal
@@ -114,33 +184,50 @@ template<typename T, int Size, int _Rows, int _Cols, int _Options> class DenseSt
 {
     internal::plain_array<T,Size,_Options> m_data;
   public:
-    inline DenseStorage() {}
-    inline DenseStorage(internal::constructor_without_unaligned_array_assert)
+    EIGEN_DEVICE_FUNC DenseStorage() {}
+    EIGEN_DEVICE_FUNC
+    explicit DenseStorage(internal::constructor_without_unaligned_array_assert)
       : m_data(internal::constructor_without_unaligned_array_assert()) {}
-    inline DenseStorage(DenseIndex,DenseIndex,DenseIndex) {}
-    inline void swap(DenseStorage& other) { std::swap(m_data,other.m_data); }
-    static inline DenseIndex rows(void) {return _Rows;}
-    static inline DenseIndex cols(void) {return _Cols;}
-    inline void conservativeResize(DenseIndex,DenseIndex,DenseIndex) {}
-    inline void resize(DenseIndex,DenseIndex,DenseIndex) {}
-    inline const T *data() const { return m_data.array; }
-    inline T *data() { return m_data.array; }
+    EIGEN_DEVICE_FUNC 
+    DenseStorage(const DenseStorage& other) : m_data(other.m_data) {}
+    EIGEN_DEVICE_FUNC 
+    DenseStorage& operator=(const DenseStorage& other)
+    { 
+      if (this != &other) m_data = other.m_data;
+      return *this; 
+    }
+    EIGEN_DEVICE_FUNC DenseStorage(Index size, Index rows, Index cols) {
+      EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN
+      eigen_internal_assert(size==rows*cols && rows==_Rows && cols==_Cols);
+      EIGEN_UNUSED_VARIABLE(size);
+      EIGEN_UNUSED_VARIABLE(rows);
+      EIGEN_UNUSED_VARIABLE(cols);
+    }
+    EIGEN_DEVICE_FUNC void swap(DenseStorage& other) { std::swap(m_data,other.m_data); }
+    EIGEN_DEVICE_FUNC static Index rows(void) {return _Rows;}
+    EIGEN_DEVICE_FUNC static Index cols(void) {return _Cols;}
+    EIGEN_DEVICE_FUNC void conservativeResize(Index,Index,Index) {}
+    EIGEN_DEVICE_FUNC void resize(Index,Index,Index) {}
+    EIGEN_DEVICE_FUNC const T *data() const { return m_data.array; }
+    EIGEN_DEVICE_FUNC T *data() { return m_data.array; }
 };
 
 // null matrix
 template<typename T, int _Rows, int _Cols, int _Options> class DenseStorage<T, 0, _Rows, _Cols, _Options>
 {
   public:
-    inline DenseStorage() {}
-    inline DenseStorage(internal::constructor_without_unaligned_array_assert) {}
-    inline DenseStorage(DenseIndex,DenseIndex,DenseIndex) {}
-    inline void swap(DenseStorage& ) {}
-    static inline DenseIndex rows(void) {return _Rows;}
-    static inline DenseIndex cols(void) {return _Cols;}
-    inline void conservativeResize(DenseIndex,DenseIndex,DenseIndex) {}
-    inline void resize(DenseIndex,DenseIndex,DenseIndex) {}
-    inline const T *data() const { return 0; }
-    inline T *data() { return 0; }
+    EIGEN_DEVICE_FUNC DenseStorage() {}
+    EIGEN_DEVICE_FUNC explicit DenseStorage(internal::constructor_without_unaligned_array_assert) {}
+    EIGEN_DEVICE_FUNC DenseStorage(const DenseStorage&) {}
+    EIGEN_DEVICE_FUNC DenseStorage& operator=(const DenseStorage&) { return *this; }
+    EIGEN_DEVICE_FUNC DenseStorage(Index,Index,Index) {}
+    EIGEN_DEVICE_FUNC void swap(DenseStorage& ) {}
+    EIGEN_DEVICE_FUNC static Index rows(void) {return _Rows;}
+    EIGEN_DEVICE_FUNC static Index cols(void) {return _Cols;}
+    EIGEN_DEVICE_FUNC void conservativeResize(Index,Index,Index) {}
+    EIGEN_DEVICE_FUNC void resize(Index,Index,Index) {}
+    EIGEN_DEVICE_FUNC const T *data() const { return 0; }
+    EIGEN_DEVICE_FUNC T *data() { return 0; }
 };
 
 // more specializations for null matrices; these are necessary to resolve ambiguities
@@ -157,86 +244,157 @@ template<typename T, int _Cols, int _Options> class DenseStorage<T, 0, Dynamic,
 template<typename T, int Size, int _Options> class DenseStorage<T, Size, Dynamic, Dynamic, _Options>
 {
     internal::plain_array<T,Size,_Options> m_data;
-    DenseIndex m_rows;
-    DenseIndex m_cols;
+    Index m_rows;
+    Index m_cols;
   public:
-    inline DenseStorage() : m_rows(0), m_cols(0) {}
-    inline DenseStorage(internal::constructor_without_unaligned_array_assert)
+    EIGEN_DEVICE_FUNC DenseStorage() : m_rows(0), m_cols(0) {}
+    EIGEN_DEVICE_FUNC explicit DenseStorage(internal::constructor_without_unaligned_array_assert)
       : m_data(internal::constructor_without_unaligned_array_assert()), m_rows(0), m_cols(0) {}
-    inline DenseStorage(DenseIndex, DenseIndex nbRows, DenseIndex nbCols) : m_rows(nbRows), m_cols(nbCols) {}
-    inline void swap(DenseStorage& other)
+    EIGEN_DEVICE_FUNC DenseStorage(const DenseStorage& other) : m_data(other.m_data), m_rows(other.m_rows), m_cols(other.m_cols) {}
+    EIGEN_DEVICE_FUNC DenseStorage& operator=(const DenseStorage& other) 
+    { 
+      if (this != &other)
+      {
+        m_data = other.m_data;
+        m_rows = other.m_rows;
+        m_cols = other.m_cols;
+      }
+      return *this; 
+    }
+    EIGEN_DEVICE_FUNC DenseStorage(Index, Index rows, Index cols) : m_rows(rows), m_cols(cols) {}
+    EIGEN_DEVICE_FUNC void swap(DenseStorage& other)
     { std::swap(m_data,other.m_data); std::swap(m_rows,other.m_rows); std::swap(m_cols,other.m_cols); }
-    inline DenseIndex rows() const {return m_rows;}
-    inline DenseIndex cols() const {return m_cols;}
-    inline void conservativeResize(DenseIndex, DenseIndex nbRows, DenseIndex nbCols) { m_rows = nbRows; m_cols = nbCols; }
-    inline void resize(DenseIndex, DenseIndex nbRows, DenseIndex nbCols) { m_rows = nbRows; m_cols = nbCols; }
-    inline const T *data() const { return m_data.array; }
-    inline T *data() { return m_data.array; }
+    EIGEN_DEVICE_FUNC Index rows() const {return m_rows;}
+    EIGEN_DEVICE_FUNC Index cols() const {return m_cols;}
+    EIGEN_DEVICE_FUNC void conservativeResize(Index, Index rows, Index cols) { m_rows = rows; m_cols = cols; }
+    EIGEN_DEVICE_FUNC void resize(Index, Index rows, Index cols) { m_rows = rows; m_cols = cols; }
+    EIGEN_DEVICE_FUNC const T *data() const { return m_data.array; }
+    EIGEN_DEVICE_FUNC T *data() { return m_data.array; }
 };
 
 // dynamic-size matrix with fixed-size storage and fixed width
 template<typename T, int Size, int _Cols, int _Options> class DenseStorage<T, Size, Dynamic, _Cols, _Options>
 {
     internal::plain_array<T,Size,_Options> m_data;
-    DenseIndex m_rows;
+    Index m_rows;
   public:
-    inline DenseStorage() : m_rows(0) {}
-    inline DenseStorage(internal::constructor_without_unaligned_array_assert)
+    EIGEN_DEVICE_FUNC DenseStorage() : m_rows(0) {}
+    EIGEN_DEVICE_FUNC explicit DenseStorage(internal::constructor_without_unaligned_array_assert)
       : m_data(internal::constructor_without_unaligned_array_assert()), m_rows(0) {}
-    inline DenseStorage(DenseIndex, DenseIndex nbRows, DenseIndex) : m_rows(nbRows) {}
-    inline void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_rows,other.m_rows); }
-    inline DenseIndex rows(void) const {return m_rows;}
-    inline DenseIndex cols(void) const {return _Cols;}
-    inline void conservativeResize(DenseIndex, DenseIndex nbRows, DenseIndex) { m_rows = nbRows; }
-    inline void resize(DenseIndex, DenseIndex nbRows, DenseIndex) { m_rows = nbRows; }
-    inline const T *data() const { return m_data.array; }
-    inline T *data() { return m_data.array; }
+    EIGEN_DEVICE_FUNC DenseStorage(const DenseStorage& other) : m_data(other.m_data), m_rows(other.m_rows) {}
+    EIGEN_DEVICE_FUNC DenseStorage& operator=(const DenseStorage& other) 
+    {
+      if (this != &other)
+      {
+        m_data = other.m_data;
+        m_rows = other.m_rows;
+      }
+      return *this; 
+    }
+    EIGEN_DEVICE_FUNC DenseStorage(Index, Index rows, Index) : m_rows(rows) {}
+    EIGEN_DEVICE_FUNC void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_rows,other.m_rows); }
+    EIGEN_DEVICE_FUNC Index rows(void) const {return m_rows;}
+    EIGEN_DEVICE_FUNC Index cols(void) const {return _Cols;}
+    EIGEN_DEVICE_FUNC void conservativeResize(Index, Index rows, Index) { m_rows = rows; }
+    EIGEN_DEVICE_FUNC void resize(Index, Index rows, Index) { m_rows = rows; }
+    EIGEN_DEVICE_FUNC const T *data() const { return m_data.array; }
+    EIGEN_DEVICE_FUNC T *data() { return m_data.array; }
 };
 
 // dynamic-size matrix with fixed-size storage and fixed height
 template<typename T, int Size, int _Rows, int _Options> class DenseStorage<T, Size, _Rows, Dynamic, _Options>
 {
     internal::plain_array<T,Size,_Options> m_data;
-    DenseIndex m_cols;
+    Index m_cols;
   public:
-    inline DenseStorage() : m_cols(0) {}
-    inline DenseStorage(internal::constructor_without_unaligned_array_assert)
+    EIGEN_DEVICE_FUNC DenseStorage() : m_cols(0) {}
+    EIGEN_DEVICE_FUNC explicit DenseStorage(internal::constructor_without_unaligned_array_assert)
       : m_data(internal::constructor_without_unaligned_array_assert()), m_cols(0) {}
-    inline DenseStorage(DenseIndex, DenseIndex, DenseIndex nbCols) : m_cols(nbCols) {}
-    inline void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_cols,other.m_cols); }
-    inline DenseIndex rows(void) const {return _Rows;}
-    inline DenseIndex cols(void) const {return m_cols;}
-    inline void conservativeResize(DenseIndex, DenseIndex, DenseIndex nbCols) { m_cols = nbCols; }
-    inline void resize(DenseIndex, DenseIndex, DenseIndex nbCols) { m_cols = nbCols; }
-    inline const T *data() const { return m_data.array; }
-    inline T *data() { return m_data.array; }
+    EIGEN_DEVICE_FUNC DenseStorage(const DenseStorage& other) : m_data(other.m_data), m_cols(other.m_cols) {}
+    EIGEN_DEVICE_FUNC DenseStorage& operator=(const DenseStorage& other)
+    {
+      if (this != &other)
+      {
+        m_data = other.m_data;
+        m_cols = other.m_cols;
+      }
+      return *this;
+    }
+    EIGEN_DEVICE_FUNC DenseStorage(Index, Index, Index cols) : m_cols(cols) {}
+    EIGEN_DEVICE_FUNC void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_cols,other.m_cols); }
+    EIGEN_DEVICE_FUNC Index rows(void) const {return _Rows;}
+    EIGEN_DEVICE_FUNC Index cols(void) const {return m_cols;}
+    void conservativeResize(Index, Index, Index cols) { m_cols = cols; }
+    void resize(Index, Index, Index cols) { m_cols = cols; }
+    EIGEN_DEVICE_FUNC const T *data() const { return m_data.array; }
+    EIGEN_DEVICE_FUNC T *data() { return m_data.array; }
 };
 
 // purely dynamic matrix.
 template<typename T, int _Options> class DenseStorage<T, Dynamic, Dynamic, Dynamic, _Options>
 {
     T *m_data;
-    DenseIndex m_rows;
-    DenseIndex m_cols;
+    Index m_rows;
+    Index m_cols;
   public:
-    inline DenseStorage() : m_data(0), m_rows(0), m_cols(0) {}
-    inline DenseStorage(internal::constructor_without_unaligned_array_assert)
+    EIGEN_DEVICE_FUNC DenseStorage() : m_data(0), m_rows(0), m_cols(0) {}
+    EIGEN_DEVICE_FUNC explicit DenseStorage(internal::constructor_without_unaligned_array_assert)
        : m_data(0), m_rows(0), m_cols(0) {}
-    inline DenseStorage(DenseIndex size, DenseIndex nbRows, DenseIndex nbCols)
-      : m_data(internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(size)), m_rows(nbRows), m_cols(nbCols)
-    { EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN }
-    inline ~DenseStorage() { internal::conditional_aligned_delete_auto<T,(_Options&DontAlign)==0>(m_data, m_rows*m_cols); }
-    inline void swap(DenseStorage& other)
+    EIGEN_DEVICE_FUNC DenseStorage(Index size, Index rows, Index cols)
+      : m_data(internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(size)), m_rows(rows), m_cols(cols)
+    {
+      EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN
+      eigen_internal_assert(size==rows*cols && rows>=0 && cols >=0);
+    }
+    EIGEN_DEVICE_FUNC DenseStorage(const DenseStorage& other)
+      : m_data(internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(other.m_rows*other.m_cols))
+      , m_rows(other.m_rows)
+      , m_cols(other.m_cols)
+    {
+      internal::smart_copy(other.m_data, other.m_data+other.m_rows*other.m_cols, m_data);
+    }
+    EIGEN_DEVICE_FUNC DenseStorage& operator=(const DenseStorage& other)
+    {
+      if (this != &other)
+      {
+        DenseStorage tmp(other);
+        this->swap(tmp);
+      }
+      return *this;
+    }
+#ifdef EIGEN_HAVE_RVALUE_REFERENCES
+    EIGEN_DEVICE_FUNC
+    DenseStorage(DenseStorage&& other)
+      : m_data(std::move(other.m_data))
+      , m_rows(std::move(other.m_rows))
+      , m_cols(std::move(other.m_cols))
+    {
+      other.m_data = nullptr;
+      other.m_rows = 0;
+      other.m_cols = 0;
+    }
+    EIGEN_DEVICE_FUNC
+    DenseStorage& operator=(DenseStorage&& other)
+    {
+      using std::swap;
+      swap(m_data, other.m_data);
+      swap(m_rows, other.m_rows);
+      swap(m_cols, other.m_cols);
+      return *this;
+    }
+#endif
+    EIGEN_DEVICE_FUNC ~DenseStorage() { internal::conditional_aligned_delete_auto<T,(_Options&DontAlign)==0>(m_data, m_rows*m_cols); }
+    EIGEN_DEVICE_FUNC void swap(DenseStorage& other)
     { std::swap(m_data,other.m_data); std::swap(m_rows,other.m_rows); std::swap(m_cols,other.m_cols); }
-    inline DenseIndex rows(void) const {return m_rows;}
-    inline DenseIndex cols(void) const {return m_cols;}
-    inline void conservativeResize(DenseIndex size, DenseIndex nbRows, DenseIndex nbCols)
+    EIGEN_DEVICE_FUNC Index rows(void) const {return m_rows;}
+    EIGEN_DEVICE_FUNC Index cols(void) const {return m_cols;}
+    void conservativeResize(Index size, Index rows, Index cols)
     {
       m_data = internal::conditional_aligned_realloc_new_auto<T,(_Options&DontAlign)==0>(m_data, size, m_rows*m_cols);
-      m_rows = nbRows;
-      m_cols = nbCols;
+      m_rows = rows;
+      m_cols = cols;
     }
-    void resize(DenseIndex size, DenseIndex nbRows, DenseIndex nbCols)
+    EIGEN_DEVICE_FUNC void resize(Index size, Index rows, Index cols)
     {
       if(size != m_rows*m_cols)
       {
@@ -247,33 +405,70 @@ template<typename T, int _Options> class DenseStorage<T, Dynamic, Dynamic, Dynam
           m_data = 0;
         EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN
       }
-      m_rows = nbRows;
-      m_cols = nbCols;
+      m_rows = rows;
+      m_cols = cols;
     }
-    inline const T *data() const { return m_data; }
-    inline T *data() { return m_data; }
+    EIGEN_DEVICE_FUNC const T *data() const { return m_data; }
+    EIGEN_DEVICE_FUNC T *data() { return m_data; }
 };
 
 // matrix with dynamic width and fixed height (so that matrix has dynamic size).
 template<typename T, int _Rows, int _Options> class DenseStorage<T, Dynamic, _Rows, Dynamic, _Options>
 {
     T *m_data;
-    DenseIndex m_cols;
+    Index m_cols;
   public:
-    inline DenseStorage() : m_data(0), m_cols(0) {}
-    inline DenseStorage(internal::constructor_without_unaligned_array_assert) : m_data(0), m_cols(0) {}
-    inline DenseStorage(DenseIndex size, DenseIndex, DenseIndex nbCols) : m_data(internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(size)), m_cols(nbCols)
-    { EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN }
-    inline ~DenseStorage() { internal::conditional_aligned_delete_auto<T,(_Options&DontAlign)==0>(m_data, _Rows*m_cols); }
-    inline void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_cols,other.m_cols); }
-    static inline DenseIndex rows(void) {return _Rows;}
-    inline DenseIndex cols(void) const {return m_cols;}
-    inline void conservativeResize(DenseIndex size, DenseIndex, DenseIndex nbCols)
+    EIGEN_DEVICE_FUNC DenseStorage() : m_data(0), m_cols(0) {}
+    explicit DenseStorage(internal::constructor_without_unaligned_array_assert) : m_data(0), m_cols(0) {}
+    EIGEN_DEVICE_FUNC DenseStorage(Index size, Index rows, Index cols) : m_data(internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(size)), m_cols(cols)
+    {
+      EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN
+      eigen_internal_assert(size==rows*cols && rows==_Rows && cols >=0);
+      EIGEN_UNUSED_VARIABLE(rows);
+    }
+    EIGEN_DEVICE_FUNC DenseStorage(const DenseStorage& other)
+      : m_data(internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(_Rows*other.m_cols))
+      , m_cols(other.m_cols)
+    {
+      internal::smart_copy(other.m_data, other.m_data+_Rows*m_cols, m_data);
+    }
+    EIGEN_DEVICE_FUNC DenseStorage& operator=(const DenseStorage& other)
+    {
+      if (this != &other)
+      {
+        DenseStorage tmp(other);
+        this->swap(tmp);
+      }
+      return *this;
+    }    
+#ifdef EIGEN_HAVE_RVALUE_REFERENCES
+    EIGEN_DEVICE_FUNC
+    DenseStorage(DenseStorage&& other)
+      : m_data(std::move(other.m_data))
+      , m_cols(std::move(other.m_cols))
+    {
+      other.m_data = nullptr;
+      other.m_cols = 0;
+    }
+    EIGEN_DEVICE_FUNC
+    DenseStorage& operator=(DenseStorage&& other)
+    {
+      using std::swap;
+      swap(m_data, other.m_data);
+      swap(m_cols, other.m_cols);
+      return *this;
+    }
+#endif
+    EIGEN_DEVICE_FUNC ~DenseStorage() { internal::conditional_aligned_delete_auto<T,(_Options&DontAlign)==0>(m_data, _Rows*m_cols); }
+    EIGEN_DEVICE_FUNC void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_cols,other.m_cols); }
+    EIGEN_DEVICE_FUNC static Index rows(void) {return _Rows;}
+    EIGEN_DEVICE_FUNC Index cols(void) const {return m_cols;}
+    EIGEN_DEVICE_FUNC void conservativeResize(Index size, Index, Index cols)
     {
       m_data = internal::conditional_aligned_realloc_new_auto<T,(_Options&DontAlign)==0>(m_data, size, _Rows*m_cols);
-      m_cols = nbCols;
+      m_cols = cols;
     }
-    EIGEN_STRONG_INLINE void resize(DenseIndex size, DenseIndex, DenseIndex nbCols)
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void resize(Index size, Index, Index cols)
     {
       if(size != _Rows*m_cols)
       {
@@ -284,32 +479,69 @@ template<typename T, int _Rows, int _Options> class DenseStorage<T, Dynamic, _Ro
           m_data = 0;
         EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN
       }
-      m_cols = nbCols;
+      m_cols = cols;
     }
-    inline const T *data() const { return m_data; }
-    inline T *data() { return m_data; }
+    EIGEN_DEVICE_FUNC const T *data() const { return m_data; }
+    EIGEN_DEVICE_FUNC T *data() { return m_data; }
 };
 
 // matrix with dynamic height and fixed width (so that matrix has dynamic size).
 template<typename T, int _Cols, int _Options> class DenseStorage<T, Dynamic, Dynamic, _Cols, _Options>
 {
     T *m_data;
-    DenseIndex m_rows;
+    Index m_rows;
   public:
-    inline DenseStorage() : m_data(0), m_rows(0) {}
-    inline DenseStorage(internal::constructor_without_unaligned_array_assert) : m_data(0), m_rows(0) {}
-    inline DenseStorage(DenseIndex size, DenseIndex nbRows, DenseIndex) : m_data(internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(size)), m_rows(nbRows)
-    { EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN }
-    inline ~DenseStorage() { internal::conditional_aligned_delete_auto<T,(_Options&DontAlign)==0>(m_data, _Cols*m_rows); }
-    inline void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_rows,other.m_rows); }
-    inline DenseIndex rows(void) const {return m_rows;}
-    static inline DenseIndex cols(void) {return _Cols;}
-    inline void conservativeResize(DenseIndex size, DenseIndex nbRows, DenseIndex)
+    EIGEN_DEVICE_FUNC DenseStorage() : m_data(0), m_rows(0) {}
+    explicit DenseStorage(internal::constructor_without_unaligned_array_assert) : m_data(0), m_rows(0) {}
+    EIGEN_DEVICE_FUNC DenseStorage(Index size, Index rows, Index cols) : m_data(internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(size)), m_rows(rows)
+    {
+      EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN
+      eigen_internal_assert(size==rows*cols && rows>=0 && cols == _Cols);
+      EIGEN_UNUSED_VARIABLE(cols);
+    }
+    EIGEN_DEVICE_FUNC DenseStorage(const DenseStorage& other)
+      : m_data(internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(other.m_rows*_Cols))
+      , m_rows(other.m_rows)
+    {
+      internal::smart_copy(other.m_data, other.m_data+other.m_rows*_Cols, m_data);
+    }
+    EIGEN_DEVICE_FUNC DenseStorage& operator=(const DenseStorage& other)
+    {
+      if (this != &other)
+      {
+        DenseStorage tmp(other);
+        this->swap(tmp);
+      }
+      return *this;
+    }    
+#ifdef EIGEN_HAVE_RVALUE_REFERENCES
+    EIGEN_DEVICE_FUNC
+    DenseStorage(DenseStorage&& other)
+      : m_data(std::move(other.m_data))
+      , m_rows(std::move(other.m_rows))
+    {
+      other.m_data = nullptr;
+      other.m_rows = 0;
+    }
+    EIGEN_DEVICE_FUNC
+    DenseStorage& operator=(DenseStorage&& other)
+    {
+      using std::swap;
+      swap(m_data, other.m_data);
+      swap(m_rows, other.m_rows);
+      return *this;
+    }
+#endif
+    EIGEN_DEVICE_FUNC ~DenseStorage() { internal::conditional_aligned_delete_auto<T,(_Options&DontAlign)==0>(m_data, _Cols*m_rows); }
+    EIGEN_DEVICE_FUNC void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_rows,other.m_rows); }
+    EIGEN_DEVICE_FUNC Index rows(void) const {return m_rows;}
+    EIGEN_DEVICE_FUNC static Index cols(void) {return _Cols;}
+    void conservativeResize(Index size, Index rows, Index)
     {
       m_data = internal::conditional_aligned_realloc_new_auto<T,(_Options&DontAlign)==0>(m_data, size, m_rows*_Cols);
-      m_rows = nbRows;
+      m_rows = rows;
     }
-    EIGEN_STRONG_INLINE void resize(DenseIndex size, DenseIndex nbRows, DenseIndex)
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void resize(Index size, Index rows, Index)
     {
       if(size != m_rows*_Cols)
       {
@@ -320,10 +552,10 @@ template<typename T, int _Cols, int _Options> class DenseStorage<T, Dynamic, Dyn
           m_data = 0;
         EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN
       }
-      m_rows = nbRows;
+      m_rows = rows;
     }
-    inline const T *data() const { return m_data; }
-    inline T *data() { return m_data; }
+    EIGEN_DEVICE_FUNC const T *data() const { return m_data; }
+    EIGEN_DEVICE_FUNC T *data() { return m_data; }
 };
 
 } // end namespace Eigen

Eigen/src/Core/Diagonal.h

@@ -37,7 +37,7 @@ template<typename MatrixType, int DiagIndex>
 struct traits<Diagonal<MatrixType,DiagIndex> >
  : traits<MatrixType>
 {
-  typedef typename nested<MatrixType>::type MatrixTypeNested;
+  typedef typename ref_selector<MatrixType>::type MatrixTypeNested;
   typedef typename remove_reference<MatrixTypeNested>::type _MatrixTypeNested;
   typedef typename MatrixType::StorageKind StorageKind;
   enum {
@@ -52,8 +52,7 @@ struct traits<Diagonal<MatrixType,DiagIndex> >
                                                  MatrixType::MaxColsAtCompileTime - EIGEN_PLAIN_ENUM_MAX( DiagIndex, 0))),
     MaxColsAtCompileTime = 1,
     MaskLvalueBit = is_lvalue<MatrixType>::value ? LvalueBit : 0,
-    Flags = (unsigned int)_MatrixTypeNested::Flags & (HereditaryBits | LinearAccessBit | MaskLvalueBit | DirectAccessBit) & ~RowMajorBit,
-    CoeffReadCost = _MatrixTypeNested::CoeffReadCost,
+    Flags = (unsigned int)_MatrixTypeNested::Flags & (RowMajorBit | MaskLvalueBit | DirectAccessBit) & ~RowMajorBit, // FIXME DirectAccessBit should not be handled by expressions
     MatrixTypeOuterStride = outer_stride_at_compile_time<MatrixType>::ret,
     InnerStrideAtCompileTime = MatrixTypeOuterStride == Dynamic ? Dynamic : MatrixTypeOuterStride+1,
     OuterStrideAtCompileTime = 0
@@ -70,20 +69,28 @@ template<typename MatrixType, int _DiagIndex> class Diagonal
     typedef typename internal::dense_xpr_base<Diagonal>::type Base;
     EIGEN_DENSE_PUBLIC_INTERFACE(Diagonal)
 
-    inline Diagonal(MatrixType& matrix, Index a_index = DiagIndex) : m_matrix(matrix), m_index(a_index) {}
+    EIGEN_DEVICE_FUNC
+    explicit inline Diagonal(MatrixType& matrix, Index a_index = DiagIndex) : m_matrix(matrix), m_index(a_index) {}
 
     EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Diagonal)
 
+    EIGEN_DEVICE_FUNC
     inline Index rows() const
-    { return m_index.value()<0 ? (std::min<Index>)(m_matrix.cols(),m_matrix.rows()+m_index.value()) : (std::min<Index>)(m_matrix.rows(),m_matrix.cols()-m_index.value()); }
+    {
+      return m_index.value()<0 ? numext::mini<Index>(m_matrix.cols(),m_matrix.rows()+m_index.value())
+                               : numext::mini<Index>(m_matrix.rows(),m_matrix.cols()-m_index.value());
+    }
 
+    EIGEN_DEVICE_FUNC
     inline Index cols() const { return 1; }
 
+    EIGEN_DEVICE_FUNC
     inline Index innerStride() const
     {
       return m_matrix.outerStride() + 1;
     }
 
+    EIGEN_DEVICE_FUNC
     inline Index outerStride() const
     {
       return 0;
@@ -95,48 +102,58 @@ template<typename MatrixType, int _DiagIndex> class Diagonal
                        const Scalar
                      >::type ScalarWithConstIfNotLvalue;
 
+    EIGEN_DEVICE_FUNC
     inline ScalarWithConstIfNotLvalue* data() { return &(m_matrix.const_cast_derived().coeffRef(rowOffset(), colOffset())); }
+    EIGEN_DEVICE_FUNC
     inline const Scalar* data() const { return &(m_matrix.const_cast_derived().coeffRef(rowOffset(), colOffset())); }
 
+    EIGEN_DEVICE_FUNC
     inline Scalar& coeffRef(Index row, Index)
     {
       EIGEN_STATIC_ASSERT_LVALUE(MatrixType)
       return m_matrix.const_cast_derived().coeffRef(row+rowOffset(), row+colOffset());
     }
 
+    EIGEN_DEVICE_FUNC
     inline const Scalar& coeffRef(Index row, Index) const
     {
       return m_matrix.const_cast_derived().coeffRef(row+rowOffset(), row+colOffset());
     }
 
+    EIGEN_DEVICE_FUNC
     inline CoeffReturnType coeff(Index row, Index) const
     {
       return m_matrix.coeff(row+rowOffset(), row+colOffset());
     }
 
+    EIGEN_DEVICE_FUNC
     inline Scalar& coeffRef(Index idx)
     {
       EIGEN_STATIC_ASSERT_LVALUE(MatrixType)
       return m_matrix.const_cast_derived().coeffRef(idx+rowOffset(), idx+colOffset());
     }
 
+    EIGEN_DEVICE_FUNC
     inline const Scalar& coeffRef(Index idx) const
     {
       return m_matrix.const_cast_derived().coeffRef(idx+rowOffset(), idx+colOffset());
     }
 
+    EIGEN_DEVICE_FUNC
     inline CoeffReturnType coeff(Index idx) const
     {
       return m_matrix.coeff(idx+rowOffset(), idx+colOffset());
     }
 
-    const typename internal::remove_all<typename MatrixType::Nested>::type& 
+    EIGEN_DEVICE_FUNC
+    inline const typename internal::remove_all<typename MatrixType::Nested>::type& 
     nestedExpression() const 
     {
       return m_matrix;
     }
 
-    int index() const
+    EIGEN_DEVICE_FUNC
+    inline Index index() const
     {
       return m_index.value();
     }
@@ -147,10 +164,13 @@ template<typename MatrixType, int _DiagIndex> class Diagonal
 
   private:
     // some compilers may fail to optimize std::max etc in case of compile-time constants...
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Index absDiagIndex() const { return m_index.value()>0 ? m_index.value() : -m_index.value(); }
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Index rowOffset() const { return m_index.value()>0 ? 0 : -m_index.value(); }
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Index colOffset() const { return m_index.value()>0 ? m_index.value() : 0; }
-    // triger a compile time error is someone try to call packet
+    // trigger a compile-time error if someone try to call packet
     template<int LoadMode> typename MatrixType::PacketReturnType packet(Index) const;
     template<int LoadMode> typename MatrixType::PacketReturnType packet(Index,Index) const;
 };
@@ -167,7 +187,7 @@ template<typename Derived>
 inline typename MatrixBase<Derived>::DiagonalReturnType
 MatrixBase<Derived>::diagonal()
 {
-  return derived();
+  return DiagonalReturnType(derived());
 }
 
 /** This is the const version of diagonal(). */
@@ -190,18 +210,18 @@ MatrixBase<Derived>::diagonal() const
   *
   * \sa MatrixBase::diagonal(), class Diagonal */
 template<typename Derived>
-inline typename MatrixBase<Derived>::template DiagonalIndexReturnType<DynamicIndex>::Type
+inline typename MatrixBase<Derived>::DiagonalDynamicIndexReturnType
 MatrixBase<Derived>::diagonal(Index index)
 {
-  return typename DiagonalIndexReturnType<DynamicIndex>::Type(derived(), index);
+  return DiagonalDynamicIndexReturnType(derived(), index);
 }
 
 /** This is the const version of diagonal(Index). */
 template<typename Derived>
-inline typename MatrixBase<Derived>::template ConstDiagonalIndexReturnType<DynamicIndex>::Type
+inline typename MatrixBase<Derived>::ConstDiagonalDynamicIndexReturnType
 MatrixBase<Derived>::diagonal(Index index) const
 {
-  return typename ConstDiagonalIndexReturnType<DynamicIndex>::Type(derived(), index);
+  return ConstDiagonalDynamicIndexReturnType(derived(), index);
 }
 
 /** \returns an expression of the \a DiagIndex-th sub or super diagonal of the matrix \c *this
@@ -216,20 +236,20 @@ MatrixBase<Derived>::diagonal(Index index) const
   *
   * \sa MatrixBase::diagonal(), class Diagonal */
 template<typename Derived>
-template<int Index>
-inline typename MatrixBase<Derived>::template DiagonalIndexReturnType<Index>::Type
+template<int Index_>
+inline typename MatrixBase<Derived>::template DiagonalIndexReturnType<Index_>::Type
 MatrixBase<Derived>::diagonal()
 {
-  return derived();
+  return typename DiagonalIndexReturnType<Index_>::Type(derived());
 }
 
 /** This is the const version of diagonal<int>(). */
 template<typename Derived>
-template<int Index>
-inline typename MatrixBase<Derived>::template ConstDiagonalIndexReturnType<Index>::Type
+template<int Index_>
+inline typename MatrixBase<Derived>::template ConstDiagonalIndexReturnType<Index_>::Type
 MatrixBase<Derived>::diagonal() const
 {
-  return derived();
+  return typename ConstDiagonalIndexReturnType<Index_>::Type(derived());
 }
 
 } // end namespace Eigen

Eigen/src/Core/DiagonalMatrix.h

@@ -22,7 +22,7 @@ class DiagonalBase : public EigenBase<Derived>
     typedef typename DiagonalVectorType::Scalar Scalar;
     typedef typename DiagonalVectorType::RealScalar RealScalar;
     typedef typename internal::traits<Derived>::StorageKind StorageKind;
-    typedef typename internal::traits<Derived>::Index Index;
+    typedef typename internal::traits<Derived>::StorageIndex StorageIndex;
 
     enum {
       RowsAtCompileTime = DiagonalVectorType::SizeAtCompileTime,
@@ -30,79 +30,62 @@ class DiagonalBase : public EigenBase<Derived>
       MaxRowsAtCompileTime = DiagonalVectorType::MaxSizeAtCompileTime,
       MaxColsAtCompileTime = DiagonalVectorType::MaxSizeAtCompileTime,
       IsVectorAtCompileTime = 0,
-      Flags = 0
+      Flags = NoPreferredStorageOrderBit
     };
 
     typedef Matrix<Scalar, RowsAtCompileTime, ColsAtCompileTime, 0, MaxRowsAtCompileTime, MaxColsAtCompileTime> DenseMatrixType;
     typedef DenseMatrixType DenseType;
     typedef DiagonalMatrix<Scalar,DiagonalVectorType::SizeAtCompileTime,DiagonalVectorType::MaxSizeAtCompileTime> PlainObject;
 
+    EIGEN_DEVICE_FUNC
     inline const Derived& derived() const { return *static_cast<const Derived*>(this); }
+    EIGEN_DEVICE_FUNC
     inline Derived& derived() { return *static_cast<Derived*>(this); }
 
+    EIGEN_DEVICE_FUNC
     DenseMatrixType toDenseMatrix() const { return derived(); }
-    template<typename DenseDerived>
-    void evalTo(MatrixBase<DenseDerived> &other) const;
-    template<typename DenseDerived>
-    void addTo(MatrixBase<DenseDerived> &other) const
-    { other.diagonal() += diagonal(); }
-    template<typename DenseDerived>
-    void subTo(MatrixBase<DenseDerived> &other) const
-    { other.diagonal() -= diagonal(); }
-
+    
+    EIGEN_DEVICE_FUNC
     inline const DiagonalVectorType& diagonal() const { return derived().diagonal(); }
+    EIGEN_DEVICE_FUNC
     inline DiagonalVectorType& diagonal() { return derived().diagonal(); }
 
+    EIGEN_DEVICE_FUNC
     inline Index rows() const { return diagonal().size(); }
+    EIGEN_DEVICE_FUNC
     inline Index cols() const { return diagonal().size(); }
 
-    /** \returns the diagonal matrix product of \c *this by the matrix \a matrix.
-      */
     template<typename MatrixDerived>
-    const DiagonalProduct<MatrixDerived, Derived, OnTheLeft>
+    EIGEN_DEVICE_FUNC
+    const Product<Derived,MatrixDerived,LazyProduct>
     operator*(const MatrixBase<MatrixDerived> &matrix) const
     {
-      return DiagonalProduct<MatrixDerived, Derived, OnTheLeft>(matrix.derived(), derived());
+      return Product<Derived, MatrixDerived, LazyProduct>(derived(),matrix.derived());
     }
 
-    inline const DiagonalWrapper<const CwiseUnaryOp<internal::scalar_inverse_op<Scalar>, const DiagonalVectorType> >
+    typedef DiagonalWrapper<const CwiseUnaryOp<internal::scalar_inverse_op<Scalar>, const DiagonalVectorType> > InverseReturnType;
+    EIGEN_DEVICE_FUNC
+    inline const InverseReturnType
     inverse() const
     {
-      return diagonal().cwiseInverse();
+      return InverseReturnType(diagonal().cwiseInverse());
     }
     
-    inline const DiagonalWrapper<const CwiseUnaryOp<internal::scalar_multiple_op<Scalar>, const DiagonalVectorType> >
+    typedef DiagonalWrapper<const CwiseUnaryOp<internal::scalar_multiple_op<Scalar>, const DiagonalVectorType> > ScalarMultipleReturnType;
+    EIGEN_DEVICE_FUNC
+    inline const ScalarMultipleReturnType
     operator*(const Scalar& scalar) const
     {
-      return diagonal() * scalar;
+      return ScalarMultipleReturnType(diagonal() * scalar);
     }
-    friend inline const DiagonalWrapper<const CwiseUnaryOp<internal::scalar_multiple_op<Scalar>, const DiagonalVectorType> >
+    EIGEN_DEVICE_FUNC
+    friend inline const ScalarMultipleReturnType
     operator*(const Scalar& scalar, const DiagonalBase& other)
     {
-      return other.diagonal() * scalar;
+      return ScalarMultipleReturnType(other.diagonal() * scalar);
     }
-    
-    #ifdef EIGEN2_SUPPORT
-    template<typename OtherDerived>
-    bool isApprox(const DiagonalBase<OtherDerived>& other, typename NumTraits<Scalar>::Real precision = NumTraits<Scalar>::dummy_precision()) const
-    {
-      return diagonal().isApprox(other.diagonal(), precision);
-    }
-    template<typename OtherDerived>
-    bool isApprox(const MatrixBase<OtherDerived>& other, typename NumTraits<Scalar>::Real precision = NumTraits<Scalar>::dummy_precision()) const
-    {
-      return toDenseMatrix().isApprox(other, precision);
-    }
-    #endif
 };
 
-template<typename Derived>
-template<typename DenseDerived>
-void DiagonalBase<Derived>::evalTo(MatrixBase<DenseDerived> &other) const
-{
-  other.setZero();
-  other.diagonal() = diagonal();
-}
 #endif
 
 /** \class DiagonalMatrix
@@ -124,10 +107,9 @@ struct traits<DiagonalMatrix<_Scalar,SizeAtCompileTime,MaxSizeAtCompileTime> >
  : traits<Matrix<_Scalar,SizeAtCompileTime,SizeAtCompileTime,0,MaxSizeAtCompileTime,MaxSizeAtCompileTime> >
 {
   typedef Matrix<_Scalar,SizeAtCompileTime,1,0,MaxSizeAtCompileTime,1> DiagonalVectorType;
-  typedef Dense StorageKind;
-  typedef DenseIndex Index;
+  typedef DiagonalShape StorageKind;
   enum {
-    Flags = LvalueBit
+    Flags = LvalueBit | NoPreferredStorageOrderBit
   };
 };
 }
@@ -141,7 +123,7 @@ class DiagonalMatrix
     typedef const DiagonalMatrix& Nested;
     typedef _Scalar Scalar;
     typedef typename internal::traits<DiagonalMatrix>::StorageKind StorageKind;
-    typedef typename internal::traits<DiagonalMatrix>::Index Index;
+    typedef typename internal::traits<DiagonalMatrix>::StorageIndex StorageIndex;
     #endif
 
   protected:
@@ -151,24 +133,31 @@ class DiagonalMatrix
   public:
 
     /** const version of diagonal(). */
+    EIGEN_DEVICE_FUNC
     inline const DiagonalVectorType& diagonal() const { return m_diagonal; }
     /** \returns a reference to the stored vector of diagonal coefficients. */
+    EIGEN_DEVICE_FUNC
     inline DiagonalVectorType& diagonal() { return m_diagonal; }
 
     /** Default constructor without initialization */
+    EIGEN_DEVICE_FUNC
     inline DiagonalMatrix() {}
 
     /** Constructs a diagonal matrix with given dimension  */
-    inline DiagonalMatrix(Index dim) : m_diagonal(dim) {}
+    EIGEN_DEVICE_FUNC
+    explicit inline DiagonalMatrix(Index dim) : m_diagonal(dim) {}
 
     /** 2D constructor. */
+    EIGEN_DEVICE_FUNC
     inline DiagonalMatrix(const Scalar& x, const Scalar& y) : m_diagonal(x,y) {}
 
     /** 3D constructor. */
+    EIGEN_DEVICE_FUNC
     inline DiagonalMatrix(const Scalar& x, const Scalar& y, const Scalar& z) : m_diagonal(x,y,z) {}
 
     /** Copy constructor. */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     inline DiagonalMatrix(const DiagonalBase<OtherDerived>& other) : m_diagonal(other.diagonal()) {}
 
     #ifndef EIGEN_PARSED_BY_DOXYGEN
@@ -178,11 +167,13 @@ class DiagonalMatrix
 
     /** generic constructor from expression of the diagonal coefficients */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     explicit inline DiagonalMatrix(const MatrixBase<OtherDerived>& other) : m_diagonal(other)
     {}
 
     /** Copy operator. */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     DiagonalMatrix& operator=(const DiagonalBase<OtherDerived>& other)
     {
       m_diagonal = other.diagonal();
@@ -193,6 +184,7 @@ class DiagonalMatrix
     /** This is a special case of the templated operator=. Its purpose is to
       * prevent a default operator= from hiding the templated operator=.
       */
+    EIGEN_DEVICE_FUNC
     DiagonalMatrix& operator=(const DiagonalMatrix& other)
     {
       m_diagonal = other.diagonal();
@@ -201,14 +193,19 @@ class DiagonalMatrix
     #endif
 
     /** Resizes to given size. */
+    EIGEN_DEVICE_FUNC
     inline void resize(Index size) { m_diagonal.resize(size); }
     /** Sets all coefficients to zero. */
+    EIGEN_DEVICE_FUNC
     inline void setZero() { m_diagonal.setZero(); }
     /** Resizes and sets all coefficients to zero. */
+    EIGEN_DEVICE_FUNC
     inline void setZero(Index size) { m_diagonal.setZero(size); }
     /** Sets this matrix to be the identity matrix of the current size. */
+    EIGEN_DEVICE_FUNC
     inline void setIdentity() { m_diagonal.setOnes(); }
     /** Sets this matrix to be the identity matrix of the given size. */
+    EIGEN_DEVICE_FUNC
     inline void setIdentity(Index size) { m_diagonal.setOnes(size); }
 };
 
@@ -232,14 +229,15 @@ struct traits<DiagonalWrapper<_DiagonalVectorType> >
 {
   typedef _DiagonalVectorType DiagonalVectorType;
   typedef typename DiagonalVectorType::Scalar Scalar;
-  typedef typename DiagonalVectorType::Index Index;
-  typedef typename DiagonalVectorType::StorageKind StorageKind;
+  typedef typename DiagonalVectorType::StorageIndex StorageIndex;
+  typedef DiagonalShape StorageKind;
+  typedef typename traits<DiagonalVectorType>::XprKind XprKind;
   enum {
     RowsAtCompileTime = DiagonalVectorType::SizeAtCompileTime,
     ColsAtCompileTime = DiagonalVectorType::SizeAtCompileTime,
-    MaxRowsAtCompileTime = DiagonalVectorType::SizeAtCompileTime,
-    MaxColsAtCompileTime = DiagonalVectorType::SizeAtCompileTime,
-    Flags =  traits<DiagonalVectorType>::Flags & LvalueBit
+    MaxRowsAtCompileTime = DiagonalVectorType::MaxSizeAtCompileTime,
+    MaxColsAtCompileTime = DiagonalVectorType::MaxSizeAtCompileTime,
+    Flags =  (traits<DiagonalVectorType>::Flags & LvalueBit) | NoPreferredStorageOrderBit
   };
 };
 }
@@ -255,9 +253,11 @@ class DiagonalWrapper
     #endif
 
     /** Constructor from expression of diagonal coefficients to wrap. */
-    inline DiagonalWrapper(DiagonalVectorType& a_diagonal) : m_diagonal(a_diagonal) {}
+    EIGEN_DEVICE_FUNC
+    explicit inline DiagonalWrapper(DiagonalVectorType& a_diagonal) : m_diagonal(a_diagonal) {}
 
     /** \returns a const reference to the wrapped expression of diagonal coefficients. */
+    EIGEN_DEVICE_FUNC
     const DiagonalVectorType& diagonal() const { return m_diagonal; }
 
   protected:
@@ -277,7 +277,7 @@ template<typename Derived>
 inline const DiagonalWrapper<const Derived>
 MatrixBase<Derived>::asDiagonal() const
 {
-  return derived();
+  return DiagonalWrapper<const Derived>(derived());
 }
 
 /** \returns true if *this is approximately equal to a diagonal matrix,
@@ -308,6 +308,33 @@ bool MatrixBase<Derived>::isDiagonal(const RealScalar& prec) const
   return true;
 }
 
+namespace internal {
+
+template<> struct storage_kind_to_shape<DiagonalShape> { typedef DiagonalShape Shape; };
+
+struct Diagonal2Dense {};
+
+template<> struct AssignmentKind<DenseShape,DiagonalShape> { typedef Diagonal2Dense Kind; };
+
+// Diagonal matrix to Dense assignment
+template< typename DstXprType, typename SrcXprType, typename Functor, typename Scalar>
+struct Assignment<DstXprType, SrcXprType, Functor, Diagonal2Dense, Scalar>
+{
+  static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<typename DstXprType::Scalar> &/*func*/)
+  {
+    dst.setZero();
+    dst.diagonal() = src.diagonal();
+  }
+  
+  static void run(DstXprType &dst, const SrcXprType &src, const internal::add_assign_op<typename DstXprType::Scalar> &/*func*/)
+  { dst.diagonal() += src.diagonal(); }
+  
+  static void run(DstXprType &dst, const SrcXprType &src, const internal::sub_assign_op<typename DstXprType::Scalar> &/*func*/)
+  { dst.diagonal() -= src.diagonal(); }
+};
+
+} // namespace internal
+
 } // end namespace Eigen
 
 #endif // EIGEN_DIAGONALMATRIX_H

Eigen/src/Core/DiagonalProduct.h

@@ -13,116 +13,14 @@
 
 namespace Eigen { 
 
-namespace internal {
-template<typename MatrixType, typename DiagonalType, int ProductOrder>
-struct traits<DiagonalProduct<MatrixType, DiagonalType, ProductOrder> >
- : traits<MatrixType>
-{
-  typedef typename scalar_product_traits<typename MatrixType::Scalar, typename DiagonalType::Scalar>::ReturnType Scalar;
-  enum {
-    RowsAtCompileTime = MatrixType::RowsAtCompileTime,
-    ColsAtCompileTime = MatrixType::ColsAtCompileTime,
-    MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime,
-    MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime,
-
-    _StorageOrder = MatrixType::Flags & RowMajorBit ? RowMajor : ColMajor,
-    _ScalarAccessOnDiag =  !((int(_StorageOrder) == ColMajor && int(ProductOrder) == OnTheLeft)
-                          ||(int(_StorageOrder) == RowMajor && int(ProductOrder) == OnTheRight)),
-    _SameTypes = is_same<typename MatrixType::Scalar, typename DiagonalType::Scalar>::value,
-    // FIXME currently we need same types, but in the future the next rule should be the one
-    //_Vectorizable = bool(int(MatrixType::Flags)&PacketAccessBit) && ((!_PacketOnDiag) || (_SameTypes && bool(int(DiagonalType::DiagonalVectorType::Flags)&PacketAccessBit))),
-    _Vectorizable = bool(int(MatrixType::Flags)&PacketAccessBit) && _SameTypes && (_ScalarAccessOnDiag || (bool(int(DiagonalType::DiagonalVectorType::Flags)&PacketAccessBit))),
-    _LinearAccessMask = (RowsAtCompileTime==1 || ColsAtCompileTime==1) ? LinearAccessBit : 0,
-
-    Flags = ((HereditaryBits|_LinearAccessMask) & (unsigned int)(MatrixType::Flags)) | (_Vectorizable ? PacketAccessBit : 0) | AlignedBit,//(int(MatrixType::Flags)&int(DiagonalType::DiagonalVectorType::Flags)&AlignedBit),
-    CoeffReadCost = NumTraits<Scalar>::MulCost + MatrixType::CoeffReadCost + DiagonalType::DiagonalVectorType::CoeffReadCost
-  };
-};
-}
-
-template<typename MatrixType, typename DiagonalType, int ProductOrder>
-class DiagonalProduct : internal::no_assignment_operator,
-                        public MatrixBase<DiagonalProduct<MatrixType, DiagonalType, ProductOrder> >
-{
-  public:
-
-    typedef MatrixBase<DiagonalProduct> Base;
-    EIGEN_DENSE_PUBLIC_INTERFACE(DiagonalProduct)
-
-    inline DiagonalProduct(const MatrixType& matrix, const DiagonalType& diagonal)
-      : m_matrix(matrix), m_diagonal(diagonal)
-    {
-      eigen_assert(diagonal.diagonal().size() == (ProductOrder == OnTheLeft ? matrix.rows() : matrix.cols()));
-    }
-
-    EIGEN_STRONG_INLINE Index rows() const { return m_matrix.rows(); }
-    EIGEN_STRONG_INLINE Index cols() const { return m_matrix.cols(); }
-
-    EIGEN_STRONG_INLINE const Scalar coeff(Index row, Index col) const
-    {
-      return m_diagonal.diagonal().coeff(ProductOrder == OnTheLeft ? row : col) * m_matrix.coeff(row, col);
-    }
-    
-    EIGEN_STRONG_INLINE const Scalar coeff(Index idx) const
-    {
-      enum {
-        StorageOrder = int(MatrixType::Flags) & RowMajorBit ? RowMajor : ColMajor
-      };
-      return coeff(int(StorageOrder)==ColMajor?idx:0,int(StorageOrder)==ColMajor?0:idx);
-    }
-
-    template<int LoadMode>
-    EIGEN_STRONG_INLINE PacketScalar packet(Index row, Index col) const
-    {
-      enum {
-        StorageOrder = Flags & RowMajorBit ? RowMajor : ColMajor
-      };
-      const Index indexInDiagonalVector = ProductOrder == OnTheLeft ? row : col;
-      return packet_impl<LoadMode>(row,col,indexInDiagonalVector,typename internal::conditional<
-        ((int(StorageOrder) == RowMajor && int(ProductOrder) == OnTheLeft)
-       ||(int(StorageOrder) == ColMajor && int(ProductOrder) == OnTheRight)), internal::true_type, internal::false_type>::type());
-    }
-    
-    template<int LoadMode>
-    EIGEN_STRONG_INLINE PacketScalar packet(Index idx) const
-    {
-      enum {
-        StorageOrder = int(MatrixType::Flags) & RowMajorBit ? RowMajor : ColMajor
-      };
-      return packet<LoadMode>(int(StorageOrder)==ColMajor?idx:0,int(StorageOrder)==ColMajor?0:idx);
-    }
-
-  protected:
-    template<int LoadMode>
-    EIGEN_STRONG_INLINE PacketScalar packet_impl(Index row, Index col, Index id, internal::true_type) const
-    {
-      return internal::pmul(m_matrix.template packet<LoadMode>(row, col),
-                     internal::pset1<PacketScalar>(m_diagonal.diagonal().coeff(id)));
-    }
-
-    template<int LoadMode>
-    EIGEN_STRONG_INLINE PacketScalar packet_impl(Index row, Index col, Index id, internal::false_type) const
-    {
-      enum {
-        InnerSize = (MatrixType::Flags & RowMajorBit) ? MatrixType::ColsAtCompileTime : MatrixType::RowsAtCompileTime,
-        DiagonalVectorPacketLoadMode = (LoadMode == Aligned && (((InnerSize%16) == 0) || (int(DiagonalType::DiagonalVectorType::Flags)&AlignedBit)==AlignedBit) ? Aligned : Unaligned)
-      };
-      return internal::pmul(m_matrix.template packet<LoadMode>(row, col),
-                     m_diagonal.diagonal().template packet<DiagonalVectorPacketLoadMode>(id));
-    }
-
-    typename MatrixType::Nested m_matrix;
-    typename DiagonalType::Nested m_diagonal;
-};
-
 /** \returns the diagonal matrix product of \c *this by the diagonal matrix \a diagonal.
   */
 template<typename Derived>
 template<typename DiagonalDerived>
-inline const DiagonalProduct<Derived, DiagonalDerived, OnTheRight>
+inline const Product<Derived, DiagonalDerived, LazyProduct>
 MatrixBase<Derived>::operator*(const DiagonalBase<DiagonalDerived> &a_diagonal) const
 {
-  return DiagonalProduct<Derived, DiagonalDerived, OnTheRight>(derived(), a_diagonal.derived());
+  return Product<Derived, DiagonalDerived, LazyProduct>(derived(),a_diagonal.derived());
 }
 
 } // end namespace Eigen

Eigen/src/Core/Dot.h

@@ -29,6 +29,7 @@ template<typename T, typename U,
 struct dot_nocheck
 {
   typedef typename scalar_product_traits<typename traits<T>::Scalar,typename traits<U>::Scalar>::ReturnType ResScalar;
+  EIGEN_DEVICE_FUNC
   static inline ResScalar run(const MatrixBase<T>& a, const MatrixBase<U>& b)
   {
     return a.template binaryExpr<scalar_conj_product_op<typename traits<T>::Scalar,typename traits<U>::Scalar> >(b).sum();
@@ -39,6 +40,7 @@ template<typename T, typename U>
 struct dot_nocheck<T, U, true>
 {
   typedef typename scalar_product_traits<typename traits<T>::Scalar,typename traits<U>::Scalar>::ReturnType ResScalar;
+  EIGEN_DEVICE_FUNC
   static inline ResScalar run(const MatrixBase<T>& a, const MatrixBase<U>& b)
   {
     return a.transpose().template binaryExpr<scalar_conj_product_op<typename traits<T>::Scalar,typename traits<U>::Scalar> >(b).sum();
@@ -59,6 +61,7 @@ struct dot_nocheck<T, U, true>
   */
 template<typename Derived>
 template<typename OtherDerived>
+EIGEN_DEVICE_FUNC
 typename internal::scalar_product_traits<typename internal::traits<Derived>::Scalar,typename internal::traits<OtherDerived>::Scalar>::ReturnType
 MatrixBase<Derived>::dot(const MatrixBase<OtherDerived>& other) const
 {
@@ -73,34 +76,6 @@ MatrixBase<Derived>::dot(const MatrixBase<OtherDerived>& other) const
   return internal::dot_nocheck<Derived,OtherDerived>::run(*this, other);
 }
 
-#ifdef EIGEN2_SUPPORT
-/** \returns the dot product of *this with other, with the Eigen2 convention that the dot product is linear in the first variable
-  * (conjugating the second variable). Of course this only makes a difference in the complex case.
-  *
-  * This method is only available in EIGEN2_SUPPORT mode.
-  *
-  * \only_for_vectors
-  *
-  * \sa dot()
-  */
-template<typename Derived>
-template<typename OtherDerived>
-typename internal::traits<Derived>::Scalar
-MatrixBase<Derived>::eigen2_dot(const MatrixBase<OtherDerived>& other) const
-{
-  EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
-  EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
-  EIGEN_STATIC_ASSERT_SAME_VECTOR_SIZE(Derived,OtherDerived)
-  EIGEN_STATIC_ASSERT((internal::is_same<Scalar, typename OtherDerived::Scalar>::value),
-    YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY)
-
-  eigen_assert(size() == other.size());
-
-  return internal::dot_nocheck<OtherDerived,Derived>::run(other,*this);
-}
-#endif
-
-
 //---------- implementation of L2 norm and related functions ----------
 
 /** \returns, for vectors, the squared \em l2 norm of \c *this, and for matrices the Frobenius norm.
@@ -124,7 +99,7 @@ EIGEN_STRONG_INLINE typename NumTraits<typename internal::traits<Derived>::Scala
 template<typename Derived>
 inline typename NumTraits<typename internal::traits<Derived>::Scalar>::Real MatrixBase<Derived>::norm() const
 {
-  using std::sqrt;
+  EIGEN_USING_STD_MATH(sqrt)
   return sqrt(squaredNorm());
 }
 
@@ -138,8 +113,7 @@ template<typename Derived>
 inline const typename MatrixBase<Derived>::PlainObject
 MatrixBase<Derived>::normalized() const
 {
-  typedef typename internal::nested<Derived>::type Nested;
-  typedef typename internal::remove_reference<Nested>::type _Nested;
+  typedef typename internal::nested_eval<Derived,2>::type _Nested;
   _Nested n(derived());
   return n / n.norm();
 }
@@ -164,9 +138,10 @@ template<typename Derived, int p>
 struct lpNorm_selector
 {
   typedef typename NumTraits<typename traits<Derived>::Scalar>::Real RealScalar;
+  EIGEN_DEVICE_FUNC
   static inline RealScalar run(const MatrixBase<Derived>& m)
   {
-    using std::pow;
+    EIGEN_USING_STD_MATH(pow)
     return pow(m.cwiseAbs().array().pow(p).sum(), RealScalar(1)/p);
   }
 };
@@ -174,6 +149,7 @@ struct lpNorm_selector
 template<typename Derived>
 struct lpNorm_selector<Derived, 1>
 {
+  EIGEN_DEVICE_FUNC
   static inline typename NumTraits<typename traits<Derived>::Scalar>::Real run(const MatrixBase<Derived>& m)
   {
     return m.cwiseAbs().sum();
@@ -183,6 +159,7 @@ struct lpNorm_selector<Derived, 1>
 template<typename Derived>
 struct lpNorm_selector<Derived, 2>
 {
+  EIGEN_DEVICE_FUNC
   static inline typename NumTraits<typename traits<Derived>::Scalar>::Real run(const MatrixBase<Derived>& m)
   {
     return m.norm();
@@ -192,6 +169,7 @@ struct lpNorm_selector<Derived, 2>
 template<typename Derived>
 struct lpNorm_selector<Derived, Infinity>
 {
+  EIGEN_DEVICE_FUNC
   static inline typename NumTraits<typename traits<Derived>::Scalar>::Real run(const MatrixBase<Derived>& m)
   {
     return m.cwiseAbs().maxCoeff();
@@ -227,8 +205,8 @@ template<typename OtherDerived>
 bool MatrixBase<Derived>::isOrthogonal
 (const MatrixBase<OtherDerived>& other, const RealScalar& prec) const
 {
-  typename internal::nested<Derived,2>::type nested(derived());
-  typename internal::nested<OtherDerived,2>::type otherNested(other.derived());
+  typename internal::nested_eval<Derived,2>::type nested(derived());
+  typename internal::nested_eval<OtherDerived,2>::type otherNested(other.derived());
   return numext::abs2(nested.dot(otherNested)) <= prec * prec * nested.squaredNorm() * otherNested.squaredNorm();
 }
 
@@ -246,13 +224,13 @@ bool MatrixBase<Derived>::isOrthogonal
 template<typename Derived>
 bool MatrixBase<Derived>::isUnitary(const RealScalar& prec) const
 {
-  typename Derived::Nested nested(derived());
+  typename internal::nested_eval<Derived,1>::type self(derived());
   for(Index i = 0; i < cols(); ++i)
   {
-    if(!internal::isApprox(nested.col(i).squaredNorm(), static_cast<RealScalar>(1), prec))
+    if(!internal::isApprox(self.col(i).squaredNorm(), static_cast<RealScalar>(1), prec))
       return false;
     for(Index j = 0; j < i; ++j)
-      if(!internal::isMuchSmallerThan(nested.col(i).dot(nested.col(j)), static_cast<Scalar>(1), prec))
+      if(!internal::isMuchSmallerThan(self.col(i).dot(self.col(j)), static_cast<Scalar>(1), prec))
         return false;
   }
   return true;

Eigen/src/Core/EigenBase.h

@@ -13,7 +13,9 @@
 
 namespace Eigen {
 
-/** Common base class for all classes T such that MatrixBase has an operator=(T) and a constructor MatrixBase(T).
+/** \class EigenBase
+  * 
+  * Common base class for all classes T such that MatrixBase has an operator=(T) and a constructor MatrixBase(T).
   *
   * In other words, an EigenBase object is an object that can be copied into a MatrixBase.
   *
@@ -26,34 +28,52 @@ namespace Eigen {
 template<typename Derived> struct EigenBase
 {
 //   typedef typename internal::plain_matrix_type<Derived>::type PlainObject;
+  
+  /** \brief The interface type of indices
+    * \details To change this, \c \#define the preprocessor symbol \c EIGEN_DEFAULT_DENSE_INDEX_TYPE.
+    * \deprecated Since Eigen 3.3, its usage is deprecated. Use Eigen::Index instead.
+    * \sa StorageIndex, \ref TopicPreprocessorDirectives.
+    */
+  typedef Eigen::Index Index;
 
+  // FIXME is it needed?
   typedef typename internal::traits<Derived>::StorageKind StorageKind;
-  typedef typename internal::traits<Derived>::Index Index;
 
   /** \returns a reference to the derived object */
+  EIGEN_DEVICE_FUNC
   Derived& derived() { return *static_cast<Derived*>(this); }
   /** \returns a const reference to the derived object */
+  EIGEN_DEVICE_FUNC
   const Derived& derived() const { return *static_cast<const Derived*>(this); }
 
+  EIGEN_DEVICE_FUNC
   inline Derived& const_cast_derived() const
   { return *static_cast<Derived*>(const_cast<EigenBase*>(this)); }
+  EIGEN_DEVICE_FUNC
   inline const Derived& const_derived() const
   { return *static_cast<const Derived*>(this); }
 
   /** \returns the number of rows. \sa cols(), RowsAtCompileTime */
+  EIGEN_DEVICE_FUNC
   inline Index rows() const { return derived().rows(); }
   /** \returns the number of columns. \sa rows(), ColsAtCompileTime*/
+  EIGEN_DEVICE_FUNC
   inline Index cols() const { return derived().cols(); }
   /** \returns the number of coefficients, which is rows()*cols().
     * \sa rows(), cols(), SizeAtCompileTime. */
+  EIGEN_DEVICE_FUNC
   inline Index size() const { return rows() * cols(); }
 
   /** \internal Don't use it, but do the equivalent: \code dst = *this; \endcode */
-  template<typename Dest> inline void evalTo(Dest& dst) const
+  template<typename Dest>
+  EIGEN_DEVICE_FUNC
+  inline void evalTo(Dest& dst) const
   { derived().evalTo(dst); }
 
   /** \internal Don't use it, but do the equivalent: \code dst += *this; \endcode */
-  template<typename Dest> inline void addTo(Dest& dst) const
+  template<typename Dest>
+  EIGEN_DEVICE_FUNC
+  inline void addTo(Dest& dst) const
   {
     // This is the default implementation,
     // derived class can reimplement it in a more optimized way.
@@ -63,7 +83,9 @@ template<typename Derived> struct EigenBase
   }
 
   /** \internal Don't use it, but do the equivalent: \code dst -= *this; \endcode */
-  template<typename Dest> inline void subTo(Dest& dst) const
+  template<typename Dest>
+  EIGEN_DEVICE_FUNC
+  inline void subTo(Dest& dst) const
   {
     // This is the default implementation,
     // derived class can reimplement it in a more optimized way.
@@ -73,7 +95,8 @@ template<typename Derived> struct EigenBase
   }
 
   /** \internal Don't use it, but do the equivalent: \code dst.applyOnTheRight(*this); \endcode */
-  template<typename Dest> inline void applyThisOnTheRight(Dest& dst) const
+  template<typename Dest>
+  EIGEN_DEVICE_FUNC inline void applyThisOnTheRight(Dest& dst) const
   {
     // This is the default implementation,
     // derived class can reimplement it in a more optimized way.
@@ -81,7 +104,8 @@ template<typename Derived> struct EigenBase
   }
 
   /** \internal Don't use it, but do the equivalent: \code dst.applyOnTheLeft(*this); \endcode */
-  template<typename Dest> inline void applyThisOnTheLeft(Dest& dst) const
+  template<typename Dest>
+  EIGEN_DEVICE_FUNC inline void applyThisOnTheLeft(Dest& dst) const
   {
     // This is the default implementation,
     // derived class can reimplement it in a more optimized way.
@@ -106,7 +130,7 @@ template<typename Derived>
 template<typename OtherDerived>
 Derived& DenseBase<Derived>::operator=(const EigenBase<OtherDerived> &other)
 {
-  other.derived().evalTo(derived());
+  call_assignment(derived(), other.derived());
   return derived();
 }
 
@@ -114,7 +138,7 @@ template<typename Derived>
 template<typename OtherDerived>
 Derived& DenseBase<Derived>::operator+=(const EigenBase<OtherDerived> &other)
 {
-  other.derived().addTo(derived());
+  call_assignment(derived(), other.derived(), internal::add_assign_op<Scalar>());
   return derived();
 }
 
@@ -122,40 +146,10 @@ template<typename Derived>
 template<typename OtherDerived>
 Derived& DenseBase<Derived>::operator-=(const EigenBase<OtherDerived> &other)
 {
-  other.derived().subTo(derived());
+  call_assignment(derived(), other.derived(), internal::sub_assign_op<Scalar>());
   return derived();
 }
 
-/** replaces \c *this by \c *this * \a other.
-  *
-  * \returns a reference to \c *this
-  */
-template<typename Derived>
-template<typename OtherDerived>
-inline Derived&
-MatrixBase<Derived>::operator*=(const EigenBase<OtherDerived> &other)
-{
-  other.derived().applyThisOnTheRight(derived());
-  return derived();
-}
-
-/** replaces \c *this by \c *this * \a other. It is equivalent to MatrixBase::operator*=().
-  */
-template<typename Derived>
-template<typename OtherDerived>
-inline void MatrixBase<Derived>::applyOnTheRight(const EigenBase<OtherDerived> &other)
-{
-  other.derived().applyThisOnTheRight(derived());
-}
-
-/** replaces \c *this by \c *this * \a other. */
-template<typename Derived>
-template<typename OtherDerived>
-inline void MatrixBase<Derived>::applyOnTheLeft(const EigenBase<OtherDerived> &other)
-{
-  other.derived().applyThisOnTheLeft(derived());
-}
-
 } // end namespace Eigen
 
