bump to 3.3.3

(grafted from cbbf88c4d7
)

CMakeLists.txt

@@ -1,4 +1,4 @@
-project(Eigen)
+project(Eigen3)
 
 cmake_minimum_required(VERSION 2.8.5)
 
@@ -8,6 +8,11 @@ if(${CMAKE_SOURCE_DIR} STREQUAL ${CMAKE_BINARY_DIR})
   message(FATAL_ERROR "In-source builds not allowed. Please make a new directory (called a build directory) and run CMake from there. You may need to remove CMakeCache.txt. ")
 endif()
 
+# Alias Eigen_*_DIR to Eigen3_*_DIR:
+
+set(Eigen_SOURCE_DIR ${Eigen3_SOURCE_DIR})
+set(Eigen_BINARY_DIR ${Eigen3_BINARY_DIR})
+
 # guard against bad build-type strings
 
 if (NOT CMAKE_BUILD_TYPE)
@@ -93,9 +98,11 @@ else()
 endif()
 
 option(EIGEN_BUILD_BTL "Build benchmark suite" OFF)
-if(NOT WIN32)
+
+# Disable pkgconfig only for native Windows builds
+if(NOT WIN32 OR NOT CMAKE_HOST_SYSTEM_NAME MATCHES Windows)
   option(EIGEN_BUILD_PKGCONFIG "Build pkg-config .pc file for Eigen" ON)
-endif(NOT WIN32)
+endif()
 
 set(CMAKE_INCLUDE_CURRENT_DIR ON)
 
@@ -126,7 +133,6 @@ if(NOT MSVC)
   if(COMPILER_SUPPORT_WERROR)
     set(CMAKE_REQUIRED_FLAGS "-Werror")
   endif()
-  
   ei_add_cxx_compiler_flag("-pedantic")
   ei_add_cxx_compiler_flag("-Wall")
   ei_add_cxx_compiler_flag("-Wextra")
@@ -224,6 +230,12 @@ if(NOT MSVC)
     message(STATUS "Enabling FMA in tests/examples")
   endif()
 
+  option(EIGEN_TEST_AVX512 "Enable/Disable AVX512 in tests/examples" OFF)
+  if(EIGEN_TEST_AVX512)
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mavx512f -fabi-version=6 -DEIGEN_ENABLE_AVX512")
+    message(STATUS "Enabling AVX512 in tests/examples")
+  endif()
+
   option(EIGEN_TEST_F16C "Enable/Disable F16C in tests/examples" OFF)
   if(EIGEN_TEST_F16C)
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mf16c")
@@ -368,7 +380,7 @@ else()
       )
 endif()
 set(CMAKEPACKAGE_INSTALL_DIR
-    "${CMAKE_INSTALL_LIBDIR}/cmake/eigen3"
+    "${CMAKE_INSTALL_DATADIR}/eigen3/cmake"
     CACHE PATH "The directory relative to CMAKE_PREFIX_PATH where Eigen3Config.cmake is installed"
     )
 set(PKGCONFIG_INSTALL_DIR
@@ -398,7 +410,7 @@ if(EIGEN_BUILD_PKGCONFIG)
     install(FILES ${CMAKE_CURRENT_BINARY_DIR}/eigen3.pc
         DESTINATION ${PKGCONFIG_INSTALL_DIR}
         )
-endif(EIGEN_BUILD_PKGCONFIG)
+endif()
 
 add_subdirectory(Eigen)
 
@@ -424,6 +436,13 @@ else()
   add_subdirectory(lapack EXCLUDE_FROM_ALL)
 endif()
 
+# add SYCL
+option(EIGEN_TEST_SYCL "Add Sycl support." OFF)
+if(EIGEN_TEST_SYCL)
+  set (CMAKE_MODULE_PATH "${CMAKE_ROOT}/Modules" "cmake/Modules/" "${CMAKE_MODULE_PATH}")
+  include(FindComputeCpp)
+endif()
+
 add_subdirectory(unsupported)
 
 add_subdirectory(demos EXCLUDE_FROM_ALL)
@@ -488,18 +507,89 @@ set ( EIGEN_VERSION_MINOR  ${EIGEN_MAJOR_VERSION} )
 set ( EIGEN_VERSION_PATCH  ${EIGEN_MINOR_VERSION} )
 set ( EIGEN_DEFINITIONS "")
 set ( EIGEN_INCLUDE_DIR "${CMAKE_INSTALL_PREFIX}/${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
-               )
+# Interface libraries require at least CMake 3.0
+if (NOT CMAKE_VERSION VERSION_LESS 3.0)
+  include (CMakePackageConfigHelpers)
+
+  # Imported target support
+  add_library (eigen INTERFACE)
+
+  target_compile_definitions (eigen INTERFACE ${EIGEN_DEFINITIONS})
+  target_include_directories (eigen INTERFACE
+    $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}>
+    $<INSTALL_INTERFACE:${INCLUDE_INSTALL_DIR}>
+  )
+
+  # Export as title case Eigen
+  set_target_properties (eigen PROPERTIES EXPORT_NAME Eigen)
+
+  install (TARGETS eigen EXPORT Eigen3Targets)
+
+  configure_package_config_file (
+    ${CMAKE_CURRENT_SOURCE_DIR}/cmake/Eigen3Config.cmake.in
+    ${CMAKE_CURRENT_BINARY_DIR}/Eigen3Config.cmake
+    PATH_VARS EIGEN_INCLUDE_DIR EIGEN_ROOT_DIR
+    INSTALL_DESTINATION ${CMAKEPACKAGE_INSTALL_DIR}
+    NO_CHECK_REQUIRED_COMPONENTS_MACRO # Eigen does not provide components
+  )
+  # Remove CMAKE_SIZEOF_VOID_P from Eigen3ConfigVersion.cmake since Eigen does
+  # not depend on architecture specific settings or libraries. More
+  # specifically, an Eigen3Config.cmake generated from a 64 bit target can be
+  # used for 32 bit targets as well (and vice versa).
+  set (_Eigen3_CMAKE_SIZEOF_VOID_P ${CMAKE_SIZEOF_VOID_P})
+  unset (CMAKE_SIZEOF_VOID_P)
+  write_basic_package_version_file (Eigen3ConfigVersion.cmake
+                                    VERSION ${EIGEN_VERSION_NUMBER}
+                                    COMPATIBILITY SameMajorVersion)
+  set (CMAKE_SIZEOF_VOID_P ${_Eigen3_CMAKE_SIZEOF_VOID_P})
+
+  # The Eigen target will be located in the Eigen3 namespace. Other CMake
+  # targets can refer to it using Eigen3::Eigen.
+  export (TARGETS eigen NAMESPACE Eigen3:: FILE Eigen3Targets.cmake)
+  # Export Eigen3 package to CMake registry such that it can be easily found by
+  # CMake even if it has not been installed to a standard directory.
+  export (PACKAGE Eigen3)
+
+  install (EXPORT Eigen3Targets NAMESPACE Eigen3:: DESTINATION ${CMAKEPACKAGE_INSTALL_DIR})
+
+else (NOT CMAKE_VERSION VERSION_LESS 3.0)
+  # Fallback to legacy Eigen3Config.cmake without the imported target
+  
+  # If CMakePackageConfigHelpers module is available (CMake >= 2.8.8)
+  # create a relocatable Config file, otherwise leave the hardcoded paths       
+  include(CMakePackageConfigHelpers OPTIONAL RESULT_VARIABLE CPCH_PATH)
+  
+  if(CPCH_PATH)
+    configure_package_config_file (
+      ${CMAKE_CURRENT_SOURCE_DIR}/cmake/Eigen3ConfigLegacy.cmake.in
+      ${CMAKE_CURRENT_BINARY_DIR}/Eigen3Config.cmake
+      PATH_VARS EIGEN_INCLUDE_DIR EIGEN_ROOT_DIR
+      INSTALL_DESTINATION ${CMAKEPACKAGE_INSTALL_DIR}
+      NO_CHECK_REQUIRED_COMPONENTS_MACRO # Eigen does not provide components
+    )
+  else() 
+    # The PACKAGE_* variables are defined by the configure_package_config_file
+    # but without it we define them manually to the hardcoded paths
+    set(PACKAGE_INIT "")
+    set(PACKAGE_EIGEN_INCLUDE_DIR ${EIGEN_INCLUDE_DIR})
+    set(PACKAGE_EIGEN_ROOT_DIR ${EIGEN_ROOT_DIR})
+    configure_file ( ${CMAKE_CURRENT_SOURCE_DIR}/cmake/Eigen3ConfigLegacy.cmake.in
+                     ${CMAKE_CURRENT_BINARY_DIR}/Eigen3Config.cmake
+                     @ONLY ESCAPE_QUOTES )
+  endif()
+
+  write_basic_package_version_file( Eigen3ConfigVersion.cmake
+                                    VERSION ${EIGEN_VERSION_NUMBER}
+                                    COMPATIBILITY SameMajorVersion )
+
+endif (NOT CMAKE_VERSION VERSION_LESS 3.0)
 
 install ( FILES ${CMAKE_CURRENT_SOURCE_DIR}/cmake/UseEigen3.cmake
                 ${CMAKE_CURRENT_BINARY_DIR}/Eigen3Config.cmake
-          DESTINATION ${CMAKEPACKAGE_INSTALL_DIR}
-        )
+                ${CMAKE_CURRENT_BINARY_DIR}/Eigen3ConfigVersion.cmake
+          DESTINATION ${CMAKEPACKAGE_INSTALL_DIR} )
 
 # Add uninstall target
 add_custom_target ( uninstall

CTestConfig.cmake

@@ -4,14 +4,10 @@
 ## # The following are required to uses Dart and the Cdash dashboard
 ##   ENABLE_TESTING()
 ##   INCLUDE(CTest)
-set(CTEST_PROJECT_NAME "Eigen")
+set(CTEST_PROJECT_NAME "Eigen3.3")
 set(CTEST_NIGHTLY_START_TIME "00:00:00 UTC")
 
 set(CTEST_DROP_METHOD "http")
 set(CTEST_DROP_SITE "manao.inria.fr")
-set(CTEST_DROP_LOCATION "/CDash/submit.php?project=Eigen")
+set(CTEST_DROP_LOCATION "/CDash/submit.php?project=Eigen3.3")
 set(CTEST_DROP_SITE_CDASH TRUE)
-set(CTEST_PROJECT_SUBPROJECTS
-Official
-Unsupported
-)

Eigen/CMakeLists.txt

@@ -16,4 +16,4 @@ install(FILES
   DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen COMPONENT Devel
   )
 
-add_subdirectory(src)
+install(DIRECTORY src DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen COMPONENT Devel FILES_MATCHING PATTERN "*.h")

Eigen/Core

@@ -14,9 +14,9 @@
 // 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!
+// Handle NVCC/CUDA/SYCL
+#if defined(__CUDACC__) || defined(__SYCL_DEVICE_ONLY__)
+  // Do not try asserts on CUDA and SYCL!
   #ifndef EIGEN_NO_DEBUG
   #define EIGEN_NO_DEBUG
   #endif
@@ -25,17 +25,24 @@
   #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__
+  #ifdef __CUDACC__
+    // Do not try to vectorize on CUDA and SYCL!
+    #ifndef EIGEN_DONT_VECTORIZE
+    #define EIGEN_DONT_VECTORIZE
+    #endif
+
+    #define EIGEN_DEVICE_FUNC __host__ __device__
+    // We need math_functions.hpp to ensure that that EIGEN_USING_STD_MATH macro
+    // works properly on the device side
+    #include <math_functions.hpp>
+  #else
+    #define EIGEN_DEVICE_FUNC
+  #endif
 
 #else
   #define EIGEN_DEVICE_FUNC
@@ -51,7 +58,7 @@
   #define EIGEN_USING_STD_MATH(FUNC) using std::FUNC;
 #endif
 
-#if (defined(_CPPUNWIND) || defined(__EXCEPTIONS)) && !defined(__CUDA_ARCH__) && !defined(EIGEN_EXCEPTIONS)
+#if (defined(_CPPUNWIND) || defined(__EXCEPTIONS)) && !defined(__CUDA_ARCH__) && !defined(EIGEN_EXCEPTIONS) && !defined(EIGEN_USE_SYCL)
   #define EIGEN_EXCEPTIONS
 #endif
 
@@ -140,6 +147,15 @@
     #ifdef __FMA__
       #define EIGEN_VECTORIZE_FMA
     #endif
+    #if defined(__AVX512F__) && defined(EIGEN_ENABLE_AVX512)
+      #define EIGEN_VECTORIZE_AVX512
+      #define EIGEN_VECTORIZE_AVX2
+      #define EIGEN_VECTORIZE_AVX
+      #define EIGEN_VECTORIZE_FMA
+      #ifdef __AVX512DQ__
+        #define EIGEN_VECTORIZE_AVX512DQ
+      #endif
+    #endif
 
     // include files
 
@@ -171,7 +187,7 @@
         #ifdef EIGEN_VECTORIZE_SSE4_2
         #include <nmmintrin.h>
         #endif
-        #ifdef EIGEN_VECTORIZE_AVX
+        #if defined(EIGEN_VECTORIZE_AVX) || defined(EIGEN_VECTORIZE_AVX512)
         #include <immintrin.h>
         #endif
       #endif
@@ -271,7 +287,9 @@
 namespace Eigen {
 
 inline static const char *SimdInstructionSetsInUse(void) {
-#if defined(EIGEN_VECTORIZE_AVX)
+#if defined(EIGEN_VECTORIZE_AVX512)
+  return "AVX512, FMA, AVX2, AVX, SSE, SSE2, SSE3, SSSE3, SSE4.1, SSE4.2";
+#elif 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";
@@ -303,12 +321,16 @@ inline static const char *SimdInstructionSetsInUse(void) {
 #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
 
+namespace Eigen {
+
 // we use size_t frequently and we'll never remember to prepend it with std:: everytime just to
 // ensure QNX/QCC support
 using std::size_t;
 // gcc 4.6.0 wants std:: for ptrdiff_t
 using std::ptrdiff_t;
 
+}
+
 /** \defgroup Core_Module Core module
   * This is the main module of Eigen providing dense matrix and vector support
   * (both fixed and dynamic size) with all the features corresponding to a BLAS library
@@ -329,8 +351,14 @@ using std::ptrdiff_t;
 #include "src/Core/NumTraits.h"
 #include "src/Core/MathFunctions.h"
 #include "src/Core/GenericPacketMath.h"
+#include "src/Core/MathFunctionsImpl.h"
 
-#if defined EIGEN_VECTORIZE_AVX
+#if defined EIGEN_VECTORIZE_AVX512
+  #include "src/Core/arch/SSE/PacketMath.h"
+  #include "src/Core/arch/AVX/PacketMath.h"
+  #include "src/Core/arch/AVX512/PacketMath.h"
+  #include "src/Core/arch/AVX512/MathFunctions.h"
+#elif 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"
@@ -377,6 +405,11 @@ using std::ptrdiff_t;
 #include "src/Core/functors/StlFunctors.h"
 #include "src/Core/functors/AssignmentFunctors.h"
 
+// Specialized functors to enable the processing of complex numbers
+// on CUDA devices
+#include "src/Core/arch/CUDA/Complex.h"
+
+#include "src/Core/IO.h"
 #include "src/Core/DenseCoeffsBase.h"
 #include "src/Core/DenseBase.h"
 #include "src/Core/MatrixBase.h"
@@ -424,7 +457,6 @@ using std::ptrdiff_t;
 #include "src/Core/Redux.h"
 #include "src/Core/Visitor.h"
 #include "src/Core/Fuzzy.h"
-#include "src/Core/IO.h"
 #include "src/Core/Swap.h"
 #include "src/Core/CommaInitializer.h"
 #include "src/Core/GeneralProduct.h"

Eigen/QtAlignedMalloc

@@ -14,7 +14,7 @@
 
 #include "src/Core/util/DisableStupidWarnings.h"
 
-void *qMalloc(size_t size)
+void *qMalloc(std::size_t size)
 {
   return Eigen::internal::aligned_malloc(size);
 }
@@ -24,7 +24,7 @@ void qFree(void *ptr)
   Eigen::internal::aligned_free(ptr);
 }
 
-void *qRealloc(void *ptr, size_t size)
+void *qRealloc(void *ptr, std::size_t size)
 {
   void* newPtr = Eigen::internal::aligned_malloc(size);
   memcpy(newPtr, ptr, size);

Eigen/Sparse

@@ -25,7 +25,9 @@
 
 #include "SparseCore"
 #include "OrderingMethods"
+#ifndef EIGEN_MPL2_ONLY
 #include "SparseCholesky"
+#endif
 #include "SparseLU"
 #include "SparseQR"
 #include "IterativeLinearSolvers"

Eigen/StdDeque

@@ -14,7 +14,7 @@
 #include "Core"
 #include <deque>
 
-#if EIGEN_COMP_MSVC && EIGEN_OS_WIN64 /* MSVC auto aligns in 64 bit builds */
+#if EIGEN_COMP_MSVC && EIGEN_OS_WIN64 && (EIGEN_MAX_STATIC_ALIGN_BYTES<=16) /* MSVC auto aligns up to 16 bytes in 64 bit builds */
 
 #define EIGEN_DEFINE_STL_DEQUE_SPECIALIZATION(...)
 

Eigen/StdList

@@ -13,7 +13,7 @@
 #include "Core"
 #include <list>
 
-#if EIGEN_COMP_MSVC && EIGEN_OS_WIN64 /* MSVC auto aligns in 64 bit builds */    
+#if EIGEN_COMP_MSVC && EIGEN_OS_WIN64 && (EIGEN_MAX_STATIC_ALIGN_BYTES<=16) /* MSVC auto aligns up to 16 bytes in 64 bit builds */
 
 #define EIGEN_DEFINE_STL_LIST_SPECIALIZATION(...)
 

Eigen/StdVector

@@ -14,7 +14,7 @@
 #include "Core"
 #include <vector>
 
-#if EIGEN_COMP_MSVC && EIGEN_OS_WIN64 /* MSVC auto aligns in 64 bit builds */
+#if EIGEN_COMP_MSVC && EIGEN_OS_WIN64 && (EIGEN_MAX_STATIC_ALIGN_BYTES<=16) /* MSVC auto aligns up to 16 bytes in 64 bit builds */
 
 #define EIGEN_DEFINE_STL_VECTOR_SPECIALIZATION(...)
 

Eigen/src/CMakeLists.txt

@@ -1,7 +0,0 @@
-file(GLOB Eigen_src_subdirectories "*")
-escape_string_as_regex(ESCAPED_CMAKE_CURRENT_SOURCE_DIR "${CMAKE_CURRENT_SOURCE_DIR}")
-foreach(f ${Eigen_src_subdirectories})
-  if(NOT f MATCHES "\\.txt" AND NOT f MATCHES "${ESCAPED_CMAKE_CURRENT_SOURCE_DIR}/[.].+" )
-    add_subdirectory(${f})
-  endif()
-endforeach()

Eigen/src/Cholesky/CMakeLists.txt

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

Eigen/src/Cholesky/LDLT.h

@@ -253,7 +253,7 @@ template<typename _MatrixType, int _UpLo> class LDLT
     ComputationInfo info() const
     {
       eigen_assert(m_isInitialized && "LDLT is not initialized.");
-      return Success;
+      return m_info;
     }
 
     #ifndef EIGEN_PARSED_BY_DOXYGEN
@@ -281,6 +281,7 @@ template<typename _MatrixType, int _UpLo> class LDLT
     TmpMatrixType m_temporary;
     internal::SignMatrix m_sign;
     bool m_isInitialized;
+    ComputationInfo m_info;
 };
 
 namespace internal {
@@ -298,6 +299,8 @@ template<> struct ldlt_inplace<Lower>
     typedef typename TranspositionType::StorageIndex IndexType;
     eigen_assert(mat.rows()==mat.cols());
     const Index size = mat.rows();
+    bool found_zero_pivot = false;
+    bool ret = true;
 
     if (size <= 1)
     {
@@ -356,9 +359,27 @@ template<> struct ldlt_inplace<Lower>
       // 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)))
+      bool pivot_is_valid = (abs(realAkk) > RealScalar(0));
+
+      if(k==0 && !pivot_is_valid)
+      {
+        // The entire diagonal is zero, there is nothing more to do
+        // except filling the transpositions, and checking whether the matrix is zero.
+        sign = ZeroSign;
+        for(Index j = 0; j<size; ++j)
+        {
+          transpositions.coeffRef(j) = IndexType(j);
+          ret = ret && (mat.col(j).tail(size-j-1).array()==Scalar(0)).all();
+        }
+        return ret;
+      }
+
+      if((rs>0) && pivot_is_valid)
         A21 /= realAkk;
 
+      if(found_zero_pivot && pivot_is_valid) ret = false; // factorization failed
+      else if(!pivot_is_valid) found_zero_pivot = true;
+
       if (sign == PositiveSemiDef) {
         if (realAkk < static_cast<RealScalar>(0)) sign = Indefinite;
       } else if (sign == NegativeSemiDef) {
@@ -369,7 +390,7 @@ template<> struct ldlt_inplace<Lower>
       }
     }
 
-    return true;
+    return ret;
   }
 
   // Reference for the algorithm: Davis and Hager, "Multiple Rank
@@ -493,7 +514,7 @@ LDLT<MatrixType,_UpLo>& LDLT<MatrixType,_UpLo>::compute(const EigenBase<InputTyp
   m_temporary.resize(size);
   m_sign = internal::ZeroSign;
 
-  internal::ldlt_inplace<UpLo>::unblocked(m_matrix, m_transpositions, m_temporary, m_sign);
+  m_info = internal::ldlt_inplace<UpLo>::unblocked(m_matrix, m_transpositions, m_temporary, m_sign) ? Success : NumericalIssue;
 
   m_isInitialized = true;
   return *this;
@@ -621,7 +642,6 @@ 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()
@@ -643,7 +663,6 @@ MatrixBase<Derived>::ldlt() const
 {
   return LDLT<PlainObject>(derived());
 }
-#endif // __CUDACC__
 
 } // end namespace Eigen
 

Eigen/src/Cholesky/LLT.h

@@ -351,7 +351,7 @@ template<typename Scalar> struct llt_inplace<Scalar, Lower>
       Index ret;
       if((ret=unblocked(A11))>=0) return k+ret;
       if(rs>0) A11.adjoint().template triangularView<Upper>().template solveInPlace<OnTheRight>(A21);
-      if(rs>0) A22.template selfadjointView<Lower>().rankUpdate(A21,-1); // bottleneck
+      if(rs>0) A22.template selfadjointView<Lower>().rankUpdate(A21,typename NumTraits<RealScalar>::Literal(-1)); // bottleneck
     }
     return -1;
   }
@@ -507,7 +507,6 @@ 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()
@@ -529,7 +528,6 @@ SelfAdjointView<MatrixType, UpLo>::llt() const
 {
   return LLT<PlainObject,UpLo>(m_matrix);
 }
-#endif // __CUDACC__
 
 } // end namespace Eigen
 

Eigen/src/CholmodSupport/CMakeLists.txt

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

Eigen/src/CholmodSupport/CholmodSupport.h

@@ -14,34 +14,40 @@ namespace Eigen {
 
 namespace internal {
 
-template<typename Scalar, typename CholmodType>
-void cholmod_configure_matrix(CholmodType& mat)
-{
-  if (internal::is_same<Scalar,float>::value)
-  {
-    mat.xtype = CHOLMOD_REAL;
-    mat.dtype = CHOLMOD_SINGLE;
-  }
-  else if (internal::is_same<Scalar,double>::value)
-  {
+template<typename Scalar> struct cholmod_configure_matrix;
+
+template<> struct cholmod_configure_matrix<double> {
+  template<typename CholmodType>
+  static void run(CholmodType& mat) {
     mat.xtype = CHOLMOD_REAL;
     mat.dtype = CHOLMOD_DOUBLE;
   }
-  else if (internal::is_same<Scalar,std::complex<float> >::value)
-  {
-    mat.xtype = CHOLMOD_COMPLEX;
-    mat.dtype = CHOLMOD_SINGLE;
-  }
-  else if (internal::is_same<Scalar,std::complex<double> >::value)
-  {
+};
+
+template<> struct cholmod_configure_matrix<std::complex<double> > {
+  template<typename CholmodType>
+  static void run(CholmodType& mat) {
     mat.xtype = CHOLMOD_COMPLEX;
     mat.dtype = CHOLMOD_DOUBLE;
   }
-  else
-  {
-    eigen_assert(false && "Scalar type not supported by CHOLMOD");
-  }
-}
+};
+
+// Other scalar types are not yet suppotred by Cholmod
+// template<> struct cholmod_configure_matrix<float> {
+//   template<typename CholmodType>
+//   static void run(CholmodType& mat) {
+//     mat.xtype = CHOLMOD_REAL;
+//     mat.dtype = CHOLMOD_SINGLE;
+//   }
+// };
+//
+// template<> struct cholmod_configure_matrix<std::complex<float> > {
+//   template<typename CholmodType>
+//   static void run(CholmodType& mat) {
+//     mat.xtype = CHOLMOD_COMPLEX;
+//     mat.dtype = CHOLMOD_SINGLE;
+//   }
+// };
 
 } // namespace internal
 
@@ -49,11 +55,11 @@ void cholmod_configure_matrix(CholmodType& mat)
   * Note that the data are shared.
   */
 template<typename _Scalar, int _Options, typename _StorageIndex>
-cholmod_sparse viewAsCholmod(SparseMatrix<_Scalar,_Options,_StorageIndex>& mat)
+cholmod_sparse viewAsCholmod(Ref<SparseMatrix<_Scalar,_Options,_StorageIndex> > mat)
 {
   cholmod_sparse res;
   res.nzmax   = mat.nonZeros();
-  res.nrow    = mat.rows();;
+  res.nrow    = mat.rows();
   res.ncol    = mat.cols();
   res.p       = mat.outerIndexPtr();
   res.i       = mat.innerIndexPtr();
@@ -88,7 +94,7 @@ cholmod_sparse viewAsCholmod(SparseMatrix<_Scalar,_Options,_StorageIndex>& mat)
   }
 
   // setup res.xtype
-  internal::cholmod_configure_matrix<_Scalar>(res);
+  internal::cholmod_configure_matrix<_Scalar>::run(res);
   
   res.stype = 0;
   
@@ -98,7 +104,14 @@ cholmod_sparse viewAsCholmod(SparseMatrix<_Scalar,_Options,_StorageIndex>& mat)
 template<typename _Scalar, int _Options, typename _Index>
 const cholmod_sparse viewAsCholmod(const SparseMatrix<_Scalar,_Options,_Index>& mat)
 {
-  cholmod_sparse res = viewAsCholmod(mat.const_cast_derived());
+  cholmod_sparse res = viewAsCholmod(Ref<SparseMatrix<_Scalar,_Options,_Index> >(mat.const_cast_derived()));
+  return res;
+}
+
+template<typename _Scalar, int _Options, typename _Index>
+const cholmod_sparse viewAsCholmod(const SparseVector<_Scalar,_Options,_Index>& mat)
+{
+  cholmod_sparse res = viewAsCholmod(Ref<SparseMatrix<_Scalar,_Options,_Index> >(mat.const_cast_derived()));
   return res;
 }
 
@@ -107,7 +120,7 @@ const cholmod_sparse viewAsCholmod(const SparseMatrix<_Scalar,_Options,_Index>&
 template<typename _Scalar, int _Options, typename _Index, unsigned int UpLo>
 cholmod_sparse viewAsCholmod(const SparseSelfAdjointView<const SparseMatrix<_Scalar,_Options,_Index>, UpLo>& mat)
 {
-  cholmod_sparse res = viewAsCholmod(mat.matrix().const_cast_derived());
+  cholmod_sparse res = viewAsCholmod(Ref<SparseMatrix<_Scalar,_Options,_Index> >(mat.matrix().const_cast_derived()));
   
   if(UpLo==Upper) res.stype =  1;
   if(UpLo==Lower) res.stype = -1;
@@ -131,7 +144,7 @@ cholmod_dense viewAsCholmod(MatrixBase<Derived>& mat)
   res.x      = (void*)(mat.derived().data());
   res.z      = 0;
 
-  internal::cholmod_configure_matrix<Scalar>(res);
+  internal::cholmod_configure_matrix<Scalar>::run(res);
 
   return res;
 }
@@ -180,14 +193,16 @@ class CholmodBase : public SparseSolverBase<Derived>
     CholmodBase()
       : m_cholmodFactor(0), m_info(Success), m_factorizationIsOk(false), m_analysisIsOk(false)
     {
-      m_shiftOffset[0] = m_shiftOffset[1] = RealScalar(0.0);
+      EIGEN_STATIC_ASSERT((internal::is_same<double,RealScalar>::value), CHOLMOD_SUPPORTS_DOUBLE_PRECISION_ONLY);
+      m_shiftOffset[0] = m_shiftOffset[1] = 0.0;
       cholmod_start(&m_cholmod);
     }
 
     explicit CholmodBase(const MatrixType& matrix)
       : m_cholmodFactor(0), m_info(Success), m_factorizationIsOk(false), m_analysisIsOk(false)
     {
-      m_shiftOffset[0] = m_shiftOffset[1] = RealScalar(0.0);
+      EIGEN_STATIC_ASSERT((internal::is_same<double,RealScalar>::value), CHOLMOD_SUPPORTS_DOUBLE_PRECISION_ONLY);
+      m_shiftOffset[0] = m_shiftOffset[1] = 0.0;
       cholmod_start(&m_cholmod);
       compute(matrix);
     }
@@ -254,7 +269,7 @@ class CholmodBase : public SparseSolverBase<Derived>
       eigen_assert(m_analysisIsOk && "You must first call analyzePattern()");
       cholmod_sparse A = viewAsCholmod(matrix.template selfadjointView<UpLo>());
       cholmod_factorize_p(&A, m_shiftOffset, 0, 0, m_cholmodFactor, &m_cholmod);
-      
+
       // If the factorization failed, minor is the column at which it did. On success minor == n.
       this->m_info = (m_cholmodFactor->minor == m_cholmodFactor->n ? Success : NumericalIssue);
       m_factorizationIsOk = true;
@@ -290,8 +305,8 @@ class CholmodBase : public SparseSolverBase<Derived>
     }
     
     /** \internal */
-    template<typename RhsScalar, int RhsOptions, typename RhsIndex, typename DestScalar, int DestOptions, typename DestIndex>
-    void _solve_impl(const SparseMatrix<RhsScalar,RhsOptions,RhsIndex> &b, SparseMatrix<DestScalar,DestOptions,DestIndex> &dest) const
+    template<typename RhsDerived, typename DestDerived>
+    void _solve_impl(const SparseMatrixBase<RhsDerived> &b, SparseMatrixBase<DestDerived> &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;
@@ -299,7 +314,8 @@ class CholmodBase : public SparseSolverBase<Derived>
       eigen_assert(size==b.rows());
 
       // note: cs stands for Cholmod Sparse
-      cholmod_sparse b_cs = viewAsCholmod(b);
+      Ref<SparseMatrix<typename RhsDerived::Scalar,ColMajor,typename RhsDerived::StorageIndex> > b_ref(b.const_cast_derived());
+      cholmod_sparse b_cs = viewAsCholmod(b_ref);
       cholmod_sparse* x_cs = cholmod_spsolve(CHOLMOD_A, m_cholmodFactor, &b_cs, &m_cholmod);
       if(!x_cs)
       {
@@ -307,7 +323,7 @@ class CholmodBase : public SparseSolverBase<Derived>
         return;
       }
       // TODO optimize this copy by swapping when possible (be careful with alignment, etc.)
-      dest = viewAsEigen<DestScalar,DestOptions,DestIndex>(*x_cs);
+      dest.derived() = viewAsEigen<typename DestDerived::Scalar,ColMajor,typename DestDerived::StorageIndex>(*x_cs);
       cholmod_free_sparse(&x_cs, &m_cholmod);
     }
     #endif // EIGEN_PARSED_BY_DOXYGEN
@@ -324,7 +340,7 @@ class CholmodBase : public SparseSolverBase<Derived>
       */
     Derived& setShift(const RealScalar& offset)
     {
-      m_shiftOffset[0] = offset;
+      m_shiftOffset[0] = double(offset);
       return derived();
     }
     
@@ -386,7 +402,7 @@ class CholmodBase : public SparseSolverBase<Derived>
   protected:
     mutable cholmod_common m_cholmod;
     cholmod_factor* m_cholmodFactor;
-    RealScalar m_shiftOffset[2];
+    double m_shiftOffset[2];
     mutable ComputationInfo m_info;
     int m_factorizationIsOk;
     int m_analysisIsOk;
@@ -410,6 +426,8 @@ class CholmodBase : public SparseSolverBase<Derived>
   *
   * This class supports all kind of SparseMatrix<>: row or column major; upper, lower, or both; compressed or non compressed.
   *
+  * \warning Only double precision real and complex scalar types are supported by Cholmod.
+  *
   * \sa \ref TutorialSparseSolverConcept, class CholmodSupernodalLLT, class SimplicialLLT
   */
 template<typename _MatrixType, int _UpLo = Lower>
@@ -459,6 +477,8 @@ class CholmodSimplicialLLT : public CholmodBase<_MatrixType, _UpLo, CholmodSimpl
   *
   * This class supports all kind of SparseMatrix<>: row or column major; upper, lower, or both; compressed or non compressed.
   *
+  * \warning Only double precision real and complex scalar types are supported by Cholmod.
+  *
   * \sa \ref TutorialSparseSolverConcept, class CholmodSupernodalLLT, class SimplicialLDLT
   */
 template<typename _MatrixType, int _UpLo = Lower>
@@ -506,6 +526,8 @@ class CholmodSimplicialLDLT : public CholmodBase<_MatrixType, _UpLo, CholmodSimp
   *
   * This class supports all kind of SparseMatrix<>: row or column major; upper, lower, or both; compressed or non compressed.
   *
+  * \warning Only double precision real and complex scalar types are supported by Cholmod.
+  *
   * \sa \ref TutorialSparseSolverConcept
   */
 template<typename _MatrixType, int _UpLo = Lower>
@@ -555,6 +577,8 @@ class CholmodSupernodalLLT : public CholmodBase<_MatrixType, _UpLo, CholmodSuper
   *
   * This class supports all kind of SparseMatrix<>: row or column major; upper, lower, or both; compressed or non compressed.
   *
+  * \warning Only double precision real and complex scalar types are supported by Cholmod.
+  *
   * \sa \ref TutorialSparseSolverConcept
   */
 template<typename _MatrixType, int _UpLo = Lower>