 #endif // EIGEN_EIGENBASE_H

Eigen/src/Core/Flagged.h

@@ -1,140 +0,0 @@
-// This file is part of Eigen, a lightweight C++ template library
-// for linear algebra.
-//
-// Copyright (C) 2008 Benoit Jacob <jacob.benoit.1@gmail.com>
-//
-// This Source Code Form is subject to the terms of the Mozilla
-// Public License v. 2.0. If a copy of the MPL was not distributed
-// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
-
-#ifndef EIGEN_FLAGGED_H
-#define EIGEN_FLAGGED_H
-
-namespace Eigen { 
-
-/** \class Flagged
-  * \ingroup Core_Module
-  *
-  * \brief Expression with modified flags
-  *
-  * \param ExpressionType the type of the object of which we are modifying the flags
-  * \param Added the flags added to the expression
-  * \param Removed the flags removed from the expression (has priority over Added).
-  *
-  * This class represents an expression whose flags have been modified.
-  * It is the return type of MatrixBase::flagged()
-  * and most of the time this is the only way it is used.
-  *
-  * \sa MatrixBase::flagged()
-  */
-
-namespace internal {
-template<typename ExpressionType, unsigned int Added, unsigned int Removed>
-struct traits<Flagged<ExpressionType, Added, Removed> > : traits<ExpressionType>
-{
-  enum { Flags = (ExpressionType::Flags | Added) & ~Removed };
-};
-}
-
-template<typename ExpressionType, unsigned int Added, unsigned int Removed> class Flagged
-  : public MatrixBase<Flagged<ExpressionType, Added, Removed> >
-{
-  public:
-
-    typedef MatrixBase<Flagged> Base;
-    
-    EIGEN_DENSE_PUBLIC_INTERFACE(Flagged)
-    typedef typename internal::conditional<internal::must_nest_by_value<ExpressionType>::ret,
-        ExpressionType, const ExpressionType&>::type ExpressionTypeNested;
-    typedef typename ExpressionType::InnerIterator InnerIterator;
-
-    inline Flagged(const ExpressionType& matrix) : m_matrix(matrix) {}
-
-    inline Index rows() const { return m_matrix.rows(); }
-    inline Index cols() const { return m_matrix.cols(); }
-    inline Index outerStride() const { return m_matrix.outerStride(); }
-    inline Index innerStride() const { return m_matrix.innerStride(); }
-
-    inline CoeffReturnType coeff(Index row, Index col) const
-    {
-      return m_matrix.coeff(row, col);
-    }
-
-    inline CoeffReturnType coeff(Index index) const
-    {
-      return m_matrix.coeff(index);
-    }
-    
-    inline const Scalar& coeffRef(Index row, Index col) const
-    {
-      return m_matrix.const_cast_derived().coeffRef(row, col);
-    }
-
-    inline const Scalar& coeffRef(Index index) const
-    {
-      return m_matrix.const_cast_derived().coeffRef(index);
-    }
-
-    inline Scalar& coeffRef(Index row, Index col)
-    {
-      return m_matrix.const_cast_derived().coeffRef(row, col);
-    }
-
-    inline Scalar& coeffRef(Index index)
-    {
-      return m_matrix.const_cast_derived().coeffRef(index);
-    }
-
-    template<int LoadMode>
-    inline const PacketScalar packet(Index row, Index col) const
-    {
-      return m_matrix.template packet<LoadMode>(row, col);
-    }
-
-    template<int LoadMode>
-    inline void writePacket(Index row, Index col, const PacketScalar& x)
-    {
-      m_matrix.const_cast_derived().template writePacket<LoadMode>(row, col, x);
-    }
-
-    template<int LoadMode>
-    inline const PacketScalar packet(Index index) const
-    {
-      return m_matrix.template packet<LoadMode>(index);
-    }
-
-    template<int LoadMode>
-    inline void writePacket(Index index, const PacketScalar& x)
-    {
-      m_matrix.const_cast_derived().template writePacket<LoadMode>(index, x);
-    }
-
-    const ExpressionType& _expression() const { return m_matrix; }
-
-    template<typename OtherDerived>
-    typename ExpressionType::PlainObject solveTriangular(const MatrixBase<OtherDerived>& other) const;
-
-    template<typename OtherDerived>
-    void solveTriangularInPlace(const MatrixBase<OtherDerived>& other) const;
-
-  protected:
-    ExpressionTypeNested m_matrix;
-};
-
-/** \returns an expression of *this with added and removed flags
-  *
-  * This is mostly for internal use.
-  *
-  * \sa class Flagged
-  */
-template<typename Derived>
-template<unsigned int Added,unsigned int Removed>
-inline const Flagged<Derived, Added, Removed>
-DenseBase<Derived>::flagged() const
-{
-  return derived();
-}
-
-} // end namespace Eigen
-
-#endif // EIGEN_FLAGGED_H

Eigen/src/Core/ForceAlignedAccess.h

@@ -39,29 +39,29 @@ template<typename ExpressionType> class ForceAlignedAccess
     typedef typename internal::dense_xpr_base<ForceAlignedAccess>::type Base;
     EIGEN_DENSE_PUBLIC_INTERFACE(ForceAlignedAccess)
 
-    inline ForceAlignedAccess(const ExpressionType& matrix) : m_expression(matrix) {}
+    EIGEN_DEVICE_FUNC explicit inline ForceAlignedAccess(const ExpressionType& matrix) : m_expression(matrix) {}
 
-    inline Index rows() const { return m_expression.rows(); }
-    inline Index cols() const { return m_expression.cols(); }
-    inline Index outerStride() const { return m_expression.outerStride(); }
-    inline Index innerStride() const { return m_expression.innerStride(); }
+    EIGEN_DEVICE_FUNC inline Index rows() const { return m_expression.rows(); }
+    EIGEN_DEVICE_FUNC inline Index cols() const { return m_expression.cols(); }
+    EIGEN_DEVICE_FUNC inline Index outerStride() const { return m_expression.outerStride(); }
+    EIGEN_DEVICE_FUNC inline Index innerStride() const { return m_expression.innerStride(); }
 
-    inline const CoeffReturnType coeff(Index row, Index col) const
+    EIGEN_DEVICE_FUNC inline const CoeffReturnType coeff(Index row, Index col) const
     {
       return m_expression.coeff(row, col);
     }
 
-    inline Scalar& coeffRef(Index row, Index col)
+    EIGEN_DEVICE_FUNC inline Scalar& coeffRef(Index row, Index col)
     {
       return m_expression.const_cast_derived().coeffRef(row, col);
     }
 
-    inline const CoeffReturnType coeff(Index index) const
+    EIGEN_DEVICE_FUNC inline const CoeffReturnType coeff(Index index) const
     {
       return m_expression.coeff(index);
     }
 
-    inline Scalar& coeffRef(Index index)
+    EIGEN_DEVICE_FUNC inline Scalar& coeffRef(Index index)
     {
       return m_expression.const_cast_derived().coeffRef(index);
     }
@@ -90,7 +90,7 @@ template<typename ExpressionType> class ForceAlignedAccess
       m_expression.const_cast_derived().template writePacket<Aligned>(index, x);
     }
 
-    operator const ExpressionType&() const { return m_expression; }
+    EIGEN_DEVICE_FUNC operator const ExpressionType&() const { return m_expression; }
 
   protected:
     const ExpressionType& m_expression;
@@ -127,7 +127,7 @@ template<bool Enable>
 inline typename internal::add_const_on_value_type<typename internal::conditional<Enable,ForceAlignedAccess<Derived>,Derived&>::type>::type
 MatrixBase<Derived>::forceAlignedAccessIf() const
 {
-  return derived();
+  return derived();  // FIXME This should not work but apparently is never used
 }
 
 /** \returns an expression of *this with forced aligned access if \a Enable is true.
@@ -138,7 +138,7 @@ template<bool Enable>
 inline typename internal::conditional<Enable,ForceAlignedAccess<Derived>,Derived&>::type
 MatrixBase<Derived>::forceAlignedAccessIf()
 {
-  return derived();
+  return derived();  // FIXME This should not work but apparently is never used
 }
 
 } // end namespace Eigen

Eigen/src/Core/Functors.h

@@ -1,985 +0,0 @@
-// This file is part of Eigen, a lightweight C++ template library
-// for linear algebra.
-//
-// Copyright (C) 2008-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
-//
-// This Source Code Form is subject to the terms of the Mozilla
-// Public License v. 2.0. If a copy of the MPL was not distributed
-// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
-
-#ifndef EIGEN_FUNCTORS_H
-#define EIGEN_FUNCTORS_H
-
-namespace Eigen {
-
-namespace internal {
-
-// associative functors:
-
-/** \internal
-  * \brief Template functor to compute the sum of two scalars
-  *
-  * \sa class CwiseBinaryOp, MatrixBase::operator+, class VectorwiseOp, MatrixBase::sum()
-  */
-template<typename Scalar> struct scalar_sum_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_sum_op)
-  EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a, const Scalar& b) const { return a + b; }
-  template<typename Packet>
-  EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const
-  { return internal::padd(a,b); }
-  template<typename Packet>
-  EIGEN_STRONG_INLINE const Scalar predux(const Packet& a) const
-  { return internal::predux(a); }
-};
-template<typename Scalar>
-struct functor_traits<scalar_sum_op<Scalar> > {
-  enum {
-    Cost = NumTraits<Scalar>::AddCost,
-    PacketAccess = packet_traits<Scalar>::HasAdd
-  };
-};
-
-/** \internal
-  * \brief Template functor to compute the product of two scalars
-  *
-  * \sa class CwiseBinaryOp, Cwise::operator*(), class VectorwiseOp, MatrixBase::redux()
-  */
-template<typename LhsScalar,typename RhsScalar> struct scalar_product_op {
-  enum {
-    // TODO vectorize mixed product
-    Vectorizable = is_same<LhsScalar,RhsScalar>::value && packet_traits<LhsScalar>::HasMul && packet_traits<RhsScalar>::HasMul
-  };
-  typedef typename scalar_product_traits<LhsScalar,RhsScalar>::ReturnType result_type;
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_product_op)
-  EIGEN_STRONG_INLINE const result_type operator() (const LhsScalar& a, const RhsScalar& b) const { return a * b; }
-  template<typename Packet>
-  EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const
-  { return internal::pmul(a,b); }
-  template<typename Packet>
-  EIGEN_STRONG_INLINE const result_type predux(const Packet& a) const
-  { return internal::predux_mul(a); }
-};
-template<typename LhsScalar,typename RhsScalar>
-struct functor_traits<scalar_product_op<LhsScalar,RhsScalar> > {
-  enum {
-    Cost = (NumTraits<LhsScalar>::MulCost + NumTraits<RhsScalar>::MulCost)/2, // rough estimate!
-    PacketAccess = scalar_product_op<LhsScalar,RhsScalar>::Vectorizable
-  };
-};
-
-/** \internal
-  * \brief Template functor to compute the conjugate product of two scalars
-  *
-  * This is a short cut for conj(x) * y which is needed for optimization purpose; in Eigen2 support mode, this becomes x * conj(y)
-  */
-template<typename LhsScalar,typename RhsScalar> struct scalar_conj_product_op {
-
-  enum {
-    Conj = NumTraits<LhsScalar>::IsComplex
-  };
-  
-  typedef typename scalar_product_traits<LhsScalar,RhsScalar>::ReturnType result_type;
-  
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_conj_product_op)
-  EIGEN_STRONG_INLINE const result_type operator() (const LhsScalar& a, const RhsScalar& b) const
-  { return conj_helper<LhsScalar,RhsScalar,Conj,false>().pmul(a,b); }
-  
-  template<typename Packet>
-  EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const
-  { return conj_helper<Packet,Packet,Conj,false>().pmul(a,b); }
-};
-template<typename LhsScalar,typename RhsScalar>
-struct functor_traits<scalar_conj_product_op<LhsScalar,RhsScalar> > {
-  enum {
-    Cost = NumTraits<LhsScalar>::MulCost,
-    PacketAccess = internal::is_same<LhsScalar, RhsScalar>::value && packet_traits<LhsScalar>::HasMul
-  };
-};
-
-/** \internal
-  * \brief Template functor to compute the min of two scalars
-  *
-  * \sa class CwiseBinaryOp, MatrixBase::cwiseMin, class VectorwiseOp, MatrixBase::minCoeff()
-  */
-template<typename Scalar> struct scalar_min_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_min_op)
-  EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a, const Scalar& b) const { using std::min; return (min)(a, b); }
-  template<typename Packet>
-  EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const
-  { return internal::pmin(a,b); }
-  template<typename Packet>
-  EIGEN_STRONG_INLINE const Scalar predux(const Packet& a) const
-  { return internal::predux_min(a); }
-};
-template<typename Scalar>
-struct functor_traits<scalar_min_op<Scalar> > {
-  enum {
-    Cost = NumTraits<Scalar>::AddCost,
-    PacketAccess = packet_traits<Scalar>::HasMin
-  };
-};
-
-/** \internal
-  * \brief Template functor to compute the max of two scalars
-  *
-  * \sa class CwiseBinaryOp, MatrixBase::cwiseMax, class VectorwiseOp, MatrixBase::maxCoeff()
-  */
-template<typename Scalar> struct scalar_max_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_max_op)
-  EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a, const Scalar& b) const { using std::max; return (max)(a, b); }
-  template<typename Packet>
-  EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const
-  { return internal::pmax(a,b); }
-  template<typename Packet>
-  EIGEN_STRONG_INLINE const Scalar predux(const Packet& a) const
-  { return internal::predux_max(a); }
-};
-template<typename Scalar>
-struct functor_traits<scalar_max_op<Scalar> > {
-  enum {
-    Cost = NumTraits<Scalar>::AddCost,
-    PacketAccess = packet_traits<Scalar>::HasMax
-  };
-};
-
-/** \internal
-  * \brief Template functor to compute the hypot of two scalars
-  *
-  * \sa MatrixBase::stableNorm(), class Redux
-  */
-template<typename Scalar> struct scalar_hypot_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_hypot_op)
-//   typedef typename NumTraits<Scalar>::Real result_type;
-  EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& _x, const Scalar& _y) const
-  {
-    using std::max;
-    using std::min;
-    using std::sqrt;
-    Scalar p = (max)(_x, _y);
-    Scalar q = (min)(_x, _y);
-    Scalar qp = q/p;
-    return p * sqrt(Scalar(1) + qp*qp);
-  }
-};
-template<typename Scalar>
-struct functor_traits<scalar_hypot_op<Scalar> > {
-  enum { Cost = 5 * NumTraits<Scalar>::MulCost, PacketAccess=0 };
-};
-
-/** \internal
-  * \brief Template functor to compute the pow of two scalars
-  */
-template<typename Scalar, typename OtherScalar> struct scalar_binary_pow_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_binary_pow_op)
-  inline Scalar operator() (const Scalar& a, const OtherScalar& b) const { return numext::pow(a, b); }
-};
-template<typename Scalar, typename OtherScalar>
-struct functor_traits<scalar_binary_pow_op<Scalar,OtherScalar> > {
-  enum { Cost = 5 * NumTraits<Scalar>::MulCost, PacketAccess = false };
-};
-
-// other binary functors:
-
-/** \internal
-  * \brief Template functor to compute the difference of two scalars
-  *
-  * \sa class CwiseBinaryOp, MatrixBase::operator-
-  */
-template<typename Scalar> struct scalar_difference_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_difference_op)
-  EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a, const Scalar& b) const { return a - b; }
-  template<typename Packet>
-  EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const
-  { return internal::psub(a,b); }
-};
-template<typename Scalar>
-struct functor_traits<scalar_difference_op<Scalar> > {
-  enum {
-    Cost = NumTraits<Scalar>::AddCost,
-    PacketAccess = packet_traits<Scalar>::HasSub
-  };
-};
-
-/** \internal
-  * \brief Template functor to compute the quotient of two scalars
-  *
-  * \sa class CwiseBinaryOp, Cwise::operator/()
-  */
-template<typename LhsScalar,typename RhsScalar> struct scalar_quotient_op {
-  enum {
-    // TODO vectorize mixed product
-    Vectorizable = is_same<LhsScalar,RhsScalar>::value && packet_traits<LhsScalar>::HasDiv && packet_traits<RhsScalar>::HasDiv
-  };
-  typedef typename scalar_product_traits<LhsScalar,RhsScalar>::ReturnType result_type;
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_quotient_op)
-  EIGEN_STRONG_INLINE const result_type operator() (const LhsScalar& a, const RhsScalar& b) const { return a / b; }
-  template<typename Packet>
-  EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const
-  { return internal::pdiv(a,b); }
-};
-template<typename LhsScalar,typename RhsScalar>
-struct functor_traits<scalar_quotient_op<LhsScalar,RhsScalar> > {
-  enum {
-    Cost = (NumTraits<LhsScalar>::MulCost + NumTraits<RhsScalar>::MulCost), // rough estimate!
-    PacketAccess = scalar_quotient_op<LhsScalar,RhsScalar>::Vectorizable
-  };
-};
-
-
-
-/** \internal
-  * \brief Template functor to compute the and of two booleans
-  *
-  * \sa class CwiseBinaryOp, ArrayBase::operator&&
-  */
-struct scalar_boolean_and_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_boolean_and_op)
-  EIGEN_STRONG_INLINE bool operator() (const bool& a, const bool& b) const { return a && b; }
-};
-template<> struct functor_traits<scalar_boolean_and_op> {
-  enum {
-    Cost = NumTraits<bool>::AddCost,
-    PacketAccess = false
-  };
-};
-
-/** \internal
-  * \brief Template functor to compute the or of two booleans
-  *
-  * \sa class CwiseBinaryOp, ArrayBase::operator||
-  */
-struct scalar_boolean_or_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_boolean_or_op)
-  EIGEN_STRONG_INLINE bool operator() (const bool& a, const bool& b) const { return a || b; }
-};
-template<> struct functor_traits<scalar_boolean_or_op> {
-  enum {
-    Cost = NumTraits<bool>::AddCost,
-    PacketAccess = false
-  };
-};
-
-// unary functors:
-
-/** \internal
-  * \brief Template functor to compute the opposite of a scalar
-  *
-  * \sa class CwiseUnaryOp, MatrixBase::operator-
-  */
-template<typename Scalar> struct scalar_opposite_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_opposite_op)
-  EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a) const { return -a; }
-  template<typename Packet>
-  EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const
-  { return internal::pnegate(a); }
-};
-template<typename Scalar>
-struct functor_traits<scalar_opposite_op<Scalar> >
-{ enum {
-    Cost = NumTraits<Scalar>::AddCost,
-    PacketAccess = packet_traits<Scalar>::HasNegate };
-};
-
-/** \internal
-  * \brief Template functor to compute the absolute value of a scalar
-  *
-  * \sa class CwiseUnaryOp, Cwise::abs
-  */
-template<typename Scalar> struct scalar_abs_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_abs_op)
-  typedef typename NumTraits<Scalar>::Real result_type;
-  EIGEN_STRONG_INLINE const result_type operator() (const Scalar& a) const { using std::abs; return abs(a); }
-  template<typename Packet>
-  EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const
-  { return internal::pabs(a); }
-};
-template<typename Scalar>
-struct functor_traits<scalar_abs_op<Scalar> >
-{
-  enum {
-    Cost = NumTraits<Scalar>::AddCost,
-    PacketAccess = packet_traits<Scalar>::HasAbs
-  };
-};
-
-/** \internal
-  * \brief Template functor to compute the squared absolute value of a scalar
-  *
-  * \sa class CwiseUnaryOp, Cwise::abs2
-  */
-template<typename Scalar> struct scalar_abs2_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_abs2_op)
-  typedef typename NumTraits<Scalar>::Real result_type;
-  EIGEN_STRONG_INLINE const result_type operator() (const Scalar& a) const { return numext::abs2(a); }
-  template<typename Packet>
-  EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const
-  { return internal::pmul(a,a); }
-};
-template<typename Scalar>
-struct functor_traits<scalar_abs2_op<Scalar> >
-{ enum { Cost = NumTraits<Scalar>::MulCost, PacketAccess = packet_traits<Scalar>::HasAbs2 }; };
-
-/** \internal
-  * \brief Template functor to compute the conjugate of a complex value
-  *
-  * \sa class CwiseUnaryOp, MatrixBase::conjugate()
-  */
-template<typename Scalar> struct scalar_conjugate_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_conjugate_op)
-  EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a) const { using numext::conj; return conj(a); }
-  template<typename Packet>
-  EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const { return internal::pconj(a); }
-};
-template<typename Scalar>
-struct functor_traits<scalar_conjugate_op<Scalar> >
-{
-  enum {
-    Cost = NumTraits<Scalar>::IsComplex ? NumTraits<Scalar>::AddCost : 0,
-    PacketAccess = packet_traits<Scalar>::HasConj
-  };
-};
-
-/** \internal
-  * \brief Template functor to cast a scalar to another type
-  *
-  * \sa class CwiseUnaryOp, MatrixBase::cast()
-  */
-template<typename Scalar, typename NewType>
-struct scalar_cast_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_cast_op)
-  typedef NewType result_type;
-  EIGEN_STRONG_INLINE const NewType operator() (const Scalar& a) const { return cast<Scalar, NewType>(a); }
-};
-template<typename Scalar, typename NewType>
-struct functor_traits<scalar_cast_op<Scalar,NewType> >
-{ enum { Cost = is_same<Scalar, NewType>::value ? 0 : NumTraits<NewType>::AddCost, PacketAccess = false }; };
-
-/** \internal
-  * \brief Template functor to extract the real part of a complex
-  *
-  * \sa class CwiseUnaryOp, MatrixBase::real()
-  */
-template<typename Scalar>
-struct scalar_real_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_real_op)
-  typedef typename NumTraits<Scalar>::Real result_type;
-  EIGEN_STRONG_INLINE result_type operator() (const Scalar& a) const { return numext::real(a); }
-};
-template<typename Scalar>
-struct functor_traits<scalar_real_op<Scalar> >
-{ enum { Cost = 0, PacketAccess = false }; };
-
-/** \internal
-  * \brief Template functor to extract the imaginary part of a complex
-  *
-  * \sa class CwiseUnaryOp, MatrixBase::imag()
-  */
-template<typename Scalar>
-struct scalar_imag_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_imag_op)
-  typedef typename NumTraits<Scalar>::Real result_type;
-  EIGEN_STRONG_INLINE result_type operator() (const Scalar& a) const { return numext::imag(a); }
-};
-template<typename Scalar>
-struct functor_traits<scalar_imag_op<Scalar> >
-{ enum { Cost = 0, PacketAccess = false }; };
-
-/** \internal
-  * \brief Template functor to extract the real part of a complex as a reference
-  *
-  * \sa class CwiseUnaryOp, MatrixBase::real()
-  */
-template<typename Scalar>
-struct scalar_real_ref_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_real_ref_op)
-  typedef typename NumTraits<Scalar>::Real result_type;
-  EIGEN_STRONG_INLINE result_type& operator() (const Scalar& a) const { return numext::real_ref(*const_cast<Scalar*>(&a)); }
-};
-template<typename Scalar>
-struct functor_traits<scalar_real_ref_op<Scalar> >
-{ enum { Cost = 0, PacketAccess = false }; };
-
-/** \internal
-  * \brief Template functor to extract the imaginary part of a complex as a reference
-  *
-  * \sa class CwiseUnaryOp, MatrixBase::imag()
-  */
-template<typename Scalar>
-struct scalar_imag_ref_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_imag_ref_op)
-  typedef typename NumTraits<Scalar>::Real result_type;
-  EIGEN_STRONG_INLINE result_type& operator() (const Scalar& a) const { return numext::imag_ref(*const_cast<Scalar*>(&a)); }
-};
-template<typename Scalar>
-struct functor_traits<scalar_imag_ref_op<Scalar> >
-{ enum { Cost = 0, PacketAccess = false }; };
-
-/** \internal
-  *
-  * \brief Template functor to compute the exponential of a scalar
-  *
-  * \sa class CwiseUnaryOp, Cwise::exp()
-  */
-template<typename Scalar> struct scalar_exp_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_exp_op)
-  inline const Scalar operator() (const Scalar& a) const { using std::exp; return exp(a); }
-  typedef typename packet_traits<Scalar>::type Packet;
-  inline Packet packetOp(const Packet& a) const { return internal::pexp(a); }
-};
-template<typename Scalar>
-struct functor_traits<scalar_exp_op<Scalar> >
-{ enum { Cost = 5 * NumTraits<Scalar>::MulCost, PacketAccess = packet_traits<Scalar>::HasExp }; };
-
-/** \internal
-  *
-  * \brief Template functor to compute the logarithm of a scalar
-  *
-  * \sa class CwiseUnaryOp, Cwise::log()
-  */
-template<typename Scalar> struct scalar_log_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_log_op)
-  inline const Scalar operator() (const Scalar& a) const { using std::log; return log(a); }
-  typedef typename packet_traits<Scalar>::type Packet;
-  inline Packet packetOp(const Packet& a) const { return internal::plog(a); }
-};
-template<typename Scalar>
-struct functor_traits<scalar_log_op<Scalar> >
-{ enum { Cost = 5 * NumTraits<Scalar>::MulCost, PacketAccess = packet_traits<Scalar>::HasLog }; };
-
-/** \internal
-  * \brief Template functor to multiply a scalar by a fixed other one
-  *
-  * \sa class CwiseUnaryOp, MatrixBase::operator*, MatrixBase::operator/
-  */
-/* NOTE why doing the pset1() in packetOp *is* an optimization ?
- * indeed it seems better to declare m_other as a Packet and do the pset1() once
- * in the constructor. However, in practice:
- *  - GCC does not like m_other as a Packet and generate a load every time it needs it
- *  - on the other hand GCC is able to moves the pset1() outside the loop :)
- *  - simpler code ;)
- * (ICC and gcc 4.4 seems to perform well in both cases, the issue is visible with y = a*x + b*y)
- */
-template<typename Scalar>
-struct scalar_multiple_op {
-  typedef typename packet_traits<Scalar>::type Packet;
-  // FIXME default copy constructors seems bugged with std::complex<>
-  EIGEN_STRONG_INLINE scalar_multiple_op(const scalar_multiple_op& other) : m_other(other.m_other) { }
-  EIGEN_STRONG_INLINE scalar_multiple_op(const Scalar& other) : m_other(other) { }
-  EIGEN_STRONG_INLINE Scalar operator() (const Scalar& a) const { return a * m_other; }
-  EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const
-  { return internal::pmul(a, pset1<Packet>(m_other)); }
-  typename add_const_on_value_type<typename NumTraits<Scalar>::Nested>::type m_other;
-};
-template<typename Scalar>
-struct functor_traits<scalar_multiple_op<Scalar> >
-{ enum { Cost = NumTraits<Scalar>::MulCost, PacketAccess = packet_traits<Scalar>::HasMul }; };
-
-template<typename Scalar1, typename Scalar2>
-struct scalar_multiple2_op {
-  typedef typename scalar_product_traits<Scalar1,Scalar2>::ReturnType result_type;
-  EIGEN_STRONG_INLINE scalar_multiple2_op(const scalar_multiple2_op& other) : m_other(other.m_other) { }
-  EIGEN_STRONG_INLINE scalar_multiple2_op(const Scalar2& other) : m_other(other) { }
-  EIGEN_STRONG_INLINE result_type operator() (const Scalar1& a) const { return a * m_other; }
-  typename add_const_on_value_type<typename NumTraits<Scalar2>::Nested>::type m_other;
-};
-template<typename Scalar1,typename Scalar2>
-struct functor_traits<scalar_multiple2_op<Scalar1,Scalar2> >
-{ enum { Cost = NumTraits<Scalar1>::MulCost, PacketAccess = false }; };
-
-/** \internal
-  * \brief Template functor to divide a scalar by a fixed other one
-  *
-  * This functor is used to implement the quotient of a matrix by
-  * a scalar where the scalar type is not necessarily a floating point type.
-  *
-  * \sa class CwiseUnaryOp, MatrixBase::operator/
-  */
-template<typename Scalar>
-struct scalar_quotient1_op {
-  typedef typename packet_traits<Scalar>::type Packet;
-  // FIXME default copy constructors seems bugged with std::complex<>
-  EIGEN_STRONG_INLINE scalar_quotient1_op(const scalar_quotient1_op& other) : m_other(other.m_other) { }
-  EIGEN_STRONG_INLINE scalar_quotient1_op(const Scalar& other) : m_other(other) {}
-  EIGEN_STRONG_INLINE Scalar operator() (const Scalar& a) const { return a / m_other; }
-  EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const
-  { return internal::pdiv(a, pset1<Packet>(m_other)); }
-  typename add_const_on_value_type<typename NumTraits<Scalar>::Nested>::type m_other;
-};
-template<typename Scalar>
-struct functor_traits<scalar_quotient1_op<Scalar> >
-{ enum { Cost = 2 * NumTraits<Scalar>::MulCost, PacketAccess = packet_traits<Scalar>::HasDiv }; };
-
-// nullary functors
-
-template<typename Scalar>
-struct scalar_constant_op {
-  typedef typename packet_traits<Scalar>::type Packet;
-  EIGEN_STRONG_INLINE scalar_constant_op(const scalar_constant_op& other) : m_other(other.m_other) { }
-  EIGEN_STRONG_INLINE scalar_constant_op(const Scalar& other) : m_other(other) { }
-  template<typename Index>
-  EIGEN_STRONG_INLINE const Scalar operator() (Index, Index = 0) const { return m_other; }
-  template<typename Index>
-  EIGEN_STRONG_INLINE const Packet packetOp(Index, Index = 0) const { return internal::pset1<Packet>(m_other); }
-  const Scalar m_other;
-};
-template<typename Scalar>
-struct functor_traits<scalar_constant_op<Scalar> >
-// FIXME replace this packet test by a safe one
-{ enum { Cost = 1, PacketAccess = packet_traits<Scalar>::Vectorizable, IsRepeatable = true }; };
-
-template<typename Scalar> struct scalar_identity_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_identity_op)
-  template<typename Index>
-  EIGEN_STRONG_INLINE const Scalar operator() (Index row, Index col) const { return row==col ? Scalar(1) : Scalar(0); }
-};
-template<typename Scalar>
-struct functor_traits<scalar_identity_op<Scalar> >
-{ enum { Cost = NumTraits<Scalar>::AddCost, PacketAccess = false, IsRepeatable = true }; };
-
-template <typename Scalar, bool RandomAccess> struct linspaced_op_impl;
-
-// linear access for packet ops:
-// 1) initialization
-//   base = [low, ..., low] + ([step, ..., step] * [-size, ..., 0])
-// 2) each step (where size is 1 for coeff access or PacketSize for packet access)
-//   base += [size*step, ..., size*step]
-//
-// TODO: Perhaps it's better to initialize lazily (so not in the constructor but in packetOp)
-//       in order to avoid the padd() in operator() ?
-template <typename Scalar>
-struct linspaced_op_impl<Scalar,false>
-{
-  typedef typename packet_traits<Scalar>::type Packet;
-
-  linspaced_op_impl(const Scalar& low, const Scalar& step) :
-  m_low(low), m_step(step),
-  m_packetStep(pset1<Packet>(packet_traits<Scalar>::size*step)),
-  m_base(padd(pset1<Packet>(low), pmul(pset1<Packet>(step),plset<Scalar>(-packet_traits<Scalar>::size)))) {}
-
-  template<typename Index>
-  EIGEN_STRONG_INLINE const Scalar operator() (Index i) const 
-  { 
-    m_base = padd(m_base, pset1<Packet>(m_step));
-    return m_low+Scalar(i)*m_step; 
-  }
-
-  template<typename Index>
-  EIGEN_STRONG_INLINE const Packet packetOp(Index) const { return m_base = padd(m_base,m_packetStep); }
-
-  const Scalar m_low;
-  const Scalar m_step;
-  const Packet m_packetStep;
-  mutable Packet m_base;
-};
-
-// random access for packet ops:
-// 1) each step
-//   [low, ..., low] + ( [step, ..., step] * ( [i, ..., i] + [0, ..., size] ) )
-template <typename Scalar>
-struct linspaced_op_impl<Scalar,true>
-{
-  typedef typename packet_traits<Scalar>::type Packet;
-
-  linspaced_op_impl(const Scalar& low, const Scalar& step) :
-  m_low(low), m_step(step),
-  m_lowPacket(pset1<Packet>(m_low)), m_stepPacket(pset1<Packet>(m_step)), m_interPacket(plset<Scalar>(0)) {}
-
-  template<typename Index>
-  EIGEN_STRONG_INLINE const Scalar operator() (Index i) const { return m_low+i*m_step; }
-
-  template<typename Index>
-  EIGEN_STRONG_INLINE const Packet packetOp(Index i) const
-  { return internal::padd(m_lowPacket, pmul(m_stepPacket, padd(pset1<Packet>(i),m_interPacket))); }
-
-  const Scalar m_low;
-  const Scalar m_step;
-  const Packet m_lowPacket;
-  const Packet m_stepPacket;
-  const Packet m_interPacket;
-};
-
-// ----- Linspace functor ----------------------------------------------------------------
-
-// Forward declaration (we default to random access which does not really give
-// us a speed gain when using packet access but it allows to use the functor in
-// nested expressions).
-template <typename Scalar, bool RandomAccess = true> struct linspaced_op;
-template <typename Scalar, bool RandomAccess> struct functor_traits< linspaced_op<Scalar,RandomAccess> >
-{ enum { Cost = 1, PacketAccess = packet_traits<Scalar>::HasSetLinear, IsRepeatable = true }; };
-template <typename Scalar, bool RandomAccess> struct linspaced_op
-{
-  typedef typename packet_traits<Scalar>::type Packet;
-  linspaced_op(const Scalar& low, const Scalar& high, DenseIndex num_steps) : impl((num_steps==1 ? high : low), (num_steps==1 ? Scalar() : (high-low)/(num_steps-1))) {}
-
-  template<typename Index>
-  EIGEN_STRONG_INLINE const Scalar operator() (Index i) const { return impl(i); }
-
-  // We need this function when assigning e.g. a RowVectorXd to a MatrixXd since
-  // there row==0 and col is used for the actual iteration.
-  template<typename Index>
-  EIGEN_STRONG_INLINE const Scalar operator() (Index row, Index col) const 
-  {
-    eigen_assert(col==0 || row==0);
-    return impl(col + row);
-  }
-
-  template<typename Index>
-  EIGEN_STRONG_INLINE const Packet packetOp(Index i) const { return impl.packetOp(i); }
-
-  // We need this function when assigning e.g. a RowVectorXd to a MatrixXd since
-  // there row==0 and col is used for the actual iteration.
-  template<typename Index>
-  EIGEN_STRONG_INLINE const Packet packetOp(Index row, Index col) const
-  {
-    eigen_assert(col==0 || row==0);
-    return impl.packetOp(col + row);
-  }
-
-  // This proxy object handles the actual required temporaries, the different
-  // implementations (random vs. sequential access) as well as the
-  // correct piping to size 2/4 packet operations.
-  const linspaced_op_impl<Scalar,RandomAccess> impl;
-};
-
-// all functors allow linear access, except scalar_identity_op. So we fix here a quick meta
-// to indicate whether a functor allows linear access, just always answering 'yes' except for
-// scalar_identity_op.
-// FIXME move this to functor_traits adding a functor_default
-template<typename Functor> struct functor_has_linear_access { enum { ret = 1 }; };
-template<typename Scalar> struct functor_has_linear_access<scalar_identity_op<Scalar> > { enum { ret = 0 }; };
-
-// In Eigen, any binary op (Product, CwiseBinaryOp) require the Lhs and Rhs to have the same scalar type, except for multiplication
-// where the mixing of different types is handled by scalar_product_traits
-// In particular, real * complex<real> is allowed.
-// FIXME move this to functor_traits adding a functor_default
-template<typename Functor> struct functor_is_product_like { enum { ret = 0 }; };
-template<typename LhsScalar,typename RhsScalar> struct functor_is_product_like<scalar_product_op<LhsScalar,RhsScalar> > { enum { ret = 1 }; };
-template<typename LhsScalar,typename RhsScalar> struct functor_is_product_like<scalar_conj_product_op<LhsScalar,RhsScalar> > { enum { ret = 1 }; };
-template<typename LhsScalar,typename RhsScalar> struct functor_is_product_like<scalar_quotient_op<LhsScalar,RhsScalar> > { enum { ret = 1 }; };
-
-
-/** \internal
-  * \brief Template functor to add a scalar to a fixed other one
-  * \sa class CwiseUnaryOp, Array::operator+
-  */
-/* If you wonder why doing the pset1() in packetOp() is an optimization check scalar_multiple_op */
-template<typename Scalar>
-struct scalar_add_op {
-  typedef typename packet_traits<Scalar>::type Packet;
-  // FIXME default copy constructors seems bugged with std::complex<>
-  inline scalar_add_op(const scalar_add_op& other) : m_other(other.m_other) { }
-  inline scalar_add_op(const Scalar& other) : m_other(other) { }
-  inline Scalar operator() (const Scalar& a) const { return a + m_other; }
-  inline const Packet packetOp(const Packet& a) const
-  { return internal::padd(a, pset1<Packet>(m_other)); }
-  const Scalar m_other;
-};
-template<typename Scalar>
-struct functor_traits<scalar_add_op<Scalar> >
-{ enum { Cost = NumTraits<Scalar>::AddCost, PacketAccess = packet_traits<Scalar>::HasAdd }; };
-
-/** \internal
-  * \brief Template functor to compute the square root of a scalar
-  * \sa class CwiseUnaryOp, Cwise::sqrt()
-  */
-template<typename Scalar> struct scalar_sqrt_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_sqrt_op)
-  inline const Scalar operator() (const Scalar& a) const { using std::sqrt; return sqrt(a); }
-  typedef typename packet_traits<Scalar>::type Packet;
-  inline Packet packetOp(const Packet& a) const { return internal::psqrt(a); }
-};
-template<typename Scalar>
-struct functor_traits<scalar_sqrt_op<Scalar> >
-{ enum {
-    Cost = 5 * NumTraits<Scalar>::MulCost,
-    PacketAccess = packet_traits<Scalar>::HasSqrt
-  };
-};
-
-/** \internal
-  * \brief Template functor to compute the cosine of a scalar
-  * \sa class CwiseUnaryOp, ArrayBase::cos()
-  */
-template<typename Scalar> struct scalar_cos_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_cos_op)
-  inline Scalar operator() (const Scalar& a) const { using std::cos; return cos(a); }
-  typedef typename packet_traits<Scalar>::type Packet;
-  inline Packet packetOp(const Packet& a) const { return internal::pcos(a); }
-};
-template<typename Scalar>
-struct functor_traits<scalar_cos_op<Scalar> >
-{
-  enum {
-    Cost = 5 * NumTraits<Scalar>::MulCost,
-    PacketAccess = packet_traits<Scalar>::HasCos
-  };
-};
-
-/** \internal
-  * \brief Template functor to compute the sine of a scalar
-  * \sa class CwiseUnaryOp, ArrayBase::sin()
-  */
-template<typename Scalar> struct scalar_sin_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_sin_op)
-  inline const Scalar operator() (const Scalar& a) const { using std::sin; return sin(a); }
-  typedef typename packet_traits<Scalar>::type Packet;
-  inline Packet packetOp(const Packet& a) const { return internal::psin(a); }
-};
-template<typename Scalar>
-struct functor_traits<scalar_sin_op<Scalar> >
-{
-  enum {
-    Cost = 5 * NumTraits<Scalar>::MulCost,
-    PacketAccess = packet_traits<Scalar>::HasSin
-  };
-};
-
-
-/** \internal
-  * \brief Template functor to compute the tan of a scalar
-  * \sa class CwiseUnaryOp, ArrayBase::tan()
-  */
-template<typename Scalar> struct scalar_tan_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_tan_op)
-  inline const Scalar operator() (const Scalar& a) const { using std::tan; return tan(a); }
-  typedef typename packet_traits<Scalar>::type Packet;
-  inline Packet packetOp(const Packet& a) const { return internal::ptan(a); }
-};
-template<typename Scalar>
-struct functor_traits<scalar_tan_op<Scalar> >
-{
-  enum {
-    Cost = 5 * NumTraits<Scalar>::MulCost,
-    PacketAccess = packet_traits<Scalar>::HasTan
-  };
-};
-
-/** \internal
-  * \brief Template functor to compute the arc cosine of a scalar
-  * \sa class CwiseUnaryOp, ArrayBase::acos()
-  */
-template<typename Scalar> struct scalar_acos_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_acos_op)
-  inline const Scalar operator() (const Scalar& a) const { using std::acos; return acos(a); }
-  typedef typename packet_traits<Scalar>::type Packet;
-  inline Packet packetOp(const Packet& a) const { return internal::pacos(a); }
-};
-template<typename Scalar>
-struct functor_traits<scalar_acos_op<Scalar> >
-{
-  enum {
-    Cost = 5 * NumTraits<Scalar>::MulCost,
-    PacketAccess = packet_traits<Scalar>::HasACos
-  };
-};
-
-/** \internal
-  * \brief Template functor to compute the arc sine of a scalar
-  * \sa class CwiseUnaryOp, ArrayBase::asin()
-  */
-template<typename Scalar> struct scalar_asin_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_asin_op)
-  inline const Scalar operator() (const Scalar& a) const { using std::asin; return asin(a); }
-  typedef typename packet_traits<Scalar>::type Packet;
-  inline Packet packetOp(const Packet& a) const { return internal::pasin(a); }
-};
-template<typename Scalar>
-struct functor_traits<scalar_asin_op<Scalar> >
-{
-  enum {
-    Cost = 5 * NumTraits<Scalar>::MulCost,
-    PacketAccess = packet_traits<Scalar>::HasASin
-  };
-};
-
-/** \internal
-  * \brief Template functor to raise a scalar to a power
-  * \sa class CwiseUnaryOp, Cwise::pow
-  */
-template<typename Scalar>
-struct scalar_pow_op {
-  // FIXME default copy constructors seems bugged with std::complex<>
-  inline scalar_pow_op(const scalar_pow_op& other) : m_exponent(other.m_exponent) { }
-  inline scalar_pow_op(const Scalar& exponent) : m_exponent(exponent) {}
-  inline Scalar operator() (const Scalar& a) const { return numext::pow(a, m_exponent); }
-  const Scalar m_exponent;
-};
-template<typename Scalar>
-struct functor_traits<scalar_pow_op<Scalar> >
-{ enum { Cost = 5 * NumTraits<Scalar>::MulCost, PacketAccess = false }; };
-
-/** \internal
-  * \brief Template functor to compute the quotient between a scalar and array entries.
-  * \sa class CwiseUnaryOp, Cwise::inverse()
-  */
-template<typename Scalar>
-struct scalar_inverse_mult_op {
-  scalar_inverse_mult_op(const Scalar& other) : m_other(other) {}
-  inline Scalar operator() (const Scalar& a) const { return m_other / a; }
-  template<typename Packet>
-  inline const Packet packetOp(const Packet& a) const
-  { return internal::pdiv(pset1<Packet>(m_other),a); }
-  Scalar m_other;
-};
-
-/** \internal
-  * \brief Template functor to compute the inverse of a scalar
-  * \sa class CwiseUnaryOp, Cwise::inverse()
-  */
-template<typename Scalar>
-struct scalar_inverse_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_inverse_op)
-  inline Scalar operator() (const Scalar& a) const { return Scalar(1)/a; }
-  template<typename Packet>
-  inline const Packet packetOp(const Packet& a) const
-  { return internal::pdiv(pset1<Packet>(Scalar(1)),a); }
-};
-template<typename Scalar>
-struct functor_traits<scalar_inverse_op<Scalar> >
-{ enum { Cost = NumTraits<Scalar>::MulCost, PacketAccess = packet_traits<Scalar>::HasDiv }; };
-
-/** \internal
-  * \brief Template functor to compute the square of a scalar
-  * \sa class CwiseUnaryOp, Cwise::square()
-  */
-template<typename Scalar>
-struct scalar_square_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_square_op)
-  inline Scalar operator() (const Scalar& a) const { return a*a; }
-  template<typename Packet>
-  inline const Packet packetOp(const Packet& a) const
-  { return internal::pmul(a,a); }
-};
-template<typename Scalar>
-struct functor_traits<scalar_square_op<Scalar> >
-{ enum { Cost = NumTraits<Scalar>::MulCost, PacketAccess = packet_traits<Scalar>::HasMul }; };
-
-/** \internal
-  * \brief Template functor to compute the cube of a scalar
-  * \sa class CwiseUnaryOp, Cwise::cube()
-  */
-template<typename Scalar>
-struct scalar_cube_op {
-  EIGEN_EMPTY_STRUCT_CTOR(scalar_cube_op)
-  inline Scalar operator() (const Scalar& a) const { return a*a*a; }
-  template<typename Packet>
-  inline const Packet packetOp(const Packet& a) const
-  { return internal::pmul(a,pmul(a,a)); }
-};
-template<typename Scalar>
-struct functor_traits<scalar_cube_op<Scalar> >
-{ enum { Cost = 2*NumTraits<Scalar>::MulCost, PacketAccess = packet_traits<Scalar>::HasMul }; };
-
-// default functor traits for STL functors:
-
-template<typename T>
-struct functor_traits<std::multiplies<T> >
-{ enum { Cost = NumTraits<T>::MulCost, PacketAccess = false }; };
-
-template<typename T>
-struct functor_traits<std::divides<T> >
-{ enum { Cost = NumTraits<T>::MulCost, PacketAccess = false }; };
-
-template<typename T>
-struct functor_traits<std::plus<T> >
-{ enum { Cost = NumTraits<T>::AddCost, PacketAccess = false }; };
-
-template<typename T>
-struct functor_traits<std::minus<T> >
-{ enum { Cost = NumTraits<T>::AddCost, PacketAccess = false }; };
-
-template<typename T>
-struct functor_traits<std::negate<T> >
-{ enum { Cost = NumTraits<T>::AddCost, PacketAccess = false }; };
-
-template<typename T>
-struct functor_traits<std::logical_or<T> >
-{ enum { Cost = 1, PacketAccess = false }; };
-
-template<typename T>
-struct functor_traits<std::logical_and<T> >
-{ enum { Cost = 1, PacketAccess = false }; };
-
-template<typename T>
-struct functor_traits<std::logical_not<T> >
-{ enum { Cost = 1, PacketAccess = false }; };
-
-template<typename T>
-struct functor_traits<std::greater<T> >
-{ enum { Cost = 1, PacketAccess = false }; };
-
-template<typename T>
-struct functor_traits<std::less<T> >
-{ enum { Cost = 1, PacketAccess = false }; };
-
-template<typename T>
-struct functor_traits<std::greater_equal<T> >
-{ enum { Cost = 1, PacketAccess = false }; };
-
-template<typename T>
-struct functor_traits<std::less_equal<T> >
-{ enum { Cost = 1, PacketAccess = false }; };
-
-template<typename T>
-struct functor_traits<std::equal_to<T> >
-{ enum { Cost = 1, PacketAccess = false }; };
-
-template<typename T>
-struct functor_traits<std::not_equal_to<T> >
-{ enum { Cost = 1, PacketAccess = false }; };
-
-template<typename T>
-struct functor_traits<std::binder2nd<T> >
-{ enum { Cost = functor_traits<T>::Cost, PacketAccess = false }; };
-
-template<typename T>
-struct functor_traits<std::binder1st<T> >
-{ enum { Cost = functor_traits<T>::Cost, PacketAccess = false }; };
-
-template<typename T>
-struct functor_traits<std::unary_negate<T> >
-{ enum { Cost = 1 + functor_traits<T>::Cost, PacketAccess = false }; };
-
-template<typename T>
-struct functor_traits<std::binary_negate<T> >
-{ enum { Cost = 1 + functor_traits<T>::Cost, PacketAccess = false }; };
-
-#ifdef EIGEN_STDEXT_SUPPORT
-
-template<typename T0,typename T1>
-struct functor_traits<std::project1st<T0,T1> >
-{ enum { Cost = 0, PacketAccess = false }; };
-
-template<typename T0,typename T1>
-struct functor_traits<std::project2nd<T0,T1> >
-{ enum { Cost = 0, PacketAccess = false }; };
-
-template<typename T0,typename T1>
-struct functor_traits<std::select2nd<std::pair<T0,T1> > >
-{ enum { Cost = 0, PacketAccess = false }; };
-
-template<typename T0,typename T1>
-struct functor_traits<std::select1st<std::pair<T0,T1> > >
-{ enum { Cost = 0, PacketAccess = false }; };
-
-template<typename T0,typename T1>
-struct functor_traits<std::unary_compose<T0,T1> >
-{ enum { Cost = functor_traits<T0>::Cost + functor_traits<T1>::Cost, PacketAccess = false }; };
-
-template<typename T0,typename T1,typename T2>
-struct functor_traits<std::binary_compose<T0,T1,T2> >
-{ enum { Cost = functor_traits<T0>::Cost + functor_traits<T1>::Cost + functor_traits<T2>::Cost, PacketAccess = false }; };
-
-#endif // EIGEN_STDEXT_SUPPORT
-
-// allow to add new functors and specializations of functor_traits from outside Eigen.
-// this macro is really needed because functor_traits must be specialized after it is declared but before it is used...
-#ifdef EIGEN_FUNCTORS_PLUGIN
-#include EIGEN_FUNCTORS_PLUGIN
-#endif
-
-} // end namespace internal
-
-} // end namespace Eigen
-
-#endif // EIGEN_FUNCTORS_H

Eigen/src/Core/Fuzzy.h

@@ -19,18 +19,19 @@ namespace internal
 template<typename Derived, typename OtherDerived, bool is_integer = NumTraits<typename Derived::Scalar>::IsInteger>
 struct isApprox_selector
 {
+  EIGEN_DEVICE_FUNC
   static bool run(const Derived& x, const OtherDerived& y, const typename Derived::RealScalar& prec)
   {
-    using std::min;
-    typename internal::nested<Derived,2>::type nested(x);
-    typename internal::nested<OtherDerived,2>::type otherNested(y);
-    return (nested - otherNested).cwiseAbs2().sum() <= prec * prec * (min)(nested.cwiseAbs2().sum(), otherNested.cwiseAbs2().sum());
+    typename internal::nested_eval<Derived,2>::type nested(x);
+    typename internal::nested_eval<OtherDerived,2>::type otherNested(y);
+    return (nested - otherNested).cwiseAbs2().sum() <= prec * prec * numext::mini(nested.cwiseAbs2().sum(), otherNested.cwiseAbs2().sum());
   }
 };
 
 template<typename Derived, typename OtherDerived>
 struct isApprox_selector<Derived, OtherDerived, true>
 {
+  EIGEN_DEVICE_FUNC
   static bool run(const Derived& x, const OtherDerived& y, const typename Derived::RealScalar&)
   {
     return x.matrix() == y.matrix();
@@ -40,6 +41,7 @@ struct isApprox_selector<Derived, OtherDerived, true>
 template<typename Derived, typename OtherDerived, bool is_integer = NumTraits<typename Derived::Scalar>::IsInteger>
 struct isMuchSmallerThan_object_selector
 {
+  EIGEN_DEVICE_FUNC
   static bool run(const Derived& x, const OtherDerived& y, const typename Derived::RealScalar& prec)
   {
     return x.cwiseAbs2().sum() <= numext::abs2(prec) * y.cwiseAbs2().sum();
@@ -49,6 +51,7 @@ struct isMuchSmallerThan_object_selector
 template<typename Derived, typename OtherDerived>
 struct isMuchSmallerThan_object_selector<Derived, OtherDerived, true>
 {
+  EIGEN_DEVICE_FUNC
   static bool run(const Derived& x, const OtherDerived&, const typename Derived::RealScalar&)
   {
     return x.matrix() == Derived::Zero(x.rows(), x.cols()).matrix();
@@ -58,6 +61,7 @@ struct isMuchSmallerThan_object_selector<Derived, OtherDerived, true>
 template<typename Derived, bool is_integer = NumTraits<typename Derived::Scalar>::IsInteger>
 struct isMuchSmallerThan_scalar_selector
 {
+  EIGEN_DEVICE_FUNC
   static bool run(const Derived& x, const typename Derived::RealScalar& y, const typename Derived::RealScalar& prec)
   {
     return x.cwiseAbs2().sum() <= numext::abs2(prec * y);
@@ -67,6 +71,7 @@ struct isMuchSmallerThan_scalar_selector
 template<typename Derived>
 struct isMuchSmallerThan_scalar_selector<Derived, true>
 {
+  EIGEN_DEVICE_FUNC
   static bool run(const Derived& x, const typename Derived::RealScalar&, const typename Derived::RealScalar&)
   {
     return x.matrix() == Derived::Zero(x.rows(), x.cols()).matrix();

Eigen/src/Core/GeneralProduct.h

@@ -11,29 +11,7 @@
 #ifndef EIGEN_GENERAL_PRODUCT_H
 #define EIGEN_GENERAL_PRODUCT_H
 
-namespace Eigen { 
-
-/** \class GeneralProduct
-  * \ingroup Core_Module
-  *
-  * \brief Expression of the product of two general matrices or vectors
-  *
-  * \param LhsNested the type used to store the left-hand side
-  * \param RhsNested the type used to store the right-hand side
-  * \param ProductMode the type of the product
-  *
-  * This class represents an expression of the product of two general matrices.
-  * We call a general matrix, a dense matrix with full storage. For instance,
-  * This excludes triangular, selfadjoint, and sparse matrices.
-  * It is the return type of the operator* between general matrices. Its template
-  * arguments are determined automatically by ProductReturnType. Therefore,
-  * GeneralProduct should never be used direclty. To determine the result type of a
-  * function which involves a matrix product, use ProductReturnType::Type.
-  *
-  * \sa ProductReturnType, MatrixBase::operator*(const MatrixBase<OtherDerived>&)
-  */
-template<typename Lhs, typename Rhs, int ProductType = internal::product_type<Lhs,Rhs>::value>
-class GeneralProduct;
+namespace Eigen {
 
 enum {
   Large = 2,
@@ -59,15 +37,14 @@ template<typename Lhs, typename Rhs> struct product_type
   typedef typename remove_all<Lhs>::type _Lhs;
   typedef typename remove_all<Rhs>::type _Rhs;
   enum {
-    MaxRows  = _Lhs::MaxRowsAtCompileTime,
-    Rows  = _Lhs::RowsAtCompileTime,
-    MaxCols  = _Rhs::MaxColsAtCompileTime,
-    Cols  = _Rhs::ColsAtCompileTime,
-    MaxDepth = EIGEN_SIZE_MIN_PREFER_FIXED(_Lhs::MaxColsAtCompileTime,
-                                           _Rhs::MaxRowsAtCompileTime),
-    Depth = EIGEN_SIZE_MIN_PREFER_FIXED(_Lhs::ColsAtCompileTime,
-                                        _Rhs::RowsAtCompileTime),
-    LargeThreshold = EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD
+    MaxRows = traits<_Lhs>::MaxRowsAtCompileTime,
+    Rows    = traits<_Lhs>::RowsAtCompileTime,
+    MaxCols = traits<_Rhs>::MaxColsAtCompileTime,
+    Cols    = traits<_Rhs>::ColsAtCompileTime,
+    MaxDepth = EIGEN_SIZE_MIN_PREFER_FIXED(traits<_Lhs>::MaxColsAtCompileTime,
+                                           traits<_Rhs>::MaxRowsAtCompileTime),
+    Depth = EIGEN_SIZE_MIN_PREFER_FIXED(traits<_Lhs>::ColsAtCompileTime,
+                                        traits<_Rhs>::RowsAtCompileTime)
   };
 
   // the splitting into different lines of code here, introducing the _select enums and the typedef below,
@@ -82,7 +59,8 @@ private:
 
 public:
   enum {
-    value = selector::ret
+    value = selector::ret,
+    ret = selector::ret
   };
 #ifdef EIGEN_DEBUG_PRODUCT
   static void debug()
@@ -98,6 +76,31 @@ public:
 #endif
 };
 
+// template<typename Lhs, typename Rhs> struct product_tag
+// {
+// private:
+//   
+//   typedef typename remove_all<Lhs>::type _Lhs;
+//   typedef typename remove_all<Rhs>::type _Rhs;
+//   enum {
+//     Rows  = _Lhs::RowsAtCompileTime,
+//     Cols  = _Rhs::ColsAtCompileTime,
+//     Depth = EIGEN_SIZE_MIN_PREFER_FIXED(_Lhs::ColsAtCompileTime, _Rhs::RowsAtCompileTime)
+//   };
+// 
+//   enum {
+//     rows_select = Rows==1 ? int(Rows) : int(Large),
+//     cols_select = Cols==1 ? int(Cols) : int(Large),
+//     depth_select = Depth==1 ? int(Depth) : int(Large)
+//   };
+//   typedef product_type_selector<rows_select, cols_select, depth_select> selector;
+// 
+// public:
+//   enum {
+//     ret = selector::ret
+//   };
+// 
+// };
 
 /* The following allows to select the kind of product at compile time
  * based on the three dimensions of the product.
@@ -128,54 +131,6 @@ template<>              struct product_type_selector<Large,Large,Small>  { enum
 
 } // end namespace internal
 
-/** \class ProductReturnType
-  * \ingroup Core_Module
-  *
-  * \brief Helper class to get the correct and optimized returned type of operator*
-  *
-  * \param Lhs the type of the left-hand side
-  * \param Rhs the type of the right-hand side
-  * \param ProductMode the type of the product (determined automatically by internal::product_mode)
-  *
-  * This class defines the typename Type representing the optimized product expression
-  * between two matrix expressions. In practice, using ProductReturnType<Lhs,Rhs>::Type
-  * is the recommended way to define the result type of a function returning an expression
-  * which involve a matrix product. The class Product should never be
-  * used directly.
-  *
-  * \sa class Product, MatrixBase::operator*(const MatrixBase<OtherDerived>&)
-  */
-template<typename Lhs, typename Rhs, int ProductType>
-struct ProductReturnType
-{
-  // TODO use the nested type to reduce instanciations ????
-//   typedef typename internal::nested<Lhs,Rhs::ColsAtCompileTime>::type LhsNested;
-//   typedef typename internal::nested<Rhs,Lhs::RowsAtCompileTime>::type RhsNested;
-
-  typedef GeneralProduct<Lhs/*Nested*/, Rhs/*Nested*/, ProductType> Type;
-};
-
-template<typename Lhs, typename Rhs>
-struct ProductReturnType<Lhs,Rhs,CoeffBasedProductMode>
-{
-  typedef typename internal::nested<Lhs, Rhs::ColsAtCompileTime, typename internal::plain_matrix_type<Lhs>::type >::type LhsNested;
-  typedef typename internal::nested<Rhs, Lhs::RowsAtCompileTime, typename internal::plain_matrix_type<Rhs>::type >::type RhsNested;
-  typedef CoeffBasedProduct<LhsNested, RhsNested, EvalBeforeAssigningBit | EvalBeforeNestingBit> Type;
-};
-
-template<typename Lhs, typename Rhs>
-struct ProductReturnType<Lhs,Rhs,LazyCoeffBasedProductMode>
-{
-  typedef typename internal::nested<Lhs, Rhs::ColsAtCompileTime, typename internal::plain_matrix_type<Lhs>::type >::type LhsNested;
-  typedef typename internal::nested<Rhs, Lhs::RowsAtCompileTime, typename internal::plain_matrix_type<Rhs>::type >::type RhsNested;
-  typedef CoeffBasedProduct<LhsNested, RhsNested, NestByRefBit> Type;
-};
-
-// this is a workaround for sun CC
-template<typename Lhs, typename Rhs>
-struct LazyProductReturnType : public ProductReturnType<Lhs,Rhs,LazyCoeffBasedProductMode>
-{};
-
 /***********************************************************************
 *  Implementation of Inner Vector Vector Product
 ***********************************************************************/
@@ -187,119 +142,10 @@ struct LazyProductReturnType : public ProductReturnType<Lhs,Rhs,LazyCoeffBasedPr
 // product ends up to a row-vector times col-vector product... To tackle this use
 // case, we could have a specialization for Block<MatrixType,1,1> with: operator=(Scalar x);
 
-namespace internal {
-
-template<typename Lhs, typename Rhs>
-struct traits<GeneralProduct<Lhs,Rhs,InnerProduct> >
- : traits<Matrix<typename scalar_product_traits<typename Lhs::Scalar, typename Rhs::Scalar>::ReturnType,1,1> >
-{};
-
-}
-
-template<typename Lhs, typename Rhs>
-class GeneralProduct<Lhs, Rhs, InnerProduct>
-  : internal::no_assignment_operator,
-    public Matrix<typename internal::scalar_product_traits<typename Lhs::Scalar, typename Rhs::Scalar>::ReturnType,1,1>
-{
-    typedef Matrix<typename internal::scalar_product_traits<typename Lhs::Scalar, typename Rhs::Scalar>::ReturnType,1,1> Base;
-  public:
-    GeneralProduct(const Lhs& lhs, const Rhs& rhs)
-    {
-      EIGEN_STATIC_ASSERT((internal::is_same<typename Lhs::RealScalar, typename Rhs::RealScalar>::value),
-        YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY)
-
-      Base::coeffRef(0,0) = (lhs.transpose().cwiseProduct(rhs)).sum();
-    }
-
-    /** Convertion to scalar */
-    operator const typename Base::Scalar() const {
-      return Base::coeff(0,0);
-    }
-};
-
 /***********************************************************************
 *  Implementation of Outer Vector Vector Product
 ***********************************************************************/
 
-namespace internal {
-
-// Column major
-template<typename ProductType, typename Dest, typename Func>
-EIGEN_DONT_INLINE void outer_product_selector_run(const ProductType& prod, Dest& dest, const Func& func, const false_type&)
-{
-  typedef typename Dest::Index Index;
-  // FIXME make sure lhs is sequentially stored
-  // FIXME not very good if rhs is real and lhs complex while alpha is real too
-  const Index cols = dest.cols();
-  for (Index j=0; j<cols; ++j)
-    func(dest.col(j), prod.rhs().coeff(j) * prod.lhs());
-}
-
-// Row major
-template<typename ProductType, typename Dest, typename Func>
-EIGEN_DONT_INLINE void outer_product_selector_run(const ProductType& prod, Dest& dest, const Func& func, const true_type&) {
-  typedef typename Dest::Index Index;
-  // FIXME make sure rhs is sequentially stored
-  // FIXME not very good if lhs is real and rhs complex while alpha is real too
-  const Index rows = dest.rows();
-  for (Index i=0; i<rows; ++i)
-    func(dest.row(i), prod.lhs().coeff(i) * prod.rhs());
-}
-
-template<typename Lhs, typename Rhs>
-struct traits<GeneralProduct<Lhs,Rhs,OuterProduct> >
- : traits<ProductBase<GeneralProduct<Lhs,Rhs,OuterProduct>, Lhs, Rhs> >
-{};
-
-}
-
-template<typename Lhs, typename Rhs>
-class GeneralProduct<Lhs, Rhs, OuterProduct>
-  : public ProductBase<GeneralProduct<Lhs,Rhs,OuterProduct>, Lhs, Rhs>
-{
-    template<typename T> struct IsRowMajor : internal::conditional<(int(T::Flags)&RowMajorBit), internal::true_type, internal::false_type>::type {};
-    
-  public:
-    EIGEN_PRODUCT_PUBLIC_INTERFACE(GeneralProduct)
-
-    GeneralProduct(const Lhs& lhs, const Rhs& rhs) : Base(lhs,rhs)
-    {
-      EIGEN_STATIC_ASSERT((internal::is_same<typename Lhs::RealScalar, typename Rhs::RealScalar>::value),
-        YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY)
-    }
-    
-    struct set  { template<typename Dst, typename Src> void operator()(const Dst& dst, const Src& src) const { dst.const_cast_derived()  = src; } };
-    struct add  { template<typename Dst, typename Src> void operator()(const Dst& dst, const Src& src) const { dst.const_cast_derived() += src; } };
-    struct sub  { template<typename Dst, typename Src> void operator()(const Dst& dst, const Src& src) const { dst.const_cast_derived() -= src; } };
-    struct adds {
-      Scalar m_scale;
-      adds(const Scalar& s) : m_scale(s) {}
-      template<typename Dst, typename Src> void operator()(const Dst& dst, const Src& src) const {
-        dst.const_cast_derived() += m_scale * src;
-      }
-    };
-    
-    template<typename Dest>
-    inline void evalTo(Dest& dest) const {
-      internal::outer_product_selector_run(*this, dest, set(), IsRowMajor<Dest>());
-    }
-    
-    template<typename Dest>
-    inline void addTo(Dest& dest) const {
-      internal::outer_product_selector_run(*this, dest, add(), IsRowMajor<Dest>());
-    }
-
-    template<typename Dest>
-    inline void subTo(Dest& dest) const {
-      internal::outer_product_selector_run(*this, dest, sub(), IsRowMajor<Dest>());
-    }
-
-    template<typename Dest> void scaleAndAddTo(Dest& dest, const Scalar& alpha) const
-    {
-      internal::outer_product_selector_run(*this, dest, adds(alpha), IsRowMajor<Dest>());
-    }
-};
-
 /***********************************************************************
 *  Implementation of General Matrix Vector Product
 ***********************************************************************/
@@ -313,60 +159,13 @@ class GeneralProduct<Lhs, Rhs, OuterProduct>
  */
 namespace internal {
 
-template<typename Lhs, typename Rhs>
-struct traits<GeneralProduct<Lhs,Rhs,GemvProduct> >
- : traits<ProductBase<GeneralProduct<Lhs,Rhs,GemvProduct>, Lhs, Rhs> >
-{};
-
 template<int Side, int StorageOrder, bool BlasCompatible>
-struct gemv_selector;
+struct gemv_dense_sense_selector;
 
 } // end namespace internal
 
-template<typename Lhs, typename Rhs>
-class GeneralProduct<Lhs, Rhs, GemvProduct>
-  : public ProductBase<GeneralProduct<Lhs,Rhs,GemvProduct>, Lhs, Rhs>
-{
-  public:
-    EIGEN_PRODUCT_PUBLIC_INTERFACE(GeneralProduct)
-
-    typedef typename Lhs::Scalar LhsScalar;
-    typedef typename Rhs::Scalar RhsScalar;
-
-    GeneralProduct(const Lhs& a_lhs, const Rhs& a_rhs) : Base(a_lhs,a_rhs)
-    {
-//       EIGEN_STATIC_ASSERT((internal::is_same<typename Lhs::Scalar, typename Rhs::Scalar>::value),
-//         YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY)
-    }
-
-    enum { Side = Lhs::IsVectorAtCompileTime ? OnTheLeft : OnTheRight };
-    typedef typename internal::conditional<int(Side)==OnTheRight,_LhsNested,_RhsNested>::type MatrixType;
-
-    template<typename Dest> void scaleAndAddTo(Dest& dst, const Scalar& alpha) const
-    {
-      eigen_assert(m_lhs.rows() == dst.rows() && m_rhs.cols() == dst.cols());
-      internal::gemv_selector<Side,(int(MatrixType::Flags)&RowMajorBit) ? RowMajor : ColMajor,
-                       bool(internal::blas_traits<MatrixType>::HasUsableDirectAccess)>::run(*this, dst, alpha);
-    }
-};
-
 namespace internal {
 
-// The vector is on the left => transposition
-template<int StorageOrder, bool BlasCompatible>
-struct gemv_selector<OnTheLeft,StorageOrder,BlasCompatible>
-{
-  template<typename ProductType, typename Dest>
-  static void run(const ProductType& prod, Dest& dest, const typename ProductType::Scalar& alpha)
-  {
-    Transpose<Dest> destT(dest);
-    enum { OtherStorageOrder = StorageOrder == RowMajor ? ColMajor : RowMajor };
-    gemv_selector<OnTheRight,OtherStorageOrder,BlasCompatible>
-      ::run(GeneralProduct<Transpose<const typename ProductType::_RhsNested>,Transpose<const typename ProductType::_LhsNested>, GemvProduct>
-        (prod.rhs().transpose(), prod.lhs().transpose()), destT, alpha);
-  }
-};
-
 template<typename Scalar,int Size,int MaxSize,bool Cond> struct gemv_static_vector_if;
 
 template<typename Scalar,int Size,int MaxSize>
@@ -384,7 +183,7 @@ struct gemv_static_vector_if<Scalar,Size,Dynamic,true>
 template<typename Scalar,int Size,int MaxSize>
 struct gemv_static_vector_if<Scalar,Size,MaxSize,true>
 {
-  #if EIGEN_ALIGN_STATICALLY
+  #if EIGEN_MAX_STATIC_ALIGN_BYTES!=0
   internal::plain_array<Scalar,EIGEN_SIZE_MIN_PREFER_FIXED(Size,MaxSize),0> m_data;
   EIGEN_STRONG_INLINE Scalar* data() { return m_data.array; }
   #else
@@ -397,33 +196,48 @@ struct gemv_static_vector_if<Scalar,Size,MaxSize,true>
   internal::plain_array<Scalar,EIGEN_SIZE_MIN_PREFER_FIXED(Size,MaxSize)+(ForceAlignment?PacketSize:0),0> m_data;
   EIGEN_STRONG_INLINE Scalar* data() {
     return ForceAlignment
-            ? reinterpret_cast<Scalar*>((reinterpret_cast<size_t>(m_data.array) & ~(size_t(15))) + 16)
+            ? reinterpret_cast<Scalar*>((reinterpret_cast<size_t>(m_data.array) & ~(size_t(EIGEN_MAX_ALIGN_BYTES-1))) + EIGEN_MAX_ALIGN_BYTES)
             : m_data.array;
   }
   #endif
 };
 
-template<> struct gemv_selector<OnTheRight,ColMajor,true>
+// The vector is on the left => transposition
+template<int StorageOrder, bool BlasCompatible>
+struct gemv_dense_sense_selector<OnTheLeft,StorageOrder,BlasCompatible>
 {
-  template<typename ProductType, typename Dest>
-  static inline void run(const ProductType& prod, Dest& dest, const typename ProductType::Scalar& alpha)
+  template<typename Lhs, typename Rhs, typename Dest>
+  static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha)
   {
-    typedef typename ProductType::Index Index;
-    typedef typename ProductType::LhsScalar   LhsScalar;
-    typedef typename ProductType::RhsScalar   RhsScalar;
-    typedef typename ProductType::Scalar      ResScalar;
-    typedef typename ProductType::RealScalar  RealScalar;
-    typedef typename ProductType::ActualLhsType ActualLhsType;
-    typedef typename ProductType::ActualRhsType ActualRhsType;
-    typedef typename ProductType::LhsBlasTraits LhsBlasTraits;
-    typedef typename ProductType::RhsBlasTraits RhsBlasTraits;
+    Transpose<Dest> destT(dest);
+    enum { OtherStorageOrder = StorageOrder == RowMajor ? ColMajor : RowMajor };
+    gemv_dense_sense_selector<OnTheRight,OtherStorageOrder,BlasCompatible>
+      ::run(rhs.transpose(), lhs.transpose(), destT, alpha);
+  }
+};
+
+template<> struct gemv_dense_sense_selector<OnTheRight,ColMajor,true>
+{
+  template<typename Lhs, typename Rhs, typename Dest>
+  static inline void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha)
+  {
+    typedef typename Lhs::Scalar   LhsScalar;
+    typedef typename Rhs::Scalar   RhsScalar;
+    typedef typename Dest::Scalar  ResScalar;
+    typedef typename Dest::RealScalar  RealScalar;
+    
+    typedef internal::blas_traits<Lhs> LhsBlasTraits;
+    typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType;
+    typedef internal::blas_traits<Rhs> RhsBlasTraits;
+    typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType;
+  
     typedef Map<Matrix<ResScalar,Dynamic,1>, Aligned> MappedDest;
 
-    ActualLhsType actualLhs = LhsBlasTraits::extract(prod.lhs());
-    ActualRhsType actualRhs = RhsBlasTraits::extract(prod.rhs());
+    ActualLhsType actualLhs = LhsBlasTraits::extract(lhs);
+    ActualRhsType actualRhs = RhsBlasTraits::extract(rhs);
 
-    ResScalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(prod.lhs())
-                                  * RhsBlasTraits::extractScalarFactor(prod.rhs());
+    ResScalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(lhs)
+                                  * RhsBlasTraits::extractScalarFactor(rhs);
 
     enum {
       // FIXME find a way to allow an inner stride on the result if packet_traits<Scalar>::size==1
@@ -435,18 +249,18 @@ template<> struct gemv_selector<OnTheRight,ColMajor,true>
 
     gemv_static_vector_if<ResScalar,Dest::SizeAtCompileTime,Dest::MaxSizeAtCompileTime,MightCannotUseDest> static_dest;
 
-    bool alphaIsCompatible = (!ComplexByReal) || (numext::imag(actualAlpha)==RealScalar(0));
-    bool evalToDest = EvalToDestAtCompileTime && alphaIsCompatible;
-    
+    const bool alphaIsCompatible = (!ComplexByReal) || (numext::imag(actualAlpha)==RealScalar(0));
+    const bool evalToDest = EvalToDestAtCompileTime && alphaIsCompatible;
+
     RhsScalar compatibleAlpha = get_factor<ResScalar,RhsScalar>::run(actualAlpha);
 
     ei_declare_aligned_stack_constructed_variable(ResScalar,actualDestPtr,dest.size(),
                                                   evalToDest ? dest.data() : static_dest.data());
-    
+
     if(!evalToDest)
     {
       #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN
-      int size = dest.size();
+      Index size = dest.size();
       EIGEN_DENSE_STORAGE_CTOR_PLUGIN
       #endif
       if(!alphaIsCompatible)
@@ -458,11 +272,13 @@ template<> struct gemv_selector<OnTheRight,ColMajor,true>
         MappedDest(actualDestPtr, dest.size()) = dest;
     }
 
+    typedef const_blas_data_mapper<LhsScalar,Index,ColMajor> LhsMapper;
+    typedef const_blas_data_mapper<RhsScalar,Index,RowMajor> RhsMapper;
     general_matrix_vector_product
-      <Index,LhsScalar,ColMajor,LhsBlasTraits::NeedToConjugate,RhsScalar,RhsBlasTraits::NeedToConjugate>::run(
+        <Index,LhsScalar,LhsMapper,ColMajor,LhsBlasTraits::NeedToConjugate,RhsScalar,RhsMapper,RhsBlasTraits::NeedToConjugate>::run(
         actualLhs.rows(), actualLhs.cols(),
-        actualLhs.data(), actualLhs.outerStride(),
-        actualRhs.data(), actualRhs.innerStride(),
+        LhsMapper(actualLhs.data(), actualLhs.outerStride()),
+        RhsMapper(actualRhs.data(), actualRhs.innerStride()),
         actualDestPtr, 1,
         compatibleAlpha);
 
@@ -476,34 +292,34 @@ template<> struct gemv_selector<OnTheRight,ColMajor,true>
   }
 };
 
-template<> struct gemv_selector<OnTheRight,RowMajor,true>
+template<> struct gemv_dense_sense_selector<OnTheRight,RowMajor,true>
 {
-  template<typename ProductType, typename Dest>
-  static void run(const ProductType& prod, Dest& dest, const typename ProductType::Scalar& alpha)
+  template<typename Lhs, typename Rhs, typename Dest>
+  static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha)
   {
-    typedef typename ProductType::LhsScalar LhsScalar;
-    typedef typename ProductType::RhsScalar RhsScalar;
-    typedef typename ProductType::Scalar    ResScalar;
-    typedef typename ProductType::Index Index;
-    typedef typename ProductType::ActualLhsType ActualLhsType;
-    typedef typename ProductType::ActualRhsType ActualRhsType;
-    typedef typename ProductType::_ActualRhsType _ActualRhsType;
-    typedef typename ProductType::LhsBlasTraits LhsBlasTraits;
-    typedef typename ProductType::RhsBlasTraits RhsBlasTraits;
+    typedef typename Lhs::Scalar   LhsScalar;
+    typedef typename Rhs::Scalar   RhsScalar;
+    typedef typename Dest::Scalar  ResScalar;
+    
+    typedef internal::blas_traits<Lhs> LhsBlasTraits;
+    typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType;
+    typedef internal::blas_traits<Rhs> RhsBlasTraits;
+    typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType;
+    typedef typename internal::remove_all<ActualRhsType>::type ActualRhsTypeCleaned;
 
-    typename add_const<ActualLhsType>::type actualLhs = LhsBlasTraits::extract(prod.lhs());
-    typename add_const<ActualRhsType>::type actualRhs = RhsBlasTraits::extract(prod.rhs());
+    typename add_const<ActualLhsType>::type actualLhs = LhsBlasTraits::extract(lhs);
+    typename add_const<ActualRhsType>::type actualRhs = RhsBlasTraits::extract(rhs);
 
-    ResScalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(prod.lhs())
-                                  * RhsBlasTraits::extractScalarFactor(prod.rhs());
+    ResScalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(lhs)
+                                  * RhsBlasTraits::extractScalarFactor(rhs);
 
     enum {
       // FIXME find a way to allow an inner stride on the result if packet_traits<Scalar>::size==1
       // on, the other hand it is good for the cache to pack the vector anyways...
-      DirectlyUseRhs = _ActualRhsType::InnerStrideAtCompileTime==1
+      DirectlyUseRhs = ActualRhsTypeCleaned::InnerStrideAtCompileTime==1
     };
 
-    gemv_static_vector_if<RhsScalar,_ActualRhsType::SizeAtCompileTime,_ActualRhsType::MaxSizeAtCompileTime,!DirectlyUseRhs> static_rhs;
+    gemv_static_vector_if<RhsScalar,ActualRhsTypeCleaned::SizeAtCompileTime,ActualRhsTypeCleaned::MaxSizeAtCompileTime,!DirectlyUseRhs> static_rhs;
 
     ei_declare_aligned_stack_constructed_variable(RhsScalar,actualRhsPtr,actualRhs.size(),
         DirectlyUseRhs ? const_cast<RhsScalar*>(actualRhs.data()) : static_rhs.data());
@@ -511,45 +327,45 @@ template<> struct gemv_selector<OnTheRight,RowMajor,true>
     if(!DirectlyUseRhs)
     {
       #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN
-      int size = actualRhs.size();
+      Index size = actualRhs.size();
       EIGEN_DENSE_STORAGE_CTOR_PLUGIN
       #endif
-      Map<typename _ActualRhsType::PlainObject>(actualRhsPtr, actualRhs.size()) = actualRhs;
+      Map<typename ActualRhsTypeCleaned::PlainObject>(actualRhsPtr, actualRhs.size()) = actualRhs;
     }
 
+    typedef const_blas_data_mapper<LhsScalar,Index,RowMajor> LhsMapper;
+    typedef const_blas_data_mapper<RhsScalar,Index,ColMajor> RhsMapper;
     general_matrix_vector_product
-      <Index,LhsScalar,RowMajor,LhsBlasTraits::NeedToConjugate,RhsScalar,RhsBlasTraits::NeedToConjugate>::run(
+        <Index,LhsScalar,LhsMapper,RowMajor,LhsBlasTraits::NeedToConjugate,RhsScalar,RhsMapper,RhsBlasTraits::NeedToConjugate>::run(
         actualLhs.rows(), actualLhs.cols(),
-        actualLhs.data(), actualLhs.outerStride(),
-        actualRhsPtr, 1,
+        LhsMapper(actualLhs.data(), actualLhs.outerStride()),
+        RhsMapper(actualRhsPtr, 1),
         dest.data(), dest.innerStride(),
         actualAlpha);
   }
 };
 
-template<> struct gemv_selector<OnTheRight,ColMajor,false>
+template<> struct gemv_dense_sense_selector<OnTheRight,ColMajor,false>
 {
-  template<typename ProductType, typename Dest>
-  static void run(const ProductType& prod, Dest& dest, const typename ProductType::Scalar& alpha)
+  template<typename Lhs, typename Rhs, typename Dest>
+  static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha)
   {
-    typedef typename Dest::Index Index;
     // TODO makes sure dest is sequentially stored in memory, otherwise use a temp
-    const Index size = prod.rhs().rows();
+    const Index size = rhs.rows();
     for(Index k=0; k<size; ++k)
-      dest += (alpha*prod.rhs().coeff(k)) * prod.lhs().col(k);
+      dest += (alpha*rhs.coeff(k)) * lhs.col(k);
   }
 };
 
-template<> struct gemv_selector<OnTheRight,RowMajor,false>
+template<> struct gemv_dense_sense_selector<OnTheRight,RowMajor,false>
 {
-  template<typename ProductType, typename Dest>
-  static void run(const ProductType& prod, Dest& dest, const typename ProductType::Scalar& alpha)
+  template<typename Lhs, typename Rhs, typename Dest>
+  static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha)
   {
-    typedef typename Dest::Index Index;
     // TODO makes sure rhs is sequentially stored in memory, otherwise use a temp
-    const Index rows = prod.rows();
+    const Index rows = dest.rows();
     for(Index i=0; i<rows; ++i)
-      dest.coeffRef(i) += alpha * (prod.lhs().row(i).cwiseProduct(prod.rhs().transpose())).sum();
+      dest.coeffRef(i) += alpha * (lhs.row(i).cwiseProduct(rhs.transpose())).sum();
   }
 };
 
@@ -565,9 +381,11 @@ template<> struct gemv_selector<OnTheRight,RowMajor,false>
   *
   * \sa lazyProduct(), operator*=(const MatrixBase&), Cwise::operator*()
   */
+#ifndef __CUDACC__
+
 template<typename Derived>
 template<typename OtherDerived>
-inline const typename ProductReturnType<Derived, OtherDerived>::Type
+inline const Product<Derived, OtherDerived>
 MatrixBase<Derived>::operator*(const MatrixBase<OtherDerived> &other) const
 {
   // A note regarding the function declaration: In MSVC, this function will sometimes
@@ -592,9 +410,12 @@ MatrixBase<Derived>::operator*(const MatrixBase<OtherDerived> &other) const
 #ifdef EIGEN_DEBUG_PRODUCT
   internal::product_type<Derived,OtherDerived>::debug();
 #endif
-  return typename ProductReturnType<Derived,OtherDerived>::Type(derived(), other.derived());
+
+  return Product<Derived, OtherDerived>(derived(), other.derived());
 }
 
+#endif // __CUDACC__
+
 /** \returns an expression of the matrix product of \c *this and \a other without implicit evaluation.
   *
   * The returned product will behave like any other expressions: the coefficients of the product will be
@@ -608,7 +429,7 @@ MatrixBase<Derived>::operator*(const MatrixBase<OtherDerived> &other) const
   */
 template<typename Derived>
 template<typename OtherDerived>
-const typename LazyProductReturnType<Derived,OtherDerived>::Type
+const Product<Derived,OtherDerived,LazyProduct>
 MatrixBase<Derived>::lazyProduct(const MatrixBase<OtherDerived> &other) const
 {
   enum {
@@ -627,7 +448,7 @@ MatrixBase<Derived>::lazyProduct(const MatrixBase<OtherDerived> &other) const
     INVALID_MATRIX_PRODUCT__IF_YOU_WANTED_A_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTION)
   EIGEN_STATIC_ASSERT(ProductIsValid || SameSizes, INVALID_MATRIX_PRODUCT)
 
-  return typename LazyProductReturnType<Derived,OtherDerived>::Type(derived(), other.derived());
+  return Product<Derived,OtherDerived,LazyProduct>(derived(), other.derived());
 }
 
 } // end namespace Eigen

Eigen/src/Core/GenericPacketMath.h

@@ -42,21 +42,27 @@ namespace internal {
 struct default_packet_traits
 {
   enum {
+    HasHalfPacket = 0,
+    
     HasAdd    = 1,
     HasSub    = 1,
     HasMul    = 1,
     HasNegate = 1,
     HasAbs    = 1,
+    HasArg    = 0,
     HasAbs2   = 1,
     HasMin    = 1,
     HasMax    = 1,
     HasConj   = 1,
     HasSetLinear = 1,
+    HasBlend  = 0,
 
     HasDiv    = 0,
     HasSqrt   = 0,
+    HasRsqrt  = 0,
     HasExp    = 0,
     HasLog    = 0,
+    HasLog10    = 0,
     HasPow    = 0,
 
     HasSin    = 0,
@@ -64,17 +70,26 @@ struct default_packet_traits
     HasTan    = 0,
     HasASin   = 0,
     HasACos   = 0,
-    HasATan   = 0
+    HasATan   = 0,
+    HasSinh    = 0,
+    HasCosh    = 0,
+    HasTanh    = 0,
+
+    HasRound  = 0,
+    HasFloor  = 0,
+    HasCeil   = 0
   };
 };
 
 template<typename T> struct packet_traits : default_packet_traits
 {
   typedef T type;
+  typedef T half;
   enum {
     Vectorizable = 0,
     size = 1,
-    AlignedOnScalar = 0
+    AlignedOnScalar = 0,
+    HasHalfPacket = 0
   };
   enum {
     HasAdd    = 0,
@@ -90,135 +105,250 @@ template<typename T> struct packet_traits : default_packet_traits
   };
 };
 
+template<typename T> struct packet_traits<const T> : packet_traits<T> { };
+
+template <typename Src, typename Tgt> struct type_casting_traits {
+  enum {
+    VectorizedCast = 0,
+    SrcCoeffRatio = 1,
+    TgtCoeffRatio = 1
+  };
+};
+
+
+/** \internal \returns static_cast<TgtType>(a) (coeff-wise) */
+template <typename SrcPacket, typename TgtPacket>
+EIGEN_DEVICE_FUNC inline TgtPacket
+pcast(const SrcPacket& a) {
+  return static_cast<TgtPacket>(a);
+}
+template <typename SrcPacket, typename TgtPacket>
+EIGEN_DEVICE_FUNC inline TgtPacket
+pcast(const SrcPacket& a, const SrcPacket& /*b*/) {
+  return static_cast<TgtPacket>(a);
+}
+
+
 /** \internal \returns a + b (coeff-wise) */
-template<typename Packet> inline Packet
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
 padd(const Packet& a,
         const Packet& b) { return a+b; }
 
 /** \internal \returns a - b (coeff-wise) */
-template<typename Packet> inline Packet
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
 psub(const Packet& a,
         const Packet& b) { return a-b; }
 
 /** \internal \returns -a (coeff-wise) */
-template<typename Packet> inline Packet
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
 pnegate(const Packet& a) { return -a; }
 
 /** \internal \returns conj(a) (coeff-wise) */
-template<typename Packet> inline Packet
+
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
 pconj(const Packet& a) { return numext::conj(a); }
 
 /** \internal \returns a * b (coeff-wise) */
-template<typename Packet> inline Packet
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
 pmul(const Packet& a,
         const Packet& b) { return a*b; }
 
 /** \internal \returns a / b (coeff-wise) */
-template<typename Packet> inline Packet
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
 pdiv(const Packet& a,
         const Packet& b) { return a/b; }
 
 /** \internal \returns the min of \a a and \a b  (coeff-wise) */
-template<typename Packet> inline Packet
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
 pmin(const Packet& a,
-        const Packet& b) { using std::min; return (min)(a, b); }
+        const Packet& b) { return numext::mini(a, b); }
 
 /** \internal \returns the max of \a a and \a b  (coeff-wise) */
-template<typename Packet> inline Packet
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
 pmax(const Packet& a,
-        const Packet& b) { using std::max; return (max)(a, b); }
+        const Packet& b) { return numext::maxi(a, b); }
 
 /** \internal \returns the absolute value of \a a */
-template<typename Packet> inline Packet
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
 pabs(const Packet& a) { using std::abs; return abs(a); }
 
+/** \internal \returns the phase angle of \a a */
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
+parg(const Packet& a) { using numext::arg; return arg(a); }
+
 /** \internal \returns the bitwise and of \a a and \a b */
-template<typename Packet> inline Packet
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
 pand(const Packet& a, const Packet& b) { return a & b; }
 
 /** \internal \returns the bitwise or of \a a and \a b */
-template<typename Packet> inline Packet
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
 por(const Packet& a, const Packet& b) { return a | b; }
 
 /** \internal \returns the bitwise xor of \a a and \a b */
-template<typename Packet> inline Packet
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
 pxor(const Packet& a, const Packet& b) { return a ^ b; }
 
 /** \internal \returns the bitwise andnot of \a a and \a b */
-template<typename Packet> inline Packet
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
 pandnot(const Packet& a, const Packet& b) { return a & (!b); }
 
 /** \internal \returns a packet version of \a *from, from must be 16 bytes aligned */
-template<typename Packet> inline Packet
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
 pload(const typename unpacket_traits<Packet>::type* from) { return *from; }
 
 /** \internal \returns a packet version of \a *from, (un-aligned load) */
-template<typename Packet> inline Packet
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
 ploadu(const typename unpacket_traits<Packet>::type* from) { return *from; }
 
-/** \internal \returns a packet with elements of \a *from duplicated.
-  * For instance, for a packet of 8 elements, 4 scalar will be read from \a *from and
-  * duplicated to form: {from[0],from[0],from[1],from[1],,from[2],from[2],,from[3],from[3]}
-  * Currently, this function is only used for scalar * complex products.
- */
-template<typename Packet> inline Packet
-ploaddup(const typename unpacket_traits<Packet>::type* from) { return *from; }
-
 /** \internal \returns a packet with constant coefficients \a a, e.g.: (a,a,a,a) */
-template<typename Packet> inline Packet
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
 pset1(const typename unpacket_traits<Packet>::type& a) { return a; }
 
+/** \internal \returns a packet with constant coefficients \a a[0], e.g.: (a[0],a[0],a[0],a[0]) */
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
+pload1(const typename unpacket_traits<Packet>::type  *a) { return pset1<Packet>(*a); }
+
+/** \internal \returns a packet with elements of \a *from duplicated.
+  * For instance, for a packet of 8 elements, 4 scalars will be read from \a *from and
+  * duplicated to form: {from[0],from[0],from[1],from[1],from[2],from[2],from[3],from[3]}
+  * Currently, this function is only used for scalar * complex products.
+  */
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
+ploaddup(const typename unpacket_traits<Packet>::type* from) { return *from; }
+
+/** \internal \returns a packet with elements of \a *from quadrupled.
+  * For instance, for a packet of 8 elements, 2 scalars will be read from \a *from and
+  * replicated to form: {from[0],from[0],from[0],from[0],from[1],from[1],from[1],from[1]}
+  * Currently, this function is only used in matrix products.
+  * For packet-size smaller or equal to 4, this function is equivalent to pload1 
+  */
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
+ploadquad(const typename unpacket_traits<Packet>::type* from)
+{ return pload1<Packet>(from); }
+
+/** \internal equivalent to
+  * \code
+  * a0 = pload1(a+0);
+  * a1 = pload1(a+1);
+  * a2 = pload1(a+2);
+  * a3 = pload1(a+3);
+  * \endcode
+  * \sa pset1, pload1, ploaddup, pbroadcast2
+  */
+template<typename Packet> EIGEN_DEVICE_FUNC
+inline void pbroadcast4(const typename unpacket_traits<Packet>::type *a,
+                        Packet& a0, Packet& a1, Packet& a2, Packet& a3)
+{
+  a0 = pload1<Packet>(a+0);
+  a1 = pload1<Packet>(a+1);
+  a2 = pload1<Packet>(a+2);
+  a3 = pload1<Packet>(a+3);
+}
+
+/** \internal equivalent to
+  * \code
+  * a0 = pload1(a+0);
+  * a1 = pload1(a+1);
+  * \endcode
+  * \sa pset1, pload1, ploaddup, pbroadcast4
+  */
+template<typename Packet> EIGEN_DEVICE_FUNC
+inline void pbroadcast2(const typename unpacket_traits<Packet>::type *a,
+                        Packet& a0, Packet& a1)
+{
+  a0 = pload1<Packet>(a+0);
+  a1 = pload1<Packet>(a+1);
+}
+
 /** \internal \brief Returns a packet with coefficients (a,a+1,...,a+packet_size-1). */
-template<typename Scalar> inline typename packet_traits<Scalar>::type
-plset(const Scalar& a) { return a; }
+template<typename Packet> inline Packet
+plset(const typename unpacket_traits<Packet>::type& a) { return a; }
 
 /** \internal copy the packet \a from to \a *to, \a to must be 16 bytes aligned */
-template<typename Scalar, typename Packet> inline void pstore(Scalar* to, const Packet& from)
+template<typename Scalar, typename Packet> EIGEN_DEVICE_FUNC inline void pstore(Scalar* to, const Packet& from)
 { (*to) = from; }
 
 /** \internal copy the packet \a from to \a *to, (un-aligned store) */
-template<typename Scalar, typename Packet> inline void pstoreu(Scalar* to, const Packet& from)
-{ (*to) = from; }
+template<typename Scalar, typename Packet> EIGEN_DEVICE_FUNC inline void pstoreu(Scalar* to, const Packet& from)
+{  (*to) = from; }
+
+ template<typename Scalar, typename Packet> EIGEN_DEVICE_FUNC inline Packet pgather(const Scalar* from, Index /*stride*/)
+ { return ploadu<Packet>(from); }
+
+ template<typename Scalar, typename Packet> EIGEN_DEVICE_FUNC inline void pscatter(Scalar* to, const Packet& from, Index /*stride*/)
+ { pstore(to, from); }
 
 /** \internal tries to do cache prefetching of \a addr */
 template<typename Scalar> inline void prefetch(const Scalar* addr)
 {
-#if !defined(_MSC_VER)
-__builtin_prefetch(addr);
+#ifdef __CUDA_ARCH__
+#if defined(__LP64__)
+  // 64-bit pointer operand constraint for inlined asm
+  asm(" prefetch.L1 [ %1 ];" : "=l"(addr) : "l"(addr));
+#else
+  // 32-bit pointer operand constraint for inlined asm
+  asm(" prefetch.L1 [ %1 ];" : "=r"(addr) : "r"(addr));
+#endif
+#elif !EIGEN_COMP_MSVC
+  __builtin_prefetch(addr);
 #endif
 }
 
 /** \internal \returns the first element of a packet */
-template<typename Packet> inline typename unpacket_traits<Packet>::type pfirst(const Packet& a)
+template<typename Packet> EIGEN_DEVICE_FUNC inline typename unpacket_traits<Packet>::type pfirst(const Packet& a)
 { return a; }
 
 /** \internal \returns a packet where the element i contains the sum of the packet of \a vec[i] */
-template<typename Packet> inline Packet
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
 preduxp(const Packet* vecs) { return vecs[0]; }
 
 /** \internal \returns the sum of the elements of \a a*/
-template<typename Packet> inline typename unpacket_traits<Packet>::type predux(const Packet& a)
+template<typename Packet> EIGEN_DEVICE_FUNC inline typename unpacket_traits<Packet>::type predux(const Packet& a)
+{ return a; }
+
+/** \internal \returns the sum of the elements of \a a by block of 4 elements.
+  * For a packet {a0, a1, a2, a3, a4, a5, a6, a7}, it returns a half packet {a0+a4, a1+a5, a2+a6, a3+a7}
+  * For packet-size smaller or equal to 4, this boils down to a noop.
+  */
+template<typename Packet> EIGEN_DEVICE_FUNC inline
+typename conditional<(unpacket_traits<Packet>::size%8)==0,typename unpacket_traits<Packet>::half,Packet>::type
+predux4(const Packet& a)
 { return a; }
 
 /** \internal \returns the product of the elements of \a a*/
-template<typename Packet> inline typename unpacket_traits<Packet>::type predux_mul(const Packet& a)
+template<typename Packet> EIGEN_DEVICE_FUNC inline typename unpacket_traits<Packet>::type predux_mul(const Packet& a)
 { return a; }
 
 /** \internal \returns the min of the elements of \a a*/
-template<typename Packet> inline typename unpacket_traits<Packet>::type predux_min(const Packet& a)
+template<typename Packet> EIGEN_DEVICE_FUNC inline typename unpacket_traits<Packet>::type predux_min(const Packet& a)
 { return a; }
 
 /** \internal \returns the max of the elements of \a a*/
-template<typename Packet> inline typename unpacket_traits<Packet>::type predux_max(const Packet& a)
+template<typename Packet> EIGEN_DEVICE_FUNC inline typename unpacket_traits<Packet>::type predux_max(const Packet& a)
 { return a; }
 
 /** \internal \returns the reversed elements of \a a*/
-template<typename Packet> inline Packet preverse(const Packet& a)
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet preverse(const Packet& a)
 { return a; }
 
+template<size_t offset, typename Packet>
+struct protate_impl
+{
+  // Empty so attempts to use this unimplemented path will fail to compile.
+  // Only specializations of this template should be used.
+};
+
+/** \internal \returns a packet with the coefficients rotated to the right in little-endian convention,
+  * by the given offset, e.g. for offset == 1:
+  *     (packet[3], packet[2], packet[1], packet[0]) becomes (packet[0], packet[3], packet[2], packet[1])
+  */
+template<size_t offset, typename Packet> EIGEN_DEVICE_FUNC inline Packet protate(const Packet& a)
+{
+  return offset ? protate_impl<offset, Packet>::run(a) : a;
+}
 
 /** \internal \returns \a a with real and imaginary part flipped (for complex type only) */
-template<typename Packet> inline Packet pcplxflip(const Packet& a)
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet pcplxflip(const Packet& a)
 {
   // FIXME: uncomment the following in case we drop the internal imag and real functions.
 //   using std::imag;
@@ -250,6 +380,22 @@ Packet pasin(const Packet& a) { using std::asin; return asin(a); }
 template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
 Packet pacos(const Packet& a) { using std::acos; return acos(a); }
 
+/** \internal \returns the arc tangent of \a a (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet patan(const Packet& a) { using std::atan; return atan(a); }
+
+/** \internal \returns the hyperbolic sine of \a a (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet psinh(const Packet& a) { using std::sinh; return sinh(a); }
+
+/** \internal \returns the hyperbolic cosine of \a a (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet pcosh(const Packet& a) { using std::cosh; return cosh(a); }
+
+/** \internal \returns the hyperbolic tan of \a a (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet ptanh(const Packet& a) { using std::tanh; return tanh(a); }
+
 /** \internal \returns the exp of \a a (coeff-wise) */
 template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
 Packet pexp(const Packet& a) { using std::exp; return exp(a); }
@@ -258,10 +404,32 @@ Packet pexp(const Packet& a) { using std::exp; return exp(a); }
 template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
 Packet plog(const Packet& a) { using std::log; return log(a); }
 
+/** \internal \returns the log10 of \a a (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet plog10(const Packet& a) { using std::log10; return log10(a); }
+
 /** \internal \returns the square-root of \a a (coeff-wise) */
 template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
 Packet psqrt(const Packet& a) { using std::sqrt; return sqrt(a); }
 
+/** \internal \returns the reciprocal square-root of \a a (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet prsqrt(const Packet& a) {
+  return pdiv(pset1<Packet>(1), psqrt(a));
+}
+
+/** \internal \returns the rounded value of \a a (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet pround(const Packet& a) { using numext::round; return round(a); }
+
+/** \internal \returns the floor of \a a (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet pfloor(const Packet& a) { using numext::floor; return floor(a); }
+
+/** \internal \returns the ceil of \a a (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet pceil(const Packet& a) { using numext::ceil; return ceil(a); }
+
 /***************************************************************************
 * The following functions might not have to be overwritten for vectorized types
 ***************************************************************************/
@@ -275,34 +443,45 @@ inline void pstore1(typename unpacket_traits<Packet>::type* to, const typename u
 }
 
 /** \internal \returns a * b + c (coeff-wise) */
-template<typename Packet> inline Packet
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
 pmadd(const Packet&  a,
          const Packet&  b,
          const Packet&  c)
 { return padd(pmul(a, b),c); }
 
 /** \internal \returns a packet version of \a *from.
-  * If LoadMode equals #Aligned, \a from must be 16 bytes aligned */
-template<typename Packet, int LoadMode>
-inline Packet ploadt(const typename unpacket_traits<Packet>::type* from)
+  * The pointer \a from must be aligned on a \a Alignment bytes boundary. */
+template<typename Packet, int Alignment>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Packet ploadt(const typename unpacket_traits<Packet>::type* from)
 {
-  if(LoadMode == Aligned)
+  if(Alignment >= unpacket_traits<Packet>::alignment)
     return pload<Packet>(from);
   else
     return ploadu<Packet>(from);
 }
 
 /** \internal copy the packet \a from to \a *to.
-  * If StoreMode equals #Aligned, \a to must be 16 bytes aligned */
-template<typename Scalar, typename Packet, int LoadMode>
-inline void pstoret(Scalar* to, const Packet& from)
+  * The pointer \a from must be aligned on a \a Alignment bytes boundary. */
+template<typename Scalar, typename Packet, int Alignment>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void pstoret(Scalar* to, const Packet& from)
 {
-  if(LoadMode == Aligned)
+  if(Alignment >= unpacket_traits<Packet>::alignment)
     pstore(to, from);
   else
     pstoreu(to, from);
 }
 
+/** \internal \returns a packet version of \a *from.
+  * Unlike ploadt, ploadt_ro takes advantage of the read-only memory path on the
+  * hardware if available to speedup the loading of data that won't be modified
+  * by the current computation.
+  */
+template<typename Packet, int LoadMode>
+inline Packet ploadt_ro(const typename unpacket_traits<Packet>::type* from)
+{
+  return ploadt<Packet, LoadMode>(from);
+}
+
 /** \internal default implementation of palign() allowing partial specialization */
 template<int Offset,typename PacketType>
 struct palign_impl
@@ -336,15 +515,46 @@ inline void palign(PacketType& first, const PacketType& second)
 * Fast complex products (GCC generates a function call which is very slow)
 ***************************************************************************/
 
+// Eigen+CUDA does not support complexes.
+#ifndef __CUDACC__
+
 template<> inline std::complex<float> pmul(const std::complex<float>& a, const std::complex<float>& b)
 { return std::complex<float>(real(a)*real(b) - imag(a)*imag(b), imag(a)*real(b) + real(a)*imag(b)); }
 
 template<> inline std::complex<double> pmul(const std::complex<double>& a, const std::complex<double>& b)
 { return std::complex<double>(real(a)*real(b) - imag(a)*imag(b), imag(a)*real(b) + real(a)*imag(b)); }
 
+#endif
+
+
+/***************************************************************************
+ * PacketBlock, that is a collection of N packets where the number of words
+ * in the packet is a multiple of N.
+***************************************************************************/
+template <typename Packet,int N=unpacket_traits<Packet>::size> struct PacketBlock {
+  Packet packet[N];
+};
+
+template<typename Packet> EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<Packet,1>& /*kernel*/) {
+  // Nothing to do in the scalar case, i.e. a 1x1 matrix.
+}
+
+/***************************************************************************
+ * Selector, i.e. vector of N boolean values used to select (i.e. blend)
+ * words from 2 packets.
+***************************************************************************/
+template <size_t N> struct Selector {
+  bool select[N];
+};
+
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
+pblend(const Selector<unpacket_traits<Packet>::size>& ifPacket, const Packet& thenPacket, const Packet& elsePacket) {
+  return ifPacket.select[0] ? thenPacket : elsePacket;
+}
+
 } // end namespace internal
 
 } // end namespace Eigen
 
 #endif // EIGEN_GENERIC_PACKET_MATH_H
-

Eigen/src/Core/GlobalFunctions.h

@@ -14,8 +14,8 @@
 #define EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(NAME,FUNCTOR) \
   template<typename Derived> \
   inline const Eigen::CwiseUnaryOp<Eigen::internal::FUNCTOR<typename Derived::Scalar>, const Derived> \
-  NAME(const Eigen::ArrayBase<Derived>& x) { \
-    return x.derived(); \
+  (NAME)(const Eigen::ArrayBase<Derived>& x) { \
+    return Eigen::CwiseUnaryOp<Eigen::internal::FUNCTOR<typename Derived::Scalar>, const Derived>(x.derived()); \
   }
 
 #define EIGEN_ARRAY_DECLARE_GLOBAL_EIGEN_UNARY(NAME,FUNCTOR) \
@@ -30,25 +30,40 @@
   { \
     static inline typename NAME##_retval<ArrayBase<Derived> >::type run(const Eigen::ArrayBase<Derived>& x) \
     { \
-      return x.derived(); \
+      return typename NAME##_retval<ArrayBase<Derived> >::type(x.derived()); \
     } \
   };
 
-
 namespace Eigen
 {
   EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(real,scalar_real_op)
   EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(imag,scalar_imag_op)
   EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(conj,scalar_conjugate_op)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(inverse,scalar_inverse_op)
   EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(sin,scalar_sin_op)
   EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(cos,scalar_cos_op)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(tan,scalar_tan_op)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(atan,scalar_atan_op)
   EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(asin,scalar_asin_op)
   EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(acos,scalar_acos_op)
-  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(tan,scalar_tan_op)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(sinh,scalar_sinh_op)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(cosh,scalar_cosh_op)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(tanh,scalar_tanh_op)
   EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(exp,scalar_exp_op)
   EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(log,scalar_log_op)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(log10,scalar_log10_op)
   EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(abs,scalar_abs_op)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(abs2,scalar_abs2_op)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(arg,scalar_arg_op)
   EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(sqrt,scalar_sqrt_op)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(square,scalar_square_op)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(cube,scalar_cube_op)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(round,scalar_round_op)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(floor,scalar_floor_op)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(ceil,scalar_ceil_op)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(isnan,scalar_isnan_op)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(isinf,scalar_isinf_op)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(isfinite,scalar_isfinite_op)
   
   template<typename Derived>
   inline const Eigen::CwiseUnaryOp<Eigen::internal::scalar_pow_op<typename Derived::Scalar>, const Derived>
@@ -56,16 +71,46 @@ namespace Eigen
     return x.derived().pow(exponent);
   }
 
-  template<typename Derived>
-  inline const Eigen::CwiseBinaryOp<Eigen::internal::scalar_binary_pow_op<typename Derived::Scalar, typename Derived::Scalar>, const Derived, const Derived>
-  pow(const Eigen::ArrayBase<Derived>& x, const Eigen::ArrayBase<Derived>& exponents) 
+  /** \returns an expression of the coefficient-wise power of \a x to the given array of \a exponents.
+    *
+    * This function computes the coefficient-wise power.
+    *
+    * Example: \include Cwise_array_power_array.cpp
+    * Output: \verbinclude Cwise_array_power_array.out
+    * 
+    * \sa ArrayBase::pow()
+    */
+  template<typename Derived,typename ExponentDerived>
+  inline const Eigen::CwiseBinaryOp<Eigen::internal::scalar_binary_pow_op<typename Derived::Scalar, typename ExponentDerived::Scalar>, const Derived, const ExponentDerived>
+  pow(const Eigen::ArrayBase<Derived>& x, const Eigen::ArrayBase<ExponentDerived>& exponents) 
   {
-    return Eigen::CwiseBinaryOp<Eigen::internal::scalar_binary_pow_op<typename Derived::Scalar, typename Derived::Scalar>, const Derived, const Derived>(
+    return Eigen::CwiseBinaryOp<Eigen::internal::scalar_binary_pow_op<typename Derived::Scalar, typename ExponentDerived::Scalar>, const Derived, const ExponentDerived>(
       x.derived(),
       exponents.derived()
     );
   }
   
+  /** \returns an expression of the coefficient-wise power of the scalar \a x to the given array of \a exponents.
+    *
+    * This function computes the coefficient-wise power between a scalar and an array of exponents.
+    * Beaware that the scalar type of the input scalar \a x and the exponents \a exponents must be the same.
+    *
+    * Example: \include Cwise_scalar_power_array.cpp
+    * Output: \verbinclude Cwise_scalar_power_array.out
+    * 
+    * \sa ArrayBase::pow()
+    */
+  template<typename Derived>
+  inline const Eigen::CwiseBinaryOp<Eigen::internal::scalar_binary_pow_op<typename Derived::Scalar, typename Derived::Scalar>, const typename Derived::ConstantReturnType, const Derived>
+  pow(const typename Derived::Scalar& x, const Eigen::ArrayBase<Derived>& exponents) 
+  {
+    typename Derived::ConstantReturnType constant_x(exponents.rows(), exponents.cols(), x);
+    return Eigen::CwiseBinaryOp<Eigen::internal::scalar_binary_pow_op<typename Derived::Scalar, typename Derived::Scalar>, const typename Derived::ConstantReturnType, const Derived>(
+      constant_x,
+      exponents.derived()
+    );
+  }
+  
   /**
   * \brief Component-wise division of a scalar by array elements.
   **/

Eigen/src/Core/IO.h

@@ -49,7 +49,7 @@ std::ostream & print_matrix(std::ostream & s, const Derived& _m, const IOFormat&
   */
 struct IOFormat
 {
-  /** Default contructor, see class IOFormat for the meaning of the parameters */
+  /** Default constructor, see class IOFormat for the meaning of the parameters */
   IOFormat(int _precision = StreamPrecision, int _flags = 0,
     const std::string& _coeffSeparator = " ",
     const std::string& _rowSeparator = "\n", const std::string& _rowPrefix="", const std::string& _rowSuffix="",
@@ -57,6 +57,10 @@ struct IOFormat
   : matPrefix(_matPrefix), matSuffix(_matSuffix), rowPrefix(_rowPrefix), rowSuffix(_rowSuffix), rowSeparator(_rowSeparator),
     rowSpacer(""), coeffSeparator(_coeffSeparator), precision(_precision), flags(_flags)
   {
+    // TODO check if rowPrefix, rowSuffix or rowSeparator contains a newline
+    // don't add rowSpacer if columns are not to be aligned
+    if((flags & DontAlignCols))
+      return;
     int i = int(matSuffix.length())-1;
     while (i>=0 && matSuffix[i]!='\n')
     {
@@ -160,7 +164,6 @@ std::ostream & print_matrix(std::ostream & s, const Derived& _m, const IOFormat&
   
   typename Derived::Nested m = _m;
   typedef typename Derived::Scalar Scalar;
-  typedef typename Derived::Index Index;
 
   Index width = 0;
 
@@ -185,21 +188,22 @@ std::ostream & print_matrix(std::ostream & s, const Derived& _m, const IOFormat&
     explicit_precision = fmt.precision;
   }
 
+  std::streamsize old_precision = 0;
+  if(explicit_precision) old_precision = s.precision(explicit_precision);
+
   bool align_cols = !(fmt.flags & DontAlignCols);
   if(align_cols)
   {
     // compute the largest width
-    for(Index j = 1; j < m.cols(); ++j)
+    for(Index j = 0; j < m.cols(); ++j)
       for(Index i = 0; i < m.rows(); ++i)
       {
         std::stringstream sstr;
-        if(explicit_precision) sstr.precision(explicit_precision);
+        sstr.copyfmt(s);
         sstr << m.coeff(i,j);
         width = std::max<Index>(width, Index(sstr.str().length()));
       }
   }
-  std::streamsize old_precision = 0;
-  if(explicit_precision) old_precision = s.precision(explicit_precision);
   s << fmt.matPrefix;
   for(Index i = 0; i < m.rows(); ++i)
   {

Eigen/src/Core/Inverse.h

@@ -0,0 +1,126 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_INVERSE_H
+#define EIGEN_INVERSE_H
+
+namespace Eigen { 
+
+// TODO move the general declaration in Core, and rename this file DenseInverseImpl.h, or something like this...
+
+template<typename XprType,typename StorageKind> class InverseImpl;
+
+namespace internal {
+
+template<typename XprType>
+struct traits<Inverse<XprType> >
+  : traits<typename XprType::PlainObject>
+{
+  typedef typename XprType::PlainObject PlainObject;
+  typedef traits<PlainObject> BaseTraits;
+  enum {
+    Flags = BaseTraits::Flags & RowMajorBit
+  };
+};
+
+} // end namespace internal
+
+/** \class Inverse
+  *
+  * \brief Expression of the inverse of another expression
+  *
+  * \tparam XprType the type of the expression we are taking the inverse
+  *
+  * This class represents an abstract expression of A.inverse()
+  * and most of the time this is the only way it is used.
+  *
+  */
+template<typename XprType>
+class Inverse : public InverseImpl<XprType,typename internal::traits<XprType>::StorageKind>
+{
+public:
+  typedef typename XprType::StorageIndex StorageIndex;
+  typedef typename XprType::PlainObject                       PlainObject;
+  typedef typename internal::ref_selector<XprType>::type      XprTypeNested;
+  typedef typename internal::remove_all<XprTypeNested>::type  XprTypeNestedCleaned;
+  
+  explicit Inverse(const XprType &xpr)
+    : m_xpr(xpr)
+  {}
+  
+  EIGEN_DEVICE_FUNC Index rows() const { return m_xpr.rows(); }
+  EIGEN_DEVICE_FUNC Index cols() const { return m_xpr.cols(); }
+
+  EIGEN_DEVICE_FUNC const XprTypeNestedCleaned& nestedExpression() const { return m_xpr; }
+
+protected:
+  XprTypeNested m_xpr;
+};
+
+/** \internal
+  * Specialization of the Inverse expression for dense expressions.
+  * Direct access to the coefficients are discared.
+  * FIXME this intermediate class is probably not needed anymore.
+  */
+template<typename XprType>
+class InverseImpl<XprType,Dense>
+  : public MatrixBase<Inverse<XprType> >
+{
+  typedef Inverse<XprType> Derived;
+  
+public:
+  
+  typedef MatrixBase<Derived> Base;
+  EIGEN_DENSE_PUBLIC_INTERFACE(Derived)
+  typedef typename internal::remove_all<XprType>::type NestedExpression;
+
+private:
+  
+  Scalar coeff(Index row, Index col) const;
+  Scalar coeff(Index i) const;
+};
+
+namespace internal {
+
+/** \internal
+  * \brief Default evaluator for Inverse expression.
+  * 
+  * This default evaluator for Inverse expression simply evaluate the inverse into a temporary
+  * by a call to internal::call_assignment_no_alias.
+  * Therefore, inverse implementers only have to specialize Assignment<Dst,Inverse<...>, ...> for
+  * there own nested expression.
+  *
+  * \sa class Inverse
+  */
+template<typename ArgType>
+struct unary_evaluator<Inverse<ArgType> >
+  : public evaluator<typename Inverse<ArgType>::PlainObject>
+{
+  typedef Inverse<ArgType> InverseType;
+  typedef typename InverseType::PlainObject PlainObject;
+  typedef evaluator<PlainObject> Base;
+  
+  enum { Flags = Base::Flags | EvalBeforeNestingBit };
+
+  unary_evaluator(const InverseType& inv_xpr)
+    : m_result(inv_xpr.rows(), inv_xpr.cols())
+  {
+    ::new (static_cast<Base*>(this)) Base(m_result);
+    internal::call_assignment_no_alias(m_result, inv_xpr);
+  }
+  
+protected:
+  PlainObject m_result;
+};
+  
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_INVERSE_H

Eigen/src/Core/Map.h

@@ -19,7 +19,7 @@ namespace Eigen {
   * \brief A matrix or vector expression mapping an existing array of data.
   *
   * \tparam PlainObjectType the equivalent matrix type of the mapped data
-  * \tparam MapOptions specifies whether the pointer is \c #Aligned, or \c #Unaligned.
+  * \tparam MapOptions specifies the pointer alignment in bytes. It can be: \c #Aligned128, , \c #Aligned64, \c #Aligned32, \c #Aligned16, \c #Aligned8 or \c #Unaligned.
   *                The default is \c #Unaligned.
   * \tparam StrideType optionally specifies strides. By default, Map assumes the memory layout
   *                   of an ordinary, contiguous array. This can be overridden by specifying strides.
@@ -70,8 +70,6 @@ struct traits<Map<PlainObjectType, MapOptions, StrideType> >
   : public traits<PlainObjectType>
 {
   typedef traits<PlainObjectType> TraitsBase;
-  typedef typename PlainObjectType::Index Index;
-  typedef typename PlainObjectType::Scalar Scalar;
   enum {
     InnerStrideAtCompileTime = StrideType::InnerStrideAtCompileTime == 0
                              ? int(PlainObjectType::InnerStrideAtCompileTime)
@@ -79,22 +77,9 @@ struct traits<Map<PlainObjectType, MapOptions, StrideType> >
     OuterStrideAtCompileTime = StrideType::OuterStrideAtCompileTime == 0
                              ? int(PlainObjectType::OuterStrideAtCompileTime)
                              : int(StrideType::OuterStrideAtCompileTime),
-    HasNoInnerStride = InnerStrideAtCompileTime == 1,
-    HasNoOuterStride = StrideType::OuterStrideAtCompileTime == 0,
-    HasNoStride = HasNoInnerStride && HasNoOuterStride,
-    IsAligned = bool(EIGEN_ALIGN) && ((int(MapOptions)&Aligned)==Aligned),
-    IsDynamicSize = PlainObjectType::SizeAtCompileTime==Dynamic,
-    KeepsPacketAccess = bool(HasNoInnerStride)
-                        && ( bool(IsDynamicSize)
-                           || HasNoOuterStride
-                           || ( OuterStrideAtCompileTime!=Dynamic
-                           && ((static_cast<int>(sizeof(Scalar))*OuterStrideAtCompileTime)%16)==0 ) ),
+    Alignment = int(MapOptions)&int(AlignedMask),
     Flags0 = TraitsBase::Flags & (~NestByRefBit),
-    Flags1 = IsAligned ? (int(Flags0) | AlignedBit) : (int(Flags0) & ~AlignedBit),
-    Flags2 = (bool(HasNoStride) || bool(PlainObjectType::IsVectorAtCompileTime))
-           ? int(Flags1) : int(Flags1 & ~LinearAccessBit),
-    Flags3 = is_lvalue<PlainObjectType>::value ? int(Flags2) : (int(Flags2) & ~LvalueBit),
-    Flags = KeepsPacketAccess ? int(Flags3) : (int(Flags3) & ~PacketAccessBit)
+    Flags = is_lvalue<PlainObjectType>::value ? int(Flags0) : (int(Flags0) & ~LvalueBit)
   };
 private:
   enum { Options }; // Expressions don't have Options
@@ -110,19 +95,17 @@ template<typename PlainObjectType, int MapOptions, typename StrideType> class Ma
     EIGEN_DENSE_PUBLIC_INTERFACE(Map)
 
     typedef typename Base::PointerType PointerType;
-#if EIGEN2_SUPPORT_STAGE <= STAGE30_FULL_EIGEN3_API
-    typedef const Scalar* PointerArgType;
-    inline PointerType cast_to_pointer_type(PointerArgType ptr) { return const_cast<PointerType>(ptr); }
-#else
     typedef PointerType PointerArgType;
+    EIGEN_DEVICE_FUNC
     inline PointerType cast_to_pointer_type(PointerArgType ptr) { return ptr; }
-#endif
 
+    EIGEN_DEVICE_FUNC
     inline Index innerStride() const
     {
       return StrideType::InnerStrideAtCompileTime != 0 ? m_stride.inner() : 1;
     }
 
+    EIGEN_DEVICE_FUNC
     inline Index outerStride() const
     {
       return StrideType::OuterStrideAtCompileTime != 0 ? m_stride.outer()
@@ -134,10 +117,11 @@ template<typename PlainObjectType, int MapOptions, typename StrideType> class Ma
     /** Constructor in the fixed-size case.
       *
       * \param dataPtr pointer to the array to map
-      * \param a_stride optional Stride object, passing the strides.
+      * \param stride optional Stride object, passing the strides.
       */
-    inline Map(PointerArgType dataPtr, const StrideType& a_stride = StrideType())
-      : Base(cast_to_pointer_type(dataPtr)), m_stride(a_stride)
+    EIGEN_DEVICE_FUNC
+    explicit inline Map(PointerArgType dataPtr, const StrideType& stride = StrideType())
+      : Base(cast_to_pointer_type(dataPtr)), m_stride(stride)
     {
       PlainObjectType::Base::_check_template_params();
     }
@@ -145,11 +129,12 @@ template<typename PlainObjectType, int MapOptions, typename StrideType> class Ma
     /** Constructor in the dynamic-size vector case.
       *
       * \param dataPtr pointer to the array to map
-      * \param a_size the size of the vector expression
-      * \param a_stride optional Stride object, passing the strides.
+      * \param size the size of the vector expression
+      * \param stride optional Stride object, passing the strides.
       */
-    inline Map(PointerArgType dataPtr, Index a_size, const StrideType& a_stride = StrideType())
-      : Base(cast_to_pointer_type(dataPtr), a_size), m_stride(a_stride)
+    EIGEN_DEVICE_FUNC
+    inline Map(PointerArgType dataPtr, Index size, const StrideType& stride = StrideType())
+      : Base(cast_to_pointer_type(dataPtr), size), m_stride(stride)
     {
       PlainObjectType::Base::_check_template_params();
     }
@@ -157,12 +142,13 @@ template<typename PlainObjectType, int MapOptions, typename StrideType> class Ma
     /** Constructor in the dynamic-size matrix case.
       *
       * \param dataPtr pointer to the array to map
-      * \param nbRows the number of rows of the matrix expression
-      * \param nbCols the number of columns of the matrix expression
-      * \param a_stride optional Stride object, passing the strides.
+      * \param rows the number of rows of the matrix expression
+      * \param cols the number of columns of the matrix expression
+      * \param stride optional Stride object, passing the strides.
       */
-    inline Map(PointerArgType dataPtr, Index nbRows, Index nbCols, const StrideType& a_stride = StrideType())
-      : Base(cast_to_pointer_type(dataPtr), nbRows, nbCols), m_stride(a_stride)
+    EIGEN_DEVICE_FUNC
+    inline Map(PointerArgType dataPtr, Index rows, Index cols, const StrideType& stride = StrideType())
+      : Base(cast_to_pointer_type(dataPtr), rows, cols), m_stride(stride)
     {
       PlainObjectType::Base::_check_template_params();
     }
@@ -173,19 +159,6 @@ template<typename PlainObjectType, int MapOptions, typename StrideType> class Ma
     StrideType m_stride;
 };
 
-template<typename _Scalar, int _Rows, int _Cols, int _Options, int _MaxRows, int _MaxCols>
-inline Array<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols>
-  ::Array(const Scalar *data)
-{
-  this->_set_noalias(Eigen::Map<const Array>(data));
-}
-
-template<typename _Scalar, int _Rows, int _Cols, int _Options, int _MaxRows, int _MaxCols>
-inline Matrix<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols>
-  ::Matrix(const Scalar *data)
-{
-  this->_set_noalias(Eigen::Map<const Matrix>(data));
-}
 
 } // end namespace Eigen
 

Eigen/src/Core/MapBase.h

@@ -12,7 +12,7 @@
 #define EIGEN_MAPBASE_H
 
 #define EIGEN_STATIC_ASSERT_INDEX_BASED_ACCESS(Derived) \
-      EIGEN_STATIC_ASSERT((int(internal::traits<Derived>::Flags) & LinearAccessBit) || Derived::IsVectorAtCompileTime, \
+      EIGEN_STATIC_ASSERT((int(internal::evaluator<Derived>::Flags) & LinearAccessBit) || Derived::IsVectorAtCompileTime, \
                           YOU_ARE_TRYING_TO_USE_AN_INDEX_BASED_ACCESSOR_ON_AN_EXPRESSION_THAT_DOES_NOT_SUPPORT_THAT)
 
 namespace Eigen { 
@@ -37,7 +37,6 @@ template<typename Derived> class MapBase<Derived, ReadOnlyAccessors>
     };
 
     typedef typename internal::traits<Derived>::StorageKind StorageKind;
-    typedef typename internal::traits<Derived>::Index Index;
     typedef typename internal::traits<Derived>::Scalar Scalar;
     typedef typename internal::packet_traits<Scalar>::type PacketScalar;
     typedef typename NumTraits<Scalar>::Real RealScalar;
@@ -76,8 +75,8 @@ template<typename Derived> class MapBase<Derived, ReadOnlyAccessors>
 
     typedef typename Base::CoeffReturnType CoeffReturnType;
 
-    inline Index rows() const { return m_rows.value(); }
-    inline Index cols() const { return m_cols.value(); }
+    EIGEN_DEVICE_FUNC inline Index rows() const { return m_rows.value(); }
+    EIGEN_DEVICE_FUNC inline Index cols() const { return m_cols.value(); }
 
     /** Returns a pointer to the first coefficient of the matrix or vector.
       *
@@ -85,24 +84,28 @@ template<typename Derived> class MapBase<Derived, ReadOnlyAccessors>
       *
       * \sa innerStride(), outerStride()
       */
-    inline const Scalar* data() const { return m_data; }
+    EIGEN_DEVICE_FUNC inline const Scalar* data() const { return m_data; }
 
+    EIGEN_DEVICE_FUNC
     inline const Scalar& coeff(Index rowId, Index colId) const
     {
       return m_data[colId * colStride() + rowId * rowStride()];
     }
 
+    EIGEN_DEVICE_FUNC
     inline const Scalar& coeff(Index index) const
     {
       EIGEN_STATIC_ASSERT_INDEX_BASED_ACCESS(Derived)
       return m_data[index * innerStride()];
     }
 
+    EIGEN_DEVICE_FUNC
     inline const Scalar& coeffRef(Index rowId, Index colId) const
     {
       return this->m_data[colId * colStride() + rowId * rowStride()];
     }
 
+    EIGEN_DEVICE_FUNC
     inline const Scalar& coeffRef(Index index) const
     {
       EIGEN_STATIC_ASSERT_INDEX_BASED_ACCESS(Derived)
@@ -123,12 +126,14 @@ template<typename Derived> class MapBase<Derived, ReadOnlyAccessors>
       return internal::ploadt<PacketScalar, LoadMode>(m_data + index * innerStride());
     }
 
-    inline MapBase(PointerType dataPtr) : m_data(dataPtr), m_rows(RowsAtCompileTime), m_cols(ColsAtCompileTime)
+    EIGEN_DEVICE_FUNC
+    explicit inline MapBase(PointerType dataPtr) : m_data(dataPtr), m_rows(RowsAtCompileTime), m_cols(ColsAtCompileTime)
     {
       EIGEN_STATIC_ASSERT_FIXED_SIZE(Derived)
       checkSanity();
     }
 
+    EIGEN_DEVICE_FUNC
     inline MapBase(PointerType dataPtr, Index vecSize)
             : m_data(dataPtr),
               m_rows(RowsAtCompileTime == Dynamic ? vecSize : Index(RowsAtCompileTime)),
@@ -140,24 +145,24 @@ template<typename Derived> class MapBase<Derived, ReadOnlyAccessors>
       checkSanity();
     }
 
-    inline MapBase(PointerType dataPtr, Index nbRows, Index nbCols)
-            : m_data(dataPtr), m_rows(nbRows), m_cols(nbCols)
+    EIGEN_DEVICE_FUNC
+    inline MapBase(PointerType dataPtr, Index rows, Index cols)
+            : m_data(dataPtr), m_rows(rows), m_cols(cols)
     {
       eigen_assert( (dataPtr == 0)
-              || (   nbRows >= 0 && (RowsAtCompileTime == Dynamic || RowsAtCompileTime == nbRows)
-                  && nbCols >= 0 && (ColsAtCompileTime == Dynamic || ColsAtCompileTime == nbCols)));
+              || (   rows >= 0 && (RowsAtCompileTime == Dynamic || RowsAtCompileTime == rows)
+                  && cols >= 0 && (ColsAtCompileTime == Dynamic || ColsAtCompileTime == cols)));
       checkSanity();
     }
 
   protected:
 
+    EIGEN_DEVICE_FUNC
     void checkSanity() const
     {
-      EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(internal::traits<Derived>::Flags&PacketAccessBit,
-                                        internal::inner_stride_at_compile_time<Derived>::ret==1),
-                          PACKET_ACCESS_REQUIRES_TO_HAVE_INNER_STRIDE_FIXED_TO_1);
-      eigen_assert(EIGEN_IMPLIES(internal::traits<Derived>::Flags&AlignedBit, (size_t(m_data) % 16) == 0)
-                   && "data is not aligned");
+#if EIGEN_MAX_ALIGN_BYTES>0
+      eigen_assert(((size_t(m_data) % EIGEN_PLAIN_ENUM_MAX(1,internal::traits<Derived>::Alignment)) == 0) && "data is not aligned");
+#endif
     }
 
     PointerType m_data;
@@ -168,13 +173,14 @@ template<typename Derived> class MapBase<Derived, ReadOnlyAccessors>
 template<typename Derived> class MapBase<Derived, WriteAccessors>
   : public MapBase<Derived, ReadOnlyAccessors>
 {
+    typedef MapBase<Derived, ReadOnlyAccessors> ReadOnlyMapBase;
   public:
 
     typedef MapBase<Derived, ReadOnlyAccessors> Base;
 
     typedef typename Base::Scalar Scalar;
     typedef typename Base::PacketScalar PacketScalar;
-    typedef typename Base::Index Index;
+    typedef typename Base::StorageIndex StorageIndex;
     typedef typename Base::PointerType PointerType;
 
     using Base::derived;
@@ -195,14 +201,18 @@ template<typename Derived> class MapBase<Derived, WriteAccessors>
                     const Scalar
                   >::type ScalarWithConstIfNotLvalue;
 
+    EIGEN_DEVICE_FUNC
     inline const Scalar* data() const { return this->m_data; }
+    EIGEN_DEVICE_FUNC
     inline ScalarWithConstIfNotLvalue* data() { return this->m_data; } // no const-cast here so non-const-correct code will give a compile error
 
+    EIGEN_DEVICE_FUNC
     inline ScalarWithConstIfNotLvalue& coeffRef(Index row, Index col)
     {
       return this->m_data[col * colStride() + row * rowStride()];
     }
 
+    EIGEN_DEVICE_FUNC
     inline ScalarWithConstIfNotLvalue& coeffRef(Index index)
     {
       EIGEN_STATIC_ASSERT_INDEX_BASED_ACCESS(Derived)
@@ -224,19 +234,24 @@ template<typename Derived> class MapBase<Derived, WriteAccessors>
                 (this->m_data + index * innerStride(), val);
     }
 
-    explicit inline MapBase(PointerType dataPtr) : Base(dataPtr) {}
-    inline MapBase(PointerType dataPtr, Index vecSize) : Base(dataPtr, vecSize) {}
-    inline MapBase(PointerType dataPtr, Index nbRows, Index nbCols) : Base(dataPtr, nbRows, nbCols) {}
+    EIGEN_DEVICE_FUNC explicit inline MapBase(PointerType dataPtr) : Base(dataPtr) {}
+    EIGEN_DEVICE_FUNC inline MapBase(PointerType dataPtr, Index vecSize) : Base(dataPtr, vecSize) {}
+    EIGEN_DEVICE_FUNC inline MapBase(PointerType dataPtr, Index rows, Index cols) : Base(dataPtr, rows, cols) {}
 
+    EIGEN_DEVICE_FUNC
     Derived& operator=(const MapBase& other)
     {
-      Base::Base::operator=(other);
+      ReadOnlyMapBase::Base::operator=(other);
       return derived();
     }
 
-    using Base::Base::operator=;
+    // In theory we could simply refer to Base:Base::operator=, but MSVC does not like Base::Base,
+    // see bugs 821 and 920.
+    using ReadOnlyMapBase::Base::operator=;
 };
 
+#undef EIGEN_STATIC_ASSERT_INDEX_BASED_ACCESS
+
 } // end namespace Eigen
 
 #endif // EIGEN_MAPBASE_H

Eigen/src/Core/MathFunctions.h

@@ -10,8 +10,20 @@
 #ifndef EIGEN_MATHFUNCTIONS_H
 #define EIGEN_MATHFUNCTIONS_H
 
+// source: http://www.geom.uiuc.edu/~huberty/math5337/groupe/digits.html
+#define EIGEN_PI 3.141592653589793238462643383279502884197169399375105820974944592307816406
+
 namespace Eigen {
 
+// On WINCE, std::abs is defined for int only, so let's defined our own overloads:
+// This issue has been confirmed with MSVC 2008 only, but the issue might exist for more recent versions too.
+#if EIGEN_OS_WINCE && EIGEN_COMP_MSVC && EIGEN_COMP_MSVC<=1500
+long        abs(long        x) { return (labs(x));  }
+double      abs(double      x) { return (fabs(x));  }
+float       abs(float       x) { return (fabsf(x)); }
+long double abs(long double x) { return (fabsl(x)); }
+#endif
+  
 namespace internal {
 
 /** \internal \struct global_math_functions_filtering_base
@@ -62,6 +74,7 @@ template<typename Scalar, bool IsComplex = NumTraits<Scalar>::IsComplex>
 struct real_default_impl
 {
   typedef typename NumTraits<Scalar>::Real RealScalar;
+  EIGEN_DEVICE_FUNC
   static inline RealScalar run(const Scalar& x)
   {
     return x;
@@ -72,6 +85,7 @@ template<typename Scalar>
 struct real_default_impl<Scalar,true>
 {
   typedef typename NumTraits<Scalar>::Real RealScalar;
+  EIGEN_DEVICE_FUNC
   static inline RealScalar run(const Scalar& x)
   {
     using std::real;
@@ -87,7 +101,6 @@ struct real_retval
   typedef typename NumTraits<Scalar>::Real type;
 };
 
-
 /****************************************************************************
 * Implementation of imag                                                 *
 ****************************************************************************/
@@ -96,6 +109,7 @@ template<typename Scalar, bool IsComplex = NumTraits<Scalar>::IsComplex>
 struct imag_default_impl
 {
   typedef typename NumTraits<Scalar>::Real RealScalar;
+  EIGEN_DEVICE_FUNC
   static inline RealScalar run(const Scalar&)
   {
     return RealScalar(0);
@@ -106,6 +120,7 @@ template<typename Scalar>
 struct imag_default_impl<Scalar,true>
 {
   typedef typename NumTraits<Scalar>::Real RealScalar;
+  EIGEN_DEVICE_FUNC
   static inline RealScalar run(const Scalar& x)
   {
     using std::imag;
@@ -129,10 +144,12 @@ template<typename Scalar>
 struct real_ref_impl
 {
   typedef typename NumTraits<Scalar>::Real RealScalar;
+  EIGEN_DEVICE_FUNC
   static inline RealScalar& run(Scalar& x)
   {
     return reinterpret_cast<RealScalar*>(&x)[0];
   }
+  EIGEN_DEVICE_FUNC
   static inline const RealScalar& run(const Scalar& x)
   {
     return reinterpret_cast<const RealScalar*>(&x)[0];
@@ -153,10 +170,12 @@ template<typename Scalar, bool IsComplex>
 struct imag_ref_default_impl
 {
   typedef typename NumTraits<Scalar>::Real RealScalar;
+  EIGEN_DEVICE_FUNC
   static inline RealScalar& run(Scalar& x)
   {
     return reinterpret_cast<RealScalar*>(&x)[1];
   }
+  EIGEN_DEVICE_FUNC
   static inline const RealScalar& run(const Scalar& x)
   {
     return reinterpret_cast<RealScalar*>(&x)[1];
@@ -166,10 +185,12 @@ struct imag_ref_default_impl
 template<typename Scalar>
 struct imag_ref_default_impl<Scalar, false>
 {
+  EIGEN_DEVICE_FUNC
   static inline Scalar run(Scalar&)
   {
     return Scalar(0);
   }
+  EIGEN_DEVICE_FUNC
   static inline const Scalar run(const Scalar&)
   {
     return Scalar(0);
@@ -192,6 +213,7 @@ struct imag_ref_retval
 template<typename Scalar, bool IsComplex = NumTraits<Scalar>::IsComplex>
 struct conj_impl
 {
+  EIGEN_DEVICE_FUNC
   static inline Scalar run(const Scalar& x)
   {
     return x;
@@ -201,6 +223,7 @@ struct conj_impl
 template<typename Scalar>
 struct conj_impl<Scalar,true>
 {
+  EIGEN_DEVICE_FUNC
   static inline Scalar run(const Scalar& x)
   {
     using std::conj;
@@ -222,6 +245,7 @@ template<typename Scalar>
 struct abs2_impl
 {
   typedef typename NumTraits<Scalar>::Real RealScalar;
+  EIGEN_DEVICE_FUNC
   static inline RealScalar run(const Scalar& x)
   {
     return x*x;
@@ -231,6 +255,7 @@ struct abs2_impl
 template<typename RealScalar>
 struct abs2_impl<std::complex<RealScalar> >
 {
+  EIGEN_DEVICE_FUNC
   static inline RealScalar run(const std::complex<RealScalar>& x)
   {
     return real(x)*real(x) + imag(x)*imag(x);
@@ -251,9 +276,10 @@ template<typename Scalar, bool IsComplex>
 struct norm1_default_impl
 {
   typedef typename NumTraits<Scalar>::Real RealScalar;
+  EIGEN_DEVICE_FUNC
   static inline RealScalar run(const Scalar& x)
   {
-    using std::abs;
+    EIGEN_USING_STD_MATH(abs);
     return abs(real(x)) + abs(imag(x));
   }
 };
@@ -261,9 +287,10 @@ struct norm1_default_impl
 template<typename Scalar>
 struct norm1_default_impl<Scalar, false>
 {
+  EIGEN_DEVICE_FUNC
   static inline Scalar run(const Scalar& x)
   {
-    using std::abs;
+    EIGEN_USING_STD_MATH(abs);
     return abs(x);
   }
 };
@@ -287,16 +314,24 @@ struct hypot_impl
   typedef typename NumTraits<Scalar>::Real RealScalar;
   static inline RealScalar run(const Scalar& x, const Scalar& y)
   {
-    using std::max;
-    using std::min;
-    using std::abs;
-    using std::sqrt;
+    EIGEN_USING_STD_MATH(max);
+    EIGEN_USING_STD_MATH(min);
+    EIGEN_USING_STD_MATH(abs);
+    EIGEN_USING_STD_MATH(sqrt);
     RealScalar _x = abs(x);
     RealScalar _y = abs(y);
-    RealScalar p = (max)(_x, _y);
-    if(p==RealScalar(0)) return 0;
-    RealScalar q = (min)(_x, _y);
-    RealScalar qp = q/p;
+    Scalar p, qp;
+    if(_x>_y)
+    {
+      p = _x;
+      qp = _y / p;
+    }
+    else
+    {
+      p = _y;
+      qp = _x / p;
+    }
+    if(p==RealScalar(0)) return RealScalar(0);
     return p * sqrt(RealScalar(1) + qp*qp);
   }
 };
@@ -314,6 +349,7 @@ struct hypot_retval
 template<typename OldType, typename NewType>
 struct cast_impl
 {
+  EIGEN_DEVICE_FUNC
   static inline NewType run(const OldType& x)
   {
     return static_cast<NewType>(x);
@@ -323,48 +359,121 @@ struct cast_impl
 // here, for once, we're plainly returning NewType: we don't want cast to do weird things.
 
 template<typename OldType, typename NewType>
+EIGEN_DEVICE_FUNC
 inline NewType cast(const OldType& x)
 {
   return cast_impl<OldType, NewType>::run(x);
 }
 
 /****************************************************************************
-* Implementation of atanh2                                                *
+* Implementation of round                                                   *
 ****************************************************************************/
 