Eigen/src/Core/Array.h

@@ -37,7 +37,7 @@ struct traits<Array<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> > : tra
   * 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.
+  * \ref TopicCustomizing_Plugins by defining the preprocessor symbol \c EIGEN_ARRAY_PLUGIN.
   *
   * \sa \blank \ref TutorialArrayClass, \ref TopicClassHierarchy
   */

Eigen/src/Core/ArrayBase.h

@@ -32,7 +32,7 @@ template<typename ExpressionType> class MatrixWrapper;
   * \tparam Derived is the derived type, e.g., an array or an expression type.
   *
   * 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_ARRAYBASE_PLUGIN.
+  * \ref TopicCustomizing_Plugins by defining the preprocessor symbol \c EIGEN_ARRAYBASE_PLUGIN.
   *
   * \sa class MatrixBase, \ref TopicClassHierarchy
   */
@@ -87,6 +87,7 @@ template<typename Derived> class ArrayBase
 #endif // not EIGEN_PARSED_BY_DOXYGEN
 
 #define EIGEN_CURRENT_STORAGE_BASE_CLASS Eigen::ArrayBase
+#define EIGEN_DOC_UNARY_ADDONS(X,Y)
 #   include "../plugins/CommonCwiseUnaryOps.h"
 #   include "../plugins/MatrixCwiseUnaryOps.h"
 #   include "../plugins/ArrayCwiseUnaryOps.h"
@@ -97,6 +98,7 @@ template<typename Derived> class ArrayBase
 #     include EIGEN_ARRAYBASE_PLUGIN
 #   endif
 #undef EIGEN_CURRENT_STORAGE_BASE_CLASS
+#undef EIGEN_DOC_UNARY_ADDONS
 
     /** Special case of the template operator=, in order to prevent the compiler
       * from generating a default operator= (issue hit with g++ 4.1)

Eigen/src/Core/ArrayWrapper.h

@@ -54,6 +54,8 @@ class ArrayWrapper : public ArrayBase<ArrayWrapper<ExpressionType> >
 
     typedef typename internal::ref_selector<ExpressionType>::non_const_type NestedExpressionType;
 
+    using Base::coeffRef;
+
     EIGEN_DEVICE_FUNC
     explicit EIGEN_STRONG_INLINE ArrayWrapper(ExpressionType& matrix) : m_expression(matrix) {}
 
@@ -71,66 +73,18 @@ class ArrayWrapper : public ArrayBase<ArrayWrapper<ExpressionType> >
     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.coeffRef(rowId, colId);
-    }
-
     EIGEN_DEVICE_FUNC
     inline const Scalar& coeffRef(Index rowId, Index colId) const
     {
       return m_expression.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.coeffRef(index);
-    }
-
     EIGEN_DEVICE_FUNC
     inline const Scalar& coeffRef(Index index) const
     {
       return m_expression.coeffRef(index);
     }
 
-    template<int LoadMode>
-    inline const PacketScalar packet(Index rowId, Index colId) const
-    {
-      return m_expression.template packet<LoadMode>(rowId, colId);
-    }
-
-    template<int LoadMode>
-    inline void writePacket(Index rowId, Index colId, const PacketScalar& val)
-    {
-      m_expression.template writePacket<LoadMode>(rowId, colId, val);
-    }
-
-    template<int LoadMode>
-    inline const PacketScalar packet(Index index) const
-    {
-      return m_expression.template packet<LoadMode>(index);
-    }
-
-    template<int LoadMode>
-    inline void writePacket(Index index, const PacketScalar& val)
-    {
-      m_expression.template writePacket<LoadMode>(index, val);
-    }
-
     template<typename Dest>
     EIGEN_DEVICE_FUNC
     inline void evalTo(Dest& dst) const { dst = m_expression; }
@@ -197,6 +151,8 @@ class MatrixWrapper : public MatrixBase<MatrixWrapper<ExpressionType> >
 
     typedef typename internal::ref_selector<ExpressionType>::non_const_type NestedExpressionType;
 
+    using Base::coeffRef;
+
     EIGEN_DEVICE_FUNC
     explicit inline MatrixWrapper(ExpressionType& matrix) : m_expression(matrix) {}
 
@@ -214,66 +170,18 @@ class MatrixWrapper : public MatrixBase<MatrixWrapper<ExpressionType> >
     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.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.coeffRef(index);
-    }
-
     EIGEN_DEVICE_FUNC
     inline const Scalar& coeffRef(Index index) const
     {
       return m_expression.coeffRef(index);
     }
 
-    template<int LoadMode>
-    inline const PacketScalar packet(Index rowId, Index colId) const
-    {
-      return m_expression.template packet<LoadMode>(rowId, colId);
-    }
-
-    template<int LoadMode>
-    inline void writePacket(Index rowId, Index colId, const PacketScalar& val)
-    {
-      m_expression.template writePacket<LoadMode>(rowId, colId, val);
-    }
-
-    template<int LoadMode>
-    inline const PacketScalar packet(Index index) const
-    {
-      return m_expression.template packet<LoadMode>(index);
-    }
-
-    template<int LoadMode>
-    inline void writePacket(Index index, const PacketScalar& val)
-    {
-      m_expression.template writePacket<LoadMode>(index, val);
-    }
-
     EIGEN_DEVICE_FUNC
     const typename internal::remove_all<NestedExpressionType>::type& 
     nestedExpression() const 

Eigen/src/Core/AssignEvaluator.h

@@ -88,10 +88,11 @@ private:
       /* 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  = bool(MightVectorize) && bool(DstHasDirectAccess)
-                       && (int(InnerMaxSize)==Dynamic || int(InnerMaxSize)>=3*InnerPacketSize)
+                       && (int(InnerMaxSize)==Dynamic || int(InnerMaxSize)>=(EIGEN_UNALIGNED_VECTORIZE?InnerPacketSize:(3*InnerPacketSize)))
       /* 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 */
+         in a fixed-size matrix
+         However, with EIGEN_UNALIGNED_VECTORIZE and unrolling, slice vectorization is still worth it */
   };
 
 public:
@@ -136,6 +137,11 @@ public:
               : int(Traversal) == int(LinearTraversal)
                 ? ( bool(MayUnrollCompletely) ? int(CompleteUnrolling) 
                                               : int(NoUnrolling) )
+#if EIGEN_UNALIGNED_VECTORIZE
+              : int(Traversal) == int(SliceVectorizedTraversal)
+                ? ( bool(MayUnrollInner) ? int(InnerUnrolling)
+                                         : int(NoUnrolling) )
+#endif
               : int(NoUnrolling)
   };
 
@@ -277,24 +283,20 @@ struct copy_using_evaluator_innervec_CompleteUnrolling<Kernel, Stop, Stop>
   EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel&) { }
 };
 
-template<typename Kernel, int Index_, int Stop>
+template<typename Kernel, int Index_, int Stop, int SrcAlignment, int DstAlignment>
 struct copy_using_evaluator_innervec_InnerUnrolling
 {
   typedef typename Kernel::PacketType PacketType;
-  enum {
-    SrcAlignment = Kernel::AssignmentTraits::SrcAlignment,
-    DstAlignment = Kernel::AssignmentTraits::DstAlignment
-  };
   EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel, Index outer)
   {
     kernel.template assignPacketByOuterInner<DstAlignment, SrcAlignment, PacketType>(outer, Index_);
     enum { NextIndex = Index_ + unpacket_traits<PacketType>::size };
-    copy_using_evaluator_innervec_InnerUnrolling<Kernel, NextIndex, Stop>::run(kernel, outer);
+    copy_using_evaluator_innervec_InnerUnrolling<Kernel, NextIndex, Stop, SrcAlignment, DstAlignment>::run(kernel, outer);
   }
 };
 
-template<typename Kernel, int Stop>
-struct copy_using_evaluator_innervec_InnerUnrolling<Kernel, Stop, Stop>
+template<typename Kernel, int Stop, int SrcAlignment, int DstAlignment>
+struct copy_using_evaluator_innervec_InnerUnrolling<Kernel, Stop, Stop, SrcAlignment, DstAlignment>
 {
   EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &, Index) { }
 };
@@ -405,7 +407,7 @@ struct dense_assignment_loop<Kernel, LinearVectorizedTraversal, NoUnrolling>
                                                             : int(Kernel::AssignmentTraits::DstAlignment),
       srcAlignment = Kernel::AssignmentTraits::JointAlignment
     };
-    const Index alignedStart = dstIsAligned ? 0 : internal::first_aligned<requestedAlignment>(&kernel.dstEvaluator().coeffRef(0), size);
+    const Index alignedStart = dstIsAligned ? 0 : internal::first_aligned<requestedAlignment>(kernel.dstDataPtr(), size);
     const Index alignedEnd = alignedStart + ((size-alignedStart)/packetSize)*packetSize;
 
     unaligned_dense_assignment_loop<dstIsAligned!=0>::run(kernel, 0, alignedStart);
@@ -423,9 +425,10 @@ struct dense_assignment_loop<Kernel, LinearVectorizedTraversal, CompleteUnrollin
   EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel)
   {
     typedef typename Kernel::DstEvaluatorType::XprType DstXprType;
+    typedef typename Kernel::PacketType PacketType;
     
     enum { size = DstXprType::SizeAtCompileTime,
-           packetSize = packet_traits<typename Kernel::Scalar>::size,
+           packetSize =unpacket_traits<PacketType>::size,
            alignedSize = (size/packetSize)*packetSize };
 
     copy_using_evaluator_innervec_CompleteUnrolling<Kernel, 0, alignedSize>::run(kernel);
@@ -472,9 +475,11 @@ struct dense_assignment_loop<Kernel, InnerVectorizedTraversal, InnerUnrolling>
   EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel)
   {
     typedef typename Kernel::DstEvaluatorType::XprType DstXprType;
+    typedef typename Kernel::AssignmentTraits Traits;
     const Index outerSize = kernel.outerSize();
     for(Index outer = 0; outer < outerSize; ++outer)
-      copy_using_evaluator_innervec_InnerUnrolling<Kernel, 0, DstXprType::InnerSizeAtCompileTime>::run(kernel, outer);
+      copy_using_evaluator_innervec_InnerUnrolling<Kernel, 0, DstXprType::InnerSizeAtCompileTime,
+                                                   Traits::SrcAlignment, Traits::DstAlignment>::run(kernel, outer);
   }
 };
 
@@ -510,7 +515,7 @@ struct dense_assignment_loop<Kernel, LinearTraversal, CompleteUnrolling>
 template<typename Kernel>
 struct dense_assignment_loop<Kernel, SliceVectorizedTraversal, NoUnrolling>
 {
-  EIGEN_DEVICE_FUNC static inline void run(Kernel &kernel)
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel)
   {
     typedef typename Kernel::Scalar Scalar;
     typedef typename Kernel::PacketType PacketType;
@@ -522,7 +527,7 @@ struct dense_assignment_loop<Kernel, SliceVectorizedTraversal, NoUnrolling>
       dstAlignment = alignable ? int(requestedAlignment)
                                : int(Kernel::AssignmentTraits::DstAlignment)
     };
-    const Scalar *dst_ptr = &kernel.dstEvaluator().coeffRef(0,0);
+    const Scalar *dst_ptr = kernel.dstDataPtr();
     if((!bool(dstIsAligned)) && (UIntPtr(dst_ptr) % sizeof(Scalar))>0)
     {
       // the pointer is not aligend-on scalar, so alignment is not possible
@@ -549,11 +554,34 @@ struct dense_assignment_loop<Kernel, SliceVectorizedTraversal, NoUnrolling>
       for(Index inner = alignedEnd; inner<innerSize ; ++inner)
         kernel.assignCoeffByOuterInner(outer, inner);
 
-      alignedStart = std::min<Index>((alignedStart+alignedStep)%packetSize, innerSize);
+      alignedStart = numext::mini((alignedStart+alignedStep)%packetSize, innerSize);
     }
   }
 };
 
+#if EIGEN_UNALIGNED_VECTORIZE
+template<typename Kernel>
+struct dense_assignment_loop<Kernel, SliceVectorizedTraversal, InnerUnrolling>
+{
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel)
+  {
+    typedef typename Kernel::DstEvaluatorType::XprType DstXprType;
+    typedef typename Kernel::PacketType PacketType;
+
+    enum { size = DstXprType::InnerSizeAtCompileTime,
+           packetSize =unpacket_traits<PacketType>::size,
+           vectorizableSize = (size/packetSize)*packetSize };
+
+    for(Index outer = 0; outer < kernel.outerSize(); ++outer)
+    {
+      copy_using_evaluator_innervec_InnerUnrolling<Kernel, 0, vectorizableSize, 0, 0>::run(kernel, outer);
+      copy_using_evaluator_DefaultTraversal_InnerUnrolling<Kernel, vectorizableSize, size>::run(kernel, outer);
+    }
+  }
+};
+#endif
+
+
 /***************************************************************************
 * Part 4 : Generic dense assignment kernel
 ***************************************************************************/
@@ -655,6 +683,11 @@ public:
       : int(DstEvaluatorType::Flags)&RowMajorBit ? inner
       : outer;
   }
+
+  EIGEN_DEVICE_FUNC const Scalar* dstDataPtr() const
+  {
+    return m_dstExpr.data();
+  }
   
 protected:
   DstEvaluatorType& m_dst;
@@ -668,25 +701,48 @@ protected:
 * Part 5 : Entry point for dense rectangular assignment
 ***************************************************************************/
 
-template<typename DstXprType, typename SrcXprType, typename Functor>
-EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void call_dense_assignment_loop(const DstXprType& dst, const SrcXprType& src, const Functor &func)
+template<typename DstXprType,typename SrcXprType, typename Functor>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+void resize_if_allowed(DstXprType &dst, const SrcXprType& src, const Functor &/*func*/)
 {
+  EIGEN_ONLY_USED_FOR_DEBUG(dst);
+  EIGEN_ONLY_USED_FOR_DEBUG(src);
   eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols());
-  
+}
+
+template<typename DstXprType,typename SrcXprType, typename T1, typename T2>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+void resize_if_allowed(DstXprType &dst, const SrcXprType& src, const internal::assign_op<T1,T2> &/*func*/)
+{
+  Index dstRows = src.rows();
+  Index dstCols = src.cols();
+  if(((dst.rows()!=dstRows) || (dst.cols()!=dstCols)))
+    dst.resize(dstRows, dstCols);
+  eigen_assert(dst.rows() == dstRows && dst.cols() == dstCols);
+}
+
+template<typename DstXprType, typename SrcXprType, typename Functor>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void call_dense_assignment_loop(DstXprType& dst, const SrcXprType& src, const Functor &func)
+{
   typedef evaluator<DstXprType> DstEvaluatorType;
   typedef evaluator<SrcXprType> SrcEvaluatorType;
 
-  DstEvaluatorType dstEvaluator(dst);
   SrcEvaluatorType srcEvaluator(src);
+
+  // NOTE To properly handle A = (A*A.transpose())/s with A rectangular,
+  // we need to resize the destination after the source evaluator has been created.
+  resize_if_allowed(dst, src, func);
+
+  DstEvaluatorType dstEvaluator(dst);
     
   typedef generic_dense_assignment_kernel<DstEvaluatorType,SrcEvaluatorType,Functor> Kernel;
   Kernel kernel(dstEvaluator, srcEvaluator, func, dst.const_cast_derived());
-  
+
   dense_assignment_loop<Kernel>::run(kernel);
 }
 
 template<typename DstXprType, typename SrcXprType>
-EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void call_dense_assignment_loop(const DstXprType& dst, const SrcXprType& src)
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void call_dense_assignment_loop(DstXprType& dst, const SrcXprType& src)
 {
   call_dense_assignment_loop(dst, src, internal::assign_op<typename DstXprType::Scalar,typename SrcXprType::Scalar>());
 }
@@ -768,11 +824,6 @@ void call_assignment_no_alias(Dst& dst, const Src& src, const Func& func)
                       ) && int(Dst::SizeAtCompileTime) != 1
   };
 
-  Index dstRows = NeedToTranspose ? src.cols() : src.rows();
-  Index dstCols = NeedToTranspose ? src.rows() : src.cols();
-  if((dst.rows()!=dstRows) || (dst.cols()!=dstCols))
-    dst.resize(dstRows, dstCols);
-  
   typedef typename internal::conditional<NeedToTranspose, Transpose<Dst>, Dst>::type ActualDstTypeCleaned;
   typedef typename internal::conditional<NeedToTranspose, Transpose<Dst>, Dst&>::type ActualDstType;
   ActualDstType actualDst(dst);
@@ -795,15 +846,11 @@ template<typename Dst, typename Src, typename Func>
 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
 void call_assignment_no_alias_no_transpose(Dst& dst, const Src& src, const Func& func)
 {
-  Index dstRows = src.rows();
-  Index dstCols = src.cols();
-  if((dst.rows()!=dstRows) || (dst.cols()!=dstCols))
-    dst.resize(dstRows, dstCols);
-  
   // TODO check whether this is the right place to perform these checks:
   EIGEN_STATIC_ASSERT_LVALUE(Dst)
   EIGEN_STATIC_ASSERT_SAME_MATRIX_SIZE(Dst,Src)
-  
+  EIGEN_CHECK_BINARY_COMPATIBILIY(Func,typename Dst::Scalar,typename Src::Scalar);
+
   Assignment<Dst,Src,Func>::run(dst, src, func);
 }
 template<typename Dst, typename Src>
@@ -825,8 +872,6 @@ struct Assignment<DstXprType, SrcXprType, Functor, Dense2Dense, Weak>
   EIGEN_DEVICE_FUNC
   static EIGEN_STRONG_INLINE void run(DstXprType &dst, const SrcXprType &src, const Functor &func)
   {
-    eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols());
-    
 #ifndef EIGEN_NO_DEBUG
     internal::check_for_aliasing(dst, src);
 #endif
@@ -845,9 +890,42 @@ struct Assignment<DstXprType, SrcXprType, Functor, EigenBase2EigenBase, Weak>
   EIGEN_DEVICE_FUNC
   static EIGEN_STRONG_INLINE void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<typename DstXprType::Scalar,typename SrcXprType::Scalar> &/*func*/)
   {
+    Index dstRows = src.rows();
+    Index dstCols = src.cols();
+    if((dst.rows()!=dstRows) || (dst.cols()!=dstCols))
+      dst.resize(dstRows, dstCols);
+
     eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols());
     src.evalTo(dst);
   }
+
+  // NOTE The following two functions are templated to avoid their instanciation if not needed
+  //      This is needed because some expressions supports evalTo only and/or have 'void' as scalar type.
+  template<typename SrcScalarType>
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE void run(DstXprType &dst, const SrcXprType &src, const internal::add_assign_op<typename DstXprType::Scalar,SrcScalarType> &/*func*/)
+  {
+    Index dstRows = src.rows();
+    Index dstCols = src.cols();
+    if((dst.rows()!=dstRows) || (dst.cols()!=dstCols))
+      dst.resize(dstRows, dstCols);
+
+    eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols());
+    src.addTo(dst);
+  }
+
+  template<typename SrcScalarType>
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE void run(DstXprType &dst, const SrcXprType &src, const internal::sub_assign_op<typename DstXprType::Scalar,SrcScalarType> &/*func*/)
+  {
+    Index dstRows = src.rows();
+    Index dstCols = src.cols();
+    if((dst.rows()!=dstRows) || (dst.cols()!=dstCols))
+      dst.resize(dstRows, dstCols);
+
+    eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols());
+    src.subTo(dst);
+  }
 };
 
 } // namespace internal

Eigen/src/Core/CMakeLists.txt

@@ -1,11 +0,0 @@
-FILE(GLOB Eigen_Core_SRCS "*.h")
-
-INSTALL(FILES
-  ${Eigen_Core_SRCS}
-  DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/Core COMPONENT Devel
-  )
-
-ADD_SUBDIRECTORY(products)
-ADD_SUBDIRECTORY(util)
-ADD_SUBDIRECTORY(arch)
-ADD_SUBDIRECTORY(functors)

Eigen/src/Core/CommaInitializer.h

@@ -80,12 +80,7 @@ struct CommaInitializer
   EIGEN_DEVICE_FUNC
   CommaInitializer& operator,(const DenseBase<OtherDerived>& other)
   {
-    if(other.rows()==0)
-    {
-      m_col += other.cols();
-      return *this;
-    }
-    if (m_col==m_xpr.cols())
+    if (m_col==m_xpr.cols() && (other.cols()!=0 || other.rows()!=m_currentBlockRows))
     {
       m_row+=m_currentBlockRows;
       m_col = 0;
@@ -93,15 +88,11 @@ struct CommaInitializer
       eigen_assert(m_row+m_currentBlockRows<=m_xpr.rows()
         && "Too many rows passed to comma initializer (operator<<)");
     }
-    eigen_assert((m_col<m_xpr.cols() || (m_xpr.cols()==0 && m_col==0))
+    eigen_assert((m_col + other.cols() <= m_xpr.cols())
       && "Too many coefficients passed to comma initializer (operator<<)");
     eigen_assert(m_currentBlockRows==other.rows());
-    if (OtherDerived::SizeAtCompileTime != Dynamic)
-      m_xpr.template block<OtherDerived::RowsAtCompileTime != Dynamic ? OtherDerived::RowsAtCompileTime : 1,
-                              OtherDerived::ColsAtCompileTime != Dynamic ? OtherDerived::ColsAtCompileTime : 1>
-                    (m_row, m_col) = other;
-    else
-      m_xpr.block(m_row, m_col, other.rows(), other.cols()) = other;
+    m_xpr.template block<OtherDerived::RowsAtCompileTime, OtherDerived::ColsAtCompileTime>
+                    (m_row, m_col, other.rows(), other.cols()) = other;
     m_col += other.cols();
     return *this;
   }
@@ -112,9 +103,7 @@ struct CommaInitializer
   EIGEN_EXCEPTION_SPEC(Eigen::eigen_assert_exception)
 #endif
   {
-    eigen_assert((m_row+m_currentBlockRows) == m_xpr.rows()
-         && m_col == m_xpr.cols()
-         && "Too few coefficients passed to comma initializer (operator<<)");
+      finished();
   }
 
   /** \returns the built matrix once all its coefficients have been set.
@@ -125,7 +114,12 @@ struct CommaInitializer
     * \endcode
     */
   EIGEN_DEVICE_FUNC
-  inline XprType& finished() { return m_xpr; }
+  inline XprType& finished() {
+      eigen_assert(((m_row+m_currentBlockRows) == m_xpr.rows() || m_xpr.cols() == 0)
+           && m_col == m_xpr.cols()
+           && "Too few coefficients passed to comma initializer (operator<<)");
+      return m_xpr;
+  }
 
   XprType& m_xpr;           // target expression
   Index m_row;              // current row id

Eigen/src/Core/CoreEvaluators.h

@@ -337,6 +337,120 @@ protected:
 // Like Matrix and Array, this is not really a unary expression, so we directly specialize evaluator.
 // Likewise, there is not need to more sophisticated dispatching here.
 
+template<typename Scalar,typename NullaryOp,
+         bool has_nullary = has_nullary_operator<NullaryOp>::value,
+         bool has_unary   = has_unary_operator<NullaryOp>::value,
+         bool has_binary  = has_binary_operator<NullaryOp>::value>
+struct nullary_wrapper
+{
+  template <typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar operator()(const NullaryOp& op, IndexType i, IndexType j) const { return op(i,j); }
+  template <typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar operator()(const NullaryOp& op, IndexType i) const { return op(i); }
+
+  template <typename T, typename IndexType> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T packetOp(const NullaryOp& op, IndexType i, IndexType j) const { return op.template packetOp<T>(i,j); }
+  template <typename T, typename IndexType> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T packetOp(const NullaryOp& op, IndexType i) const { return op.template packetOp<T>(i); }
+};
+
+template<typename Scalar,typename NullaryOp>
+struct nullary_wrapper<Scalar,NullaryOp,true,false,false>
+{
+  template <typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar operator()(const NullaryOp& op, IndexType=0, IndexType=0) const { return op(); }
+  template <typename T, typename IndexType> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T packetOp(const NullaryOp& op, IndexType=0, IndexType=0) const { return op.template packetOp<T>(); }
+};
+
+template<typename Scalar,typename NullaryOp>
+struct nullary_wrapper<Scalar,NullaryOp,false,false,true>
+{
+  template <typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar operator()(const NullaryOp& op, IndexType i, IndexType j=0) const { return op(i,j); }
+  template <typename T, typename IndexType> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T packetOp(const NullaryOp& op, IndexType i, IndexType j=0) const { return op.template packetOp<T>(i,j); }
+};
+
+// We need the following specialization for vector-only functors assigned to a runtime vector,
+// for instance, using linspace and assigning a RowVectorXd to a MatrixXd or even a row of a MatrixXd.
+// In this case, i==0 and j is used for the actual iteration.
+template<typename Scalar,typename NullaryOp>
+struct nullary_wrapper<Scalar,NullaryOp,false,true,false>
+{
+  template <typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar operator()(const NullaryOp& op, IndexType i, IndexType j) const {
+    eigen_assert(i==0 || j==0);
+    return op(i+j);
+  }
+  template <typename T, typename IndexType> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T packetOp(const NullaryOp& op, IndexType i, IndexType j) const {
+    eigen_assert(i==0 || j==0);
+    return op.template packetOp<T>(i+j);
+  }
+
+  template <typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar operator()(const NullaryOp& op, IndexType i) const { return op(i); }
+  template <typename T, typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T packetOp(const NullaryOp& op, IndexType i) const { return op.template packetOp<T>(i); }
+};
+
+template<typename Scalar,typename NullaryOp>
+struct nullary_wrapper<Scalar,NullaryOp,false,false,false> {};
+
+#if 0 && EIGEN_COMP_MSVC>0
+// Disable this ugly workaround. This is now handled in traits<Ref>::match,
+// but this piece of code might still become handly if some other weird compilation
+// erros pop up again.
+
+// MSVC exhibits a weird compilation error when
+// compiling:
+//    Eigen::MatrixXf A = MatrixXf::Random(3,3);
+//    Ref<const MatrixXf> R = 2.f*A;
+// and that has_*ary_operator<scalar_constant_op<float>> have not been instantiated yet.
+// The "problem" is that evaluator<2.f*A> is instantiated by traits<Ref>::match<2.f*A>
+// and at that time has_*ary_operator<T> returns true regardless of T.
+// Then nullary_wrapper is badly instantiated as nullary_wrapper<.,.,true,true,true>.
+// The trick is thus to defer the proper instantiation of nullary_wrapper when coeff(),
+// and packet() are really instantiated as implemented below:
+
+// This is a simple wrapper around Index to enforce the re-instantiation of
+// has_*ary_operator when needed.
+template<typename T> struct nullary_wrapper_workaround_msvc {
+  nullary_wrapper_workaround_msvc(const T&);
+  operator T()const;
+};
+
+template<typename Scalar,typename NullaryOp>
+struct nullary_wrapper<Scalar,NullaryOp,true,true,true>
+{
+  template <typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar operator()(const NullaryOp& op, IndexType i, IndexType j) const {
+    return nullary_wrapper<Scalar,NullaryOp,
+    has_nullary_operator<NullaryOp,nullary_wrapper_workaround_msvc<IndexType> >::value,
+    has_unary_operator<NullaryOp,nullary_wrapper_workaround_msvc<IndexType> >::value,
+    has_binary_operator<NullaryOp,nullary_wrapper_workaround_msvc<IndexType> >::value>().operator()(op,i,j);
+  }
+  template <typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar operator()(const NullaryOp& op, IndexType i) const {
+    return nullary_wrapper<Scalar,NullaryOp,
+    has_nullary_operator<NullaryOp,nullary_wrapper_workaround_msvc<IndexType> >::value,
+    has_unary_operator<NullaryOp,nullary_wrapper_workaround_msvc<IndexType> >::value,
+    has_binary_operator<NullaryOp,nullary_wrapper_workaround_msvc<IndexType> >::value>().operator()(op,i);
+  }
+
+  template <typename T, typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T packetOp(const NullaryOp& op, IndexType i, IndexType j) const {
+    return nullary_wrapper<Scalar,NullaryOp,
+    has_nullary_operator<NullaryOp,nullary_wrapper_workaround_msvc<IndexType> >::value,
+    has_unary_operator<NullaryOp,nullary_wrapper_workaround_msvc<IndexType> >::value,
+    has_binary_operator<NullaryOp,nullary_wrapper_workaround_msvc<IndexType> >::value>().template packetOp<T>(op,i,j);
+  }
+  template <typename T, typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T packetOp(const NullaryOp& op, IndexType i) const {
+    return nullary_wrapper<Scalar,NullaryOp,
+    has_nullary_operator<NullaryOp,nullary_wrapper_workaround_msvc<IndexType> >::value,
+    has_unary_operator<NullaryOp,nullary_wrapper_workaround_msvc<IndexType> >::value,
+    has_binary_operator<NullaryOp,nullary_wrapper_workaround_msvc<IndexType> >::value>().template packetOp<T>(op,i);
+  }
+};
+#endif // MSVC workaround
+
 template<typename NullaryOp, typename PlainObjectType>
 struct evaluator<CwiseNullaryOp<NullaryOp,PlainObjectType> >
   : evaluator_base<CwiseNullaryOp<NullaryOp,PlainObjectType> >
@@ -356,41 +470,44 @@ struct evaluator<CwiseNullaryOp<NullaryOp,PlainObjectType> >
   };
 
   EIGEN_DEVICE_FUNC explicit evaluator(const XprType& n)
-    : m_functor(n.functor()) 
+    : m_functor(n.functor()), m_wrapper()
   {
     EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
   }
 
   typedef typename XprType::CoeffReturnType CoeffReturnType;
 
+  template <typename IndexType>
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
-  CoeffReturnType coeff(Index row, Index col) const
+  CoeffReturnType coeff(IndexType row, IndexType col) const
   {
-    return m_functor(row, col);
+    return m_wrapper(m_functor, row, col);
   }
 
+  template <typename IndexType>
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
-  CoeffReturnType coeff(Index index) const
+  CoeffReturnType coeff(IndexType index) const
   {
-    return m_functor(index);
+    return m_wrapper(m_functor,index);
   }
 
-  template<int LoadMode, typename PacketType>
+  template<int LoadMode, typename PacketType, typename IndexType>
   EIGEN_STRONG_INLINE
-  PacketType packet(Index row, Index col) const
+  PacketType packet(IndexType row, IndexType col) const
   {
-    return m_functor.template packetOp<Index,PacketType>(row, col);
+    return m_wrapper.template packetOp<PacketType>(m_functor, row, col);
   }
 
-  template<int LoadMode, typename PacketType>
+  template<int LoadMode, typename PacketType, typename IndexType>
   EIGEN_STRONG_INLINE
-  PacketType packet(Index index) const
+  PacketType packet(IndexType index) const
   {
-    return m_functor.template packetOp<Index,PacketType>(index);
+    return m_wrapper.template packetOp<PacketType>(m_functor, index);
   }
 
 protected:
   const NullaryOp m_functor;
+  const internal::nullary_wrapper<CoeffReturnType,NullaryOp> m_wrapper;
 };
 
 // -------------------- CwiseUnaryOp --------------------
@@ -700,73 +817,79 @@ struct mapbase_evaluator : evaluator_base<Derived>
     ColsAtCompileTime = XprType::ColsAtCompileTime,
     CoeffReadCost = NumTraits<Scalar>::ReadCost
   };
-  
+
   EIGEN_DEVICE_FUNC explicit mapbase_evaluator(const XprType& map)
-    : m_data(const_cast<PointerType>(map.data())),  
-      m_xpr(map)
+    : m_data(const_cast<PointerType>(map.data())),
+      m_innerStride(map.innerStride()),
+      m_outerStride(map.outerStride())
   {
     EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(evaluator<Derived>::Flags&PacketAccessBit, internal::inner_stride_at_compile_time<Derived>::ret==1),
                         PACKET_ACCESS_REQUIRES_TO_HAVE_INNER_STRIDE_FIXED_TO_1);
     EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
   }
- 
+
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
   CoeffReturnType coeff(Index row, Index col) const
   {
-    return m_data[col * m_xpr.colStride() + row * m_xpr.rowStride()];
+    return m_data[col * colStride() + row * rowStride()];
   }
-  
+
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
   CoeffReturnType coeff(Index index) const
   {
-    return m_data[index * m_xpr.innerStride()];
+    return m_data[index * m_innerStride.value()];
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
   Scalar& coeffRef(Index row, Index col)
   {
-    return m_data[col * m_xpr.colStride() + row * m_xpr.rowStride()];
+    return m_data[col * colStride() + row * rowStride()];
   }
-  
+
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
   Scalar& coeffRef(Index index)
   {
-    return m_data[index * m_xpr.innerStride()];
+    return m_data[index * m_innerStride.value()];
   }
- 
+
   template<int LoadMode, typename PacketType>
   EIGEN_STRONG_INLINE
-  PacketType packet(Index row, Index col) const 
+  PacketType packet(Index row, Index col) const
   {
-    PointerType ptr = m_data + row * m_xpr.rowStride() + col * m_xpr.colStride();
+    PointerType ptr = m_data + row * rowStride() + col * colStride();
     return internal::ploadt<PacketType, LoadMode>(ptr);
   }
 
   template<int LoadMode, typename PacketType>
   EIGEN_STRONG_INLINE
-  PacketType packet(Index index) const 
+  PacketType packet(Index index) const
   {
-    return internal::ploadt<PacketType, LoadMode>(m_data + index * m_xpr.innerStride());
+    return internal::ploadt<PacketType, LoadMode>(m_data + index * m_innerStride.value());
   }
-  
+
   template<int StoreMode, typename PacketType>
   EIGEN_STRONG_INLINE
-  void writePacket(Index row, Index col, const PacketType& x) 
+  void writePacket(Index row, Index col, const PacketType& x)
   {
-    PointerType ptr = m_data + row * m_xpr.rowStride() + col * m_xpr.colStride();
+    PointerType ptr = m_data + row * rowStride() + col * colStride();
     return internal::pstoret<Scalar, PacketType, StoreMode>(ptr, x);
   }
-  
+
   template<int StoreMode, typename PacketType>
   EIGEN_STRONG_INLINE
-  void writePacket(Index index, const PacketType& x) 
+  void writePacket(Index index, const PacketType& x)
   {
-    internal::pstoret<Scalar, PacketType, StoreMode>(m_data + index * m_xpr.innerStride(), x);
+    internal::pstoret<Scalar, PacketType, StoreMode>(m_data + index * m_innerStride.value(), x);
   }
- 
 protected:
+  EIGEN_DEVICE_FUNC
+  inline Index rowStride() const { return XprType::IsRowMajor ? m_outerStride.value() : m_innerStride.value(); }
+  EIGEN_DEVICE_FUNC
+  inline Index colStride() const { return XprType::IsRowMajor ? m_innerStride.value() : m_outerStride.value(); }
+
   PointerType m_data;
-  const XprType& m_xpr;
+  const internal::variable_if_dynamic<Index, XprType::InnerStrideAtCompileTime> m_innerStride;
+  const internal::variable_if_dynamic<Index, XprType::OuterStrideAtCompileTime> m_outerStride;
 };
 
 template<typename PlainObjectType, int MapOptions, typename StrideType> 
@@ -1179,7 +1302,7 @@ struct evaluator<PartialReduxExpr<ArgType, MemberOp, Direction> >
   }
 
 protected:
-  const ArgTypeNested m_arg;
+  typename internal::add_const_on_value_type<ArgTypeNested>::type m_arg;
   const MemberOp m_functor;
 };
 
@@ -1433,9 +1556,7 @@ struct evaluator<Diagonal<ArgType, DiagIndex> >
   { }
  
   typedef typename XprType::Scalar Scalar;
-  // FIXME having to check whether ArgType is sparse here i not very nice.
-  typedef typename internal::conditional<!internal::is_same<typename ArgType::StorageKind,Sparse>::value,
-                                         typename XprType::CoeffReturnType,Scalar>::type CoeffReturnType;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
   CoeffReturnType coeff(Index row, Index) const

Eigen/src/Core/CwiseBinaryOp.h

@@ -46,7 +46,7 @@ struct traits<CwiseBinaryOp<BinaryOp, Lhs, Rhs> >
   typedef typename remove_reference<LhsNested>::type _LhsNested;
   typedef typename remove_reference<RhsNested>::type _RhsNested;
   enum {
-    Flags = _LhsNested::Flags & RowMajorBit
+    Flags = cwise_promote_storage_order<typename traits<Lhs>::StorageKind,typename traits<Rhs>::StorageKind,_LhsNested::Flags & RowMajorBit,_RhsNested::Flags & RowMajorBit>::value
   };
 };
 } // end namespace internal
@@ -84,6 +84,7 @@ class CwiseBinaryOp :
 {
   public:
     
+    typedef typename internal::remove_all<BinaryOp>::type Functor;
     typedef typename internal::remove_all<LhsType>::type Lhs;
     typedef typename internal::remove_all<RhsType>::type Rhs;
 

Eigen/src/Core/CwiseNullaryOp.h

@@ -20,7 +20,8 @@ struct traits<CwiseNullaryOp<NullaryOp, PlainObjectType> > : traits<PlainObjectT
     Flags = traits<PlainObjectType>::Flags & RowMajorBit
   };
 };
-}
+
+} // namespace internal
 
 /** \class CwiseNullaryOp
   * \ingroup Core_Module
@@ -37,7 +38,23 @@ struct traits<CwiseNullaryOp<NullaryOp, PlainObjectType> > : traits<PlainObjectT
   * However, if you want to write a function returning such an expression, you
   * will need to use this class.
   *
-  * \sa class CwiseUnaryOp, class CwiseBinaryOp, DenseBase::NullaryExpr()
+  * The functor NullaryOp must expose one of the following method:
+    <table class="manual">
+    <tr            ><td>\c operator()() </td><td>if the procedural generation does not depend on the coefficient entries (e.g., random numbers)</td></tr>
+    <tr class="alt"><td>\c operator()(Index i)</td><td>if the procedural generation makes sense for vectors only and that it depends on the coefficient index \c i (e.g., linspace) </td></tr>
+    <tr            ><td>\c operator()(Index i,Index j)</td><td>if the procedural generation depends on the matrix coordinates \c i, \c j (e.g., to generate a checkerboard with 0 and 1)</td></tr>
+    </table>
+  * It is also possible to expose the last two operators if the generation makes sense for matrices but can be optimized for vectors.
+  *
+  * See DenseBase::NullaryExpr(Index,const CustomNullaryOp&) for an example binding
+  * C++11 random number generators.
+  *
+  * A nullary expression can also be used to implement custom sophisticated matrix manipulations
+  * that cannot be covered by the existing set of natively supported matrix manipulations.
+  * See this \ref TopicCustomizing_NullaryExpr "page" for some examples and additional explanations
+  * on the behavior of CwiseNullaryOp.
+  *
+  * \sa class CwiseUnaryOp, class CwiseBinaryOp, DenseBase::NullaryExpr
   */
 template<typename NullaryOp, typename PlainObjectType>
 class CwiseNullaryOp : public internal::dense_xpr_base< CwiseNullaryOp<NullaryOp, PlainObjectType> >::type, internal::no_assignment_operator
@@ -62,30 +79,6 @@ class CwiseNullaryOp : public internal::dense_xpr_base< CwiseNullaryOp<NullaryOp
     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);
-    }
-
-    template<int LoadMode>
-    EIGEN_STRONG_INLINE PacketScalar packet(Index rowId, Index colId) const
-    {
-      return m_functor.packetOp(rowId, colId);
-    }
-
-    EIGEN_DEVICE_FUNC
-    EIGEN_STRONG_INLINE const Scalar coeff(Index index) const
-    {
-      return m_functor(index);
-    }
-
-    template<int LoadMode>
-    EIGEN_STRONG_INLINE PacketScalar packet(Index index) const
-    {
-      return m_functor.packetOp(index);
-    }
-
     /** \returns the functor representing the nullary operation */
     EIGEN_DEVICE_FUNC
     const NullaryOp& functor() const { return m_functor; }
@@ -222,42 +215,29 @@ DenseBase<Derived>::Constant(const Scalar& value)
   return DenseBase<Derived>::NullaryExpr(RowsAtCompileTime, ColsAtCompileTime, internal::scalar_constant_op<Scalar>(value));
 }
 
-/**
-  * \brief Sets a linearly spaced vector.
+/** \deprecated because of accuracy loss. In Eigen 3.3, it is an alias for LinSpaced(Index,const Scalar&,const Scalar&)
   *
-  * The function generates 'size' equally spaced values in the closed interval [low,high].
-  * This particular version of LinSpaced() uses sequential access, i.e. vector access is
-  * assumed to be a(0), a(1), ..., a(size). This assumption allows for better vectorization
-  * and yields faster code than the random access version.
-  *
-  * When size is set to 1, a vector of length 1 containing 'high' is returned.
-  *
-  * \only_for_vectors
-  *
-  * Example: \include DenseBase_LinSpaced_seq.cpp
-  * Output: \verbinclude DenseBase_LinSpaced_seq.out
-  *
-  * \sa setLinSpaced(Index,const Scalar&,const Scalar&), LinSpaced(Index,Scalar,Scalar), CwiseNullaryOp
+  * \sa LinSpaced(Index,Scalar,Scalar), setLinSpaced(Index,const Scalar&,const Scalar&)
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE const typename DenseBase<Derived>::SequentialLinSpacedReturnType
+EIGEN_STRONG_INLINE const typename DenseBase<Derived>::RandomAccessLinSpacedReturnType
 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,PacketScalar,false>(low,high,size));
+  return DenseBase<Derived>::NullaryExpr(size, internal::linspaced_op<Scalar,PacketScalar>(low,high,size));
 }
 
-/**
-  * \copydoc DenseBase::LinSpaced(Sequential_t, Index, const Scalar&, const Scalar&)
-  * Special version for fixed size types which does not require the size parameter.
+/** \deprecated because of accuracy loss. In Eigen 3.3, it is an alias for LinSpaced(const Scalar&,const Scalar&)
+  *
+  * \sa LinSpaced(Scalar,Scalar)
   */
 template<typename Derived>
-EIGEN_STRONG_INLINE const typename DenseBase<Derived>::SequentialLinSpacedReturnType
+EIGEN_STRONG_INLINE const typename DenseBase<Derived>::RandomAccessLinSpacedReturnType
 DenseBase<Derived>::LinSpaced(Sequential_t, 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,PacketScalar,false>(low,high,Derived::SizeAtCompileTime));
+  return DenseBase<Derived>::NullaryExpr(Derived::SizeAtCompileTime, internal::linspaced_op<Scalar,PacketScalar>(low,high,Derived::SizeAtCompileTime));
 }
 
 /**
@@ -271,14 +251,24 @@ DenseBase<Derived>::LinSpaced(Sequential_t, const Scalar& low, const Scalar& hig
   * Example: \include DenseBase_LinSpaced.cpp
   * Output: \verbinclude DenseBase_LinSpaced.out
   *
-  * \sa setLinSpaced(Index,const Scalar&,const Scalar&), LinSpaced(Sequential_t,Index,const Scalar&,const Scalar&,Index), CwiseNullaryOp
+  * For integer scalar types, an even spacing is possible if and only if the length of the range,
+  * i.e., \c high-low is a scalar multiple of \c size-1, or if \c size is a scalar multiple of the
+  * number of values \c high-low+1 (meaning each value can be repeated the same number of time).
+  * If one of these two considions is not satisfied, then \c high is lowered to the largest value
+  * satisfying one of this constraint.
+  * Here are some examples:
+  *
+  * Example: \include DenseBase_LinSpacedInt.cpp
+  * Output: \verbinclude DenseBase_LinSpacedInt.out
+  *
+  * \sa setLinSpaced(Index,const Scalar&,const Scalar&), CwiseNullaryOp
   */
 template<typename Derived>
 EIGEN_STRONG_INLINE const typename DenseBase<Derived>::RandomAccessLinSpacedReturnType
 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,PacketScalar,true>(low,high,size));
+  return DenseBase<Derived>::NullaryExpr(size, internal::linspaced_op<Scalar,PacketScalar>(low,high,size));
 }
 
 /**
@@ -291,7 +281,7 @@ 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,PacketScalar,true>(low,high,Derived::SizeAtCompileTime));
+  return DenseBase<Derived>::NullaryExpr(Derived::SizeAtCompileTime, internal::linspaced_op<Scalar,PacketScalar>(low,high,Derived::SizeAtCompileTime));
 }
 
 /** \returns true if all coefficients in this matrix are approximately equal to \a val, to within precision \a prec */
@@ -384,24 +374,30 @@ PlainObjectBase<Derived>::setConstant(Index rows, Index cols, const Scalar& val)
   * Example: \include DenseBase_setLinSpaced.cpp
   * Output: \verbinclude DenseBase_setLinSpaced.out
   *
-  * \sa CwiseNullaryOp
+  * For integer scalar types, do not miss the explanations on the definition
+  * of \link LinSpaced(Index,const Scalar&,const Scalar&) even spacing \endlink.
+  *
+  * \sa LinSpaced(Index,const Scalar&,const Scalar&), CwiseNullaryOp
   */
 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,PacketScalar,false>(low,high,newSize));
+  return derived() = Derived::NullaryExpr(newSize, internal::linspaced_op<Scalar,PacketScalar>(low,high,newSize));
 }
 
 /**
   * \brief Sets a linearly spaced vector.
   *
-  * The function fill *this with equally spaced values in the closed interval [low,high].
+  * The function fills \c *this with equally spaced values in the closed interval [low,high].
   * When size is set to 1, a vector of length 1 containing 'high' is returned.
   *
   * \only_for_vectors
   *
-  * \sa setLinSpaced(Index, const Scalar&, const Scalar&), CwiseNullaryOp
+  * For integer scalar types, do not miss the explanations on the definition
+  * of \link LinSpaced(Index,const Scalar&,const Scalar&) even spacing \endlink.
+  *
+  * \sa LinSpaced(Index,const Scalar&,const Scalar&), setLinSpaced(Index, const Scalar&, const Scalar&), CwiseNullaryOp
   */
 template<typename Derived>
 EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::setLinSpaced(const Scalar& low, const Scalar& high)
@@ -759,7 +755,7 @@ struct setIdentity_impl<Derived, true>
   static EIGEN_STRONG_INLINE Derived& run(Derived& m)
   {
     m.setZero();
-    const Index size = (std::min)(m.rows(), m.cols());
+    const Index size = numext::mini(m.rows(), m.cols());
     for(Index i = 0; i < size; ++i) m.coeffRef(i,i) = typename Derived::Scalar(1);
     return m;
   }

Eigen/src/Core/DenseBase.h

@@ -34,7 +34,7 @@ static inline void check_DenseIndex_is_signed() {
   * \tparam Derived is the derived type, e.g., a matrix type or an expression.
   *
   * 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_DENSEBASE_PLUGIN.
+  * \ref TopicCustomizing_Plugins by defining the preprocessor symbol \c EIGEN_DENSEBASE_PLUGIN.
   *
   * \sa \blank \ref TopicClassHierarchy
   */
@@ -260,10 +260,10 @@ template<typename Derived> class DenseBase
 #ifndef EIGEN_PARSED_BY_DOXYGEN
     /** \internal Represents a matrix with all coefficients equal to one another*/
     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,PacketScalar,false>,PlainObject> SequentialLinSpacedReturnType;
+    /** \internal \deprecated Represents a vector with linearly spaced coefficients that allows sequential access only. */
+    typedef CwiseNullaryOp<internal::linspaced_op<Scalar,PacketScalar>,PlainObject> SequentialLinSpacedReturnType;
     /** \internal Represents a vector with linearly spaced coefficients that allows random access. */
-    typedef CwiseNullaryOp<internal::linspaced_op<Scalar,PacketScalar,true>,PlainObject> RandomAccessLinSpacedReturnType;
+    typedef CwiseNullaryOp<internal::linspaced_op<Scalar,PacketScalar>,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;
 
@@ -463,7 +463,17 @@ template<typename Derived> class DenseBase
     EIGEN_DEVICE_FUNC
     void visit(Visitor& func) const;
 
-    inline const WithFormat<Derived> format(const IOFormat& fmt) const;
+    /** \returns a WithFormat proxy object allowing to print a matrix the with given
+      * format \a fmt.
+      *
+      * See class IOFormat for some examples.
+      *
+      * \sa class IOFormat, class WithFormat
+      */
+    inline const WithFormat<Derived> format(const IOFormat& fmt) const
+    {
+      return WithFormat<Derived>(derived(), fmt);
+    }
 
     /** \returns the unique coefficient of a 1x1 expression */
     EIGEN_DEVICE_FUNC
@@ -558,12 +568,15 @@ template<typename Derived> class DenseBase
     EIGEN_DEVICE_FUNC void reverseInPlace();
 
 #define EIGEN_CURRENT_STORAGE_BASE_CLASS Eigen::DenseBase
+#define EIGEN_DOC_BLOCK_ADDONS_NOT_INNER_PANEL
+#define EIGEN_DOC_BLOCK_ADDONS_INNER_PANEL_IF(COND)
 #   include "../plugins/BlockMethods.h"
 #   ifdef EIGEN_DENSEBASE_PLUGIN
 #     include EIGEN_DENSEBASE_PLUGIN
 #   endif
 #undef EIGEN_CURRENT_STORAGE_BASE_CLASS
-
+#undef EIGEN_DOC_BLOCK_ADDONS_NOT_INNER_PANEL
+#undef EIGEN_DOC_BLOCK_ADDONS_INNER_PANEL_IF
 
     // disable the use of evalTo for dense objects with a nice compilation error
     template<typename Dest>

Eigen/src/Core/DenseCoeffsBase.h

@@ -624,7 +624,7 @@ struct first_aligned_impl<Alignment, Derived, false>
 {
   static inline Index run(const Derived& m)
   {
-    return internal::first_aligned<Alignment>(&m.const_cast_derived().coeffRef(0,0), m.size());
+    return internal::first_aligned<Alignment>(m.data(), m.size());
   }
 };
 

Eigen/src/Core/DenseStorage.h

@@ -13,9 +13,9 @@
 #define EIGEN_MATRIXSTORAGE_H
 
 #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN
-  #define EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN EIGEN_DENSE_STORAGE_CTOR_PLUGIN;
+  #define EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN(X) X; EIGEN_DENSE_STORAGE_CTOR_PLUGIN;
 #else
-  #define EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN
+  #define EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN(X)
 #endif
 
 namespace Eigen {
@@ -184,12 +184,16 @@ template<typename T, int Size, int _Rows, int _Cols, int _Options> class DenseSt
 {
     internal::plain_array<T,Size,_Options> m_data;
   public:
-    EIGEN_DEVICE_FUNC DenseStorage() {}
+    EIGEN_DEVICE_FUNC DenseStorage() {
+      EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN(Index size = Size)
+    }
     EIGEN_DEVICE_FUNC
     explicit DenseStorage(internal::constructor_without_unaligned_array_assert)
       : m_data(internal::constructor_without_unaligned_array_assert()) {}
     EIGEN_DEVICE_FUNC 
-    DenseStorage(const DenseStorage& other) : m_data(other.m_data) {}
+    DenseStorage(const DenseStorage& other) : m_data(other.m_data) {
+      EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN(Index size = Size)
+    }
     EIGEN_DEVICE_FUNC 
     DenseStorage& operator=(const DenseStorage& other)
     { 
@@ -197,7 +201,7 @@ template<typename T, int Size, int _Rows, int _Cols, int _Options> class DenseSt
       return *this; 
     }
     EIGEN_DEVICE_FUNC DenseStorage(Index size, Index rows, Index cols) {
-      EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN
+      EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN({})
       eigen_internal_assert(size==rows*cols && rows==_Rows && cols==_Cols);
       EIGEN_UNUSED_VARIABLE(size);
       EIGEN_UNUSED_VARIABLE(rows);
@@ -343,7 +347,7 @@ template<typename T, int _Options> class DenseStorage<T, Dynamic, Dynamic, Dynam
     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_DENSE_STORAGE_CTOR_PLUGIN({})
       eigen_internal_assert(size==rows*cols && rows>=0 && cols >=0);
     }
     EIGEN_DEVICE_FUNC DenseStorage(const DenseStorage& other)
@@ -351,6 +355,7 @@ template<typename T, int _Options> class DenseStorage<T, Dynamic, Dynamic, Dynam
       , m_rows(other.m_rows)
       , m_cols(other.m_cols)
     {
+      EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN(Index size = m_rows*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)
@@ -403,7 +408,7 @@ template<typename T, int _Options> class DenseStorage<T, Dynamic, Dynamic, Dynam
           m_data = internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(size);
         else
           m_data = 0;
-        EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN
+        EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN({})
       }
       m_rows = rows;
       m_cols = cols;
@@ -422,7 +427,7 @@ template<typename T, int _Rows, int _Options> class DenseStorage<T, Dynamic, _Ro
     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_DENSE_STORAGE_CTOR_PLUGIN({})
       eigen_internal_assert(size==rows*cols && rows==_Rows && cols >=0);
       EIGEN_UNUSED_VARIABLE(rows);
     }
@@ -430,6 +435,7 @@ template<typename T, int _Rows, int _Options> class DenseStorage<T, Dynamic, _Ro
       : m_data(internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(_Rows*other.m_cols))
       , m_cols(other.m_cols)
     {
+      EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN(Index size = m_cols*_Rows)
       internal::smart_copy(other.m_data, other.m_data+_Rows*m_cols, m_data);
     }
     EIGEN_DEVICE_FUNC DenseStorage& operator=(const DenseStorage& other)
@@ -477,7 +483,7 @@ template<typename T, int _Rows, int _Options> class DenseStorage<T, Dynamic, _Ro
           m_data = internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(size);
         else
           m_data = 0;
-        EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN
+        EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN({})
       }
       m_cols = cols;
     }
@@ -495,7 +501,7 @@ template<typename T, int _Cols, int _Options> class DenseStorage<T, Dynamic, Dyn
     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_DENSE_STORAGE_CTOR_PLUGIN({})
       eigen_internal_assert(size==rows*cols && rows>=0 && cols == _Cols);
       EIGEN_UNUSED_VARIABLE(cols);
     }
@@ -503,6 +509,7 @@ template<typename T, int _Cols, int _Options> class DenseStorage<T, Dynamic, Dyn
       : m_data(internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(other.m_rows*_Cols))
       , m_rows(other.m_rows)
     {
+      EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN(Index size = m_rows*_Cols)
       internal::smart_copy(other.m_data, other.m_data+other.m_rows*_Cols, m_data);
     }
     EIGEN_DEVICE_FUNC DenseStorage& operator=(const DenseStorage& other)
@@ -550,7 +557,7 @@ template<typename T, int _Cols, int _Options> class DenseStorage<T, Dynamic, Dyn
           m_data = internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(size);
         else
           m_data = 0;
-        EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN
+        EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN({})
       }
       m_rows = rows;
     }

Eigen/src/Core/Diagonal.h

@@ -21,7 +21,7 @@ namespace Eigen {
   * \param MatrixType the type of the object in which we are taking a sub/main/super diagonal
   * \param DiagIndex the index of the sub/super diagonal. The default is 0 and it means the main diagonal.
   *              A positive value means a superdiagonal, a negative value means a subdiagonal.
-  *              You can also use Dynamic so the index can be set at runtime.
+  *              You can also use DynamicIndex so the index can be set at runtime.
   *
   * The matrix is not required to be square.
   *

Eigen/src/Core/DiagonalMatrix.h

@@ -290,12 +290,11 @@ MatrixBase<Derived>::asDiagonal() const
 template<typename Derived>
 bool MatrixBase<Derived>::isDiagonal(const RealScalar& prec) const
 {
-  using std::abs;
   if(cols() != rows()) return false;
   RealScalar maxAbsOnDiagonal = static_cast<RealScalar>(-1);
   for(Index j = 0; j < cols(); ++j)
   {
-    RealScalar absOnDiagonal = abs(coeff(j,j));
+    RealScalar absOnDiagonal = numext::abs(coeff(j,j));
     if(absOnDiagonal > maxAbsOnDiagonal) maxAbsOnDiagonal = absOnDiagonal;
   }
   for(Index j = 0; j < cols(); ++j)
@@ -321,6 +320,11 @@ struct Assignment<DstXprType, SrcXprType, Functor, Diagonal2Dense>
 {
   static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<typename DstXprType::Scalar,typename SrcXprType::Scalar> &/*func*/)
   {
+    Index dstRows = src.rows();
+    Index dstCols = src.cols();
+    if((dst.rows()!=dstRows) || (dst.cols()!=dstCols))
+      dst.resize(dstRows, dstCols);
+    
     dst.setZero();
     dst.diagonal() = src.diagonal();
   }

Eigen/src/Core/Dot.h

@@ -51,7 +51,8 @@ struct dot_nocheck<T, U, true>
 
 } // end namespace internal
 
-/** \returns the dot product of *this with other.
+/** \fn MatrixBase::dot
+  * \returns the dot product of *this with other.
   *
   * \only_for_vectors
   *
@@ -70,9 +71,11 @@ MatrixBase<Derived>::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)
+#if !(defined(EIGEN_NO_STATIC_ASSERT) && defined(EIGEN_NO_DEBUG))
   typedef internal::scalar_conj_product_op<Scalar,typename OtherDerived::Scalar> func;
   EIGEN_CHECK_BINARY_COMPATIBILIY(func,Scalar,typename OtherDerived::Scalar);
-
+#endif
+  
   eigen_assert(size() == other.size());
 
   return internal::dot_nocheck<Derived,OtherDerived>::run(*this, other);

Eigen/src/Core/GeneralProduct.h

@@ -25,7 +25,8 @@ template<int Rows, int Cols, int Depth> struct product_type_selector;
 template<int Size, int MaxSize> struct product_size_category
 {
   enum { is_large = MaxSize == Dynamic ||
-                    Size >= EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD,
+                    Size >= EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD ||
+                    (Size==Dynamic && MaxSize>=EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD),
          value = is_large  ? Large
                : Size == 1 ? 1
                            : Small
@@ -159,20 +160,20 @@ 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_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
-  // Some architectures cannot align on the stack,
-  // => let's manually enforce alignment by allocating more data and return the address of the first aligned element.
   enum {
     ForceAlignment  = internal::packet_traits<Scalar>::Vectorizable,
     PacketSize      = internal::packet_traits<Scalar>::size
   };
-  internal::plain_array<Scalar,EIGEN_SIZE_MIN_PREFER_FIXED(Size,MaxSize)+(ForceAlignment?PacketSize:0),0> m_data;
+  #if EIGEN_MAX_STATIC_ALIGN_BYTES!=0
+  internal::plain_array<Scalar,EIGEN_SIZE_MIN_PREFER_FIXED(Size,MaxSize),0,EIGEN_PLAIN_ENUM_MIN(AlignedMax,PacketSize)> m_data;
+  EIGEN_STRONG_INLINE Scalar* data() { return m_data.array; }
+  #else
+  // Some architectures cannot align on the stack,
+  // => let's manually enforce alignment by allocating more data and return the address of the first aligned element.
+  internal::plain_array<Scalar,EIGEN_SIZE_MIN_PREFER_FIXED(Size,MaxSize)+(ForceAlignment?EIGEN_MAX_ALIGN_BYTES:0),0> m_data;
   EIGEN_STRONG_INLINE Scalar* data() {
     return ForceAlignment
-            ? reinterpret_cast<Scalar*>((internal::UIntPtr(m_data.array) & ~(size_t(EIGEN_MAX_ALIGN_BYTES-1))) + EIGEN_MAX_ALIGN_BYTES)
+            ? reinterpret_cast<Scalar*>((internal::UIntPtr(m_data.array) & ~(std::size_t(EIGEN_MAX_ALIGN_BYTES-1))) + EIGEN_MAX_ALIGN_BYTES)
             : m_data.array;
   }
   #endif
@@ -207,7 +208,7 @@ template<> struct gemv_dense_selector<OnTheRight,ColMajor,true>
     typedef internal::blas_traits<Rhs> RhsBlasTraits;
     typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType;
   
-    typedef Map<Matrix<ResScalar,Dynamic,1>, Aligned> MappedDest;
+    typedef Map<Matrix<ResScalar,Dynamic,1>, EIGEN_PLAIN_ENUM_MIN(AlignedMax,internal::packet_traits<ResScalar>::size)> MappedDest;
 
     ActualLhsType actualLhs = LhsBlasTraits::extract(lhs);
     ActualRhsType actualRhs = RhsBlasTraits::extract(rhs);
@@ -223,50 +224,65 @@ template<> struct gemv_dense_selector<OnTheRight,ColMajor,true>
       // on, the other hand it is good for the cache to pack the vector anyways...
       EvalToDestAtCompileTime = (ActualDest::InnerStrideAtCompileTime==1),
       ComplexByReal = (NumTraits<LhsScalar>::IsComplex) && (!NumTraits<RhsScalar>::IsComplex),
-      MightCannotUseDest = (ActualDest::InnerStrideAtCompileTime!=1) || ComplexByReal
+      MightCannotUseDest = (!EvalToDestAtCompileTime) || ComplexByReal
     };
 
-    gemv_static_vector_if<ResScalar,ActualDest::SizeAtCompileTime,ActualDest::MaxSizeAtCompileTime,MightCannotUseDest> static_dest;
-
-    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
-      Index size = dest.size();
-      EIGEN_DENSE_STORAGE_CTOR_PLUGIN
-      #endif
-      if(!alphaIsCompatible)
-      {
-        MappedDest(actualDestPtr, dest.size()).setZero();
-        compatibleAlpha = RhsScalar(1);
-      }
-      else
-        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,LhsMapper,ColMajor,LhsBlasTraits::NeedToConjugate,RhsScalar,RhsMapper,RhsBlasTraits::NeedToConjugate>::run(
-        actualLhs.rows(), actualLhs.cols(),
-        LhsMapper(actualLhs.data(), actualLhs.outerStride()),
-        RhsMapper(actualRhs.data(), actualRhs.innerStride()),
-        actualDestPtr, 1,
-        compatibleAlpha);
+    RhsScalar compatibleAlpha = get_factor<ResScalar,RhsScalar>::run(actualAlpha);
 
-    if (!evalToDest)
+    if(!MightCannotUseDest)
     {
-      if(!alphaIsCompatible)
-        dest += actualAlpha * MappedDest(actualDestPtr, dest.size());
-      else
-        dest = MappedDest(actualDestPtr, dest.size());
+      // shortcut if we are sure to be able to use dest directly,
+      // this ease the compiler to generate cleaner and more optimzized code for most common cases
+      general_matrix_vector_product
+          <Index,LhsScalar,LhsMapper,ColMajor,LhsBlasTraits::NeedToConjugate,RhsScalar,RhsMapper,RhsBlasTraits::NeedToConjugate>::run(
+          actualLhs.rows(), actualLhs.cols(),
+          LhsMapper(actualLhs.data(), actualLhs.outerStride()),
+          RhsMapper(actualRhs.data(), actualRhs.innerStride()),
+          dest.data(), 1,
+          compatibleAlpha);
+    }
+    else
+    {
+      gemv_static_vector_if<ResScalar,ActualDest::SizeAtCompileTime,ActualDest::MaxSizeAtCompileTime,MightCannotUseDest> static_dest;
+
+      const bool alphaIsCompatible = (!ComplexByReal) || (numext::imag(actualAlpha)==RealScalar(0));
+      const bool evalToDest = EvalToDestAtCompileTime && alphaIsCompatible;
+
+      ei_declare_aligned_stack_constructed_variable(ResScalar,actualDestPtr,dest.size(),
+                                                    evalToDest ? dest.data() : static_dest.data());
+
+      if(!evalToDest)
+      {
+        #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN
+        Index size = dest.size();
+        EIGEN_DENSE_STORAGE_CTOR_PLUGIN
+        #endif
+        if(!alphaIsCompatible)
+        {
+          MappedDest(actualDestPtr, dest.size()).setZero();
+          compatibleAlpha = RhsScalar(1);
+        }
+        else
+          MappedDest(actualDestPtr, dest.size()) = dest;
+      }
+
+      general_matrix_vector_product
+          <Index,LhsScalar,LhsMapper,ColMajor,LhsBlasTraits::NeedToConjugate,RhsScalar,RhsMapper,RhsBlasTraits::NeedToConjugate>::run(
+          actualLhs.rows(), actualLhs.cols(),
+          LhsMapper(actualLhs.data(), actualLhs.outerStride()),
+          RhsMapper(actualRhs.data(), actualRhs.innerStride()),
+          actualDestPtr, 1,
+          compatibleAlpha);
+
+      if (!evalToDest)
+      {
+        if(!alphaIsCompatible)
+          dest.matrix() += actualAlpha * MappedDest(actualDestPtr, dest.size());
+        else
+          dest = MappedDest(actualDestPtr, dest.size());
+      }
     }
   }
 };
@@ -329,6 +345,7 @@ template<> struct gemv_dense_selector<OnTheRight,ColMajor,false>
   template<typename Lhs, typename Rhs, typename Dest>
   static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha)
   {
+    EIGEN_STATIC_ASSERT((!nested_eval<Lhs,1>::Evaluate),EIGEN_INTERNAL_COMPILATION_ERROR_OR_YOU_MADE_A_PROGRAMMING_MISTAKE);
     // TODO if rhs is large enough it might be beneficial to make sure that dest is sequentially stored in memory, otherwise use a temp
     typename nested_eval<Rhs,1>::type actual_rhs(rhs);
     const Index size = rhs.rows();
@@ -342,6 +359,7 @@ template<> struct gemv_dense_selector<OnTheRight,RowMajor,false>
   template<typename Lhs, typename Rhs, typename Dest>
   static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha)
   {
+    EIGEN_STATIC_ASSERT((!nested_eval<Lhs,1>::Evaluate),EIGEN_INTERNAL_COMPILATION_ERROR_OR_YOU_MADE_A_PROGRAMMING_MISTAKE);
     typename nested_eval<Rhs,Lhs::RowsAtCompileTime>::type actual_rhs(rhs);
     const Index rows = dest.rows();
     for(Index i=0; i<rows; ++i)

Eigen/src/Core/GenericPacketMath.h

@@ -329,7 +329,7 @@ template<typename Packet> EIGEN_DEVICE_FUNC inline typename unpacket_traits<Pack
   */
 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)
+predux_downto4(const Packet& a)
 { return a; }
 
 /** \internal \returns the product of the elements of \a a*/
@@ -558,6 +558,34 @@ pblend(const Selector<unpacket_traits<Packet>::size>& ifPacket, const Packet& th
   return ifPacket.select[0] ? thenPacket : elsePacket;
 }
 
+/** \internal \returns \a a with the first coefficient replaced by the scalar b */
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
+pinsertfirst(const Packet& a, typename unpacket_traits<Packet>::type b)
+{
+  // Default implementation based on pblend.
+  // It must be specialized for higher performance.
+  Selector<unpacket_traits<Packet>::size> mask;
+  mask.select[0] = true;
+  // This for loop should be optimized away by the compiler.
+  for(Index i=1; i<unpacket_traits<Packet>::size; ++i)
+    mask.select[i] = false;
+  return pblend(mask, pset1<Packet>(b), a);
+}
+
+/** \internal \returns \a a with the last coefficient replaced by the scalar b */
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
+pinsertlast(const Packet& a, typename unpacket_traits<Packet>::type b)
+{
+  // Default implementation based on pblend.
+  // It must be specialized for higher performance.
+  Selector<unpacket_traits<Packet>::size> mask;
+  // This for loop should be optimized away by the compiler.
+  for(Index i=0; i<unpacket_traits<Packet>::size-1; ++i)
+    mask.select[i] = false;
+  mask.select[unpacket_traits<Packet>::size-1] = true;
+  return pblend(mask, pset1<Packet>(b), a);
+}
+
 } // end namespace internal
 
 } // end namespace Eigen

Eigen/src/Core/IO.h

@@ -105,24 +105,10 @@ class WithFormat
     }
 
   protected:
-    const typename ExpressionType::Nested m_matrix;
+    typename ExpressionType::Nested m_matrix;
     IOFormat m_format;
 };
 
-/** \returns a WithFormat proxy object allowing to print a matrix the with given
-  * format \a fmt.
-  *
-  * See class IOFormat for some examples.
-  *
-  * \sa class IOFormat, class WithFormat
-  */
-template<typename Derived>
-inline const WithFormat<Derived>
-DenseBase<Derived>::format(const IOFormat& fmt) const
-{
-  return WithFormat<Derived>(derived(), fmt);
-}
-
 namespace internal {
 
 // NOTE: This helper is kept for backward compatibility with previous code specializing

Eigen/src/Core/Inverse.h

@@ -45,12 +45,13 @@ class Inverse : public InverseImpl<XprType,typename internal::traits<XprType>::S
 public:
   typedef typename XprType::StorageIndex StorageIndex;
   typedef typename XprType::PlainObject                       PlainObject;
+  typedef typename XprType::Scalar                            Scalar;
   typedef typename internal::ref_selector<XprType>::type      XprTypeNested;
   typedef typename internal::remove_all<XprTypeNested>::type  XprTypeNestedCleaned;
   typedef typename internal::ref_selector<Inverse>::type Nested;
   typedef typename internal::remove_all<XprType>::type NestedExpression;
   
-  explicit Inverse(const XprType &xpr)
+  explicit EIGEN_DEVICE_FUNC Inverse(const XprType &xpr)
     : m_xpr(xpr)
   {}
 