-template<typename Scalar, bool IsInteger>
-struct atanh2_default_impl
-{
-  typedef Scalar retval;
-  typedef typename NumTraits<Scalar>::Real RealScalar;
-  static inline Scalar run(const Scalar& x, const Scalar& y)
+#if EIGEN_HAS_CXX11_MATH
+  template<typename Scalar>
+  struct round_impl {
+    static inline Scalar run(const Scalar& x)
+    {
+      EIGEN_STATIC_ASSERT((!NumTraits<Scalar>::IsComplex), NUMERIC_TYPE_MUST_BE_REAL)
+      using std::round;
+      return round(x);
+    }
+  };
+#else
+  template<typename Scalar>
+  struct round_impl
   {
-    using std::abs;
-    using std::log;
-    using std::sqrt;
-    Scalar z = x / y;
-    if (y == Scalar(0) || abs(z) > sqrt(NumTraits<RealScalar>::epsilon()))
-      return RealScalar(0.5) * log((y + x) / (y - x));
-    else
-      return z + z*z*z / RealScalar(3);
-  }
-};
+    static inline Scalar run(const Scalar& x)
+    {
+      EIGEN_STATIC_ASSERT((!NumTraits<Scalar>::IsComplex), NUMERIC_TYPE_MUST_BE_REAL)
+      EIGEN_USING_STD_MATH(floor);
+      EIGEN_USING_STD_MATH(ceil);
+      return (x > Scalar(0)) ? floor(x + Scalar(0.5)) : ceil(x - Scalar(0.5));
+    }
+  };
+#endif
 
 template<typename Scalar>
-struct atanh2_default_impl<Scalar, true>
+struct round_retval
 {
-  static inline Scalar run(const Scalar&, const Scalar&)
+  typedef Scalar type;
+};
+
+/****************************************************************************
+* Implementation of arg                                                     *
+****************************************************************************/
+
+#if EIGEN_HAS_CXX11_MATH
+  template<typename Scalar>
+  struct arg_impl {
+    static inline Scalar run(const Scalar& x)
+    {
+      EIGEN_USING_STD_MATH(arg);
+      return arg(x);
+    }
+  };
+#else
+  template<typename Scalar, bool IsComplex = NumTraits<Scalar>::IsComplex>
+  struct arg_default_impl
+  {
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+    EIGEN_DEVICE_FUNC
+    static inline RealScalar run(const Scalar& x)
+    {
+      return (x < Scalar(0)) ? Scalar(EIGEN_PI) : Scalar(0); }
+  };
+
+  template<typename Scalar>
+  struct arg_default_impl<Scalar,true>
+  {
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+    EIGEN_DEVICE_FUNC
+    static inline RealScalar run(const Scalar& x)
+    {
+      EIGEN_USING_STD_MATH(arg);
+      return arg(x);
+    }
+  };
+
+  template<typename Scalar> struct arg_impl : arg_default_impl<Scalar> {};
+#endif
+
+template<typename Scalar>
+struct arg_retval
+{
+  typedef typename NumTraits<Scalar>::Real type;
+};
+
+/****************************************************************************
+* Implementation of log1p                                                   *
+****************************************************************************/
+template<typename Scalar, bool isComplex = NumTraits<Scalar>::IsComplex >
+struct log1p_impl
+{
+  static inline Scalar run(const Scalar& x)
   {
     EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar)
-    return Scalar(0);
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+    EIGEN_USING_STD_MATH(log);
+    Scalar x1p = RealScalar(1) + x;
+    return ( x1p == Scalar(1) ) ? x : x * ( log(x1p) / (x1p - RealScalar(1)) );
   }
 };
 
+#if EIGEN_HAS_CXX11_MATH
 template<typename Scalar>
-struct atanh2_impl : atanh2_default_impl<Scalar, NumTraits<Scalar>::IsInteger> {};
+struct log1p_impl<Scalar, false> {
+  static inline Scalar run(const Scalar& x)
+  {
+    EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar)
+    using std::log1p;
+    return log1p(x);
+  }
+};
+#endif
 
 template<typename Scalar>
-struct atanh2_retval
+struct log1p_retval
 {
   typedef Scalar type;
 };
@@ -379,7 +488,7 @@ struct pow_default_impl
   typedef Scalar retval;
   static inline Scalar run(const Scalar& x, const Scalar& y)
   {
-    using std::pow;
+    EIGEN_USING_STD_MATH(pow);
     return pow(x, y);
   }
 };
@@ -447,48 +556,48 @@ struct random_default_impl<Scalar, false, false>
 };
 
 enum {
-  floor_log2_terminate,
-  floor_log2_move_up,
-  floor_log2_move_down,
-  floor_log2_bogus
+  meta_floor_log2_terminate,
+  meta_floor_log2_move_up,
+  meta_floor_log2_move_down,
+  meta_floor_log2_bogus
 };
 
-template<unsigned int n, int lower, int upper> struct floor_log2_selector
+template<unsigned int n, int lower, int upper> struct meta_floor_log2_selector
 {
   enum { middle = (lower + upper) / 2,
-         value = (upper <= lower + 1) ? int(floor_log2_terminate)
-               : (n < (1 << middle)) ? int(floor_log2_move_down)
-               : (n==0) ? int(floor_log2_bogus)
-               : int(floor_log2_move_up)
+         value = (upper <= lower + 1) ? int(meta_floor_log2_terminate)
+               : (n < (1 << middle)) ? int(meta_floor_log2_move_down)
+               : (n==0) ? int(meta_floor_log2_bogus)
+               : int(meta_floor_log2_move_up)
   };
 };
 
 template<unsigned int n,
          int lower = 0,
          int upper = sizeof(unsigned int) * CHAR_BIT - 1,
-         int selector = floor_log2_selector<n, lower, upper>::value>
-struct floor_log2 {};
+         int selector = meta_floor_log2_selector<n, lower, upper>::value>
+struct meta_floor_log2 {};
 
 template<unsigned int n, int lower, int upper>
-struct floor_log2<n, lower, upper, floor_log2_move_down>
+struct meta_floor_log2<n, lower, upper, meta_floor_log2_move_down>
 {
-  enum { value = floor_log2<n, lower, floor_log2_selector<n, lower, upper>::middle>::value };
+  enum { value = meta_floor_log2<n, lower, meta_floor_log2_selector<n, lower, upper>::middle>::value };
 };
 
 template<unsigned int n, int lower, int upper>
-struct floor_log2<n, lower, upper, floor_log2_move_up>
+struct meta_floor_log2<n, lower, upper, meta_floor_log2_move_up>
 {
-  enum { value = floor_log2<n, floor_log2_selector<n, lower, upper>::middle, upper>::value };
+  enum { value = meta_floor_log2<n, meta_floor_log2_selector<n, lower, upper>::middle, upper>::value };
 };
 
 template<unsigned int n, int lower, int upper>
-struct floor_log2<n, lower, upper, floor_log2_terminate>
+struct meta_floor_log2<n, lower, upper, meta_floor_log2_terminate>
 {
   enum { value = (n >= ((unsigned int)(1) << (lower+1))) ? lower+1 : lower };
 };
 
 template<unsigned int n, int lower, int upper>
-struct floor_log2<n, lower, upper, floor_log2_bogus>
+struct meta_floor_log2<n, lower, upper, meta_floor_log2_bogus>
 {
   // no value, error at compile time
 };
@@ -496,11 +605,24 @@ struct floor_log2<n, lower, upper, floor_log2_bogus>
 template<typename Scalar>
 struct random_default_impl<Scalar, false, true>
 {
-  typedef typename NumTraits<Scalar>::NonInteger NonInteger;
-
   static inline Scalar run(const Scalar& x, const Scalar& y)
-  {
-    return x + Scalar((NonInteger(y)-x+1) * std::rand() / (RAND_MAX + NonInteger(1)));
+  { 
+    using std::max;
+    using std::min;
+    typedef typename conditional<NumTraits<Scalar>::IsSigned,std::ptrdiff_t,std::size_t>::type ScalarX;
+    if(y<x)
+      return x;
+    std::size_t range = ScalarX(y)-ScalarX(x);
+    std::size_t offset = 0;
+    // rejection sampling
+    std::size_t divisor    = (range+RAND_MAX-1)/(range+1);
+    std::size_t multiplier = (range+RAND_MAX-1)/std::size_t(RAND_MAX);
+
+    do {
+      offset = ( (std::size_t(std::rand()) * multiplier) / divisor );
+    } while (offset > range);
+
+    return Scalar(ScalarX(x) + offset);
   }
 
   static inline Scalar run()
@@ -508,7 +630,7 @@ struct random_default_impl<Scalar, false, true>
 #ifdef EIGEN_MAKING_DOCS
     return run(Scalar(NumTraits<Scalar>::IsSigned ? -10 : 0), Scalar(10));
 #else
-    enum { rand_bits = floor_log2<(unsigned int)(RAND_MAX)+1>::value,
+    enum { rand_bits = meta_floor_log2<(unsigned int)(RAND_MAX)+1>::value,
            scalar_bits = sizeof(Scalar) * CHAR_BIT,
            shift = EIGEN_PLAIN_ENUM_MAX(0, int(rand_bits) - int(scalar_bits)),
            offset = NumTraits<Scalar>::IsSigned ? (1 << (EIGEN_PLAIN_ENUM_MIN(rand_bits,scalar_bits)-1)) : 0
@@ -548,88 +670,236 @@ inline EIGEN_MATHFUNC_RETVAL(random, Scalar) random()
 } // end namespace internal
 
 /****************************************************************************
-* Generic math function                                                    *
+* Generic math functions                                                    *
 ****************************************************************************/
 
 namespace numext {
 
+#ifndef __CUDA_ARCH__
+template<typename T>
+EIGEN_DEVICE_FUNC
+EIGEN_ALWAYS_INLINE T mini(const T& x, const T& y)
+{
+  EIGEN_USING_STD_MATH(min);
+  return min EIGEN_NOT_A_MACRO (x,y);
+}
+
+template<typename T>
+EIGEN_DEVICE_FUNC
+EIGEN_ALWAYS_INLINE T maxi(const T& x, const T& y)
+{
+  EIGEN_USING_STD_MATH(max);
+  return max EIGEN_NOT_A_MACRO (x,y);
+}
+#else
+template<typename T>
+EIGEN_DEVICE_FUNC
+EIGEN_ALWAYS_INLINE T mini(const T& x, const T& y)
+{
+  return y < x ? y : x;
+}
+template<>
+EIGEN_DEVICE_FUNC
+EIGEN_ALWAYS_INLINE float mini(const float& x, const float& y)
+{
+  return fmin(x, y);
+}
+template<typename T>
+EIGEN_DEVICE_FUNC
+EIGEN_ALWAYS_INLINE T maxi(const T& x, const T& y)
+{
+  return x < y ? y : x;
+}
+template<>
+EIGEN_DEVICE_FUNC
+EIGEN_ALWAYS_INLINE float maxi(const float& x, const float& y)
+{
+  return fmax(x, y);
+}
+#endif
+
+
 template<typename Scalar>
+EIGEN_DEVICE_FUNC
 inline EIGEN_MATHFUNC_RETVAL(real, Scalar) real(const Scalar& x)
 {
   return EIGEN_MATHFUNC_IMPL(real, Scalar)::run(x);
 }  
 
 template<typename Scalar>
+EIGEN_DEVICE_FUNC
 inline typename internal::add_const_on_value_type< EIGEN_MATHFUNC_RETVAL(real_ref, Scalar) >::type real_ref(const Scalar& x)
 {
   return internal::real_ref_impl<Scalar>::run(x);
 }
 
 template<typename Scalar>
+EIGEN_DEVICE_FUNC
 inline EIGEN_MATHFUNC_RETVAL(real_ref, Scalar) real_ref(Scalar& x)
 {
   return EIGEN_MATHFUNC_IMPL(real_ref, Scalar)::run(x);
 }
 
 template<typename Scalar>
+EIGEN_DEVICE_FUNC
 inline EIGEN_MATHFUNC_RETVAL(imag, Scalar) imag(const Scalar& x)
 {
   return EIGEN_MATHFUNC_IMPL(imag, Scalar)::run(x);
 }
 
 template<typename Scalar>
+EIGEN_DEVICE_FUNC
+inline EIGEN_MATHFUNC_RETVAL(arg, Scalar) arg(const Scalar& x)
+{
+  return EIGEN_MATHFUNC_IMPL(arg, Scalar)::run(x);
+}
+
+template<typename Scalar>
+EIGEN_DEVICE_FUNC
 inline typename internal::add_const_on_value_type< EIGEN_MATHFUNC_RETVAL(imag_ref, Scalar) >::type imag_ref(const Scalar& x)
 {
   return internal::imag_ref_impl<Scalar>::run(x);
 }
 
 template<typename Scalar>
+EIGEN_DEVICE_FUNC
 inline EIGEN_MATHFUNC_RETVAL(imag_ref, Scalar) imag_ref(Scalar& x)
 {
   return EIGEN_MATHFUNC_IMPL(imag_ref, Scalar)::run(x);
 }
 
 template<typename Scalar>
+EIGEN_DEVICE_FUNC
 inline EIGEN_MATHFUNC_RETVAL(conj, Scalar) conj(const Scalar& x)
 {
   return EIGEN_MATHFUNC_IMPL(conj, Scalar)::run(x);
 }
 
 template<typename Scalar>
+EIGEN_DEVICE_FUNC
 inline EIGEN_MATHFUNC_RETVAL(abs2, Scalar) abs2(const Scalar& x)
 {
   return EIGEN_MATHFUNC_IMPL(abs2, Scalar)::run(x);
 }
 
 template<typename Scalar>
+EIGEN_DEVICE_FUNC
 inline EIGEN_MATHFUNC_RETVAL(norm1, Scalar) norm1(const Scalar& x)
 {
   return EIGEN_MATHFUNC_IMPL(norm1, Scalar)::run(x);
 }
 
 template<typename Scalar>
+EIGEN_DEVICE_FUNC
 inline EIGEN_MATHFUNC_RETVAL(hypot, Scalar) hypot(const Scalar& x, const Scalar& y)
 {
   return EIGEN_MATHFUNC_IMPL(hypot, Scalar)::run(x, y);
 }
 
 template<typename Scalar>
-inline EIGEN_MATHFUNC_RETVAL(atanh2, Scalar) atanh2(const Scalar& x, const Scalar& y)
+EIGEN_DEVICE_FUNC
+inline EIGEN_MATHFUNC_RETVAL(log1p, Scalar) log1p(const Scalar& x)
 {
-  return EIGEN_MATHFUNC_IMPL(atanh2, Scalar)::run(x, y);
+  return EIGEN_MATHFUNC_IMPL(log1p, Scalar)::run(x);
 }
 
 template<typename Scalar>
+EIGEN_DEVICE_FUNC
 inline EIGEN_MATHFUNC_RETVAL(pow, Scalar) pow(const Scalar& x, const Scalar& y)
 {
   return EIGEN_MATHFUNC_IMPL(pow, Scalar)::run(x, y);
 }
 
-// std::isfinite is non standard, so let's define our own version,
-// even though it is not very efficient.
-template<typename T> bool (isfinite)(const T& x)
+template<typename T>
+EIGEN_DEVICE_FUNC
+bool (isfinite)(const T& x)
 {
-  return x<NumTraits<T>::highest() && x>NumTraits<T>::lowest();
+  #if EIGEN_HAS_CXX11_MATH
+    using std::isfinite;
+    return isfinite EIGEN_NOT_A_MACRO (x);
+  #else
+    return x<NumTraits<T>::highest() && x>NumTraits<T>::lowest();
+  #endif
+}
+
+template<typename T>
+EIGEN_DEVICE_FUNC
+bool (isnan)(const T& x)
+{
+  #if EIGEN_HAS_CXX11_MATH
+    using std::isnan;
+    return isnan EIGEN_NOT_A_MACRO (x);
+  #else
+    return x != x;
+  #endif
+}
+
+template<typename T>
+EIGEN_DEVICE_FUNC
+bool (isinf)(const T& x)
+{
+  #if EIGEN_HAS_CXX11_MATH
+    using std::isinf;
+    return isinf EIGEN_NOT_A_MACRO (x);
+  #else
+    return x>NumTraits<T>::highest() || x<NumTraits<T>::lowest();
+  #endif
+}
+
+template<typename T>
+bool (isfinite)(const std::complex<T>& x)
+{
+  return (numext::isfinite)(numext::real(x)) && (numext::isfinite)(numext::imag(x));
+}
+
+template<typename T>
+bool (isnan)(const std::complex<T>& x)
+{
+  return (numext::isnan)(numext::real(x)) || (numext::isnan)(numext::imag(x));
+}
+
+template<typename T>
+bool (isinf)(const std::complex<T>& x)
+{
+  return ((numext::isinf)(numext::real(x)) || (numext::isinf)(numext::imag(x))) && (!(numext::isnan)(x));
+}
+
+template<typename Scalar>
+EIGEN_DEVICE_FUNC
+inline EIGEN_MATHFUNC_RETVAL(round, Scalar) round(const Scalar& x)
+{
+  return EIGEN_MATHFUNC_IMPL(round, Scalar)::run(x);
+}
+
+template<typename T>
+EIGEN_DEVICE_FUNC
+T (floor)(const T& x)
+{
+  EIGEN_USING_STD_MATH(floor);
+  return floor(x);
+}
+
+template<typename T>
+EIGEN_DEVICE_FUNC
+T (ceil)(const T& x)
+{
+  EIGEN_USING_STD_MATH(ceil);
+  return ceil(x);
+}
+
+// Log base 2 for 32 bits positive integers.
+// Conveniently returns 0 for x==0.
+inline int log2(int x)
+{
+  eigen_assert(x>=0);
+  unsigned int v(x);
+  static const int table[32] = { 0, 9, 1, 10, 13, 21, 2, 29, 11, 14, 16, 18, 22, 25, 3, 30, 8, 12, 20, 28, 15, 17, 24, 7, 19, 27, 23, 6, 26, 5, 4, 31 };
+  v |= v >> 1;
+  v |= v >> 2;
+  v |= v >> 4;
+  v |= v >> 8;
+  v |= v >> 16;
+  return table[(v * 0x07C4ACDDU) >> 27];
 }
 
 } // end namespace numext
@@ -649,18 +919,20 @@ template<typename Scalar>
 struct scalar_fuzzy_default_impl<Scalar, false, false>
 {
   typedef typename NumTraits<Scalar>::Real RealScalar;
-  template<typename OtherScalar>
+  template<typename OtherScalar> EIGEN_DEVICE_FUNC
   static inline bool isMuchSmallerThan(const Scalar& x, const OtherScalar& y, const RealScalar& prec)
   {
-    using std::abs;
+    EIGEN_USING_STD_MATH(abs);
     return abs(x) <= abs(y) * prec;
   }
+  EIGEN_DEVICE_FUNC
   static inline bool isApprox(const Scalar& x, const Scalar& y, const RealScalar& prec)
   {
-    using std::min;
-    using std::abs;
+    EIGEN_USING_STD_MATH(min);
+    EIGEN_USING_STD_MATH(abs);
     return abs(x - y) <= (min)(abs(x), abs(y)) * prec;
   }
+  EIGEN_DEVICE_FUNC
   static inline bool isApproxOrLessThan(const Scalar& x, const Scalar& y, const RealScalar& prec)
   {
     return x <= y || isApprox(x, y, prec);
@@ -671,15 +943,17 @@ template<typename Scalar>
 struct scalar_fuzzy_default_impl<Scalar, false, true>
 {
   typedef typename NumTraits<Scalar>::Real RealScalar;
-  template<typename OtherScalar>
+  template<typename OtherScalar> EIGEN_DEVICE_FUNC
   static inline bool isMuchSmallerThan(const Scalar& x, const Scalar&, const RealScalar&)
   {
     return x == Scalar(0);
   }
+  EIGEN_DEVICE_FUNC
   static inline bool isApprox(const Scalar& x, const Scalar& y, const RealScalar&)
   {
     return x == y;
   }
+  EIGEN_DEVICE_FUNC
   static inline bool isApproxOrLessThan(const Scalar& x, const Scalar& y, const RealScalar&)
   {
     return x <= y;
@@ -697,7 +971,7 @@ struct scalar_fuzzy_default_impl<Scalar, true, false>
   }
   static inline bool isApprox(const Scalar& x, const Scalar& y, const RealScalar& prec)
   {
-    using std::min;
+    EIGEN_USING_STD_MATH(min);
     return numext::abs2(x - y) <= (min)(numext::abs2(x), numext::abs2(y)) * prec * prec;
   }
 };
@@ -705,21 +979,21 @@ struct scalar_fuzzy_default_impl<Scalar, true, false>
 template<typename Scalar>
 struct scalar_fuzzy_impl : scalar_fuzzy_default_impl<Scalar, NumTraits<Scalar>::IsComplex, NumTraits<Scalar>::IsInteger> {};
 
-template<typename Scalar, typename OtherScalar>
+template<typename Scalar, typename OtherScalar> EIGEN_DEVICE_FUNC
 inline bool isMuchSmallerThan(const Scalar& x, const OtherScalar& y,
                                    typename NumTraits<Scalar>::Real precision = NumTraits<Scalar>::dummy_precision())
 {
   return scalar_fuzzy_impl<Scalar>::template isMuchSmallerThan<OtherScalar>(x, y, precision);
 }
 
-template<typename Scalar>
+template<typename Scalar> EIGEN_DEVICE_FUNC
 inline bool isApprox(const Scalar& x, const Scalar& y,
                           typename NumTraits<Scalar>::Real precision = NumTraits<Scalar>::dummy_precision())
 {
   return scalar_fuzzy_impl<Scalar>::isApprox(x, y, precision);
 }
 
-template<typename Scalar>
+template<typename Scalar> EIGEN_DEVICE_FUNC
 inline bool isApproxOrLessThan(const Scalar& x, const Scalar& y,
                                     typename NumTraits<Scalar>::Real precision = NumTraits<Scalar>::dummy_precision())
 {
@@ -742,17 +1016,19 @@ template<> struct scalar_fuzzy_impl<bool>
 {
   typedef bool RealScalar;
   
-  template<typename OtherScalar>
+  template<typename OtherScalar> EIGEN_DEVICE_FUNC
   static inline bool isMuchSmallerThan(const bool& x, const bool&, const bool&)
   {
     return !x;
   }
   
+  EIGEN_DEVICE_FUNC
   static inline bool isApprox(bool x, bool y, bool)
   {
     return x == y;
   }
 
+  EIGEN_DEVICE_FUNC
   static inline bool isApproxOrLessThan(const bool& x, const bool& y, const bool&)
   {
     return (!x) || y;

Eigen/src/Core/Matrix.h

@@ -24,13 +24,13 @@ namespace Eigen {
   * The %Matrix class encompasses \em both fixed-size and dynamic-size objects (\ref fixedsize "note").
   *
   * The first three template parameters are required:
-  * \tparam _Scalar \anchor matrix_tparam_scalar Numeric type, e.g. float, double, int or std::complex<float>.
-  *                 User defined sclar types are supported as well (see \ref user_defined_scalars "here").
+  * \tparam _Scalar Numeric type, e.g. float, double, int or std::complex<float>.
+  *                 User defined scalar types are supported as well (see \ref user_defined_scalars "here").
   * \tparam _Rows Number of rows, or \b Dynamic
   * \tparam _Cols Number of columns, or \b Dynamic
   *
   * The remaining template parameters are optional -- in most cases you don't have to worry about them.
-  * \tparam _Options \anchor matrix_tparam_options A combination of either \b #RowMajor or \b #ColMajor, and of either
+  * \tparam _Options A combination of either \b #RowMajor or \b #ColMajor, and of either
   *                 \b #AutoAlign or \b #DontAlign.
   *                 The former controls \ref TopicStorageOrders "storage order", and defaults to column-major. The latter controls alignment, which is required
   *                 for vectorization. It defaults to aligning matrices except for fixed sizes that aren't a multiple of the packet size.
@@ -97,6 +97,40 @@ namespace Eigen {
   * are the dimensions of the original matrix, while _Rows and _Cols are Dynamic.</dd>
   * </dl>
   *
+  * <i><b>ABI and storage layout</b></i>
+  * 
+  * The table below summarizes the ABI of some possible Matrix instances which is fixed thorough the lifetime of Eigen 3.
+  * <table  class="manual">
+  * <tr><th>Matrix type</th><th>Equivalent C structure</th></tr>
+  * <tr><td>\code Matrix<T,Dynamic,Dynamic> \endcode</td><td>\code
+  * struct {
+  *   T *data;                  // with (size_t(data)%EIGEN_MAX_ALIGN_BYTES)==0
+  *   Eigen::Index rows, cols;
+  *  };
+  * \endcode</td></tr>
+  * <tr class="alt"><td>\code
+  * Matrix<T,Dynamic,1>
+  * Matrix<T,1,Dynamic> \endcode</td><td>\code
+  * struct {
+  *   T *data;                  // with (size_t(data)%EIGEN_MAX_ALIGN_BYTES)==0
+  *   Eigen::Index size;
+  *  };
+  * \endcode</td></tr>
+  * <tr><td>\code Matrix<T,Rows,Cols> \endcode</td><td>\code
+  * struct {
+  *   T data[Rows*Cols];        // with (size_t(data)%A(Rows*Cols*sizeof(T)))==0
+  *  };
+  * \endcode</td></tr>
+  * <tr class="alt"><td>\code Matrix<T,Dynamic,Dynamic,0,MaxRows,MaxCols> \endcode</td><td>\code
+  * struct {
+  *   T data[MaxRows*MaxCols];  // with (size_t(data)%A(MaxRows*MaxCols*sizeof(T)))==0
+  *   Eigen::Index rows, cols;
+  *  };
+  * \endcode</td></tr>
+  * </table>
+  * Note that in this table Rows, Cols, MaxRows and MaxCols are all positive integers. A(S) is defined to the largest possible power-of-two
+  * smaller to EIGEN_MAX_STATIC_ALIGN_BYTES.
+  * 
   * \see MatrixBase for the majority of the API methods for matrices, \ref TopicClassHierarchy, 
   * \ref TopicStorageOrders 
   */
@@ -105,9 +139,23 @@ namespace internal {
 template<typename _Scalar, int _Rows, int _Cols, int _Options, int _MaxRows, int _MaxCols>
 struct traits<Matrix<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> >
 {
+private:
+  enum { size = internal::size_at_compile_time<_Rows,_Cols>::ret };
+  typedef typename find_best_packet<_Scalar,size>::type PacketScalar;
+  enum {
+      row_major_bit = _Options&RowMajor ? RowMajorBit : 0,
+      is_dynamic_size_storage = _MaxRows==Dynamic || _MaxCols==Dynamic,
+      max_size = is_dynamic_size_storage ? Dynamic : _MaxRows*_MaxCols,
+      default_alignment = compute_default_alignment<_Scalar,max_size>::value,
+      actual_alignment = ((_Options&DontAlign)==0) ? default_alignment : 0,
+      required_alignment = unpacket_traits<PacketScalar>::alignment,
+      packet_access_bit = packet_traits<_Scalar>::Vectorizable && (actual_alignment>=required_alignment) ? PacketAccessBit : 0
+    };
+    
+public:
   typedef _Scalar Scalar;
   typedef Dense StorageKind;
-  typedef DenseIndex Index;
+  typedef Eigen::Index StorageIndex;
   typedef MatrixXpr XprKind;
   enum {
     RowsAtCompileTime = _Rows,
@@ -115,10 +163,13 @@ struct traits<Matrix<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> >
     MaxRowsAtCompileTime = _MaxRows,
     MaxColsAtCompileTime = _MaxCols,
     Flags = compute_matrix_flags<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols>::ret,
-    CoeffReadCost = NumTraits<Scalar>::ReadCost,
     Options = _Options,
     InnerStrideAtCompileTime = 1,
-    OuterStrideAtCompileTime = (Options&RowMajor) ? ColsAtCompileTime : RowsAtCompileTime
+    OuterStrideAtCompileTime = (Options&RowMajor) ? ColsAtCompileTime : RowsAtCompileTime,
+    
+    // FIXME, the following flag in only used to define NeedsToAlign in PlainObjectBase
+    EvaluatorFlags = LinearAccessBit | DirectAccessBit | packet_access_bit | row_major_bit,
+    Alignment = actual_alignment
   };
 };
 }
@@ -151,6 +202,7 @@ class Matrix
       *
       * \callgraph
       */
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Matrix& operator=(const Matrix& other)
     {
       return Base::_set(other);
@@ -167,7 +219,8 @@ class Matrix
       * remain row-vectors and vectors remain vectors.
       */
     template<typename OtherDerived>
-    EIGEN_STRONG_INLINE Matrix& operator=(const MatrixBase<OtherDerived>& other)
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Matrix& operator=(const DenseBase<OtherDerived>& other)
     {
       return Base::_set(other);
     }
@@ -179,12 +232,14 @@ class Matrix
       * \copydetails DenseBase::operator=(const EigenBase<OtherDerived> &other)
       */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Matrix& operator=(const EigenBase<OtherDerived> &other)
     {
       return Base::operator=(other);
     }
 
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Matrix& operator=(const ReturnByValue<OtherDerived>& func)
     {
       return Base::operator=(func);
@@ -200,6 +255,7 @@ class Matrix
       *
       * \sa resize(Index,Index)
       */
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Matrix() : Base()
     {
       Base::_check_template_params();
@@ -207,45 +263,87 @@ class Matrix
     }
 
     // FIXME is it still needed
-    Matrix(internal::constructor_without_unaligned_array_assert)
+    EIGEN_DEVICE_FUNC
+    explicit Matrix(internal::constructor_without_unaligned_array_assert)
       : Base(internal::constructor_without_unaligned_array_assert())
     { Base::_check_template_params(); EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED }
 
-    /** \brief Constructs a vector or row-vector with given dimension. \only_for_vectors
-      *
-      * Note that this is only useful for dynamic-size vectors. For fixed-size vectors,
-      * it is redundant to pass the dimension here, so it makes more sense to use the default
-      * constructor Matrix() instead.
-      */
-    EIGEN_STRONG_INLINE explicit Matrix(Index dim)
-      : Base(dim, RowsAtCompileTime == 1 ? 1 : dim, ColsAtCompileTime == 1 ? 1 : dim)
+#ifdef EIGEN_HAVE_RVALUE_REFERENCES
+    EIGEN_DEVICE_FUNC
+    Matrix(Matrix&& other)
+      : Base(std::move(other))
     {
       Base::_check_template_params();
-      EIGEN_STATIC_ASSERT_VECTOR_ONLY(Matrix)
-      eigen_assert(dim >= 0);
-      eigen_assert(SizeAtCompileTime == Dynamic || SizeAtCompileTime == dim);
-      EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED
+      if (RowsAtCompileTime!=Dynamic && ColsAtCompileTime!=Dynamic)
+        Base::_set_noalias(other);
     }
+    EIGEN_DEVICE_FUNC
+    Matrix& operator=(Matrix&& other)
+    {
+      other.swap(*this);
+      return *this;
+    }
+#endif
 
     #ifndef EIGEN_PARSED_BY_DOXYGEN
+
+    // This constructor is for both 1x1 matrices and dynamic vectors
+    template<typename T>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE explicit Matrix(const T& x)
+    {
+      Base::_check_template_params();
+      Base::template _init1<T>(x);
+    }
+
     template<typename T0, typename T1>
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Matrix(const T0& x, const T1& y)
     {
       Base::_check_template_params();
       Base::template _init2<T0,T1>(x, y);
     }
     #else
+    /** \brief Constructs a fixed-sized matrix initialized with coefficients starting at \a data */
+    EIGEN_DEVICE_FUNC
+    explicit Matrix(const Scalar *data);
+
+    /** \brief Constructs a vector or row-vector with given dimension. \only_for_vectors
+      *
+      * This is useful for dynamic-size vectors. For fixed-size vectors,
+      * it is redundant to pass these parameters, so one should use the default constructor
+      * Matrix() instead.
+      * 
+      * \warning This constructor is disabled for fixed-size \c 1x1 matrices. For instance,
+      * calling Matrix<double,1,1>(1) will call the initialization constructor: Matrix(const Scalar&).
+      * For fixed-size \c 1x1 matrices it is therefore recommended to use the default
+      * constructor Matrix() instead, especially when using one of the non standard
+      * \c EIGEN_INITIALIZE_MATRICES_BY_{ZERO,\c NAN} macros (see \ref TopicPreprocessorDirectives).
+      */
+    EIGEN_STRONG_INLINE explicit Matrix(Index dim);
+    /** \brief Constructs an initialized 1x1 matrix with the given coefficient */
+    Matrix(const Scalar& x);
     /** \brief Constructs an uninitialized matrix with \a rows rows and \a cols columns.
       *
       * This is useful for dynamic-size matrices. For fixed-size matrices,
       * it is redundant to pass these parameters, so one should use the default constructor
-      * Matrix() instead. */
+      * Matrix() instead.
+      * 
+      * \warning This constructor is disabled for fixed-size \c 1x2 and \c 2x1 vectors. For instance,
+      * calling Matrix2f(2,1) will call the initialization constructor: Matrix(const Scalar& x, const Scalar& y).
+      * For fixed-size \c 1x2 or \c 2x1 vectors it is therefore recommended to use the default
+      * constructor Matrix() instead, especially when using one of the non standard
+      * \c EIGEN_INITIALIZE_MATRICES_BY_{ZERO,\c NAN} macros (see \ref TopicPreprocessorDirectives).
+      */
+    EIGEN_DEVICE_FUNC
     Matrix(Index rows, Index cols);
+    
     /** \brief Constructs an initialized 2D vector with given coefficients */
     Matrix(const Scalar& x, const Scalar& y);
     #endif
 
     /** \brief Constructs an initialized 3D vector with given coefficients */
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Matrix(const Scalar& x, const Scalar& y, const Scalar& z)
     {
       Base::_check_template_params();
@@ -255,6 +353,7 @@ class Matrix
       m_storage.data()[2] = z;
     }
     /** \brief Constructs an initialized 4D vector with given coefficients */
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Matrix(const Scalar& x, const Scalar& y, const Scalar& z, const Scalar& w)
     {
       Base::_check_template_params();
@@ -265,76 +364,33 @@ class Matrix
       m_storage.data()[3] = w;
     }
 
-    explicit Matrix(const Scalar *data);
 
-    /** \brief Constructor copying the value of the expression \a other */
-    template<typename OtherDerived>
-    EIGEN_STRONG_INLINE Matrix(const MatrixBase<OtherDerived>& other)
-             : Base(other.rows() * other.cols(), other.rows(), other.cols())
-    {
-      // This test resides here, to bring the error messages closer to the user. Normally, these checks
-      // are performed deeply within the library, thus causing long and scary error traces.
-      EIGEN_STATIC_ASSERT((internal::is_same<Scalar, typename OtherDerived::Scalar>::value),
-        YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY)
-
-      Base::_check_template_params();
-      Base::_set_noalias(other);
-    }
     /** \brief Copy constructor */
-    EIGEN_STRONG_INLINE Matrix(const Matrix& other)
-            : Base(other.rows() * other.cols(), other.rows(), other.cols())
-    {
-      Base::_check_template_params();
-      Base::_set_noalias(other);
-    }
-    /** \brief Copy constructor with in-place evaluation */
-    template<typename OtherDerived>
-    EIGEN_STRONG_INLINE Matrix(const ReturnByValue<OtherDerived>& other)
-    {
-      Base::_check_template_params();
-      Base::resize(other.rows(), other.cols());
-      other.evalTo(*this);
-    }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Matrix(const Matrix& other) : Base(other)
+    { }
 
     /** \brief Copy constructor for generic expressions.
       * \sa MatrixBase::operator=(const EigenBase<OtherDerived>&)
       */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Matrix(const EigenBase<OtherDerived> &other)
-      : Base(other.derived().rows() * other.derived().cols(), other.derived().rows(), other.derived().cols())
-    {
-      Base::_check_template_params();
-      Base::resize(other.rows(), other.cols());
-      // FIXME/CHECK: isn't *this = other.derived() more efficient. it allows to
-      //              go for pure _set() implementations, right?
-      *this = other;
-    }
+      : Base(other.derived())
+    { }
 
-    /** \internal
-      * \brief Override MatrixBase::swap() since for dynamic-sized matrices
-      * of same type it is enough to swap the data pointers.
-      */
-    template<typename OtherDerived>
-    void swap(MatrixBase<OtherDerived> const & other)
-    { this->_swap(other.derived()); }
-
-    inline Index innerStride() const { return 1; }
-    inline Index outerStride() const { return this->innerSize(); }
+    EIGEN_DEVICE_FUNC inline Index innerStride() const { return 1; }
+    EIGEN_DEVICE_FUNC inline Index outerStride() const { return this->innerSize(); }
 
     /////////// Geometry module ///////////
 
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     explicit Matrix(const RotationBase<OtherDerived,ColsAtCompileTime>& r);
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     Matrix& operator=(const RotationBase<OtherDerived,ColsAtCompileTime>& r);
 
-    #ifdef EIGEN2_SUPPORT
-    template<typename OtherDerived>
-    explicit Matrix(const eigen2_RotationBase<OtherDerived,ColsAtCompileTime>& r);
-    template<typename OtherDerived>
-    Matrix& operator=(const eigen2_RotationBase<OtherDerived,ColsAtCompileTime>& r);
-    #endif
-
     // allow to extend Matrix outside Eigen
     #ifdef EIGEN_MATRIX_PLUGIN
     #include EIGEN_MATRIX_PLUGIN

Eigen/src/Core/MatrixBase.h

@@ -52,7 +52,7 @@ template<typename Derived> class MatrixBase
 #ifndef EIGEN_PARSED_BY_DOXYGEN
     typedef MatrixBase StorageBaseType;
     typedef typename internal::traits<Derived>::StorageKind StorageKind;
-    typedef typename internal::traits<Derived>::Index Index;
+    typedef typename internal::traits<Derived>::StorageIndex StorageIndex;
     typedef typename internal::traits<Derived>::Scalar Scalar;
     typedef typename internal::packet_traits<Scalar>::type PacketScalar;
     typedef typename NumTraits<Scalar>::Real RealScalar;
@@ -66,8 +66,7 @@ template<typename Derived> class MatrixBase
     using Base::MaxSizeAtCompileTime;
     using Base::IsVectorAtCompileTime;
     using Base::Flags;
-    using Base::CoeffReadCost;
-
+    
     using Base::derived;
     using Base::const_cast_derived;
     using Base::rows;
@@ -81,6 +80,8 @@ template<typename Derived> class MatrixBase
     using Base::operator-=;
     using Base::operator*=;
     using Base::operator/=;
+    using Base::operator*;
+    using Base::operator/;
 
     typedef typename Base::CoeffReturnType CoeffReturnType;
     typedef typename Base::ConstTransposeReturnType ConstTransposeReturnType;
@@ -98,25 +99,14 @@ template<typename Derived> class MatrixBase
 
     /** \returns the size of the main diagonal, which is min(rows(),cols()).
       * \sa rows(), cols(), SizeAtCompileTime. */
+    EIGEN_DEVICE_FUNC
     inline Index diagonalSize() const { return (std::min)(rows(),cols()); }
 
-    /** \brief The plain matrix type corresponding to this expression.
-      *
-      * This is not necessarily exactly the return type of eval(). In the case of plain matrices,
-      * the return type of eval() is a const reference to a matrix, not a matrix! It is however guaranteed
-      * that the return type of eval() is either PlainObject or const PlainObject&.
-      */
-    typedef Matrix<typename internal::traits<Derived>::Scalar,
-                internal::traits<Derived>::RowsAtCompileTime,
-                internal::traits<Derived>::ColsAtCompileTime,
-                AutoAlign | (internal::traits<Derived>::Flags&RowMajorBit ? RowMajor : ColMajor),
-                internal::traits<Derived>::MaxRowsAtCompileTime,
-                internal::traits<Derived>::MaxColsAtCompileTime
-          > PlainObject;
+    typedef typename Base::PlainObject PlainObject;
 
 #ifndef EIGEN_PARSED_BY_DOXYGEN
     /** \internal Represents a matrix with all coefficients equal to one another*/
-    typedef CwiseNullaryOp<internal::scalar_constant_op<Scalar>,Derived> ConstantReturnType;
+    typedef CwiseNullaryOp<internal::scalar_constant_op<Scalar>,PlainObject> ConstantReturnType;
     /** \internal the return type of MatrixBase::adjoint() */
     typedef typename internal::conditional<NumTraits<Scalar>::IsComplex,
                         CwiseUnaryOp<internal::scalar_conjugate_op<Scalar>, ConstTransposeReturnType>,
@@ -125,7 +115,7 @@ template<typename Derived> class MatrixBase
     /** \internal Return type of eigenvalues() */
     typedef Matrix<std::complex<RealScalar>, internal::traits<Derived>::ColsAtCompileTime, 1, ColMajor> EigenvaluesReturnType;
     /** \internal the return type of identity */
-    typedef CwiseNullaryOp<internal::scalar_identity_op<Scalar>,Derived> IdentityReturnType;
+    typedef CwiseNullaryOp<internal::scalar_identity_op<Scalar>,PlainObject> IdentityReturnType;
     /** \internal the return type of unit vectors */
     typedef Block<const CwiseNullaryOp<internal::scalar_identity_op<Scalar>, SquareMatrixType>,
                   internal::traits<Derived>::RowsAtCompileTime,
@@ -145,39 +135,48 @@ template<typename Derived> class MatrixBase
     /** Special case of the template operator=, in order to prevent the compiler
       * from generating a default operator= (issue hit with g++ 4.1)
       */
+    EIGEN_DEVICE_FUNC
     Derived& operator=(const MatrixBase& other);
 
     // We cannot inherit here via Base::operator= since it is causing
     // trouble with MSVC.
 
     template <typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     Derived& operator=(const DenseBase<OtherDerived>& other);
 
     template <typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     Derived& operator=(const EigenBase<OtherDerived>& other);
 
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     Derived& operator=(const ReturnByValue<OtherDerived>& other);
 
-#ifndef EIGEN_PARSED_BY_DOXYGEN
-    template<typename ProductDerived, typename Lhs, typename Rhs>
-    Derived& lazyAssign(const ProductBase<ProductDerived, Lhs,Rhs>& other);
-
-    template<typename MatrixPower, typename Lhs, typename Rhs>
-    Derived& lazyAssign(const MatrixPowerProduct<MatrixPower, Lhs,Rhs>& other);
-#endif // not EIGEN_PARSED_BY_DOXYGEN
-
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     Derived& operator+=(const MatrixBase<OtherDerived>& other);
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     Derived& operator-=(const MatrixBase<OtherDerived>& other);
 
+#ifdef __CUDACC__
     template<typename OtherDerived>
-    const typename ProductReturnType<Derived,OtherDerived>::Type
-    operator*(const MatrixBase<OtherDerived> &other) const;
+    EIGEN_DEVICE_FUNC
+    const Product<Derived,OtherDerived,LazyProduct>
+    operator*(const MatrixBase<OtherDerived> &other) const
+    { return this->lazyProduct(other); }
+#else
 
     template<typename OtherDerived>
-    const typename LazyProductReturnType<Derived,OtherDerived>::Type
+    const Product<Derived,OtherDerived>
+    operator*(const MatrixBase<OtherDerived> &other) const;
+
+#endif
+
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC 
+    const Product<Derived,OtherDerived,LazyProduct>
     lazyProduct(const MatrixBase<OtherDerived> &other) const;
 
     template<typename OtherDerived>
@@ -190,88 +189,91 @@ template<typename Derived> class MatrixBase
     void applyOnTheRight(const EigenBase<OtherDerived>& other);
 
     template<typename DiagonalDerived>
-    const DiagonalProduct<Derived, DiagonalDerived, OnTheRight>
+    EIGEN_DEVICE_FUNC
+    const Product<Derived, DiagonalDerived, LazyProduct>
     operator*(const DiagonalBase<DiagonalDerived> &diagonal) const;
 
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     typename internal::scalar_product_traits<typename internal::traits<Derived>::Scalar,typename internal::traits<OtherDerived>::Scalar>::ReturnType
     dot(const MatrixBase<OtherDerived>& other) const;
 
-    #ifdef EIGEN2_SUPPORT
-      template<typename OtherDerived>
-      Scalar eigen2_dot(const MatrixBase<OtherDerived>& other) const;
-    #endif
-
-    RealScalar squaredNorm() const;
-    RealScalar norm() const;
+    EIGEN_DEVICE_FUNC RealScalar squaredNorm() const;
+    EIGEN_DEVICE_FUNC RealScalar norm() const;
     RealScalar stableNorm() const;
     RealScalar blueNorm() const;
     RealScalar hypotNorm() const;
-    const PlainObject normalized() const;
-    void normalize();
+    EIGEN_DEVICE_FUNC const PlainObject normalized() const;
+    EIGEN_DEVICE_FUNC void normalize();
 
-    const AdjointReturnType adjoint() const;
-    void adjointInPlace();
+    EIGEN_DEVICE_FUNC const AdjointReturnType adjoint() const;
+    EIGEN_DEVICE_FUNC void adjointInPlace();
 
     typedef Diagonal<Derived> DiagonalReturnType;
+    EIGEN_DEVICE_FUNC
     DiagonalReturnType diagonal();
-	typedef typename internal::add_const<Diagonal<const Derived> >::type ConstDiagonalReturnType;
+    
+    typedef typename internal::add_const<Diagonal<const Derived> >::type ConstDiagonalReturnType;
+    EIGEN_DEVICE_FUNC
     ConstDiagonalReturnType diagonal() const;
 
     template<int Index> struct DiagonalIndexReturnType { typedef Diagonal<Derived,Index> Type; };
     template<int Index> struct ConstDiagonalIndexReturnType { typedef const Diagonal<const Derived,Index> Type; };
 
-    template<int Index> typename DiagonalIndexReturnType<Index>::Type diagonal();
-    template<int Index> typename ConstDiagonalIndexReturnType<Index>::Type diagonal() const;
+    template<int Index> 
+    EIGEN_DEVICE_FUNC
+    typename DiagonalIndexReturnType<Index>::Type diagonal();
 
-    // Note: The "MatrixBase::" prefixes are added to help MSVC9 to match these declarations with the later implementations.
-    // On the other hand they confuse MSVC8...
-    #if (defined _MSC_VER) && (_MSC_VER >= 1500) // 2008 or later
-    typename MatrixBase::template DiagonalIndexReturnType<DynamicIndex>::Type diagonal(Index index);
-    typename MatrixBase::template ConstDiagonalIndexReturnType<DynamicIndex>::Type diagonal(Index index) const;
-    #else
-    typename DiagonalIndexReturnType<DynamicIndex>::Type diagonal(Index index);
-    typename ConstDiagonalIndexReturnType<DynamicIndex>::Type diagonal(Index index) const;
-    #endif
-
-    #ifdef EIGEN2_SUPPORT
-    template<unsigned int Mode> typename internal::eigen2_part_return_type<Derived, Mode>::type part();
-    template<unsigned int Mode> const typename internal::eigen2_part_return_type<Derived, Mode>::type part() const;
+    template<int Index>
+    EIGEN_DEVICE_FUNC
+    typename ConstDiagonalIndexReturnType<Index>::Type diagonal() const;
     
-    // huuuge hack. make Eigen2's matrix.part<Diagonal>() work in eigen3. Problem: Diagonal is now a class template instead
-    // of an integer constant. Solution: overload the part() method template wrt template parameters list.
-    template<template<typename T, int N> class U>
-    const DiagonalWrapper<ConstDiagonalReturnType> part() const
-    { return diagonal().asDiagonal(); }
-    #endif // EIGEN2_SUPPORT
+    typedef Diagonal<Derived,DynamicIndex> DiagonalDynamicIndexReturnType;
+    typedef typename internal::add_const<Diagonal<const Derived,DynamicIndex> >::type ConstDiagonalDynamicIndexReturnType;
+
+    EIGEN_DEVICE_FUNC
+    DiagonalDynamicIndexReturnType diagonal(Index index);
+    EIGEN_DEVICE_FUNC
+    ConstDiagonalDynamicIndexReturnType diagonal(Index index) const;
 
     template<unsigned int Mode> struct TriangularViewReturnType { typedef TriangularView<Derived, Mode> Type; };
     template<unsigned int Mode> struct ConstTriangularViewReturnType { typedef const TriangularView<const Derived, Mode> Type; };
 
-    template<unsigned int Mode> typename TriangularViewReturnType<Mode>::Type triangularView();
-    template<unsigned int Mode> typename ConstTriangularViewReturnType<Mode>::Type triangularView() const;
+    template<unsigned int Mode>
+    EIGEN_DEVICE_FUNC
+    typename TriangularViewReturnType<Mode>::Type triangularView();
+    template<unsigned int Mode>
+    EIGEN_DEVICE_FUNC
+    typename ConstTriangularViewReturnType<Mode>::Type triangularView() const;
 
     template<unsigned int UpLo> struct SelfAdjointViewReturnType { typedef SelfAdjointView<Derived, UpLo> Type; };
     template<unsigned int UpLo> struct ConstSelfAdjointViewReturnType { typedef const SelfAdjointView<const Derived, UpLo> Type; };
 
-    template<unsigned int UpLo> typename SelfAdjointViewReturnType<UpLo>::Type selfadjointView();
-    template<unsigned int UpLo> typename ConstSelfAdjointViewReturnType<UpLo>::Type selfadjointView() const;
+    template<unsigned int UpLo> 
+    EIGEN_DEVICE_FUNC
+    typename SelfAdjointViewReturnType<UpLo>::Type selfadjointView();
+    template<unsigned int UpLo>
+    EIGEN_DEVICE_FUNC
+    typename ConstSelfAdjointViewReturnType<UpLo>::Type selfadjointView() const;
 
     const SparseView<Derived> sparseView(const Scalar& m_reference = Scalar(0),
                                          const typename NumTraits<Scalar>::Real& m_epsilon = NumTraits<Scalar>::dummy_precision()) const;
-    static const IdentityReturnType Identity();
-    static const IdentityReturnType Identity(Index rows, Index cols);
-    static const BasisReturnType Unit(Index size, Index i);
-    static const BasisReturnType Unit(Index i);
-    static const BasisReturnType UnitX();
-    static const BasisReturnType UnitY();
-    static const BasisReturnType UnitZ();
-    static const BasisReturnType UnitW();
+    EIGEN_DEVICE_FUNC static const IdentityReturnType Identity();
+    EIGEN_DEVICE_FUNC static const IdentityReturnType Identity(Index rows, Index cols);
+    EIGEN_DEVICE_FUNC static const BasisReturnType Unit(Index size, Index i);
+    EIGEN_DEVICE_FUNC static const BasisReturnType Unit(Index i);
+    EIGEN_DEVICE_FUNC static const BasisReturnType UnitX();
+    EIGEN_DEVICE_FUNC static const BasisReturnType UnitY();
+    EIGEN_DEVICE_FUNC static const BasisReturnType UnitZ();
+    EIGEN_DEVICE_FUNC static const BasisReturnType UnitW();
 
+    EIGEN_DEVICE_FUNC
     const DiagonalWrapper<const Derived> asDiagonal() const;
     const PermutationWrapper<const Derived> asPermutation() const;
 
+    EIGEN_DEVICE_FUNC
     Derived& setIdentity();
+    EIGEN_DEVICE_FUNC
     Derived& setIdentity(Index rows, Index cols);
 
     bool isIdentity(const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const;
@@ -303,50 +305,37 @@ template<typename Derived> class MatrixBase
 
     NoAlias<Derived,Eigen::MatrixBase > noalias();
 
-    inline const ForceAlignedAccess<Derived> forceAlignedAccess() const;
-    inline ForceAlignedAccess<Derived> forceAlignedAccess();
-    template<bool Enable> inline typename internal::add_const_on_value_type<typename internal::conditional<Enable,ForceAlignedAccess<Derived>,Derived&>::type>::type forceAlignedAccessIf() const;
-    template<bool Enable> inline typename internal::conditional<Enable,ForceAlignedAccess<Derived>,Derived&>::type forceAlignedAccessIf();
+    // TODO forceAlignedAccess is temporarily disabled
+    // Need to find a nicer workaround.
+    inline const Derived& forceAlignedAccess() const { return derived(); }
+    inline Derived& forceAlignedAccess() { return derived(); }
+    template<bool Enable> inline const Derived& forceAlignedAccessIf() const { return derived(); }
+    template<bool Enable> inline Derived& forceAlignedAccessIf() { return derived(); }
 
-    Scalar trace() const;
+    EIGEN_DEVICE_FUNC Scalar trace() const;
 
-/////////// Array module ///////////
+    template<int p> EIGEN_DEVICE_FUNC RealScalar lpNorm() const;
 
-    template<int p> RealScalar lpNorm() const;
-
-    MatrixBase<Derived>& matrix() { return *this; }
-    const MatrixBase<Derived>& matrix() const { return *this; }
+    EIGEN_DEVICE_FUNC MatrixBase<Derived>& matrix() { return *this; }
+    EIGEN_DEVICE_FUNC const MatrixBase<Derived>& matrix() const { return *this; }
 
     /** \returns an \link Eigen::ArrayBase Array \endlink expression of this matrix
       * \sa ArrayBase::matrix() */
-    ArrayWrapper<Derived> array() { return derived(); }
-    const ArrayWrapper<const Derived> array() const { return derived(); }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ArrayWrapper<Derived> array() { return ArrayWrapper<Derived>(derived()); }
+    /** \returns a const \link Eigen::ArrayBase Array \endlink expression of this matrix
+      * \sa ArrayBase::matrix() */
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const ArrayWrapper<const Derived> array() const { return ArrayWrapper<const Derived>(derived()); }
 
 /////////// LU module ///////////
 
-    const FullPivLU<PlainObject> fullPivLu() const;
-    const PartialPivLU<PlainObject> partialPivLu() const;
+    EIGEN_DEVICE_FUNC const FullPivLU<PlainObject> fullPivLu() const;
+    EIGEN_DEVICE_FUNC const PartialPivLU<PlainObject> partialPivLu() const;
 
-    #if EIGEN2_SUPPORT_STAGE < STAGE20_RESOLVE_API_CONFLICTS
-    const LU<PlainObject> lu() const;
-    #endif
-
-    #ifdef EIGEN2_SUPPORT
-    const LU<PlainObject> eigen2_lu() const;
-    #endif
-
-    #if EIGEN2_SUPPORT_STAGE > STAGE20_RESOLVE_API_CONFLICTS
     const PartialPivLU<PlainObject> lu() const;
-    #endif
-    
-    #ifdef EIGEN2_SUPPORT
-    template<typename ResultType>
-    void computeInverse(MatrixBase<ResultType> *result) const {
-      *result = this->inverse();
-    }
-    #endif
 
-    const internal::inverse_impl<Derived> inverse() const;
+    EIGEN_DEVICE_FUNC
+    const Inverse<Derived> inverse() const;
+    
     template<typename ResultType>
     void computeInverseAndDetWithCheck(
       ResultType& inverse,
@@ -372,10 +361,6 @@ template<typename Derived> class MatrixBase
     const HouseholderQR<PlainObject> householderQr() const;
     const ColPivHouseholderQR<PlainObject> colPivHouseholderQr() const;
     const FullPivHouseholderQR<PlainObject> fullPivHouseholderQr() const;
-    
-    #ifdef EIGEN2_SUPPORT
-    const QR<PlainObject> qr() const;
-    #endif
 
     EigenvaluesReturnType eigenvalues() const;
     RealScalar operatorNorm() const;
@@ -383,10 +368,7 @@ template<typename Derived> class MatrixBase
 /////////// SVD module ///////////
 
     JacobiSVD<PlainObject> jacobiSvd(unsigned int computationOptions = 0) const;
-
-    #ifdef EIGEN2_SUPPORT
-    SVD<PlainObject> svd() const;
-    #endif
+    BDCSVD<PlainObject>    bdcSvd(unsigned int computationOptions = 0) const;
 
 /////////// Geometry module ///////////
 
@@ -398,20 +380,25 @@ template<typename Derived> class MatrixBase
     };
     #endif // EIGEN_PARSED_BY_DOXYGEN
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     typename cross_product_return_type<OtherDerived>::type
     cross(const MatrixBase<OtherDerived>& other) const;
+    
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     PlainObject cross3(const MatrixBase<OtherDerived>& other) const;
+    
+    EIGEN_DEVICE_FUNC
     PlainObject unitOrthogonal(void) const;
+    
     Matrix<Scalar,3,1> eulerAngles(Index a0, Index a1, Index a2) const;
     
-    #if EIGEN2_SUPPORT_STAGE > STAGE20_RESOLVE_API_CONFLICTS
     ScalarMultipleReturnType operator*(const UniformScaling<Scalar>& s) const;
     // put this as separate enum value to work around possible GCC 4.3 bug (?)
-    enum { HomogeneousReturnTypeDirection = ColsAtCompileTime==1?Vertical:Horizontal };
+    enum { HomogeneousReturnTypeDirection = ColsAtCompileTime==1&&RowsAtCompileTime==1 ? ((internal::traits<Derived>::Flags&RowMajorBit)==RowMajorBit ? Horizontal : Vertical)
+                                          : ColsAtCompileTime==1 ? Vertical : Horizontal };
     typedef Homogeneous<Derived, HomogeneousReturnTypeDirection> HomogeneousReturnType;
     HomogeneousReturnType homogeneous() const;
-    #endif
     
     enum {
       SizeMinusOne = SizeAtCompileTime==Dynamic ? Dynamic : SizeAtCompileTime-1
@@ -458,49 +445,15 @@ template<typename Derived> class MatrixBase
     const MatrixSquareRootReturnValue<Derived> sqrt() const;
     const MatrixLogarithmReturnValue<Derived> log() const;
     const MatrixPowerReturnValue<Derived> pow(const RealScalar& p) const;
-
-#ifdef EIGEN2_SUPPORT
-    template<typename ProductDerived, typename Lhs, typename Rhs>
-    Derived& operator+=(const Flagged<ProductBase<ProductDerived, Lhs,Rhs>, 0,
-                                      EvalBeforeAssigningBit>& other);
-
-    template<typename ProductDerived, typename Lhs, typename Rhs>
-    Derived& operator-=(const Flagged<ProductBase<ProductDerived, Lhs,Rhs>, 0,
-                                      EvalBeforeAssigningBit>& other);
-
-    /** \deprecated because .lazy() is deprecated
-      * Overloaded for cache friendly product evaluation */
-    template<typename OtherDerived>
-    Derived& lazyAssign(const Flagged<OtherDerived, 0, EvalBeforeAssigningBit>& other)
-    { return lazyAssign(other._expression()); }
-
-    template<unsigned int Added>
-    const Flagged<Derived, Added, 0> marked() const;
-    const Flagged<Derived, 0, EvalBeforeAssigningBit> lazy() const;
-
-    inline const Cwise<Derived> cwise() const;
-    inline Cwise<Derived> cwise();
-
-    VectorBlock<Derived> start(Index size);
-    const VectorBlock<const Derived> start(Index size) const;
-    VectorBlock<Derived> end(Index size);
-    const VectorBlock<const Derived> end(Index size) const;
-    template<int Size> VectorBlock<Derived,Size> start();
-    template<int Size> const VectorBlock<const Derived,Size> start() const;
-    template<int Size> VectorBlock<Derived,Size> end();
-    template<int Size> const VectorBlock<const Derived,Size> end() const;
-
-    Minor<Derived> minor(Index row, Index col);
-    const Minor<Derived> minor(Index row, Index col) const;
-#endif
+    const MatrixComplexPowerReturnValue<Derived> pow(const std::complex<RealScalar>& p) const;
 
   protected:
-    MatrixBase() : Base() {}
+    EIGEN_DEVICE_FUNC MatrixBase() : Base() {}
 
   private:
-    explicit MatrixBase(int);
-    MatrixBase(int,int);
-    template<typename OtherDerived> explicit MatrixBase(const MatrixBase<OtherDerived>&);
+    EIGEN_DEVICE_FUNC explicit MatrixBase(int);
+    EIGEN_DEVICE_FUNC MatrixBase(int,int);
+    template<typename OtherDerived> EIGEN_DEVICE_FUNC explicit MatrixBase(const MatrixBase<OtherDerived>&);
   protected:
     // mixing arrays and matrices is not legal
     template<typename OtherDerived> Derived& operator+=(const ArrayBase<OtherDerived>& )
@@ -510,6 +463,51 @@ template<typename Derived> class MatrixBase
     {EIGEN_STATIC_ASSERT(std::ptrdiff_t(sizeof(typename OtherDerived::Scalar))==-1,YOU_CANNOT_MIX_ARRAYS_AND_MATRICES); return *this;}
 };
 
+
+/***************************************************************************
+* Implementation of matrix base methods
+***************************************************************************/
+
+/** replaces \c *this by \c *this * \a other.
+  *
+  * \returns a reference to \c *this
+  *
+  * Example: \include MatrixBase_applyOnTheRight.cpp
+  * Output: \verbinclude MatrixBase_applyOnTheRight.out
+  */
+template<typename Derived>
+template<typename OtherDerived>
+inline Derived&
+MatrixBase<Derived>::operator*=(const EigenBase<OtherDerived> &other)
+{
+  other.derived().applyThisOnTheRight(derived());
+  return derived();
+}
+
+/** replaces \c *this by \c *this * \a other. It is equivalent to MatrixBase::operator*=().
+  *
+  * Example: \include MatrixBase_applyOnTheRight.cpp
+  * Output: \verbinclude MatrixBase_applyOnTheRight.out
+  */
+template<typename Derived>
+template<typename OtherDerived>
+inline void MatrixBase<Derived>::applyOnTheRight(const EigenBase<OtherDerived> &other)
+{
+  other.derived().applyThisOnTheRight(derived());
+}
+
+/** replaces \c *this by \a other * \c *this.
+  *
+  * Example: \include MatrixBase_applyOnTheLeft.cpp
+  * Output: \verbinclude MatrixBase_applyOnTheLeft.out
+  */
+template<typename Derived>
+template<typename OtherDerived>
+inline void MatrixBase<Derived>::applyOnTheLeft(const EigenBase<OtherDerived> &other)
+{
+  other.derived().applyThisOnTheLeft(derived());
+}
+
 } // end namespace Eigen
 
 #endif // EIGEN_MATRIXBASE_H

Eigen/src/Core/NestByValue.h

@@ -40,29 +40,29 @@ template<typename ExpressionType> class NestByValue
     typedef typename internal::dense_xpr_base<NestByValue>::type Base;
     EIGEN_DENSE_PUBLIC_INTERFACE(NestByValue)
 
-    inline NestByValue(const ExpressionType& matrix) : m_expression(matrix) {}
+    EIGEN_DEVICE_FUNC explicit inline NestByValue(const ExpressionType& matrix) : m_expression(matrix) {}
 
-    inline Index rows() const { return m_expression.rows(); }
-    inline Index cols() const { return m_expression.cols(); }
-    inline Index outerStride() const { return m_expression.outerStride(); }
-    inline Index innerStride() const { return m_expression.innerStride(); }
+    EIGEN_DEVICE_FUNC inline Index rows() const { return m_expression.rows(); }
+    EIGEN_DEVICE_FUNC inline Index cols() const { return m_expression.cols(); }
+    EIGEN_DEVICE_FUNC inline Index outerStride() const { return m_expression.outerStride(); }
+    EIGEN_DEVICE_FUNC inline Index innerStride() const { return m_expression.innerStride(); }
 
-    inline const CoeffReturnType coeff(Index row, Index col) const
+    EIGEN_DEVICE_FUNC inline const CoeffReturnType coeff(Index row, Index col) const
     {
       return m_expression.coeff(row, col);
     }
 
-    inline Scalar& coeffRef(Index row, Index col)
+    EIGEN_DEVICE_FUNC inline Scalar& coeffRef(Index row, Index col)
     {
       return m_expression.const_cast_derived().coeffRef(row, col);
     }
 
-    inline const CoeffReturnType coeff(Index index) const
+    EIGEN_DEVICE_FUNC inline const CoeffReturnType coeff(Index index) const
     {
       return m_expression.coeff(index);
     }
 
-    inline Scalar& coeffRef(Index index)
+    EIGEN_DEVICE_FUNC inline Scalar& coeffRef(Index index)
     {
       return m_expression.const_cast_derived().coeffRef(index);
     }
@@ -91,7 +91,7 @@ template<typename ExpressionType> class NestByValue
       m_expression.const_cast_derived().template writePacket<LoadMode>(index, x);
     }
 
-    operator const ExpressionType&() const { return m_expression; }
+    EIGEN_DEVICE_FUNC operator const ExpressionType&() const { return m_expression; }
 
   protected:
     const ExpressionType m_expression;

Eigen/src/Core/NoAlias.h

@@ -30,62 +30,36 @@ namespace Eigen {
 template<typename ExpressionType, template <typename> class StorageBase>
 class NoAlias
 {
-    typedef typename ExpressionType::Scalar Scalar;
   public:
-    NoAlias(ExpressionType& expression) : m_expression(expression) {}
-
-    /** Behaves like MatrixBase::lazyAssign(other)
-      * \sa MatrixBase::lazyAssign() */
-    template<typename OtherDerived>
-    EIGEN_STRONG_INLINE ExpressionType& operator=(const StorageBase<OtherDerived>& other)
-    { return internal::assign_selector<ExpressionType,OtherDerived,false>::run(m_expression,other.derived()); }
-
-    /** \sa MatrixBase::operator+= */
-    template<typename OtherDerived>
-    EIGEN_STRONG_INLINE ExpressionType& operator+=(const StorageBase<OtherDerived>& other)
-    {
-      typedef SelfCwiseBinaryOp<internal::scalar_sum_op<Scalar>, ExpressionType, OtherDerived> SelfAdder;
-      SelfAdder tmp(m_expression);
-      typedef typename internal::nested<OtherDerived>::type OtherDerivedNested;
-      typedef typename internal::remove_all<OtherDerivedNested>::type _OtherDerivedNested;
-      internal::assign_selector<SelfAdder,_OtherDerivedNested,false>::run(tmp,OtherDerivedNested(other.derived()));
-      return m_expression;
-    }
-
-    /** \sa MatrixBase::operator-= */
-    template<typename OtherDerived>
-    EIGEN_STRONG_INLINE ExpressionType& operator-=(const StorageBase<OtherDerived>& other)
-    {
-      typedef SelfCwiseBinaryOp<internal::scalar_difference_op<Scalar>, ExpressionType, OtherDerived> SelfAdder;
-      SelfAdder tmp(m_expression);
-      typedef typename internal::nested<OtherDerived>::type OtherDerivedNested;
-      typedef typename internal::remove_all<OtherDerivedNested>::type _OtherDerivedNested;
-      internal::assign_selector<SelfAdder,_OtherDerivedNested,false>::run(tmp,OtherDerivedNested(other.derived()));
-      return m_expression;
-    }
-
-#ifndef EIGEN_PARSED_BY_DOXYGEN
-    template<typename ProductDerived, typename Lhs, typename Rhs>
-    EIGEN_STRONG_INLINE ExpressionType& operator+=(const ProductBase<ProductDerived, Lhs,Rhs>& other)
-    { other.derived().addTo(m_expression); return m_expression; }
-
-    template<typename ProductDerived, typename Lhs, typename Rhs>
-    EIGEN_STRONG_INLINE ExpressionType& operator-=(const ProductBase<ProductDerived, Lhs,Rhs>& other)
-    { other.derived().subTo(m_expression); return m_expression; }
-
-    template<typename Lhs, typename Rhs, int NestingFlags>
-    EIGEN_STRONG_INLINE ExpressionType& operator+=(const CoeffBasedProduct<Lhs,Rhs,NestingFlags>& other)
-    { return m_expression.derived() += CoeffBasedProduct<Lhs,Rhs,NestByRefBit>(other.lhs(), other.rhs()); }
-
-    template<typename Lhs, typename Rhs, int NestingFlags>
-    EIGEN_STRONG_INLINE ExpressionType& operator-=(const CoeffBasedProduct<Lhs,Rhs,NestingFlags>& other)
-    { return m_expression.derived() -= CoeffBasedProduct<Lhs,Rhs,NestByRefBit>(other.lhs(), other.rhs()); }
+    typedef typename ExpressionType::Scalar Scalar;
+    
+    explicit NoAlias(ExpressionType& expression) : m_expression(expression) {}
     
     template<typename OtherDerived>
-    ExpressionType& operator=(const ReturnByValue<OtherDerived>& func)
-    { return m_expression = func; }
-#endif
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE ExpressionType& operator=(const StorageBase<OtherDerived>& other)
+    {
+      call_assignment_no_alias(m_expression, other.derived(), internal::assign_op<Scalar>());
+      return m_expression;
+    }
+    
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE ExpressionType& operator+=(const StorageBase<OtherDerived>& other)
+    {
+      call_assignment_no_alias(m_expression, other.derived(), internal::add_assign_op<Scalar>());
+      return m_expression;
+    }
+    
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE ExpressionType& operator-=(const StorageBase<OtherDerived>& other)
+    {
+      call_assignment_no_alias(m_expression, other.derived(), internal::sub_assign_op<Scalar>());
+      return m_expression;
+    }
 
+    EIGEN_DEVICE_FUNC
     ExpressionType& expression() const
     {
       return m_expression;
@@ -126,7 +100,7 @@ class NoAlias
 template<typename Derived>
 NoAlias<Derived,MatrixBase> MatrixBase<Derived>::noalias()
 {
-  return derived();
+  return NoAlias<Derived, Eigen::MatrixBase >(derived());
 }
 
 } // end namespace Eigen

Eigen/src/Core/NumTraits.h

@@ -68,21 +68,40 @@ template<typename T> struct GenericNumTraits
                    >::type NonInteger;
   typedef T Nested;
 
-  static inline Real epsilon() { return std::numeric_limits<T>::epsilon(); }
+  EIGEN_DEVICE_FUNC
+  static inline Real epsilon()
+  {
+    #if defined(__CUDA_ARCH__)
+    return internal::device::numeric_limits<T>::epsilon();
+    #else
+    return std::numeric_limits<T>::epsilon();
+    #endif
+  }
+  EIGEN_DEVICE_FUNC
   static inline Real dummy_precision()
   {
     // make sure to override this for floating-point types
     return Real(0);
   }
-  static inline T highest() { return (std::numeric_limits<T>::max)(); }
-  static inline T lowest()  { return IsInteger ? (std::numeric_limits<T>::min)() : (-(std::numeric_limits<T>::max)()); }
-  
-#ifdef EIGEN2_SUPPORT
-  enum {
-    HasFloatingPoint = !IsInteger
-  };
-  typedef NonInteger FloatingPoint;
+
+
+  EIGEN_DEVICE_FUNC
+  static inline T highest() {
+#if defined(__CUDA_ARCH__)
+    return (internal::device::numeric_limits<T>::max)();
+#else
+    return (std::numeric_limits<T>::max)();
 #endif
+  }
+
+  EIGEN_DEVICE_FUNC
+  static inline T lowest()  {
+#if defined(__CUDA_ARCH__)
+    return IsInteger ? (internal::device::numeric_limits<T>::min)() : (-(internal::device::numeric_limits<T>::max)());
+#else
+    return IsInteger ? (std::numeric_limits<T>::min)() : (-(std::numeric_limits<T>::max)());
+#endif
+  }
 };
 
 template<typename T> struct NumTraits : GenericNumTraits<T>
@@ -91,11 +110,13 @@ template<typename T> struct NumTraits : GenericNumTraits<T>
 template<> struct NumTraits<float>
   : GenericNumTraits<float>
 {
+  EIGEN_DEVICE_FUNC
   static inline float dummy_precision() { return 1e-5f; }
 };
 
 template<> struct NumTraits<double> : GenericNumTraits<double>
 {
+  EIGEN_DEVICE_FUNC
   static inline double dummy_precision() { return 1e-12; }
 };
 