Eigen/src/Core/MapBase.h

@@ -26,7 +26,7 @@ namespace Eigen {
   * Typical users do not have to directly deal with this class.
   *
   * This class can be extended by through the macro plugin \c EIGEN_MAPBASE_PLUGIN.
-  * See \link TopicCustomizingEigen customizing Eigen \endlink for details.
+  * See \link TopicCustomizing_Plugins customizing Eigen \endlink for details.
   *
   * The \c Derived class has to provide the following two methods describing the memory layout:
   *  \code Index innerStride() const; \endcode

Eigen/src/Core/MathFunctions.h

@@ -97,6 +97,19 @@ struct real_default_impl<Scalar,true>
 
 template<typename Scalar> struct real_impl : real_default_impl<Scalar> {};
 
+#ifdef __CUDA_ARCH__
+template<typename T>
+struct real_impl<std::complex<T> >
+{
+  typedef T RealScalar;
+  EIGEN_DEVICE_FUNC
+  static inline T run(const std::complex<T>& x)
+  {
+    return x.real();
+  }
+};
+#endif
+
 template<typename Scalar>
 struct real_retval
 {
@@ -132,6 +145,19 @@ struct imag_default_impl<Scalar,true>
 
 template<typename Scalar> struct imag_impl : imag_default_impl<Scalar> {};
 
+#ifdef __CUDA_ARCH__
+template<typename T>
+struct imag_impl<std::complex<T> >
+{
+  typedef T RealScalar;
+  EIGEN_DEVICE_FUNC
+  static inline T run(const std::complex<T>& x)
+  {
+    return x.imag();
+  }
+};
+#endif
+
 template<typename Scalar>
 struct imag_retval
 {
@@ -459,30 +485,33 @@ struct arg_retval
 /****************************************************************************
 * Implementation of log1p                                                   *
 ****************************************************************************/
-template<typename Scalar, bool isComplex = NumTraits<Scalar>::IsComplex >
-struct log1p_impl
-{
-  static EIGEN_DEVICE_FUNC inline Scalar run(const Scalar& x)
-  {
+
+namespace std_fallback {
+  // fallback log1p implementation in case there is no log1p(Scalar) function in namespace of Scalar,
+  // or that there is no suitable std::log1p function available
+  template<typename Scalar>
+  EIGEN_DEVICE_FUNC inline Scalar log1p(const Scalar& x) {
     EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar)
     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 && !defined(__CUDACC__)
 template<typename Scalar>
-struct log1p_impl<Scalar, false> {
+struct log1p_impl {
   static inline Scalar run(const Scalar& x)
   {
     EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar)
+    #if EIGEN_HAS_CXX11_MATH
     using std::log1p;
+    #endif
+    using std_fallback::log1p;
     return log1p(x);
   }
 };
-#endif
+
 
 template<typename Scalar>
 struct log1p_retval
@@ -615,16 +644,18 @@ struct random_default_impl<Scalar, false, true>
     typedef typename conditional<NumTraits<Scalar>::IsSigned,std::ptrdiff_t,std::size_t>::type ScalarX;
     if(y<x)
       return x;
+    // the following difference might overflow on a 32 bits system,
+    // but since y>=x the result converted to an unsigned long is still correct.
     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);
-
+    std::size_t divisor = 1;
+    std::size_t multiplier = 1;
+    if(range<RAND_MAX) divisor = (std::size_t(RAND_MAX)+1)/(range+1);
+    else               multiplier = 1 + range/(std::size_t(RAND_MAX)+1);
     do {
-      offset = ( (std::size_t(std::rand()) * multiplier) / divisor );
+      offset = (std::size_t(std::rand()) * multiplier) / divisor;
     } while (offset > range);
-
     return Scalar(ScalarX(x) + offset);
   }
 
@@ -785,6 +816,8 @@ template<typename T> EIGEN_DEVICE_FUNC bool isfinite_impl(const std::complex<T>&
 template<typename T> EIGEN_DEVICE_FUNC bool isnan_impl(const std::complex<T>& x);
 template<typename T> EIGEN_DEVICE_FUNC bool isinf_impl(const std::complex<T>& x);
 
+template<typename T> T generic_fast_tanh_float(const T& a_x);
+
 } // end namespace internal
 
 /****************************************************************************
@@ -921,6 +954,14 @@ inline EIGEN_MATHFUNC_RETVAL(log1p, Scalar) log1p(const Scalar& x)
   return EIGEN_MATHFUNC_IMPL(log1p, Scalar)::run(x);
 }
 
+#ifdef __CUDACC__
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+float log1p(const float &x) { return ::log1pf(x); }
+
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+double log1p(const double &x) { return ::log1p(x); }
+#endif
+
 template<typename ScalarX,typename ScalarY>
 EIGEN_DEVICE_FUNC
 inline typename internal::pow_impl<ScalarX,ScalarY>::result_type pow(const ScalarX& x, const ScalarY& y)
@@ -1031,6 +1072,16 @@ float abs(const float &x) { return ::fabsf(x); }
 
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 double abs(const double &x) { return ::fabs(x); }
+
+template <> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+float abs(const std::complex<float>& x) {
+  return ::hypotf(x.real(), x.imag());
+}
+
+template <> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+double abs(const std::complex<double>& x) {
+  return ::hypot(x.real(), x.imag());
+}
 #endif
 
 template<typename T>
@@ -1176,6 +1227,11 @@ T tanh(const T &x) {
   return tanh(x);
 }
 
+#if (!defined(__CUDACC__)) && EIGEN_FAST_MATH
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+float tanh(float x) { return internal::generic_fast_tanh_float(x); }
+#endif
+
 #ifdef __CUDACC__
 template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 float tanh(const float &x) { return ::tanhf(x); }
@@ -1282,11 +1338,12 @@ template<typename Scalar>
 struct scalar_fuzzy_default_impl<Scalar, true, 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)
   {
     return numext::abs2(x) <= numext::abs2(y) * prec * prec;
   }
+  EIGEN_DEVICE_FUNC
   static inline bool isApprox(const Scalar& x, const Scalar& y, const RealScalar& prec)
   {
     return numext::abs2(x - y) <= numext::mini(numext::abs2(x), numext::abs2(y)) * prec * prec;

Eigen/src/Core/MathFunctionsImpl.h

@@ -0,0 +1,78 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Pedro Gonnet (pedro.gonnet@gmail.com)
+// Copyright (C) 2016 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_MATHFUNCTIONSIMPL_H
+#define EIGEN_MATHFUNCTIONSIMPL_H
+
+namespace Eigen {
+
+namespace internal {
+
+/** \internal \returns the hyperbolic tan of \a a (coeff-wise)
+    Doesn't do anything fancy, just a 13/6-degree rational interpolant which
+    is accurate up to a couple of ulp in the range [-9, 9], outside of which
+    the tanh(x) = +/-1.
+
+    This implementation works on both scalars and packets.
+*/
+template<typename T>
+T generic_fast_tanh_float(const T& a_x)
+{
+  // Clamp the inputs to the range [-9, 9] since anything outside
+  // this range is +/-1.0f in single-precision.
+  const T plus_9 = pset1<T>(9.f);
+  const T minus_9 = pset1<T>(-9.f);
+  // NOTE GCC prior to 6.3 might improperly optimize this max/min
+  //      step such that if a_x is nan, x will be either 9 or -9,
+  //      and tanh will return 1 or -1 instead of nan.
+  //      This is supposed to be fixed in gcc6.3,
+  //      see: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=72867
+  const T x = pmax(minus_9,pmin(plus_9,a_x));
+  // The monomial coefficients of the numerator polynomial (odd).
+  const T alpha_1 = pset1<T>(4.89352455891786e-03f);
+  const T alpha_3 = pset1<T>(6.37261928875436e-04f);
+  const T alpha_5 = pset1<T>(1.48572235717979e-05f);
+  const T alpha_7 = pset1<T>(5.12229709037114e-08f);
+  const T alpha_9 = pset1<T>(-8.60467152213735e-11f);
+  const T alpha_11 = pset1<T>(2.00018790482477e-13f);
+  const T alpha_13 = pset1<T>(-2.76076847742355e-16f);
+
+  // The monomial coefficients of the denominator polynomial (even).
+  const T beta_0 = pset1<T>(4.89352518554385e-03f);
+  const T beta_2 = pset1<T>(2.26843463243900e-03f);
+  const T beta_4 = pset1<T>(1.18534705686654e-04f);
+  const T beta_6 = pset1<T>(1.19825839466702e-06f);
+
+  // Since the polynomials are odd/even, we need x^2.
+  const T x2 = pmul(x, x);
+
+  // Evaluate the numerator polynomial p.
+  T p = pmadd(x2, alpha_13, alpha_11);
+  p = pmadd(x2, p, alpha_9);
+  p = pmadd(x2, p, alpha_7);
+  p = pmadd(x2, p, alpha_5);
+  p = pmadd(x2, p, alpha_3);
+  p = pmadd(x2, p, alpha_1);
+  p = pmul(x, p);
+
+  // Evaluate the denominator polynomial p.
+  T q = pmadd(x2, beta_6, beta_4);
+  q = pmadd(x2, q, beta_2);
+  q = pmadd(x2, q, beta_0);
+
+  // Divide the numerator by the denominator.
+  return pdiv(p, q);
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_MATHFUNCTIONSIMPL_H

Eigen/src/Core/Matrix.h

@@ -106,7 +106,7 @@ public:
   * \endcode
   *
   * 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_MATRIX_PLUGIN.
+  * \ref TopicCustomizing_Plugins by defining the preprocessor symbol \c EIGEN_MATRIX_PLUGIN.
   *
   * <i><b>Some notes:</b></i>
   *

Eigen/src/Core/MatrixBase.h

@@ -41,7 +41,7 @@ namespace Eigen {
   * \endcode
   *
   * 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_MATRIXBASE_PLUGIN.
+  * \ref TopicCustomizing_Plugins by defining the preprocessor symbol \c EIGEN_MATRIXBASE_PLUGIN.
   *
   * \sa \blank \ref TopicClassHierarchy
   */
@@ -98,7 +98,7 @@ 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()); }
+    inline Index diagonalSize() const { return (numext::mini)(rows(),cols()); }
 
     typedef typename Base::PlainObject PlainObject;
 
@@ -121,6 +121,7 @@ template<typename Derived> class MatrixBase
 #endif // not EIGEN_PARSED_BY_DOXYGEN
 
 #define EIGEN_CURRENT_STORAGE_BASE_CLASS Eigen::MatrixBase
+#define EIGEN_DOC_UNARY_ADDONS(X,Y)
 #   include "../plugins/CommonCwiseUnaryOps.h"
 #   include "../plugins/CommonCwiseBinaryOps.h"
 #   include "../plugins/MatrixCwiseUnaryOps.h"
@@ -129,6 +130,7 @@ template<typename Derived> class MatrixBase
 #     include EIGEN_MATRIXBASE_PLUGIN
 #   endif
 #undef EIGEN_CURRENT_STORAGE_BASE_CLASS
+#undef EIGEN_DOC_UNARY_ADDONS
 
     /** Special case of the template operator=, in order to prevent the compiler
       * from generating a default operator= (issue hit with g++ 4.1)
@@ -328,15 +330,11 @@ template<typename Derived> class MatrixBase
 
 /////////// LU module ///////////
 
-    EIGEN_DEVICE_FUNC
     inline const FullPivLU<PlainObject> fullPivLu() const;
-    EIGEN_DEVICE_FUNC
     inline const PartialPivLU<PlainObject> partialPivLu() const;
 
-    EIGEN_DEVICE_FUNC
     inline const PartialPivLU<PlainObject> lu() const;
 
-    EIGEN_DEVICE_FUNC
     inline const Inverse<Derived> inverse() const;
 
     template<typename ResultType>
@@ -401,12 +399,14 @@ template<typename Derived> class MatrixBase
     EIGEN_DEVICE_FUNC
     inline PlainObject unitOrthogonal(void) const;
 
+    EIGEN_DEVICE_FUNC
     inline Matrix<Scalar,3,1> eulerAngles(Index a0, Index a1, Index a2) const;
 
     // put this as separate enum value to work around possible GCC 4.3 bug (?)
     enum { HomogeneousReturnTypeDirection = ColsAtCompileTime==1&&RowsAtCompileTime==1 ? ((internal::traits<Derived>::Flags&RowMajorBit)==RowMajorBit ? Horizontal : Vertical)
                                           : ColsAtCompileTime==1 ? Vertical : Horizontal };
     typedef Homogeneous<Derived, HomogeneousReturnTypeDirection> HomogeneousReturnType;
+    EIGEN_DEVICE_FUNC
     inline HomogeneousReturnType homogeneous() const;
 
     enum {
@@ -416,7 +416,7 @@ template<typename Derived> class MatrixBase
                   internal::traits<Derived>::ColsAtCompileTime==1 ? SizeMinusOne : 1,
                   internal::traits<Derived>::ColsAtCompileTime==1 ? 1 : SizeMinusOne> ConstStartMinusOne;
     typedef EIGEN_EXPR_BINARYOP_SCALAR_RETURN_TYPE(ConstStartMinusOne,Scalar,quotient) HNormalizedReturnType;
-
+    EIGEN_DEVICE_FUNC
     inline const HNormalizedReturnType hnormalized() const;
 
 ////////// Householder module ///////////

Eigen/src/Core/NumTraits.h

@@ -97,23 +97,6 @@ template<typename T> struct GenericNumTraits
     MulCost = 1
   };
 
-  // Division is messy but important, because it is expensive and throughput
-  // varies significantly. The following numbers are based on min division
-  // throughput on Haswell.
-  template<bool Vectorized>
-  struct Div {
-    enum {
-#ifdef EIGEN_VECTORIZE_AVX
-      AVX = true,
-#else
-      AVX = false,
-#endif
-      Cost = IsInteger ? (sizeof(T) == 8 ? (IsSigned ? 24 : 21) : (IsSigned ? 8 : 9)):
-          Vectorized ? (sizeof(T) == 8 ? (AVX ? 16 : 8) : (AVX ? 14 : 7)) : 8
-    };
-  };
-
-
   typedef T Real;
   typedef typename internal::conditional<
                      IsInteger,
@@ -255,6 +238,9 @@ private:
   static inline std::string quiet_NaN();
 };
 
+// Empty specialization for void to allow template specialization based on NumTraits<T>::Real with T==void and SFINAE.
+template<> struct NumTraits<void> {};
+
 } // end namespace Eigen
 
 #endif // EIGEN_NUMTRAITS_H

Eigen/src/Core/PlainObjectBase.h

@@ -41,7 +41,7 @@ template<> struct check_rows_cols_for_overflow<Dynamic> {
   {
     // http://hg.mozilla.org/mozilla-central/file/6c8a909977d3/xpcom/ds/CheckedInt.h#l242
     // we assume Index is signed
-    Index max_index = (size_t(1) << (8 * sizeof(Index) - 1)) - 1; // assume Index is signed
+    Index max_index = (std::size_t(1) << (8 * sizeof(Index) - 1)) - 1; // assume Index is signed
     bool error = (rows == 0 || cols == 0) ? false
                : (rows > max_index / cols);
     if (error)
@@ -58,33 +58,39 @@ template<typename MatrixTypeA, typename MatrixTypeB, bool SwapPointers> struct m
 
 } // end namespace internal
 
-/** \class PlainObjectBase
-  * \ingroup Core_Module
-  * \brief %Dense storage base class for matrices and arrays.
-  *
-  * 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_PLAINOBJECTBASE_PLUGIN.
-  *
-  * \sa \ref TopicClassHierarchy
-  */
 #ifdef EIGEN_PARSED_BY_DOXYGEN
 namespace doxygen {
 
-// this is a workaround to doxygen not being able to understand the inheritance 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.
+// Moreover, doxygen fails to include members that are not documented in the declaration body of
+// MatrixBase if we inherits MatrixBase<Matrix<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> >,
+// this is why we simply inherits MatrixBase, though this does not make sense.
+
 /** This class is just a workaround for Doxygen and it does not not actually exist. */
 template<typename Derived> struct dense_xpr_base_dispatcher;
 /** 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>
 struct dense_xpr_base_dispatcher<Matrix<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> >
-    : public MatrixBase<Matrix<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> > {};
+    : public MatrixBase {};
 /** 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>
 struct dense_xpr_base_dispatcher<Array<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> >
-    : public ArrayBase<Array<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> > {};
+    : public ArrayBase {};
 
 } // namespace doxygen
 
+/** \class PlainObjectBase
+  * \ingroup Core_Module
+  * \brief %Dense storage base class for matrices and arrays.
+  *
+  * This class can be extended with the help of the plugin mechanism described on the page
+  * \ref TopicCustomizing_Plugins by defining the preprocessor symbol \c EIGEN_PLAINOBJECTBASE_PLUGIN.
+  *
+  * \tparam Derived is the derived type, e.g., a Matrix or Array
+  *
+  * \sa \ref TopicClassHierarchy
+  */
 template<typename Derived>
 class PlainObjectBase : public doxygen::dense_xpr_base_dispatcher<Derived>
 #else
@@ -554,7 +560,8 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
 
   public:
 
-    /** \copydoc DenseBase::operator=(const EigenBase<OtherDerived>&)
+    /** \brief Copies the generic expression \a other into *this.
+      * \copydetails DenseBase::operator=(const EigenBase<OtherDerived> &other)
       */
     template<typename OtherDerived>
     EIGEN_DEVICE_FUNC 
@@ -763,6 +770,7 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
     {
       // 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_UNUSED_VARIABLE(is_integer);
       EIGEN_STATIC_ASSERT(is_integer,
                           FLOATING_POINT_ARGUMENT_PASSED__INTEGER_WAS_EXPECTED)
       resize(size);
@@ -804,6 +812,13 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       this->_set_noalias(other);
     }
 
+    // Initialize an arbitrary matrix from an object convertible to the Derived type.
+    template<typename T>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void _init1(const Derived& other){
+      this->_set_noalias(other);
+    }
+
     // Initialize an arbitrary matrix from a generic Eigen expression
     template<typename T, typename OtherDerived>
     EIGEN_DEVICE_FUNC
@@ -826,7 +841,7 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       this->derived() = r;
     }
     
-    // For fixed -size arrays:
+    // For fixed-size Array<Scalar,...>
     template<typename T>
     EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE void _init1(const Scalar& val0,
@@ -838,6 +853,7 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       Base::setConstant(val0);
     }
     
+    // For fixed-size Array<Index,...>
     template<typename T>
     EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE void _init1(const Index& val0,
@@ -916,8 +932,8 @@ struct conservative_resize_like_impl
     {
       // The storage order does not allow us to use reallocation.
       typename Derived::PlainObject tmp(rows,cols);
-      const Index common_rows = (std::min)(rows, _this.rows());
-      const Index common_cols = (std::min)(cols, _this.cols());
+      const Index common_rows = numext::mini(rows, _this.rows());
+      const Index common_cols = numext::mini(cols, _this.cols());
       tmp.block(0,0,common_rows,common_cols) = _this.block(0,0,common_rows,common_cols);
       _this.derived().swap(tmp);
     }
@@ -950,8 +966,8 @@ struct conservative_resize_like_impl
     {
       // The storage order does not allow us to use reallocation.
       typename Derived::PlainObject tmp(other);
-      const Index common_rows = (std::min)(tmp.rows(), _this.rows());
-      const Index common_cols = (std::min)(tmp.cols(), _this.cols());
+      const Index common_rows = numext::mini(tmp.rows(), _this.rows());
+      const Index common_cols = numext::mini(tmp.cols(), _this.cols());
       tmp.block(0,0,common_rows,common_cols) = _this.block(0,0,common_rows,common_cols);
       _this.derived().swap(tmp);
     }

Eigen/src/Core/ProductEvaluators.h

@@ -140,6 +140,10 @@ struct Assignment<DstXprType, Product<Lhs,Rhs,Options>, internal::assign_op<Scal
   static EIGEN_STRONG_INLINE
   void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<Scalar,Scalar> &)
   {
+    Index dstRows = src.rows();
+    Index dstCols = src.cols();
+    if((dst.rows()!=dstRows) || (dst.cols()!=dstCols))
+      dst.resize(dstRows, dstCols);
     // FIXME shall we handle nested_eval here?
     generic_product_impl<Lhs, Rhs>::evalTo(dst, src.lhs(), src.rhs());
   }
@@ -154,6 +158,7 @@ struct Assignment<DstXprType, Product<Lhs,Rhs,Options>, internal::add_assign_op<
   static EIGEN_STRONG_INLINE
   void run(DstXprType &dst, const SrcXprType &src, const internal::add_assign_op<Scalar,Scalar> &)
   {
+    eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols());
     // FIXME shall we handle nested_eval here?
     generic_product_impl<Lhs, Rhs>::addTo(dst, src.lhs(), src.rhs());
   }
@@ -168,6 +173,7 @@ struct Assignment<DstXprType, Product<Lhs,Rhs,Options>, internal::sub_assign_op<
   static EIGEN_STRONG_INLINE
   void run(DstXprType &dst, const SrcXprType &src, const internal::sub_assign_op<Scalar,Scalar> &)
   {
+    eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols());
     // FIXME shall we handle nested_eval here?
     generic_product_impl<Lhs, Rhs>::subTo(dst, src.lhs(), src.rhs());
   }
@@ -194,7 +200,6 @@ struct Assignment<DstXprType, CwiseBinaryOp<internal::scalar_product_op<ScalarBi
 //----------------------------------------
 // Catch "Dense ?= xpr + Product<>" expression to save one temporary
 // FIXME we could probably enable these rules for any product, i.e., not only Dense and DefaultProduct
-// TODO enable it for "Dense ?= xpr - Product<>" as well.
 
 template<typename OtherXpr, typename Lhs, typename Rhs>
 struct evaluator_assume_aliasing<CwiseBinaryOp<internal::scalar_sum_op<typename OtherXpr::Scalar,typename Product<Lhs,Rhs,DefaultProduct>::Scalar>, const OtherXpr,
@@ -203,10 +208,9 @@ struct evaluator_assume_aliasing<CwiseBinaryOp<internal::scalar_sum_op<typename
 };
 
 template<typename DstXprType, typename OtherXpr, typename ProductType, typename Func1, typename Func2>
-struct assignment_from_xpr_plus_product
+struct assignment_from_xpr_op_product
 {
-  typedef CwiseBinaryOp<internal::scalar_sum_op<typename OtherXpr::Scalar,typename ProductType::Scalar>, const OtherXpr, const ProductType> SrcXprType;
-  template<typename InitialFunc>
+  template<typename SrcXprType, typename InitialFunc>
   static EIGEN_STRONG_INLINE
   void run(DstXprType &dst, const SrcXprType &src, const InitialFunc& /*func*/)
   {
@@ -215,21 +219,21 @@ struct assignment_from_xpr_plus_product
   }
 };
 
-template< typename DstXprType, typename OtherXpr, typename Lhs, typename Rhs, typename DstScalar, typename SrcScalar, typename OtherScalar,typename ProdScalar>
-struct Assignment<DstXprType, CwiseBinaryOp<internal::scalar_sum_op<OtherScalar,ProdScalar>, const OtherXpr,
-                                           const Product<Lhs,Rhs,DefaultProduct> >, internal::assign_op<DstScalar,SrcScalar>, Dense2Dense>
-  : assignment_from_xpr_plus_product<DstXprType, OtherXpr, Product<Lhs,Rhs,DefaultProduct>, internal::assign_op<DstScalar,OtherScalar>, internal::add_assign_op<DstScalar,ProdScalar> >
-{};
-template< typename DstXprType, typename OtherXpr, typename Lhs, typename Rhs, typename DstScalar, typename SrcScalar, typename OtherScalar,typename ProdScalar>
-struct Assignment<DstXprType, CwiseBinaryOp<internal::scalar_sum_op<OtherScalar,ProdScalar>, const OtherXpr,
-                                           const Product<Lhs,Rhs,DefaultProduct> >, internal::add_assign_op<DstScalar,SrcScalar>, Dense2Dense>
-  : assignment_from_xpr_plus_product<DstXprType, OtherXpr, Product<Lhs,Rhs,DefaultProduct>, internal::add_assign_op<DstScalar,OtherScalar>, internal::add_assign_op<DstScalar,ProdScalar> >
-{};
-template< typename DstXprType, typename OtherXpr, typename Lhs, typename Rhs, typename DstScalar, typename SrcScalar, typename OtherScalar,typename ProdScalar>
-struct Assignment<DstXprType, CwiseBinaryOp<internal::scalar_sum_op<OtherScalar,ProdScalar>, const OtherXpr,
-                                           const Product<Lhs,Rhs,DefaultProduct> >, internal::sub_assign_op<DstScalar,SrcScalar>, Dense2Dense>
-  : assignment_from_xpr_plus_product<DstXprType, OtherXpr, Product<Lhs,Rhs,DefaultProduct>, internal::sub_assign_op<DstScalar,OtherScalar>, internal::sub_assign_op<DstScalar,ProdScalar> >
-{};
+#define EIGEN_CATCH_ASSIGN_XPR_OP_PRODUCT(ASSIGN_OP,BINOP,ASSIGN_OP2) \
+  template< typename DstXprType, typename OtherXpr, typename Lhs, typename Rhs, typename DstScalar, typename SrcScalar, typename OtherScalar,typename ProdScalar> \
+  struct Assignment<DstXprType, CwiseBinaryOp<internal::BINOP<OtherScalar,ProdScalar>, const OtherXpr, \
+                                            const Product<Lhs,Rhs,DefaultProduct> >, internal::ASSIGN_OP<DstScalar,SrcScalar>, Dense2Dense> \
+    : assignment_from_xpr_op_product<DstXprType, OtherXpr, Product<Lhs,Rhs,DefaultProduct>, internal::ASSIGN_OP<DstScalar,OtherScalar>, internal::ASSIGN_OP2<DstScalar,ProdScalar> > \
+  {}
+
+EIGEN_CATCH_ASSIGN_XPR_OP_PRODUCT(assign_op,    scalar_sum_op,add_assign_op);
+EIGEN_CATCH_ASSIGN_XPR_OP_PRODUCT(add_assign_op,scalar_sum_op,add_assign_op);
+EIGEN_CATCH_ASSIGN_XPR_OP_PRODUCT(sub_assign_op,scalar_sum_op,sub_assign_op);
+
+EIGEN_CATCH_ASSIGN_XPR_OP_PRODUCT(assign_op,    scalar_difference_op,sub_assign_op);
+EIGEN_CATCH_ASSIGN_XPR_OP_PRODUCT(add_assign_op,scalar_difference_op,sub_assign_op);
+EIGEN_CATCH_ASSIGN_XPR_OP_PRODUCT(sub_assign_op,scalar_difference_op,add_assign_op);
+
 //----------------------------------------
 
 template<typename Lhs, typename Rhs>
@@ -267,7 +271,7 @@ void outer_product_selector_run(Dst& dst, const Lhs &lhs, const Rhs &rhs, const
   // FIXME not very good if rhs is real and lhs complex while alpha is real too
   const Index cols = dst.cols();
   for (Index j=0; j<cols; ++j)
-    func(dst.col(j), rhsEval.coeff(0,j) * actual_lhs);
+    func(dst.col(j), rhsEval.coeff(Index(0),j) * actual_lhs);
 }
 
 // Row major result
@@ -280,7 +284,7 @@ void outer_product_selector_run(Dst& dst, const Lhs &lhs, const Rhs &rhs, const
   // FIXME not very good if lhs is real and rhs complex while alpha is real too
   const Index rows = dst.rows();
   for (Index i=0; i<rows; ++i)
-    func(dst.row(i), lhsEval.coeff(i,0) * actual_rhs);
+    func(dst.row(i), lhsEval.coeff(i,Index(0)) * actual_rhs);
 }
 
 template<typename Lhs, typename Rhs>
@@ -356,17 +360,21 @@ template<typename Lhs, typename Rhs>
 struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,GemvProduct>
   : generic_product_impl_base<Lhs,Rhs,generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,GemvProduct> >
 {
+  typedef typename nested_eval<Lhs,1>::type LhsNested;
+  typedef typename nested_eval<Rhs,1>::type RhsNested;
   typedef typename Product<Lhs,Rhs>::Scalar Scalar;
   enum { Side = Lhs::IsVectorAtCompileTime ? OnTheLeft : OnTheRight };
-  typedef typename internal::conditional<int(Side)==OnTheRight,Lhs,Rhs>::type MatrixType;
+  typedef typename internal::remove_all<typename internal::conditional<int(Side)==OnTheRight,LhsNested,RhsNested>::type>::type MatrixType;
 
   template<typename Dest>
   static EIGEN_STRONG_INLINE void scaleAndAddTo(Dest& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha)
   {
+    LhsNested actual_lhs(lhs);
+    RhsNested actual_rhs(rhs);
     internal::gemv_dense_selector<Side,
                             (int(MatrixType::Flags)&RowMajorBit) ? RowMajor : ColMajor,
                             bool(internal::blas_traits<MatrixType>::HasUsableDirectAccess)
-                           >::run(lhs, rhs, dst, alpha);
+                           >::run(actual_lhs, actual_rhs, dst, alpha);
   }
 };
 
@@ -439,6 +447,18 @@ struct product_evaluator<Product<Lhs, Rhs, LazyProduct>, ProductTag, DenseShape,
     EIGEN_INTERNAL_CHECK_COST_VALUE(NumTraits<Scalar>::MulCost);
     EIGEN_INTERNAL_CHECK_COST_VALUE(NumTraits<Scalar>::AddCost);
     EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+#if 0
+    std::cerr << "LhsOuterStrideBytes=  " << LhsOuterStrideBytes << "\n";
+    std::cerr << "RhsOuterStrideBytes=  " << RhsOuterStrideBytes << "\n";
+    std::cerr << "LhsAlignment=         " << LhsAlignment << "\n";
+    std::cerr << "RhsAlignment=         " << RhsAlignment << "\n";
+    std::cerr << "CanVectorizeLhs=      " << CanVectorizeLhs << "\n";
+    std::cerr << "CanVectorizeRhs=      " << CanVectorizeRhs << "\n";
+    std::cerr << "CanVectorizeInner=    " << CanVectorizeInner << "\n";
+    std::cerr << "EvalToRowMajor=       " << EvalToRowMajor << "\n";
+    std::cerr << "Alignment=            " << Alignment << "\n";
+    std::cerr << "Flags=                " << Flags << "\n";
+#endif
   }
 
   // Everything below here is taken from CoeffBasedProduct.h
@@ -489,11 +509,8 @@ struct product_evaluator<Product<Lhs, Rhs, LazyProduct>, ProductTag, DenseShape,
       
     SameType = is_same<typename LhsNestedCleaned::Scalar,typename RhsNestedCleaned::Scalar>::value,
 
-    CanVectorizeRhs = bool(RhsRowMajor) && (RhsFlags & PacketAccessBit)
-                    && (ColsAtCompileTime == Dynamic || ((ColsAtCompileTime % RhsVecPacketSize) == 0) ),
-
-    CanVectorizeLhs = (!LhsRowMajor) && (LhsFlags & PacketAccessBit)
-                    && (RowsAtCompileTime == Dynamic || ((RowsAtCompileTime % LhsVecPacketSize) == 0) ),
+    CanVectorizeRhs = bool(RhsRowMajor) && (RhsFlags & PacketAccessBit) && (ColsAtCompileTime!=1),
+    CanVectorizeLhs = (!LhsRowMajor) && (LhsFlags & PacketAccessBit) && (RowsAtCompileTime!=1),
 
     EvalToRowMajor = (MaxRowsAtCompileTime==1&&MaxColsAtCompileTime!=1) ? 1
                     : (MaxColsAtCompileTime==1&&MaxRowsAtCompileTime!=1) ? 0
@@ -508,8 +525,8 @@ struct product_evaluator<Product<Lhs, Rhs, LazyProduct>, ProductTag, DenseShape,
     LhsOuterStrideBytes = int(LhsNestedCleaned::OuterStrideAtCompileTime) * int(sizeof(typename LhsNestedCleaned::Scalar)),
     RhsOuterStrideBytes = int(RhsNestedCleaned::OuterStrideAtCompileTime) * int(sizeof(typename RhsNestedCleaned::Scalar)),
 
-    Alignment = bool(CanVectorizeLhs) ? (LhsOuterStrideBytes<0 || (int(LhsOuterStrideBytes) % EIGEN_PLAIN_ENUM_MAX(1,LhsAlignment))!=0 ? 0 : LhsAlignment)
-              : bool(CanVectorizeRhs) ? (RhsOuterStrideBytes<0 || (int(RhsOuterStrideBytes) % EIGEN_PLAIN_ENUM_MAX(1,RhsAlignment))!=0 ? 0 : RhsAlignment)
+    Alignment = bool(CanVectorizeLhs) ? (LhsOuterStrideBytes<=0 || (int(LhsOuterStrideBytes) % EIGEN_PLAIN_ENUM_MAX(1,LhsAlignment))!=0 ? 0 : LhsAlignment)
+              : bool(CanVectorizeRhs) ? (RhsOuterStrideBytes<=0 || (int(RhsOuterStrideBytes) % EIGEN_PLAIN_ENUM_MAX(1,RhsAlignment))!=0 ? 0 : RhsAlignment)
               : 0,
 
     /* CanVectorizeInner deserves special explanation. It does not affect the product flags. It is not used outside
@@ -535,8 +552,8 @@ struct product_evaluator<Product<Lhs, Rhs, LazyProduct>, ProductTag, DenseShape,
    */
   EIGEN_DEVICE_FUNC const CoeffReturnType coeff(Index index) const
   {
-    const Index row = RowsAtCompileTime == 1 ? 0 : index;
-    const Index col = RowsAtCompileTime == 1 ? index : 0;
+    const Index row = (RowsAtCompileTime == 1 || MaxRowsAtCompileTime==1) ? 0 : index;
+    const Index col = (RowsAtCompileTime == 1 || MaxRowsAtCompileTime==1) ? index : 0;
     return (m_lhs.row(row).transpose().cwiseProduct( m_rhs.col(col) )).sum();
   }
 
@@ -554,14 +571,14 @@ struct product_evaluator<Product<Lhs, Rhs, LazyProduct>, ProductTag, DenseShape,
   template<int LoadMode, typename PacketType>
   const PacketType packet(Index index) const
   {
-    const Index row = RowsAtCompileTime == 1 ? 0 : index;
-    const Index col = RowsAtCompileTime == 1 ? index : 0;
+    const Index row = (RowsAtCompileTime == 1 || MaxRowsAtCompileTime==1) ? 0 : index;
+    const Index col = (RowsAtCompileTime == 1 || MaxRowsAtCompileTime==1) ? index : 0;
     return packet<LoadMode,PacketType>(row,col);
   }
 
 protected:
-  const LhsNested m_lhs;
-  const RhsNested m_rhs;
+  typename internal::add_const_on_value_type<LhsNested>::type m_lhs;
+  typename internal::add_const_on_value_type<RhsNested>::type m_rhs;
   
   LhsEtorType m_lhsImpl;
   RhsEtorType m_rhsImpl;
@@ -595,7 +612,7 @@ struct etor_product_packet_impl<RowMajor, UnrollingIndex, Lhs, Rhs, Packet, Load
   static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Index innerDim, Packet &res)
   {
     etor_product_packet_impl<RowMajor, UnrollingIndex-1, Lhs, Rhs, Packet, LoadMode>::run(row, col, lhs, rhs, innerDim, res);
-    res =  pmadd(pset1<Packet>(lhs.coeff(row, UnrollingIndex-1)), rhs.template packet<LoadMode,Packet>(UnrollingIndex-1, col), res);
+    res =  pmadd(pset1<Packet>(lhs.coeff(row, Index(UnrollingIndex-1))), rhs.template packet<LoadMode,Packet>(Index(UnrollingIndex-1), col), res);
   }
 };
 
@@ -605,7 +622,7 @@ struct etor_product_packet_impl<ColMajor, UnrollingIndex, Lhs, Rhs, Packet, Load
   static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Index innerDim, Packet &res)
   {
     etor_product_packet_impl<ColMajor, UnrollingIndex-1, Lhs, Rhs, Packet, LoadMode>::run(row, col, lhs, rhs, innerDim, res);
-    res =  pmadd(lhs.template packet<LoadMode,Packet>(row, UnrollingIndex-1), pset1<Packet>(rhs.coeff(UnrollingIndex-1, col)), res);
+    res =  pmadd(lhs.template packet<LoadMode,Packet>(row, Index(UnrollingIndex-1)), pset1<Packet>(rhs.coeff(Index(UnrollingIndex-1), col)), res);
   }
 };
 
@@ -614,7 +631,7 @@ struct etor_product_packet_impl<RowMajor, 1, Lhs, Rhs, Packet, LoadMode>
 {
   static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Index /*innerDim*/, Packet &res)
   {
-    res = pmul(pset1<Packet>(lhs.coeff(row, 0)),rhs.template packet<LoadMode,Packet>(0, col));
+    res = pmul(pset1<Packet>(lhs.coeff(row, Index(0))),rhs.template packet<LoadMode,Packet>(Index(0), col));
   }
 };
 
@@ -623,7 +640,7 @@ struct etor_product_packet_impl<ColMajor, 1, Lhs, Rhs, Packet, LoadMode>
 {
   static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Index /*innerDim*/, Packet &res)
   {
-    res = pmul(lhs.template packet<LoadMode,Packet>(row, 0), pset1<Packet>(rhs.coeff(0, col)));
+    res = pmul(lhs.template packet<LoadMode,Packet>(row, Index(0)), pset1<Packet>(rhs.coeff(Index(0), col)));
   }
 };
 
@@ -632,7 +649,7 @@ struct etor_product_packet_impl<RowMajor, 0, Lhs, Rhs, Packet, LoadMode>
 {
   static EIGEN_STRONG_INLINE void run(Index /*row*/, Index /*col*/, const Lhs& /*lhs*/, const Rhs& /*rhs*/, Index /*innerDim*/, Packet &res)
   {
-    res = pset1<Packet>(0);
+    res = pset1<Packet>(typename unpacket_traits<Packet>::type(0));
   }
 };
 
@@ -641,7 +658,7 @@ struct etor_product_packet_impl<ColMajor, 0, Lhs, Rhs, Packet, LoadMode>
 {
   static EIGEN_STRONG_INLINE void run(Index /*row*/, Index /*col*/, const Lhs& /*lhs*/, const Rhs& /*rhs*/, Index /*innerDim*/, Packet &res)
   {
-    res = pset1<Packet>(0);
+    res = pset1<Packet>(typename unpacket_traits<Packet>::type(0));
   }
 };
 
@@ -650,7 +667,7 @@ struct etor_product_packet_impl<RowMajor, Dynamic, Lhs, Rhs, Packet, LoadMode>
 {
   static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Index innerDim, Packet& res)
   {
-    res = pset1<Packet>(0);
+    res = pset1<Packet>(typename unpacket_traits<Packet>::type(0));
     for(Index i = 0; i < innerDim; ++i)
       res =  pmadd(pset1<Packet>(lhs.coeff(row, i)), rhs.template packet<LoadMode,Packet>(i, col), res);
   }
@@ -661,7 +678,7 @@ struct etor_product_packet_impl<ColMajor, Dynamic, Lhs, Rhs, Packet, LoadMode>
 {
   static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Index innerDim, Packet& res)
   {
-    res = pset1<Packet>(0);
+    res = pset1<Packet>(typename unpacket_traits<Packet>::type(0));
     for(Index i = 0; i < innerDim; ++i)
       res =  pmadd(lhs.template packet<LoadMode,Packet>(row, i), pset1<Packet>(rhs.coeff(i, col)), res);
   }

Eigen/src/Core/Random.h

@@ -16,8 +16,7 @@ namespace internal {
 
 template<typename Scalar> struct scalar_random_op {
   EIGEN_EMPTY_STRUCT_CTOR(scalar_random_op)
-  template<typename Index>
-  inline const Scalar operator() (Index, Index = 0) const { return random<Scalar>(); }
+  inline const Scalar operator() () const { return random<Scalar>(); }
 };
 
 template<typename Scalar>

Eigen/src/Core/Ref.h

@@ -35,7 +35,13 @@ struct traits<Ref<_PlainObjectType, _Options, _StrideType> >
                       || (int(StrideType::InnerStrideAtCompileTime)==0 && int(Derived::InnerStrideAtCompileTime)==1),
       OuterStrideMatch = Derived::IsVectorAtCompileTime
                       || int(StrideType::OuterStrideAtCompileTime)==int(Dynamic) || int(StrideType::OuterStrideAtCompileTime)==int(Derived::OuterStrideAtCompileTime),
-      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
+      // NOTE, this indirection of evaluator<Derived>::Alignment is needed
+      // to workaround a very strange bug in MSVC related to the instantiation
+      // of has_*ary_operator in evaluator<CwiseNullaryOp>.
+      // This line is surprisingly very sensitive. For instance, simply adding parenthesis
+      // as "DerivedAlignment = (int(evaluator<Derived>::Alignment))," will make MSVC fail...
+      DerivedAlignment = int(evaluator<Derived>::Alignment),
+      AlignmentMatch = (int(traits<PlainObjectType>::Alignment)==int(Unaligned)) || (DerivedAlignment >= 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
     };

Eigen/src/Core/SelfAdjointView.h

@@ -45,7 +45,7 @@ struct traits<SelfAdjointView<MatrixType, UpLo> > : traits<MatrixType>
 };
 }
 
-// FIXME could also be called SelfAdjointWrapper to be consistent with DiagonalWrapper ??
+
 template<typename _MatrixType, unsigned int UpLo> class SelfAdjointView
   : public TriangularBase<SelfAdjointView<_MatrixType, UpLo> >
 {
@@ -60,10 +60,12 @@ template<typename _MatrixType, unsigned int UpLo> class SelfAdjointView
     /** \brief The type of coefficients in this matrix */
     typedef typename internal::traits<SelfAdjointView>::Scalar Scalar; 
     typedef typename MatrixType::StorageIndex StorageIndex;
+    typedef typename internal::remove_all<typename MatrixType::ConjugateReturnType>::type MatrixConjugateReturnType;
 
     enum {
       Mode = internal::traits<SelfAdjointView>::Mode,
-      Flags = internal::traits<SelfAdjointView>::Flags
+      Flags = internal::traits<SelfAdjointView>::Flags,
+      TransposeMode = ((Mode & Upper) ? Lower : 0) | ((Mode & Lower) ? Upper : 0)
     };
     typedef typename MatrixType::PlainObject PlainObject;
 
@@ -187,6 +189,36 @@ template<typename _MatrixType, unsigned int UpLo> class SelfAdjointView
                                    TriangularView<typename MatrixType::AdjointReturnType,TriMode> >::type(tmp2);
     }
 
+    typedef SelfAdjointView<const MatrixConjugateReturnType,Mode> ConjugateReturnType;
+    /** \sa MatrixBase::conjugate() const */
+    EIGEN_DEVICE_FUNC
+    inline const ConjugateReturnType conjugate() const
+    { return ConjugateReturnType(m_matrix.conjugate()); }
+
+    typedef SelfAdjointView<const typename MatrixType::AdjointReturnType,TransposeMode> AdjointReturnType;
+    /** \sa MatrixBase::adjoint() const */
+    EIGEN_DEVICE_FUNC
+    inline const AdjointReturnType adjoint() const
+    { return AdjointReturnType(m_matrix.adjoint()); }
+
+    typedef SelfAdjointView<typename MatrixType::TransposeReturnType,TransposeMode> TransposeReturnType;
+     /** \sa MatrixBase::transpose() */
+    EIGEN_DEVICE_FUNC
+    inline TransposeReturnType transpose()
+    {
+      EIGEN_STATIC_ASSERT_LVALUE(MatrixType)
+      typename MatrixType::TransposeReturnType tmp(m_matrix);
+      return TransposeReturnType(tmp);
+    }
+
+    typedef SelfAdjointView<const typename MatrixType::ConstTransposeReturnType,TransposeMode> ConstTransposeReturnType;
+    /** \sa MatrixBase::transpose() const */
+    EIGEN_DEVICE_FUNC
+    inline const ConstTransposeReturnType transpose() const
+    {
+      return ConstTransposeReturnType(m_matrix.transpose());
+    }
+
     /** \returns a const expression of the main diagonal of the matrix \c *this
       *
       * This method simply returns the diagonal of the nested expression, thus by-passing the SelfAdjointView decorator.
@@ -287,6 +319,7 @@ public:
 * Implementation of MatrixBase methods
 ***************************************************************************/
 
+/** This is the const version of MatrixBase::selfadjointView() */
 template<typename Derived>
 template<unsigned int UpLo>
 typename MatrixBase<Derived>::template ConstSelfAdjointViewReturnType<UpLo>::Type
@@ -295,6 +328,15 @@ MatrixBase<Derived>::selfadjointView() const
   return typename ConstSelfAdjointViewReturnType<UpLo>::Type(derived());
 }
 
+/** \returns an expression of a symmetric/self-adjoint view extracted from the upper or lower triangular part of the current matrix
+  *
+  * The parameter \a UpLo can be either \c #Upper or \c #Lower
+  *
+  * Example: \include MatrixBase_selfadjointView.cpp
+  * Output: \verbinclude MatrixBase_selfadjointView.out
+  *
+  * \sa class SelfAdjointView
+  */
 template<typename Derived>
 template<unsigned int UpLo>
 typename MatrixBase<Derived>::template SelfAdjointViewReturnType<UpLo>::Type

Eigen/src/Core/Solve.h

@@ -139,7 +139,11 @@ struct Assignment<DstXprType, Solve<DecType,RhsType>, internal::assign_op<Scalar
   typedef Solve<DecType,RhsType> SrcXprType;
   static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<Scalar,Scalar> &)
   {
-    // FIXME shall we resize dst here?
+    Index dstRows = src.rows();
+    Index dstCols = src.cols();
+    if((dst.rows()!=dstRows) || (dst.cols()!=dstCols))
+      dst.resize(dstRows, dstCols);
+
     src.dec()._solve_impl(src.rhs(), dst);
   }
 };
@@ -151,6 +155,11 @@ struct Assignment<DstXprType, Solve<Transpose<const DecType>,RhsType>, internal:
   typedef Solve<Transpose<const DecType>,RhsType> SrcXprType;
   static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<Scalar,Scalar> &)
   {
+    Index dstRows = src.rows();
+    Index dstCols = src.cols();
+    if((dst.rows()!=dstRows) || (dst.cols()!=dstCols))
+      dst.resize(dstRows, dstCols);
+
     src.dec().nestedExpression().template _solve_impl_transposed<false>(src.rhs(), dst);
   }
 };
@@ -163,6 +172,11 @@ struct Assignment<DstXprType, Solve<CwiseUnaryOp<internal::scalar_conjugate_op<t
   typedef Solve<CwiseUnaryOp<internal::scalar_conjugate_op<typename DecType::Scalar>, const Transpose<const DecType> >,RhsType> SrcXprType;
   static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<Scalar,Scalar> &)
   {
+    Index dstRows = src.rows();
+    Index dstCols = src.cols();
+    if((dst.rows()!=dstRows) || (dst.cols()!=dstCols))
+      dst.resize(dstRows, dstCols);
+    
     src.dec().nestedExpression().nestedExpression().template _solve_impl_transposed<true>(src.rhs(), dst);
   }
 };

Eigen/src/Core/SolveTriangular.h

@@ -161,6 +161,7 @@ struct triangular_solver_selector<Lhs,Rhs,OnTheRight,Mode,CompleteUnrolling,1> {
 * TriangularView methods
 ***************************************************************************/
 
+#ifndef EIGEN_PARSED_BY_DOXYGEN
 template<typename MatrixType, unsigned int Mode>
 template<int Side, typename OtherDerived>
 void TriangularViewImpl<MatrixType,Mode,Dense>::solveInPlace(const MatrixBase<OtherDerived>& _other) const
@@ -188,6 +189,7 @@ TriangularViewImpl<Derived,Mode,Dense>::solve(const MatrixBase<Other>& other) co
 {
   return internal::triangular_solve_retval<Side,TriangularViewType,Other>(derived(), other.derived());
 }
+#endif
 
 namespace internal {
 

Eigen/src/Core/Transpose.h

@@ -78,6 +78,11 @@ template<typename MatrixType> class Transpose
     typename internal::remove_reference<MatrixTypeNested>::type&
     nestedExpression() { return m_matrix; }
 
+    /** \internal */
+    void resize(Index nrows, Index ncols) {
+      m_matrix.resize(ncols,nrows);
+    }
+
   protected:
     typename internal::ref_selector<MatrixType>::non_const_type m_matrix;
 };

Eigen/src/Core/TriangularMatrix.h

@@ -470,6 +470,8 @@ template<typename _MatrixType, unsigned int _Mode> class TriangularViewImpl<_Mat
       * \a Side==OnTheLeft (the default), or the right-inverse-multiply  \a other * inverse(\c *this) if
       * \a Side==OnTheRight.
       *
+      * Note that the template parameter \c Side can be ommitted, in which case \c Side==OnTheLeft
+      *
       * The matrix \c *this must be triangular and invertible (i.e., all the coefficients of the
       * diagonal must be non zero). It works as a forward (resp. backward) substitution if \c *this
       * is an upper (resp. lower) triangular matrix.
@@ -495,6 +497,8 @@ template<typename _MatrixType, unsigned int _Mode> class TriangularViewImpl<_Mat
       * \warning The parameter is only marked 'const' to make the C++ compiler accept a temporary expression here.
       * This function will const_cast it, so constness isn't honored here.
       *
+      * Note that the template parameter \c Side can be ommitted, in which case \c Side==OnTheLeft
+      *
       * See TriangularView:solve() for the details.
       */
     template<int Side, typename OtherDerived>
@@ -539,13 +543,14 @@ template<typename _MatrixType, unsigned int _Mode> class TriangularViewImpl<_Mat
 
     template<typename ProductType>
     EIGEN_DEVICE_FUNC
-    EIGEN_STRONG_INLINE TriangularViewType& _assignProduct(const ProductType& prod, const Scalar& alpha);
+    EIGEN_STRONG_INLINE TriangularViewType& _assignProduct(const ProductType& prod, const Scalar& alpha, bool beta);
 };
 
 /***************************************************************************
 * Implementation of triangular evaluation/assignment
 ***************************************************************************/
 
+#ifndef EIGEN_PARSED_BY_DOXYGEN
 // FIXME should we keep that possibility
 template<typename MatrixType, unsigned int Mode>
 template<typename OtherDerived>
@@ -583,6 +588,7 @@ void TriangularViewImpl<MatrixType, Mode, Dense>::lazyAssign(const TriangularBas
   eigen_assert(Mode == int(OtherDerived::Mode));
   internal::call_assignment_no_alias(derived(), other.derived());
 }
+#endif
 
 /***************************************************************************
 * Implementation of TriangularBase methods
@@ -641,21 +647,20 @@ MatrixBase<Derived>::triangularView() const
 template<typename Derived>
 bool MatrixBase<Derived>::isUpperTriangular(const RealScalar& prec) const
 {
-  using std::abs;
   RealScalar maxAbsOnUpperPart = static_cast<RealScalar>(-1);
   for(Index j = 0; j < cols(); ++j)
   {
-    Index maxi = (std::min)(j, rows()-1);
+    Index maxi = numext::mini(j, rows()-1);
     for(Index i = 0; i <= maxi; ++i)
     {
-      RealScalar absValue = abs(coeff(i,j));
+      RealScalar absValue = numext::abs(coeff(i,j));
       if(absValue > maxAbsOnUpperPart) maxAbsOnUpperPart = absValue;
     }
   }
   RealScalar threshold = maxAbsOnUpperPart * prec;
   for(Index j = 0; j < cols(); ++j)
     for(Index i = j+1; i < rows(); ++i)
-      if(abs(coeff(i, j)) > threshold) return false;
+      if(numext::abs(coeff(i, j)) > threshold) return false;
   return true;
 }
 
@@ -667,20 +672,19 @@ bool MatrixBase<Derived>::isUpperTriangular(const RealScalar& prec) const
 template<typename Derived>
 bool MatrixBase<Derived>::isLowerTriangular(const RealScalar& prec) const
 {
-  using std::abs;
   RealScalar maxAbsOnLowerPart = static_cast<RealScalar>(-1);
   for(Index j = 0; j < cols(); ++j)
     for(Index i = j; i < rows(); ++i)
     {
-      RealScalar absValue = abs(coeff(i,j));
+      RealScalar absValue = numext::abs(coeff(i,j));
       if(absValue > maxAbsOnLowerPart) maxAbsOnLowerPart = absValue;
     }
   RealScalar threshold = maxAbsOnLowerPart * prec;
   for(Index j = 1; j < cols(); ++j)
   {
-    Index maxi = (std::min)(j, rows()-1);
+    Index maxi = numext::mini(j, rows()-1);
     for(Index i = 0; i < maxi; ++i)
-      if(abs(coeff(i, j)) > threshold) return false;
+      if(numext::abs(coeff(i, j)) > threshold) return false;
   }
   return true;
 }
@@ -777,15 +781,18 @@ public:
 
 template<int Mode, bool SetOpposite, typename DstXprType, typename SrcXprType, typename Functor>
 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
-void call_triangular_assignment_loop(const DstXprType& dst, const SrcXprType& src, const Functor &func)
+void call_triangular_assignment_loop(DstXprType& dst, const SrcXprType& src, const Functor &func)
 {
-  eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols());
-  
   typedef evaluator<DstXprType> DstEvaluatorType;
   typedef evaluator<SrcXprType> SrcEvaluatorType;
 
-  DstEvaluatorType dstEvaluator(dst);
   SrcEvaluatorType srcEvaluator(src);
+
+  Index dstRows = src.rows();
+  Index dstCols = src.cols();
+  if((dst.rows()!=dstRows) || (dst.cols()!=dstCols))
+    dst.resize(dstRows, dstCols);
+  DstEvaluatorType dstEvaluator(dst);
     
   typedef triangular_dense_assignment_kernel< Mode&(Lower|Upper),Mode&(UnitDiag|ZeroDiag|SelfAdjoint),SetOpposite,
                                               DstEvaluatorType,SrcEvaluatorType,Functor> Kernel;
@@ -802,7 +809,7 @@ void call_triangular_assignment_loop(const DstXprType& dst, const SrcXprType& sr
 
 template<int Mode, bool SetOpposite, typename DstXprType, typename SrcXprType>
 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
-void call_triangular_assignment_loop(const DstXprType& dst, const SrcXprType& src)
+void call_triangular_assignment_loop(DstXprType& dst, const SrcXprType& src)
 {
   call_triangular_assignment_loop<Mode,SetOpposite>(dst, src, internal::assign_op<typename DstXprType::Scalar,typename SrcXprType::Scalar>());
 }
@@ -893,7 +900,7 @@ struct triangular_assignment_loop<Kernel, Mode, Dynamic, SetOpposite>
   {
     for(Index j = 0; j < kernel.cols(); ++j)
     {
-      Index maxi = (std::min)(j, kernel.rows());
+      Index maxi = numext::mini(j, kernel.rows());
       Index i = 0;
       if (((Mode&Lower) && SetOpposite) || (Mode&Upper))
       {
@@ -938,8 +945,12 @@ struct Assignment<DstXprType, Product<Lhs,Rhs,DefaultProduct>, internal::assign_
   typedef Product<Lhs,Rhs,DefaultProduct> SrcXprType;
   static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<Scalar,typename SrcXprType::Scalar> &)
   {
-    dst.setZero();
-    dst._assignProduct(src, 1);
+    Index dstRows = src.rows();
+    Index dstCols = src.cols();
+    if((dst.rows()!=dstRows) || (dst.cols()!=dstCols))
+      dst.resize(dstRows, dstCols);
+
+    dst._assignProduct(src, 1, 0);
   }
 };
 
@@ -950,7 +961,7 @@ struct Assignment<DstXprType, Product<Lhs,Rhs,DefaultProduct>, internal::add_ass
   typedef Product<Lhs,Rhs,DefaultProduct> SrcXprType;
   static void run(DstXprType &dst, const SrcXprType &src, const internal::add_assign_op<Scalar,typename SrcXprType::Scalar> &)
   {
-    dst._assignProduct(src, 1);
+    dst._assignProduct(src, 1, 1);
   }
 };
 
@@ -961,7 +972,7 @@ struct Assignment<DstXprType, Product<Lhs,Rhs,DefaultProduct>, internal::sub_ass
   typedef Product<Lhs,Rhs,DefaultProduct> SrcXprType;
   static void run(DstXprType &dst, const SrcXprType &src, const internal::sub_assign_op<Scalar,typename SrcXprType::Scalar> &)
   {
-    dst._assignProduct(src, -1);
+    dst._assignProduct(src, -1, 1);
   }
 };
 

Eigen/src/Core/VectorwiseOp.h

@@ -602,7 +602,7 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
       return m_matrix / extendedTo(other.derived());
     }
 
-    /** \returns an expression where each column of row of the referenced matrix are normalized.
+    /** \returns an expression where each column (or row) of the referenced matrix are normalized.
       * The referenced matrix is \b not modified.
       * \sa MatrixBase::normalized(), normalize()
       */
@@ -625,6 +625,7 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
 /////////// Geometry module ///////////
 
     typedef Homogeneous<ExpressionType,Direction> HomogeneousReturnType;
+    EIGEN_DEVICE_FUNC
     HomogeneousReturnType homogeneous() const;
 
     typedef typename ExpressionType::PlainObject CrossReturnType;
@@ -654,6 +655,7 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
                   Direction==Horizontal ? HNormalized_SizeMinusOne : 1> >
             HNormalizedReturnType;
 
+    EIGEN_DEVICE_FUNC
     const HNormalizedReturnType hnormalized() const;
 
   protected:

Eigen/src/Core/Visitor.h

@@ -194,7 +194,8 @@ struct functor_traits<max_coeff_visitor<Scalar> > {
 
 } // end namespace internal
 
-/** \returns the minimum of all coefficients of *this and puts in *row and *col its location.
+/** \fn DenseBase<Derived>::minCoeff(IndexType* rowId, IndexType* colId) const
+  * \returns the minimum of all coefficients of *this and puts in *row and *col its location.
   * \warning the result is undefined if \c *this contains NaN.
   *
   * \sa DenseBase::minCoeff(Index*), DenseBase::maxCoeff(Index*,Index*), DenseBase::visit(), DenseBase::minCoeff()
@@ -230,7 +231,8 @@ DenseBase<Derived>::minCoeff(IndexType* index) const
   return minVisitor.res;
 }
 
-/** \returns the maximum of all coefficients of *this and puts in *row and *col its location.
+/** \fn DenseBase<Derived>::maxCoeff(IndexType* rowId, IndexType* colId) const
+  * \returns the maximum of all coefficients of *this and puts in *row and *col its location.
   * \warning the result is undefined if \c *this contains NaN. 
   *
   * \sa DenseBase::minCoeff(IndexType*,IndexType*), DenseBase::visit(), DenseBase::maxCoeff()

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

@@ -1,6 +0,0 @@
-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

@@ -456,6 +456,26 @@ ptranspose(PacketBlock<Packet2cd,2>& kernel) {
  kernel.packet[0].v = tmp;
 }
 
+template<> EIGEN_STRONG_INLINE Packet4cf pinsertfirst(const Packet4cf& a, std::complex<float> b)
+{
+  return Packet4cf(_mm256_blend_ps(a.v,pset1<Packet4cf>(b).v,1|2));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cd pinsertfirst(const Packet2cd& a, std::complex<double> b)
+{
+  return Packet2cd(_mm256_blend_pd(a.v,pset1<Packet2cd>(b).v,1|2));
+}
+
+template<> EIGEN_STRONG_INLINE Packet4cf pinsertlast(const Packet4cf& a, std::complex<float> b)
+{
+  return Packet4cf(_mm256_blend_ps(a.v,pset1<Packet4cf>(b).v,(1<<7)|(1<<6)));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cd pinsertlast(const Packet2cd& a, std::complex<double> b)
+{
+  return Packet2cd(_mm256_blend_pd(a.v,pset1<Packet2cd>(b).v,(1<<3)|(1<<2)));
+}
+
 } // end namespace internal
 
 } // end namespace Eigen

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

@@ -266,52 +266,10 @@ pexp<Packet8f>(const Packet8f& _x) {
 }
 
 // Hyperbolic Tangent function.
-// Doesn't do anything fancy, just a 13/6-degree rational interpolant which
-// is accurate up to a couple of ulp in the range [-9, 9], outside of which the
-// fl(tanh(x)) = +/-1.
 template <>
 EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet8f
-ptanh<Packet8f>(const Packet8f& _x) {
-  // Clamp the inputs to the range [-9, 9] since anything outside
-  // this range is +/-1.0f in single-precision.
-  _EIGEN_DECLARE_CONST_Packet8f(plus_9, 9.0f);
-  _EIGEN_DECLARE_CONST_Packet8f(minus_9, -9.0f);
-  const Packet8f x = pmax(p8f_minus_9, pmin(p8f_plus_9, _x));
-
-  // The monomial coefficients of the numerator polynomial (odd).
-  _EIGEN_DECLARE_CONST_Packet8f(alpha_1, 4.89352455891786e-03f);
-  _EIGEN_DECLARE_CONST_Packet8f(alpha_3, 6.37261928875436e-04f);
-  _EIGEN_DECLARE_CONST_Packet8f(alpha_5, 1.48572235717979e-05f);
-  _EIGEN_DECLARE_CONST_Packet8f(alpha_7, 5.12229709037114e-08f);
-  _EIGEN_DECLARE_CONST_Packet8f(alpha_9, -8.60467152213735e-11f);
-  _EIGEN_DECLARE_CONST_Packet8f(alpha_11, 2.00018790482477e-13f);
-  _EIGEN_DECLARE_CONST_Packet8f(alpha_13, -2.76076847742355e-16f);
-
-  // The monomial coefficients of the denominator polynomial (even).
-  _EIGEN_DECLARE_CONST_Packet8f(beta_0, 4.89352518554385e-03f);
-  _EIGEN_DECLARE_CONST_Packet8f(beta_2, 2.26843463243900e-03f);
-  _EIGEN_DECLARE_CONST_Packet8f(beta_4, 1.18534705686654e-04f);
-  _EIGEN_DECLARE_CONST_Packet8f(beta_6, 1.19825839466702e-06f);
-
-  // Since the polynomials are odd/even, we need x^2.
-  const Packet8f x2 = pmul(x, x);
-
-  // Evaluate the numerator polynomial p.
-  Packet8f p = pmadd(x2, p8f_alpha_13, p8f_alpha_11);
-  p = pmadd(x2, p, p8f_alpha_9);
-  p = pmadd(x2, p, p8f_alpha_7);
-  p = pmadd(x2, p, p8f_alpha_5);
-  p = pmadd(x2, p, p8f_alpha_3);
-  p = pmadd(x2, p, p8f_alpha_1);
-  p = pmul(x, p);
-
-  // Evaluate the denominator polynomial p.
-  Packet8f q = pmadd(x2, p8f_beta_6, p8f_beta_4);
-  q = pmadd(x2, q, p8f_beta_2);
-  q = pmadd(x2, q, p8f_beta_0);
-
-  // Divide the numerator by the denominator.
-  return pdiv(p, q);
+ptanh<Packet8f>(const Packet8f& x) {
+  return internal::generic_fast_tanh_float(x);
 }
 
 template <>
@@ -397,30 +355,27 @@ pexp<Packet4d>(const Packet4d& _x) {
 // Functions for sqrt.
 // The EIGEN_FAST_MATH version uses the _mm_rsqrt_ps approximation and one step
 // of Newton's method, at a cost of 1-2 bits of precision as opposed to the
-// exact solution. The main advantage of this approach is not just speed, but
-// also the fact that it can be inlined and pipelined with other computations,
-// further reducing its effective latency.
+// exact solution. It does not handle +inf, or denormalized numbers correctly.
+// The main advantage of this approach is not just speed, but also the fact that
+// it can be inlined and pipelined with other computations, further reducing its
+// effective latency. This is similar to Quake3's fast inverse square root.
+// For detail see here: http://www.beyond3d.com/content/articles/8/
 #if EIGEN_FAST_MATH
 template <>
 EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet8f
 psqrt<Packet8f>(const Packet8f& _x) {
-  _EIGEN_DECLARE_CONST_Packet8f(one_point_five, 1.5f);
-  _EIGEN_DECLARE_CONST_Packet8f(minus_half, -0.5f);
-  _EIGEN_DECLARE_CONST_Packet8f_FROM_INT(flt_min, 0x00800000);
-
-  Packet8f neg_half = pmul(_x, p8f_minus_half);
-
-  // select only the inverse sqrt of positive normal inputs (denormals are
-  // flushed to zero and cause infs as well).
-  Packet8f non_zero_mask = _mm256_cmp_ps(_x, p8f_flt_min, _CMP_GE_OQ);
-  Packet8f x = _mm256_and_ps(non_zero_mask, _mm256_rsqrt_ps(_x));
+  Packet8f half = pmul(_x, pset1<Packet8f>(.5f));
+  Packet8f denormal_mask = _mm256_and_ps(
+      _mm256_cmp_ps(_x, pset1<Packet8f>((std::numeric_limits<float>::min)()),
+                    _CMP_LT_OQ),
+      _mm256_cmp_ps(_x, _mm256_setzero_ps(), _CMP_GE_OQ));
 
+  // Compute approximate reciprocal sqrt.
+  Packet8f x = _mm256_rsqrt_ps(_x);
   // Do a single step of Newton's iteration.
-  x = pmul(x, pmadd(neg_half, pmul(x, x), p8f_one_point_five));
-
-  // Multiply the original _x by it's reciprocal square root to extract the
-  // square root.
-  return pmul(_x, x);
+  x = pmul(x, psub(pset1<Packet8f>(1.5f), pmul(half, pmul(x,x))));
+  // Flush results for denormals to zero.
+  return _mm256_andnot_ps(denormal_mask, pmul(_x,x));
 }
 #else
 template <> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED

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

@@ -48,7 +48,9 @@ template<> struct is_arithmetic<__m256d> { enum { value = true }; };
 #define _EIGEN_DECLARE_CONST_Packet8i(NAME,X) \
   const Packet8i p8i_##NAME = pset1<Packet8i>(X)
 
-
+// Use the packet_traits defined in AVX512/PacketMath.h instead if we're going
+// to leverage AVX512 instructions.
+#ifndef EIGEN_VECTORIZE_AVX512
 template<> struct packet_traits<float>  : default_packet_traits
 {
   typedef Packet8f type;
@@ -93,6 +95,10 @@ template<> struct packet_traits<double> : default_packet_traits
     HasCeil = 1
   };
 };
+#endif
+
+template<> struct scalar_div_cost<float,true> { enum { value = 14 }; };
+template<> struct scalar_div_cost<double,true> { enum { value = 16 }; };
 
 /* Proper support for integers is only provided by AVX2. In the meantime, we'll
    use SSE instructions and packets to deal with integers.
@@ -153,7 +159,7 @@ template<> EIGEN_STRONG_INLINE Packet8i pdiv<Packet8i>(const Packet8i& /*a*/, co
 
 #ifdef __FMA__
 template<> EIGEN_STRONG_INLINE Packet8f pmadd(const Packet8f& a, const Packet8f& b, const Packet8f& c) {
-#if EIGEN_COMP_GNUC || EIGEN_COMP_CLANG
+#if ( EIGEN_COMP_GNUC_STRICT || (EIGEN_COMP_CLANG && (EIGEN_COMP_CLANG<308)) )
   // clang stupidly generates a vfmadd213ps instruction plus some vmovaps on registers,
   // and gcc stupidly generates a vfmadd132ps instruction,
   // so let's enforce it to generate a vfmadd231ps instruction since the most common use case is to accumulate
@@ -166,7 +172,7 @@ template<> EIGEN_STRONG_INLINE Packet8f pmadd(const Packet8f& a, const Packet8f&
 #endif
 }
 template<> EIGEN_STRONG_INLINE Packet4d pmadd(const Packet4d& a, const Packet4d& b, const Packet4d& c) {
-#if EIGEN_COMP_GNUC || EIGEN_COMP_CLANG
+#if ( EIGEN_COMP_GNUC_STRICT || (EIGEN_COMP_CLANG && (EIGEN_COMP_CLANG<308)) )
   // see above
   Packet4d res = c;
   __asm__("vfmadd231pd %[a], %[b], %[c]" : [c] "+x" (res) : [a] "x" (a), [b] "x" (b));
@@ -301,9 +307,11 @@ template<> EIGEN_STRONG_INLINE void pstore1<Packet8i>(int* to, const int& a)
   pstore(to, pa);
 }
 