Eigen/src/Core/PermutationMatrix.h

@@ -13,7 +13,8 @@
 
 namespace Eigen { 
 
-template<int RowCol,typename IndicesType,typename MatrixType, typename StorageKind> class PermutedImpl;
+// TODO: this does not seems to be needed at all:
+// template<int RowCol,typename IndicesType,typename MatrixType, typename StorageKind> class PermutedImpl;
 
 /** \class PermutationBase
   * \ingroup Core_Module
@@ -41,10 +42,6 @@ template<int RowCol,typename IndicesType,typename MatrixType, typename StorageKi
 
 namespace internal {
 
-template<typename PermutationType, typename MatrixType, int Side, bool Transposed=false>
-struct permut_matrix_product_retval;
-template<typename PermutationType, typename MatrixType, int Side, bool Transposed=false>
-struct permut_sparsematrix_product_retval;
 enum PermPermProduct_t {PermPermProduct};
 
 } // end namespace internal
@@ -60,19 +57,18 @@ class PermutationBase : public EigenBase<Derived>
     typedef typename Traits::IndicesType IndicesType;
     enum {
       Flags = Traits::Flags,
-      CoeffReadCost = Traits::CoeffReadCost,
       RowsAtCompileTime = Traits::RowsAtCompileTime,
       ColsAtCompileTime = Traits::ColsAtCompileTime,
       MaxRowsAtCompileTime = Traits::MaxRowsAtCompileTime,
       MaxColsAtCompileTime = Traits::MaxColsAtCompileTime
     };
-    typedef typename Traits::Scalar Scalar;
-    typedef typename Traits::Index Index;
-    typedef Matrix<Scalar,RowsAtCompileTime,ColsAtCompileTime,0,MaxRowsAtCompileTime,MaxColsAtCompileTime>
+    typedef typename Traits::StorageIndex StorageIndex;
+    typedef Matrix<StorageIndex,RowsAtCompileTime,ColsAtCompileTime,0,MaxRowsAtCompileTime,MaxColsAtCompileTime>
             DenseMatrixType;
-    typedef PermutationMatrix<IndicesType::SizeAtCompileTime,IndicesType::MaxSizeAtCompileTime,Index>
+    typedef PermutationMatrix<IndicesType::SizeAtCompileTime,IndicesType::MaxSizeAtCompileTime,StorageIndex>
             PlainPermutationType;
     using Base::derived;
+    typedef Transpose<PermutationBase> TransposeReturnType;
     #endif
 
     /** Copies the other permutation into *this */
@@ -118,7 +114,7 @@ class PermutationBase : public EigenBase<Derived>
     void evalTo(MatrixBase<DenseDerived>& other) const
     {
       other.setZero();
-      for (int i=0; i<rows();++i)
+      for (Index i=0; i<rows(); ++i)
         other.coeffRef(indices().coeff(i),i) = typename DenseDerived::Scalar(1);
     }
     #endif
@@ -147,7 +143,8 @@ class PermutationBase : public EigenBase<Derived>
     /** Sets *this to be the identity permutation matrix */
     void setIdentity()
     {
-      for(Index i = 0; i < size(); ++i)
+      StorageIndex n = StorageIndex(size());
+      for(StorageIndex i = 0; i < n; ++i)
         indices().coeffRef(i) = i;
     }
 
@@ -163,18 +160,18 @@ class PermutationBase : public EigenBase<Derived>
       *
       * \returns a reference to *this.
       *
-      * \warning This is much slower than applyTranspositionOnTheRight(int,int):
+      * \warning This is much slower than applyTranspositionOnTheRight(Index,Index):
       * this has linear complexity and requires a lot of branching.
       *
-      * \sa applyTranspositionOnTheRight(int,int)
+      * \sa applyTranspositionOnTheRight(Index,Index)
       */
     Derived& applyTranspositionOnTheLeft(Index i, Index j)
     {
       eigen_assert(i>=0 && j>=0 && i<size() && j<size());
       for(Index k = 0; k < size(); ++k)
       {
-        if(indices().coeff(k) == i) indices().coeffRef(k) = j;
-        else if(indices().coeff(k) == j) indices().coeffRef(k) = i;
+        if(indices().coeff(k) == i) indices().coeffRef(k) = StorageIndex(j);
+        else if(indices().coeff(k) == j) indices().coeffRef(k) = StorageIndex(i);
       }
       return derived();
     }
@@ -185,7 +182,7 @@ class PermutationBase : public EigenBase<Derived>
       *
       * This is a fast operation, it only consists in swapping two indices.
       *
-      * \sa applyTranspositionOnTheLeft(int,int)
+      * \sa applyTranspositionOnTheLeft(Index,Index)
       */
     Derived& applyTranspositionOnTheRight(Index i, Index j)
     {
@@ -198,14 +195,14 @@ class PermutationBase : public EigenBase<Derived>
       *
       * \note \note_try_to_help_rvo
       */
-    inline Transpose<PermutationBase> inverse() const
-    { return derived(); }
+    inline TransposeReturnType inverse() const
+    { return TransposeReturnType(derived()); }
     /** \returns the tranpose permutation matrix.
       *
       * \note \note_try_to_help_rvo
       */
-    inline Transpose<PermutationBase> transpose() const
-    { return derived(); }
+    inline TransposeReturnType transpose() const
+    { return TransposeReturnType(derived()); }
 
     /**** multiplication helpers to hopefully get RVO ****/
 
@@ -215,13 +212,13 @@ class PermutationBase : public EigenBase<Derived>
     template<typename OtherDerived>
     void assignTranspose(const PermutationBase<OtherDerived>& other)
     {
-      for (int i=0; i<rows();++i) indices().coeffRef(other.indices().coeff(i)) = i;
+      for (Index i=0; i<rows();++i) indices().coeffRef(other.indices().coeff(i)) = i;
     }
     template<typename Lhs,typename Rhs>
     void assignProduct(const Lhs& lhs, const Rhs& rhs)
     {
       eigen_assert(lhs.cols() == rhs.rows());
-      for (int i=0; i<rows();++i) indices().coeffRef(i) = lhs.indices().coeff(rhs.indices().coeff(i));
+      for (Index i=0; i<rows();++i) indices().coeffRef(i) = lhs.indices().coeff(rhs.indices().coeff(i));
     }
 #endif
 
@@ -250,6 +247,35 @@ class PermutationBase : public EigenBase<Derived>
     template<typename Other> friend
     inline PlainPermutationType operator*(const Transpose<PermutationBase<Other> >& other, const PermutationBase& perm)
     { return PlainPermutationType(internal::PermPermProduct, other.eval(), perm); }
+    
+    /** \returns the determinant of the permutation matrix, which is either 1 or -1 depending on the parity of the permutation.
+      *
+      * This function is O(\c n) procedure allocating a buffer of \c n booleans.
+      */
+    Index determinant() const
+    {
+      Index res = 1;
+      Index n = size();
+      Matrix<bool,RowsAtCompileTime,1,0,MaxRowsAtCompileTime> mask(n);
+      mask.fill(false);
+      Index r = 0;
+      while(r < n)
+      {
+        // search for the next seed
+        while(r<n && mask[r]) r++;
+        if(r>=n)
+          break;
+        // we got one, let's follow it until we are back to the seed
+        Index k0 = r++;
+        mask.coeffRef(k0) = true;
+        for(Index k=indices().coeff(k0); k!=k0; k=indices().coeff(k))
+        {
+          mask.coeffRef(k) = true;
+          res = -res;
+        }
+      }
+      return res;
+    }
 
   protected:
 
@@ -262,7 +288,7 @@ class PermutationBase : public EigenBase<Derived>
   *
   * \param SizeAtCompileTime the number of rows/cols, or Dynamic
   * \param MaxSizeAtCompileTime the maximum number of rows/cols, or Dynamic. This optional parameter defaults to SizeAtCompileTime. Most of the time, you should not have to specify it.
-  * \param IndexType the interger type of the indices
+  * \param StorageIndex the integer type of the indices
   *
   * This class represents a permutation matrix, internally stored as a vector of integers.
   *
@@ -270,24 +296,28 @@ class PermutationBase : public EigenBase<Derived>
   */
 
 namespace internal {
-template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename IndexType>
-struct traits<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, IndexType> >
- : traits<Matrix<IndexType,SizeAtCompileTime,SizeAtCompileTime,0,MaxSizeAtCompileTime,MaxSizeAtCompileTime> >
+template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename _StorageIndex>
+struct traits<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, _StorageIndex> >
+ : traits<Matrix<_StorageIndex,SizeAtCompileTime,SizeAtCompileTime,0,MaxSizeAtCompileTime,MaxSizeAtCompileTime> >
 {
-  typedef IndexType Index;
-  typedef Matrix<IndexType, SizeAtCompileTime, 1, 0, MaxSizeAtCompileTime, 1> IndicesType;
+  typedef PermutationStorage StorageKind;
+  typedef Matrix<_StorageIndex, SizeAtCompileTime, 1, 0, MaxSizeAtCompileTime, 1> IndicesType;
+  typedef _StorageIndex StorageIndex;
 };
 }
 
-template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename IndexType>
-class PermutationMatrix : public PermutationBase<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, IndexType> >
+template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename _StorageIndex>
+class PermutationMatrix : public PermutationBase<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, _StorageIndex> >
 {
     typedef PermutationBase<PermutationMatrix> Base;
     typedef internal::traits<PermutationMatrix> Traits;
   public:
 
+    typedef const PermutationMatrix& Nested;
+
     #ifndef EIGEN_PARSED_BY_DOXYGEN
     typedef typename Traits::IndicesType IndicesType;
+    typedef typename Traits::StorageIndex StorageIndex;
     #endif
 
     inline PermutationMatrix()
@@ -295,8 +325,10 @@ class PermutationMatrix : public PermutationBase<PermutationMatrix<SizeAtCompile
 
     /** Constructs an uninitialized permutation matrix of given size.
       */
-    inline PermutationMatrix(int size) : m_indices(size)
-    {}
+    explicit inline PermutationMatrix(Index size) : m_indices(size)
+    {
+      eigen_internal_assert(size <= NumTraits<StorageIndex>::highest());
+    }
 
     /** Copy constructor. */
     template<typename OtherDerived>
@@ -317,7 +349,7 @@ class PermutationMatrix : public PermutationBase<PermutationMatrix<SizeAtCompile
       * array's size.
       */
     template<typename Other>
-    explicit inline PermutationMatrix(const MatrixBase<Other>& a_indices) : m_indices(a_indices)
+    explicit inline PermutationMatrix(const MatrixBase<Other>& indices) : m_indices(indices)
     {}
 
     /** Convert the Transpositions \a tr to a permutation matrix */
@@ -365,9 +397,12 @@ class PermutationMatrix : public PermutationBase<PermutationMatrix<SizeAtCompile
 #ifndef EIGEN_PARSED_BY_DOXYGEN
     template<typename Other>
     PermutationMatrix(const Transpose<PermutationBase<Other> >& other)
-      : m_indices(other.nestedPermutation().size())
+      : m_indices(other.nestedExpression().size())
     {
-      for (int i=0; i<m_indices.size();++i) m_indices.coeffRef(other.nestedPermutation().indices().coeff(i)) = i;
+      eigen_internal_assert(m_indices.size() <= NumTraits<StorageIndex>::highest());
+      StorageIndex end = StorageIndex(m_indices.size());
+      for (StorageIndex i=0; i<end;++i)
+        m_indices.coeffRef(other.nestedExpression().indices().coeff(i)) = i;
     }
     template<typename Lhs,typename Rhs>
     PermutationMatrix(internal::PermPermProduct_t, const Lhs& lhs, const Rhs& rhs)
@@ -384,18 +419,19 @@ class PermutationMatrix : public PermutationBase<PermutationMatrix<SizeAtCompile
 
 
 namespace internal {
-template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename IndexType, int _PacketAccess>
-struct traits<Map<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, IndexType>,_PacketAccess> >
- : traits<Matrix<IndexType,SizeAtCompileTime,SizeAtCompileTime,0,MaxSizeAtCompileTime,MaxSizeAtCompileTime> >
+template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename _StorageIndex, int _PacketAccess>
+struct traits<Map<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, _StorageIndex>,_PacketAccess> >
+ : traits<Matrix<_StorageIndex,SizeAtCompileTime,SizeAtCompileTime,0,MaxSizeAtCompileTime,MaxSizeAtCompileTime> >
 {
-  typedef IndexType Index;
-  typedef Map<const Matrix<IndexType, SizeAtCompileTime, 1, 0, MaxSizeAtCompileTime, 1>, _PacketAccess> IndicesType;
+  typedef PermutationStorage StorageKind;
+  typedef Map<const Matrix<_StorageIndex, SizeAtCompileTime, 1, 0, MaxSizeAtCompileTime, 1>, _PacketAccess> IndicesType;
+  typedef _StorageIndex StorageIndex;
 };
 }
 
-template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename IndexType, int _PacketAccess>
-class Map<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, IndexType>,_PacketAccess>
-  : public PermutationBase<Map<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, IndexType>,_PacketAccess> >
+template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename _StorageIndex, int _PacketAccess>
+class Map<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, _StorageIndex>,_PacketAccess>
+  : public PermutationBase<Map<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, _StorageIndex>,_PacketAccess> >
 {
     typedef PermutationBase<Map> Base;
     typedef internal::traits<Map> Traits;
@@ -403,14 +439,14 @@ class Map<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, IndexType>,
 
     #ifndef EIGEN_PARSED_BY_DOXYGEN
     typedef typename Traits::IndicesType IndicesType;
-    typedef typename IndicesType::Scalar Index;
+    typedef typename IndicesType::Scalar StorageIndex;
     #endif
 
-    inline Map(const Index* indicesPtr)
+    inline Map(const StorageIndex* indicesPtr)
       : m_indices(indicesPtr)
     {}
 
-    inline Map(const Index* indicesPtr, Index size)
+    inline Map(const StorageIndex* indicesPtr, Index size)
       : m_indices(indicesPtr,size)
     {}
 
@@ -457,8 +493,6 @@ class Map<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, IndexType>,
   * \sa class PermutationBase, class PermutationMatrix
   */
 
-struct PermutationStorage {};
-
 template<typename _IndicesType> class TranspositionsWrapper;
 namespace internal {
 template<typename _IndicesType>
@@ -466,15 +500,14 @@ struct traits<PermutationWrapper<_IndicesType> >
 {
   typedef PermutationStorage StorageKind;
   typedef typename _IndicesType::Scalar Scalar;
-  typedef typename _IndicesType::Scalar Index;
+  typedef typename _IndicesType::Scalar StorageIndex;
   typedef _IndicesType IndicesType;
   enum {
     RowsAtCompileTime = _IndicesType::SizeAtCompileTime,
     ColsAtCompileTime = _IndicesType::SizeAtCompileTime,
-    MaxRowsAtCompileTime = IndicesType::MaxRowsAtCompileTime,
-    MaxColsAtCompileTime = IndicesType::MaxColsAtCompileTime,
-    Flags = 0,
-    CoeffReadCost = _IndicesType::CoeffReadCost
+    MaxRowsAtCompileTime = IndicesType::MaxSizeAtCompileTime,
+    MaxColsAtCompileTime = IndicesType::MaxSizeAtCompileTime,
+    Flags = 0
   };
 };
 }
@@ -490,8 +523,8 @@ class PermutationWrapper : public PermutationBase<PermutationWrapper<_IndicesTyp
     typedef typename Traits::IndicesType IndicesType;
     #endif
 
-    inline PermutationWrapper(const IndicesType& a_indices)
-      : m_indices(a_indices)
+    inline PermutationWrapper(const IndicesType& indices)
+      : m_indices(indices)
     {}
 
     /** const version of indices(). */
@@ -503,104 +536,33 @@ class PermutationWrapper : public PermutationBase<PermutationWrapper<_IndicesTyp
     typename IndicesType::Nested m_indices;
 };
 
+
 /** \returns the matrix with the permutation applied to the columns.
   */
-template<typename Derived, typename PermutationDerived>
-inline const internal::permut_matrix_product_retval<PermutationDerived, Derived, OnTheRight>
-operator*(const MatrixBase<Derived>& matrix,
-          const PermutationBase<PermutationDerived> &permutation)
+template<typename MatrixDerived, typename PermutationDerived>
+EIGEN_DEVICE_FUNC
+const Product<MatrixDerived, PermutationDerived, AliasFreeProduct>
+operator*(const MatrixBase<MatrixDerived> &matrix,
+          const PermutationBase<PermutationDerived>& permutation)
 {
-  return internal::permut_matrix_product_retval
-           <PermutationDerived, Derived, OnTheRight>
-           (permutation.derived(), matrix.derived());
+  return Product<MatrixDerived, PermutationDerived, AliasFreeProduct>
+            (matrix.derived(), permutation.derived());
 }
 
 /** \returns the matrix with the permutation applied to the rows.
   */
-template<typename Derived, typename PermutationDerived>
-inline const internal::permut_matrix_product_retval
-               <PermutationDerived, Derived, OnTheLeft>
+template<typename PermutationDerived, typename MatrixDerived>
+EIGEN_DEVICE_FUNC
+const Product<PermutationDerived, MatrixDerived, AliasFreeProduct>
 operator*(const PermutationBase<PermutationDerived> &permutation,
-          const MatrixBase<Derived>& matrix)
+          const MatrixBase<MatrixDerived>& matrix)
 {
-  return internal::permut_matrix_product_retval
-           <PermutationDerived, Derived, OnTheLeft>
-           (permutation.derived(), matrix.derived());
+  return Product<PermutationDerived, MatrixDerived, AliasFreeProduct>
+            (permutation.derived(), matrix.derived());
 }
 
 namespace internal {
 
-template<typename PermutationType, typename MatrixType, int Side, bool Transposed>
-struct traits<permut_matrix_product_retval<PermutationType, MatrixType, Side, Transposed> >
-{
-  typedef typename MatrixType::PlainObject ReturnType;
-};
-
-template<typename PermutationType, typename MatrixType, int Side, bool Transposed>
-struct permut_matrix_product_retval
- : public ReturnByValue<permut_matrix_product_retval<PermutationType, MatrixType, Side, Transposed> >
-{
-    typedef typename remove_all<typename MatrixType::Nested>::type MatrixTypeNestedCleaned;
-    typedef typename MatrixType::Index Index;
-
-    permut_matrix_product_retval(const PermutationType& perm, const MatrixType& matrix)
-      : m_permutation(perm), m_matrix(matrix)
-    {}
-
-    inline Index rows() const { return m_matrix.rows(); }
-    inline Index cols() const { return m_matrix.cols(); }
-
-    template<typename Dest> inline void evalTo(Dest& dst) const
-    {
-      const Index n = Side==OnTheLeft ? rows() : cols();
-
-      if(is_same<MatrixTypeNestedCleaned,Dest>::value && extract_data(dst) == extract_data(m_matrix))
-      {
-        // apply the permutation inplace
-        Matrix<bool,PermutationType::RowsAtCompileTime,1,0,PermutationType::MaxRowsAtCompileTime> mask(m_permutation.size());
-        mask.fill(false);
-        Index r = 0;
-        while(r < m_permutation.size())
-        {
-          // search for the next seed
-          while(r<m_permutation.size() && mask[r]) r++;
-          if(r>=m_permutation.size())
-            break;
-          // we got one, let's follow it until we are back to the seed
-          Index k0 = r++;
-          Index kPrev = k0;
-          mask.coeffRef(k0) = true;
-          for(Index k=m_permutation.indices().coeff(k0); k!=k0; k=m_permutation.indices().coeff(k))
-          {
-                  Block<Dest, Side==OnTheLeft ? 1 : Dest::RowsAtCompileTime, Side==OnTheRight ? 1 : Dest::ColsAtCompileTime>(dst, k)
-            .swap(Block<Dest, Side==OnTheLeft ? 1 : Dest::RowsAtCompileTime, Side==OnTheRight ? 1 : Dest::ColsAtCompileTime>
-                       (dst,((Side==OnTheLeft) ^ Transposed) ? k0 : kPrev));
-
-            mask.coeffRef(k) = true;
-            kPrev = k;
-          }
-        }
-      }
-      else
-      {
-        for(int i = 0; i < n; ++i)
-        {
-          Block<Dest, Side==OnTheLeft ? 1 : Dest::RowsAtCompileTime, Side==OnTheRight ? 1 : Dest::ColsAtCompileTime>
-               (dst, ((Side==OnTheLeft) ^ Transposed) ? m_permutation.indices().coeff(i) : i)
-
-          =
-
-          Block<const MatrixTypeNestedCleaned,Side==OnTheLeft ? 1 : MatrixType::RowsAtCompileTime,Side==OnTheRight ? 1 : MatrixType::ColsAtCompileTime>
-               (m_matrix, ((Side==OnTheRight) ^ Transposed) ? m_permutation.indices().coeff(i) : i);
-        }
-      }
-    }
-
-  protected:
-    const PermutationType& m_permutation;
-    typename MatrixType::Nested m_matrix;
-};
-
 /* Template partial specialization for transposed/inverse permutations */
 
 template<typename Derived>
@@ -610,6 +572,8 @@ struct traits<Transpose<PermutationBase<Derived> > >
 
 } // end namespace internal
 
+// TODO: the specificties should be handled by the evaluator,
+// at the very least we should only specialize TransposeImpl
 template<typename Derived>
 class Transpose<PermutationBase<Derived> >
   : public EigenBase<Transpose<PermutationBase<Derived> > >
@@ -624,26 +588,26 @@ class Transpose<PermutationBase<Derived> >
     typedef typename Derived::DenseMatrixType DenseMatrixType;
     enum {
       Flags = Traits::Flags,
-      CoeffReadCost = Traits::CoeffReadCost,
       RowsAtCompileTime = Traits::RowsAtCompileTime,
       ColsAtCompileTime = Traits::ColsAtCompileTime,
       MaxRowsAtCompileTime = Traits::MaxRowsAtCompileTime,
       MaxColsAtCompileTime = Traits::MaxColsAtCompileTime
     };
     typedef typename Traits::Scalar Scalar;
+    typedef typename Traits::StorageIndex StorageIndex;
     #endif
 
     Transpose(const PermutationType& p) : m_permutation(p) {}
 
-    inline int rows() const { return m_permutation.rows(); }
-    inline int cols() const { return m_permutation.cols(); }
+    inline Index rows() const { return m_permutation.rows(); }
+    inline Index cols() const { return m_permutation.cols(); }
 
     #ifndef EIGEN_PARSED_BY_DOXYGEN
     template<typename DenseDerived>
     void evalTo(MatrixBase<DenseDerived>& other) const
     {
       other.setZero();
-      for (int i=0; i<rows();++i)
+      for (Index i=0; i<rows();++i)
         other.coeffRef(i, m_permutation.indices().coeff(i)) = typename DenseDerived::Scalar(1);
     }
     #endif
@@ -656,22 +620,22 @@ class Transpose<PermutationBase<Derived> >
     /** \returns the matrix with the inverse permutation applied to the columns.
       */
     template<typename OtherDerived> friend
-    inline const internal::permut_matrix_product_retval<PermutationType, OtherDerived, OnTheRight, true>
+    const Product<OtherDerived, Transpose, AliasFreeProduct>
     operator*(const MatrixBase<OtherDerived>& matrix, const Transpose& trPerm)
     {
-      return internal::permut_matrix_product_retval<PermutationType, OtherDerived, OnTheRight, true>(trPerm.m_permutation, matrix.derived());
+      return Product<OtherDerived, Transpose, AliasFreeProduct>(matrix.derived(), trPerm.derived());
     }
 
     /** \returns the matrix with the inverse permutation applied to the rows.
       */
     template<typename OtherDerived>
-    inline const internal::permut_matrix_product_retval<PermutationType, OtherDerived, OnTheLeft, true>
+    const Product<Transpose, OtherDerived, AliasFreeProduct>
     operator*(const MatrixBase<OtherDerived>& matrix) const
     {
-      return internal::permut_matrix_product_retval<PermutationType, OtherDerived, OnTheLeft, true>(m_permutation, matrix.derived());
+      return Product<Transpose, OtherDerived, AliasFreeProduct>(*this, matrix.derived());
     }
 
-    const PermutationType& nestedPermutation() const { return m_permutation; }
+    const PermutationType& nestedExpression() const { return m_permutation; }
 
   protected:
     const PermutationType& m_permutation;
@@ -683,6 +647,12 @@ const PermutationWrapper<const Derived> MatrixBase<Derived>::asPermutation() con
   return derived();
 }
 
+namespace internal {
+
+template<> struct AssignmentKind<DenseShape,PermutationShape> { typedef EigenBase2EigenBase Kind; };
+
+} // end namespace internal
+
 } // end namespace Eigen
 
 #endif // EIGEN_PERMUTATIONMATRIX_H

Eigen/src/Core/PlainObjectBase.h

@@ -28,6 +28,7 @@ namespace internal {
 
 template<int MaxSizeAtCompileTime> struct check_rows_cols_for_overflow {
   template<typename Index>
+  EIGEN_DEVICE_FUNC
   static EIGEN_ALWAYS_INLINE void run(Index, Index)
   {
   }
@@ -35,6 +36,7 @@ template<int MaxSizeAtCompileTime> struct check_rows_cols_for_overflow {
 
 template<> struct check_rows_cols_for_overflow<Dynamic> {
   template<typename Index>
+  EIGEN_DEVICE_FUNC
   static EIGEN_ALWAYS_INLINE void run(Index rows, Index cols)
   {
     // http://hg.mozilla.org/mozilla-central/file/6c8a909977d3/xpcom/ds/CheckedInt.h#l242
@@ -47,7 +49,10 @@ template<> struct check_rows_cols_for_overflow<Dynamic> {
   }
 };
 
-template <typename Derived, typename OtherDerived = Derived, bool IsVector = bool(Derived::IsVectorAtCompileTime)> struct conservative_resize_like_impl;
+template <typename Derived,
+          typename OtherDerived = Derived,
+          bool IsVector = bool(Derived::IsVectorAtCompileTime) && bool(OtherDerived::IsVectorAtCompileTime)>
+struct conservative_resize_like_impl;
 
 template<typename MatrixTypeA, typename MatrixTypeB, bool SwapPointers> struct matrix_swap_impl;
 
@@ -64,8 +69,9 @@ template<typename MatrixTypeA, typename MatrixTypeB, bool SwapPointers> struct m
 #ifdef EIGEN_PARSED_BY_DOXYGEN
 namespace internal {
 
-// this is a warkaround to doxygen not being able to understand the inheritence logic
+// this is a workaround to doxygen not being able to understand the inheritance logic
 // when it is hidden by the dense_xpr_base helper struct.
+/** This class is just a workaround for Doxygen and it does not not actually exist. */
 template<typename Derived> struct dense_xpr_base_dispatcher_for_doxygen;// : public MatrixBase<Derived> {};
 /** This class is just a workaround for Doxygen and it does not not actually exist. */
 template<typename _Scalar, int _Rows, int _Cols, int _Options, int _MaxRows, int _MaxCols>
@@ -90,8 +96,8 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
     typedef typename internal::dense_xpr_base<Derived>::type Base;
 
     typedef typename internal::traits<Derived>::StorageKind StorageKind;
-    typedef typename internal::traits<Derived>::Index Index;
     typedef typename internal::traits<Derived>::Scalar Scalar;
+    
     typedef typename internal::packet_traits<Scalar>::type PacketScalar;
     typedef typename NumTraits<Scalar>::Real RealScalar;
     typedef Derived DenseType;
@@ -110,28 +116,36 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
     typedef Eigen::Map<Derived, Unaligned>  MapType;
     friend  class Eigen::Map<const Derived, Unaligned>;
     typedef const Eigen::Map<const Derived, Unaligned> ConstMapType;
-    friend  class Eigen::Map<Derived, Aligned>;
-    typedef Eigen::Map<Derived, Aligned> AlignedMapType;
-    friend  class Eigen::Map<const Derived, Aligned>;
-    typedef const Eigen::Map<const Derived, Aligned> ConstAlignedMapType;
+#if EIGEN_MAX_ALIGN_BYTES>0
+    // for EIGEN_MAX_ALIGN_BYTES==0, AlignedMax==Unaligned, and many compilers generate warnings for friend-ing a class twice.
+    friend  class Eigen::Map<Derived, AlignedMax>;
+    friend  class Eigen::Map<const Derived, AlignedMax>;
+#endif
+    typedef Eigen::Map<Derived, AlignedMax> AlignedMapType;
+    typedef const Eigen::Map<const Derived, AlignedMax> ConstAlignedMapType;
     template<typename StrideType> struct StridedMapType { typedef Eigen::Map<Derived, Unaligned, StrideType> type; };
     template<typename StrideType> struct StridedConstMapType { typedef Eigen::Map<const Derived, Unaligned, StrideType> type; };
-    template<typename StrideType> struct StridedAlignedMapType { typedef Eigen::Map<Derived, Aligned, StrideType> type; };
-    template<typename StrideType> struct StridedConstAlignedMapType { typedef Eigen::Map<const Derived, Aligned, StrideType> type; };
+    template<typename StrideType> struct StridedAlignedMapType { typedef Eigen::Map<Derived, AlignedMax, StrideType> type; };
+    template<typename StrideType> struct StridedConstAlignedMapType { typedef Eigen::Map<const Derived, AlignedMax, StrideType> type; };
 
   protected:
     DenseStorage<Scalar, Base::MaxSizeAtCompileTime, Base::RowsAtCompileTime, Base::ColsAtCompileTime, Options> m_storage;
 
   public:
-    enum { NeedsToAlign = SizeAtCompileTime != Dynamic && (internal::traits<Derived>::Flags & AlignedBit) != 0 };
+    enum { NeedsToAlign = (SizeAtCompileTime != Dynamic) && (internal::traits<Derived>::Alignment>0) };
     EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign)
 
+    EIGEN_DEVICE_FUNC
     Base& base() { return *static_cast<Base*>(this); }
+    EIGEN_DEVICE_FUNC
     const Base& base() const { return *static_cast<const Base*>(this); }
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Index rows() const { return m_storage.rows(); }
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Index cols() const { return m_storage.cols(); }
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE const Scalar& coeff(Index rowId, Index colId) const
     {
       if(Flags & RowMajorBit)
@@ -140,11 +154,13 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
         return m_storage.data()[rowId + colId * m_storage.rows()];
     }
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE const Scalar& coeff(Index index) const
     {
       return m_storage.data()[index];
     }
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Scalar& coeffRef(Index rowId, Index colId)
     {
       if(Flags & RowMajorBit)
@@ -153,11 +169,13 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
         return m_storage.data()[rowId + colId * m_storage.rows()];
     }
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Scalar& coeffRef(Index index)
     {
       return m_storage.data()[index];
     }
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE const Scalar& coeffRef(Index rowId, Index colId) const
     {
       if(Flags & RowMajorBit)
@@ -166,6 +184,7 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
         return m_storage.data()[rowId + colId * m_storage.rows()];
     }
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE const Scalar& coeffRef(Index index) const
     {
       return m_storage.data()[index];
@@ -206,11 +225,11 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
     }
 
     /** \returns a const pointer to the data array of this matrix */
-    EIGEN_STRONG_INLINE const Scalar *data() const
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar *data() const
     { return m_storage.data(); }
 
     /** \returns a pointer to the data array of this matrix */
-    EIGEN_STRONG_INLINE Scalar *data()
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar *data()
     { return m_storage.data(); }
 
     /** Resizes \c *this to a \a rows x \a cols matrix.
@@ -229,22 +248,23 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       *
       * \sa resize(Index) for vectors, resize(NoChange_t, Index), resize(Index, NoChange_t)
       */
-    EIGEN_STRONG_INLINE void resize(Index nbRows, Index nbCols)
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void resize(Index rows, Index cols)
     {
-      eigen_assert(   EIGEN_IMPLIES(RowsAtCompileTime!=Dynamic,nbRows==RowsAtCompileTime)
-                   && EIGEN_IMPLIES(ColsAtCompileTime!=Dynamic,nbCols==ColsAtCompileTime)
-                   && EIGEN_IMPLIES(RowsAtCompileTime==Dynamic && MaxRowsAtCompileTime!=Dynamic,nbRows<=MaxRowsAtCompileTime)
-                   && EIGEN_IMPLIES(ColsAtCompileTime==Dynamic && MaxColsAtCompileTime!=Dynamic,nbCols<=MaxColsAtCompileTime)
-                   && nbRows>=0 && nbCols>=0 && "Invalid sizes when resizing a matrix or array.");
-      internal::check_rows_cols_for_overflow<MaxSizeAtCompileTime>::run(nbRows, nbCols);
+      eigen_assert(   EIGEN_IMPLIES(RowsAtCompileTime!=Dynamic,rows==RowsAtCompileTime)
+                   && EIGEN_IMPLIES(ColsAtCompileTime!=Dynamic,cols==ColsAtCompileTime)
+                   && EIGEN_IMPLIES(RowsAtCompileTime==Dynamic && MaxRowsAtCompileTime!=Dynamic,rows<=MaxRowsAtCompileTime)
+                   && EIGEN_IMPLIES(ColsAtCompileTime==Dynamic && MaxColsAtCompileTime!=Dynamic,cols<=MaxColsAtCompileTime)
+                   && rows>=0 && cols>=0 && "Invalid sizes when resizing a matrix or array.");
+      internal::check_rows_cols_for_overflow<MaxSizeAtCompileTime>::run(rows, cols);
       #ifdef EIGEN_INITIALIZE_COEFFS
-        Index size = nbRows*nbCols;
+        Index size = rows*cols;
         bool size_changed = size != this->size();
-        m_storage.resize(size, nbRows, nbCols);
+        m_storage.resize(size, rows, cols);
         if(size_changed) EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED
       #else
-        internal::check_rows_cols_for_overflow<MaxSizeAtCompileTime>::run(nbRows, nbCols);
-        m_storage.resize(nbRows*nbCols, nbRows, nbCols);
+        internal::check_rows_cols_for_overflow<MaxSizeAtCompileTime>::run(rows, cols);
+        m_storage.resize(rows*cols, rows, cols);
       #endif
     }
 
@@ -259,6 +279,7 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       *
       * \sa resize(Index,Index), resize(NoChange_t, Index), resize(Index, NoChange_t)
       */
+    EIGEN_DEVICE_FUNC
     inline void resize(Index size)
     {
       EIGEN_STATIC_ASSERT_VECTOR_ONLY(PlainObjectBase)
@@ -283,9 +304,10 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       *
       * \sa resize(Index,Index)
       */
-    inline void resize(NoChange_t, Index nbCols)
+    EIGEN_DEVICE_FUNC
+    inline void resize(NoChange_t, Index cols)
     {
-      resize(rows(), nbCols);
+      resize(rows(), cols);
     }
 
     /** Resizes the matrix, changing only the number of rows. For the parameter of type NoChange_t, just pass the special value \c NoChange
@@ -296,9 +318,10 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       *
       * \sa resize(Index,Index)
       */
-    inline void resize(Index nbRows, NoChange_t)
+    EIGEN_DEVICE_FUNC
+    inline void resize(Index rows, NoChange_t)
     {
-      resize(nbRows, cols());
+      resize(rows, cols());
     }
 
     /** Resizes \c *this to have the same dimensions as \a other.
@@ -309,6 +332,7 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       * remain row-vectors and vectors remain vectors.
       */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC 
     EIGEN_STRONG_INLINE void resizeLike(const EigenBase<OtherDerived>& _other)
     {
       const OtherDerived& other = _other.derived();
@@ -336,9 +360,10 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       * Matrices are resized relative to the top-left element. In case values need to be 
       * appended to the matrix they will be uninitialized.
       */
-    EIGEN_STRONG_INLINE void conservativeResize(Index nbRows, Index nbCols)
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void conservativeResize(Index rows, Index cols)
     {
-      internal::conservative_resize_like_impl<Derived>::run(*this, nbRows, nbCols);
+      internal::conservative_resize_like_impl<Derived>::run(*this, rows, cols);
     }
 
     /** Resizes the matrix to \a rows x \a cols while leaving old values untouched.
@@ -348,10 +373,11 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       *
       * In case the matrix is growing, new rows will be uninitialized.
       */
-    EIGEN_STRONG_INLINE void conservativeResize(Index nbRows, NoChange_t)
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void conservativeResize(Index rows, NoChange_t)
     {
       // Note: see the comment in conservativeResize(Index,Index)
-      conservativeResize(nbRows, cols());
+      conservativeResize(rows, cols());
     }
 
     /** Resizes the matrix to \a rows x \a cols while leaving old values untouched.
@@ -361,10 +387,11 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       *
       * In case the matrix is growing, new columns will be uninitialized.
       */
-    EIGEN_STRONG_INLINE void conservativeResize(NoChange_t, Index nbCols)
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void conservativeResize(NoChange_t, Index cols)
     {
       // Note: see the comment in conservativeResize(Index,Index)
-      conservativeResize(rows(), nbCols);
+      conservativeResize(rows(), cols);
     }
 
     /** Resizes the vector to \a size while retaining old values.
@@ -375,6 +402,7 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       *
       * When values are appended, they will be uninitialized.
       */
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE void conservativeResize(Index size)
     {
       internal::conservative_resize_like_impl<Derived>::run(*this, size);
@@ -390,6 +418,7 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       * appended to the matrix they will copied from \c other.
       */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE void conservativeResizeLike(const DenseBase<OtherDerived>& other)
     {
       internal::conservative_resize_like_impl<Derived,OtherDerived>::run(*this, other);
@@ -398,6 +427,7 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
     /** This is a special case of the templated operator=. Its purpose is to
       * prevent a default operator= from hiding the templated operator=.
       */
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Derived& operator=(const PlainObjectBase& other)
     {
       return _set(other);
@@ -405,6 +435,7 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
 
     /** \sa MatrixBase::lazyAssign() */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Derived& lazyAssign(const DenseBase<OtherDerived>& other)
     {
       _resize_to_match(other);
@@ -412,12 +443,14 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
     }
 
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE Derived& operator=(const ReturnByValue<OtherDerived>& func)
     {
       resize(func.rows(), func.cols());
       return Base::operator=(func);
     }
 
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE PlainObjectBase() : m_storage()
     {
 //       _check_template_params();
@@ -427,15 +460,37 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
 #ifndef EIGEN_PARSED_BY_DOXYGEN
     // FIXME is it still needed ?
     /** \internal */
-    PlainObjectBase(internal::constructor_without_unaligned_array_assert)
+    EIGEN_DEVICE_FUNC
+    explicit PlainObjectBase(internal::constructor_without_unaligned_array_assert)
       : m_storage(internal::constructor_without_unaligned_array_assert())
     {
 //       _check_template_params(); EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED
     }
 #endif
 
-    EIGEN_STRONG_INLINE PlainObjectBase(Index a_size, Index nbRows, Index nbCols)
-      : m_storage(a_size, nbRows, nbCols)
+#ifdef EIGEN_HAVE_RVALUE_REFERENCES
+    EIGEN_DEVICE_FUNC
+    PlainObjectBase(PlainObjectBase&& other)
+      : m_storage( std::move(other.m_storage) )
+    {
+    }
+
+    EIGEN_DEVICE_FUNC
+    PlainObjectBase& operator=(PlainObjectBase&& other)
+    {
+      using std::swap;
+      swap(m_storage, other.m_storage);
+      return *this;
+    }
+#endif
+
+    /** Copy constructor */
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE PlainObjectBase(const PlainObjectBase& other)
+      : Base(), m_storage(other.m_storage) { }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE PlainObjectBase(Index size, Index rows, Index cols)
+      : m_storage(size, rows, cols)
     {
 //       _check_template_params();
 //       EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED
@@ -444,6 +499,7 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
     /** \copydoc MatrixBase::operator=(const EigenBase<OtherDerived>&)
       */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC 
     EIGEN_STRONG_INLINE Derived& operator=(const EigenBase<OtherDerived> &other)
     {
       _resize_to_match(other);
@@ -451,14 +507,36 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       return this->derived();
     }
 
-    /** \sa MatrixBase::operator=(const EigenBase<OtherDerived>&) */
+    /** \sa PlainObjectBase::operator=(const EigenBase<OtherDerived>&) */
     template<typename OtherDerived>
-    EIGEN_STRONG_INLINE PlainObjectBase(const EigenBase<OtherDerived> &other)
-      : m_storage(other.derived().rows() * other.derived().cols(), other.derived().rows(), other.derived().cols())
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE PlainObjectBase(const DenseBase<OtherDerived> &other)
+      : m_storage()
     {
       _check_template_params();
-      internal::check_rows_cols_for_overflow<MaxSizeAtCompileTime>::run(other.derived().rows(), other.derived().cols());
-      Base::operator=(other.derived());
+      resizeLike(other);
+      _set_noalias(other);
+    }
+
+    /** \sa PlainObjectBase::operator=(const EigenBase<OtherDerived>&) */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC 
+    EIGEN_STRONG_INLINE PlainObjectBase(const EigenBase<OtherDerived> &other)
+      : m_storage()
+    {
+      _check_template_params();
+      resizeLike(other);
+      *this = other.derived();
+    }
+    /** \brief Copy constructor with in-place evaluation */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE PlainObjectBase(const ReturnByValue<OtherDerived>& other)
+    {
+      _check_template_params();
+      // FIXME this does not automatically transpose vectors if necessary
+      resize(other.rows(), other.cols());
+      other.evalTo(this->derived());
     }
 
     /** \name Map
@@ -535,16 +613,16 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
     //@}
 
     using Base::setConstant;
-    Derived& setConstant(Index size, const Scalar& value);
-    Derived& setConstant(Index rows, Index cols, const Scalar& value);
+    EIGEN_DEVICE_FUNC Derived& setConstant(Index size, const Scalar& value);
+    EIGEN_DEVICE_FUNC Derived& setConstant(Index rows, Index cols, const Scalar& value);
 
     using Base::setZero;
-    Derived& setZero(Index size);
-    Derived& setZero(Index rows, Index cols);
+    EIGEN_DEVICE_FUNC Derived& setZero(Index size);
+    EIGEN_DEVICE_FUNC Derived& setZero(Index rows, Index cols);
 
     using Base::setOnes;
-    Derived& setOnes(Index size);
-    Derived& setOnes(Index rows, Index cols);
+    EIGEN_DEVICE_FUNC Derived& setOnes(Index size);
+    EIGEN_DEVICE_FUNC Derived& setOnes(Index rows, Index cols);
 
     using Base::setRandom;
     Derived& setRandom(Index size);
@@ -563,6 +641,7 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       * remain row-vectors and vectors remain vectors.
       */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC 
     EIGEN_STRONG_INLINE void _resize_to_match(const EigenBase<OtherDerived>& other)
     {
       #ifdef EIGEN_NO_AUTOMATIC_RESIZING
@@ -589,25 +668,23 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       *
       * \internal
       */
+    // aliasing is dealt once in internall::call_assignment
+    // so at this stage we have to assume aliasing... and resising has to be done later.
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC 
     EIGEN_STRONG_INLINE Derived& _set(const DenseBase<OtherDerived>& other)
     {
-      _set_selector(other.derived(), typename internal::conditional<static_cast<bool>(int(OtherDerived::Flags) & EvalBeforeAssigningBit), internal::true_type, internal::false_type>::type());
+      internal::call_assignment(this->derived(), other.derived());
       return this->derived();
     }
 
-    template<typename OtherDerived>
-    EIGEN_STRONG_INLINE void _set_selector(const OtherDerived& other, const internal::true_type&) { _set_noalias(other.eval()); }
-
-    template<typename OtherDerived>
-    EIGEN_STRONG_INLINE void _set_selector(const OtherDerived& other, const internal::false_type&) { _set_noalias(other); }
-
     /** \internal Like _set() but additionally makes the assumption that no aliasing effect can happen (which
       * is the case when creating a new matrix) so one can enforce lazy evaluation.
       *
       * \sa operator=(const MatrixBase<OtherDerived>&), _set()
       */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC 
     EIGEN_STRONG_INLINE Derived& _set_noalias(const DenseBase<OtherDerived>& other)
     {
       // I don't think we need this resize call since the lazyAssign will anyways resize
@@ -615,40 +692,166 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       //_resize_to_match(other);
       // the 'false' below means to enforce lazy evaluation. We don't use lazyAssign() because
       // it wouldn't allow to copy a row-vector into a column-vector.
-      return internal::assign_selector<Derived,OtherDerived,false>::run(this->derived(), other.derived());
+      internal::call_assignment_no_alias(this->derived(), other.derived(), internal::assign_op<Scalar>());
+      return this->derived();
     }
 
     template<typename T0, typename T1>
-    EIGEN_STRONG_INLINE void _init2(Index nbRows, Index nbCols, typename internal::enable_if<Base::SizeAtCompileTime!=2,T0>::type* = 0)
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void _init2(Index rows, Index cols, typename internal::enable_if<Base::SizeAtCompileTime!=2,T0>::type* = 0)
     {
       EIGEN_STATIC_ASSERT(bool(NumTraits<T0>::IsInteger) &&
                           bool(NumTraits<T1>::IsInteger),
                           FLOATING_POINT_ARGUMENT_PASSED__INTEGER_WAS_EXPECTED)
-      resize(nbRows,nbCols);
+      resize(rows,cols);
     }
+    
     template<typename T0, typename T1>
+    EIGEN_DEVICE_FUNC 
     EIGEN_STRONG_INLINE void _init2(const Scalar& val0, const Scalar& val1, typename internal::enable_if<Base::SizeAtCompileTime==2,T0>::type* = 0)
     {
       EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(PlainObjectBase, 2)
       m_storage.data()[0] = val0;
       m_storage.data()[1] = val1;
     }
+    
+    template<typename T0, typename T1>
+    EIGEN_DEVICE_FUNC 
+    EIGEN_STRONG_INLINE void _init2(const Index& val0, const Index& val1,
+                                    typename internal::enable_if<    (!internal::is_same<Index,Scalar>::value)
+                                                                  && (internal::is_same<T0,Index>::value)
+                                                                  && (internal::is_same<T1,Index>::value)
+                                                                  && Base::SizeAtCompileTime==2,T1>::type* = 0)
+    {
+      EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(PlainObjectBase, 2)
+      m_storage.data()[0] = Scalar(val0);
+      m_storage.data()[1] = Scalar(val1);
+    }
 
+    // The argument is convertible to the Index type and we either have a non 1x1 Matrix, or a dynamic-sized Array,
+    // then the argument is meant to be the size of the object.
+    template<typename T>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void _init1(Index size, typename internal::enable_if<    (Base::SizeAtCompileTime!=1 || !internal::is_convertible<T, Scalar>::value)
+                                                                              && ((!internal::is_same<typename internal::traits<Derived>::XprKind,ArrayXpr>::value || Base::SizeAtCompileTime==Dynamic)),T>::type* = 0)
+    {
+      // NOTE MSVC 2008 complains if we directly put bool(NumTraits<T>::IsInteger) as the EIGEN_STATIC_ASSERT argument.
+      const bool is_integer = NumTraits<T>::IsInteger;
+      EIGEN_STATIC_ASSERT(is_integer,
+                          FLOATING_POINT_ARGUMENT_PASSED__INTEGER_WAS_EXPECTED)
+      resize(size);
+    }
+    
+    // We have a 1x1 matrix/array => the argument is interpreted as the value of the unique coefficient (case where scalar type can be implicitely converted)
+    template<typename T>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void _init1(const Scalar& val0, typename internal::enable_if<Base::SizeAtCompileTime==1 && internal::is_convertible<T, Scalar>::value,T>::type* = 0)
+    {
+      EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(PlainObjectBase, 1)
+      m_storage.data()[0] = val0;
+    }
+    
+    // We have a 1x1 matrix/array => the argument is interpreted as the value of the unique coefficient (case where scalar type match the index type)
+    template<typename T>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void _init1(const Index& val0,
+                                    typename internal::enable_if<    (!internal::is_same<Index,Scalar>::value)
+                                                                  && (internal::is_same<Index,T>::value)
+                                                                  && Base::SizeAtCompileTime==1
+                                                                  && internal::is_convertible<T, Scalar>::value,T*>::type* = 0)
+    {
+      EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(PlainObjectBase, 1)
+      m_storage.data()[0] = Scalar(val0);
+    }
+
+    // Initialize a fixed size matrix from a pointer to raw data
+    template<typename T>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void _init1(const Scalar* data){
+      this->_set_noalias(ConstMapType(data));
+    }
+
+    // Initialize an arbitrary matrix from a dense expression
+    template<typename T, typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void _init1(const DenseBase<OtherDerived>& other){
+      this->_set_noalias(other);
+    }
+
+    // Initialize an arbitrary matrix from a generic Eigen expression
+    template<typename T, typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void _init1(const EigenBase<OtherDerived>& other){
+      this->derived() = other;
+    }
+
+    template<typename T, typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void _init1(const ReturnByValue<OtherDerived>& other)
+    {
+      resize(other.rows(), other.cols());
+      other.evalTo(this->derived());
+    }
+
+    template<typename T, typename OtherDerived, int ColsAtCompileTime>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void _init1(const RotationBase<OtherDerived,ColsAtCompileTime>& r)
+    {
+      this->derived() = r;
+    }
+    
+    // For fixed -size arrays:
+    template<typename T>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void _init1(const Scalar& val0,
+                                    typename internal::enable_if<    Base::SizeAtCompileTime!=Dynamic
+                                                                  && Base::SizeAtCompileTime!=1
+                                                                  && internal::is_convertible<T, Scalar>::value
+                                                                  && internal::is_same<typename internal::traits<Derived>::XprKind,ArrayXpr>::value,T>::type* = 0)
+    {
+      Base::setConstant(val0);
+    }
+    
+    template<typename T>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void _init1(const Index& val0,
+                                    typename internal::enable_if<    (!internal::is_same<Index,Scalar>::value)
+                                                                  && (internal::is_same<Index,T>::value)
+                                                                  && Base::SizeAtCompileTime!=Dynamic
+                                                                  && Base::SizeAtCompileTime!=1
+                                                                  && internal::is_convertible<T, Scalar>::value
+                                                                  && internal::is_same<typename internal::traits<Derived>::XprKind,ArrayXpr>::value,T*>::type* = 0)
+    {
+      Base::setConstant(val0);
+    }
+    
     template<typename MatrixTypeA, typename MatrixTypeB, bool SwapPointers>
     friend struct internal::matrix_swap_impl;
 
-    /** \internal generic implementation of swap for dense storage since for dynamic-sized matrices of same type it is enough to swap the
-      * data pointers.
+  public:
+    
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+    /** \internal
+      * \brief Override DenseBase::swap() since for dynamic-sized matrices
+      * of same type it is enough to swap the data pointers.
       */
     template<typename OtherDerived>
-    void _swap(DenseBase<OtherDerived> const & other)
+    EIGEN_DEVICE_FUNC
+    void swap(DenseBase<OtherDerived> & other)
     {
       enum { SwapPointers = internal::is_same<Derived, OtherDerived>::value && Base::SizeAtCompileTime==Dynamic };
-      internal::matrix_swap_impl<Derived, OtherDerived, bool(SwapPointers)>::run(this->derived(), other.const_cast_derived());
+      internal::matrix_swap_impl<Derived, OtherDerived, bool(SwapPointers)>::run(this->derived(), other.derived());
     }
-
-  public:
-#ifndef EIGEN_PARSED_BY_DOXYGEN
+    
+    /** \internal
+      * \brief const version forwarded to DenseBase::swap
+      */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    void swap(DenseBase<OtherDerived> const & other)
+    { Base::swap(other.derived()); }
+    
+    EIGEN_DEVICE_FUNC 
     static EIGEN_STRONG_INLINE void _check_template_params()
     {
       EIGEN_STATIC_ASSERT((EIGEN_IMPLIES(MaxRowsAtCompileTime==1 && MaxColsAtCompileTime!=1, (Options&RowMajor)==RowMajor)
@@ -662,16 +865,16 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
                         && (Options & (DontAlign|RowMajor)) == Options),
         INVALID_MATRIX_TEMPLATE_PARAMETERS)
     }
-#endif
 
-private:
-    enum { ThisConstantIsPrivateInPlainObjectBase };
+    enum { IsPlainObjectBase = 1 };
+#endif
 };
 
+namespace internal {
+
 template <typename Derived, typename OtherDerived, bool IsVector>
-struct internal::conservative_resize_like_impl
+struct conservative_resize_like_impl
 {
-  typedef typename Derived::Index Index;
   static void run(DenseBase<Derived>& _this, Index rows, Index cols)
   {
     if (_this.rows() == rows && _this.cols() == cols) return;
@@ -729,12 +932,14 @@ struct internal::conservative_resize_like_impl
   }
 };
 
-namespace internal {
-
+// Here, the specialization for vectors inherits from the general matrix case
+// to allow calling .conservativeResize(rows,cols) on vectors.
 template <typename Derived, typename OtherDerived>
 struct conservative_resize_like_impl<Derived,OtherDerived,true>
+  : conservative_resize_like_impl<Derived,OtherDerived,false>
 {
-  typedef typename Derived::Index Index;
+  using conservative_resize_like_impl<Derived,OtherDerived,false>::run;
+  
   static void run(DenseBase<Derived>& _this, Index size)
   {
     const Index new_rows = Derived::RowsAtCompileTime==1 ? 1 : size;
@@ -760,6 +965,7 @@ struct conservative_resize_like_impl<Derived,OtherDerived,true>
 template<typename MatrixTypeA, typename MatrixTypeB, bool SwapPointers>
 struct matrix_swap_impl
 {
+  EIGEN_DEVICE_FUNC
   static inline void run(MatrixTypeA& a, MatrixTypeB& b)
   {
     a.base().swap(b);
@@ -769,6 +975,7 @@ struct matrix_swap_impl
 template<typename MatrixTypeA, typename MatrixTypeB>
 struct matrix_swap_impl<MatrixTypeA, MatrixTypeB, true>
 {
+  EIGEN_DEVICE_FUNC
   static inline void run(MatrixTypeA& a, MatrixTypeB& b)
   {
     static_cast<typename MatrixTypeA::Base&>(a).m_storage.swap(static_cast<typename MatrixTypeB::Base&>(b).m_storage);

Eigen/src/Core/Product.h

@@ -12,8 +12,7 @@
 
 namespace Eigen {
 
-template<typename Lhs, typename Rhs> class Product;
-template<typename Lhs, typename Rhs, typename StorageKind> class ProductImpl;
+template<typename Lhs, typename Rhs, int Option, typename StorageKind> class ProductImpl;
 
 /** \class Product
   * \ingroup Core_Module
@@ -24,53 +23,121 @@ template<typename Lhs, typename Rhs, typename StorageKind> class ProductImpl;
   * \param Rhs the type of the right-hand side expression
   *
   * This class represents an expression of the product of two arbitrary matrices.
+  * 
+  * The other template parameters are:
+  * \tparam Option     can be DefaultProduct, AliasFreeProduct, or LazyProduct
   *
   */
 
-// Use ProductReturnType to get correct traits, in particular vectorization flags
+
 namespace internal {
-template<typename Lhs, typename Rhs>
-struct traits<Product<Lhs, Rhs> >
-  : traits<typename ProductReturnType<Lhs, Rhs>::Type>
-{ 
-  // We want A+B*C to be of type Product<Matrix, Sum> and not Product<Matrix, Matrix>
-  // TODO: This flag should eventually go in a separate evaluator traits class
+
+// Determine the scalar of Product<Lhs, Rhs>. This is normally the same as Lhs::Scalar times
+// Rhs::Scalar, but product with permutation matrices inherit the scalar of the other factor.
+template<typename Lhs, typename Rhs, typename LhsShape = typename evaluator_traits<Lhs>::Shape, 
+         typename RhsShape = typename evaluator_traits<Rhs>::Shape >
+struct product_result_scalar
+{
+  typedef typename scalar_product_traits<typename Lhs::Scalar, typename Rhs::Scalar>::ReturnType Scalar;
+};
+
+template<typename Lhs, typename Rhs, typename RhsShape>
+struct product_result_scalar<Lhs, Rhs, PermutationShape, RhsShape>
+{
+  typedef typename Rhs::Scalar Scalar;
+};
+
+template<typename Lhs, typename Rhs, typename LhsShape>
+  struct product_result_scalar<Lhs, Rhs, LhsShape, PermutationShape>
+{
+  typedef typename Lhs::Scalar Scalar;
+};
+
+template<typename Lhs, typename Rhs, typename RhsShape>
+struct product_result_scalar<Lhs, Rhs, TranspositionsShape, RhsShape>
+{
+  typedef typename Rhs::Scalar Scalar;
+};
+
+template<typename Lhs, typename Rhs, typename LhsShape>
+  struct product_result_scalar<Lhs, Rhs, LhsShape, TranspositionsShape>
+{
+  typedef typename Lhs::Scalar Scalar;
+};
+
+template<typename Lhs, typename Rhs, int Option>
+struct traits<Product<Lhs, Rhs, Option> >
+{
+  typedef typename remove_all<Lhs>::type LhsCleaned;
+  typedef typename remove_all<Rhs>::type RhsCleaned;
+  typedef traits<LhsCleaned> LhsTraits;
+  typedef traits<RhsCleaned> RhsTraits;
+  
+  typedef MatrixXpr XprKind;
+  
+  typedef typename product_result_scalar<LhsCleaned,RhsCleaned>::Scalar Scalar;
+  typedef typename product_promote_storage_type<typename LhsTraits::StorageKind,
+                                                typename RhsTraits::StorageKind,
+                                                internal::product_type<Lhs,Rhs>::ret>::ret StorageKind;
+  typedef typename promote_index_type<typename LhsTraits::StorageIndex,
+                                      typename RhsTraits::StorageIndex>::type StorageIndex;
+  
   enum {
-    Flags = traits<typename ProductReturnType<Lhs, Rhs>::Type>::Flags & ~EvalBeforeNestingBit
+    RowsAtCompileTime    = LhsTraits::RowsAtCompileTime,
+    ColsAtCompileTime    = RhsTraits::ColsAtCompileTime,
+    MaxRowsAtCompileTime = LhsTraits::MaxRowsAtCompileTime,
+    MaxColsAtCompileTime = RhsTraits::MaxColsAtCompileTime,
+    
+    // FIXME: only needed by GeneralMatrixMatrixTriangular
+    InnerSize = EIGEN_SIZE_MIN_PREFER_FIXED(LhsTraits::ColsAtCompileTime, RhsTraits::RowsAtCompileTime),
+    
+    // The storage order is somewhat arbitrary here. The correct one will be determined through the evaluator.
+    Flags = (MaxRowsAtCompileTime==1 && MaxColsAtCompileTime!=1) ? RowMajorBit
+          : (MaxColsAtCompileTime==1 && MaxRowsAtCompileTime!=1) ? 0
+          : (   ((LhsTraits::Flags&NoPreferredStorageOrderBit) && (RhsTraits::Flags&RowMajorBit))
+             || ((RhsTraits::Flags&NoPreferredStorageOrderBit) && (LhsTraits::Flags&RowMajorBit)) ) ? RowMajorBit
+          : NoPreferredStorageOrderBit
   };
 };
+
 } // end namespace internal
 
 
-template<typename Lhs, typename Rhs>
-class Product : public ProductImpl<Lhs,Rhs,typename internal::promote_storage_type<typename internal::traits<Lhs>::StorageKind,
-                                                                            typename internal::traits<Rhs>::StorageKind>::ret>
+template<typename _Lhs, typename _Rhs, int Option>
+class Product : public ProductImpl<_Lhs,_Rhs,Option,
+                                   typename internal::product_promote_storage_type<typename internal::traits<_Lhs>::StorageKind,
+                                                                                   typename internal::traits<_Rhs>::StorageKind,
+                                                                                   internal::product_type<_Lhs,_Rhs>::ret>::ret>
 {
   public:
     
+    typedef _Lhs Lhs;
+    typedef _Rhs Rhs;
+    
     typedef typename ProductImpl<
-        Lhs, Rhs,
-        typename internal::promote_storage_type<typename Lhs::StorageKind,
-                                                typename Rhs::StorageKind>::ret>::Base Base;
+        Lhs, Rhs, Option,
+        typename internal::product_promote_storage_type<typename internal::traits<Lhs>::StorageKind,
+                                                        typename internal::traits<Rhs>::StorageKind,
+                                                        internal::product_type<Lhs,Rhs>::ret>::ret>::Base Base;
     EIGEN_GENERIC_PUBLIC_INTERFACE(Product)
 
-    typedef typename Lhs::Nested LhsNested;
-    typedef typename Rhs::Nested RhsNested;
+    typedef typename internal::ref_selector<Lhs>::type LhsNested;
+    typedef typename internal::ref_selector<Rhs>::type RhsNested;
     typedef typename internal::remove_all<LhsNested>::type LhsNestedCleaned;
     typedef typename internal::remove_all<RhsNested>::type RhsNestedCleaned;
 
-    Product(const Lhs& lhs, const Rhs& rhs) : m_lhs(lhs), m_rhs(rhs)
+    EIGEN_DEVICE_FUNC Product(const Lhs& lhs, const Rhs& rhs) : m_lhs(lhs), m_rhs(rhs)
     {
       eigen_assert(lhs.cols() == rhs.rows()
         && "invalid matrix product"
         && "if you wanted a coeff-wise or a dot product use the respective explicit functions");
     }
 
-    inline Index rows() const { return m_lhs.rows(); }
-    inline Index cols() const { return m_rhs.cols(); }
+    EIGEN_DEVICE_FUNC inline Index rows() const { return m_lhs.rows(); }
+    EIGEN_DEVICE_FUNC inline Index cols() const { return m_rhs.cols(); }
 
-    const LhsNestedCleaned& lhs() const { return m_lhs; }
-    const RhsNestedCleaned& rhs() const { return m_rhs; }
+    EIGEN_DEVICE_FUNC const LhsNestedCleaned& lhs() const { return m_lhs; }
+    EIGEN_DEVICE_FUNC const RhsNestedCleaned& rhs() const { return m_rhs; }
 
   protected:
 
@@ -78,14 +145,76 @@ class Product : public ProductImpl<Lhs,Rhs,typename internal::promote_storage_ty
     RhsNested m_rhs;
 };
 
-template<typename Lhs, typename Rhs>
-class ProductImpl<Lhs,Rhs,Dense> : public internal::dense_xpr_base<Product<Lhs,Rhs> >::type
-{
-    typedef Product<Lhs, Rhs> Derived;
-  public:
+namespace internal {
+  
+template<typename Lhs, typename Rhs, int Option, int ProductTag = internal::product_type<Lhs,Rhs>::ret>
+class dense_product_base
+ : public internal::dense_xpr_base<Product<Lhs,Rhs,Option> >::type
+{};
 
-    typedef typename internal::dense_xpr_base<Product<Lhs, Rhs> >::type Base;
+/** Convertion to scalar for inner-products */
+template<typename Lhs, typename Rhs, int Option>
+class dense_product_base<Lhs, Rhs, Option, InnerProduct>
+ : public internal::dense_xpr_base<Product<Lhs,Rhs,Option> >::type
+{
+  typedef Product<Lhs,Rhs,Option> ProductXpr;
+  typedef typename internal::dense_xpr_base<ProductXpr>::type Base;
+public:
+  using Base::derived;
+  typedef typename Base::Scalar Scalar;
+  
+  operator const Scalar() const
+  {
+    return internal::evaluator<ProductXpr>(derived()).coeff(0,0);
+  }
+};
+
+} // namespace internal
+
+// Generic API dispatcher
+template<typename Lhs, typename Rhs, int Option, typename StorageKind>
+class ProductImpl : public internal::generic_xpr_base<Product<Lhs,Rhs,Option>, MatrixXpr, StorageKind>::type
+{
+  public:
+    typedef typename internal::generic_xpr_base<Product<Lhs,Rhs,Option>, MatrixXpr, StorageKind>::type Base;
+};
+
+template<typename Lhs, typename Rhs, int Option>
+class ProductImpl<Lhs,Rhs,Option,Dense>
+  : public internal::dense_product_base<Lhs,Rhs,Option>
+{
+    typedef Product<Lhs, Rhs, Option> Derived;
+    
+  public:
+    
+    typedef typename internal::dense_product_base<Lhs, Rhs, Option> Base;
     EIGEN_DENSE_PUBLIC_INTERFACE(Derived)
+  protected:
+    enum {
+      IsOneByOne = (RowsAtCompileTime == 1 || RowsAtCompileTime == Dynamic) && 
+                   (ColsAtCompileTime == 1 || ColsAtCompileTime == Dynamic),
+      EnableCoeff = IsOneByOne || Option==LazyProduct
+    };
+    
+  public:
+  
+    EIGEN_DEVICE_FUNC Scalar coeff(Index row, Index col) const
+    {
+      EIGEN_STATIC_ASSERT(EnableCoeff, THIS_METHOD_IS_ONLY_FOR_INNER_OR_LAZY_PRODUCTS);
+      eigen_assert( (Option==LazyProduct) || (this->rows() == 1 && this->cols() == 1) );
+      
+      return internal::evaluator<Derived>(derived()).coeff(row,col);
+    }
+
+    EIGEN_DEVICE_FUNC Scalar coeff(Index i) const
+    {
+      EIGEN_STATIC_ASSERT(EnableCoeff, THIS_METHOD_IS_ONLY_FOR_INNER_OR_LAZY_PRODUCTS);
+      eigen_assert( (Option==LazyProduct) || (this->rows() == 1 && this->cols() == 1) );
+      
+      return internal::evaluator<Derived>(derived()).coeff(i);
+    }
+    
+  
 };
 
 /***************************************************************************
@@ -102,6 +231,15 @@ prod(const Lhs& lhs, const Rhs& rhs)
   return Product<Lhs,Rhs>(lhs,rhs);
 }
 
+/** \internal used to test the evaluator only
+  */
+template<typename Lhs,typename Rhs>
+const Product<Lhs,Rhs,LazyProduct>
+lazyprod(const Lhs& lhs, const Rhs& rhs)
+{
+  return Product<Lhs,Rhs,LazyProduct>(lhs,rhs);
+}
+
 } // end namespace Eigen
 