+#ifndef EIGEN_VECTORIZE_AVX512
 template<> EIGEN_STRONG_INLINE void prefetch<float>(const float*   addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); }
 template<> EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); }
 template<> EIGEN_STRONG_INLINE void prefetch<int>(const int*       addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); }
+#endif
 
 template<> EIGEN_STRONG_INLINE float  pfirst<Packet8f>(const Packet8f& a) {
   return _mm_cvtss_f32(_mm256_castps256_ps128(a));
@@ -387,17 +395,14 @@ template<> EIGEN_STRONG_INLINE Packet4d preduxp<Packet4d>(const Packet4d* vecs)
 
 template<> EIGEN_STRONG_INLINE float predux<Packet8f>(const Packet8f& a)
 {
-  Packet8f tmp0 = _mm256_hadd_ps(a,_mm256_permute2f128_ps(a,a,1));
-  tmp0 = _mm256_hadd_ps(tmp0,tmp0);
-  return pfirst(_mm256_hadd_ps(tmp0, tmp0));
+  return predux(Packet4f(_mm_add_ps(_mm256_castps256_ps128(a),_mm256_extractf128_ps(a,1))));
 }
 template<> EIGEN_STRONG_INLINE double predux<Packet4d>(const Packet4d& a)
 {
-  Packet4d tmp0 = _mm256_hadd_pd(a,_mm256_permute2f128_pd(a,a,1));
-  return pfirst(_mm256_hadd_pd(tmp0,tmp0));
+  return predux(Packet2d(_mm_add_pd(_mm256_castpd256_pd128(a),_mm256_extractf128_pd(a,1))));
 }
 
-template<> EIGEN_STRONG_INLINE Packet4f predux4<Packet8f>(const Packet8f& a)
+template<> EIGEN_STRONG_INLINE Packet4f predux_downto4<Packet8f>(const Packet8f& a)
 {
   return _mm_add_ps(_mm256_castps256_ps128(a),_mm256_extractf128_ps(a,1));
 }
@@ -601,6 +606,26 @@ template<> EIGEN_STRONG_INLINE Packet4d pblend(const Selector<4>& ifPacket, cons
   return _mm256_blendv_pd(thenPacket, elsePacket, false_mask);
 }
 
+template<> EIGEN_STRONG_INLINE Packet8f pinsertfirst(const Packet8f& a, float b)
+{
+  return _mm256_blend_ps(a,pset1<Packet8f>(b),1);
+}
+
+template<> EIGEN_STRONG_INLINE Packet4d pinsertfirst(const Packet4d& a, double b)
+{
+  return _mm256_blend_pd(a,pset1<Packet4d>(b),1);
+}
+
+template<> EIGEN_STRONG_INLINE Packet8f pinsertlast(const Packet8f& a, float b)
+{
+  return _mm256_blend_ps(a,pset1<Packet8f>(b),(1<<7));
+}
+
+template<> EIGEN_STRONG_INLINE Packet4d pinsertlast(const Packet4d& a, double b)
+{
+  return _mm256_blend_pd(a,pset1<Packet4d>(b),(1<<3));
+}
+
 } // end namespace internal
 
 } // end namespace Eigen

Eigen/src/Core/arch/AVX512/MathFunctions.h

@@ -0,0 +1,396 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016 Pedro Gonnet (pedro.gonnet@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 THIRD_PARTY_EIGEN3_EIGEN_SRC_CORE_ARCH_AVX512_MATHFUNCTIONS_H_
+#define THIRD_PARTY_EIGEN3_EIGEN_SRC_CORE_ARCH_AVX512_MATHFUNCTIONS_H_
+
+namespace Eigen {
+
+namespace internal {
+
+// Disable the code for older versions of gcc that don't support many of the required avx512 instrinsics.
+#if EIGEN_GNUC_AT_LEAST(5, 3)
+
+#define _EIGEN_DECLARE_CONST_Packet16f(NAME, X) \
+  const Packet16f p16f_##NAME = pset1<Packet16f>(X)
+
+#define _EIGEN_DECLARE_CONST_Packet16f_FROM_INT(NAME, X) \
+  const Packet16f p16f_##NAME = (__m512)pset1<Packet16i>(X)
+
+#define _EIGEN_DECLARE_CONST_Packet8d(NAME, X) \
+  const Packet8d p8d_##NAME = pset1<Packet8d>(X)
+
+#define _EIGEN_DECLARE_CONST_Packet8d_FROM_INT64(NAME, X) \
+  const Packet8d p8d_##NAME = _mm512_castsi512_pd(_mm512_set1_epi64(X))
+
+// Natural logarithm
+// Computes log(x) as log(2^e * m) = C*e + log(m), where the constant C =log(2)
+// and m is in the range [sqrt(1/2),sqrt(2)). In this range, the logarithm can
+// be easily approximated by a polynomial centered on m=1 for stability.
+#if defined(EIGEN_VECTORIZE_AVX512DQ)
+template <>
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet16f
+plog<Packet16f>(const Packet16f& _x) {
+  Packet16f x = _x;
+  _EIGEN_DECLARE_CONST_Packet16f(1, 1.0f);
+  _EIGEN_DECLARE_CONST_Packet16f(half, 0.5f);
+  _EIGEN_DECLARE_CONST_Packet16f(126f, 126.0f);
+
+  _EIGEN_DECLARE_CONST_Packet16f_FROM_INT(inv_mant_mask, ~0x7f800000);
+
+  // The smallest non denormalized float number.
+  _EIGEN_DECLARE_CONST_Packet16f_FROM_INT(min_norm_pos, 0x00800000);
+  _EIGEN_DECLARE_CONST_Packet16f_FROM_INT(minus_inf, 0xff800000);
+  _EIGEN_DECLARE_CONST_Packet16f_FROM_INT(nan, 0x7fc00000);
+
+  // Polynomial coefficients.
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_SQRTHF, 0.707106781186547524f);
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_log_p0, 7.0376836292E-2f);
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_log_p1, -1.1514610310E-1f);
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_log_p2, 1.1676998740E-1f);
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_log_p3, -1.2420140846E-1f);
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_log_p4, +1.4249322787E-1f);
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_log_p5, -1.6668057665E-1f);
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_log_p6, +2.0000714765E-1f);
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_log_p7, -2.4999993993E-1f);
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_log_p8, +3.3333331174E-1f);
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_log_q1, -2.12194440e-4f);
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_log_q2, 0.693359375f);
+
+  // invalid_mask is set to true when x is NaN
+  __mmask16 invalid_mask =
+      _mm512_cmp_ps_mask(x, _mm512_setzero_ps(), _CMP_NGE_UQ);
+  __mmask16 iszero_mask =
+      _mm512_cmp_ps_mask(x, _mm512_setzero_ps(), _CMP_EQ_UQ);
+
+  // Truncate input values to the minimum positive normal.
+  x = pmax(x, p16f_min_norm_pos);
+
+  // Extract the shifted exponents.
+  Packet16f emm0 = _mm512_cvtepi32_ps(_mm512_srli_epi32((__m512i)x, 23));
+  Packet16f e = _mm512_sub_ps(emm0, p16f_126f);
+
+  // Set the exponents to -1, i.e. x are in the range [0.5,1).
+  x = _mm512_and_ps(x, p16f_inv_mant_mask);
+  x = _mm512_or_ps(x, p16f_half);
+
+  // part2: Shift the inputs from the range [0.5,1) to [sqrt(1/2),sqrt(2))
+  // and shift by -1. The values are then centered around 0, which improves
+  // the stability of the polynomial evaluation.
+  //   if( x < SQRTHF ) {
+  //     e -= 1;
+  //     x = x + x - 1.0;
+  //   } else { x = x - 1.0; }
+  __mmask16 mask = _mm512_cmp_ps_mask(x, p16f_cephes_SQRTHF, _CMP_LT_OQ);
+  Packet16f tmp = _mm512_mask_blend_ps(mask, x, _mm512_setzero_ps());
+  x = psub(x, p16f_1);
+  e = psub(e, _mm512_mask_blend_ps(mask, p16f_1, _mm512_setzero_ps()));
+  x = padd(x, tmp);
+
+  Packet16f x2 = pmul(x, x);
+  Packet16f x3 = pmul(x2, x);
+
+  // Evaluate the polynomial approximant of degree 8 in three parts, probably
+  // to improve instruction-level parallelism.
+  Packet16f y, y1, y2;
+  y = pmadd(p16f_cephes_log_p0, x, p16f_cephes_log_p1);
+  y1 = pmadd(p16f_cephes_log_p3, x, p16f_cephes_log_p4);
+  y2 = pmadd(p16f_cephes_log_p6, x, p16f_cephes_log_p7);
+  y = pmadd(y, x, p16f_cephes_log_p2);
+  y1 = pmadd(y1, x, p16f_cephes_log_p5);
+  y2 = pmadd(y2, x, p16f_cephes_log_p8);
+  y = pmadd(y, x3, y1);
+  y = pmadd(y, x3, y2);
+  y = pmul(y, x3);
+
+  // Add the logarithm of the exponent back to the result of the interpolation.
+  y1 = pmul(e, p16f_cephes_log_q1);
+  tmp = pmul(x2, p16f_half);
+  y = padd(y, y1);
+  x = psub(x, tmp);
+  y2 = pmul(e, p16f_cephes_log_q2);
+  x = padd(x, y);
+  x = padd(x, y2);
+
+  // Filter out invalid inputs, i.e. negative arg will be NAN, 0 will be -INF.
+  return _mm512_mask_blend_ps(iszero_mask, p16f_minus_inf,
+                              _mm512_mask_blend_ps(invalid_mask, p16f_nan, x));
+}
+#endif
+
+// Exponential function. Works by writing "x = m*log(2) + r" where
+// "m = floor(x/log(2)+1/2)" and "r" is the remainder. The result is then
+// "exp(x) = 2^m*exp(r)" where exp(r) is in the range [-1,1).
+template <>
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet16f
+pexp<Packet16f>(const Packet16f& _x) {
+  _EIGEN_DECLARE_CONST_Packet16f(1, 1.0f);
+  _EIGEN_DECLARE_CONST_Packet16f(half, 0.5f);
+  _EIGEN_DECLARE_CONST_Packet16f(127, 127.0f);
+
+  _EIGEN_DECLARE_CONST_Packet16f(exp_hi, 88.3762626647950f);
+  _EIGEN_DECLARE_CONST_Packet16f(exp_lo, -88.3762626647949f);
+
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_LOG2EF, 1.44269504088896341f);
+
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_exp_p0, 1.9875691500E-4f);
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_exp_p1, 1.3981999507E-3f);
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_exp_p2, 8.3334519073E-3f);
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_exp_p3, 4.1665795894E-2f);
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_exp_p4, 1.6666665459E-1f);
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_exp_p5, 5.0000001201E-1f);
+
+  // Clamp x.
+  Packet16f x = pmax(pmin(_x, p16f_exp_hi), p16f_exp_lo);
+
+  // Express exp(x) as exp(m*ln(2) + r), start by extracting
+  // m = floor(x/ln(2) + 0.5).
+  Packet16f m = _mm512_floor_ps(pmadd(x, p16f_cephes_LOG2EF, p16f_half));
+
+  // Get r = x - m*ln(2). Note that we can do this without losing more than one
+  // ulp precision due to the FMA instruction.
+  _EIGEN_DECLARE_CONST_Packet16f(nln2, -0.6931471805599453f);
+  Packet16f r = _mm512_fmadd_ps(m, p16f_nln2, x);
+  Packet16f r2 = pmul(r, r);
+
+  // TODO(gonnet): Split into odd/even polynomials and try to exploit
+  //               instruction-level parallelism.
+  Packet16f y = p16f_cephes_exp_p0;
+  y = pmadd(y, r, p16f_cephes_exp_p1);
+  y = pmadd(y, r, p16f_cephes_exp_p2);
+  y = pmadd(y, r, p16f_cephes_exp_p3);
+  y = pmadd(y, r, p16f_cephes_exp_p4);
+  y = pmadd(y, r, p16f_cephes_exp_p5);
+  y = pmadd(y, r2, r);
+  y = padd(y, p16f_1);
+
+  // Build emm0 = 2^m.
+  Packet16i emm0 = _mm512_cvttps_epi32(padd(m, p16f_127));
+  emm0 = _mm512_slli_epi32(emm0, 23);
+
+  // Return 2^m * exp(r).
+  return pmax(pmul(y, _mm512_castsi512_ps(emm0)), _x);
+}
+
+/*template <>
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet8d
+pexp<Packet8d>(const Packet8d& _x) {
+  Packet8d x = _x;
+
+  _EIGEN_DECLARE_CONST_Packet8d(1, 1.0);
+  _EIGEN_DECLARE_CONST_Packet8d(2, 2.0);
+
+  _EIGEN_DECLARE_CONST_Packet8d(exp_hi, 709.437);
+  _EIGEN_DECLARE_CONST_Packet8d(exp_lo, -709.436139303);
+
+  _EIGEN_DECLARE_CONST_Packet8d(cephes_LOG2EF, 1.4426950408889634073599);
+
+  _EIGEN_DECLARE_CONST_Packet8d(cephes_exp_p0, 1.26177193074810590878e-4);
+  _EIGEN_DECLARE_CONST_Packet8d(cephes_exp_p1, 3.02994407707441961300e-2);
+  _EIGEN_DECLARE_CONST_Packet8d(cephes_exp_p2, 9.99999999999999999910e-1);
+
+  _EIGEN_DECLARE_CONST_Packet8d(cephes_exp_q0, 3.00198505138664455042e-6);
+  _EIGEN_DECLARE_CONST_Packet8d(cephes_exp_q1, 2.52448340349684104192e-3);
+  _EIGEN_DECLARE_CONST_Packet8d(cephes_exp_q2, 2.27265548208155028766e-1);
+  _EIGEN_DECLARE_CONST_Packet8d(cephes_exp_q3, 2.00000000000000000009e0);
+
+  _EIGEN_DECLARE_CONST_Packet8d(cephes_exp_C1, 0.693145751953125);
+  _EIGEN_DECLARE_CONST_Packet8d(cephes_exp_C2, 1.42860682030941723212e-6);
+
+  // clamp x
+  x = pmax(pmin(x, p8d_exp_hi), p8d_exp_lo);
+
+  // Express exp(x) as exp(g + n*log(2)).
+  const Packet8d n =
+      _mm512_mul_round_pd(p8d_cephes_LOG2EF, x, _MM_FROUND_TO_NEAREST_INT);
+
+  // Get the remainder modulo log(2), i.e. the "g" described above. Subtract
+  // n*log(2) out in two steps, i.e. n*C1 + n*C2, C1+C2=log2 to get the last
+  // digits right.
+  const Packet8d nC1 = pmul(n, p8d_cephes_exp_C1);
+  const Packet8d nC2 = pmul(n, p8d_cephes_exp_C2);
+  x = psub(x, nC1);
+  x = psub(x, nC2);
+
+  const Packet8d x2 = pmul(x, x);
+
+  // Evaluate the numerator polynomial of the rational interpolant.
+  Packet8d px = p8d_cephes_exp_p0;
+  px = pmadd(px, x2, p8d_cephes_exp_p1);
+  px = pmadd(px, x2, p8d_cephes_exp_p2);
+  px = pmul(px, x);
+
+  // Evaluate the denominator polynomial of the rational interpolant.
+  Packet8d qx = p8d_cephes_exp_q0;
+  qx = pmadd(qx, x2, p8d_cephes_exp_q1);
+  qx = pmadd(qx, x2, p8d_cephes_exp_q2);
+  qx = pmadd(qx, x2, p8d_cephes_exp_q3);
+
+  // I don't really get this bit, copied from the SSE2 routines, so...
+  // TODO(gonnet): Figure out what is going on here, perhaps find a better
+  // rational interpolant?
+  x = _mm512_div_pd(px, psub(qx, px));
+  x = pmadd(p8d_2, x, p8d_1);
+
+  // Build e=2^n.
+  const Packet8d e = _mm512_castsi512_pd(_mm512_slli_epi64(
+      _mm512_add_epi64(_mm512_cvtpd_epi64(n), _mm512_set1_epi64(1023)), 52));
+
+  // Construct the result 2^n * exp(g) = e * x. The max is used to catch
+  // non-finite values in the input.
+  return pmax(pmul(x, e), _x);
+  }*/
+
+// Functions for sqrt.
+// The EIGEN_FAST_MATH version uses the _mm_rsqrt_ps approximation and one step
+// of Newton's method, at a cost of 1-2 bits of precision as opposed to the
+// exact solution. The main advantage of this approach is not just speed, but
+// also the fact that it can be inlined and pipelined with other computations,
+// further reducing its effective latency.
+#if EIGEN_FAST_MATH
+template <>
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet16f
+psqrt<Packet16f>(const Packet16f& _x) {
+  _EIGEN_DECLARE_CONST_Packet16f(one_point_five, 1.5f);
+  _EIGEN_DECLARE_CONST_Packet16f(minus_half, -0.5f);
+  _EIGEN_DECLARE_CONST_Packet16f_FROM_INT(flt_min, 0x00800000);
+
+  Packet16f neg_half = pmul(_x, p16f_minus_half);
+
+  // select only the inverse sqrt of positive normal inputs (denormals are
+  // flushed to zero and cause infs as well).
+  __mmask16 non_zero_mask = _mm512_cmp_ps_mask(_x, p16f_flt_min, _CMP_GE_OQ);
+  Packet16f x = _mm512_mask_blend_ps(non_zero_mask, _mm512_rsqrt14_ps(_x),
+                                     _mm512_setzero_ps());
+
+  // Do a single step of Newton's iteration.
+  x = pmul(x, pmadd(neg_half, pmul(x, x), p16f_one_point_five));
+
+  // Multiply the original _x by it's reciprocal square root to extract the
+  // square root.
+  return pmul(_x, x);
+}
+
+template <>
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet8d
+psqrt<Packet8d>(const Packet8d& _x) {
+  _EIGEN_DECLARE_CONST_Packet8d(one_point_five, 1.5);
+  _EIGEN_DECLARE_CONST_Packet8d(minus_half, -0.5);
+  _EIGEN_DECLARE_CONST_Packet8d_FROM_INT64(dbl_min, 0x0010000000000000LL);
+
+  Packet8d neg_half = pmul(_x, p8d_minus_half);
+
+  // select only the inverse sqrt of positive normal inputs (denormals are
+  // flushed to zero and cause infs as well).
+  __mmask8 non_zero_mask = _mm512_cmp_pd_mask(_x, p8d_dbl_min, _CMP_GE_OQ);
+  Packet8d x = _mm512_mask_blend_pd(non_zero_mask, _mm512_rsqrt14_pd(_x),
+                                    _mm512_setzero_pd());
+
+  // Do a first step of Newton's iteration.
+  x = pmul(x, pmadd(neg_half, pmul(x, x), p8d_one_point_five));
+
+  // Do a second step of Newton's iteration.
+  x = pmul(x, pmadd(neg_half, pmul(x, x), p8d_one_point_five));
+
+  // Multiply the original _x by it's reciprocal square root to extract the
+  // square root.
+  return pmul(_x, x);
+}
+#else
+template <>
+EIGEN_STRONG_INLINE Packet16f psqrt<Packet16f>(const Packet16f& x) {
+  return _mm512_sqrt_ps(x);
+}
+template <>
+EIGEN_STRONG_INLINE Packet8d psqrt<Packet8d>(const Packet8d& x) {
+  return _mm512_sqrt_pd(x);
+}
+#endif
+
+// Functions for rsqrt.
+// Almost identical to the sqrt routine, just leave out the last multiplication
+// and fill in NaN/Inf where needed. Note that this function only exists as an
+// iterative version for doubles since there is no instruction for diretly
+// computing the reciprocal square root in AVX-512.
+#ifdef EIGEN_FAST_MATH
+template <>
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet16f
+prsqrt<Packet16f>(const Packet16f& _x) {
+  _EIGEN_DECLARE_CONST_Packet16f_FROM_INT(inf, 0x7f800000);
+  _EIGEN_DECLARE_CONST_Packet16f_FROM_INT(nan, 0x7fc00000);
+  _EIGEN_DECLARE_CONST_Packet16f(one_point_five, 1.5f);
+  _EIGEN_DECLARE_CONST_Packet16f(minus_half, -0.5f);
+  _EIGEN_DECLARE_CONST_Packet16f_FROM_INT(flt_min, 0x00800000);
+
+  Packet16f neg_half = pmul(_x, p16f_minus_half);
+
+  // select only the inverse sqrt of positive normal inputs (denormals are
+  // flushed to zero and cause infs as well).
+  __mmask16 le_zero_mask = _mm512_cmp_ps_mask(_x, p16f_flt_min, _CMP_LT_OQ);
+  Packet16f x = _mm512_mask_blend_ps(le_zero_mask, _mm512_setzero_ps(),
+                                     _mm512_rsqrt14_ps(_x));
+
+  // Fill in NaNs and Infs for the negative/zero entries.
+  __mmask16 neg_mask = _mm512_cmp_ps_mask(_x, _mm512_setzero_ps(), _CMP_LT_OQ);
+  Packet16f infs_and_nans = _mm512_mask_blend_ps(
+      neg_mask, p16f_nan,
+      _mm512_mask_blend_ps(le_zero_mask, p16f_inf, _mm512_setzero_ps()));
+
+  // Do a single step of Newton's iteration.
+  x = pmul(x, pmadd(neg_half, pmul(x, x), p16f_one_point_five));
+
+  // Insert NaNs and Infs in all the right places.
+  return _mm512_mask_blend_ps(le_zero_mask, infs_and_nans, x);
+}
+
+template <>
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet8d
+prsqrt<Packet8d>(const Packet8d& _x) {
+  _EIGEN_DECLARE_CONST_Packet8d_FROM_INT64(inf, 0x7ff0000000000000LL);
+  _EIGEN_DECLARE_CONST_Packet8d_FROM_INT64(nan, 0x7ff1000000000000LL);
+  _EIGEN_DECLARE_CONST_Packet8d(one_point_five, 1.5);
+  _EIGEN_DECLARE_CONST_Packet8d(minus_half, -0.5);
+  _EIGEN_DECLARE_CONST_Packet8d_FROM_INT64(dbl_min, 0x0010000000000000LL);
+
+  Packet8d neg_half = pmul(_x, p8d_minus_half);
+
+  // select only the inverse sqrt of positive normal inputs (denormals are
+  // flushed to zero and cause infs as well).
+  __mmask8 le_zero_mask = _mm512_cmp_pd_mask(_x, p8d_dbl_min, _CMP_LT_OQ);
+  Packet8d x = _mm512_mask_blend_pd(le_zero_mask, _mm512_setzero_pd(),
+                                    _mm512_rsqrt14_pd(_x));
+
+  // Fill in NaNs and Infs for the negative/zero entries.
+  __mmask8 neg_mask = _mm512_cmp_pd_mask(_x, _mm512_setzero_pd(), _CMP_LT_OQ);
+  Packet8d infs_and_nans = _mm512_mask_blend_pd(
+      neg_mask, p8d_nan,
+      _mm512_mask_blend_pd(le_zero_mask, p8d_inf, _mm512_setzero_pd()));
+
+  // Do a first step of Newton's iteration.
+  x = pmul(x, pmadd(neg_half, pmul(x, x), p8d_one_point_five));
+
+  // Do a second step of Newton's iteration.
+  x = pmul(x, pmadd(neg_half, pmul(x, x), p8d_one_point_five));
+
+  // Insert NaNs and Infs in all the right places.
+  return _mm512_mask_blend_pd(le_zero_mask, infs_and_nans, x);
+}
+#else
+template <>
+EIGEN_STRONG_INLINE Packet16f prsqrt<Packet16f>(const Packet16f& x) {
+  return _mm512_rsqrt28_ps(x);
+}
+#endif
+#endif
+
+}  // end namespace internal
+
+}  // end namespace Eigen
+
+#endif  // THIRD_PARTY_EIGEN3_EIGEN_SRC_CORE_ARCH_AVX512_MATHFUNCTIONS_H_

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

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

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

@@ -15,14 +15,14 @@ namespace Eigen {
 
 namespace internal {
 
-static Packet4ui  p4ui_CONJ_XOR = vec_mergeh((Packet4ui)p4i_ZERO, (Packet4ui)p4f_ZERO_);//{ 0x00000000, 0x80000000, 0x00000000, 0x80000000 };
+static Packet4ui  p4ui_CONJ_XOR = vec_mergeh((Packet4ui)p4i_ZERO, (Packet4ui)p4f_MZERO);//{ 0x00000000, 0x80000000, 0x00000000, 0x80000000 };
 #ifdef __VSX__
 #if defined(_BIG_ENDIAN)
-static Packet2ul  p2ul_CONJ_XOR1 = (Packet2ul) vec_sld((Packet4ui) p2d_ZERO_, (Packet4ui) p2l_ZERO, 8);//{ 0x8000000000000000, 0x0000000000000000 };
-static Packet2ul  p2ul_CONJ_XOR2 = (Packet2ul) vec_sld((Packet4ui) p2l_ZERO,  (Packet4ui) p2d_ZERO_, 8);//{ 0x8000000000000000, 0x0000000000000000 };
+static Packet2ul  p2ul_CONJ_XOR1 = (Packet2ul) vec_sld((Packet4ui) p2d_MZERO, (Packet4ui) p2l_ZERO, 8);//{ 0x8000000000000000, 0x0000000000000000 };
+static Packet2ul  p2ul_CONJ_XOR2 = (Packet2ul) vec_sld((Packet4ui) p2l_ZERO,  (Packet4ui) p2d_MZERO, 8);//{ 0x8000000000000000, 0x0000000000000000 };
 #else
-static Packet2ul  p2ul_CONJ_XOR1 = (Packet2ul) vec_sld((Packet4ui) p2l_ZERO,  (Packet4ui) p2d_ZERO_, 8);//{ 0x8000000000000000, 0x0000000000000000 };
-static Packet2ul  p2ul_CONJ_XOR2 = (Packet2ul) vec_sld((Packet4ui) p2d_ZERO_, (Packet4ui) p2l_ZERO, 8);//{ 0x8000000000000000, 0x0000000000000000 };
+static Packet2ul  p2ul_CONJ_XOR1 = (Packet2ul) vec_sld((Packet4ui) p2l_ZERO,  (Packet4ui) p2d_MZERO, 8);//{ 0x8000000000000000, 0x0000000000000000 };
+static Packet2ul  p2ul_CONJ_XOR2 = (Packet2ul) vec_sld((Packet4ui) p2d_MZERO, (Packet4ui) p2l_ZERO, 8);//{ 0x8000000000000000, 0x0000000000000000 };
 #endif
 #endif
 
@@ -65,7 +65,7 @@ template<> struct unpacket_traits<Packet2cf> { typedef std::complex<float> type;
 template<> EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(const std::complex<float>&  from)
 {
   Packet2cf res;
-  if((ptrdiff_t(&from) % 16) == 0)
+  if((std::ptrdiff_t(&from) % 16) == 0)
     res.v = pload<Packet4f>((const float *)&from);
   else
     res.v = ploadu<Packet4f>((const float *)&from);

Eigen/src/Core/arch/AltiVec/MathFunctions.h

@@ -84,8 +84,10 @@ static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q3, 2.00000000000000000009e0);
 static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C1, 0.693145751953125);
 static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C2, 1.42860682030941723212e-6);
 
+#ifdef __POWER8_VECTOR__
 static Packet2l p2l_1023 = { 1023, 1023 };
 static Packet2ul p2ul_52 = { 52, 52 };
+#endif
 
 #endif
 

Eigen/src/Core/arch/AltiVec/PacketMath.h

@@ -72,7 +72,7 @@ static _EIGEN_DECLARE_CONST_FAST_Packet4i(ZERO, 0); //{ 0, 0, 0, 0,}
 static _EIGEN_DECLARE_CONST_FAST_Packet4i(ONE,1); //{ 1, 1, 1, 1}
 static _EIGEN_DECLARE_CONST_FAST_Packet4i(MINUS16,-16); //{ -16, -16, -16, -16}
 static _EIGEN_DECLARE_CONST_FAST_Packet4i(MINUS1,-1); //{ -1, -1, -1, -1}
-static Packet4f p4f_ZERO_ = (Packet4f) vec_sl((Packet4ui)p4i_MINUS1, (Packet4ui)p4i_MINUS1); //{ 0x80000000, 0x80000000, 0x80000000, 0x80000000}
+static Packet4f p4f_MZERO = (Packet4f) vec_sl((Packet4ui)p4i_MINUS1, (Packet4ui)p4i_MINUS1); //{ 0x80000000, 0x80000000, 0x80000000, 0x80000000}
 #ifndef __VSX__
 static Packet4f p4f_ONE = vec_ctf(p4i_ONE, 0); //{ 1.0, 1.0, 1.0, 1.0}
 #endif
@@ -90,7 +90,7 @@ static Packet16uc p16uc_DUPLICATE32_HI = { 0,1,2,3, 0,1,2,3, 4,5,6,7, 4,5,6,7 };
 #define _EIGEN_MASK_ALIGNMENT	0xfffffff0
 #endif
 
-#define _EIGEN_ALIGNED_PTR(x)	((ptrdiff_t)(x) & _EIGEN_MASK_ALIGNMENT)
+#define _EIGEN_ALIGNED_PTR(x)	((std::ptrdiff_t)(x) & _EIGEN_MASK_ALIGNMENT)
 
 // Handle endianness properly while loading constants
 // Define global static constants:
@@ -358,7 +358,7 @@ template<> EIGEN_STRONG_INLINE Packet4i pnegate(const Packet4i& a) { return p4i_
 template<> EIGEN_STRONG_INLINE Packet4f pconj(const Packet4f& a) { return a; }
 template<> EIGEN_STRONG_INLINE Packet4i pconj(const Packet4i& a) { return a; }
 
-template<> EIGEN_STRONG_INLINE Packet4f pmul<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_madd(a,b, p4f_ZERO); }
+template<> EIGEN_STRONG_INLINE Packet4f pmul<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_madd(a,b, p4f_MZERO); }
 template<> EIGEN_STRONG_INLINE Packet4i pmul<Packet4i>(const Packet4i& a, const Packet4i& b) { return a * b; }
 
 template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const Packet4f& b)
@@ -373,7 +373,7 @@ template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const
   t   = vec_nmsub(y_0, b, p4f_ONE);
   y_1 = vec_madd(y_0, t, y_0);
 
-  return vec_madd(a, y_1, p4f_ZERO);
+  return vec_madd(a, y_1, p4f_MZERO);
 #else
   return vec_div(a, b);
 #endif
@@ -450,15 +450,15 @@ template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from)
 template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float*   from)
 {
   Packet4f p;
-  if((ptrdiff_t(from) % 16) == 0)  p = pload<Packet4f>(from);
-  else                             p = ploadu<Packet4f>(from);
+  if((std::ptrdiff_t(from) % 16) == 0)  p = pload<Packet4f>(from);
+  else                                  p = ploadu<Packet4f>(from);
   return vec_perm(p, p, p16uc_DUPLICATE32_HI);
 }
 template<> EIGEN_STRONG_INLINE Packet4i ploaddup<Packet4i>(const int*     from)
 {
   Packet4i p;
-  if((ptrdiff_t(from) % 16) == 0)  p = pload<Packet4i>(from);
-  else                             p = ploadu<Packet4i>(from);
+  if((std::ptrdiff_t(from) % 16) == 0)  p = pload<Packet4i>(from);
+  else                                  p = ploadu<Packet4i>(from);
   return vec_perm(p, p, p16uc_DUPLICATE32_HI);
 }
 
@@ -766,7 +766,7 @@ static Packet2l  p2l_ONE  = { 1, 1 };
 static Packet2l  p2l_ZERO = reinterpret_cast<Packet2l>(p4i_ZERO);
 static Packet2d  p2d_ONE  = { 1.0, 1.0 }; 
 static Packet2d  p2d_ZERO = reinterpret_cast<Packet2d>(p4f_ZERO);
-static Packet2d  p2d_ZERO_ = { -0.0, -0.0 };
+static Packet2d  p2d_MZERO = { -0.0, -0.0 };
 
 #ifdef _BIG_ENDIAN
 static Packet2d p2d_COUNTDOWN = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4f>(p2d_ZERO), reinterpret_cast<Packet4f>(p2d_ONE), 8));
@@ -904,7 +904,7 @@ template<> EIGEN_STRONG_INLINE Packet2d pnegate(const Packet2d& a) { return p2d_
 
 template<> EIGEN_STRONG_INLINE Packet2d pconj(const Packet2d& a) { return a; }
 
-template<> EIGEN_STRONG_INLINE Packet2d pmul<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_madd(a,b,p2d_ZERO); }
+template<> EIGEN_STRONG_INLINE Packet2d pmul<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_madd(a,b,p2d_MZERO); }
 template<> EIGEN_STRONG_INLINE Packet2d pdiv<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_div(a,b); }
 
 // for some weird raisons, it has to be overloaded for packet of integers
@@ -935,8 +935,8 @@ template<> EIGEN_STRONG_INLINE Packet2d ploadu<Packet2d>(const double* from)
 template<> EIGEN_STRONG_INLINE Packet2d ploaddup<Packet2d>(const double*   from)
 {
   Packet2d p;
-  if((ptrdiff_t(from) % 16) == 0)  p = pload<Packet2d>(from);
-  else                             p = ploadu<Packet2d>(from);
+  if((std::ptrdiff_t(from) % 16) == 0)  p = pload<Packet2d>(from);
+  else                                  p = ploadu<Packet2d>(from);
   return vec_splat_dbl<0>(p);
 }
 

Eigen/src/Core/arch/CMakeLists.txt

@@ -1,9 +0,0 @@
-ADD_SUBDIRECTORY(AltiVec)
-ADD_SUBDIRECTORY(AVX)
-ADD_SUBDIRECTORY(CUDA)
-ADD_SUBDIRECTORY(Default)
-ADD_SUBDIRECTORY(NEON)
-ADD_SUBDIRECTORY(SSE)
-
-
-