 #endif // EIGEN_PRODUCT_H

Eigen/src/Core/ProductBase.h

@@ -1,278 +0,0 @@
-// This file is part of Eigen, a lightweight C++ template library
-// for linear algebra.
-//
-// Copyright (C) 2009-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
-//
-// This Source Code Form is subject to the terms of the Mozilla
-// Public License v. 2.0. If a copy of the MPL was not distributed
-// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
-
-#ifndef EIGEN_PRODUCTBASE_H
-#define EIGEN_PRODUCTBASE_H
-
-namespace Eigen { 
-
-/** \class ProductBase
-  * \ingroup Core_Module
-  *
-  */
-
-namespace internal {
-template<typename Derived, typename _Lhs, typename _Rhs>
-struct traits<ProductBase<Derived,_Lhs,_Rhs> >
-{
-  typedef MatrixXpr XprKind;
-  typedef typename remove_all<_Lhs>::type Lhs;
-  typedef typename remove_all<_Rhs>::type Rhs;
-  typedef typename scalar_product_traits<typename Lhs::Scalar, typename Rhs::Scalar>::ReturnType Scalar;
-  typedef typename promote_storage_type<typename traits<Lhs>::StorageKind,
-                                           typename traits<Rhs>::StorageKind>::ret StorageKind;
-  typedef typename promote_index_type<typename traits<Lhs>::Index,
-                                         typename traits<Rhs>::Index>::type Index;
-  enum {
-    RowsAtCompileTime = traits<Lhs>::RowsAtCompileTime,
-    ColsAtCompileTime = traits<Rhs>::ColsAtCompileTime,
-    MaxRowsAtCompileTime = traits<Lhs>::MaxRowsAtCompileTime,
-    MaxColsAtCompileTime = traits<Rhs>::MaxColsAtCompileTime,
-    Flags = (MaxRowsAtCompileTime==1 ? RowMajorBit : 0)
-          | EvalBeforeNestingBit | EvalBeforeAssigningBit | NestByRefBit,
-                  // Note that EvalBeforeNestingBit and NestByRefBit
-                  // are not used in practice because nested is overloaded for products
-    CoeffReadCost = 0 // FIXME why is it needed ?
-  };
-};
-}
-
-#define EIGEN_PRODUCT_PUBLIC_INTERFACE(Derived) \
-  typedef ProductBase<Derived, Lhs, Rhs > Base; \
-  EIGEN_DENSE_PUBLIC_INTERFACE(Derived) \
-  typedef typename Base::LhsNested LhsNested; \
-  typedef typename Base::_LhsNested _LhsNested; \
-  typedef typename Base::LhsBlasTraits LhsBlasTraits; \
-  typedef typename Base::ActualLhsType ActualLhsType; \
-  typedef typename Base::_ActualLhsType _ActualLhsType; \
-  typedef typename Base::RhsNested RhsNested; \
-  typedef typename Base::_RhsNested _RhsNested; \
-  typedef typename Base::RhsBlasTraits RhsBlasTraits; \
-  typedef typename Base::ActualRhsType ActualRhsType; \
-  typedef typename Base::_ActualRhsType _ActualRhsType; \
-  using Base::m_lhs; \
-  using Base::m_rhs;
-
-template<typename Derived, typename Lhs, typename Rhs>
-class ProductBase : public MatrixBase<Derived>
-{
-  public:
-    typedef MatrixBase<Derived> Base;
-    EIGEN_DENSE_PUBLIC_INTERFACE(ProductBase)
-    
-    typedef typename Lhs::Nested LhsNested;
-    typedef typename internal::remove_all<LhsNested>::type _LhsNested;
-    typedef internal::blas_traits<_LhsNested> LhsBlasTraits;
-    typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType;
-    typedef typename internal::remove_all<ActualLhsType>::type _ActualLhsType;
-    typedef typename internal::traits<Lhs>::Scalar LhsScalar;
-
-    typedef typename Rhs::Nested RhsNested;
-    typedef typename internal::remove_all<RhsNested>::type _RhsNested;
-    typedef internal::blas_traits<_RhsNested> RhsBlasTraits;
-    typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType;
-    typedef typename internal::remove_all<ActualRhsType>::type _ActualRhsType;
-    typedef typename internal::traits<Rhs>::Scalar RhsScalar;
-
-    // Diagonal of a product: no need to evaluate the arguments because they are going to be evaluated only once
-    typedef CoeffBasedProduct<LhsNested, RhsNested, 0> FullyLazyCoeffBaseProductType;
-
-  public:
-
-    typedef typename Base::PlainObject PlainObject;
-
-    ProductBase(const Lhs& a_lhs, const Rhs& a_rhs)
-      : m_lhs(a_lhs), m_rhs(a_rhs)
-    {
-      eigen_assert(a_lhs.cols() == a_rhs.rows()
-        && "invalid matrix product"
-        && "if you wanted a coeff-wise or a dot product use the respective explicit functions");
-    }
-
-    inline Index rows() const { return m_lhs.rows(); }
-    inline Index cols() const { return m_rhs.cols(); }
-
-    template<typename Dest>
-    inline void evalTo(Dest& dst) const { dst.setZero(); scaleAndAddTo(dst,Scalar(1)); }
-
-    template<typename Dest>
-    inline void addTo(Dest& dst) const { scaleAndAddTo(dst,Scalar(1)); }
-
-    template<typename Dest>
-    inline void subTo(Dest& dst) const { scaleAndAddTo(dst,Scalar(-1)); }
-
-    template<typename Dest>
-    inline void scaleAndAddTo(Dest& dst, const Scalar& alpha) const { derived().scaleAndAddTo(dst,alpha); }
-
-    const _LhsNested& lhs() const { return m_lhs; }
-    const _RhsNested& rhs() const { return m_rhs; }
-
-    // Implicit conversion to the nested type (trigger the evaluation of the product)
-    operator const PlainObject& () const
-    {
-      m_result.resize(m_lhs.rows(), m_rhs.cols());
-      derived().evalTo(m_result);
-      return m_result;
-    }
-
-    const Diagonal<const FullyLazyCoeffBaseProductType,0> diagonal() const
-    { return FullyLazyCoeffBaseProductType(m_lhs, m_rhs); }
-
-    template<int Index>
-    const Diagonal<FullyLazyCoeffBaseProductType,Index> diagonal() const
-    { return FullyLazyCoeffBaseProductType(m_lhs, m_rhs); }
-
-    const Diagonal<FullyLazyCoeffBaseProductType,Dynamic> diagonal(Index index) const
-    { return FullyLazyCoeffBaseProductType(m_lhs, m_rhs).diagonal(index); }
-
-    // restrict coeff accessors to 1x1 expressions. No need to care about mutators here since this isnt a Lvalue expression
-    typename Base::CoeffReturnType coeff(Index row, Index col) const
-    {
-#ifdef EIGEN2_SUPPORT
-      return lhs().row(row).cwiseProduct(rhs().col(col).transpose()).sum();
-#else
-      EIGEN_STATIC_ASSERT_SIZE_1x1(Derived)
-      eigen_assert(this->rows() == 1 && this->cols() == 1);
-      Matrix<Scalar,1,1> result = *this;
-      return result.coeff(row,col);
-#endif
-    }
-
-    typename Base::CoeffReturnType coeff(Index i) const
-    {
-      EIGEN_STATIC_ASSERT_SIZE_1x1(Derived)
-      eigen_assert(this->rows() == 1 && this->cols() == 1);
-      Matrix<Scalar,1,1> result = *this;
-      return result.coeff(i);
-    }
-
-    const Scalar& coeffRef(Index row, Index col) const
-    {
-      EIGEN_STATIC_ASSERT_SIZE_1x1(Derived)
-      eigen_assert(this->rows() == 1 && this->cols() == 1);
-      return derived().coeffRef(row,col);
-    }
-
-    const Scalar& coeffRef(Index i) const
-    {
-      EIGEN_STATIC_ASSERT_SIZE_1x1(Derived)
-      eigen_assert(this->rows() == 1 && this->cols() == 1);
-      return derived().coeffRef(i);
-    }
-
-  protected:
-
-    LhsNested m_lhs;
-    RhsNested m_rhs;
-
-    mutable PlainObject m_result;
-};
-
-// here we need to overload the nested rule for products
-// such that the nested type is a const reference to a plain matrix
-namespace internal {
-template<typename Lhs, typename Rhs, int Mode, int N, typename PlainObject>
-struct nested<GeneralProduct<Lhs,Rhs,Mode>, N, PlainObject>
-{
-  typedef PlainObject const& type;
-};
-}
-
-template<typename NestedProduct>
-class ScaledProduct;
-
-// Note that these two operator* functions are not defined as member
-// functions of ProductBase, because, otherwise we would have to
-// define all overloads defined in MatrixBase. Furthermore, Using
-// "using Base::operator*" would not work with MSVC.
-//
-// Also note that here we accept any compatible scalar types
-template<typename Derived,typename Lhs,typename Rhs>
-const ScaledProduct<Derived>
-operator*(const ProductBase<Derived,Lhs,Rhs>& prod, const typename Derived::Scalar& x)
-{ return ScaledProduct<Derived>(prod.derived(), x); }
-
-template<typename Derived,typename Lhs,typename Rhs>
-typename internal::enable_if<!internal::is_same<typename Derived::Scalar,typename Derived::RealScalar>::value,
-                      const ScaledProduct<Derived> >::type
-operator*(const ProductBase<Derived,Lhs,Rhs>& prod, const typename Derived::RealScalar& x)
-{ return ScaledProduct<Derived>(prod.derived(), x); }
-
-
-template<typename Derived,typename Lhs,typename Rhs>
-const ScaledProduct<Derived>
-operator*(const typename Derived::Scalar& x,const ProductBase<Derived,Lhs,Rhs>& prod)
-{ return ScaledProduct<Derived>(prod.derived(), x); }
-
-template<typename Derived,typename Lhs,typename Rhs>
-typename internal::enable_if<!internal::is_same<typename Derived::Scalar,typename Derived::RealScalar>::value,
-                      const ScaledProduct<Derived> >::type
-operator*(const typename Derived::RealScalar& x,const ProductBase<Derived,Lhs,Rhs>& prod)
-{ return ScaledProduct<Derived>(prod.derived(), x); }
-
-namespace internal {
-template<typename NestedProduct>
-struct traits<ScaledProduct<NestedProduct> >
- : traits<ProductBase<ScaledProduct<NestedProduct>,
-                         typename NestedProduct::_LhsNested,
-                         typename NestedProduct::_RhsNested> >
-{
-  typedef typename traits<NestedProduct>::StorageKind StorageKind;
-};
-}
-
-template<typename NestedProduct>
-class ScaledProduct
-  : public ProductBase<ScaledProduct<NestedProduct>,
-                       typename NestedProduct::_LhsNested,
-                       typename NestedProduct::_RhsNested>
-{
-  public:
-    typedef ProductBase<ScaledProduct<NestedProduct>,
-                       typename NestedProduct::_LhsNested,
-                       typename NestedProduct::_RhsNested> Base;
-    typedef typename Base::Scalar Scalar;
-    typedef typename Base::PlainObject PlainObject;
-//     EIGEN_PRODUCT_PUBLIC_INTERFACE(ScaledProduct)
-
-    ScaledProduct(const NestedProduct& prod, const Scalar& x)
-    : Base(prod.lhs(),prod.rhs()), m_prod(prod), m_alpha(x) {}
-
-    template<typename Dest>
-    inline void evalTo(Dest& dst) const { dst.setZero(); scaleAndAddTo(dst, Scalar(1)); }
-
-    template<typename Dest>
-    inline void addTo(Dest& dst) const { scaleAndAddTo(dst, Scalar(1)); }
-
-    template<typename Dest>
-    inline void subTo(Dest& dst) const { scaleAndAddTo(dst, Scalar(-1)); }
-
-    template<typename Dest>
-    inline void scaleAndAddTo(Dest& dst, const Scalar& a_alpha) const { m_prod.derived().scaleAndAddTo(dst,a_alpha * m_alpha); }
-
-    const Scalar& alpha() const { return m_alpha; }
-    
-  protected:
-    const NestedProduct& m_prod;
-    Scalar m_alpha;
-};
-
-/** \internal
-  * Overloaded to perform an efficient C = (A*B).lazy() */
-template<typename Derived>
-template<typename ProductDerived, typename Lhs, typename Rhs>
-Derived& MatrixBase<Derived>::lazyAssign(const ProductBase<ProductDerived, Lhs,Rhs>& other)
-{
-  other.derived().evalTo(derived());
-  return derived();
-}
-
-} // end namespace Eigen
-
-#endif // EIGEN_PRODUCTBASE_H

Eigen/src/Core/Random.h

@@ -28,12 +28,18 @@ struct functor_traits<scalar_random_op<Scalar> >
 
 /** \returns a random matrix expression
   *
+  * Numbers are uniformly spread through their whole definition range for integer types,
+  * and in the [-1:1] range for floating point scalar types.
+  * 
   * The parameters \a rows and \a cols are the number of rows and of columns of
   * the returned matrix. Must be compatible with this MatrixBase type.
   *
+  * \not_reentrant
+  * 
   * This variant is meant to be used for dynamic-size matrix types. For fixed-size types,
   * it is redundant to pass \a rows and \a cols as arguments, so Random() should be used
   * instead.
+  * 
   *
   * Example: \include MatrixBase_random_int_int.cpp
   * Output: \verbinclude MatrixBase_random_int_int.out
@@ -41,22 +47,28 @@ struct functor_traits<scalar_random_op<Scalar> >
   * This expression has the "evaluate before nesting" flag so that it will be evaluated into
   * a temporary matrix whenever it is nested in a larger expression. This prevents unexpected
   * behavior with expressions involving random matrices.
+  * 
+  * See DenseBase::NullaryExpr(Index, const CustomNullaryOp&) for an example using C++11 random generators.
   *
-  * \sa MatrixBase::setRandom(), MatrixBase::Random(Index), MatrixBase::Random()
+  * \sa DenseBase::setRandom(), DenseBase::Random(Index), DenseBase::Random()
   */
 template<typename Derived>
-inline const CwiseNullaryOp<internal::scalar_random_op<typename internal::traits<Derived>::Scalar>, Derived>
+inline const typename DenseBase<Derived>::RandomReturnType
 DenseBase<Derived>::Random(Index rows, Index cols)
 {
   return NullaryExpr(rows, cols, internal::scalar_random_op<Scalar>());
 }
 
 /** \returns a random vector expression
+  *
+  * Numbers are uniformly spread through their whole definition range for integer types,
+  * and in the [-1:1] range for floating point scalar types.
   *
   * The parameter \a size is the size of the returned vector.
   * Must be compatible with this MatrixBase type.
   *
   * \only_for_vectors
+  * \not_reentrant
   *
   * This variant is meant to be used for dynamic-size vector types. For fixed-size types,
   * it is redundant to pass \a size as argument, so Random() should be used
@@ -69,10 +81,10 @@ DenseBase<Derived>::Random(Index rows, Index cols)
   * a temporary vector whenever it is nested in a larger expression. This prevents unexpected
   * behavior with expressions involving random matrices.
   *
-  * \sa MatrixBase::setRandom(), MatrixBase::Random(Index,Index), MatrixBase::Random()
+  * \sa DenseBase::setRandom(), DenseBase::Random(Index,Index), DenseBase::Random()
   */
 template<typename Derived>
-inline const CwiseNullaryOp<internal::scalar_random_op<typename internal::traits<Derived>::Scalar>, Derived>
+inline const typename DenseBase<Derived>::RandomReturnType
 DenseBase<Derived>::Random(Index size)
 {
   return NullaryExpr(size, internal::scalar_random_op<Scalar>());
@@ -80,6 +92,9 @@ DenseBase<Derived>::Random(Index size)
 
 /** \returns a fixed-size random matrix or vector expression
   *
+  * Numbers are uniformly spread through their whole definition range for integer types,
+  * and in the [-1:1] range for floating point scalar types.
+  * 
   * This variant is only for fixed-size MatrixBase types. For dynamic-size types, you
   * need to use the variants taking size arguments.
   *
@@ -89,11 +104,13 @@ DenseBase<Derived>::Random(Index size)
   * This expression has the "evaluate before nesting" flag so that it will be evaluated into
   * a temporary matrix whenever it is nested in a larger expression. This prevents unexpected
   * behavior with expressions involving random matrices.
+  * 
+  * \not_reentrant
   *
-  * \sa MatrixBase::setRandom(), MatrixBase::Random(Index,Index), MatrixBase::Random(Index)
+  * \sa DenseBase::setRandom(), DenseBase::Random(Index,Index), DenseBase::Random(Index)
   */
 template<typename Derived>
-inline const CwiseNullaryOp<internal::scalar_random_op<typename internal::traits<Derived>::Scalar>, Derived>
+inline const typename DenseBase<Derived>::RandomReturnType
 DenseBase<Derived>::Random()
 {
   return NullaryExpr(RowsAtCompileTime, ColsAtCompileTime, internal::scalar_random_op<Scalar>());
@@ -101,6 +118,11 @@ DenseBase<Derived>::Random()
 
 /** Sets all coefficients in this expression to random values.
   *
+  * Numbers are uniformly spread through their whole definition range for integer types,
+  * and in the [-1:1] range for floating point scalar types.
+  * 
+  * \not_reentrant
+  * 
   * Example: \include MatrixBase_setRandom.cpp
   * Output: \verbinclude MatrixBase_setRandom.out
   *
@@ -114,12 +136,16 @@ inline Derived& DenseBase<Derived>::setRandom()
 
 /** Resizes to the given \a newSize, and sets all coefficients in this expression to random values.
   *
+  * Numbers are uniformly spread through their whole definition range for integer types,
+  * and in the [-1:1] range for floating point scalar types.
+  * 
   * \only_for_vectors
+  * \not_reentrant
   *
   * Example: \include Matrix_setRandom_int.cpp
   * Output: \verbinclude Matrix_setRandom_int.out
   *
-  * \sa MatrixBase::setRandom(), setRandom(Index,Index), class CwiseNullaryOp, MatrixBase::Random()
+  * \sa DenseBase::setRandom(), setRandom(Index,Index), class CwiseNullaryOp, DenseBase::Random()
   */
 template<typename Derived>
 EIGEN_STRONG_INLINE Derived&
@@ -131,19 +157,24 @@ PlainObjectBase<Derived>::setRandom(Index newSize)
 
 /** Resizes to the given size, and sets all coefficients in this expression to random values.
   *
-  * \param nbRows the new number of rows
-  * \param nbCols the new number of columns
+  * Numbers are uniformly spread through their whole definition range for integer types,
+  * and in the [-1:1] range for floating point scalar types.
+  *
+  * \not_reentrant
+  * 
+  * \param rows the new number of rows
+  * \param cols the new number of columns
   *
   * Example: \include Matrix_setRandom_int_int.cpp
   * Output: \verbinclude Matrix_setRandom_int_int.out
   *
-  * \sa MatrixBase::setRandom(), setRandom(Index), class CwiseNullaryOp, MatrixBase::Random()
+  * \sa DenseBase::setRandom(), setRandom(Index), class CwiseNullaryOp, DenseBase::Random()
   */
 template<typename Derived>
 EIGEN_STRONG_INLINE Derived&
-PlainObjectBase<Derived>::setRandom(Index nbRows, Index nbCols)
+PlainObjectBase<Derived>::setRandom(Index rows, Index cols)
 {
-  resize(nbRows, nbCols);
+  resize(rows, cols);
   return setRandom();
 }
 

Eigen/src/Core/Redux.h

@@ -65,6 +65,25 @@ public:
               ? CompleteUnrolling
               : NoUnrolling
   };
+  
+#ifdef EIGEN_DEBUG_ASSIGN
+  static void debug()
+  {
+    std::cerr << "Xpr: " << typeid(typename Derived::XprType).name() << std::endl;
+    std::cerr.setf(std::ios::hex, std::ios::basefield);
+    EIGEN_DEBUG_VAR(Derived::Flags)
+    std::cerr.unsetf(std::ios::hex);
+    EIGEN_DEBUG_VAR(InnerMaxSize)
+    EIGEN_DEBUG_VAR(PacketSize)
+    EIGEN_DEBUG_VAR(MightVectorize)
+    EIGEN_DEBUG_VAR(MayLinearVectorize)
+    EIGEN_DEBUG_VAR(MaySliceVectorize)
+    EIGEN_DEBUG_VAR(Traversal)
+    EIGEN_DEBUG_VAR(UnrollingLimit)
+    EIGEN_DEBUG_VAR(Unrolling)
+    std::cerr << std::endl;
+  }
+#endif
 };
 
 /***************************************************************************
@@ -82,6 +101,7 @@ struct redux_novec_unroller
 
   typedef typename Derived::Scalar Scalar;
 
+  EIGEN_DEVICE_FUNC
   static EIGEN_STRONG_INLINE Scalar run(const Derived &mat, const Func& func)
   {
     return func(redux_novec_unroller<Func, Derived, Start, HalfLength>::run(mat,func),
@@ -99,6 +119,7 @@ struct redux_novec_unroller<Func, Derived, Start, 1>
 
   typedef typename Derived::Scalar Scalar;
 
+  EIGEN_DEVICE_FUNC
   static EIGEN_STRONG_INLINE Scalar run(const Derived &mat, const Func&)
   {
     return mat.coeffByOuterInner(outer, inner);
@@ -112,6 +133,7 @@ template<typename Func, typename Derived, int Start>
 struct redux_novec_unroller<Func, Derived, Start, 0>
 {
   typedef typename Derived::Scalar Scalar;
+  EIGEN_DEVICE_FUNC 
   static EIGEN_STRONG_INLINE Scalar run(const Derived&, const Func&) { return Scalar(); }
 };
 
@@ -143,7 +165,7 @@ struct redux_vec_unroller<Func, Derived, Start, 1>
     index = Start * packet_traits<typename Derived::Scalar>::size,
     outer = index / int(Derived::InnerSizeAtCompileTime),
     inner = index % int(Derived::InnerSizeAtCompileTime),
-    alignment = (Derived::Flags & AlignedBit) ? Aligned : Unaligned
+    alignment = Derived::Alignment
   };
 
   typedef typename Derived::Scalar Scalar;
@@ -151,7 +173,7 @@ struct redux_vec_unroller<Func, Derived, Start, 1>
 
   static EIGEN_STRONG_INLINE PacketScalar run(const Derived &mat, const Func&)
   {
-    return mat.template packetByOuterInner<alignment>(outer, inner);
+    return mat.template packetByOuterInner<alignment,PacketScalar>(outer, inner);
   }
 };
 
@@ -169,8 +191,8 @@ template<typename Func, typename Derived>
 struct redux_impl<Func, Derived, DefaultTraversal, NoUnrolling>
 {
   typedef typename Derived::Scalar Scalar;
-  typedef typename Derived::Index Index;
-  static EIGEN_STRONG_INLINE Scalar run(const Derived& mat, const Func& func)
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE Scalar run(const Derived &mat, const Func& func)
   {
     eigen_assert(mat.rows()>0 && mat.cols()>0 && "you are using an empty matrix");
     Scalar res;
@@ -194,18 +216,18 @@ struct redux_impl<Func, Derived, LinearVectorizedTraversal, NoUnrolling>
 {
   typedef typename Derived::Scalar Scalar;
   typedef typename packet_traits<Scalar>::type PacketScalar;
-  typedef typename Derived::Index Index;
 
-  static Scalar run(const Derived& mat, const Func& func)
+  static Scalar run(const Derived &mat, const Func& func)
   {
     const Index size = mat.size();
-    eigen_assert(size && "you are using an empty matrix");
+    
     const Index packetSize = packet_traits<Scalar>::size;
-    const Index alignedStart = internal::first_aligned(mat);
+    const int packetAlignment = unpacket_traits<PacketScalar>::alignment;
     enum {
-      alignment = bool(Derived::Flags & DirectAccessBit) || bool(Derived::Flags & AlignedBit)
-                ? Aligned : Unaligned
+      alignment0 = (bool(Derived::Flags & DirectAccessBit) && bool(packet_traits<Scalar>::AlignedOnScalar)) ? int(packetAlignment) : int(Unaligned),
+      alignment = EIGEN_PLAIN_ENUM_MAX(alignment0, Derived::Alignment)
     };
+    const Index alignedStart = internal::first_default_aligned(mat.nestedExpression());
     const Index alignedSize2 = ((size-alignedStart)/(2*packetSize))*(2*packetSize);
     const Index alignedSize = ((size-alignedStart)/(packetSize))*(packetSize);
     const Index alignedEnd2 = alignedStart + alignedSize2;
@@ -213,19 +235,19 @@ struct redux_impl<Func, Derived, LinearVectorizedTraversal, NoUnrolling>
     Scalar res;
     if(alignedSize)
     {
-      PacketScalar packet_res0 = mat.template packet<alignment>(alignedStart);
+      PacketScalar packet_res0 = mat.template packet<alignment,PacketScalar>(alignedStart);
       if(alignedSize>packetSize) // we have at least two packets to partly unroll the loop
       {
-        PacketScalar packet_res1 = mat.template packet<alignment>(alignedStart+packetSize);
+        PacketScalar packet_res1 = mat.template packet<alignment,PacketScalar>(alignedStart+packetSize);
         for(Index index = alignedStart + 2*packetSize; index < alignedEnd2; index += 2*packetSize)
         {
-          packet_res0 = func.packetOp(packet_res0, mat.template packet<alignment>(index));
-          packet_res1 = func.packetOp(packet_res1, mat.template packet<alignment>(index+packetSize));
+          packet_res0 = func.packetOp(packet_res0, mat.template packet<alignment,PacketScalar>(index));
+          packet_res1 = func.packetOp(packet_res1, mat.template packet<alignment,PacketScalar>(index+packetSize));
         }
 
         packet_res0 = func.packetOp(packet_res0,packet_res1);
         if(alignedEnd>alignedEnd2)
-          packet_res0 = func.packetOp(packet_res0, mat.template packet<alignment>(alignedEnd2));
+          packet_res0 = func.packetOp(packet_res0, mat.template packet<alignment,PacketScalar>(alignedEnd2));
       }
       res = func.predux(packet_res0);
 
@@ -251,10 +273,9 @@ template<typename Func, typename Derived>
 struct redux_impl<Func, Derived, SliceVectorizedTraversal, NoUnrolling>
 {
   typedef typename Derived::Scalar Scalar;
-  typedef typename packet_traits<Scalar>::type PacketScalar;
-  typedef typename Derived::Index Index;
+  typedef typename packet_traits<Scalar>::type PacketType;
 
-  static Scalar run(const Derived& mat, const Func& func)
+  EIGEN_DEVICE_FUNC static Scalar run(const Derived &mat, const Func& func)
   {
     eigen_assert(mat.rows()>0 && mat.cols()>0 && "you are using an empty matrix");
     const Index innerSize = mat.innerSize();
@@ -266,10 +287,10 @@ struct redux_impl<Func, Derived, SliceVectorizedTraversal, NoUnrolling>
     Scalar res;
     if(packetedInnerSize)
     {
-      PacketScalar packet_res = mat.template packet<Unaligned>(0,0);
+      PacketType packet_res = mat.template packet<Unaligned,PacketType>(0,0);
       for(Index j=0; j<outerSize; ++j)
         for(Index i=(j==0?packetSize:0); i<packetedInnerSize; i+=Index(packetSize))
-          packet_res = func.packetOp(packet_res, mat.template packetByOuterInner<Unaligned>(j,i));
+          packet_res = func.packetOp(packet_res, mat.template packetByOuterInner<Unaligned,PacketType>(j,i));
 
       res = func.predux(packet_res);
       for(Index j=0; j<outerSize; ++j)
@@ -296,16 +317,83 @@ struct redux_impl<Func, Derived, LinearVectorizedTraversal, CompleteUnrolling>
     Size = Derived::SizeAtCompileTime,
     VectorizedSize = (Size / PacketSize) * PacketSize
   };
-  static EIGEN_STRONG_INLINE Scalar run(const Derived& mat, const Func& func)
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Scalar run(const Derived &mat, const Func& func)
   {
     eigen_assert(mat.rows()>0 && mat.cols()>0 && "you are using an empty matrix");
-    Scalar res = func.predux(redux_vec_unroller<Func, Derived, 0, Size / PacketSize>::run(mat,func));
-    if (VectorizedSize != Size)
-      res = func(res,redux_novec_unroller<Func, Derived, VectorizedSize, Size-VectorizedSize>::run(mat,func));
-    return res;
+    if (VectorizedSize > 0) {
+      Scalar res = func.predux(redux_vec_unroller<Func, Derived, 0, Size / PacketSize>::run(mat,func));
+      if (VectorizedSize != Size)
+        res = func(res,redux_novec_unroller<Func, Derived, VectorizedSize, Size-VectorizedSize>::run(mat,func));
+      return res;
+    }
+    else {
+      return redux_novec_unroller<Func, Derived, 0, Size>::run(mat,func);
+    }
   }
 };
 
+// evaluator adaptor
+template<typename _XprType>
+class redux_evaluator
+{
+public:
+  typedef _XprType XprType;
+  EIGEN_DEVICE_FUNC explicit redux_evaluator(const XprType &xpr) : m_evaluator(xpr), m_xpr(xpr) {}
+  
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename XprType::PacketScalar PacketScalar;
+  typedef typename XprType::PacketReturnType PacketReturnType;
+  
+  enum {
+    MaxRowsAtCompileTime = XprType::MaxRowsAtCompileTime,
+    MaxColsAtCompileTime = XprType::MaxColsAtCompileTime,
+    // TODO we should not remove DirectAccessBit and rather find an elegant way to query the alignment offset at runtime from the evaluator
+    Flags = evaluator<XprType>::Flags & ~DirectAccessBit,
+    IsRowMajor = XprType::IsRowMajor,
+    SizeAtCompileTime = XprType::SizeAtCompileTime,
+    InnerSizeAtCompileTime = XprType::InnerSizeAtCompileTime,
+    CoeffReadCost = evaluator<XprType>::CoeffReadCost,
+    Alignment = evaluator<XprType>::Alignment
+  };
+  
+  EIGEN_DEVICE_FUNC Index rows() const { return m_xpr.rows(); }
+  EIGEN_DEVICE_FUNC Index cols() const { return m_xpr.cols(); }
+  EIGEN_DEVICE_FUNC Index size() const { return m_xpr.size(); }
+  EIGEN_DEVICE_FUNC Index innerSize() const { return m_xpr.innerSize(); }
+  EIGEN_DEVICE_FUNC Index outerSize() const { return m_xpr.outerSize(); }
+
+  EIGEN_DEVICE_FUNC
+  CoeffReturnType coeff(Index row, Index col) const
+  { return m_evaluator.coeff(row, col); }
+
+  EIGEN_DEVICE_FUNC
+  CoeffReturnType coeff(Index index) const
+  { return m_evaluator.coeff(index); }
+
+  template<int LoadMode, typename PacketType>
+  PacketReturnType packet(Index row, Index col) const
+  { return m_evaluator.template packet<LoadMode,PacketType>(row, col); }
+
+  template<int LoadMode, typename PacketType>
+  PacketReturnType packet(Index index) const
+  { return m_evaluator.template packet<LoadMode,PacketType>(index); }
+  
+  EIGEN_DEVICE_FUNC
+  CoeffReturnType coeffByOuterInner(Index outer, Index inner) const
+  { return m_evaluator.coeff(IsRowMajor ? outer : inner, IsRowMajor ? inner : outer); }
+  
+  template<int LoadMode, typename PacketType>
+  PacketReturnType packetByOuterInner(Index outer, Index inner) const
+  { return m_evaluator.template packet<LoadMode,PacketType>(IsRowMajor ? outer : inner, IsRowMajor ? inner : outer); }
+  
+  const XprType & nestedExpression() const { return m_xpr; }
+  
+protected:
+  internal::evaluator<XprType> m_evaluator;
+  const XprType &m_xpr;
+};
+
 } // end namespace internal
 
 /***************************************************************************
@@ -316,18 +404,31 @@ struct redux_impl<Func, Derived, LinearVectorizedTraversal, CompleteUnrolling>
 /** \returns the result of a full redux operation on the whole matrix or vector using \a func
   *
   * The template parameter \a BinaryOp is the type of the functor \a func which must be
-  * an associative operator. Both current STL and TR1 functor styles are handled.
+  * an associative operator. Both current C++98 and C++11 functor styles are handled.
   *
   * \sa DenseBase::sum(), DenseBase::minCoeff(), DenseBase::maxCoeff(), MatrixBase::colwise(), MatrixBase::rowwise()
   */
 template<typename Derived>
 template<typename Func>
-EIGEN_STRONG_INLINE typename internal::result_of<Func(typename internal::traits<Derived>::Scalar)>::type
+typename internal::traits<Derived>::Scalar
 DenseBase<Derived>::redux(const Func& func) const
 {
-  typedef typename internal::remove_all<typename Derived::Nested>::type ThisNested;
-  return internal::redux_impl<Func, ThisNested>
-            ::run(derived(), func);
+  eigen_assert(this->rows()>0 && this->cols()>0 && "you are using an empty matrix");
+  
+  // FIXME, eval_nest should be handled by redux_evaluator, however:
+  //  - it is currently difficult to provide the right Flags since they are still handled by the expressions
+  //  - handling it here might reduce the number of template instantiations
+//   typedef typename internal::nested_eval<Derived,1>::type ThisNested;
+//   typedef typename internal::remove_all<ThisNested>::type ThisNestedCleaned;
+//   typedef typename internal::redux_evaluator<ThisNestedCleaned> ThisEvaluator;
+//   
+//   ThisNested thisNested(derived());
+//   ThisEvaluator thisEval(thisNested);
+  
+  typedef typename internal::redux_evaluator<Derived> ThisEvaluator;
+  ThisEvaluator thisEval(derived());
+  
+  return internal::redux_impl<Func, ThisEvaluator>::run(thisEval, func);
 }
 
 /** \returns the minimum of all coefficients of \c *this.
@@ -337,7 +438,7 @@ template<typename Derived>
 EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar
 DenseBase<Derived>::minCoeff() const
 {
-  return this->redux(Eigen::internal::scalar_min_op<Scalar>());
+  return derived().redux(Eigen::internal::scalar_min_op<Scalar>());
 }
 
 /** \returns the maximum of all coefficients of \c *this.
@@ -347,7 +448,7 @@ template<typename Derived>
 EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar
 DenseBase<Derived>::maxCoeff() const
 {
-  return this->redux(Eigen::internal::scalar_max_op<Scalar>());
+  return derived().redux(Eigen::internal::scalar_max_op<Scalar>());
 }
 
 /** \returns the sum of all coefficients of *this
@@ -360,7 +461,7 @@ DenseBase<Derived>::sum() const
 {
   if(SizeAtCompileTime==0 || (SizeAtCompileTime==Dynamic && size()==0))
     return Scalar(0);
-  return this->redux(Eigen::internal::scalar_sum_op<Scalar>());
+  return derived().redux(Eigen::internal::scalar_sum_op<Scalar>());
 }
 
 /** \returns the mean of all coefficients of *this
@@ -371,7 +472,7 @@ template<typename Derived>
 EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar
 DenseBase<Derived>::mean() const
 {
-  return Scalar(this->redux(Eigen::internal::scalar_sum_op<Scalar>())) / Scalar(this->size());
+  return Scalar(derived().redux(Eigen::internal::scalar_sum_op<Scalar>())) / Scalar(this->size());
 }
 
 /** \returns the product of all coefficients of *this
@@ -387,7 +488,7 @@ DenseBase<Derived>::prod() const
 {
   if(SizeAtCompileTime==0 || (SizeAtCompileTime==Dynamic && size()==0))
     return Scalar(1);
-  return this->redux(Eigen::internal::scalar_product_op<Scalar>());
+  return derived().redux(Eigen::internal::scalar_product_op<Scalar>());
 }
 
 /** \returns the trace of \c *this, i.e. the sum of the coefficients on the main diagonal.

Eigen/src/Core/Ref.h

@@ -12,24 +12,20 @@
 
 namespace Eigen { 
 
-template<typename Derived> class RefBase;
-template<typename PlainObjectType, int Options = 0,
-         typename StrideType = typename internal::conditional<PlainObjectType::IsVectorAtCompileTime,InnerStride<1>,OuterStride<> >::type > class Ref;
-
 /** \class Ref
   * \ingroup Core_Module
   *
-  * \brief A matrix or vector expression mapping an existing expressions
+  * \brief A matrix or vector expression mapping an existing expression
   *
   * \tparam PlainObjectType the equivalent matrix type of the mapped data
-  * \tparam Options specifies whether the pointer is \c #Aligned, or \c #Unaligned.
+  * \tparam MapOptions specifies the pointer alignment in bytes. It can be: \c #Aligned128, , \c #Aligned64, \c #Aligned32, \c #Aligned16, \c #Aligned8 or \c #Unaligned.
   *                The default is \c #Unaligned.
   * \tparam StrideType optionally specifies strides. By default, Ref implies a contiguous storage along the inner dimension (inner stride==1),
-  *                   but accept a variable outer stride (leading dimension).
+  *                   but accepts a variable outer stride (leading dimension).
   *                   This can be overridden by specifying strides.
   *                   The type passed here must be a specialization of the Stride template, see examples below.
   *
-  * This class permits to write non template functions taking Eigen's object as parameters while limiting the number of copies.
+  * This class provides a way to write non-template functions taking Eigen objects as parameters while limiting the number of copies.
   * A Ref<> object can represent either a const expression or a l-value:
   * \code
   * // in-out argument:
@@ -39,10 +35,10 @@ template<typename PlainObjectType, int Options = 0,
   * void foo2(const Ref<const VectorXf>& x);
   * \endcode
   *
-  * In the in-out case, the input argument must satisfies the constraints of the actual Ref<> type, otherwise a compilation issue will be triggered.
+  * In the in-out case, the input argument must satisfy the constraints of the actual Ref<> type, otherwise a compilation issue will be triggered.
   * By default, a Ref<VectorXf> can reference any dense vector expression of float having a contiguous memory layout.
-  * Likewise, a Ref<MatrixXf> can reference any column major dense matrix expression of float whose column's elements are contiguously stored with
-  * the possibility to have a constant space inbetween each column, i.e.: the inner stride mmust be equal to 1, but the outer-stride (or leading dimension),
+  * Likewise, a Ref<MatrixXf> can reference any column-major dense matrix expression of float whose column's elements are contiguously stored with
+  * the possibility to have a constant space in-between each column, i.e. the inner stride must be equal to 1, but the outer stride (or leading dimension)
   * can be greater than the number of rows.
   *
   * In the const case, if the input expression does not match the above requirement, then it is evaluated into a temporary before being passed to the function.
@@ -52,21 +48,22 @@ template<typename PlainObjectType, int Options = 0,
   * VectorXf a;
   * foo1(a.head());             // OK
   * foo1(A.col());              // OK
-  * foo1(A.row());              // compilation error because here innerstride!=1
-  * foo2(A.row());              // The row is copied into a contiguous temporary
+  * foo1(A.row());              // Compilation error because here innerstride!=1
+  * foo2(A.row());              // Compilation error because A.row() is a 1xN object while foo2 is expecting a Nx1 object
+  * foo2(A.row().transpose());  // The row is copied into a contiguous temporary
   * foo2(2*a);                  // The expression is evaluated into a temporary
   * foo2(A.col().segment(2,4)); // No temporary
   * \endcode
   *
-  * The range of inputs that can be referenced without temporary can be enlarged using the last two template parameter.
+  * The range of inputs that can be referenced without temporary can be enlarged using the last two template parameters.
   * Here is an example accepting an innerstride!=1:
   * \code
   * // in-out argument:
   * void foo3(Ref<VectorXf,0,InnerStride<> > x);
   * foo3(A.row());              // OK
   * \endcode
-  * The downside here is that the function foo3 might be significantly slower than foo1 because it won't be able to exploit vectorization, and will involved more
-  * expensive address computations even if the input is contiguously stored in memory. To overcome this issue, one might propose to overloads internally calling a
+  * The downside here is that the function foo3 might be significantly slower than foo1 because it won't be able to exploit vectorization, and will involve more
+  * expensive address computations even if the input is contiguously stored in memory. To overcome this issue, one might propose to overload internally calling a
   * template function, e.g.:
   * \code
   * // in the .h:
@@ -94,24 +91,27 @@ struct traits<Ref<_PlainObjectType, _Options, _StrideType> >
   typedef _PlainObjectType PlainObjectType;
   typedef _StrideType StrideType;
   enum {
-    Options = _Options
+    Options = _Options,
+    Flags = traits<Map<_PlainObjectType, _Options, _StrideType> >::Flags | NestByRefBit,
+    Alignment = traits<Map<_PlainObjectType, _Options, _StrideType> >::Alignment
   };
 
   template<typename Derived> struct match {
     enum {
       HasDirectAccess = internal::has_direct_access<Derived>::ret,
-      StorageOrderMatch = PlainObjectType::IsVectorAtCompileTime || ((PlainObjectType::Flags&RowMajorBit)==(Derived::Flags&RowMajorBit)),
+      StorageOrderMatch = PlainObjectType::IsVectorAtCompileTime || Derived::IsVectorAtCompileTime || ((PlainObjectType::Flags&RowMajorBit)==(Derived::Flags&RowMajorBit)),
       InnerStrideMatch = int(StrideType::InnerStrideAtCompileTime)==int(Dynamic)
                       || int(StrideType::InnerStrideAtCompileTime)==int(Derived::InnerStrideAtCompileTime)
                       || (int(StrideType::InnerStrideAtCompileTime)==0 && int(Derived::InnerStrideAtCompileTime)==1),
       OuterStrideMatch = Derived::IsVectorAtCompileTime
                       || int(StrideType::OuterStrideAtCompileTime)==int(Dynamic) || int(StrideType::OuterStrideAtCompileTime)==int(Derived::OuterStrideAtCompileTime),
-      AlignmentMatch = (_Options!=Aligned) || ((PlainObjectType::Flags&AlignedBit)==0) || ((traits<Derived>::Flags&AlignedBit)==AlignedBit),
-      MatchAtCompileTime = HasDirectAccess && StorageOrderMatch && InnerStrideMatch && OuterStrideMatch && AlignmentMatch
+      AlignmentMatch = (int(traits<PlainObjectType>::Alignment)==int(Unaligned)) || (int(evaluator<Derived>::Alignment) >= int(Alignment)), // FIXME the first condition is not very clear, it should be replaced by the required alignment
+      ScalarTypeMatch = internal::is_same<typename PlainObjectType::Scalar, typename Derived::Scalar>::value,
+      MatchAtCompileTime = HasDirectAccess && StorageOrderMatch && InnerStrideMatch && OuterStrideMatch && AlignmentMatch && ScalarTypeMatch
     };
     typedef typename internal::conditional<MatchAtCompileTime,internal::true_type,internal::false_type>::type type;
   };
-
+  
 };
 
 template<typename Derived>
@@ -130,12 +130,12 @@ public:
   typedef MapBase<Derived> Base;
   EIGEN_DENSE_PUBLIC_INTERFACE(RefBase)
 
-  inline Index innerStride() const
+  EIGEN_DEVICE_FUNC inline Index innerStride() const
   {
     return StrideType::InnerStrideAtCompileTime != 0 ? m_stride.inner() : 1;
   }
 
-  inline Index outerStride() const
+  EIGEN_DEVICE_FUNC inline Index outerStride() const
   {
     return StrideType::OuterStrideAtCompileTime != 0 ? m_stride.outer()
          : IsVectorAtCompileTime ? this->size()
@@ -143,7 +143,7 @@ public:
          : this->rows();
   }
 
-  RefBase()
+  EIGEN_DEVICE_FUNC RefBase()
     : Base(0,RowsAtCompileTime==Dynamic?0:RowsAtCompileTime,ColsAtCompileTime==Dynamic?0:ColsAtCompileTime),
       // Stride<> does not allow default ctor for Dynamic strides, so let' initialize it with dummy values:
       m_stride(StrideType::OuterStrideAtCompileTime==Dynamic?0:StrideType::OuterStrideAtCompileTime,
@@ -157,7 +157,7 @@ protected:
   typedef Stride<StrideType::OuterStrideAtCompileTime,StrideType::InnerStrideAtCompileTime> StrideBase;
 
   template<typename Expression>
-  void construct(Expression& expr)
+  EIGEN_DEVICE_FUNC void construct(Expression& expr)
   {
     if(PlainObjectType::RowsAtCompileTime==1)
     {
@@ -171,8 +171,12 @@ protected:
     }
     else
       ::new (static_cast<Base*>(this)) Base(expr.data(), expr.rows(), expr.cols());
-    ::new (&m_stride) StrideBase(StrideType::OuterStrideAtCompileTime==0?0:expr.outerStride(),
-                                 StrideType::InnerStrideAtCompileTime==0?0:expr.innerStride());    
+    
+    if(Expression::IsVectorAtCompileTime && (!PlainObjectType::IsVectorAtCompileTime) && ((Expression::Flags&RowMajorBit)!=(PlainObjectType::Flags&RowMajorBit)))
+      ::new (&m_stride) StrideBase(expr.innerStride(), StrideType::InnerStrideAtCompileTime==0?0:1);
+    else
+      ::new (&m_stride) StrideBase(StrideType::OuterStrideAtCompileTime==0?0:expr.outerStride(),
+                                   StrideType::InnerStrideAtCompileTime==0?0:expr.innerStride());    
   }
 
   StrideBase m_stride;
@@ -182,7 +186,11 @@ protected:
 template<typename PlainObjectType, int Options, typename StrideType> class Ref
   : public RefBase<Ref<PlainObjectType, Options, StrideType> >
 {
+  private:
     typedef internal::traits<Ref> Traits;
+    template<typename Derived>
+    EIGEN_DEVICE_FUNC inline Ref(const PlainObjectBase<Derived>& expr,
+                                 typename internal::enable_if<bool(Traits::template match<Derived>::MatchAtCompileTime),Derived>::type* = 0);
   public:
 
     typedef RefBase<Ref> Base;
@@ -191,20 +199,23 @@ template<typename PlainObjectType, int Options, typename StrideType> class Ref
 
     #ifndef EIGEN_PARSED_BY_DOXYGEN
     template<typename Derived>
-    inline Ref(PlainObjectBase<Derived>& expr,
-               typename internal::enable_if<bool(Traits::template match<Derived>::MatchAtCompileTime),Derived>::type* = 0)
+    EIGEN_DEVICE_FUNC inline Ref(PlainObjectBase<Derived>& expr,
+                                 typename internal::enable_if<bool(Traits::template match<Derived>::MatchAtCompileTime),Derived>::type* = 0)
     {
-      Base::construct(expr);
+      EIGEN_STATIC_ASSERT(bool(Traits::template match<Derived>::MatchAtCompileTime), STORAGE_LAYOUT_DOES_NOT_MATCH);
+      Base::construct(expr.derived());
     }
     template<typename Derived>
-    inline Ref(const DenseBase<Derived>& expr,
-               typename internal::enable_if<bool(internal::is_lvalue<Derived>::value&&bool(Traits::template match<Derived>::MatchAtCompileTime)),Derived>::type* = 0,
-               int = Derived::ThisConstantIsPrivateInPlainObjectBase)
+    EIGEN_DEVICE_FUNC inline Ref(const DenseBase<Derived>& expr,
+                                 typename internal::enable_if<bool(Traits::template match<Derived>::MatchAtCompileTime),Derived>::type* = 0)
     #else
     template<typename Derived>
     inline Ref(DenseBase<Derived>& expr)
     #endif
     {
+      EIGEN_STATIC_ASSERT(bool(internal::is_lvalue<Derived>::value), THIS_EXPRESSION_IS_NOT_A_LVALUE__IT_IS_READ_ONLY);
+      EIGEN_STATIC_ASSERT(bool(Traits::template match<Derived>::MatchAtCompileTime), STORAGE_LAYOUT_DOES_NOT_MATCH);
+      EIGEN_STATIC_ASSERT(!Derived::IsPlainObjectBase,THIS_EXPRESSION_IS_NOT_A_LVALUE__IT_IS_READ_ONLY);
       Base::construct(expr.const_cast_derived());
     }
 
@@ -223,7 +234,8 @@ template<typename TPlainObjectType, int Options, typename StrideType> class Ref<
     EIGEN_DENSE_PUBLIC_INTERFACE(Ref)
 
     template<typename Derived>
-    inline Ref(const DenseBase<Derived>& expr)
+    EIGEN_DEVICE_FUNC inline Ref(const DenseBase<Derived>& expr,
+                                 typename internal::enable_if<bool(Traits::template match<Derived>::ScalarTypeMatch),Derived>::type* = 0)
     {
 //      std::cout << match_helper<Derived>::HasDirectAccess << "," << match_helper<Derived>::OuterStrideMatch << "," << match_helper<Derived>::InnerStrideMatch << "\n";
 //      std::cout << int(StrideType::OuterStrideAtCompileTime) << " - " << int(Derived::OuterStrideAtCompileTime) << "\n";
@@ -231,18 +243,27 @@ template<typename TPlainObjectType, int Options, typename StrideType> class Ref<
       construct(expr.derived(), typename Traits::template match<Derived>::type());
     }
 
+    EIGEN_DEVICE_FUNC inline Ref(const Ref& other) : Base(other) {
+      // copy constructor shall not copy the m_object, to avoid unnecessary malloc and copy
+    }
+
+    template<typename OtherRef>
+    EIGEN_DEVICE_FUNC inline Ref(const RefBase<OtherRef>& other) {
+      construct(other.derived(), typename Traits::template match<OtherRef>::type());
+    }
+
   protected:
 
     template<typename Expression>
-    void construct(const Expression& expr,internal::true_type)
+    EIGEN_DEVICE_FUNC void construct(const Expression& expr,internal::true_type)
     {
       Base::construct(expr);
     }
 
     template<typename Expression>
-    void construct(const Expression& expr, internal::false_type)
+    EIGEN_DEVICE_FUNC void construct(const Expression& expr, internal::false_type)
     {
-      m_object.lazyAssign(expr);
+      internal::call_assignment_no_alias(m_object,expr,internal::assign_op<Scalar>());
       Base::construct(m_object);
     }
 

Eigen/src/Core/Replicate.h

@@ -35,10 +35,7 @@ struct traits<Replicate<MatrixType,RowFactor,ColFactor> >
   typedef typename MatrixType::Scalar Scalar;
   typedef typename traits<MatrixType>::StorageKind StorageKind;
   typedef typename traits<MatrixType>::XprKind XprKind;
-  enum {
-    Factor = (RowFactor==Dynamic || ColFactor==Dynamic) ? Dynamic : RowFactor*ColFactor
-  };
-  typedef typename nested<MatrixType,Factor>::type MatrixTypeNested;
+  typedef typename ref_selector<MatrixType>::type MatrixTypeNested;
   typedef typename remove_reference<MatrixTypeNested>::type _MatrixTypeNested;
   enum {
     RowsAtCompileTime = RowFactor==Dynamic || int(MatrixType::RowsAtCompileTime)==Dynamic
@@ -53,8 +50,9 @@ struct traits<Replicate<MatrixType,RowFactor,ColFactor> >
     IsRowMajor = MaxRowsAtCompileTime==1 && MaxColsAtCompileTime!=1 ? 1
                : MaxColsAtCompileTime==1 && MaxRowsAtCompileTime!=1 ? 0
                : (MatrixType::Flags & RowMajorBit) ? 1 : 0,
-    Flags = (_MatrixTypeNested::Flags & HereditaryBits & ~RowMajorBit) | (IsRowMajor ? RowMajorBit : 0),
-    CoeffReadCost = _MatrixTypeNested::CoeffReadCost
+    
+    // FIXME enable DirectAccess with negative strides?
+    Flags = IsRowMajor ? RowMajorBit : 0
   };
 };
 }
@@ -68,10 +66,12 @@ template<typename MatrixType,int RowFactor,int ColFactor> class Replicate
 
     typedef typename internal::dense_xpr_base<Replicate>::type Base;
     EIGEN_DENSE_PUBLIC_INTERFACE(Replicate)
+    typedef typename internal::remove_all<MatrixType>::type NestedExpression;
 
     template<typename OriginalMatrixType>
-    inline explicit Replicate(const OriginalMatrixType& a_matrix)
-      : m_matrix(a_matrix), m_rowFactor(RowFactor), m_colFactor(ColFactor)
+    EIGEN_DEVICE_FUNC
+    inline explicit Replicate(const OriginalMatrixType& matrix)
+      : m_matrix(matrix), m_rowFactor(RowFactor), m_colFactor(ColFactor)
     {
       EIGEN_STATIC_ASSERT((internal::is_same<typename internal::remove_const<MatrixType>::type,OriginalMatrixType>::value),
                           THE_MATRIX_OR_EXPRESSION_THAT_YOU_PASSED_DOES_NOT_HAVE_THE_EXPECTED_TYPE)
@@ -79,41 +79,20 @@ template<typename MatrixType,int RowFactor,int ColFactor> class Replicate
     }
 
     template<typename OriginalMatrixType>
-    inline Replicate(const OriginalMatrixType& a_matrix, Index rowFactor, Index colFactor)
-      : m_matrix(a_matrix), m_rowFactor(rowFactor), m_colFactor(colFactor)
+    EIGEN_DEVICE_FUNC
+    inline Replicate(const OriginalMatrixType& matrix, Index rowFactor, Index colFactor)
+      : m_matrix(matrix), m_rowFactor(rowFactor), m_colFactor(colFactor)
     {
       EIGEN_STATIC_ASSERT((internal::is_same<typename internal::remove_const<MatrixType>::type,OriginalMatrixType>::value),
                           THE_MATRIX_OR_EXPRESSION_THAT_YOU_PASSED_DOES_NOT_HAVE_THE_EXPECTED_TYPE)
     }
 
+    EIGEN_DEVICE_FUNC
     inline Index rows() const { return m_matrix.rows() * m_rowFactor.value(); }
+    EIGEN_DEVICE_FUNC
     inline Index cols() const { return m_matrix.cols() * m_colFactor.value(); }
 
-    inline Scalar coeff(Index rowId, Index colId) const
-    {
-      // try to avoid using modulo; this is a pure optimization strategy
-      const Index actual_row  = internal::traits<MatrixType>::RowsAtCompileTime==1 ? 0
-                            : RowFactor==1 ? rowId
-                            : rowId%m_matrix.rows();
-      const Index actual_col  = internal::traits<MatrixType>::ColsAtCompileTime==1 ? 0
-                            : ColFactor==1 ? colId
-                            : colId%m_matrix.cols();
-
-      return m_matrix.coeff(actual_row, actual_col);
-    }
-    template<int LoadMode>
-    inline PacketScalar packet(Index rowId, Index colId) const
-    {
-      const Index actual_row  = internal::traits<MatrixType>::RowsAtCompileTime==1 ? 0
-                            : RowFactor==1 ? rowId
-                            : rowId%m_matrix.rows();
-      const Index actual_col  = internal::traits<MatrixType>::ColsAtCompileTime==1 ? 0
-                            : ColFactor==1 ? colId
-                            : colId%m_matrix.cols();
-
-      return m_matrix.template packet<LoadMode>(actual_row, actual_col);
-    }
-
+    EIGEN_DEVICE_FUNC
     const _MatrixTypeNested& nestedExpression() const
     { 
       return m_matrix; 
@@ -135,27 +114,12 @@ template<typename MatrixType,int RowFactor,int ColFactor> class Replicate
   */
 template<typename Derived>
 template<int RowFactor, int ColFactor>
-inline const Replicate<Derived,RowFactor,ColFactor>
+const Replicate<Derived,RowFactor,ColFactor>
 DenseBase<Derived>::replicate() const
 {
   return Replicate<Derived,RowFactor,ColFactor>(derived());
 }
 
-/**
-  * \return an expression of the replication of \c *this
-  *
-  * Example: \include MatrixBase_replicate_int_int.cpp
-  * Output: \verbinclude MatrixBase_replicate_int_int.out
-  *
-  * \sa VectorwiseOp::replicate(), DenseBase::replicate<int,int>(), class Replicate
-  */
-template<typename Derived>
-inline const Replicate<Derived,Dynamic,Dynamic>
-DenseBase<Derived>::replicate(Index rowFactor,Index colFactor) const
-{
-  return Replicate<Derived,Dynamic,Dynamic>(derived(),rowFactor,colFactor);
-}
-
 /**
   * \return an expression of the replication of each column (or row) of \c *this
   *

Eigen/src/Core/ReturnByValue.h

@@ -38,9 +38,10 @@ struct traits<ReturnByValue<Derived> >
  * So internal::nested always gives the plain return matrix type.
  *
  * FIXME: I don't understand why we need this specialization: isn't this taken care of by the EvalBeforeNestingBit ??
+ * Answer: EvalBeforeNestingBit should be deprecated since we have the evaluators
  */
 template<typename Derived,int n,typename PlainObject>
-struct nested<ReturnByValue<Derived>, n, PlainObject>
+struct nested_eval<ReturnByValue<Derived>, n, PlainObject>
 {
   typedef typename traits<Derived>::ReturnType type;
 };
@@ -48,7 +49,7 @@ struct nested<ReturnByValue<Derived>, n, PlainObject>
 } // end namespace internal
 
 template<typename Derived> class ReturnByValue
-  : internal::no_assignment_operator, public internal::dense_xpr_base< ReturnByValue<Derived> >::type
+  : public internal::dense_xpr_base< ReturnByValue<Derived> >::type, internal::no_assignment_operator
 {
   public:
     typedef typename internal::traits<Derived>::ReturnType ReturnType;
@@ -57,10 +58,11 @@ template<typename Derived> class ReturnByValue
     EIGEN_DENSE_PUBLIC_INTERFACE(ReturnByValue)
 
     template<typename Dest>
+    EIGEN_DEVICE_FUNC
     inline void evalTo(Dest& dst) const
     { static_cast<const Derived*>(this)->evalTo(dst); }
-    inline Index rows() const { return static_cast<const Derived*>(this)->rows(); }
-    inline Index cols() const { return static_cast<const Derived*>(this)->cols(); }
+    EIGEN_DEVICE_FUNC inline Index rows() const { return static_cast<const Derived*>(this)->rows(); }
+    EIGEN_DEVICE_FUNC inline Index cols() const { return static_cast<const Derived*>(this)->cols(); }
 
 #ifndef EIGEN_PARSED_BY_DOXYGEN
 #define Unusable YOU_ARE_TRYING_TO_ACCESS_A_SINGLE_COEFFICIENT_IN_A_SPECIAL_EXPRESSION_WHERE_THAT_IS_NOT_ALLOWED_BECAUSE_THAT_WOULD_BE_INEFFICIENT
@@ -72,6 +74,7 @@ template<typename Derived> class ReturnByValue
     const Unusable& coeff(Index,Index) const { return *reinterpret_cast<const Unusable*>(this); }
     Unusable& coeffRef(Index) { return *reinterpret_cast<Unusable*>(this); }
     Unusable& coeffRef(Index,Index) { return *reinterpret_cast<Unusable*>(this); }
+#undef Unusable
 #endif
 };
 
@@ -83,6 +86,33 @@ Derived& DenseBase<Derived>::operator=(const ReturnByValue<OtherDerived>& other)
   return derived();
 }
 
+namespace internal {
+
+// Expression is evaluated in a temporary; default implementation of Assignment is bypassed so that
+// when a ReturnByValue expression is assigned, the evaluator is not constructed.
+// TODO: Finalize port to new regime; ReturnByValue should not exist in the expression world
+  
+template<typename Derived>
+struct evaluator<ReturnByValue<Derived> >
+  : public evaluator<typename internal::traits<Derived>::ReturnType>
+{
+  typedef ReturnByValue<Derived> XprType;
+  typedef typename internal::traits<Derived>::ReturnType PlainObject;
+  typedef evaluator<PlainObject> Base;
+  
+  EIGEN_DEVICE_FUNC explicit evaluator(const XprType& xpr)
+    : m_result(xpr.rows(), xpr.cols())
+  {
+    ::new (static_cast<Base*>(this)) Base(m_result);
+    xpr.evalTo(m_result);
+  }
+
+protected:
+  PlainObject m_result;
+};
+
+} // end namespace internal
+
 } // end namespace Eigen
 
 #endif // EIGEN_RETURNBYVALUE_H

Eigen/src/Core/Reverse.h

@@ -37,32 +37,25 @@ struct traits<Reverse<MatrixType, Direction> >
   typedef typename MatrixType::Scalar Scalar;
   typedef typename traits<MatrixType>::StorageKind StorageKind;
   typedef typename traits<MatrixType>::XprKind XprKind;
-  typedef typename nested<MatrixType>::type MatrixTypeNested;
+  typedef typename ref_selector<MatrixType>::type MatrixTypeNested;
   typedef typename remove_reference<MatrixTypeNested>::type _MatrixTypeNested;
   enum {
     RowsAtCompileTime = MatrixType::RowsAtCompileTime,
     ColsAtCompileTime = MatrixType::ColsAtCompileTime,
     MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime,
     MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime,
-
-    // let's enable LinearAccess only with vectorization because of the product overhead
-    LinearAccess = ( (Direction==BothDirections) && (int(_MatrixTypeNested::Flags)&PacketAccessBit) )
-                 ? LinearAccessBit : 0,
-
-    Flags = int(_MatrixTypeNested::Flags) & (HereditaryBits | LvalueBit | PacketAccessBit | LinearAccess),
-
-    CoeffReadCost = _MatrixTypeNested::CoeffReadCost
+    Flags = _MatrixTypeNested::Flags & (RowMajorBit | LvalueBit)
   };
 };
 
-template<typename PacketScalar, bool ReversePacket> struct reverse_packet_cond
+template<typename PacketType, bool ReversePacket> struct reverse_packet_cond
 {
-  static inline PacketScalar run(const PacketScalar& x) { return preverse(x); }
+  static inline PacketType run(const PacketType& x) { return preverse(x); }
 };
 
-template<typename PacketScalar> struct reverse_packet_cond<PacketScalar,false>
+template<typename PacketType> struct reverse_packet_cond<PacketType,false>
 {
-  static inline PacketScalar run(const PacketScalar& x) { return x; }
+  static inline PacketType run(const PacketType& x) { return x; }
 };
 
 } // end namespace internal 
@@ -74,12 +67,9 @@ template<typename MatrixType, int Direction> class Reverse
 
     typedef typename internal::dense_xpr_base<Reverse>::type Base;
     EIGEN_DENSE_PUBLIC_INTERFACE(Reverse)
+    typedef typename internal::remove_all<MatrixType>::type NestedExpression;
     using Base::IsRowMajor;
 
-    // next line is necessary because otherwise const version of operator()
-    // is hidden by non-const version defined in this file
-    using Base::operator(); 
-
   protected:
     enum {
       PacketSize = internal::packet_traits<Scalar>::size,
@@ -95,82 +85,19 @@ template<typename MatrixType, int Direction> class Reverse
     typedef internal::reverse_packet_cond<PacketScalar,ReversePacket> reverse_packet;
   public:
 
-    inline Reverse(const MatrixType& matrix) : m_matrix(matrix) { }
+    EIGEN_DEVICE_FUNC explicit inline Reverse(const MatrixType& matrix) : m_matrix(matrix) { }
 
     EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Reverse)
 
-    inline Index rows() const { return m_matrix.rows(); }
-    inline Index cols() const { return m_matrix.cols(); }
+    EIGEN_DEVICE_FUNC inline Index rows() const { return m_matrix.rows(); }
+    EIGEN_DEVICE_FUNC inline Index cols() const { return m_matrix.cols(); }
 
-    inline Index innerStride() const
+    EIGEN_DEVICE_FUNC inline Index innerStride() const
     {
       return -m_matrix.innerStride();
     }
 
-    inline Scalar& operator()(Index row, Index col)
-    {
-      eigen_assert(row >= 0 && row < rows() && col >= 0 && col < cols());
-      return coeffRef(row, col);
-    }
-
-    inline Scalar& coeffRef(Index row, Index col)
-    {
-      return m_matrix.const_cast_derived().coeffRef(ReverseRow ? m_matrix.rows() - row - 1 : row,
-                                                    ReverseCol ? m_matrix.cols() - col - 1 : col);
-    }
-
-    inline CoeffReturnType coeff(Index row, Index col) const
-    {
-      return m_matrix.coeff(ReverseRow ? m_matrix.rows() - row - 1 : row,
-                            ReverseCol ? m_matrix.cols() - col - 1 : col);
-    }
-
-    inline CoeffReturnType coeff(Index index) const
-    {
-      return m_matrix.coeff(m_matrix.size() - index - 1);
-    }
-
-    inline Scalar& coeffRef(Index index)
-    {
-      return m_matrix.const_cast_derived().coeffRef(m_matrix.size() - index - 1);
-    }
-
-    inline Scalar& operator()(Index index)
-    {
-      eigen_assert(index >= 0 && index < m_matrix.size());
-      return coeffRef(index);
-    }
-
-    template<int LoadMode>
-    inline const PacketScalar packet(Index row, Index col) const
-    {
-      return reverse_packet::run(m_matrix.template packet<LoadMode>(
-                                    ReverseRow ? m_matrix.rows() - row - OffsetRow : row,
-                                    ReverseCol ? m_matrix.cols() - col - OffsetCol : col));
-    }
-
-    template<int LoadMode>
-    inline void writePacket(Index row, Index col, const PacketScalar& x)
-    {
-      m_matrix.const_cast_derived().template writePacket<LoadMode>(
-                                      ReverseRow ? m_matrix.rows() - row - OffsetRow : row,
-                                      ReverseCol ? m_matrix.cols() - col - OffsetCol : col,
-                                      reverse_packet::run(x));
-    }
-
-    template<int LoadMode>
-    inline const PacketScalar packet(Index index) const
-    {
-      return internal::preverse(m_matrix.template packet<LoadMode>( m_matrix.size() - index - PacketSize ));
-    }
-
-    template<int LoadMode>
-    inline void writePacket(Index index, const PacketScalar& x)
-    {
-      m_matrix.const_cast_derived().template writePacket<LoadMode>(m_matrix.size() - index - PacketSize, internal::preverse(x));
-    }
-
-    const typename internal::remove_all<typename MatrixType::Nested>::type& 
+    EIGEN_DEVICE_FUNC const typename internal::remove_all<typename MatrixType::Nested>::type&
     nestedExpression() const 
     {
       return m_matrix;
@@ -190,33 +117,93 @@ template<typename Derived>
 inline typename DenseBase<Derived>::ReverseReturnType
 DenseBase<Derived>::reverse()
 {
-  return derived();
+  return ReverseReturnType(derived());
 }
 
-/** This is the const version of reverse(). */
-template<typename Derived>
-inline const typename DenseBase<Derived>::ConstReverseReturnType
-DenseBase<Derived>::reverse() const
-{
-  return derived();
-}
+
+//reverse const overload moved DenseBase.h due to a CUDA compiler bug
 
 /** This is the "in place" version of reverse: it reverses \c *this.
   *
   * In most cases it is probably better to simply use the reversed expression
   * of a matrix. However, when reversing the matrix data itself is really needed,
   * then this "in-place" version is probably the right choice because it provides
-  * the following additional features:
+  * the following additional benefits:
   *  - less error prone: doing the same operation with .reverse() requires special care:
   *    \code m = m.reverse().eval(); \endcode
-  *  - this API allows to avoid creating a temporary (the current implementation creates a temporary, but that could be avoided using swap)
+  *  - this API enables reverse operations without the need for a temporary
   *  - it allows future optimizations (cache friendliness, etc.)
   *
-  * \sa reverse() */
+  * \sa VectorwiseOp::reverseInPlace(), reverse() */
 template<typename Derived>
 inline void DenseBase<Derived>::reverseInPlace()
 {
-  derived() = derived().reverse().eval();
+  if(cols()>rows())
+  {
+    Index half = cols()/2;
+    leftCols(half).swap(rightCols(half).reverse());
+    if((cols()%2)==1)
+    {
+      Index half2 = rows()/2;
+      col(half).head(half2).swap(col(half).tail(half2).reverse());
+    }
+  }
+  else
+  {
+    Index half = rows()/2;
+    topRows(half).swap(bottomRows(half).reverse());
+    if((rows()%2)==1)
+    {
+      Index half2 = cols()/2;
+      row(half).head(half2).swap(row(half).tail(half2).reverse());
+    }
+  }
+}
+
+namespace internal {
+  
+template<int Direction>
+struct vectorwise_reverse_inplace_impl;
+
+template<>
+struct vectorwise_reverse_inplace_impl<Vertical>
+{
+  template<typename ExpressionType>
+  static void run(ExpressionType &xpr)
+  {
+    Index half = xpr.rows()/2;
+    xpr.topRows(half).swap(xpr.bottomRows(half).colwise().reverse());
+  }
+};
+
+template<>
+struct vectorwise_reverse_inplace_impl<Horizontal>
+{
+  template<typename ExpressionType>
+  static void run(ExpressionType &xpr)
+  {
+    Index half = xpr.cols()/2;
+    xpr.leftCols(half).swap(xpr.rightCols(half).rowwise().reverse());
+  }
+};
+
+} // end namespace internal
+
+/** This is the "in place" version of VectorwiseOp::reverse: it reverses each column or row of \c *this.
+  *
+  * In most cases it is probably better to simply use the reversed expression
+  * of a matrix. However, when reversing the matrix data itself is really needed,
+  * then this "in-place" version is probably the right choice because it provides
+  * the following additional benefits:
+  *  - less error prone: doing the same operation with .reverse() requires special care:
+  *    \code m = m.reverse().eval(); \endcode
+  *  - this API enables reverse operations without the need for a temporary
+  *
+  * \sa DenseBase::reverseInPlace(), reverse() */
+template<typename ExpressionType, int Direction>
+void VectorwiseOp<ExpressionType,Direction>::reverseInPlace()
+{
+  internal::vectorwise_reverse_inplace_impl<Direction>::run(_expression().const_cast_derived());
 }
 
 } // end namespace Eigen

Eigen/src/Core/Select.h

@@ -43,23 +43,21 @@ struct traits<Select<ConditionMatrixType, ThenMatrixType, ElseMatrixType> >
     ColsAtCompileTime = ConditionMatrixType::ColsAtCompileTime,
     MaxRowsAtCompileTime = ConditionMatrixType::MaxRowsAtCompileTime,
     MaxColsAtCompileTime = ConditionMatrixType::MaxColsAtCompileTime,
-    Flags = (unsigned int)ThenMatrixType::Flags & ElseMatrixType::Flags & HereditaryBits,
-    CoeffReadCost = traits<typename remove_all<ConditionMatrixNested>::type>::CoeffReadCost
-                  + EIGEN_SIZE_MAX(traits<typename remove_all<ThenMatrixNested>::type>::CoeffReadCost,
-                                   traits<typename remove_all<ElseMatrixNested>::type>::CoeffReadCost)
+    Flags = (unsigned int)ThenMatrixType::Flags & ElseMatrixType::Flags & RowMajorBit
   };
 };
 }
 
 template<typename ConditionMatrixType, typename ThenMatrixType, typename ElseMatrixType>
-class Select : internal::no_assignment_operator,
-  public internal::dense_xpr_base< Select<ConditionMatrixType, ThenMatrixType, ElseMatrixType> >::type
+class Select : public internal::dense_xpr_base< Select<ConditionMatrixType, ThenMatrixType, ElseMatrixType> >::type,
+               internal::no_assignment_operator
 {
   public:
 
     typedef typename internal::dense_xpr_base<Select>::type Base;
     EIGEN_DENSE_PUBLIC_INTERFACE(Select)
 
+    inline EIGEN_DEVICE_FUNC
     Select(const ConditionMatrixType& a_conditionMatrix,
            const ThenMatrixType& a_thenMatrix,
            const ElseMatrixType& a_elseMatrix)
@@ -69,9 +67,10 @@ class Select : internal::no_assignment_operator,
       eigen_assert(m_condition.cols() == m_then.cols() && m_condition.cols() == m_else.cols());
     }
 
-    Index rows() const { return m_condition.rows(); }
-    Index cols() const { return m_condition.cols(); }
+    inline EIGEN_DEVICE_FUNC Index rows() const { return m_condition.rows(); }
+    inline EIGEN_DEVICE_FUNC Index cols() const { return m_condition.cols(); }
 