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

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

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

@@ -0,0 +1,103 @@
+// 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_CUDA_H
+#define EIGEN_COMPLEX_CUDA_H
+
+// clang-format off
+
+namespace Eigen {
+
+namespace internal {
+
+#if defined(__CUDACC__) && defined(EIGEN_USE_GPU)
+
+// Many std::complex methods such as operator+, operator-, operator* and
+// operator/ are not constexpr. Due to this, clang does not treat them as device
+// functions and thus Eigen functors making use of these operators fail to
+// compile. Here, we manually specialize these functors for complex types when
+// building for CUDA to avoid non-constexpr methods.
+
+// Sum
+template<typename T> struct scalar_sum_op<const std::complex<T>, const std::complex<T> > : binary_op_base<const std::complex<T>, const std::complex<T> > {
+  typedef typename std::complex<T> result_type;
+
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_sum_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::complex<T> operator() (const std::complex<T>& a, const std::complex<T>& b) const {
+    return std::complex<T>(numext::real(a) + numext::real(b),
+                           numext::imag(a) + numext::imag(b));
+  }
+};
+
+template<typename T> struct scalar_sum_op<std::complex<T>, std::complex<T> > : scalar_sum_op<const std::complex<T>, const std::complex<T> > {};
+
+
+// Difference
+template<typename T> struct scalar_difference_op<const std::complex<T>, const std::complex<T> >  : binary_op_base<const std::complex<T>, const std::complex<T> > {
+  typedef typename std::complex<T> result_type;
+
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_difference_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::complex<T> operator() (const std::complex<T>& a, const std::complex<T>& b) const {
+    return std::complex<T>(numext::real(a) - numext::real(b),
+                           numext::imag(a) - numext::imag(b));
+  }
+};
+
+template<typename T> struct scalar_difference_op<std::complex<T>, std::complex<T> > : scalar_difference_op<const std::complex<T>, const std::complex<T> > {};
+
+
+// Product
+template<typename T> struct scalar_product_op<const std::complex<T>, const std::complex<T> >  : binary_op_base<const std::complex<T>, const std::complex<T> > {
+  enum {
+    Vectorizable = packet_traits<std::complex<T>>::HasMul
+  };
+  typedef typename std::complex<T> result_type;
+
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_product_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::complex<T> operator() (const std::complex<T>& a, const std::complex<T>& b) const {
+    const T a_real = numext::real(a);
+    const T a_imag = numext::imag(a);
+    const T b_real = numext::real(b);
+    const T b_imag = numext::imag(b);
+    return std::complex<T>(a_real * b_real - a_imag * b_imag,
+                           a_real * b_imag + a_imag * b_real);
+  }
+};
+
+template<typename T> struct scalar_product_op<std::complex<T>, std::complex<T> > : scalar_product_op<const std::complex<T>, const std::complex<T> > {};
+
+
+// Quotient
+template<typename T> struct scalar_quotient_op<const std::complex<T>, const std::complex<T> > : binary_op_base<const std::complex<T>, const std::complex<T> > {
+  enum {
+    Vectorizable = packet_traits<std::complex<T>>::HasDiv
+  };
+  typedef typename std::complex<T> result_type;
+
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_quotient_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::complex<T> operator() (const std::complex<T>& a, const std::complex<T>& b) const {
+    const T a_real = numext::real(a);
+    const T a_imag = numext::imag(a);
+    const T b_real = numext::real(b);
+    const T b_imag = numext::imag(b);
+    const T norm = T(1) / (b_real * b_real + b_imag * b_imag);
+    return std::complex<T>((a_real * b_real + a_imag * b_imag) * norm,
+                           (a_imag * b_real - a_real * b_imag) * norm);
+  }
+};
+
+template<typename T> struct scalar_quotient_op<std::complex<T>, std::complex<T> > : scalar_quotient_op<const std::complex<T>, const std::complex<T> > {};
+
+#endif
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_COMPLEX_CUDA_H

Eigen/src/Core/arch/CUDA/Half.h

@@ -45,6 +45,8 @@
 
 namespace Eigen {
 
+struct half;
+
 namespace half_impl {
 
 #if !defined(EIGEN_HAS_CUDA_FP16)
@@ -62,60 +64,72 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half raw_uint16_to_half(unsigned short x
 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half float_to_half_rtne(float ff);
 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC float half_to_float(__half h);
 
+struct half_base : public __half {
+  EIGEN_DEVICE_FUNC half_base() {}
+  EIGEN_DEVICE_FUNC half_base(const half_base& h) : __half(h) {}
+  EIGEN_DEVICE_FUNC half_base(const __half& h) : __half(h) {}
+};
+
+} // namespace half_impl
+
 // Class definition.
-struct half : public __half {
+struct half : public half_impl::half_base {
+  #if !defined(EIGEN_HAS_CUDA_FP16)
+    typedef half_impl::__half __half;
+  #endif
+
   EIGEN_DEVICE_FUNC half() {}
 
-  EIGEN_DEVICE_FUNC half(const __half& h) : __half(h) {}
-  EIGEN_DEVICE_FUNC half(const half& h) : __half(h) {}
+  EIGEN_DEVICE_FUNC half(const __half& h) : half_impl::half_base(h) {}
+  EIGEN_DEVICE_FUNC half(const half& h) : half_impl::half_base(h) {}
 
   explicit EIGEN_DEVICE_FUNC half(bool b)
-      : __half(raw_uint16_to_half(b ? 0x3c00 : 0)) {}
+      : half_impl::half_base(half_impl::raw_uint16_to_half(b ? 0x3c00 : 0)) {}
   template<class T>
   explicit EIGEN_DEVICE_FUNC half(const T& val)
-      : __half(float_to_half_rtne(static_cast<float>(val))) {}
+      : half_impl::half_base(half_impl::float_to_half_rtne(static_cast<float>(val))) {}
   explicit EIGEN_DEVICE_FUNC half(float f)
-      : __half(float_to_half_rtne(f)) {}
+      : half_impl::half_base(half_impl::float_to_half_rtne(f)) {}
 
   EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(bool) const {
     // +0.0 and -0.0 become false, everything else becomes true.
     return (x & 0x7fff) != 0;
   }
   EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(signed char) const {
-    return static_cast<signed char>(half_to_float(*this));
+    return static_cast<signed char>(half_impl::half_to_float(*this));
   }
   EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(unsigned char) const {
-    return static_cast<unsigned char>(half_to_float(*this));
+    return static_cast<unsigned char>(half_impl::half_to_float(*this));
   }
   EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(short) const {
-    return static_cast<short>(half_to_float(*this));
+    return static_cast<short>(half_impl::half_to_float(*this));
   }
   EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(unsigned short) const {
-    return static_cast<unsigned short>(half_to_float(*this));
+    return static_cast<unsigned short>(half_impl::half_to_float(*this));
   }
   EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(int) const {
-    return static_cast<int>(half_to_float(*this));
+    return static_cast<int>(half_impl::half_to_float(*this));
   }
   EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(unsigned int) const {
-    return static_cast<unsigned int>(half_to_float(*this));
+    return static_cast<unsigned int>(half_impl::half_to_float(*this));
   }
   EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(long) const {
-    return static_cast<long>(half_to_float(*this));
+    return static_cast<long>(half_impl::half_to_float(*this));
   }
   EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(unsigned long) const {
-    return static_cast<unsigned long>(half_to_float(*this));
+    return static_cast<unsigned long>(half_impl::half_to_float(*this));
   }
   EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(long long) const {
-    return static_cast<long long>(half_to_float(*this));
+    return static_cast<long long>(half_impl::half_to_float(*this));
   }
   EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(unsigned long long) const {
     return static_cast<unsigned long long>(half_to_float(*this));
   }
   EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(float) const {
-    return half_to_float(*this);
+    return half_impl::half_to_float(*this);
   }
   EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(double) const {
-    return static_cast<double>(half_to_float(*this));
+    return static_cast<double>(half_impl::half_to_float(*this));
   }
 
   EIGEN_DEVICE_FUNC half& operator=(const half& other) {
@@ -124,6 +138,8 @@ struct half : public __half {
   }
 };
 
+namespace half_impl {
+
 #if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
 
 // Intrinsics for native fp16 support. Note that on current hardware,
@@ -373,7 +389,14 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half exp(const half& a) {
   return half(::expf(float(a)));
 }
 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half log(const half& a) {
+#if defined(EIGEN_HAS_CUDA_FP16) && defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
+  return Eigen::half(::hlog(a));
+#else
   return half(::logf(float(a)));
+#endif
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half log1p(const half& a) {
+  return half(numext::log1p(float(a)));
 }
 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half log10(const half& a) {
   return half(::log10f(float(a)));
@@ -430,12 +453,12 @@ EIGEN_ALWAYS_INLINE std::ostream& operator << (std::ostream& os, const half& v)
 } // end namespace half_impl
 
 // import Eigen::half_impl::half into Eigen namespace
-using half_impl::half;
+// using half_impl::half;
 
 namespace internal {
 
 template<>
-struct random_default_impl<half_impl::half, false, false>
+struct random_default_impl<half, false, false>
 {
   static inline half run(const half& x, const half& y)
   {
@@ -447,27 +470,27 @@ struct random_default_impl<half_impl::half, false, false>
   }
 };
 
-template<> struct is_arithmetic<half_impl::half> { enum { value = true }; };
+template<> struct is_arithmetic<half> { enum { value = true }; };
 
 } // end namespace internal
 
-template<> struct NumTraits<Eigen::half_impl::half>
-    : GenericNumTraits<Eigen::half_impl::half>
+template<> struct NumTraits<Eigen::half>
+    : GenericNumTraits<Eigen::half>
 {
-  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half_impl::half epsilon() {
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half epsilon() {
     return half_impl::raw_uint16_to_half(0x0800);
   }
-  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half_impl::half dummy_precision() { return half_impl::half(1e-2f); }
-  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half_impl::half highest() {
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half dummy_precision() { return Eigen::half(1e-2f); }
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half highest() {
     return half_impl::raw_uint16_to_half(0x7bff);
   }
-  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half_impl::half lowest() {
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half lowest() {
     return half_impl::raw_uint16_to_half(0xfbff);
   }
-  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half_impl::half infinity() {
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half infinity() {
     return half_impl::raw_uint16_to_half(0x7c00);
   }
-  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half_impl::half quiet_NaN() {
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half quiet_NaN() {
     return half_impl::raw_uint16_to_half(0x7c01);
   }
 };
@@ -484,7 +507,11 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half exph(const Eigen::half& a) {
   return Eigen::half(::expf(float(a)));
 }
 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half logh(const Eigen::half& a) {
+#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
+  return Eigen::half(::hlog(a));
+#else
   return Eigen::half(::logf(float(a)));
+#endif
 }
 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half sqrth(const Eigen::half& a) {
   return Eigen::half(::sqrtf(float(a)));
@@ -523,9 +550,36 @@ __device__ EIGEN_STRONG_INLINE Eigen::half __shfl_xor(Eigen::half var, int laneM
 // ldg() has an overload for __half, but we also need one for Eigen::half.
 #if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 350
 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half __ldg(const Eigen::half* ptr) {
-  return Eigen::internal::raw_uint16_to_half(
+  return Eigen::half_impl::raw_uint16_to_half(
       __ldg(reinterpret_cast<const unsigned short*>(ptr)));
 }
 #endif
 
+
+#if defined(__CUDA_ARCH__)
+namespace Eigen {
+namespace numext {
+
+template<>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+bool (isnan)(const Eigen::half& h) {
+  return (half_impl::isnan)(h);
+}
+
+template<>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+bool (isinf)(const Eigen::half& h) {
+  return (half_impl::isinf)(h);
+}
+
+template<>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+bool (isfinite)(const Eigen::half& h) {
+  return (half_impl::isfinite)(h);
+}
+
+} // namespace Eigen
+}  // namespace numext
+#endif
+
 #endif // EIGEN_HALF_CUDA_H

Eigen/src/Core/arch/CUDA/MathFunctions.h

@@ -31,6 +31,18 @@ double2 plog<double2>(const double2& a)
   return make_double2(log(a.x), log(a.y));
 }
 
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+float4 plog1p<float4>(const float4& a)
+{
+  return make_float4(log1pf(a.x), log1pf(a.y), log1pf(a.z), log1pf(a.w));
+}
+
+template<>  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+double2 plog1p<double2>(const double2& a)
+{
+  return make_double2(log1p(a.x), log1p(a.y));
+}
+
 template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
 float4 pexp<float4>(const float4& a)
 {

Eigen/src/Core/arch/CUDA/PacketMathHalf.h

@@ -15,7 +15,7 @@ namespace Eigen {
 namespace internal {
 
 // Most of the following operations require arch >= 3.0
-#if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300
+#if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDACC__) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300
 
 template<> struct is_arithmetic<half2> { enum { value = true }; };
 
@@ -34,21 +34,22 @@ template<> struct packet_traits<Eigen::half> : default_packet_traits
     HasSqrt   = 1,
     HasRsqrt  = 1,
     HasExp    = 1,
-    HasLog = 1
+    HasLog    = 1,
+    HasLog1p  = 1
   };
 };
 
 template<> struct unpacket_traits<half2> { typedef Eigen::half type; enum {size=2, alignment=Aligned16}; typedef half2 half; };
 
-template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pset1<half2>(const Eigen::half& from) {
+template<> __device__ EIGEN_STRONG_INLINE half2 pset1<half2>(const Eigen::half& from) {
   return __half2half2(from);
 }
 
-template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pload<half2>(const Eigen::half* from) {
+template<> __device__ EIGEN_STRONG_INLINE half2 pload<half2>(const Eigen::half* from) {
   return *reinterpret_cast<const half2*>(from);
 }
 
-template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 ploadu<half2>(const Eigen::half* from) {
+template<> __device__ EIGEN_STRONG_INLINE half2 ploadu<half2>(const Eigen::half* from) {
   return __halves2half2(from[0], from[1]);
 }
 
@@ -56,17 +57,17 @@ template<> EIGEN_STRONG_INLINE half2 ploaddup<half2>(const Eigen::half*  from) {
   return __halves2half2(from[0], from[0]);
 }
 
-template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pstore<Eigen::half>(Eigen::half* to, const half2& from) {
+template<> __device__ EIGEN_STRONG_INLINE void pstore<Eigen::half>(Eigen::half* to, const half2& from) {
   *reinterpret_cast<half2*>(to) = from;
 }
 
-template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pstoreu<Eigen::half>(Eigen::half* to, const half2& from) {
+template<> __device__ EIGEN_STRONG_INLINE void pstoreu<Eigen::half>(Eigen::half* to, const half2& from) {
   to[0] = __low2half(from);
   to[1] = __high2half(from);
 }
 
 template<>
- EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE half2 ploadt_ro<half2, Aligned>(const Eigen::half* from) {
+ __device__ EIGEN_ALWAYS_INLINE half2 ploadt_ro<half2, Aligned>(const Eigen::half* from) {
 #if __CUDA_ARCH__ >= 350
    return __ldg((const half2*)from);
 #else
@@ -75,7 +76,7 @@ template<>
 }
 
 template<>
-EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE half2 ploadt_ro<half2, Unaligned>(const Eigen::half* from) {
+__device__ EIGEN_ALWAYS_INLINE half2 ploadt_ro<half2, Unaligned>(const Eigen::half* from) {
 #if __CUDA_ARCH__ >= 350
    return __halves2half2(__ldg(from+0), __ldg(from+1));
 #else
@@ -83,27 +84,27 @@ EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE half2 ploadt_ro<half2, Unaligned>(const Ei
 #endif
 }
 
-template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pgather<Eigen::half, half2>(const Eigen::half* from, Index stride) {
+template<> __device__ EIGEN_STRONG_INLINE half2 pgather<Eigen::half, half2>(const Eigen::half* from, Index stride) {
   return __halves2half2(from[0*stride], from[1*stride]);
 }
 
-template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pscatter<Eigen::half, half2>(Eigen::half* to, const half2& from, Index stride) {
+template<> __device__ EIGEN_STRONG_INLINE void pscatter<Eigen::half, half2>(Eigen::half* to, const half2& from, Index stride) {
   to[stride*0] = __low2half(from);
   to[stride*1] = __high2half(from);
 }
 
-template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Eigen::half pfirst<half2>(const half2& a) {
+template<> __device__ EIGEN_STRONG_INLINE Eigen::half pfirst<half2>(const half2& a) {
   return __low2half(a);
 }
 
-template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pabs<half2>(const half2& a) {
+template<> __device__ EIGEN_STRONG_INLINE half2 pabs<half2>(const half2& a) {
   half2 result;
   result.x = a.x & 0x7FFF7FFF;
   return result;
 }
 
 
-EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void
+__device__ EIGEN_STRONG_INLINE void
 ptranspose(PacketBlock<half2,2>& kernel) {
   __half a1 = __low2half(kernel.packet[0]);
   __half a2 = __high2half(kernel.packet[0]);
@@ -113,7 +114,7 @@ ptranspose(PacketBlock<half2,2>& kernel) {
   kernel.packet[1] = __halves2half2(a2, b2);
 }
 
-template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 plset<half2>(const Eigen::half& a) {
+template<> __device__ EIGEN_STRONG_INLINE half2 plset<half2>(const Eigen::half& a) {
 #if __CUDA_ARCH__ >= 530
   return __halves2half2(a, __hadd(a, __float2half(1.0f)));
 #else
@@ -122,7 +123,7 @@ template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 plset<half2>(const Eigen:
 #endif
 }
 
-template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 padd<half2>(const half2& a, const half2& b) {
+template<> __device__ EIGEN_STRONG_INLINE half2 padd<half2>(const half2& a, const half2& b) {
 #if __CUDA_ARCH__ >= 530
   return __hadd2(a, b);
 #else
@@ -136,7 +137,7 @@ template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 padd<half2>(const half2&
 #endif
 }
 
-template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 psub<half2>(const half2& a, const half2& b) {
+template<> __device__ EIGEN_STRONG_INLINE half2 psub<half2>(const half2& a, const half2& b) {
 #if __CUDA_ARCH__ >= 530
   return __hsub2(a, b);
 #else
@@ -150,7 +151,7 @@ template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 psub<half2>(const half2&
 #endif
 }
 
-template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pnegate(const half2& a) {
+template<> __device__ EIGEN_STRONG_INLINE half2 pnegate(const half2& a) {
 #if __CUDA_ARCH__ >= 530
   return __hneg2(a);
 #else
@@ -160,9 +161,9 @@ template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pnegate(const half2& a) {
 #endif
 }
 
-template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pconj(const half2& a) { return a; }
+template<> __device__ EIGEN_STRONG_INLINE half2 pconj(const half2& a) { return a; }
 
-template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pmul<half2>(const half2& a, const half2& b) {
+template<> __device__ EIGEN_STRONG_INLINE half2 pmul<half2>(const half2& a, const half2& b) {
 #if __CUDA_ARCH__ >= 530
   return __hmul2(a, b);
 #else
@@ -176,7 +177,7 @@ template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pmul<half2>(const half2&
 #endif
 }
 
-template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pmadd<half2>(const half2& a, const half2& b, const half2& c) {
+template<> __device__ EIGEN_STRONG_INLINE half2 pmadd<half2>(const half2& a, const half2& b, const half2& c) {
 #if __CUDA_ARCH__ >= 530
    return __hfma2(a, b, c);
 #else
@@ -192,7 +193,7 @@ template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pmadd<half2>(const half2&
 #endif
 }
 
-template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pdiv<half2>(const half2& a, const half2& b) {
+template<> __device__ EIGEN_STRONG_INLINE half2 pdiv<half2>(const half2& a, const half2& b) {
   float a1 = __low2float(a);
   float a2 = __high2float(a);
   float b1 = __low2float(b);
@@ -202,7 +203,7 @@ template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pdiv<half2>(const half2&
   return __floats2half2_rn(r1, r2);
 }
 
-template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pmin<half2>(const half2& a, const half2& b) {
+template<> __device__ EIGEN_STRONG_INLINE half2 pmin<half2>(const half2& a, const half2& b) {
   float a1 = __low2float(a);
   float a2 = __high2float(a);
   float b1 = __low2float(b);
@@ -212,7 +213,7 @@ template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pmin<half2>(const half2&
   return __halves2half2(r1, r2);
 }
 
-template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pmax<half2>(const half2& a, const half2& b) {
+template<> __device__ EIGEN_STRONG_INLINE half2 pmax<half2>(const half2& a, const half2& b) {
   float a1 = __low2float(a);
   float a2 = __high2float(a);
   float b1 = __low2float(b);
@@ -222,7 +223,7 @@ template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pmax<half2>(const half2&
   return __halves2half2(r1, r2);
 }
 
-template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Eigen::half predux<half2>(const half2& a) {
+template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux<half2>(const half2& a) {
 #if __CUDA_ARCH__ >= 530
   return __hadd(__low2half(a), __high2half(a));
 #else
@@ -232,7 +233,7 @@ template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Eigen::half predux<half2>(const
 #endif
 }
 
-template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Eigen::half predux_max<half2>(const half2& a) {
+template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux_max<half2>(const half2& a) {
 #if __CUDA_ARCH__ >= 530
   __half first = __low2half(a);
   __half second = __high2half(a);
@@ -244,7 +245,7 @@ template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Eigen::half predux_max<half2>(c
 #endif
 }
 
-template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Eigen::half predux_min<half2>(const half2& a) {
+template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux_min<half2>(const half2& a) {
 #if __CUDA_ARCH__ >= 530
   __half first = __low2half(a);
   __half second = __high2half(a);
@@ -256,7 +257,7 @@ template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Eigen::half predux_min<half2>(c
 #endif
 }
 
-template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Eigen::half predux_mul<half2>(const half2& a) {
+template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux_mul<half2>(const half2& a) {
 #if __CUDA_ARCH__ >= 530
   return __hmul(__low2half(a), __high2half(a));
 #else
@@ -266,31 +267,39 @@ template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Eigen::half predux_mul<half2>(c
 #endif
 }
 
+template<> __device__ EIGEN_STRONG_INLINE half2 plog1p<half2>(const half2& a) {
+  float a1 = __low2float(a);
+  float a2 = __high2float(a);
+  float r1 = log1pf(a1);
+  float r2 = log1pf(a2);
+  return __floats2half2_rn(r1, r2);
+}
+
 #if defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 530
 
-template<>  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+template<>  __device__ EIGEN_STRONG_INLINE
 half2 plog<half2>(const half2& a) {
   return h2log(a);
 }
 
-template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+template<> __device__ EIGEN_STRONG_INLINE
 half2 pexp<half2>(const half2& a) {
   return h2exp(a);
 }
 
-template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+template<> __device__ EIGEN_STRONG_INLINE
 half2 psqrt<half2>(const half2& a) {
   return h2sqrt(a);
 }
 
-template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+template<> __device__ EIGEN_STRONG_INLINE
 half2 prsqrt<half2>(const half2& a) {
   return h2rsqrt(a);
 }
 
 #else
 
-template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 plog<half2>(const half2& a) {
+template<> __device__ EIGEN_STRONG_INLINE half2 plog<half2>(const half2& a) {
   float a1 = __low2float(a);
   float a2 = __high2float(a);
   float r1 = logf(a1);
@@ -298,7 +307,7 @@ template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 plog<half2>(const half2&
   return __floats2half2_rn(r1, r2);
 }
 
-template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pexp<half2>(const half2& a) {
+template<> __device__ EIGEN_STRONG_INLINE half2 pexp<half2>(const half2& a) {
   float a1 = __low2float(a);
   float a2 = __high2float(a);
   float r1 = expf(a1);
@@ -306,7 +315,7 @@ template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pexp<half2>(const half2&
   return __floats2half2_rn(r1, r2);
 }
 
-template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 psqrt<half2>(const half2& a) {
+template<> __device__ EIGEN_STRONG_INLINE half2 psqrt<half2>(const half2& a) {
   float a1 = __low2float(a);
   float a2 = __high2float(a);
   float r1 = sqrtf(a1);
@@ -314,7 +323,7 @@ template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 psqrt<half2>(const half2&
   return __floats2half2_rn(r1, r2);
 }
 
-template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 prsqrt<half2>(const half2& a) {
+template<> __device__ EIGEN_STRONG_INLINE half2 prsqrt<half2>(const half2& a) {
   float a1 = __low2float(a);
   float a2 = __high2float(a);
   float r1 = rsqrtf(a1);
@@ -324,6 +333,374 @@ template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 prsqrt<half2>(const half2
 
 #endif
 
+#elif defined EIGEN_VECTORIZE_AVX512
+
+typedef struct {
+  __m256i x;
+} Packet16h;
+
+
+template<> struct is_arithmetic<Packet16h> { enum { value = true }; };
+
+template <>
+struct packet_traits<half> : default_packet_traits {
+  typedef Packet16h type;
+  // There is no half-size packet for Packet16h.
+  typedef Packet16h half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size = 16,
+    HasHalfPacket = 0,
+    HasAdd    = 0,
+    HasSub    = 0,
+    HasMul    = 0,
+    HasNegate = 0,
+    HasAbs    = 0,
+    HasAbs2   = 0,
+    HasMin    = 0,
+    HasMax    = 0,
+    HasConj   = 0,
+    HasSetLinear = 0,
+    HasDiv = 0,
+    HasSqrt = 0,
+    HasRsqrt = 0,
+    HasExp = 0,
+    HasLog = 0,
+    HasBlend = 0
+  };
+};
+
+
+template<> struct unpacket_traits<Packet16h> { typedef Eigen::half type; enum {size=16, alignment=Aligned32}; typedef Packet16h half; };
+
+template<> EIGEN_STRONG_INLINE Packet16h pset1<Packet16h>(const Eigen::half& from) {
+  Packet16h result;
+  result.x = _mm256_set1_epi16(from.x);
+  return result;
+}
+
+template<> EIGEN_STRONG_INLINE Eigen::half pfirst<Packet16h>(const Packet16h& from) {
+  return half_impl::raw_uint16_to_half(static_cast<unsigned short>(_mm256_extract_epi16(from.x, 0)));
+}
+
+template<> EIGEN_STRONG_INLINE Packet16h pload<Packet16h>(const Eigen::half* from) {
+  Packet16h result;
+  result.x = _mm256_load_si256(reinterpret_cast<const __m256i*>(from));
+  return result;
+}
+
+template<> EIGEN_STRONG_INLINE Packet16h ploadu<Packet16h>(const Eigen::half* from) {
+  Packet16h result;
+  result.x = _mm256_loadu_si256(reinterpret_cast<const __m256i*>(from));
+  return result;
+}
+
+template<> EIGEN_STRONG_INLINE void pstore<half>(Eigen::half* to, const Packet16h& from) {
+  _mm256_store_si256((__m256i*)to, from.x);
+}
+
+template<> EIGEN_STRONG_INLINE void pstoreu<half>(Eigen::half* to, const Packet16h& from) {
+  _mm256_storeu_si256((__m256i*)to, from.x);
+}
+
+template<> EIGEN_STRONG_INLINE Packet16h
+ploadquad(const Eigen::half* from) {
+  Packet16h result;
+  unsigned short a = from[0].x;
+  unsigned short b = from[1].x;
+  unsigned short c = from[2].x;
+  unsigned short d = from[3].x;
+  result.x = _mm256_set_epi16(d, d, d, d, c, c, c, c, b, b, b, b, a, a, a, a);
+  return result;
+}
+
+EIGEN_STRONG_INLINE Packet16f half2float(const Packet16h& a) {
+#ifdef EIGEN_HAS_FP16_C
+  return _mm512_cvtph_ps(a.x);
+#else
+  EIGEN_ALIGN64 half aux[16];
+  pstore(aux, a);
+  float f0(aux[0]);
+  float f1(aux[1]);
+  float f2(aux[2]);
+  float f3(aux[3]);
+  float f4(aux[4]);
+  float f5(aux[5]);
+  float f6(aux[6]);
+  float f7(aux[7]);
+  float f8(aux[8]);
+  float f9(aux[9]);
+  float fa(aux[10]);
+  float fb(aux[11]);
+  float fc(aux[12]);
+  float fd(aux[13]);
+  float fe(aux[14]);
+  float ff(aux[15]);
+
+  return _mm512_set_ps(
+      ff, fe, fd, fc, fb, fa, f9, f8, f7, f6, f5, f4, f3, f2, f1, f0);
+#endif
+}
+
+EIGEN_STRONG_INLINE Packet16h float2half(const Packet16f& a) {
+#ifdef EIGEN_HAS_FP16_C
+  Packet16h result;
+  result.x = _mm512_cvtps_ph(a, _MM_FROUND_TO_NEAREST_INT|_MM_FROUND_NO_EXC);
+  return result;
+#else
+  EIGEN_ALIGN64 float aux[16];
+  pstore(aux, a);
+  half h0(aux[0]);
+  half h1(aux[1]);
+  half h2(aux[2]);
+  half h3(aux[3]);
+  half h4(aux[4]);
+  half h5(aux[5]);
+  half h6(aux[6]);
+  half h7(aux[7]);
+  half h8(aux[8]);
+  half h9(aux[9]);
+  half ha(aux[10]);
+  half hb(aux[11]);
+  half hc(aux[12]);
+  half hd(aux[13]);
+  half he(aux[14]);
+  half hf(aux[15]);
+
+  Packet16h result;
+  result.x = _mm256_set_epi16(
+      hf.x, he.x, hd.x, hc.x, hb.x, ha.x, h9.x, h8.x,
+      h7.x, h6.x, h5.x, h4.x, h3.x, h2.x, h1.x, h0.x);
+  return result;
+#endif
+}
+
+template<> EIGEN_STRONG_INLINE Packet16h padd<Packet16h>(const Packet16h& a, const Packet16h& b) {
+  Packet16f af = half2float(a);
+  Packet16f bf = half2float(b);
+  Packet16f rf = padd(af, bf);
+  return float2half(rf);
+}
+
+template<> EIGEN_STRONG_INLINE Packet16h pmul<Packet16h>(const Packet16h& a, const Packet16h& b) {
+  Packet16f af = half2float(a);
+  Packet16f bf = half2float(b);
+  Packet16f rf = pmul(af, bf);
+  return float2half(rf);
+}
+
+template<> EIGEN_STRONG_INLINE half predux<Packet16h>(const Packet16h& from) {
+  Packet16f from_float = half2float(from);
+  return half(predux(from_float));
+}
+
+template<> EIGEN_STRONG_INLINE Packet16h pgather<Eigen::half, Packet16h>(const Eigen::half* from, Index stride)
+{
+  Packet16h result;
+  result.x = _mm256_set_epi16(
+      from[15*stride].x, from[14*stride].x, from[13*stride].x, from[12*stride].x,
+      from[11*stride].x, from[10*stride].x, from[9*stride].x, from[8*stride].x,
+      from[7*stride].x, from[6*stride].x, from[5*stride].x, from[4*stride].x,
+      from[3*stride].x, from[2*stride].x, from[1*stride].x, from[0*stride].x);
+  return result;
+}
+
+template<> EIGEN_STRONG_INLINE void pscatter<half, Packet16h>(half* to, const Packet16h& from, Index stride)
+{
+  EIGEN_ALIGN64 half aux[16];
+  pstore(aux, from);
+  to[stride*0].x = aux[0].x;
+  to[stride*1].x = aux[1].x;
+  to[stride*2].x = aux[2].x;
+  to[stride*3].x = aux[3].x;
+  to[stride*4].x = aux[4].x;
+  to[stride*5].x = aux[5].x;
+  to[stride*6].x = aux[6].x;
+  to[stride*7].x = aux[7].x;
+  to[stride*8].x = aux[8].x;
+  to[stride*9].x = aux[9].x;
+  to[stride*10].x = aux[10].x;
+  to[stride*11].x = aux[11].x;
+  to[stride*12].x = aux[12].x;
+  to[stride*13].x = aux[13].x;
+  to[stride*14].x = aux[14].x;
+  to[stride*15].x = aux[15].x;
+}
+
+EIGEN_STRONG_INLINE void
+ptranspose(PacketBlock<Packet16h,16>& kernel) {
+  __m256i a = kernel.packet[0].x;
+  __m256i b = kernel.packet[1].x;
+  __m256i c = kernel.packet[2].x;
+  __m256i d = kernel.packet[3].x;
+  __m256i e = kernel.packet[4].x;
+  __m256i f = kernel.packet[5].x;
+  __m256i g = kernel.packet[6].x;
+  __m256i h = kernel.packet[7].x;
+  __m256i i = kernel.packet[8].x;
+  __m256i j = kernel.packet[9].x;
+  __m256i k = kernel.packet[10].x;
+  __m256i l = kernel.packet[11].x;
+  __m256i m = kernel.packet[12].x;
+  __m256i n = kernel.packet[13].x;
+  __m256i o = kernel.packet[14].x;
+  __m256i p = kernel.packet[15].x;
+
+  __m256i ab_07 = _mm256_unpacklo_epi16(a, b);
+  __m256i cd_07 = _mm256_unpacklo_epi16(c, d);
+  __m256i ef_07 = _mm256_unpacklo_epi16(e, f);
+  __m256i gh_07 = _mm256_unpacklo_epi16(g, h);
+  __m256i ij_07 = _mm256_unpacklo_epi16(i, j);
+  __m256i kl_07 = _mm256_unpacklo_epi16(k, l);
+  __m256i mn_07 = _mm256_unpacklo_epi16(m, n);
+  __m256i op_07 = _mm256_unpacklo_epi16(o, p);
+
+  __m256i ab_8f = _mm256_unpackhi_epi16(a, b);
+  __m256i cd_8f = _mm256_unpackhi_epi16(c, d);
+  __m256i ef_8f = _mm256_unpackhi_epi16(e, f);
+  __m256i gh_8f = _mm256_unpackhi_epi16(g, h);
+  __m256i ij_8f = _mm256_unpackhi_epi16(i, j);
+  __m256i kl_8f = _mm256_unpackhi_epi16(k, l);
+  __m256i mn_8f = _mm256_unpackhi_epi16(m, n);
+  __m256i op_8f = _mm256_unpackhi_epi16(o, p);
+
+  __m256i abcd_03 = _mm256_unpacklo_epi32(ab_07, cd_07);
+  __m256i abcd_47 = _mm256_unpackhi_epi32(ab_07, cd_07);
+  __m256i efgh_03 = _mm256_unpacklo_epi32(ef_07, gh_07);
+  __m256i efgh_47 = _mm256_unpackhi_epi32(ef_07, gh_07);
+  __m256i ijkl_03 = _mm256_unpacklo_epi32(ij_07, kl_07);
+  __m256i ijkl_47 = _mm256_unpackhi_epi32(ij_07, kl_07);
+  __m256i mnop_03 = _mm256_unpacklo_epi32(mn_07, op_07);
+  __m256i mnop_47 = _mm256_unpackhi_epi32(mn_07, op_07);
+
+  __m256i abcd_8b = _mm256_unpacklo_epi32(ab_8f, cd_8f);
+  __m256i abcd_cf = _mm256_unpackhi_epi32(ab_8f, cd_8f);
+  __m256i efgh_8b = _mm256_unpacklo_epi32(ef_8f, gh_8f);
+  __m256i efgh_cf = _mm256_unpackhi_epi32(ef_8f, gh_8f);
+  __m256i ijkl_8b = _mm256_unpacklo_epi32(ij_8f, kl_8f);
+  __m256i ijkl_cf = _mm256_unpackhi_epi32(ij_8f, kl_8f);
+  __m256i mnop_8b = _mm256_unpacklo_epi32(mn_8f, op_8f);
+  __m256i mnop_cf = _mm256_unpackhi_epi32(mn_8f, op_8f);
+
+  __m256i abcdefgh_01 = _mm256_unpacklo_epi64(abcd_03, efgh_03);
+  __m256i abcdefgh_23 = _mm256_unpackhi_epi64(abcd_03, efgh_03);
+  __m256i ijklmnop_01 = _mm256_unpacklo_epi64(ijkl_03, mnop_03);
+  __m256i ijklmnop_23 = _mm256_unpackhi_epi64(ijkl_03, mnop_03);
+  __m256i abcdefgh_45 = _mm256_unpacklo_epi64(abcd_47, efgh_47);
+  __m256i abcdefgh_67 = _mm256_unpackhi_epi64(abcd_47, efgh_47);
+  __m256i ijklmnop_45 = _mm256_unpacklo_epi64(ijkl_47, mnop_47);
+  __m256i ijklmnop_67 = _mm256_unpackhi_epi64(ijkl_47, mnop_47);
+  __m256i abcdefgh_89 = _mm256_unpacklo_epi64(abcd_8b, efgh_8b);
+  __m256i abcdefgh_ab = _mm256_unpackhi_epi64(abcd_8b, efgh_8b);
+  __m256i ijklmnop_89 = _mm256_unpacklo_epi64(ijkl_8b, mnop_8b);
+  __m256i ijklmnop_ab = _mm256_unpackhi_epi64(ijkl_8b, mnop_8b);
+  __m256i abcdefgh_cd = _mm256_unpacklo_epi64(abcd_cf, efgh_cf);
+  __m256i abcdefgh_ef = _mm256_unpackhi_epi64(abcd_cf, efgh_cf);
+  __m256i ijklmnop_cd = _mm256_unpacklo_epi64(ijkl_cf, mnop_cf);
+  __m256i ijklmnop_ef = _mm256_unpackhi_epi64(ijkl_cf, mnop_cf);
+
+  // NOTE: no unpacklo/hi instr in this case, so using permute instr.
+  __m256i a_p_0 = _mm256_permute2x128_si256(abcdefgh_01, ijklmnop_01, 0x20);
+  __m256i a_p_1 = _mm256_permute2x128_si256(abcdefgh_01, ijklmnop_01, 0x31);
+  __m256i a_p_2 = _mm256_permute2x128_si256(abcdefgh_23, ijklmnop_23, 0x20);
+  __m256i a_p_3 = _mm256_permute2x128_si256(abcdefgh_23, ijklmnop_23, 0x31);
+  __m256i a_p_4 = _mm256_permute2x128_si256(abcdefgh_45, ijklmnop_45, 0x20);
+  __m256i a_p_5 = _mm256_permute2x128_si256(abcdefgh_45, ijklmnop_45, 0x31);
+  __m256i a_p_6 = _mm256_permute2x128_si256(abcdefgh_67, ijklmnop_67, 0x20);
+  __m256i a_p_7 = _mm256_permute2x128_si256(abcdefgh_67, ijklmnop_67, 0x31);
+  __m256i a_p_8 = _mm256_permute2x128_si256(abcdefgh_89, ijklmnop_89, 0x20);
+  __m256i a_p_9 = _mm256_permute2x128_si256(abcdefgh_89, ijklmnop_89, 0x31);
+  __m256i a_p_a = _mm256_permute2x128_si256(abcdefgh_ab, ijklmnop_ab, 0x20);
+  __m256i a_p_b = _mm256_permute2x128_si256(abcdefgh_ab, ijklmnop_ab, 0x31);
+  __m256i a_p_c = _mm256_permute2x128_si256(abcdefgh_cd, ijklmnop_cd, 0x20);
+  __m256i a_p_d = _mm256_permute2x128_si256(abcdefgh_cd, ijklmnop_cd, 0x31);
+  __m256i a_p_e = _mm256_permute2x128_si256(abcdefgh_ef, ijklmnop_ef, 0x20);
+  __m256i a_p_f = _mm256_permute2x128_si256(abcdefgh_ef, ijklmnop_ef, 0x31);
+
+  kernel.packet[0].x = a_p_0;
+  kernel.packet[1].x = a_p_1;
+  kernel.packet[2].x = a_p_2;
+  kernel.packet[3].x = a_p_3;
+  kernel.packet[4].x = a_p_4;
+  kernel.packet[5].x = a_p_5;
+  kernel.packet[6].x = a_p_6;
+  kernel.packet[7].x = a_p_7;
+  kernel.packet[8].x = a_p_8;
+  kernel.packet[9].x = a_p_9;
+  kernel.packet[10].x = a_p_a;
+  kernel.packet[11].x = a_p_b;
+  kernel.packet[12].x = a_p_c;
+  kernel.packet[13].x = a_p_d;
+  kernel.packet[14].x = a_p_e;
+  kernel.packet[15].x = a_p_f;
+}
+
+EIGEN_STRONG_INLINE void
+ptranspose(PacketBlock<Packet16h,8>& kernel) {
+  EIGEN_ALIGN64 half in[8][16];
+  pstore<half>(in[0], kernel.packet[0]);
+  pstore<half>(in[1], kernel.packet[1]);
+  pstore<half>(in[2], kernel.packet[2]);
+  pstore<half>(in[3], kernel.packet[3]);
+  pstore<half>(in[4], kernel.packet[4]);
+  pstore<half>(in[5], kernel.packet[5]);
+  pstore<half>(in[6], kernel.packet[6]);
+  pstore<half>(in[7], kernel.packet[7]);
+
+  EIGEN_ALIGN64 half out[8][16];
+
+  for (int i = 0; i < 8; ++i) {
+    for (int j = 0; j < 8; ++j) {
+      out[i][j] = in[j][2*i];
+    }
+    for (int j = 0; j < 8; ++j) {
+      out[i][j+8] = in[j][2*i+1];
+    }
+  }
+
+  kernel.packet[0] = pload<Packet16h>(out[0]);
+  kernel.packet[1] = pload<Packet16h>(out[1]);
+  kernel.packet[2] = pload<Packet16h>(out[2]);
+  kernel.packet[3] = pload<Packet16h>(out[3]);
+  kernel.packet[4] = pload<Packet16h>(out[4]);
+  kernel.packet[5] = pload<Packet16h>(out[5]);
+  kernel.packet[6] = pload<Packet16h>(out[6]);
+  kernel.packet[7] = pload<Packet16h>(out[7]);
+}
+
+EIGEN_STRONG_INLINE void
+ptranspose(PacketBlock<Packet16h,4>& kernel) {
+  EIGEN_ALIGN64 half in[4][16];
+  pstore<half>(in[0], kernel.packet[0]);
+  pstore<half>(in[1], kernel.packet[1]);
+  pstore<half>(in[2], kernel.packet[2]);
+  pstore<half>(in[3], kernel.packet[3]);
+
+  EIGEN_ALIGN64 half out[4][16];
+
+  for (int i = 0; i < 4; ++i) {
+    for (int j = 0; j < 4; ++j) {
+      out[i][j] = in[j][4*i];
+    }
+    for (int j = 0; j < 4; ++j) {
+      out[i][j+4] = in[j][4*i+1];
+    }
+    for (int j = 0; j < 4; ++j) {
+      out[i][j+8] = in[j][4*i+2];
+    }
+    for (int j = 0; j < 4; ++j) {
+      out[i][j+12] = in[j][4*i+3];
+    }
+  }
+
+  kernel.packet[0] = pload<Packet16h>(out[0]);
+  kernel.packet[1] = pload<Packet16h>(out[1]);
+  kernel.packet[2] = pload<Packet16h>(out[2]);
+  kernel.packet[3] = pload<Packet16h>(out[3]);
+}
+
+
 #elif defined EIGEN_VECTORIZE_AVX
 