+    inline EIGEN_DEVICE_FUNC
     const Scalar coeff(Index i, Index j) const
     {
       if (m_condition.coeff(i,j))
@@ -80,6 +79,7 @@ class Select : internal::no_assignment_operator,
         return m_else.coeff(i,j);
     }
 
+    inline EIGEN_DEVICE_FUNC
     const Scalar coeff(Index i) const
     {
       if (m_condition.coeff(i))
@@ -88,17 +88,17 @@ class Select : internal::no_assignment_operator,
         return m_else.coeff(i);
     }
 
-    const ConditionMatrixType& conditionMatrix() const
+    inline EIGEN_DEVICE_FUNC const ConditionMatrixType& conditionMatrix() const
     {
       return m_condition;
     }
 
-    const ThenMatrixType& thenMatrix() const
+    inline EIGEN_DEVICE_FUNC const ThenMatrixType& thenMatrix() const
     {
       return m_then;
     }
 
-    const ElseMatrixType& elseMatrix() const
+    inline EIGEN_DEVICE_FUNC const ElseMatrixType& elseMatrix() const
     {
       return m_else;
     }

Eigen/src/Core/SelfAdjointView.h

@@ -32,54 +32,57 @@ namespace internal {
 template<typename MatrixType, unsigned int UpLo>
 struct traits<SelfAdjointView<MatrixType, UpLo> > : traits<MatrixType>
 {
-  typedef typename nested<MatrixType>::type MatrixTypeNested;
+  typedef typename ref_selector<MatrixType>::type MatrixTypeNested;
   typedef typename remove_all<MatrixTypeNested>::type MatrixTypeNestedCleaned;
   typedef MatrixType ExpressionType;
-  typedef typename MatrixType::PlainObject DenseMatrixType;
+  typedef typename MatrixType::PlainObject FullMatrixType;
   enum {
     Mode = UpLo | SelfAdjoint,
-    Flags =  MatrixTypeNestedCleaned::Flags & (HereditaryBits)
-           & (~(PacketAccessBit | DirectAccessBit | LinearAccessBit)), // FIXME these flags should be preserved
-    CoeffReadCost = MatrixTypeNestedCleaned::CoeffReadCost
+    FlagsLvalueBit = is_lvalue<MatrixType>::value ? LvalueBit : 0,
+    Flags =  MatrixTypeNestedCleaned::Flags & (HereditaryBits|FlagsLvalueBit)
+           & (~(PacketAccessBit | DirectAccessBit | LinearAccessBit)) // FIXME these flags should be preserved
   };
 };
 }
 
-template <typename Lhs, int LhsMode, bool LhsIsVector,
-          typename Rhs, int RhsMode, bool RhsIsVector>
-struct SelfadjointProductMatrix;
-
 // FIXME could also be called SelfAdjointWrapper to be consistent with DiagonalWrapper ??
-template<typename MatrixType, unsigned int UpLo> class SelfAdjointView
-  : public TriangularBase<SelfAdjointView<MatrixType, UpLo> >
+template<typename _MatrixType, unsigned int UpLo> class SelfAdjointView
+  : public TriangularBase<SelfAdjointView<_MatrixType, UpLo> >
 {
   public:
 
+    typedef _MatrixType MatrixType;
     typedef TriangularBase<SelfAdjointView> Base;
     typedef typename internal::traits<SelfAdjointView>::MatrixTypeNested MatrixTypeNested;
     typedef typename internal::traits<SelfAdjointView>::MatrixTypeNestedCleaned MatrixTypeNestedCleaned;
 
     /** \brief The type of coefficients in this matrix */
     typedef typename internal::traits<SelfAdjointView>::Scalar Scalar; 
-
-    typedef typename MatrixType::Index Index;
+    typedef typename MatrixType::StorageIndex StorageIndex;
 
     enum {
-      Mode = internal::traits<SelfAdjointView>::Mode
+      Mode = internal::traits<SelfAdjointView>::Mode,
+      Flags = internal::traits<SelfAdjointView>::Flags
     };
     typedef typename MatrixType::PlainObject PlainObject;
 
-    inline SelfAdjointView(MatrixType& matrix) : m_matrix(matrix)
+    EIGEN_DEVICE_FUNC
+    explicit inline SelfAdjointView(MatrixType& matrix) : m_matrix(matrix)
     {}
 
+    EIGEN_DEVICE_FUNC
     inline Index rows() const { return m_matrix.rows(); }
+    EIGEN_DEVICE_FUNC
     inline Index cols() const { return m_matrix.cols(); }
+    EIGEN_DEVICE_FUNC
     inline Index outerStride() const { return m_matrix.outerStride(); }
+    EIGEN_DEVICE_FUNC
     inline Index innerStride() const { return m_matrix.innerStride(); }
 
     /** \sa MatrixBase::coeff()
       * \warning the coordinates must fit into the referenced triangular part
       */
+    EIGEN_DEVICE_FUNC
     inline Scalar coeff(Index row, Index col) const
     {
       Base::check_coordinates_internal(row, col);
@@ -89,36 +92,46 @@ template<typename MatrixType, unsigned int UpLo> class SelfAdjointView
     /** \sa MatrixBase::coeffRef()
       * \warning the coordinates must fit into the referenced triangular part
       */
+    EIGEN_DEVICE_FUNC
     inline Scalar& coeffRef(Index row, Index col)
     {
+      EIGEN_STATIC_ASSERT_LVALUE(SelfAdjointView);
       Base::check_coordinates_internal(row, col);
       return m_matrix.const_cast_derived().coeffRef(row, col);
     }
 
     /** \internal */
+    EIGEN_DEVICE_FUNC
     const MatrixTypeNestedCleaned& _expression() const { return m_matrix; }
 
+    EIGEN_DEVICE_FUNC
     const MatrixTypeNestedCleaned& nestedExpression() const { return m_matrix; }
+    EIGEN_DEVICE_FUNC
     MatrixTypeNestedCleaned& nestedExpression() { return *const_cast<MatrixTypeNestedCleaned*>(&m_matrix); }
 
-    /** Efficient self-adjoint matrix times vector/matrix product */
+    /** Efficient triangular matrix times vector/matrix product */
     template<typename OtherDerived>
-    SelfadjointProductMatrix<MatrixType,Mode,false,OtherDerived,0,OtherDerived::IsVectorAtCompileTime>
+    EIGEN_DEVICE_FUNC
+    const Product<SelfAdjointView,OtherDerived>
     operator*(const MatrixBase<OtherDerived>& rhs) const
     {
-      return SelfadjointProductMatrix
-              <MatrixType,Mode,false,OtherDerived,0,OtherDerived::IsVectorAtCompileTime>
-              (m_matrix, rhs.derived());
+      return Product<SelfAdjointView,OtherDerived>(*this, rhs.derived());
     }
 
-    /** Efficient vector/matrix times self-adjoint matrix product */
+    /** Efficient vector/matrix times triangular matrix product */
     template<typename OtherDerived> friend
-    SelfadjointProductMatrix<OtherDerived,0,OtherDerived::IsVectorAtCompileTime,MatrixType,Mode,false>
+    EIGEN_DEVICE_FUNC
+    const Product<OtherDerived,SelfAdjointView>
     operator*(const MatrixBase<OtherDerived>& lhs, const SelfAdjointView& rhs)
     {
-      return SelfadjointProductMatrix
-              <OtherDerived,0,OtherDerived::IsVectorAtCompileTime,MatrixType,Mode,false>
-              (lhs.derived(),rhs.m_matrix);
+      return Product<OtherDerived,SelfAdjointView>(lhs.derived(),rhs);
+    }
+    
+    friend EIGEN_DEVICE_FUNC
+    const SelfAdjointView<const CwiseUnaryOp<internal::scalar_multiple_op<Scalar>,MatrixType>,UpLo>
+    operator*(const Scalar& s, const SelfAdjointView& mat)
+    {
+      return (s*mat.nestedExpression()).template selfadjointView<UpLo>();
     }
 
     /** Perform a symmetric rank 2 update of the selfadjoint matrix \c *this:
@@ -132,6 +145,7 @@ template<typename MatrixType, unsigned int UpLo> class SelfAdjointView
       * \sa rankUpdate(const MatrixBase<DerivedU>&, Scalar)
       */
     template<typename DerivedU, typename DerivedV>
+    EIGEN_DEVICE_FUNC
     SelfAdjointView& rankUpdate(const MatrixBase<DerivedU>& u, const MatrixBase<DerivedV>& v, const Scalar& alpha = Scalar(1));
 
     /** Perform a symmetric rank K update of the selfadjoint matrix \c *this:
@@ -145,6 +159,7 @@ template<typename MatrixType, unsigned int UpLo> class SelfAdjointView
       * \sa rankUpdate(const MatrixBase<DerivedU>&, const MatrixBase<DerivedV>&, Scalar)
       */
     template<typename DerivedU>
+    EIGEN_DEVICE_FUNC
     SelfAdjointView& rankUpdate(const MatrixBase<DerivedU>& u, const Scalar& alpha = Scalar(1));
 
 /////////// Cholesky module ///////////
@@ -159,31 +174,10 @@ template<typename MatrixType, unsigned int UpLo> class SelfAdjointView
     /** Return type of eigenvalues() */
     typedef Matrix<RealScalar, internal::traits<MatrixType>::ColsAtCompileTime, 1> EigenvaluesReturnType;
 
+    EIGEN_DEVICE_FUNC
     EigenvaluesReturnType eigenvalues() const;
+    EIGEN_DEVICE_FUNC
     RealScalar operatorNorm() const;
-    
-    #ifdef EIGEN2_SUPPORT
-    template<typename OtherDerived>
-    SelfAdjointView& operator=(const MatrixBase<OtherDerived>& other)
-    {
-      enum {
-        OtherPart = UpLo == Upper ? StrictlyLower : StrictlyUpper
-      };
-      m_matrix.const_cast_derived().template triangularView<UpLo>() = other;
-      m_matrix.const_cast_derived().template triangularView<OtherPart>() = other.adjoint();
-      return *this;
-    }
-    template<typename OtherMatrixType, unsigned int OtherMode>
-    SelfAdjointView& operator=(const TriangularView<OtherMatrixType, OtherMode>& other)
-    {
-      enum {
-        OtherPart = UpLo == Upper ? StrictlyLower : StrictlyUpper
-      };
-      m_matrix.const_cast_derived().template triangularView<UpLo>() = other.toDenseMatrix();
-      m_matrix.const_cast_derived().template triangularView<OtherPart>() = other.toDenseMatrix().adjoint();
-      return *this;
-    }
-    #endif
 
   protected:
     MatrixTypeNested m_matrix;
@@ -201,90 +195,56 @@ template<typename MatrixType, unsigned int UpLo> class SelfAdjointView
 
 namespace internal {
 
-template<typename Derived1, typename Derived2, int UnrollCount, bool ClearOpposite>
-struct triangular_assignment_selector<Derived1, Derived2, (SelfAdjoint|Upper), UnrollCount, ClearOpposite>
+// TODO currently a selfadjoint expression has the form SelfAdjointView<.,.>
+//      in the future selfadjoint-ness should be defined by the expression traits
+//      such that Transpose<SelfAdjointView<.,.> > is valid. (currently TriangularBase::transpose() is overloaded to make it work)
+template<typename MatrixType, unsigned int Mode>
+struct evaluator_traits<SelfAdjointView<MatrixType,Mode> >
 {
-  enum {
-    col = (UnrollCount-1) / Derived1::RowsAtCompileTime,
-    row = (UnrollCount-1) % Derived1::RowsAtCompileTime
-  };
+  typedef typename storage_kind_to_evaluator_kind<typename MatrixType::StorageKind>::Kind Kind;
+  typedef SelfAdjointShape Shape;
+  
+  static const int AssumeAliasing = 0;
+};
 
-  static inline void run(Derived1 &dst, const Derived2 &src)
+template<int UpLo, int SetOpposite, typename DstEvaluatorTypeT, typename SrcEvaluatorTypeT, typename Functor, int Version>
+class triangular_dense_assignment_kernel<UpLo,SelfAdjoint,SetOpposite,DstEvaluatorTypeT,SrcEvaluatorTypeT,Functor,Version>
+  : public generic_dense_assignment_kernel<DstEvaluatorTypeT, SrcEvaluatorTypeT, Functor, Version>
+{
+protected:
+  typedef generic_dense_assignment_kernel<DstEvaluatorTypeT, SrcEvaluatorTypeT, Functor, Version> Base;
+  typedef typename Base::DstXprType DstXprType;
+  typedef typename Base::SrcXprType SrcXprType;
+  using Base::m_dst;
+  using Base::m_src;
+  using Base::m_functor;
+public:
+  
+  typedef typename Base::DstEvaluatorType DstEvaluatorType;
+  typedef typename Base::SrcEvaluatorType SrcEvaluatorType;
+  typedef typename Base::Scalar Scalar;
+  typedef typename Base::AssignmentTraits AssignmentTraits;
+  
+  
+  EIGEN_DEVICE_FUNC triangular_dense_assignment_kernel(DstEvaluatorType &dst, const SrcEvaluatorType &src, const Functor &func, DstXprType& dstExpr)
+    : Base(dst, src, func, dstExpr)
+  {}
+  
+  EIGEN_DEVICE_FUNC void assignCoeff(Index row, Index col)
   {
-    triangular_assignment_selector<Derived1, Derived2, (SelfAdjoint|Upper), UnrollCount-1, ClearOpposite>::run(dst, src);
-
-    if(row == col)
-      dst.coeffRef(row, col) = numext::real(src.coeff(row, col));
-    else if(row < col)
-      dst.coeffRef(col, row) = numext::conj(dst.coeffRef(row, col) = src.coeff(row, col));
+    eigen_internal_assert(row!=col);
+    Scalar tmp = m_src.coeff(row,col);
+    m_functor.assignCoeff(m_dst.coeffRef(row,col), tmp);
+    m_functor.assignCoeff(m_dst.coeffRef(col,row), numext::conj(tmp));
   }
-};
-
-template<typename Derived1, typename Derived2, bool ClearOpposite>
-struct triangular_assignment_selector<Derived1, Derived2, SelfAdjoint|Upper, 0, ClearOpposite>
-{
-  static inline void run(Derived1 &, const Derived2 &) {}
-};
-
-template<typename Derived1, typename Derived2, int UnrollCount, bool ClearOpposite>
-struct triangular_assignment_selector<Derived1, Derived2, (SelfAdjoint|Lower), UnrollCount, ClearOpposite>
-{
-  enum {
-    col = (UnrollCount-1) / Derived1::RowsAtCompileTime,
-    row = (UnrollCount-1) % Derived1::RowsAtCompileTime
-  };
-
-  static inline void run(Derived1 &dst, const Derived2 &src)
+  
+  EIGEN_DEVICE_FUNC void assignDiagonalCoeff(Index id)
   {
-    triangular_assignment_selector<Derived1, Derived2, (SelfAdjoint|Lower), UnrollCount-1, ClearOpposite>::run(dst, src);
-
-    if(row == col)
-      dst.coeffRef(row, col) = numext::real(src.coeff(row, col));
-    else if(row > col)
-      dst.coeffRef(col, row) = numext::conj(dst.coeffRef(row, col) = src.coeff(row, col));
-  }
-};
-
-template<typename Derived1, typename Derived2, bool ClearOpposite>
-struct triangular_assignment_selector<Derived1, Derived2, SelfAdjoint|Lower, 0, ClearOpposite>
-{
-  static inline void run(Derived1 &, const Derived2 &) {}
-};
-
-template<typename Derived1, typename Derived2, bool ClearOpposite>
-struct triangular_assignment_selector<Derived1, Derived2, SelfAdjoint|Upper, Dynamic, ClearOpposite>
-{
-  typedef typename Derived1::Index Index;
-  static inline void run(Derived1 &dst, const Derived2 &src)
-  {
-    for(Index j = 0; j < dst.cols(); ++j)
-    {
-      for(Index i = 0; i < j; ++i)
-      {
-        dst.copyCoeff(i, j, src);
-        dst.coeffRef(j,i) = numext::conj(dst.coeff(i,j));
-      }
-      dst.copyCoeff(j, j, src);
-    }
-  }
-};
-
-template<typename Derived1, typename Derived2, bool ClearOpposite>
-struct triangular_assignment_selector<Derived1, Derived2, SelfAdjoint|Lower, Dynamic, ClearOpposite>
-{
-  static inline void run(Derived1 &dst, const Derived2 &src)
-  {
-  typedef typename Derived1::Index Index;
-    for(Index i = 0; i < dst.rows(); ++i)
-    {
-      for(Index j = 0; j < i; ++j)
-      {
-        dst.copyCoeff(i, j, src);
-        dst.coeffRef(j,i) = numext::conj(dst.coeff(i,j));
-      }
-      dst.copyCoeff(i, i, src);
-    }
+    Base::assignCoeff(id,id);
   }
+  
+  EIGEN_DEVICE_FUNC void assignOppositeCoeff(Index, Index)
+  { eigen_internal_assert(false && "should never be called"); }
 };
 
 } // end namespace internal
@@ -298,7 +258,7 @@ template<unsigned int UpLo>
 typename MatrixBase<Derived>::template ConstSelfAdjointViewReturnType<UpLo>::Type
 MatrixBase<Derived>::selfadjointView() const
 {
-  return derived();
+  return typename ConstSelfAdjointViewReturnType<UpLo>::Type(derived());
 }
 
 template<typename Derived>
@@ -306,7 +266,7 @@ template<unsigned int UpLo>
 typename MatrixBase<Derived>::template SelfAdjointViewReturnType<UpLo>::Type
 MatrixBase<Derived>::selfadjointView()
 {
-  return derived();
+  return typename SelfAdjointViewReturnType<UpLo>::Type(derived());
 }
 
 } // end namespace Eigen

Eigen/src/Core/SelfCwiseBinaryOp.h

@@ -12,183 +12,35 @@
 
 namespace Eigen { 
 
-/** \class SelfCwiseBinaryOp
-  * \ingroup Core_Module
-  *
-  * \internal
-  *
-  * \brief Internal helper class for optimizing operators like +=, -=
-  *
-  * This is a pseudo expression class re-implementing the copyCoeff/copyPacket
-  * method to directly performs a +=/-= operations in an optimal way. In particular,
-  * this allows to make sure that the input/output data are loaded only once using
-  * aligned packet loads.
-  *
-  * \sa class SwapWrapper for a similar trick.
-  */
-
-namespace internal {
-template<typename BinaryOp, typename Lhs, typename Rhs>
-struct traits<SelfCwiseBinaryOp<BinaryOp,Lhs,Rhs> >
-  : traits<CwiseBinaryOp<BinaryOp,Lhs,Rhs> >
-{
-  enum {
-    // Note that it is still a good idea to preserve the DirectAccessBit
-    // so that assign can correctly align the data.
-    Flags = traits<CwiseBinaryOp<BinaryOp,Lhs,Rhs> >::Flags | (Lhs::Flags&DirectAccessBit) | (Lhs::Flags&LvalueBit),
-    OuterStrideAtCompileTime = Lhs::OuterStrideAtCompileTime,
-    InnerStrideAtCompileTime = Lhs::InnerStrideAtCompileTime
-  };
-};
-}
-
-template<typename BinaryOp, typename Lhs, typename Rhs> class SelfCwiseBinaryOp
-  : public internal::dense_xpr_base< SelfCwiseBinaryOp<BinaryOp, Lhs, Rhs> >::type
-{
-  public:
-
-    typedef typename internal::dense_xpr_base<SelfCwiseBinaryOp>::type Base;
-    EIGEN_DENSE_PUBLIC_INTERFACE(SelfCwiseBinaryOp)
-
-    typedef typename internal::packet_traits<Scalar>::type Packet;
-
-    inline SelfCwiseBinaryOp(Lhs& xpr, const BinaryOp& func = BinaryOp()) : m_matrix(xpr), m_functor(func) {}
-
-    inline Index rows() const { return m_matrix.rows(); }
-    inline Index cols() const { return m_matrix.cols(); }
-    inline Index outerStride() const { return m_matrix.outerStride(); }
-    inline Index innerStride() const { return m_matrix.innerStride(); }
-    inline const Scalar* data() const { return m_matrix.data(); }
-
-    // note that this function is needed by assign to correctly align loads/stores
-    // TODO make Assign use .data()
-    inline Scalar& coeffRef(Index row, Index col)
-    {
-      EIGEN_STATIC_ASSERT_LVALUE(Lhs)
-      return m_matrix.const_cast_derived().coeffRef(row, col);
-    }
-    inline const Scalar& coeffRef(Index row, Index col) const
-    {
-      return m_matrix.coeffRef(row, col);
-    }
-
-    // note that this function is needed by assign to correctly align loads/stores
-    // TODO make Assign use .data()
-    inline Scalar& coeffRef(Index index)
-    {
-      EIGEN_STATIC_ASSERT_LVALUE(Lhs)
-      return m_matrix.const_cast_derived().coeffRef(index);
-    }
-    inline const Scalar& coeffRef(Index index) const
-    {
-      return m_matrix.const_cast_derived().coeffRef(index);
-    }
-
-    template<typename OtherDerived>
-    void copyCoeff(Index row, Index col, const DenseBase<OtherDerived>& other)
-    {
-      OtherDerived& _other = other.const_cast_derived();
-      eigen_internal_assert(row >= 0 && row < rows()
-                         && col >= 0 && col < cols());
-      Scalar& tmp = m_matrix.coeffRef(row,col);
-      tmp = m_functor(tmp, _other.coeff(row,col));
-    }
-
-    template<typename OtherDerived>
-    void copyCoeff(Index index, const DenseBase<OtherDerived>& other)
-    {
-      OtherDerived& _other = other.const_cast_derived();
-      eigen_internal_assert(index >= 0 && index < m_matrix.size());
-      Scalar& tmp = m_matrix.coeffRef(index);
-      tmp = m_functor(tmp, _other.coeff(index));
-    }
-
-    template<typename OtherDerived, int StoreMode, int LoadMode>
-    void copyPacket(Index row, Index col, const DenseBase<OtherDerived>& other)
-    {
-      OtherDerived& _other = other.const_cast_derived();
-      eigen_internal_assert(row >= 0 && row < rows()
-                        && col >= 0 && col < cols());
-      m_matrix.template writePacket<StoreMode>(row, col,
-        m_functor.packetOp(m_matrix.template packet<StoreMode>(row, col),_other.template packet<LoadMode>(row, col)) );
-    }
-
-    template<typename OtherDerived, int StoreMode, int LoadMode>
-    void copyPacket(Index index, const DenseBase<OtherDerived>& other)
-    {
-      OtherDerived& _other = other.const_cast_derived();
-      eigen_internal_assert(index >= 0 && index < m_matrix.size());
-      m_matrix.template writePacket<StoreMode>(index,
-        m_functor.packetOp(m_matrix.template packet<StoreMode>(index),_other.template packet<LoadMode>(index)) );
-    }
-
-    // reimplement lazyAssign to handle complex *= real
-    // see CwiseBinaryOp ctor for details
-    template<typename RhsDerived>
-    EIGEN_STRONG_INLINE SelfCwiseBinaryOp& lazyAssign(const DenseBase<RhsDerived>& rhs)
-    {
-      EIGEN_STATIC_ASSERT_SAME_MATRIX_SIZE(Lhs,RhsDerived)
-      EIGEN_CHECK_BINARY_COMPATIBILIY(BinaryOp,typename Lhs::Scalar,typename RhsDerived::Scalar);
-      
-    #ifdef EIGEN_DEBUG_ASSIGN
-      internal::assign_traits<SelfCwiseBinaryOp, RhsDerived>::debug();
-    #endif
-      eigen_assert(rows() == rhs.rows() && cols() == rhs.cols());
-      internal::assign_impl<SelfCwiseBinaryOp, RhsDerived>::run(*this,rhs.derived());
-    #ifndef EIGEN_NO_DEBUG
-      this->checkTransposeAliasing(rhs.derived());
-    #endif
-      return *this;
-    }
-    
-    // overloaded to honor evaluation of special matrices
-    // maybe another solution would be to not use SelfCwiseBinaryOp
-    // at first...
-    SelfCwiseBinaryOp& operator=(const Rhs& _rhs)
-    {
-      typename internal::nested<Rhs>::type rhs(_rhs);
-      return Base::operator=(rhs);
-    }
-
-    Lhs& expression() const 
-    { 
-      return m_matrix;
-    }
-
-    const BinaryOp& functor() const 
-    { 
-      return m_functor;
-    }
-
-  protected:
-    Lhs& m_matrix;
-    const BinaryOp& m_functor;
-
-  private:
-    SelfCwiseBinaryOp& operator=(const SelfCwiseBinaryOp&);
-};
-
 template<typename Derived>
 inline Derived& DenseBase<Derived>::operator*=(const Scalar& other)
 {
   typedef typename Derived::PlainObject PlainObject;
-  SelfCwiseBinaryOp<internal::scalar_product_op<Scalar>, Derived, typename PlainObject::ConstantReturnType> tmp(derived());
-  tmp = PlainObject::Constant(rows(),cols(),other);
+  internal::call_assignment(this->derived(), PlainObject::Constant(rows(),cols(),other), internal::mul_assign_op<Scalar>());
+  return derived();
+}
+
+template<typename Derived>
+inline Derived& ArrayBase<Derived>::operator+=(const Scalar& other)
+{
+  typedef typename Derived::PlainObject PlainObject;
+  internal::call_assignment(this->derived(), PlainObject::Constant(rows(),cols(),other), internal::add_assign_op<Scalar>());
+  return derived();
+}
+
+template<typename Derived>
+inline Derived& ArrayBase<Derived>::operator-=(const Scalar& other)
+{
+  typedef typename Derived::PlainObject PlainObject;
+  internal::call_assignment(this->derived(), PlainObject::Constant(rows(),cols(),other), internal::sub_assign_op<Scalar>());
   return derived();
 }
 
 template<typename Derived>
 inline Derived& DenseBase<Derived>::operator/=(const Scalar& other)
 {
-  typedef typename internal::conditional<NumTraits<Scalar>::IsInteger,
-                                        internal::scalar_quotient_op<Scalar>,
-                                        internal::scalar_product_op<Scalar> >::type BinOp;
   typedef typename Derived::PlainObject PlainObject;
-  SelfCwiseBinaryOp<BinOp, Derived, typename PlainObject::ConstantReturnType> tmp(derived());
-  Scalar actual_other;
-  if(NumTraits<Scalar>::IsInteger)  actual_other = other;
-  else                              actual_other = Scalar(1)/other;
-  tmp = PlainObject::Constant(rows(),cols(), actual_other);
+  internal::call_assignment(this->derived(), PlainObject::Constant(rows(),cols(),other), internal::div_assign_op<Scalar>());
   return derived();
 }
 

Eigen/src/Core/Solve.h

@@ -0,0 +1,150 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SOLVE_H
+#define EIGEN_SOLVE_H
+
+namespace Eigen {
+
+template<typename Decomposition, typename RhsType, typename StorageKind> class SolveImpl;
+  
+/** \class Solve
+  * \ingroup Core_Module
+  *
+  * \brief Pseudo expression representing a solving operation
+  *
+  * \tparam Decomposition the type of the matrix or decomposion object
+  * \tparam Rhstype the type of the right-hand side
+  *
+  * This class represents an expression of A.solve(B)
+  * and most of the time this is the only way it is used.
+  *
+  */
+namespace internal {
+
+// this solve_traits class permits to determine the evaluation type with respect to storage kind (Dense vs Sparse)
+template<typename Decomposition, typename RhsType,typename StorageKind> struct solve_traits;
+
+template<typename Decomposition, typename RhsType>
+struct solve_traits<Decomposition,RhsType,Dense>
+{
+  typedef typename Decomposition::MatrixType MatrixType;
+  typedef Matrix<typename RhsType::Scalar,
+                 MatrixType::ColsAtCompileTime,
+                 RhsType::ColsAtCompileTime,
+                 RhsType::PlainObject::Options,
+                 MatrixType::MaxColsAtCompileTime,
+                 RhsType::MaxColsAtCompileTime> PlainObject;  
+};
+
+template<typename Decomposition, typename RhsType>
+struct traits<Solve<Decomposition, RhsType> >
+  : traits<typename solve_traits<Decomposition,RhsType,typename internal::traits<RhsType>::StorageKind>::PlainObject>
+{
+  typedef typename solve_traits<Decomposition,RhsType,typename internal::traits<RhsType>::StorageKind>::PlainObject PlainObject;
+  typedef typename promote_index_type<typename Decomposition::StorageIndex, typename RhsType::StorageIndex>::type StorageIndex;
+  typedef traits<PlainObject> BaseTraits;
+  enum {
+    Flags = BaseTraits::Flags & RowMajorBit,
+    CoeffReadCost = Dynamic
+  };
+};
+
+}
+
+
+template<typename Decomposition, typename RhsType>
+class Solve : public SolveImpl<Decomposition,RhsType,typename internal::traits<RhsType>::StorageKind>
+{
+public:
+  typedef typename internal::traits<Solve>::PlainObject PlainObject;
+  typedef typename internal::traits<Solve>::StorageIndex StorageIndex;
+  
+  Solve(const Decomposition &dec, const RhsType &rhs)
+    : m_dec(dec), m_rhs(rhs)
+  {}
+  
+  EIGEN_DEVICE_FUNC Index rows() const { return m_dec.cols(); }
+  EIGEN_DEVICE_FUNC Index cols() const { return m_rhs.cols(); }
+
+  EIGEN_DEVICE_FUNC const Decomposition& dec() const { return m_dec; }
+  EIGEN_DEVICE_FUNC const RhsType&       rhs() const { return m_rhs; }
+
+protected:
+  const Decomposition &m_dec;
+  const RhsType       &m_rhs;
+};
+
+
+// Specialization of the Solve expression for dense results
+template<typename Decomposition, typename RhsType>
+class SolveImpl<Decomposition,RhsType,Dense>
+  : public MatrixBase<Solve<Decomposition,RhsType> >
+{
+  typedef Solve<Decomposition,RhsType> Derived;
+  
+public:
+  
+  typedef MatrixBase<Solve<Decomposition,RhsType> > Base;
+  EIGEN_DENSE_PUBLIC_INTERFACE(Derived)
+
+private:
+  
+  Scalar coeff(Index row, Index col) const;
+  Scalar coeff(Index i) const;
+};
+
+// Generic API dispatcher
+template<typename Decomposition, typename RhsType, typename StorageKind>
+class SolveImpl : public internal::generic_xpr_base<Solve<Decomposition,RhsType>, MatrixXpr, StorageKind>::type
+{
+  public:
+    typedef typename internal::generic_xpr_base<Solve<Decomposition,RhsType>, MatrixXpr, StorageKind>::type Base;
+};
+
+namespace internal {
+
+// Evaluator of Solve -> eval into a temporary
+template<typename Decomposition, typename RhsType>
+struct evaluator<Solve<Decomposition,RhsType> >
+  : public evaluator<typename Solve<Decomposition,RhsType>::PlainObject>
+{
+  typedef Solve<Decomposition,RhsType> SolveType;
+  typedef typename SolveType::PlainObject PlainObject;
+  typedef evaluator<PlainObject> Base;
+  
+  EIGEN_DEVICE_FUNC explicit evaluator(const SolveType& solve)
+    : m_result(solve.rows(), solve.cols())
+  {
+    ::new (static_cast<Base*>(this)) Base(m_result);
+    solve.dec()._solve_impl(solve.rhs(), m_result);
+  }
+  
+protected:  
+  PlainObject m_result;
+};
+
+// Specialization for "dst = dec.solve(rhs)"
+// NOTE we need to specialize it for Dense2Dense to avoid ambiguous specialization error and a Sparse2Sparse specialization must exist somewhere
+template<typename DstXprType, typename DecType, typename RhsType, typename Scalar>
+struct Assignment<DstXprType, Solve<DecType,RhsType>, internal::assign_op<Scalar>, Dense2Dense, Scalar>
+{
+  typedef Solve<DecType,RhsType> SrcXprType;
+  static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<Scalar> &)
+  {
+    // FIXME shall we resize dst here?
+    src.dec()._solve_impl(src.rhs(), dst);
+  }
+};
+
+} // end namepsace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_SOLVE_H

Eigen/src/Core/SolveTriangular.h

@@ -68,7 +68,7 @@ struct triangular_solver_selector<Lhs,Rhs,Side,Mode,NoUnrolling,1>
     if(!useRhsDirectly)
       MappedRhs(actualRhs,rhs.size()) = rhs;
 
-    triangular_solve_vector<LhsScalar, RhsScalar, typename Lhs::Index, Side, Mode, LhsProductTraits::NeedToConjugate,
+    triangular_solve_vector<LhsScalar, RhsScalar, Index, Side, Mode, LhsProductTraits::NeedToConjugate,
                             (int(Lhs::Flags) & RowMajorBit) ? RowMajor : ColMajor>
       ::run(actualLhs.cols(), actualLhs.data(), actualLhs.outerStride(), actualRhs);
 
@@ -82,7 +82,6 @@ template<typename Lhs, typename Rhs, int Side, int Mode>
 struct triangular_solver_selector<Lhs,Rhs,Side,Mode,NoUnrolling,Dynamic>
 {
   typedef typename Rhs::Scalar Scalar;
-  typedef typename Rhs::Index Index;
   typedef blas_traits<Lhs> LhsProductTraits;
   typedef typename LhsProductTraits::DirectLinearAccessType ActualLhsType;
 
@@ -96,7 +95,7 @@ struct triangular_solver_selector<Lhs,Rhs,Side,Mode,NoUnrolling,Dynamic>
     typedef internal::gemm_blocking_space<(Rhs::Flags&RowMajorBit) ? RowMajor : ColMajor,Scalar,Scalar,
               Rhs::MaxRowsAtCompileTime, Rhs::MaxColsAtCompileTime, Lhs::MaxRowsAtCompileTime,4> BlockingType;
 
-    BlockingType blocking(rhs.rows(), rhs.cols(), size);
+    BlockingType blocking(rhs.rows(), rhs.cols(), size, 1, false);
 
     triangular_solve_matrix<Scalar,Index,Side,Mode,LhsProductTraits::NeedToConjugate,(int(Lhs::Flags) & RowMajorBit) ? RowMajor : ColMajor,
                                (Rhs::Flags&RowMajorBit) ? RowMajor : ColMajor>
@@ -171,10 +170,10 @@ struct triangular_solver_selector<Lhs,Rhs,OnTheRight,Mode,CompleteUnrolling,1> {
   */
 template<typename MatrixType, unsigned int Mode>
 template<int Side, typename OtherDerived>
-void TriangularView<MatrixType,Mode>::solveInPlace(const MatrixBase<OtherDerived>& _other) const
+void TriangularViewImpl<MatrixType,Mode,Dense>::solveInPlace(const MatrixBase<OtherDerived>& _other) const
 {
   OtherDerived& other = _other.const_cast_derived();
-  eigen_assert( cols() == rows() && ((Side==OnTheLeft && cols() == other.rows()) || (Side==OnTheRight && cols() == other.cols())) );
+  eigen_assert( derived().cols() == derived().rows() && ((Side==OnTheLeft && derived().cols() == other.rows()) || (Side==OnTheRight && derived().cols() == other.cols())) );
   eigen_assert((!(Mode & ZeroDiag)) && bool(Mode & (Upper|Lower)));
 
   enum { copy = internal::traits<OtherDerived>::Flags & RowMajorBit  && OtherDerived::IsVectorAtCompileTime };
@@ -183,7 +182,7 @@ void TriangularView<MatrixType,Mode>::solveInPlace(const MatrixBase<OtherDerived
   OtherCopy otherCopy(other);
 
   internal::triangular_solver_selector<MatrixType, typename internal::remove_reference<OtherCopy>::type,
-    Side, Mode>::run(nestedExpression(), otherCopy);
+    Side, Mode>::run(derived().nestedExpression(), otherCopy);
 
   if (copy)
     other = otherCopy;
@@ -199,8 +198,8 @@ void TriangularView<MatrixType,Mode>::solveInPlace(const MatrixBase<OtherDerived
   * diagonal must be non zero). It works as a forward (resp. backward) substitution if \c *this
   * is an upper (resp. lower) triangular matrix.
   *
-  * Example: \include MatrixBase_marked.cpp
-  * Output: \verbinclude MatrixBase_marked.out
+  * Example: \include Triangular_solve.cpp
+  * Output: \verbinclude Triangular_solve.out
   *
   * This function returns an expression of the inverse-multiply and can works in-place if it is assigned
   * to the same matrix or vector \a other.
@@ -213,9 +212,9 @@ void TriangularView<MatrixType,Mode>::solveInPlace(const MatrixBase<OtherDerived
 template<typename Derived, unsigned int Mode>
 template<int Side, typename Other>
 const internal::triangular_solve_retval<Side,TriangularView<Derived,Mode>,Other>
-TriangularView<Derived,Mode>::solve(const MatrixBase<Other>& other) const
+TriangularViewImpl<Derived,Mode,Dense>::solve(const MatrixBase<Other>& other) const
 {
-  return internal::triangular_solve_retval<Side,TriangularView,Other>(*this, other.derived());
+  return internal::triangular_solve_retval<Side,TriangularViewType,Other>(derived(), other.derived());
 }
 
 namespace internal {
@@ -232,7 +231,6 @@ template<int Side, typename TriangularType, typename Rhs> struct triangular_solv
 {
   typedef typename remove_all<typename Rhs::Nested>::type RhsNestedCleaned;
   typedef ReturnByValue<triangular_solve_retval> Base;
-  typedef typename Base::Index Index;
 
   triangular_solve_retval(const TriangularType& tri, const Rhs& rhs)
     : m_triangularMatrix(tri), m_rhs(rhs)

Eigen/src/Core/StableNorm.h

@@ -17,16 +17,37 @@ namespace internal {
 template<typename ExpressionType, typename Scalar>
 inline void stable_norm_kernel(const ExpressionType& bl, Scalar& ssq, Scalar& scale, Scalar& invScale)
 {
-  Scalar max = bl.cwiseAbs().maxCoeff();
-  if (max>scale)
+  Scalar maxCoeff = bl.cwiseAbs().maxCoeff();
+  
+  if(maxCoeff>scale)
   {
-    ssq = ssq * numext::abs2(scale/max);
-    scale = max;
-    invScale = Scalar(1)/scale;
+    ssq = ssq * numext::abs2(scale/maxCoeff);
+    Scalar tmp = Scalar(1)/maxCoeff;
+    if(tmp > NumTraits<Scalar>::highest())
+    {
+      invScale = NumTraits<Scalar>::highest();
+      scale = Scalar(1)/invScale;
+    }
+    else if(maxCoeff>NumTraits<Scalar>::highest()) // we got a INF
+    {
+      invScale = Scalar(1);
+      scale = maxCoeff;
+    }
+    else
+    {
+      scale = maxCoeff;
+      invScale = tmp;
+    }
   }
-  // TODO if the max is much much smaller than the current scale,
+  else if(maxCoeff!=maxCoeff) // we got a NaN
+  {
+    scale = maxCoeff;
+  }
+  
+  // TODO if the maxCoeff is much much smaller than the current scale,
   // then we can neglect this sub vector
-  ssq += (bl*invScale).squaredNorm();
+  if(scale>Scalar(0)) // if scale==0, then bl is 0 
+    ssq += (bl*invScale).squaredNorm();
 }
 
 template<typename Derived>
@@ -34,15 +55,12 @@ inline typename NumTraits<typename traits<Derived>::Scalar>::Real
 blueNorm_impl(const EigenBase<Derived>& _vec)
 {
   typedef typename Derived::RealScalar RealScalar;  
-  typedef typename Derived::Index Index;
   using std::pow;
-  using std::min;
-  using std::max;
   using std::sqrt;
   using std::abs;
   const Derived& vec(_vec.derived());
   static bool initialized = false;
-  static RealScalar b1, b2, s1m, s2m, overfl, rbig, relerr;
+  static RealScalar b1, b2, s1m, s2m, rbig, relerr;
   if(!initialized)
   {
     int ibeta, it, iemin, iemax, iexp;
@@ -71,7 +89,6 @@ blueNorm_impl(const EigenBase<Derived>& _vec)
     iexp  = - ((iemax+it)/2);
     s2m   = RealScalar(pow(RealScalar(ibeta),RealScalar(iexp)));    // scaling factor for upper range
 
-    overfl  = rbig*s2m;                                             // overflow boundary for abig
     eps     = RealScalar(pow(double(ibeta), 1-it));
     relerr  = sqrt(eps);                                            // tolerance for neglecting asml
     initialized = true;
@@ -88,13 +105,13 @@ blueNorm_impl(const EigenBase<Derived>& _vec)
     else if(ax < b1) asml += numext::abs2(ax*s1m);
     else             amed += numext::abs2(ax);
   }
+  if(amed!=amed)
+    return amed;  // we got a NaN
   if(abig > RealScalar(0))
   {
     abig = sqrt(abig);
-    if(abig > overfl)
-    {
-      return rbig;
-    }
+    if(abig > rbig) // overflow, or *this contains INF values
+      return abig;  // return INF
     if(amed > RealScalar(0))
     {
       abig = abig/s2m;
@@ -115,8 +132,8 @@ blueNorm_impl(const EigenBase<Derived>& _vec)
   }
   else
     return sqrt(amed);
-  asml = (min)(abig, amed);
-  abig = (max)(abig, amed);
+  asml = numext::mini(abig, amed);
+  abig = numext::maxi(abig, amed);
   if(asml <= abig*relerr)
     return abig;
   else
@@ -139,21 +156,33 @@ template<typename Derived>
 inline typename NumTraits<typename internal::traits<Derived>::Scalar>::Real
 MatrixBase<Derived>::stableNorm() const
 {
-  using std::min;
   using std::sqrt;
+  using std::abs;
   const Index blockSize = 4096;
   RealScalar scale(0);
   RealScalar invScale(1);
   RealScalar ssq(0); // sum of square
+  
+  typedef typename internal::nested_eval<Derived,2>::type DerivedCopy;
+  typedef typename internal::remove_all<DerivedCopy>::type DerivedCopyClean;
+  DerivedCopy copy(derived());
+  
   enum {
-    Alignment = (int(Flags)&DirectAccessBit) || (int(Flags)&AlignedBit) ? 1 : 0
+    CanAlign = (int(Flags)&DirectAccessBit) || (int(internal::evaluator<DerivedCopyClean>::Alignment)>0) // FIXME
   };
+  typedef typename internal::conditional<CanAlign, Ref<const Matrix<Scalar,Dynamic,1,0,blockSize,1>, internal::evaluator<DerivedCopyClean>::Alignment>,
+                                                   typename DerivedCopyClean
+                                                   ::ConstSegmentReturnType>::type SegmentWrapper;
   Index n = size();
-  Index bi = internal::first_aligned(derived());
+  
+  if(n==1)
+    return abs(this->coeff(0));
+  
+  Index bi = internal::first_default_aligned(copy);
   if (bi>0)
-    internal::stable_norm_kernel(this->head(bi), ssq, scale, invScale);
+    internal::stable_norm_kernel(copy.head(bi), ssq, scale, invScale);
   for (; bi<n; bi+=blockSize)
-    internal::stable_norm_kernel(this->segment(bi,(min)(blockSize, n - bi)).template forceAlignedAccessIf<Alignment>(), ssq, scale, invScale);
+    internal::stable_norm_kernel(SegmentWrapper(copy.segment(bi,numext::mini(blockSize, n - bi))), ssq, scale, invScale);
   return scale * sqrt(ssq);
 }
 

Eigen/src/Core/Stride.h

@@ -44,13 +44,14 @@ template<int _OuterStrideAtCompileTime, int _InnerStrideAtCompileTime>
 class Stride
 {
   public:
-    typedef DenseIndex Index;
+    typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3
     enum {
       InnerStrideAtCompileTime = _InnerStrideAtCompileTime,
       OuterStrideAtCompileTime = _OuterStrideAtCompileTime
     };
 
     /** Default constructor, for use when strides are fixed at compile time */
+    EIGEN_DEVICE_FUNC
     Stride()
       : m_outer(OuterStrideAtCompileTime), m_inner(InnerStrideAtCompileTime)
     {
@@ -58,6 +59,7 @@ class Stride
     }
 
     /** Constructor allowing to pass the strides at runtime */
+    EIGEN_DEVICE_FUNC
     Stride(Index outerStride, Index innerStride)
       : m_outer(outerStride), m_inner(innerStride)
     {
@@ -65,13 +67,16 @@ class Stride
     }
 
     /** Copy constructor */
+    EIGEN_DEVICE_FUNC
     Stride(const Stride& other)
       : m_outer(other.outer()), m_inner(other.inner())
     {}
 
     /** \returns the outer stride */
+    EIGEN_DEVICE_FUNC
     inline Index outer() const { return m_outer.value(); }
     /** \returns the inner stride */
+    EIGEN_DEVICE_FUNC
     inline Index inner() const { return m_inner.value(); }
 
   protected:
@@ -81,26 +86,24 @@ class Stride
 
 /** \brief Convenience specialization of Stride to specify only an inner stride
   * See class Map for some examples */
-template<int Value = Dynamic>
+template<int Value>
 class InnerStride : public Stride<0, Value>
 {
     typedef Stride<0, Value> Base;
   public:
-    typedef DenseIndex Index;
-    InnerStride() : Base() {}
-    InnerStride(Index v) : Base(0, v) {}
+    EIGEN_DEVICE_FUNC InnerStride() : Base() {}
+    EIGEN_DEVICE_FUNC InnerStride(Index v) : Base(0, v) {} // FIXME making this explicit could break valid code
 };
 
 /** \brief Convenience specialization of Stride to specify only an outer stride
   * See class Map for some examples */
-template<int Value = Dynamic>
+template<int Value>
 class OuterStride : public Stride<Value, 0>
 {
     typedef Stride<Value, 0> Base;
   public:
-    typedef DenseIndex Index;
-    OuterStride() : Base() {}
-    OuterStride(Index v) : Base(v,0) {}
+    EIGEN_DEVICE_FUNC OuterStride() : Base() {}
+    EIGEN_DEVICE_FUNC OuterStride(Index v) : Base(v,0) {} // FIXME making this explicit could break valid code
 };
 
 } // end namespace Eigen

Eigen/src/Core/Swap.h

@@ -12,115 +12,56 @@
 
 namespace Eigen { 
 
-/** \class SwapWrapper
-  * \ingroup Core_Module
-  *
-  * \internal
-  *
-  * \brief Internal helper class for swapping two expressions
-  */
 namespace internal {
-template<typename ExpressionType>
-struct traits<SwapWrapper<ExpressionType> > : traits<ExpressionType> {};
-}
 
-template<typename ExpressionType> class SwapWrapper
-  : public internal::dense_xpr_base<SwapWrapper<ExpressionType> >::type
+// Overload default assignPacket behavior for swapping them
+template<typename DstEvaluatorTypeT, typename SrcEvaluatorTypeT>
+class generic_dense_assignment_kernel<DstEvaluatorTypeT, SrcEvaluatorTypeT, swap_assign_op<typename DstEvaluatorTypeT::Scalar>, Specialized>
+ : public generic_dense_assignment_kernel<DstEvaluatorTypeT, SrcEvaluatorTypeT, swap_assign_op<typename DstEvaluatorTypeT::Scalar>, BuiltIn>
 {
-  public:
-
-    typedef typename internal::dense_xpr_base<SwapWrapper>::type Base;
-    EIGEN_DENSE_PUBLIC_INTERFACE(SwapWrapper)
-    typedef typename internal::packet_traits<Scalar>::type Packet;
-
-    inline SwapWrapper(ExpressionType& xpr) : m_expression(xpr) {}
-
-    inline Index rows() const { return m_expression.rows(); }
-    inline Index cols() const { return m_expression.cols(); }
-    inline Index outerStride() const { return m_expression.outerStride(); }
-    inline Index innerStride() const { return m_expression.innerStride(); }
-    
-    typedef typename internal::conditional<
-                       internal::is_lvalue<ExpressionType>::value,
-                       Scalar,
-                       const Scalar
-                     >::type ScalarWithConstIfNotLvalue;
-                     
-    inline ScalarWithConstIfNotLvalue* data() { return m_expression.data(); }
-    inline const Scalar* data() const { return m_expression.data(); }
-
-    inline Scalar& coeffRef(Index rowId, Index colId)
-    {
-      return m_expression.const_cast_derived().coeffRef(rowId, colId);
-    }
-
-    inline Scalar& coeffRef(Index index)
-    {
-      return m_expression.const_cast_derived().coeffRef(index);
-    }
-
-    inline Scalar& coeffRef(Index rowId, Index colId) const
-    {
-      return m_expression.coeffRef(rowId, colId);
-    }
-
-    inline Scalar& coeffRef(Index index) const
-    {
-      return m_expression.coeffRef(index);
-    }
-
-    template<typename OtherDerived>
-    void copyCoeff(Index rowId, Index colId, const DenseBase<OtherDerived>& other)
-    {
-      OtherDerived& _other = other.const_cast_derived();
-      eigen_internal_assert(rowId >= 0 && rowId < rows()
-                         && colId >= 0 && colId < cols());
-      Scalar tmp = m_expression.coeff(rowId, colId);
-      m_expression.coeffRef(rowId, colId) = _other.coeff(rowId, colId);
-      _other.coeffRef(rowId, colId) = tmp;
-    }
-
-    template<typename OtherDerived>
-    void copyCoeff(Index index, const DenseBase<OtherDerived>& other)
-    {
-      OtherDerived& _other = other.const_cast_derived();
-      eigen_internal_assert(index >= 0 && index < m_expression.size());
-      Scalar tmp = m_expression.coeff(index);
-      m_expression.coeffRef(index) = _other.coeff(index);
-      _other.coeffRef(index) = tmp;
-    }
-
-    template<typename OtherDerived, int StoreMode, int LoadMode>
-    void copyPacket(Index rowId, Index colId, const DenseBase<OtherDerived>& other)
-    {
-      OtherDerived& _other = other.const_cast_derived();
-      eigen_internal_assert(rowId >= 0 && rowId < rows()
-                        && colId >= 0 && colId < cols());
-      Packet tmp = m_expression.template packet<StoreMode>(rowId, colId);
-      m_expression.template writePacket<StoreMode>(rowId, colId,
-        _other.template packet<LoadMode>(rowId, colId)
-      );
-      _other.template writePacket<LoadMode>(rowId, colId, tmp);
-    }
-
-    template<typename OtherDerived, int StoreMode, int LoadMode>
-    void copyPacket(Index index, const DenseBase<OtherDerived>& other)
-    {
-      OtherDerived& _other = other.const_cast_derived();
-      eigen_internal_assert(index >= 0 && index < m_expression.size());
-      Packet tmp = m_expression.template packet<StoreMode>(index);
-      m_expression.template writePacket<StoreMode>(index,
-        _other.template packet<LoadMode>(index)
-      );
-      _other.template writePacket<LoadMode>(index, tmp);
-    }
-
-    ExpressionType& expression() const { return m_expression; }
-
-  protected:
-    ExpressionType& m_expression;
+protected:
+  typedef generic_dense_assignment_kernel<DstEvaluatorTypeT, SrcEvaluatorTypeT, swap_assign_op<typename DstEvaluatorTypeT::Scalar>, BuiltIn> Base;
+  using Base::m_dst;
+  using Base::m_src;
+  using Base::m_functor;
+  
+public:
+  typedef typename Base::Scalar Scalar;
+  typedef typename Base::DstXprType DstXprType;
+  typedef swap_assign_op<Scalar> Functor;
+  
+  EIGEN_DEVICE_FUNC generic_dense_assignment_kernel(DstEvaluatorTypeT &dst, const SrcEvaluatorTypeT &src, const Functor &func, DstXprType& dstExpr)
+    : Base(dst, src, func, dstExpr)
+  {}
+  
+  template<int StoreMode, int LoadMode, typename PacketType>
+  void assignPacket(Index row, Index col)
+  {
+    PacketType tmp = m_src.template packet<LoadMode,PacketType>(row,col);
+    const_cast<SrcEvaluatorTypeT&>(m_src).template writePacket<LoadMode>(row,col, m_dst.template packet<StoreMode,PacketType>(row,col));
+    m_dst.template writePacket<StoreMode>(row,col,tmp);
+  }
+  
+  template<int StoreMode, int LoadMode, typename PacketType>
+  void assignPacket(Index index)
+  {
+    PacketType tmp = m_src.template packet<LoadMode,PacketType>(index);
+    const_cast<SrcEvaluatorTypeT&>(m_src).template writePacket<LoadMode>(index, m_dst.template packet<StoreMode,PacketType>(index));
+    m_dst.template writePacket<StoreMode>(index,tmp);
+  }
+  
+  // TODO find a simple way not to have to copy/paste this function from generic_dense_assignment_kernel, by simple I mean no CRTP (Gael)
+  template<int StoreMode, int LoadMode, typename PacketType>
+  void assignPacketByOuterInner(Index outer, Index inner)
+  {
+    Index row = Base::rowIndexByOuterInner(outer, inner); 
+    Index col = Base::colIndexByOuterInner(outer, inner);
+    assignPacket<StoreMode,LoadMode,PacketType>(row, col);
+  }
 };
 
+} // namespace internal
+
 } // end namespace Eigen
 
 #endif // EIGEN_SWAP_H

Eigen/src/Core/Transpose.h

@@ -2,7 +2,7 @@
 // for linear algebra.
 //
 // Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
-// Copyright (C) 2009-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2009-2014 Gael Guennebaud <gael.guennebaud@inria.fr>
 //
 // This Source Code Form is subject to the terms of the Mozilla
 // Public License v. 2.0. If a copy of the MPL was not distributed
@@ -29,13 +29,10 @@ namespace Eigen {
 
 namespace internal {
 template<typename MatrixType>
-struct traits<Transpose<MatrixType> > : traits<MatrixType>
+struct traits<Transpose<MatrixType> > : public traits<MatrixType>
 {
-  typedef typename MatrixType::Scalar Scalar;
-  typedef typename nested<MatrixType>::type MatrixTypeNested;
+  typedef typename ref_selector<MatrixType>::type MatrixTypeNested;
   typedef typename remove_reference<MatrixTypeNested>::type MatrixTypeNestedPlain;
-  typedef typename traits<MatrixType>::StorageKind StorageKind;
-  typedef typename traits<MatrixType>::XprKind XprKind;
   enum {
     RowsAtCompileTime = MatrixType::ColsAtCompileTime,
     ColsAtCompileTime = MatrixType::RowsAtCompileTime,
@@ -45,7 +42,6 @@ struct traits<Transpose<MatrixType> > : traits<MatrixType>
     Flags0 = MatrixTypeNestedPlain::Flags & ~(LvalueBit | NestByRefBit),
     Flags1 = Flags0 | FlagsLvalueBit,
     Flags = Flags1 ^ RowMajorBit,
-    CoeffReadCost = MatrixTypeNestedPlain::CoeffReadCost,
     InnerStrideAtCompileTime = inner_stride_at_compile_time<MatrixType>::ret,
     OuterStrideAtCompileTime = outer_stride_at_compile_time<MatrixType>::ret
   };
@@ -61,19 +57,23 @@ template<typename MatrixType> class Transpose
 
     typedef typename TransposeImpl<MatrixType,typename internal::traits<MatrixType>::StorageKind>::Base Base;
     EIGEN_GENERIC_PUBLIC_INTERFACE(Transpose)
+    typedef typename internal::remove_all<MatrixType>::type NestedExpression;
 
-    inline Transpose(MatrixType& a_matrix) : m_matrix(a_matrix) {}
+    EIGEN_DEVICE_FUNC
+    explicit inline Transpose(MatrixType& matrix) : m_matrix(matrix) {}
 
     EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Transpose)
 
-    inline Index rows() const { return m_matrix.cols(); }
-    inline Index cols() const { return m_matrix.rows(); }
+    EIGEN_DEVICE_FUNC inline Index rows() const { return m_matrix.cols(); }
+    EIGEN_DEVICE_FUNC inline Index cols() const { return m_matrix.rows(); }
 
     /** \returns the nested expression */
+    EIGEN_DEVICE_FUNC
     const typename internal::remove_all<typename MatrixType::Nested>::type&
     nestedExpression() const { return m_matrix; }
 
     /** \returns the nested expression */
+    EIGEN_DEVICE_FUNC
     typename internal::remove_all<typename MatrixType::Nested>::type&
     nestedExpression() { return m_matrix.const_cast_derived(); }
 
@@ -97,17 +97,27 @@ struct TransposeImpl_base<MatrixType, false>
 
 } // end namespace internal
 
+// Generic API dispatcher
+template<typename XprType, typename StorageKind>
+class TransposeImpl
+  : public internal::generic_xpr_base<Transpose<XprType> >::type
+{
+public:
+  typedef typename internal::generic_xpr_base<Transpose<XprType> >::type Base;
+};
+
 template<typename MatrixType> class TransposeImpl<MatrixType,Dense>
   : public internal::TransposeImpl_base<MatrixType>::type
 {
   public:
 
     typedef typename internal::TransposeImpl_base<MatrixType>::type Base;
+    using Base::coeffRef;
     EIGEN_DENSE_PUBLIC_INTERFACE(Transpose<MatrixType>)
     EIGEN_INHERIT_ASSIGNMENT_OPERATORS(TransposeImpl)
 
-    inline Index innerStride() const { return derived().nestedExpression().innerStride(); }
-    inline Index outerStride() const { return derived().nestedExpression().outerStride(); }
+    EIGEN_DEVICE_FUNC inline Index innerStride() const { return derived().nestedExpression().innerStride(); }
+    EIGEN_DEVICE_FUNC inline Index outerStride() const { return derived().nestedExpression().outerStride(); }
 
     typedef typename internal::conditional<
                        internal::is_lvalue<MatrixType>::value,
@@ -115,64 +125,21 @@ template<typename MatrixType> class TransposeImpl<MatrixType,Dense>
                        const Scalar
                      >::type ScalarWithConstIfNotLvalue;
 
-    inline ScalarWithConstIfNotLvalue* data() { return derived().nestedExpression().data(); }
-    inline const Scalar* data() const { return derived().nestedExpression().data(); }
-
-    inline ScalarWithConstIfNotLvalue& coeffRef(Index rowId, Index colId)
-    {
-      EIGEN_STATIC_ASSERT_LVALUE(MatrixType)
-      return derived().nestedExpression().const_cast_derived().coeffRef(colId, rowId);
-    }
-
-    inline ScalarWithConstIfNotLvalue& coeffRef(Index index)
-    {
-      EIGEN_STATIC_ASSERT_LVALUE(MatrixType)
-      return derived().nestedExpression().const_cast_derived().coeffRef(index);
-    }
+    EIGEN_DEVICE_FUNC inline ScalarWithConstIfNotLvalue* data() { return derived().nestedExpression().data(); }
+    EIGEN_DEVICE_FUNC inline const Scalar* data() const { return derived().nestedExpression().data(); }
 
+    // FIXME: shall we keep the const version of coeffRef?
+    EIGEN_DEVICE_FUNC
     inline const Scalar& coeffRef(Index rowId, Index colId) const
     {
       return derived().nestedExpression().coeffRef(colId, rowId);
     }
 
+    EIGEN_DEVICE_FUNC
     inline const Scalar& coeffRef(Index index) const
     {
       return derived().nestedExpression().coeffRef(index);
     }
-
-    inline CoeffReturnType coeff(Index rowId, Index colId) const
-    {
-      return derived().nestedExpression().coeff(colId, rowId);
-    }
-
-    inline CoeffReturnType coeff(Index index) const
-    {
-      return derived().nestedExpression().coeff(index);
-    }
-
-    template<int LoadMode>
-    inline const PacketScalar packet(Index rowId, Index colId) const
-    {
-      return derived().nestedExpression().template packet<LoadMode>(colId, rowId);
-    }
-
-    template<int LoadMode>
-    inline void writePacket(Index rowId, Index colId, const PacketScalar& x)
-    {
-      derived().nestedExpression().const_cast_derived().template writePacket<LoadMode>(colId, rowId, x);
-    }
-
-    template<int LoadMode>
-    inline const PacketScalar packet(Index index) const
-    {
-      return derived().nestedExpression().template packet<LoadMode>(index);
-    }
-
-    template<int LoadMode>
-    inline void writePacket(Index index, const PacketScalar& x)
-    {
-      derived().nestedExpression().const_cast_derived().template writePacket<LoadMode>(index, x);
-    }
 };
 
 /** \returns an expression of the transpose of *this.
@@ -198,7 +165,7 @@ template<typename Derived>
 inline Transpose<Derived>
 DenseBase<Derived>::transpose()
 {
-  return derived();
+  return TransposeReturnType(derived());
 }
 
 /** This is the const version of transpose().
@@ -236,8 +203,7 @@ template<typename Derived>
 inline const typename MatrixBase<Derived>::AdjointReturnType
 MatrixBase<Derived>::adjoint() const
 {
-  return this->transpose(); // in the complex case, the .conjugate() is be implicit here
-                            // due to implicit conversion to return type
+  return AdjointReturnType(this->transpose());
 }
 
 /***************************************************************************
@@ -247,18 +213,38 @@ MatrixBase<Derived>::adjoint() const
 namespace internal {
 
 template<typename MatrixType,
-  bool IsSquare = (MatrixType::RowsAtCompileTime == MatrixType::ColsAtCompileTime) && MatrixType::RowsAtCompileTime!=Dynamic>
+  bool IsSquare = (MatrixType::RowsAtCompileTime == MatrixType::ColsAtCompileTime) && MatrixType::RowsAtCompileTime!=Dynamic,
+  bool MatchPacketSize =
+        (int(MatrixType::RowsAtCompileTime) == int(internal::packet_traits<typename MatrixType::Scalar>::size))
+    &&  (internal::evaluator<MatrixType>::Flags&PacketAccessBit) >
 struct inplace_transpose_selector;
 
 template<typename MatrixType>
-struct inplace_transpose_selector<MatrixType,true> { // square matrix
+struct inplace_transpose_selector<MatrixType,true,false> { // square matrix
   static void run(MatrixType& m) {
     m.matrix().template triangularView<StrictlyUpper>().swap(m.matrix().transpose());
   }
 };
 
+// TODO: vectorized path is currently limited to LargestPacketSize x LargestPacketSize cases only.
 template<typename MatrixType>
-struct inplace_transpose_selector<MatrixType,false> { // non square matrix
+struct inplace_transpose_selector<MatrixType,true,true> { // PacketSize x PacketSize
+  static void run(MatrixType& m) {
+    typedef typename MatrixType::Scalar Scalar;
+    typedef typename internal::packet_traits<typename MatrixType::Scalar>::type Packet;
+    const Index PacketSize = internal::packet_traits<Scalar>::size;
+    const Index Alignment = internal::evaluator<MatrixType>::Alignment;
+    PacketBlock<Packet> A;
+    for (Index i=0; i<PacketSize; ++i)
+      A.packet[i] = m.template packetByOuterInner<Alignment>(i,0);
+    internal::ptranspose(A);
+    for (Index i=0; i<PacketSize; ++i)
+      m.template writePacket<Alignment>(m.rowIndexByOuterInner(i,0), m.colIndexByOuterInner(i,0), A.packet[i]);
+  }
+};
+
+template<typename MatrixType,bool MatchPacketSize>
+struct inplace_transpose_selector<MatrixType,false,MatchPacketSize> { // non square matrix
   static void run(MatrixType& m) {
     if (m.rows()==m.cols())
       m.matrix().template triangularView<StrictlyUpper>().swap(m.matrix().transpose());
@@ -284,7 +270,8 @@ struct inplace_transpose_selector<MatrixType,false> { // non square matrix
   * Notice however that this method is only useful if you want to replace a matrix by its own transpose.
   * If you just need the transpose of a matrix, use transpose().
   *
-  * \note if the matrix is not square, then \c *this must be a resizable matrix.
+  * \note if the matrix is not square, then \c *this must be a resizable matrix. 
+  * This excludes (non-square) fixed-size matrices, block-expressions and maps.
   *
   * \sa transpose(), adjoint(), adjointInPlace() */
 template<typename Derived>
@@ -315,6 +302,7 @@ inline void DenseBase<Derived>::transposeInPlace()
   * If you just need the adjoint of a matrix, use adjoint().
   *
   * \note if the matrix is not square, then \c *this must be a resizable matrix.
+  * This excludes (non-square) fixed-size matrices, block-expressions and maps.
   *
   * \sa transpose(), adjoint(), transposeInPlace() */
 template<typename Derived>
@@ -329,14 +317,6 @@ inline void MatrixBase<Derived>::adjointInPlace()
 
 namespace internal {
 
-template<typename BinOp,typename NestedXpr,typename Rhs>
-struct blas_traits<SelfCwiseBinaryOp<BinOp,NestedXpr,Rhs> >
- : blas_traits<NestedXpr>
-{
-  typedef SelfCwiseBinaryOp<BinOp,NestedXpr,Rhs> XprType;
-  static inline const XprType extract(const XprType& x) { return x; }
-};
-
 template<bool DestIsTransposed, typename OtherDerived>
 struct check_transpose_aliasing_compile_time_selector
 {
@@ -402,15 +382,15 @@ struct checkTransposeAliasing_impl<Derived, OtherDerived, false>
     }
 };
 
+template<typename Dst, typename Src>
+void check_for_aliasing(const Dst &dst, const Src &src)
+{
+  internal::checkTransposeAliasing_impl<Dst, Src>::run(dst, src);
+}
+
 } // end namespace internal
 
-template<typename Derived>
-template<typename OtherDerived>
-void DenseBase<Derived>::checkTransposeAliasing(const OtherDerived& other) const
-{
-    internal::checkTransposeAliasing_impl<Derived, OtherDerived>::run(derived(), other);
-}
-#endif
+#endif // EIGEN_NO_DEBUG
 
 } // end namespace Eigen
 

Eigen/src/Core/Transpositions.h

@@ -41,10 +41,6 @@ namespace Eigen {
   * \sa class PermutationMatrix
   */
 
-namespace internal {
-template<typename TranspositionType, typename MatrixType, int Side, bool Transposed=false> struct transposition_matrix_product_retval;
-}
-
 template<typename Derived>
 class TranspositionsBase
 {
@@ -53,7 +49,8 @@ class TranspositionsBase
   public:
 
     typedef typename Traits::IndicesType IndicesType;
-    typedef typename IndicesType::Scalar Index;
+    typedef typename IndicesType::Scalar StorageIndex;
+    typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3
 
     Derived& derived() { return *static_cast<Derived*>(this); }
     const Derived& derived() const { return *static_cast<const Derived*>(this); }
@@ -65,7 +62,7 @@ class TranspositionsBase
       indices() = other.indices();
       return derived();
     }
-
+    
     #ifndef EIGEN_PARSED_BY_DOXYGEN
     /** This is a special case of the templated operator=. Its purpose is to
       * prevent a default operator= from hiding the templated operator=.
@@ -78,20 +75,24 @@ class TranspositionsBase
     #endif
 
     /** \returns the number of transpositions */
-    inline Index size() const { return indices().size(); }
+    Index size() const { return indices().size(); }
+    /** \returns the number of rows of the equivalent permutation matrix */
+    Index rows() const { return indices().size(); }
+    /** \returns the number of columns of the equivalent permutation matrix */
+    Index cols() const { return indices().size(); }
 
     /** Direct access to the underlying index vector */
-    inline const Index& coeff(Index i) const { return indices().coeff(i); }
+    inline const StorageIndex& coeff(Index i) const { return indices().coeff(i); }
     /** Direct access to the underlying index vector */
-    inline Index& coeffRef(Index i) { return indices().coeffRef(i); }
+    inline StorageIndex& coeffRef(Index i) { return indices().coeffRef(i); }
     /** Direct access to the underlying index vector */
-    inline const Index& operator()(Index i) const { return indices()(i); }
+    inline const StorageIndex& operator()(Index i) const { return indices()(i); }
     /** Direct access to the underlying index vector */
-    inline Index& operator()(Index i) { return indices()(i); }
+    inline StorageIndex& operator()(Index i) { return indices()(i); }
     /** Direct access to the underlying index vector */
-    inline const Index& operator[](Index i) const { return indices()(i); }
+    inline const StorageIndex& operator[](Index i) const { return indices()(i); }
     /** Direct access to the underlying index vector */
-    inline Index& operator[](Index i) { return indices()(i); }
+    inline StorageIndex& operator[](Index i) { return indices()(i); }
 
     /** const version of indices(). */
     const IndicesType& indices() const { return derived().indices(); }
@@ -99,7 +100,7 @@ class TranspositionsBase
     IndicesType& indices() { return derived().indices(); }
 
     /** Resizes to given size. */
-    inline void resize(int newSize)
+    inline void resize(Index newSize)
     {
       indices().resize(newSize);
     }
@@ -107,7 +108,7 @@ class TranspositionsBase
     /** Sets \c *this to represents an identity transformation */
     void setIdentity()
     {
-      for(int i = 0; i < indices().size(); ++i)
+      for(StorageIndex i = 0; i < indices().size(); ++i)
         coeffRef(i) = i;
     }
 
@@ -144,23 +145,24 @@ class TranspositionsBase
 };
 
 namespace internal {
-template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename IndexType>
-struct traits<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,IndexType> >
+template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename _StorageIndex>
+struct traits<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,_StorageIndex> >
+ : traits<PermutationMatrix<SizeAtCompileTime,MaxSizeAtCompileTime,_StorageIndex> >
 {
-  typedef IndexType Index;
-  typedef Matrix<Index, SizeAtCompileTime, 1, 0, MaxSizeAtCompileTime, 1> IndicesType;
+  typedef Matrix<_StorageIndex, SizeAtCompileTime, 1, 0, MaxSizeAtCompileTime, 1> IndicesType;
+  typedef TranspositionsStorage StorageKind;
 };
 }
 
-template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename IndexType>
-class Transpositions : public TranspositionsBase<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,IndexType> >
+template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename _StorageIndex>
+class Transpositions : public TranspositionsBase<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,_StorageIndex> >
 {
     typedef internal::traits<Transpositions> Traits;
   public:
 
     typedef TranspositionsBase<Transpositions> Base;
     typedef typename Traits::IndicesType IndicesType;
-    typedef typename IndicesType::Scalar Index;
+    typedef typename IndicesType::Scalar StorageIndex;
 
     inline Transpositions() {}
 
@@ -177,7 +179,7 @@ class Transpositions : public TranspositionsBase<Transpositions<SizeAtCompileTim
 
     /** Generic constructor from expression of the transposition indices. */
     template<typename Other>
-    explicit inline Transpositions(const MatrixBase<Other>& a_indices) : m_indices(a_indices)
+    explicit inline Transpositions(const MatrixBase<Other>& indices) : m_indices(indices)
     {}
 
     /** Copies the \a other transpositions into \c *this */
@@ -215,30 +217,32 @@ class Transpositions : public TranspositionsBase<Transpositions<SizeAtCompileTim
 
 
 namespace internal {
-template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename IndexType, int _PacketAccess>
-struct traits<Map<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,IndexType>,_PacketAccess> >
+template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename _StorageIndex, int _PacketAccess>
+struct traits<Map<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,_StorageIndex>,_PacketAccess> >
+ : traits<PermutationMatrix<SizeAtCompileTime,MaxSizeAtCompileTime,_StorageIndex> >
 {
-  typedef IndexType Index;
-  typedef Map<const Matrix<Index,SizeAtCompileTime,1,0,MaxSizeAtCompileTime,1>, _PacketAccess> IndicesType;
+  typedef Map<const Matrix<_StorageIndex,SizeAtCompileTime,1,0,MaxSizeAtCompileTime,1>, _PacketAccess> IndicesType;
+  typedef _StorageIndex StorageIndex;
+  typedef TranspositionsStorage StorageKind;
 };
 }
 
-template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename IndexType, int PacketAccess>
-class Map<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,IndexType>,PacketAccess>
- : public TranspositionsBase<Map<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,IndexType>,PacketAccess> >
+template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename _StorageIndex, int PacketAccess>
+class Map<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,_StorageIndex>,PacketAccess>
+ : public TranspositionsBase<Map<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,_StorageIndex>,PacketAccess> >
 {
     typedef internal::traits<Map> Traits;
   public:
 
     typedef TranspositionsBase<Map> Base;
     typedef typename Traits::IndicesType IndicesType;
-    typedef typename IndicesType::Scalar Index;
+    typedef typename IndicesType::Scalar StorageIndex;
 
-    inline Map(const Index* indicesPtr)
+    explicit inline Map(const StorageIndex* indicesPtr)
       : m_indices(indicesPtr)
     {}
 
-    inline Map(const Index* indicesPtr, Index size)
+    inline Map(const StorageIndex* indicesPtr, Index size)
       : m_indices(indicesPtr,size)
     {}
 
@@ -274,9 +278,9 @@ class Map<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,IndexType>,Packe
 namespace internal {
 template<typename _IndicesType>
 struct traits<TranspositionsWrapper<_IndicesType> >
+ : traits<PermutationWrapper<_IndicesType> >
 {
-  typedef typename _IndicesType::Scalar Index;
-  typedef _IndicesType IndicesType;
+  typedef TranspositionsStorage StorageKind;
 };
 }
 
@@ -289,10 +293,10 @@ class TranspositionsWrapper
 
     typedef TranspositionsBase<TranspositionsWrapper> Base;
     typedef typename Traits::IndicesType IndicesType;
-    typedef typename IndicesType::Scalar Index;
+    typedef typename IndicesType::Scalar StorageIndex;
 
-    inline TranspositionsWrapper(IndicesType& a_indices)
-      : m_indices(a_indices)
+    explicit inline TranspositionsWrapper(IndicesType& indices)
+      : m_indices(indices)
     {}
 
     /** Copies the \a other transpositions into \c *this */
@@ -324,80 +328,43 @@ class TranspositionsWrapper
     const typename IndicesType::Nested m_indices;
 };
 
+
+
 /** \returns the \a matrix with the \a transpositions applied to the columns.
   */
-template<typename Derived, typename TranspositionsDerived>
-inline const internal::transposition_matrix_product_retval<TranspositionsDerived, Derived, OnTheRight>
-operator*(const MatrixBase<Derived>& matrix,
-          const TranspositionsBase<TranspositionsDerived> &transpositions)
+template<typename MatrixDerived, typename TranspositionsDerived>
+EIGEN_DEVICE_FUNC
+const Product<MatrixDerived, TranspositionsDerived, AliasFreeProduct>
+operator*(const MatrixBase<MatrixDerived> &matrix,
+          const TranspositionsBase<TranspositionsDerived>& transpositions)
 {
-  return internal::transposition_matrix_product_retval
-           <TranspositionsDerived, Derived, OnTheRight>
-           (transpositions.derived(), matrix.derived());
+  return Product<MatrixDerived, TranspositionsDerived, AliasFreeProduct>
+            (matrix.derived(), transpositions.derived());
 }
 
 /** \returns the \a matrix with the \a transpositions applied to the rows.
   */
-template<typename Derived, typename TranspositionDerived>
-inline const internal::transposition_matrix_product_retval
-               <TranspositionDerived, Derived, OnTheLeft>
-operator*(const TranspositionsBase<TranspositionDerived> &transpositions,
-          const MatrixBase<Derived>& matrix)
+template<typename TranspositionsDerived, typename MatrixDerived>
+EIGEN_DEVICE_FUNC
+const Product<TranspositionsDerived, MatrixDerived, AliasFreeProduct>
+operator*(const TranspositionsBase<TranspositionsDerived> &transpositions,
+          const MatrixBase<MatrixDerived>& matrix)
 {
-  return internal::transposition_matrix_product_retval
-           <TranspositionDerived, Derived, OnTheLeft>
-           (transpositions.derived(), matrix.derived());
+  return Product<TranspositionsDerived, MatrixDerived, AliasFreeProduct>
+            (transpositions.derived(), matrix.derived());
 }
 
+// Template partial specialization for transposed/inverse transpositions
+
 namespace internal {
 
-template<typename TranspositionType, typename MatrixType, int Side, bool Transposed>
-struct traits<transposition_matrix_product_retval<TranspositionType, MatrixType, Side, Transposed> >
-{
-  typedef typename MatrixType::PlainObject ReturnType;
-};
-
-template<typename TranspositionType, typename MatrixType, int Side, bool Transposed>
-struct transposition_matrix_product_retval
- : public ReturnByValue<transposition_matrix_product_retval<TranspositionType, MatrixType, Side, Transposed> >
-{
-    typedef typename remove_all<typename MatrixType::Nested>::type MatrixTypeNestedCleaned;
-    typedef typename TranspositionType::Index Index;
-
-    transposition_matrix_product_retval(const TranspositionType& tr, const MatrixType& matrix)
-      : m_transpositions(tr), m_matrix(matrix)
-    {}
-
-    inline int rows() const { return m_matrix.rows(); }
-    inline int cols() const { return m_matrix.cols(); }
-
-    template<typename Dest> inline void evalTo(Dest& dst) const
-    {
-      const int size = m_transpositions.size();
-      Index j = 0;
-
-      if(!(is_same<MatrixTypeNestedCleaned,Dest>::value && extract_data(dst) == extract_data(m_matrix)))
-        dst = m_matrix;
-
-      for(int k=(Transposed?size-1:0) ; Transposed?k>=0:k<size ; Transposed?--k:++k)
-        if((j=m_transpositions.coeff(k))!=k)
-        {
-          if(Side==OnTheLeft)
-            dst.row(k).swap(dst.row(j));
-          else if(Side==OnTheRight)
-            dst.col(k).swap(dst.col(j));
-        }
-    }
-
-  protected:
-    const TranspositionType& m_transpositions;
-    typename MatrixType::Nested m_matrix;
-};
+template<typename Derived>
+struct traits<Transpose<TranspositionsBase<Derived> > >
+ : traits<Derived>
+{};
 
 } // end namespace internal
 
-/* Template partial specialization for transposed/inverse transpositions */
-
 template<typename TranspositionsDerived>
 class Transpose<TranspositionsBase<TranspositionsDerived> >
 {
@@ -405,27 +372,31 @@ class Transpose<TranspositionsBase<TranspositionsDerived> >
     typedef typename TranspositionType::IndicesType IndicesType;
   public:
 
-    Transpose(const TranspositionType& t) : m_transpositions(t) {}
+    explicit Transpose(const TranspositionType& t) : m_transpositions(t) {}
 
-    inline int size() const { return m_transpositions.size(); }
+    Index size() const { return m_transpositions.size(); }
+    Index rows() const { return m_transpositions.size(); }
+    Index cols() const { return m_transpositions.size(); }
 
     /** \returns the \a matrix with the inverse transpositions applied to the columns.
       */
-    template<typename Derived> friend
-    inline const internal::transposition_matrix_product_retval<TranspositionType, Derived, OnTheRight, true>
-    operator*(const MatrixBase<Derived>& matrix, const Transpose& trt)
+    template<typename OtherDerived> friend
+    const Product<OtherDerived, Transpose, AliasFreeProduct>
+    operator*(const MatrixBase<OtherDerived>& matrix, const Transpose& trt)
     {
-      return internal::transposition_matrix_product_retval<TranspositionType, Derived, OnTheRight, true>(trt.m_transpositions, matrix.derived());
+      return Product<OtherDerived, Transpose, AliasFreeProduct>(matrix.derived(), trt.derived());
     }
 
     /** \returns the \a matrix with the inverse transpositions applied to the rows.
       */
-    template<typename Derived>
-    inline const internal::transposition_matrix_product_retval<TranspositionType, Derived, OnTheLeft, true>
-    operator*(const MatrixBase<Derived>& matrix) const
+    template<typename OtherDerived>
+    const Product<Transpose, OtherDerived, AliasFreeProduct>
+    operator*(const MatrixBase<OtherDerived>& matrix) const
     {
-      return internal::transposition_matrix_product_retval<TranspositionType, Derived, OnTheLeft, true>(m_transpositions, matrix.derived());
+      return Product<Transpose, OtherDerived, AliasFreeProduct>(*this, matrix.derived());
     }
+    
+    const TranspositionType& nestedExpression() const { return m_transpositions; }
 
   protected:
     const TranspositionType& m_transpositions;

Eigen/src/Core/VectorBlock.h

@@ -72,6 +72,7 @@ template<typename VectorType, int Size> class VectorBlock
 
     /** Dynamic-size constructor
       */
+    EIGEN_DEVICE_FUNC
     inline VectorBlock(VectorType& vector, Index start, Index size)
       : Base(vector,
              IsColVector ? start : 0, IsColVector ? 0 : start,
@@ -82,6 +83,7 @@ template<typename VectorType, int Size> class VectorBlock
 
     /** Fixed-size constructor
       */
+    EIGEN_DEVICE_FUNC
     inline VectorBlock(VectorType& vector, Index start)
       : Base(vector, IsColVector ? start : 0, IsColVector ? 0 : start)
     {

Eigen/src/Core/VectorwiseOp.h

@@ -41,31 +41,22 @@ struct traits<PartialReduxExpr<MatrixType, MemberOp, Direction> >
   typedef typename traits<MatrixType>::StorageKind StorageKind;
   typedef typename traits<MatrixType>::XprKind XprKind;
   typedef typename MatrixType::Scalar InputScalar;
-  typedef typename nested<MatrixType>::type MatrixTypeNested;
+  typedef typename ref_selector<MatrixType>::type MatrixTypeNested;
   typedef typename remove_all<MatrixTypeNested>::type _MatrixTypeNested;
   enum {
     RowsAtCompileTime = Direction==Vertical   ? 1 : MatrixType::RowsAtCompileTime,
     ColsAtCompileTime = Direction==Horizontal ? 1 : MatrixType::ColsAtCompileTime,
     MaxRowsAtCompileTime = Direction==Vertical   ? 1 : MatrixType::MaxRowsAtCompileTime,
     MaxColsAtCompileTime = Direction==Horizontal ? 1 : MatrixType::MaxColsAtCompileTime,
-    Flags0 = (unsigned int)_MatrixTypeNested::Flags & HereditaryBits,
-    Flags = (Flags0 & ~RowMajorBit) | (RowsAtCompileTime == 1 ? RowMajorBit : 0),
-    TraversalSize = Direction==Vertical ? RowsAtCompileTime : ColsAtCompileTime
-  };
-  #if EIGEN_GNUC_AT_LEAST(3,4)
-  typedef typename MemberOp::template Cost<InputScalar,int(TraversalSize)> CostOpType;
-  #else
-  typedef typename MemberOp::template Cost<InputScalar,TraversalSize> CostOpType;
-  #endif
-  enum {
-    CoeffReadCost = TraversalSize * traits<_MatrixTypeNested>::CoeffReadCost + int(CostOpType::value)
+    Flags = RowsAtCompileTime == 1 ? RowMajorBit : 0,
+    TraversalSize = Direction==Vertical ? MatrixType::RowsAtCompileTime :  MatrixType::ColsAtCompileTime
   };
 };
 }
 
 template< typename MatrixType, typename MemberOp, int Direction>
-class PartialReduxExpr : internal::no_assignment_operator,
-  public internal::dense_xpr_base< PartialReduxExpr<MatrixType, MemberOp, Direction> >::type
+class PartialReduxExpr : public internal::dense_xpr_base< PartialReduxExpr<MatrixType, MemberOp, Direction> >::type,
+                         internal::no_assignment_operator
 {
   public:
 
@@ -74,13 +65,16 @@ class PartialReduxExpr : internal::no_assignment_operator,
     typedef typename internal::traits<PartialReduxExpr>::MatrixTypeNested MatrixTypeNested;
     typedef typename internal::traits<PartialReduxExpr>::_MatrixTypeNested _MatrixTypeNested;
 
-    PartialReduxExpr(const MatrixType& mat, const MemberOp& func = MemberOp())
+    EIGEN_DEVICE_FUNC
+    explicit PartialReduxExpr(const MatrixType& mat, const MemberOp& func = MemberOp())
       : m_matrix(mat), m_functor(func) {}
 
+    EIGEN_DEVICE_FUNC
     Index rows() const { return (Direction==Vertical   ? 1 : m_matrix.rows()); }
+    EIGEN_DEVICE_FUNC
     Index cols() const { return (Direction==Horizontal ? 1 : m_matrix.cols()); }
 
-    EIGEN_STRONG_INLINE const Scalar coeff(Index i, Index j) const
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar coeff(Index i, Index j) const
     {
       if (Direction==Vertical)
         return m_functor(m_matrix.col(j));
@@ -88,7 +82,7 @@ class PartialReduxExpr : internal::no_assignment_operator,
         return m_functor(m_matrix.row(i));
     }
 
-    const Scalar coeff(Index index) const
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar coeff(Index index) const
     {
       if (Direction==Vertical)
         return m_functor(m_matrix.col(index));
@@ -109,7 +103,8 @@ class PartialReduxExpr : internal::no_assignment_operator,
     template<typename Scalar, int Size> struct Cost                     \
     { enum { value = COST }; };                                         \
     template<typename XprType>                                          \
-    EIGEN_STRONG_INLINE ResultType operator()(const XprType& mat) const \
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE                               \
+    ResultType operator()(const XprType& mat) const                     \
     { return mat.MEMBER(); } \
   }
 
@@ -133,13 +128,13 @@ EIGEN_MEMBER_FUNCTOR(prod, (Size-1)*NumTraits<Scalar>::MulCost);
 template <typename BinaryOp, typename Scalar>
 struct member_redux {
   typedef typename result_of<
-                     BinaryOp(Scalar)
+                     BinaryOp(Scalar,Scalar)
                    >::type  result_type;
   template<typename _Scalar, int Size> struct Cost
   { enum { value = (Size-1) * functor_traits<BinaryOp>::Cost }; };
-  member_redux(const BinaryOp func) : m_functor(func) {}
+  EIGEN_DEVICE_FUNC explicit member_redux(const BinaryOp func) : m_functor(func) {}
   template<typename Derived>
-  inline result_type operator()(const DenseBase<Derived>& mat) const
+  EIGEN_DEVICE_FUNC inline result_type operator()(const DenseBase<Derived>& mat) const
   { return mat.redux(m_functor); }
   const BinaryOp m_functor;
 };
@@ -168,16 +163,15 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
 
     typedef typename ExpressionType::Scalar Scalar;
     typedef typename ExpressionType::RealScalar RealScalar;
-    typedef typename ExpressionType::Index Index;
-    typedef typename internal::conditional<internal::must_nest_by_value<ExpressionType>::ret,
-        ExpressionType, ExpressionType&>::type ExpressionTypeNested;
+    typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3
+    typedef typename internal::ref_selector<ExpressionType>::non_const_type ExpressionTypeNested;
     typedef typename internal::remove_all<ExpressionTypeNested>::type ExpressionTypeNestedCleaned;
 
     template<template<typename _Scalar> class Functor,
-                      typename Scalar=typename internal::traits<ExpressionType>::Scalar> struct ReturnType
+                      typename Scalar_=Scalar> struct ReturnType
     {
       typedef PartialReduxExpr<ExpressionType,
-                               Functor<Scalar>,
+                               Functor<Scalar_>,
                                Direction
                               > Type;
     };
@@ -185,23 +179,24 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
     template<typename BinaryOp> struct ReduxReturnType
     {
       typedef PartialReduxExpr<ExpressionType,
-                               internal::member_redux<BinaryOp,typename internal::traits<ExpressionType>::Scalar>,
+                               internal::member_redux<BinaryOp,Scalar>,
                                Direction
                               > Type;
     };
 
     enum {
-      IsVertical   = (Direction==Vertical) ? 1 : 0,
-      IsHorizontal = (Direction==Horizontal) ? 1 : 0
+      isVertical   = (Direction==Vertical) ? 1 : 0,
+      isHorizontal = (Direction==Horizontal) ? 1 : 0
     };
 
   protected:
 
     /** \internal
       * \returns the i-th subvector according to the \c Direction */
-    typedef typename internal::conditional<Direction==Vertical,
+    typedef typename internal::conditional<isVertical,
                                typename ExpressionType::ColXpr,
                                typename ExpressionType::RowXpr>::type SubVector;
+    EIGEN_DEVICE_FUNC
     SubVector subVector(Index i)
     {
       return SubVector(m_matrix.derived(),i);
@@ -209,58 +204,62 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
 
     /** \internal
       * \returns the number of subvectors in the direction \c Direction */
+    EIGEN_DEVICE_FUNC
     Index subVectors() const
-    { return Direction==Vertical?m_matrix.cols():m_matrix.rows(); }
+    { return isVertical?m_matrix.cols():m_matrix.rows(); }
 
     template<typename OtherDerived> struct ExtendedType {
       typedef Replicate<OtherDerived,
-                        Direction==Vertical   ? 1 : ExpressionType::RowsAtCompileTime,
-                        Direction==Horizontal ? 1 : ExpressionType::ColsAtCompileTime> Type;
+                        isVertical   ? 1 : ExpressionType::RowsAtCompileTime,
+                        isHorizontal ? 1 : ExpressionType::ColsAtCompileTime> Type;
     };
 
     /** \internal
       * Replicates a vector to match the size of \c *this */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     typename ExtendedType<OtherDerived>::Type
     extendedTo(const DenseBase<OtherDerived>& other) const
     {
-      EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(Direction==Vertical, OtherDerived::MaxColsAtCompileTime==1),
+      EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(isVertical, OtherDerived::MaxColsAtCompileTime==1),
                           YOU_PASSED_A_ROW_VECTOR_BUT_A_COLUMN_VECTOR_WAS_EXPECTED)
-      EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(Direction==Horizontal, OtherDerived::MaxRowsAtCompileTime==1),
+      EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(isHorizontal, OtherDerived::MaxRowsAtCompileTime==1),
                           YOU_PASSED_A_COLUMN_VECTOR_BUT_A_ROW_VECTOR_WAS_EXPECTED)
       return typename ExtendedType<OtherDerived>::Type
                       (other.derived(),
-                       Direction==Vertical   ? 1 : m_matrix.rows(),
-                       Direction==Horizontal ? 1 : m_matrix.cols());
+                       isVertical   ? 1 : m_matrix.rows(),
+                       isHorizontal ? 1 : m_matrix.cols());
     }
     
     template<typename OtherDerived> struct OppositeExtendedType {
       typedef Replicate<OtherDerived,
-                        Direction==Horizontal ? 1 : ExpressionType::RowsAtCompileTime,
-                        Direction==Vertical   ? 1 : ExpressionType::ColsAtCompileTime> Type;
+                        isHorizontal ? 1 : ExpressionType::RowsAtCompileTime,
+                        isVertical   ? 1 : ExpressionType::ColsAtCompileTime> Type;
     };
 
     /** \internal
       * Replicates a vector in the opposite direction to match the size of \c *this */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     typename OppositeExtendedType<OtherDerived>::Type
     extendedToOpposite(const DenseBase<OtherDerived>& other) const
     {
-      EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(Direction==Horizontal, OtherDerived::MaxColsAtCompileTime==1),
+      EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(isHorizontal, OtherDerived::MaxColsAtCompileTime==1),
                           YOU_PASSED_A_ROW_VECTOR_BUT_A_COLUMN_VECTOR_WAS_EXPECTED)
-      EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(Direction==Vertical, OtherDerived::MaxRowsAtCompileTime==1),
+      EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(isVertical, OtherDerived::MaxRowsAtCompileTime==1),
                           YOU_PASSED_A_COLUMN_VECTOR_BUT_A_ROW_VECTOR_WAS_EXPECTED)
       return typename OppositeExtendedType<OtherDerived>::Type
                       (other.derived(),
-                       Direction==Horizontal  ? 1 : m_matrix.rows(),
-                       Direction==Vertical    ? 1 : m_matrix.cols());
+                       isHorizontal  ? 1 : m_matrix.rows(),
+                       isVertical    ? 1 : m_matrix.cols());
     }
 
   public:
-
-    inline VectorwiseOp(ExpressionType& matrix) : m_matrix(matrix) {}
+    EIGEN_DEVICE_FUNC
+    explicit inline VectorwiseOp(ExpressionType& matrix) : m_matrix(matrix) {}
 
     /** \internal */
+    EIGEN_DEVICE_FUNC
     inline const ExpressionType& _expression() const { return m_matrix; }
 
     /** \returns a row or column vector expression of \c *this reduxed by \a func
@@ -271,9 +270,25 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
       * \sa class VectorwiseOp, DenseBase::colwise(), DenseBase::rowwise()
       */
     template<typename BinaryOp>
+    EIGEN_DEVICE_FUNC
     const typename ReduxReturnType<BinaryOp>::Type
     redux(const BinaryOp& func = BinaryOp()) const
-    { return typename ReduxReturnType<BinaryOp>::Type(_expression(), func); }
+    { return typename ReduxReturnType<BinaryOp>::Type(_expression(), internal::member_redux<BinaryOp,Scalar>(func)); }
+
+    typedef typename ReturnType<internal::member_minCoeff>::Type MinCoeffReturnType;
+    typedef typename ReturnType<internal::member_maxCoeff>::Type MaxCoeffReturnType;
+    typedef typename ReturnType<internal::member_squaredNorm,RealScalar>::Type SquaredNormReturnType;
+    typedef typename ReturnType<internal::member_norm,RealScalar>::Type NormReturnType;
+    typedef typename ReturnType<internal::member_blueNorm,RealScalar>::Type BlueNormReturnType;
+    typedef typename ReturnType<internal::member_stableNorm,RealScalar>::Type StableNormReturnType;
+    typedef typename ReturnType<internal::member_hypotNorm,RealScalar>::Type HypotNormReturnType;
+    typedef typename ReturnType<internal::member_sum>::Type SumReturnType;
+    typedef typename ReturnType<internal::member_mean>::Type MeanReturnType;
+    typedef typename ReturnType<internal::member_all>::Type AllReturnType;
+    typedef typename ReturnType<internal::member_any>::Type AnyReturnType;
+    typedef PartialReduxExpr<ExpressionType, internal::member_count<Index>, Direction> CountReturnType;
+    typedef typename ReturnType<internal::member_prod>::Type ProdReturnType;
+    typedef Reverse<ExpressionType, Direction> ReverseReturnType;
 
     /** \returns a row (or column) vector expression of the smallest coefficient
       * of each column (or row) of the referenced expression.
@@ -284,8 +299,9 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
       * Output: \verbinclude PartialRedux_minCoeff.out
       *
       * \sa DenseBase::minCoeff() */
-    const typename ReturnType<internal::member_minCoeff>::Type minCoeff() const
-    { return _expression(); }
+    EIGEN_DEVICE_FUNC
+    const MinCoeffReturnType minCoeff() const
+    { return MinCoeffReturnType(_expression()); }
 
     /** \returns a row (or column) vector expression of the largest coefficient
       * of each column (or row) of the referenced expression.
@@ -296,55 +312,66 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
       * Output: \verbinclude PartialRedux_maxCoeff.out
       *
       * \sa DenseBase::maxCoeff() */
-    const typename ReturnType<internal::member_maxCoeff>::Type maxCoeff() const
-    { return _expression(); }
+    EIGEN_DEVICE_FUNC
+    const MaxCoeffReturnType maxCoeff() const
+    { return MaxCoeffReturnType(_expression()); }
 
     /** \returns a row (or column) vector expression of the squared norm
       * of each column (or row) of the referenced expression.
+      * This is a vector with real entries, even if the original matrix has complex entries.
       *
       * Example: \include PartialRedux_squaredNorm.cpp
       * Output: \verbinclude PartialRedux_squaredNorm.out
       *
       * \sa DenseBase::squaredNorm() */
-    const typename ReturnType<internal::member_squaredNorm,RealScalar>::Type squaredNorm() const
-    { return _expression(); }
+    EIGEN_DEVICE_FUNC
+    const SquaredNormReturnType squaredNorm() const
+    { return SquaredNormReturnType(_expression()); }
 
     /** \returns a row (or column) vector expression of the norm
       * of each column (or row) of the referenced expression.
+      * This is a vector with real entries, even if the original matrix has complex entries.
       *
       * Example: \include PartialRedux_norm.cpp
       * Output: \verbinclude PartialRedux_norm.out
       *
       * \sa DenseBase::norm() */
-    const typename ReturnType<internal::member_norm,RealScalar>::Type norm() const
-    { return _expression(); }
+    EIGEN_DEVICE_FUNC
+    const NormReturnType norm() const
+    { return NormReturnType(_expression()); }
 
 
     /** \returns a row (or column) vector expression of the norm
       * of each column (or row) of the referenced expression, using
-      * blue's algorithm.
+      * Blue's algorithm. 
+      * This is a vector with real entries, even if the original matrix has complex entries.
       *
       * \sa DenseBase::blueNorm() */
-    const typename ReturnType<internal::member_blueNorm,RealScalar>::Type blueNorm() const
-    { return _expression(); }
+    EIGEN_DEVICE_FUNC
+    const BlueNormReturnType blueNorm() const
+    { return BlueNormReturnType(_expression()); }
 
 
     /** \returns a row (or column) vector expression of the norm
       * of each column (or row) of the referenced expression, avoiding
       * underflow and overflow.
+      * This is a vector with real entries, even if the original matrix has complex entries.
       *
       * \sa DenseBase::stableNorm() */
-    const typename ReturnType<internal::member_stableNorm,RealScalar>::Type stableNorm() const
-    { return _expression(); }
+    EIGEN_DEVICE_FUNC
+    const StableNormReturnType stableNorm() const
+    { return StableNormReturnType(_expression()); }
 
 
     /** \returns a row (or column) vector expression of the norm
       * of each column (or row) of the referenced expression, avoiding
       * underflow and overflow using a concatenation of hypot() calls.
+      * This is a vector with real entries, even if the original matrix has complex entries.
       *
       * \sa DenseBase::hypotNorm() */
-    const typename ReturnType<internal::member_hypotNorm,RealScalar>::Type hypotNorm() const
-    { return _expression(); }
+    EIGEN_DEVICE_FUNC
+    const HypotNormReturnType hypotNorm() const
+    { return HypotNormReturnType(_expression()); }
 
     /** \returns a row (or column) vector expression of the sum
       * of each column (or row) of the referenced expression.
@@ -353,39 +380,48 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
       * Output: \verbinclude PartialRedux_sum.out
       *
       * \sa DenseBase::sum() */
-    const typename ReturnType<internal::member_sum>::Type sum() const
-    { return _expression(); }
+    EIGEN_DEVICE_FUNC
+    const SumReturnType sum() const
+    { return SumReturnType(_expression()); }
 
     /** \returns a row (or column) vector expression of the mean
     * of each column (or row) of the referenced expression.
     *
     * \sa DenseBase::mean() */
-    const typename ReturnType<internal::member_mean>::Type mean() const
-    { return _expression(); }
+    EIGEN_DEVICE_FUNC
+    const MeanReturnType mean() const
+    { return MeanReturnType(_expression()); }
 
     /** \returns a row (or column) vector expression representing
       * whether \b all coefficients of each respective column (or row) are \c true.
+      * This expression can be assigned to a vector with entries of type \c bool.
       *
       * \sa DenseBase::all() */
-    const typename ReturnType<internal::member_all>::Type all() const
-    { return _expression(); }
+    EIGEN_DEVICE_FUNC
+    const AllReturnType all() const
+    { return AllReturnType(_expression()); }
 
     /** \returns a row (or column) vector expression representing
       * whether \b at \b least one coefficient of each respective column (or row) is \c true.
+      * This expression can be assigned to a vector with entries of type \c bool.
       *
       * \sa DenseBase::any() */
-    const typename ReturnType<internal::member_any>::Type any() const
-    { return _expression(); }
+    EIGEN_DEVICE_FUNC
+    const AnyReturnType any() const
+    { return Any(_expression()); }
 
     /** \returns a row (or column) vector expression representing
       * the number of \c true coefficients of each respective column (or row).
+      * This expression can be assigned to a vector whose entries have the same type as is used to
+      * index entries of the original matrix; for dense matrices, this is \c std::ptrdiff_t .
       *
       * Example: \include PartialRedux_count.cpp
       * Output: \verbinclude PartialRedux_count.out
       *
       * \sa DenseBase::count() */
-    const PartialReduxExpr<ExpressionType, internal::member_count<Index>, Direction> count() const
-    { return _expression(); }
+    EIGEN_DEVICE_FUNC
+    const CountReturnType count() const
+    { return CountReturnType(_expression()); }
 
     /** \returns a row (or column) vector expression of the product
       * of each column (or row) of the referenced expression.
@@ -394,8 +430,9 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
       * Output: \verbinclude PartialRedux_prod.out
       *
       * \sa DenseBase::prod() */
-    const typename ReturnType<internal::member_prod>::Type prod() const
-    { return _expression(); }
+    EIGEN_DEVICE_FUNC
+    const ProdReturnType prod() const
+    { return ProdReturnType(_expression()); }
 
 
     /** \returns a matrix expression
@@ -405,10 +442,12 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
       * Output: \verbinclude Vectorwise_reverse.out
       *
       * \sa DenseBase::reverse() */
-    const Reverse<ExpressionType, Direction> reverse() const
-    { return Reverse<ExpressionType, Direction>( _expression() ); }
+    EIGEN_DEVICE_FUNC
+    const ReverseReturnType reverse() const
+    { return ReverseReturnType( _expression() ); }
 
-    typedef Replicate<ExpressionType,Direction==Vertical?Dynamic:1,Direction==Horizontal?Dynamic:1> ReplicateReturnType;
+    typedef Replicate<ExpressionType,(isVertical?Dynamic:1),(isHorizontal?Dynamic:1)> ReplicateReturnType;
+    EIGEN_DEVICE_FUNC
     const ReplicateReturnType replicate(Index factor) const;
 
     /**
@@ -420,17 +459,20 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
       * \sa VectorwiseOp::replicate(Index), DenseBase::replicate(), class Replicate
       */
     // NOTE implemented here because of sunstudio's compilation errors
-    template<int Factor> const Replicate<ExpressionType,(IsVertical?Factor:1),(IsHorizontal?Factor:1)>
+    // isVertical*Factor+isHorizontal instead of (isVertical?Factor:1) to handle CUDA bug with ternary operator
+    template<int Factor> const Replicate<ExpressionType,isVertical*Factor+isHorizontal,isHorizontal*Factor+isVertical>
+    EIGEN_DEVICE_FUNC
     replicate(Index factor = Factor) const
     {
-      return Replicate<ExpressionType,Direction==Vertical?Factor:1,Direction==Horizontal?Factor:1>
-          (_expression(),Direction==Vertical?factor:1,Direction==Horizontal?factor:1);
+      return Replicate<ExpressionType,(isVertical?Factor:1),(isHorizontal?Factor:1)>
+          (_expression(),isVertical?factor:1,isHorizontal?factor:1);
     }
 
 /////////// Artithmetic operators ///////////
 
     /** Copies the vector \a other to each subvector of \c *this */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     ExpressionType& operator=(const DenseBase<OtherDerived>& other)
     {
       EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
@@ -441,6 +483,7 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
 
     /** Adds the vector \a other to each subvector of \c *this */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     ExpressionType& operator+=(const DenseBase<OtherDerived>& other)
     {
       EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
@@ -450,6 +493,7 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
 
     /** Substracts the vector \a other to each subvector of \c *this */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     ExpressionType& operator-=(const DenseBase<OtherDerived>& other)
     {
       EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
@@ -459,6 +503,7 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
 
     /** Multiples each subvector of \c *this by the vector \a other */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     ExpressionType& operator*=(const DenseBase<OtherDerived>& other)
     {
       EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
@@ -470,6 +515,7 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
 
     /** Divides each subvector of \c *this by the vector \a other */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     ExpressionType& operator/=(const DenseBase<OtherDerived>& other)
     {
       EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
@@ -480,7 +526,7 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
     }
 
     /** Returns the expression of the sum of the vector \a other to each subvector of \c *this */
-    template<typename OtherDerived> EIGEN_STRONG_INLINE
+    template<typename OtherDerived> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC
     CwiseBinaryOp<internal::scalar_sum_op<Scalar>,
                   const ExpressionTypeNestedCleaned,
                   const typename ExtendedType<OtherDerived>::Type>
@@ -493,6 +539,7 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
 
     /** Returns the expression of the difference between each subvector of \c *this and the vector \a other */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     CwiseBinaryOp<internal::scalar_difference_op<Scalar>,
                   const ExpressionTypeNestedCleaned,
                   const typename ExtendedType<OtherDerived>::Type>
@@ -505,10 +552,11 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
 
     /** Returns the expression where each subvector is the product of the vector \a other
       * by the corresponding subvector of \c *this */
-    template<typename OtherDerived> EIGEN_STRONG_INLINE
+    template<typename OtherDerived> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC
     CwiseBinaryOp<internal::scalar_product_op<Scalar>,
                   const ExpressionTypeNestedCleaned,
                   const typename ExtendedType<OtherDerived>::Type>
+    EIGEN_DEVICE_FUNC
     operator*(const DenseBase<OtherDerived>& other) const
     {
       EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
@@ -520,6 +568,7 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
     /** Returns the expression where each subvector is the quotient of the corresponding
       * subvector of \c *this by the vector \a other */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     CwiseBinaryOp<internal::scalar_quotient_op<Scalar>,
                   const ExpressionTypeNestedCleaned,
                   const typename ExtendedType<OtherDerived>::Type>
@@ -535,6 +584,7 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
       * The referenced matrix is \b not modified.
       * \sa MatrixBase::normalized(), normalize()
       */
+    EIGEN_DEVICE_FUNC
     CwiseBinaryOp<internal::scalar_quotient_op<Scalar>,
                   const ExpressionTypeNestedCleaned,
                   const typename OppositeExtendedType<typename ReturnType<internal::member_norm,RealScalar>::Type>::Type>
@@ -544,18 +594,20 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
     /** Normalize in-place each row or columns of the referenced matrix.
       * \sa MatrixBase::normalize(), normalized()
       */
-    void normalize() {
+    EIGEN_DEVICE_FUNC void normalize() {
       m_matrix = this->normalized();
     }
 
+    EIGEN_DEVICE_FUNC inline void reverseInPlace();
+
 /////////// Geometry module ///////////
 
-    #if EIGEN2_SUPPORT_STAGE > STAGE20_RESOLVE_API_CONFLICTS
-    Homogeneous<ExpressionType,Direction> homogeneous() const;
-    #endif
+    typedef Homogeneous<ExpressionType,Direction> HomogeneousReturnType;
+    HomogeneousReturnType homogeneous() const;
 
     typedef typename ExpressionType::PlainObject CrossReturnType;
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     const CrossReturnType cross(const MatrixBase<OtherDerived>& other) const;
 
     enum {
@@ -586,19 +638,8 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
     ExpressionTypeNested m_matrix;
 };
 
-/** \returns a VectorwiseOp wrapper of *this providing additional partial reduction operations
-  *
-  * Example: \include MatrixBase_colwise.cpp
-  * Output: \verbinclude MatrixBase_colwise.out
-  *
-  * \sa rowwise(), class VectorwiseOp, \ref TutorialReductionsVisitorsBroadcasting
-  */
-template<typename Derived>
-inline const typename DenseBase<Derived>::ConstColwiseReturnType
-DenseBase<Derived>::colwise() const
-{
-  return derived();
-}
+//const colwise moved to DenseBase.h due to CUDA compiler bug
+
 
 /** \returns a writable VectorwiseOp wrapper of *this providing additional partial reduction operations
   *
@@ -608,22 +649,11 @@ template<typename Derived>
 inline typename DenseBase<Derived>::ColwiseReturnType
 DenseBase<Derived>::colwise()
 {
-  return derived();
+  return ColwiseReturnType(derived());
 }
 
-/** \returns a VectorwiseOp wrapper of *this providing additional partial reduction operations
-  *
-  * Example: \include MatrixBase_rowwise.cpp
-  * Output: \verbinclude MatrixBase_rowwise.out
-  *
-  * \sa colwise(), class VectorwiseOp, \ref TutorialReductionsVisitorsBroadcasting
-  */
-template<typename Derived>
-inline const typename DenseBase<Derived>::ConstRowwiseReturnType
-DenseBase<Derived>::rowwise() const
-{
-  return derived();
-}
+//const rowwise moved to DenseBase.h due to CUDA compiler bug
+
 
 /** \returns a writable VectorwiseOp wrapper of *this providing additional partial reduction operations
   *
@@ -633,7 +663,7 @@ template<typename Derived>
 inline typename DenseBase<Derived>::RowwiseReturnType
 DenseBase<Derived>::rowwise()
 {
-  return derived();
+  return RowwiseReturnType(derived());
 }
 
 } // end namespace Eigen

Eigen/src/Core/Visitor.h

@@ -22,6 +22,7 @@ struct visitor_impl
     row = (UnrollCount-1) % Derived::RowsAtCompileTime
   };
 
+  EIGEN_DEVICE_FUNC
   static inline void run(const Derived &mat, Visitor& visitor)
   {
     visitor_impl<Visitor, Derived, UnrollCount-1>::run(mat, visitor);
@@ -32,6 +33,7 @@ struct visitor_impl
 template<typename Visitor, typename Derived>
 struct visitor_impl<Visitor, Derived, 1>
 {
+  EIGEN_DEVICE_FUNC
   static inline void run(const Derived &mat, Visitor& visitor)
   {
     return visitor.init(mat.coeff(0, 0), 0, 0);
@@ -41,7 +43,7 @@ struct visitor_impl<Visitor, Derived, 1>
 template<typename Visitor, typename Derived>
 struct visitor_impl<Visitor, Derived, Dynamic>
 {
-  typedef typename Derived::Index Index;
+  EIGEN_DEVICE_FUNC
   static inline void run(const Derived& mat, Visitor& visitor)
   {
     visitor.init(mat.coeff(0,0), 0, 0);
@@ -53,6 +55,33 @@ struct visitor_impl<Visitor, Derived, Dynamic>
   }
 };
 
+// evaluator adaptor
+template<typename XprType>
+class visitor_evaluator
+{
+public:
+  EIGEN_DEVICE_FUNC
+  explicit visitor_evaluator(const XprType &xpr) : m_evaluator(xpr), m_xpr(xpr) {}
+  
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  
+  enum {
+    RowsAtCompileTime = XprType::RowsAtCompileTime,
+    CoeffReadCost = internal::evaluator<XprType>::CoeffReadCost
+  };
+  
+  EIGEN_DEVICE_FUNC Index rows() const { return m_xpr.rows(); }
+  EIGEN_DEVICE_FUNC Index cols() const { return m_xpr.cols(); }
+  EIGEN_DEVICE_FUNC Index size() const { return m_xpr.size(); }
+
+  EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index row, Index col) const
+  { return m_evaluator.coeff(row, col); }
+  
+protected:
+  internal::evaluator<XprType> m_evaluator;
+  const XprType &m_xpr;
+};
 } // end namespace internal
 
 /** Applies the visitor \a visitor to the whole coefficients of the matrix or vector.
@@ -74,16 +103,20 @@ struct visitor_impl<Visitor, Derived, Dynamic>
   */
 template<typename Derived>
 template<typename Visitor>
+EIGEN_DEVICE_FUNC
 void DenseBase<Derived>::visit(Visitor& visitor) const
 {
-  enum { unroll = SizeAtCompileTime != Dynamic
-                   && CoeffReadCost != Dynamic
-                   && (SizeAtCompileTime == 1 || internal::functor_traits<Visitor>::Cost != Dynamic)
-                   && SizeAtCompileTime * CoeffReadCost + (SizeAtCompileTime-1) * internal::functor_traits<Visitor>::Cost
-                      <= EIGEN_UNROLLING_LIMIT };
-  return internal::visitor_impl<Visitor, Derived,
+  typedef typename internal::visitor_evaluator<Derived> ThisEvaluator;
+  ThisEvaluator thisEval(derived());
+  
+  enum { unroll =   SizeAtCompileTime != Dynamic
+                &&  ThisEvaluator::CoeffReadCost != Dynamic
+                &&  (SizeAtCompileTime == 1 || internal::functor_traits<Visitor>::Cost != Dynamic)
+                &&  SizeAtCompileTime * ThisEvaluator::CoeffReadCost + (SizeAtCompileTime-1) * internal::functor_traits<Visitor>::Cost
+                <= EIGEN_UNROLLING_LIMIT };
+  return internal::visitor_impl<Visitor, ThisEvaluator,
       unroll ? int(SizeAtCompileTime) : Dynamic
-    >::run(derived(), visitor);
+    >::run(thisEval, visitor);
 }
 
 namespace internal {
@@ -94,10 +127,10 @@ namespace internal {
 template <typename Derived>
 struct coeff_visitor
 {
-  typedef typename Derived::Index Index;
   typedef typename Derived::Scalar Scalar;
   Index row, col;
   Scalar res;
+  EIGEN_DEVICE_FUNC
   inline void init(const Scalar& value, Index i, Index j)
   {
     res = value;
@@ -114,8 +147,8 @@ struct coeff_visitor
 template <typename Derived>
 struct min_coeff_visitor : coeff_visitor<Derived>
 {
-  typedef typename Derived::Index Index;
   typedef typename Derived::Scalar Scalar;
+  EIGEN_DEVICE_FUNC
   void operator() (const Scalar& value, Index i, Index j)
   {
     if(value < this->res)
@@ -142,8 +175,8 @@ struct functor_traits<min_coeff_visitor<Scalar> > {
 template <typename Derived>
 struct max_coeff_visitor : coeff_visitor<Derived>
 {
-  typedef typename Derived::Index Index;
-  typedef typename Derived::Scalar Scalar;
+  typedef typename Derived::Scalar Scalar; 
+  EIGEN_DEVICE_FUNC
   void operator() (const Scalar& value, Index i, Index j)
   {
     if(value > this->res)
@@ -171,6 +204,7 @@ struct functor_traits<max_coeff_visitor<Scalar> > {
   */
 template<typename Derived>
 template<typename IndexType>
+EIGEN_DEVICE_FUNC
 typename internal::traits<Derived>::Scalar
 DenseBase<Derived>::minCoeff(IndexType* rowId, IndexType* colId) const
 {
@@ -188,13 +222,14 @@ DenseBase<Derived>::minCoeff(IndexType* rowId, IndexType* colId) const
   */
 template<typename Derived>
 template<typename IndexType>
+EIGEN_DEVICE_FUNC
 typename internal::traits<Derived>::Scalar
 DenseBase<Derived>::minCoeff(IndexType* index) const
 {
   EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
   internal::min_coeff_visitor<Derived> minVisitor;
   this->visit(minVisitor);
-  *index = (RowsAtCompileTime==1) ? minVisitor.col : minVisitor.row;
+  *index = IndexType((RowsAtCompileTime==1) ? minVisitor.col : minVisitor.row);
   return minVisitor.res;
 }
 
@@ -205,6 +240,7 @@ DenseBase<Derived>::minCoeff(IndexType* index) const
   */
 template<typename Derived>
 template<typename IndexType>
+EIGEN_DEVICE_FUNC
 typename internal::traits<Derived>::Scalar
 DenseBase<Derived>::maxCoeff(IndexType* rowPtr, IndexType* colPtr) const
 {
@@ -222,6 +258,7 @@ DenseBase<Derived>::maxCoeff(IndexType* rowPtr, IndexType* colPtr) const
   */
 template<typename Derived>
 template<typename IndexType>
+EIGEN_DEVICE_FUNC
 typename internal::traits<Derived>::Scalar
 DenseBase<Derived>::maxCoeff(IndexType* index) const
 {

Eigen/src/Core/arch/AVX/CMakeLists.txt

@@ -0,0 +1,6 @@
+FILE(GLOB Eigen_Core_arch_AVX_SRCS "*.h")
+
+INSTALL(FILES
+  ${Eigen_Core_arch_AVX_SRCS}
+  DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/Core/arch/AVX COMPONENT Devel
+)

Eigen/src/Core/arch/AVX/Complex.h

@@ -0,0 +1,463 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner (benoit.steiner.goog@gmail.com)
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_COMPLEX_AVX_H
+#define EIGEN_COMPLEX_AVX_H
+
+namespace Eigen {
+
+namespace internal {
+
+//---------- float ----------
+struct Packet4cf
+{
+  EIGEN_STRONG_INLINE Packet4cf() {}
+  EIGEN_STRONG_INLINE explicit Packet4cf(const __m256& a) : v(a) {}
+  __m256  v;
+};
+
+template<> struct packet_traits<std::complex<float> >  : default_packet_traits
+{
+  typedef Packet4cf type;
+  typedef Packet2cf half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size = 4,
+    HasHalfPacket = 1,
+
+    HasAdd    = 1,
+    HasSub    = 1,
+    HasMul    = 1,
+    HasDiv    = 1,
+    HasNegate = 1,
+    HasAbs    = 0,
+    HasAbs2   = 0,
+    HasMin    = 0,
+    HasMax    = 0,
+    HasSetLinear = 0
+  };
+};
+
+template<> struct unpacket_traits<Packet4cf> { typedef std::complex<float> type; enum {size=4, alignment=Aligned32}; typedef Packet2cf half; };
+
+template<> EIGEN_STRONG_INLINE Packet4cf padd<Packet4cf>(const Packet4cf& a, const Packet4cf& b) { return Packet4cf(_mm256_add_ps(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet4cf psub<Packet4cf>(const Packet4cf& a, const Packet4cf& b) { return Packet4cf(_mm256_sub_ps(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet4cf pnegate(const Packet4cf& a)
+{
+  return Packet4cf(pnegate(a.v));
+}
+template<> EIGEN_STRONG_INLINE Packet4cf pconj(const Packet4cf& a)
+{
+  const __m256 mask = _mm256_castsi256_ps(_mm256_setr_epi32(0x00000000,0x80000000,0x00000000,0x80000000,0x00000000,0x80000000,0x00000000,0x80000000));
+  return Packet4cf(_mm256_xor_ps(a.v,mask));
+}
+
+template<> EIGEN_STRONG_INLINE Packet4cf pmul<Packet4cf>(const Packet4cf& a, const Packet4cf& b)
+{
+  __m256 tmp1 = _mm256_mul_ps(_mm256_moveldup_ps(a.v), b.v);
+  __m256 tmp2 = _mm256_mul_ps(_mm256_movehdup_ps(a.v), _mm256_permute_ps(b.v, _MM_SHUFFLE(2,3,0,1)));
+  __m256 result = _mm256_addsub_ps(tmp1, tmp2);
+  return Packet4cf(result);
+}
+
+template<> EIGEN_STRONG_INLINE Packet4cf pand   <Packet4cf>(const Packet4cf& a, const Packet4cf& b) { return Packet4cf(_mm256_and_ps(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet4cf por    <Packet4cf>(const Packet4cf& a, const Packet4cf& b) { return Packet4cf(_mm256_or_ps(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet4cf pxor   <Packet4cf>(const Packet4cf& a, const Packet4cf& b) { return Packet4cf(_mm256_xor_ps(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet4cf pandnot<Packet4cf>(const Packet4cf& a, const Packet4cf& b) { return Packet4cf(_mm256_andnot_ps(a.v,b.v)); }
+
+template<> EIGEN_STRONG_INLINE Packet4cf pload <Packet4cf>(const std::complex<float>* from) { EIGEN_DEBUG_ALIGNED_LOAD return Packet4cf(pload<Packet8f>(&numext::real_ref(*from))); }
+template<> EIGEN_STRONG_INLINE Packet4cf ploadu<Packet4cf>(const std::complex<float>* from) { EIGEN_DEBUG_UNALIGNED_LOAD return Packet4cf(ploadu<Packet8f>(&numext::real_ref(*from))); }
+
+
+template<> EIGEN_STRONG_INLINE Packet4cf pset1<Packet4cf>(const std::complex<float>& from)
+{
+  return Packet4cf(_mm256_castpd_ps(_mm256_broadcast_sd((const double*)(const void*)&from)));
+}
+
+template<> EIGEN_STRONG_INLINE Packet4cf ploaddup<Packet4cf>(const std::complex<float>* from)
+{
+  // FIXME The following might be optimized using _mm256_movedup_pd
+  Packet2cf a = ploaddup<Packet2cf>(from);
+  Packet2cf b = ploaddup<Packet2cf>(from+1);
+  return  Packet4cf(_mm256_insertf128_ps(_mm256_castps128_ps256(a.v), b.v, 1));
+}
+
+template<> EIGEN_STRONG_INLINE void pstore <std::complex<float> >(std::complex<float>* to, const Packet4cf& from) { EIGEN_DEBUG_ALIGNED_STORE pstore(&numext::real_ref(*to), from.v); }
+template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<float> >(std::complex<float>* to, const Packet4cf& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(&numext::real_ref(*to), from.v); }
+
+template<> EIGEN_DEVICE_FUNC inline Packet4cf pgather<std::complex<float>, Packet4cf>(const std::complex<float>* from, Index stride)
+{
+  return Packet4cf(_mm256_set_ps(std::imag(from[3*stride]), std::real(from[3*stride]),
+                                 std::imag(from[2*stride]), std::real(from[2*stride]),
+                                 std::imag(from[1*stride]), std::real(from[1*stride]),
+                                 std::imag(from[0*stride]), std::real(from[0*stride])));
+}
+
+template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<float>, Packet4cf>(std::complex<float>* to, const Packet4cf& from, Index stride)
+{
+  __m128 low = _mm256_extractf128_ps(from.v, 0);
+  to[stride*0] = std::complex<float>(_mm_cvtss_f32(_mm_shuffle_ps(low, low, 0)),
+                                     _mm_cvtss_f32(_mm_shuffle_ps(low, low, 1)));
+  to[stride*1] = std::complex<float>(_mm_cvtss_f32(_mm_shuffle_ps(low, low, 2)),
+                                     _mm_cvtss_f32(_mm_shuffle_ps(low, low, 3)));
+
+  __m128 high = _mm256_extractf128_ps(from.v, 1);
+  to[stride*2] = std::complex<float>(_mm_cvtss_f32(_mm_shuffle_ps(high, high, 0)),
+                                     _mm_cvtss_f32(_mm_shuffle_ps(high, high, 1)));
+  to[stride*3] = std::complex<float>(_mm_cvtss_f32(_mm_shuffle_ps(high, high, 2)),
+                                     _mm_cvtss_f32(_mm_shuffle_ps(high, high, 3)));
+
+}
+
+template<> EIGEN_STRONG_INLINE std::complex<float>  pfirst<Packet4cf>(const Packet4cf& a)
+{
+  return pfirst(Packet2cf(_mm256_castps256_ps128(a.v)));
+}
+
+template<> EIGEN_STRONG_INLINE Packet4cf preverse(const Packet4cf& a) {
+  __m128 low  = _mm256_extractf128_ps(a.v, 0);
+  __m128 high = _mm256_extractf128_ps(a.v, 1);
+  __m128d lowd  = _mm_castps_pd(low);
+  __m128d highd = _mm_castps_pd(high);
+  low  = _mm_castpd_ps(_mm_shuffle_pd(lowd,lowd,0x1));
+  high = _mm_castpd_ps(_mm_shuffle_pd(highd,highd,0x1));
+  __m256 result = _mm256_setzero_ps();
+  result = _mm256_insertf128_ps(result, low, 1);
+  result = _mm256_insertf128_ps(result, high, 0);
+  return Packet4cf(result);
+}
+
+template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet4cf>(const Packet4cf& a)
+{
+  return predux(padd(Packet2cf(_mm256_extractf128_ps(a.v,0)),
+                     Packet2cf(_mm256_extractf128_ps(a.v,1))));
+}
+
+template<> EIGEN_STRONG_INLINE Packet4cf preduxp<Packet4cf>(const Packet4cf* vecs)
+{
+  Packet8f t0 = _mm256_shuffle_ps(vecs[0].v, vecs[0].v, _MM_SHUFFLE(3, 1, 2 ,0));
+  Packet8f t1 = _mm256_shuffle_ps(vecs[1].v, vecs[1].v, _MM_SHUFFLE(3, 1, 2 ,0));
+  t0 = _mm256_hadd_ps(t0,t1);
+  Packet8f t2 = _mm256_shuffle_ps(vecs[2].v, vecs[2].v, _MM_SHUFFLE(3, 1, 2 ,0));
+  Packet8f t3 = _mm256_shuffle_ps(vecs[3].v, vecs[3].v, _MM_SHUFFLE(3, 1, 2 ,0));
+  t2 = _mm256_hadd_ps(t2,t3);
+  
+  t1 = _mm256_permute2f128_ps(t0,t2, 0 + (2<<4));
+  t3 = _mm256_permute2f128_ps(t0,t2, 1 + (3<<4));
+
+  return Packet4cf(_mm256_add_ps(t1,t3));
+}
+
+template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet4cf>(const Packet4cf& a)
+{
+  return predux_mul(pmul(Packet2cf(_mm256_extractf128_ps(a.v, 0)),
+                         Packet2cf(_mm256_extractf128_ps(a.v, 1))));
+}
+
+template<int Offset>
+struct palign_impl<Offset,Packet4cf>
+{
+  static EIGEN_STRONG_INLINE void run(Packet4cf& first, const Packet4cf& second)
+  {
+    if (Offset==0) return;
+    palign_impl<Offset*2,Packet8f>::run(first.v, second.v);
+  }
+};
+
+template<> struct conj_helper<Packet4cf, Packet4cf, false,true>
+{
+  EIGEN_STRONG_INLINE Packet4cf pmadd(const Packet4cf& x, const Packet4cf& y, const Packet4cf& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet4cf pmul(const Packet4cf& a, const Packet4cf& b) const
+  {
+    return internal::pmul(a, pconj(b));
+  }
+};
+
+template<> struct conj_helper<Packet4cf, Packet4cf, true,false>
+{
+  EIGEN_STRONG_INLINE Packet4cf pmadd(const Packet4cf& x, const Packet4cf& y, const Packet4cf& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet4cf pmul(const Packet4cf& a, const Packet4cf& b) const
+  {
+    return internal::pmul(pconj(a), b);
+  }
+};
+
+template<> struct conj_helper<Packet4cf, Packet4cf, true,true>
+{
+  EIGEN_STRONG_INLINE Packet4cf pmadd(const Packet4cf& x, const Packet4cf& y, const Packet4cf& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet4cf pmul(const Packet4cf& a, const Packet4cf& b) const
+  {
+    return pconj(internal::pmul(a, b));
+  }
+};
+
+template<> struct conj_helper<Packet8f, Packet4cf, false,false>
+{
+  EIGEN_STRONG_INLINE Packet4cf pmadd(const Packet8f& x, const Packet4cf& y, const Packet4cf& c) const
+  { return padd(c, pmul(x,y)); }
+
+  EIGEN_STRONG_INLINE Packet4cf pmul(const Packet8f& x, const Packet4cf& y) const
+  { return Packet4cf(Eigen::internal::pmul(x, y.v)); }
+};
+
+template<> struct conj_helper<Packet4cf, Packet8f, false,false>
+{
+  EIGEN_STRONG_INLINE Packet4cf pmadd(const Packet4cf& x, const Packet8f& y, const Packet4cf& c) const
+  { return padd(c, pmul(x,y)); }
+
+  EIGEN_STRONG_INLINE Packet4cf pmul(const Packet4cf& x, const Packet8f& y) const
+  { return Packet4cf(Eigen::internal::pmul(x.v, y)); }
+};
+
+template<> EIGEN_STRONG_INLINE Packet4cf pdiv<Packet4cf>(const Packet4cf& a, const Packet4cf& b)
+{
+  Packet4cf num = pmul(a, pconj(b));
+  __m256 tmp = _mm256_mul_ps(b.v, b.v);
+  __m256 tmp2    = _mm256_shuffle_ps(tmp,tmp,0xB1);
+  __m256 denom = _mm256_add_ps(tmp, tmp2);
+  return Packet4cf(_mm256_div_ps(num.v, denom));
+}
+
+template<> EIGEN_STRONG_INLINE Packet4cf pcplxflip<Packet4cf>(const Packet4cf& x)
+{
+  return Packet4cf(_mm256_shuffle_ps(x.v, x.v, _MM_SHUFFLE(2, 3, 0 ,1)));
+}
+
+//---------- double ----------
+struct Packet2cd
+{
+  EIGEN_STRONG_INLINE Packet2cd() {}
+  EIGEN_STRONG_INLINE explicit Packet2cd(const __m256d& a) : v(a) {}
+  __m256d  v;
+};
+
+template<> struct packet_traits<std::complex<double> >  : default_packet_traits
+{
+  typedef Packet2cd type;
+  typedef Packet1cd half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 0,
+    size = 2,
+    HasHalfPacket = 1,
+
+    HasAdd    = 1,
+    HasSub    = 1,
+    HasMul    = 1,
+    HasDiv    = 1,
+    HasNegate = 1,
+    HasAbs    = 0,
+    HasAbs2   = 0,
+    HasMin    = 0,
+    HasMax    = 0,
+    HasSetLinear = 0
+  };
+};
+
+template<> struct unpacket_traits<Packet2cd> { typedef std::complex<double> type; enum {size=2, alignment=Aligned32}; typedef Packet1cd half; };
+
+template<> EIGEN_STRONG_INLINE Packet2cd padd<Packet2cd>(const Packet2cd& a, const Packet2cd& b) { return Packet2cd(_mm256_add_pd(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cd psub<Packet2cd>(const Packet2cd& a, const Packet2cd& b) { return Packet2cd(_mm256_sub_pd(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cd pnegate(const Packet2cd& a) { return Packet2cd(pnegate(a.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cd pconj(const Packet2cd& a)
+{
+  const __m256d mask = _mm256_castsi256_pd(_mm256_set_epi32(0x80000000,0x0,0x0,0x0,0x80000000,0x0,0x0,0x0));
+  return Packet2cd(_mm256_xor_pd(a.v,mask));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cd pmul<Packet2cd>(const Packet2cd& a, const Packet2cd& b)
+{
+  __m256d tmp1 = _mm256_shuffle_pd(a.v,a.v,0x0);
+  __m256d even = _mm256_mul_pd(tmp1, b.v);
+  __m256d tmp2 = _mm256_shuffle_pd(a.v,a.v,0xF);
+  __m256d tmp3 = _mm256_shuffle_pd(b.v,b.v,0x5);
+  __m256d odd  = _mm256_mul_pd(tmp2, tmp3);
+  return Packet2cd(_mm256_addsub_pd(even, odd));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cd pand   <Packet2cd>(const Packet2cd& a, const Packet2cd& b) { return Packet2cd(_mm256_and_pd(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cd por    <Packet2cd>(const Packet2cd& a, const Packet2cd& b) { return Packet2cd(_mm256_or_pd(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cd pxor   <Packet2cd>(const Packet2cd& a, const Packet2cd& b) { return Packet2cd(_mm256_xor_pd(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cd pandnot<Packet2cd>(const Packet2cd& a, const Packet2cd& b) { return Packet2cd(_mm256_andnot_pd(a.v,b.v)); }
+
+template<> EIGEN_STRONG_INLINE Packet2cd pload <Packet2cd>(const std::complex<double>* from)
+{ EIGEN_DEBUG_ALIGNED_LOAD return Packet2cd(pload<Packet4d>((const double*)from)); }
+template<> EIGEN_STRONG_INLINE Packet2cd ploadu<Packet2cd>(const std::complex<double>* from)
+{ EIGEN_DEBUG_UNALIGNED_LOAD return Packet2cd(ploadu<Packet4d>((const double*)from)); }
+
+template<> EIGEN_STRONG_INLINE Packet2cd pset1<Packet2cd>(const std::complex<double>& from)
+{
+  // in case casting to a __m128d* is really not safe, then we can still fallback to this version: (much slower though)
+//   return Packet2cd(_mm256_loadu2_m128d((const double*)&from,(const double*)&from));
+    return Packet2cd(_mm256_broadcast_pd((const __m128d*)(const void*)&from));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cd ploaddup<Packet2cd>(const std::complex<double>* from) { return pset1<Packet2cd>(*from); }
+
+template<> EIGEN_STRONG_INLINE void pstore <std::complex<double> >(std::complex<double> *   to, const Packet2cd& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((double*)to, from.v); }
+template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<double> >(std::complex<double> *   to, const Packet2cd& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((double*)to, from.v); }
+
+template<> EIGEN_DEVICE_FUNC inline Packet2cd pgather<std::complex<double>, Packet2cd>(const std::complex<double>* from, Index stride)
+{
+  return Packet2cd(_mm256_set_pd(std::imag(from[1*stride]), std::real(from[1*stride]),
+				 std::imag(from[0*stride]), std::real(from[0*stride])));
+}
+
+template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<double>, Packet2cd>(std::complex<double>* to, const Packet2cd& from, Index stride)
+{
+  __m128d low = _mm256_extractf128_pd(from.v, 0);
+  to[stride*0] = std::complex<double>(_mm_cvtsd_f64(low), _mm_cvtsd_f64(_mm_shuffle_pd(low, low, 1)));
+  __m128d high = _mm256_extractf128_pd(from.v, 1);
+  to[stride*1] = std::complex<double>(_mm_cvtsd_f64(high), _mm_cvtsd_f64(_mm_shuffle_pd(high, high, 1)));
+}
+
+template<> EIGEN_STRONG_INLINE std::complex<double> pfirst<Packet2cd>(const Packet2cd& a)
+{
+  __m128d low = _mm256_extractf128_pd(a.v, 0);
+  EIGEN_ALIGN16 double res[2];
+  _mm_store_pd(res, low);
+  return std::complex<double>(res[0],res[1]);
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cd preverse(const Packet2cd& a) {
+  __m256d result = _mm256_permute2f128_pd(a.v, a.v, 1);
+  return Packet2cd(result);
+}
+
+template<> EIGEN_STRONG_INLINE std::complex<double> predux<Packet2cd>(const Packet2cd& a)
+{
+  return predux(padd(Packet1cd(_mm256_extractf128_pd(a.v,0)),
+                     Packet1cd(_mm256_extractf128_pd(a.v,1))));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cd preduxp<Packet2cd>(const Packet2cd* vecs)
+{
+  Packet4d t0 = _mm256_permute2f128_pd(vecs[0].v,vecs[1].v, 0 + (2<<4));
+  Packet4d t1 = _mm256_permute2f128_pd(vecs[0].v,vecs[1].v, 1 + (3<<4));
+
+  return Packet2cd(_mm256_add_pd(t0,t1));
+}
+
+template<> EIGEN_STRONG_INLINE std::complex<double> predux_mul<Packet2cd>(const Packet2cd& a)
+{
+  return predux(pmul(Packet1cd(_mm256_extractf128_pd(a.v,0)),
+                     Packet1cd(_mm256_extractf128_pd(a.v,1))));
+}
+
+template<int Offset>
+struct palign_impl<Offset,Packet2cd>
+{
+  static EIGEN_STRONG_INLINE void run(Packet2cd& first, const Packet2cd& second)
+  {
+    if (Offset==0) return;
+    palign_impl<Offset*2,Packet4d>::run(first.v, second.v);
+  }
+};
+
+template<> struct conj_helper<Packet2cd, Packet2cd, false,true>
+{
+  EIGEN_STRONG_INLINE Packet2cd pmadd(const Packet2cd& x, const Packet2cd& y, const Packet2cd& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet2cd pmul(const Packet2cd& a, const Packet2cd& b) const
+  {
+    return internal::pmul(a, pconj(b));
+  }
+};
+
+template<> struct conj_helper<Packet2cd, Packet2cd, true,false>
+{
+  EIGEN_STRONG_INLINE Packet2cd pmadd(const Packet2cd& x, const Packet2cd& y, const Packet2cd& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet2cd pmul(const Packet2cd& a, const Packet2cd& b) const
+  {
+    return internal::pmul(pconj(a), b);
+  }
+};
+
+template<> struct conj_helper<Packet2cd, Packet2cd, true,true>
+{
+  EIGEN_STRONG_INLINE Packet2cd pmadd(const Packet2cd& x, const Packet2cd& y, const Packet2cd& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet2cd pmul(const Packet2cd& a, const Packet2cd& b) const
+  {
+    return pconj(internal::pmul(a, b));
+  }
+};
+
+template<> struct conj_helper<Packet4d, Packet2cd, false,false>
+{
+  EIGEN_STRONG_INLINE Packet2cd pmadd(const Packet4d& x, const Packet2cd& y, const Packet2cd& c) const
+  { return padd(c, pmul(x,y)); }
+
+  EIGEN_STRONG_INLINE Packet2cd pmul(const Packet4d& x, const Packet2cd& y) const
+  { return Packet2cd(Eigen::internal::pmul(x, y.v)); }
+};
+
+template<> struct conj_helper<Packet2cd, Packet4d, false,false>
+{
+  EIGEN_STRONG_INLINE Packet2cd pmadd(const Packet2cd& x, const Packet4d& y, const Packet2cd& c) const
+  { return padd(c, pmul(x,y)); }
+
+  EIGEN_STRONG_INLINE Packet2cd pmul(const Packet2cd& x, const Packet4d& y) const
+  { return Packet2cd(Eigen::internal::pmul(x.v, y)); }
+};
+
+template<> EIGEN_STRONG_INLINE Packet2cd pdiv<Packet2cd>(const Packet2cd& a, const Packet2cd& b)
+{
+  Packet2cd num = pmul(a, pconj(b));
+  __m256d tmp = _mm256_mul_pd(b.v, b.v);
+  __m256d denom = _mm256_hadd_pd(tmp, tmp);
+  return Packet2cd(_mm256_div_pd(num.v, denom));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cd pcplxflip<Packet2cd>(const Packet2cd& x)
+{
+  return Packet2cd(_mm256_shuffle_pd(x.v, x.v, 0x5));
+}
+
+EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<Packet4cf,4>& kernel) {
+  __m256d P0 = _mm256_castps_pd(kernel.packet[0].v);
+  __m256d P1 = _mm256_castps_pd(kernel.packet[1].v);
+  __m256d P2 = _mm256_castps_pd(kernel.packet[2].v);
+  __m256d P3 = _mm256_castps_pd(kernel.packet[3].v);
+
+  __m256d T0 = _mm256_shuffle_pd(P0, P1, 15);
+  __m256d T1 = _mm256_shuffle_pd(P0, P1, 0);
+  __m256d T2 = _mm256_shuffle_pd(P2, P3, 15);
+  __m256d T3 = _mm256_shuffle_pd(P2, P3, 0);
+
+  kernel.packet[1].v = _mm256_castpd_ps(_mm256_permute2f128_pd(T0, T2, 32));
+  kernel.packet[3].v = _mm256_castpd_ps(_mm256_permute2f128_pd(T0, T2, 49));
+  kernel.packet[0].v = _mm256_castpd_ps(_mm256_permute2f128_pd(T1, T3, 32));
+  kernel.packet[2].v = _mm256_castpd_ps(_mm256_permute2f128_pd(T1, T3, 49));
+}
+
+EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<Packet2cd,2>& kernel) {
+  __m256d tmp = _mm256_permute2f128_pd(kernel.packet[0].v, kernel.packet[1].v, 0+(2<<4));
+  kernel.packet[1].v = _mm256_permute2f128_pd(kernel.packet[0].v, kernel.packet[1].v, 1+(3<<4));
+ kernel.packet[0].v = tmp;
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_COMPLEX_AVX_H