 typedef struct {
@@ -483,6 +860,30 @@ template<> EIGEN_STRONG_INLINE void pscatter<Eigen::half, Packet8h>(Eigen::half*
   to[stride*7].x = aux[7].x;
 }
 
+template<> EIGEN_STRONG_INLINE Eigen::half predux<Packet8h>(const Packet8h& a) {
+  Packet8f af = half2float(a);
+  float reduced = predux<Packet8f>(af);
+  return Eigen::half(reduced);
+}
+
+template<> EIGEN_STRONG_INLINE Eigen::half predux_max<Packet8h>(const Packet8h& a) {
+  Packet8f af = half2float(a);
+  float reduced = predux_max<Packet8f>(af);
+  return Eigen::half(reduced);
+}
+
+template<> EIGEN_STRONG_INLINE Eigen::half predux_min<Packet8h>(const Packet8h& a) {
+  Packet8f af = half2float(a);
+  float reduced = predux_min<Packet8f>(af);
+  return Eigen::half(reduced);
+}
+
+template<> EIGEN_STRONG_INLINE Eigen::half predux_mul<Packet8h>(const Packet8h& a) {
+  Packet8f af = half2float(a);
+  float reduced = predux_mul<Packet8f>(af);
+  return Eigen::half(reduced);
+}
+
 EIGEN_STRONG_INLINE void
 ptranspose(PacketBlock<Packet8h,8>& kernel) {
   __m128i a = kernel.packet[0].x;
@@ -607,7 +1008,7 @@ template<> EIGEN_STRONG_INLINE Packet4h pset1<Packet4h>(const Eigen::half& from)
 }
 
 template<> EIGEN_STRONG_INLINE Eigen::half pfirst<Packet4h>(const Packet4h& from) {
-  return raw_uint16_to_half(static_cast<unsigned short>(_mm_cvtsi64_si32(from.x)));
+  return half_impl::raw_uint16_to_half(static_cast<unsigned short>(_mm_cvtsi64_si32(from.x)));
 }
 
 template<> EIGEN_STRONG_INLINE Packet4h pconj(const Packet4h& a) { return a; }
@@ -618,17 +1019,17 @@ template<> EIGEN_STRONG_INLINE Packet4h padd<Packet4h>(const Packet4h& a, const
 
   Eigen::half h[4];
 
-  Eigen::half ha = raw_uint16_to_half(static_cast<unsigned short>(a64));
-  Eigen::half hb = raw_uint16_to_half(static_cast<unsigned short>(b64));
+  Eigen::half ha = half_impl::raw_uint16_to_half(static_cast<unsigned short>(a64));
+  Eigen::half hb = half_impl::raw_uint16_to_half(static_cast<unsigned short>(b64));
   h[0] = ha + hb;
-  ha = raw_uint16_to_half(static_cast<unsigned short>(a64 >> 16));
-  hb = raw_uint16_to_half(static_cast<unsigned short>(b64 >> 16));
+  ha = half_impl::raw_uint16_to_half(static_cast<unsigned short>(a64 >> 16));
+  hb = half_impl::raw_uint16_to_half(static_cast<unsigned short>(b64 >> 16));
   h[1] = ha + hb;
-  ha = raw_uint16_to_half(static_cast<unsigned short>(a64 >> 32));
-  hb = raw_uint16_to_half(static_cast<unsigned short>(b64 >> 32));
+  ha = half_impl::raw_uint16_to_half(static_cast<unsigned short>(a64 >> 32));
+  hb = half_impl::raw_uint16_to_half(static_cast<unsigned short>(b64 >> 32));
   h[2] = ha + hb;
-  ha = raw_uint16_to_half(static_cast<unsigned short>(a64 >> 48));
-  hb = raw_uint16_to_half(static_cast<unsigned short>(b64 >> 48));
+  ha = half_impl::raw_uint16_to_half(static_cast<unsigned short>(a64 >> 48));
+  hb = half_impl::raw_uint16_to_half(static_cast<unsigned short>(b64 >> 48));
   h[3] = ha + hb;
   Packet4h result;
   result.x = _mm_set_pi16(h[3].x, h[2].x, h[1].x, h[0].x);
@@ -641,17 +1042,17 @@ template<> EIGEN_STRONG_INLINE Packet4h pmul<Packet4h>(const Packet4h& a, const
 
   Eigen::half h[4];
 
-  Eigen::half ha = raw_uint16_to_half(static_cast<unsigned short>(a64));
-  Eigen::half hb = raw_uint16_to_half(static_cast<unsigned short>(b64));
+  Eigen::half ha = half_impl::raw_uint16_to_half(static_cast<unsigned short>(a64));
+  Eigen::half hb = half_impl::raw_uint16_to_half(static_cast<unsigned short>(b64));
   h[0] = ha * hb;
-  ha = raw_uint16_to_half(static_cast<unsigned short>(a64 >> 16));
-  hb = raw_uint16_to_half(static_cast<unsigned short>(b64 >> 16));
+  ha = half_impl::raw_uint16_to_half(static_cast<unsigned short>(a64 >> 16));
+  hb = half_impl::raw_uint16_to_half(static_cast<unsigned short>(b64 >> 16));
   h[1] = ha * hb;
-  ha = raw_uint16_to_half(static_cast<unsigned short>(a64 >> 32));
-  hb = raw_uint16_to_half(static_cast<unsigned short>(b64 >> 32));
+  ha = half_impl::raw_uint16_to_half(static_cast<unsigned short>(a64 >> 32));
+  hb = half_impl::raw_uint16_to_half(static_cast<unsigned short>(b64 >> 32));
   h[2] = ha * hb;
-  ha = raw_uint16_to_half(static_cast<unsigned short>(a64 >> 48));
-  hb = raw_uint16_to_half(static_cast<unsigned short>(b64 >> 48));
+  ha = half_impl::raw_uint16_to_half(static_cast<unsigned short>(a64 >> 48));
+  hb = half_impl::raw_uint16_to_half(static_cast<unsigned short>(b64 >> 48));
   h[3] = ha * hb;
   Packet4h result;
   result.x = _mm_set_pi16(h[3].x, h[2].x, h[1].x, h[0].x);

Eigen/src/Core/arch/CUDA/TypeCasting.h

@@ -100,6 +100,33 @@ template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pcast<float4, half2>(cons
   return __floats2half2_rn(a.x, a.y);
 }
 
+#elif defined EIGEN_VECTORIZE_AVX512
+template <>
+struct type_casting_traits<half, float> {
+  enum {
+    VectorizedCast = 1,
+    SrcCoeffRatio = 1,
+    TgtCoeffRatio = 1
+  };
+};
+
+template<> EIGEN_STRONG_INLINE Packet16f pcast<Packet16h, Packet16f>(const Packet16h& a) {
+  return half2float(a);
+}
+
+template <>
+struct type_casting_traits<float, half> {
+  enum {
+    VectorizedCast = 1,
+    SrcCoeffRatio = 1,
+    TgtCoeffRatio = 1
+  };
+};
+
+template<> EIGEN_STRONG_INLINE Packet16h pcast<Packet16f, Packet16h>(const Packet16f& a) {
+  return float2half(a);
+}
+
 #elif defined EIGEN_VECTORIZE_AVX
 
 template <>

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

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

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

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

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

@@ -16,8 +16,14 @@ namespace Eigen {
 namespace internal {
 
 inline uint32x4_t p4ui_CONJ_XOR() {
+// See bug 1325, clang fails to call vld1q_u64.
+#if EIGEN_COMP_CLANG
+  uint32x4_t ret = { 0x00000000, 0x80000000, 0x00000000, 0x80000000 };
+  return ret;
+#else
   static const uint32_t conj_XOR_DATA[] = { 0x00000000, 0x80000000, 0x00000000, 0x80000000 };
   return vld1q_u32( conj_XOR_DATA );
+#endif
 }
 
 inline uint32x2_t p2ui_CONJ_XOR() {
@@ -282,8 +288,13 @@ ptranspose(PacketBlock<Packet2cf,2>& kernel) {
 //---------- double ----------
 #if EIGEN_ARCH_ARM64 && !EIGEN_APPLE_DOUBLE_NEON_BUG
 
-const uint64_t  p2ul_conj_XOR_DATA[] = { 0x0, 0x8000000000000000 };
-static uint64x2_t p2ul_CONJ_XOR = vld1q_u64( p2ul_conj_XOR_DATA );
+// See bug 1325, clang fails to call vld1q_u64.
+#if EIGEN_COMP_CLANG
+  static uint64x2_t p2ul_CONJ_XOR = {0x0, 0x8000000000000000};
+#else
+  const uint64_t  p2ul_conj_XOR_DATA[] = { 0x0, 0x8000000000000000 };
+  static uint64x2_t p2ul_CONJ_XOR = vld1q_u64( p2ul_conj_XOR_DATA );
+#endif
 
 struct Packet1cd
 {

Eigen/src/Core/arch/NEON/PacketMath.h

@@ -28,11 +28,13 @@ namespace internal {
 #define EIGEN_HAS_SINGLE_INSTRUCTION_CJMADD
 #endif
 
-// FIXME NEON has 16 quad registers, but since the current register allocator
-// is so bad, it is much better to reduce it to 8
 #ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS
+#if EIGEN_ARCH_ARM64
+#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS 32
+#else
 #define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS 16 
 #endif
+#endif
 
 typedef float32x2_t Packet2f;
 typedef float32x4_t Packet4f;
@@ -44,7 +46,7 @@ typedef uint32x4_t  Packet4ui;
   const Packet4f p4f_##NAME = pset1<Packet4f>(X)
 
 #define _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(NAME,X) \
-  const Packet4f p4f_##NAME = vreinterpretq_f32_u32(pset1<int>(X))
+  const Packet4f p4f_##NAME = vreinterpretq_f32_u32(pset1<int32_t>(X))
 
 #define _EIGEN_DECLARE_CONST_Packet4i(NAME,X) \
   const Packet4i p4i_##NAME = pset1<Packet4i>(X)
@@ -81,7 +83,7 @@ template<> struct packet_traits<float>  : default_packet_traits
     HasSqrt = 0
   };
 };
-template<> struct packet_traits<int>    : default_packet_traits
+template<> struct packet_traits<int32_t>    : default_packet_traits
 {
   typedef Packet4i type;
   typedef Packet4i half; // Packet2i intrinsics not implemented yet
@@ -103,11 +105,11 @@ EIGEN_STRONG_INLINE void        vst1q_f32(float* to, float32x4_t from) { ::vst1q
 EIGEN_STRONG_INLINE void        vst1_f32 (float* to, float32x2_t from) { ::vst1_f32 ((float32_t*)to,from); }
 #endif
 
-template<> struct unpacket_traits<Packet4f> { typedef float  type; enum {size=4, alignment=Aligned16}; typedef Packet4f half; };
-template<> struct unpacket_traits<Packet4i> { typedef int    type; enum {size=4, alignment=Aligned16}; typedef Packet4i half; };
+template<> struct unpacket_traits<Packet4f> { typedef float   type; enum {size=4, alignment=Aligned16}; typedef Packet4f half; };
+template<> struct unpacket_traits<Packet4i> { typedef int32_t type; enum {size=4, alignment=Aligned16}; typedef Packet4i half; };
 
 template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float&  from) { return vdupq_n_f32(from); }
-template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int&    from)   { return vdupq_n_s32(from); }
+template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int32_t&    from)   { return vdupq_n_s32(from); }
 
 template<> EIGEN_STRONG_INLINE Packet4f plset<Packet4f>(const float& a)
 {
@@ -115,7 +117,7 @@ template<> EIGEN_STRONG_INLINE Packet4f plset<Packet4f>(const float& a)
   Packet4f countdown = vld1q_f32(f);
   return vaddq_f32(pset1<Packet4f>(a), countdown);
 }
-template<> EIGEN_STRONG_INLINE Packet4i plset<Packet4i>(const int& a)
+template<> EIGEN_STRONG_INLINE Packet4i plset<Packet4i>(const int32_t& a)
 {
   const int32_t i[] = {0, 1, 2, 3};
   Packet4i countdown = vld1q_s32(i);
@@ -238,20 +240,20 @@ template<> EIGEN_STRONG_INLINE Packet4f pandnot<Packet4f>(const Packet4f& a, con
 }
 template<> EIGEN_STRONG_INLINE Packet4i pandnot<Packet4i>(const Packet4i& a, const Packet4i& b) { return vbicq_s32(a,b); }
 
-template<> EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(const float* from) { EIGEN_DEBUG_ALIGNED_LOAD return vld1q_f32(from); }
-template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int*   from) { EIGEN_DEBUG_ALIGNED_LOAD return vld1q_s32(from); }
+template<> EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(const float*    from) { EIGEN_DEBUG_ALIGNED_LOAD return vld1q_f32(from); }
+template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int32_t*  from) { EIGEN_DEBUG_ALIGNED_LOAD return vld1q_s32(from); }
 
-template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from) { EIGEN_DEBUG_UNALIGNED_LOAD return vld1q_f32(from); }
-template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from)   { EIGEN_DEBUG_UNALIGNED_LOAD return vld1q_s32(from); }
+template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float*   from) { EIGEN_DEBUG_UNALIGNED_LOAD return vld1q_f32(from); }
+template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int32_t* from) { EIGEN_DEBUG_UNALIGNED_LOAD return vld1q_s32(from); }
 
-template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float*   from)
+template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float* from)
 {
   float32x2_t lo, hi;
   lo = vld1_dup_f32(from);
   hi = vld1_dup_f32(from+1);
   return vcombine_f32(lo, hi);
 }
-template<> EIGEN_STRONG_INLINE Packet4i ploaddup<Packet4i>(const int*     from)
+template<> EIGEN_STRONG_INLINE Packet4i ploaddup<Packet4i>(const int32_t* from)
 {
   int32x2_t lo, hi;
   lo = vld1_dup_s32(from);
@@ -259,11 +261,11 @@ template<> EIGEN_STRONG_INLINE Packet4i ploaddup<Packet4i>(const int*     from)
   return vcombine_s32(lo, hi);
 }
 
-template<> EIGEN_STRONG_INLINE void pstore<float>(float*   to, const Packet4f& from) { EIGEN_DEBUG_ALIGNED_STORE vst1q_f32(to, from); }
-template<> EIGEN_STRONG_INLINE void pstore<int>(int*       to, const Packet4i& from) { EIGEN_DEBUG_ALIGNED_STORE vst1q_s32(to, from); }
+template<> EIGEN_STRONG_INLINE void pstore<float>  (float*    to, const Packet4f& from) { EIGEN_DEBUG_ALIGNED_STORE vst1q_f32(to, from); }
+template<> EIGEN_STRONG_INLINE void pstore<int32_t>(int32_t*  to, const Packet4i& from) { EIGEN_DEBUG_ALIGNED_STORE vst1q_s32(to, from); }
 
-template<> EIGEN_STRONG_INLINE void pstoreu<float>(float*  to, const Packet4f& from) { EIGEN_DEBUG_UNALIGNED_STORE vst1q_f32(to, from); }
-template<> EIGEN_STRONG_INLINE void pstoreu<int>(int*      to, const Packet4i& from) { EIGEN_DEBUG_UNALIGNED_STORE vst1q_s32(to, from); }
+template<> EIGEN_STRONG_INLINE void pstoreu<float>  (float*   to, const Packet4f& from) { EIGEN_DEBUG_UNALIGNED_STORE vst1q_f32(to, from); }
+template<> EIGEN_STRONG_INLINE void pstoreu<int32_t>(int32_t* to, const Packet4i& from) { EIGEN_DEBUG_UNALIGNED_STORE vst1q_s32(to, from); }
 
 template<> EIGEN_DEVICE_FUNC inline Packet4f pgather<float, Packet4f>(const float* from, Index stride)
 {
@@ -274,7 +276,7 @@ template<> EIGEN_DEVICE_FUNC inline Packet4f pgather<float, Packet4f>(const floa
   res = vsetq_lane_f32(from[3*stride], res, 3);
   return res;
 }
-template<> EIGEN_DEVICE_FUNC inline Packet4i pgather<int, Packet4i>(const int* from, Index stride)
+template<> EIGEN_DEVICE_FUNC inline Packet4i pgather<int32_t, Packet4i>(const int32_t* from, Index stride)
 {
   Packet4i res = pset1<Packet4i>(0);
   res = vsetq_lane_s32(from[0*stride], res, 0);
@@ -291,7 +293,7 @@ template<> EIGEN_DEVICE_FUNC inline void pscatter<float, Packet4f>(float* to, co
   to[stride*2] = vgetq_lane_f32(from, 2);
   to[stride*3] = vgetq_lane_f32(from, 3);
 }
-template<> EIGEN_DEVICE_FUNC inline void pscatter<int, Packet4i>(int* to, const Packet4i& from, Index stride)
+template<> EIGEN_DEVICE_FUNC inline void pscatter<int32_t, Packet4i>(int32_t* to, const Packet4i& from, Index stride)
 {
   to[stride*0] = vgetq_lane_s32(from, 0);
   to[stride*1] = vgetq_lane_s32(from, 1);
@@ -299,12 +301,12 @@ template<> EIGEN_DEVICE_FUNC inline void pscatter<int, Packet4i>(int* to, const
   to[stride*3] = vgetq_lane_s32(from, 3);
 }
 
-template<> EIGEN_STRONG_INLINE void prefetch<float>(const float* addr) { EIGEN_ARM_PREFETCH(addr); }
-template<> EIGEN_STRONG_INLINE void prefetch<int>(const int*     addr) { EIGEN_ARM_PREFETCH(addr); }
+template<> EIGEN_STRONG_INLINE void prefetch<float>  (const float*    addr) { EIGEN_ARM_PREFETCH(addr); }
+template<> EIGEN_STRONG_INLINE void prefetch<int32_t>(const int32_t*  addr) { EIGEN_ARM_PREFETCH(addr); }
 
 // FIXME only store the 2 first elements ?
-template<> EIGEN_STRONG_INLINE float  pfirst<Packet4f>(const Packet4f& a) { float EIGEN_ALIGN16 x[4]; vst1q_f32(x, a); return x[0]; }
-template<> EIGEN_STRONG_INLINE int    pfirst<Packet4i>(const Packet4i& a) { int   EIGEN_ALIGN16 x[4]; vst1q_s32(x, a); return x[0]; }
+template<> EIGEN_STRONG_INLINE float   pfirst<Packet4f>(const Packet4f& a) { float   EIGEN_ALIGN16 x[4]; vst1q_f32(x, a); return x[0]; }
+template<> EIGEN_STRONG_INLINE int32_t pfirst<Packet4i>(const Packet4i& a) { int32_t EIGEN_ALIGN16 x[4]; vst1q_s32(x, a); return x[0]; }
 
 template<> EIGEN_STRONG_INLINE Packet4f preverse(const Packet4f& a) {
   float32x2_t a_lo, a_hi;
@@ -359,7 +361,7 @@ template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs)
   return sum;
 }
 
-template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a)
+template<> EIGEN_STRONG_INLINE int32_t predux<Packet4i>(const Packet4i& a)
 {
   int32x2_t a_lo, a_hi, sum;
 
@@ -406,7 +408,7 @@ template<> EIGEN_STRONG_INLINE float predux_mul<Packet4f>(const Packet4f& a)
 
   return vget_lane_f32(prod, 0);
 }
-template<> EIGEN_STRONG_INLINE int predux_mul<Packet4i>(const Packet4i& a)
+template<> EIGEN_STRONG_INLINE int32_t predux_mul<Packet4i>(const Packet4i& a)
 {
   int32x2_t a_lo, a_hi, prod;
 
@@ -434,7 +436,7 @@ template<> EIGEN_STRONG_INLINE float predux_min<Packet4f>(const Packet4f& a)
   return vget_lane_f32(min, 0);
 }
 
-template<> EIGEN_STRONG_INLINE int predux_min<Packet4i>(const Packet4i& a)
+template<> EIGEN_STRONG_INLINE int32_t predux_min<Packet4i>(const Packet4i& a)
 {
   int32x2_t a_lo, a_hi, min;
 
@@ -459,7 +461,7 @@ template<> EIGEN_STRONG_INLINE float predux_max<Packet4f>(const Packet4f& a)
   return vget_lane_f32(max, 0);
 }
 
-template<> EIGEN_STRONG_INLINE int predux_max<Packet4i>(const Packet4i& a)
+template<> EIGEN_STRONG_INLINE int32_t predux_max<Packet4i>(const Packet4i& a)
 {
   int32x2_t a_lo, a_hi, max;
 

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

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

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

@@ -476,6 +476,26 @@ template<>  EIGEN_STRONG_INLINE Packet2cf pblend(const Selector<2>& ifPacket, co
   return Packet2cf(_mm_castpd_ps(result));
 }
 
+template<> EIGEN_STRONG_INLINE Packet2cf pinsertfirst(const Packet2cf& a, std::complex<float> b)
+{
+  return Packet2cf(_mm_loadl_pi(a.v, reinterpret_cast<const __m64*>(&b)));
+}
+
+template<> EIGEN_STRONG_INLINE Packet1cd pinsertfirst(const Packet1cd&, std::complex<double> b)
+{
+  return pset1<Packet1cd>(b);
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cf pinsertlast(const Packet2cf& a, std::complex<float> b)
+{
+  return Packet2cf(_mm_loadh_pi(a.v, reinterpret_cast<const __m64*>(&b)));
+}
+
+template<> EIGEN_STRONG_INLINE Packet1cd pinsertlast(const Packet1cd&, std::complex<double> b)
+{
+  return pset1<Packet1cd>(b);
+}
+
 } // end namespace internal
 
 } // end namespace Eigen

Eigen/src/Core/arch/SSE/MathFunctions.h

@@ -32,7 +32,7 @@ Packet4f plog<Packet4f>(const Packet4f& _x)
   /* the smallest non denormalized float number */
   _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(min_norm_pos,  0x00800000);
   _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(minus_inf,     0xff800000);//-1.f/0.f);
-  
+
   /* natural logarithm computed for 4 simultaneous float
     return NaN for x <= 0
   */
@@ -444,25 +444,33 @@ Packet4f pcos<Packet4f>(const Packet4f& _x)
 
 #if EIGEN_FAST_MATH
 
-// This is based on Quake3's fast inverse square root.
+// Functions for sqrt.
+// The EIGEN_FAST_MATH version uses the _mm_rsqrt_ps approximation and one step
+// of Newton's method, at a cost of 1-2 bits of precision as opposed to the
+// exact solution. It does not handle +inf, or denormalized numbers correctly.
+// The main advantage of this approach is not just speed, but also the fact that
+// it can be inlined and pipelined with other computations, further reducing its
+// effective latency. This is similar to Quake3's fast inverse square root.
 // For detail see here: http://www.beyond3d.com/content/articles/8/
-// It lacks 1 (or 2 bits in some rare cases) of precision, and does not handle negative, +inf, or denormalized numbers correctly.
 template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
 Packet4f psqrt<Packet4f>(const Packet4f& _x)
 {
   Packet4f half = pmul(_x, pset1<Packet4f>(.5f));
+  Packet4f denormal_mask = _mm_and_ps(
+      _mm_cmpge_ps(_x, _mm_setzero_ps()),
+      _mm_cmplt_ps(_x, pset1<Packet4f>((std::numeric_limits<float>::min)())));
 
-  /* select only the inverse sqrt of non-zero inputs */
-  Packet4f non_zero_mask = _mm_cmpge_ps(_x, pset1<Packet4f>((std::numeric_limits<float>::min)()));
-  Packet4f x = _mm_and_ps(non_zero_mask, _mm_rsqrt_ps(_x));
-
+  // Compute approximate reciprocal sqrt.
+  Packet4f x = _mm_rsqrt_ps(_x);
+  // Do a single step of Newton's iteration.
   x = pmul(x, psub(pset1<Packet4f>(1.5f), pmul(half, pmul(x,x))));
-  return pmul(_x,x);
+  // Flush results for denormals to zero.
+  return _mm_andnot_ps(denormal_mask, pmul(_x,x));
 }
 
 #else
 
-template<>EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED 
+template<>EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
 Packet4f psqrt<Packet4f>(const Packet4f& x) { return _mm_sqrt_ps(x); }
 
 #endif
@@ -491,7 +499,7 @@ Packet4f prsqrt<Packet4f>(const Packet4f& _x) {
   Packet4f neg_mask = _mm_cmplt_ps(_x, _mm_setzero_ps());
   Packet4f zero_mask = _mm_andnot_ps(neg_mask, le_zero_mask);
   Packet4f infs_and_nans = _mm_or_ps(_mm_and_ps(neg_mask, p4f_nan),
-                                        _mm_and_ps(zero_mask, p4f_inf));
+                                     _mm_and_ps(zero_mask, p4f_inf));
 
   // Do a single step of Newton's iteration.
   x = pmul(x, pmadd(neg_half, pmul(x, x), p4f_one_point_five));
@@ -517,52 +525,10 @@ Packet2d prsqrt<Packet2d>(const Packet2d& x) {
 }
 
 // Hyperbolic Tangent function.
-// Doesn't do anything fancy, just a 13/6-degree rational interpolant which
-// is accurate up to a couple of ulp in the range [-9, 9], outside of which the
-// fl(tanh(x)) = +/-1.
 template <>
 EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet4f
-ptanh<Packet4f>(const Packet4f& _x) {
-  // Clamp the inputs to the range [-9, 9] since anything outside
-  // this range is +/-1.0f in single-precision.
-  _EIGEN_DECLARE_CONST_Packet4f(plus_9, 9.0f);
-  _EIGEN_DECLARE_CONST_Packet4f(minus_9, -9.0f);
-  const Packet4f x = pmax(p4f_minus_9, pmin(p4f_plus_9, _x));
-
-  // The monomial coefficients of the numerator polynomial (odd).
-  _EIGEN_DECLARE_CONST_Packet4f(alpha_1, 4.89352455891786e-03f);
-  _EIGEN_DECLARE_CONST_Packet4f(alpha_3, 6.37261928875436e-04f);
-  _EIGEN_DECLARE_CONST_Packet4f(alpha_5, 1.48572235717979e-05f);
-  _EIGEN_DECLARE_CONST_Packet4f(alpha_7, 5.12229709037114e-08f);
-  _EIGEN_DECLARE_CONST_Packet4f(alpha_9, -8.60467152213735e-11f);
-  _EIGEN_DECLARE_CONST_Packet4f(alpha_11, 2.00018790482477e-13f);
-  _EIGEN_DECLARE_CONST_Packet4f(alpha_13, -2.76076847742355e-16f);
-
-  // The monomial coefficients of the denominator polynomial (even).
-  _EIGEN_DECLARE_CONST_Packet4f(beta_0, 4.89352518554385e-03f);
-  _EIGEN_DECLARE_CONST_Packet4f(beta_2, 2.26843463243900e-03f);
-  _EIGEN_DECLARE_CONST_Packet4f(beta_4, 1.18534705686654e-04f);
-  _EIGEN_DECLARE_CONST_Packet4f(beta_6, 1.19825839466702e-06f);
-
-  // Since the polynomials are odd/even, we need x^2.
-  const Packet4f x2 = pmul(x, x);
-
-  // Evaluate the numerator polynomial p.
-  Packet4f p = pmadd(x2, p4f_alpha_13, p4f_alpha_11);
-  p = pmadd(x2, p, p4f_alpha_9);
-  p = pmadd(x2, p, p4f_alpha_7);
-  p = pmadd(x2, p, p4f_alpha_5);
-  p = pmadd(x2, p, p4f_alpha_3);
-  p = pmadd(x2, p, p4f_alpha_1);
-  p = pmul(x, p);
-
-  // Evaluate the denominator polynomial p.
-  Packet4f q = pmadd(x2, p4f_beta_6, p4f_beta_4);
-  q = pmadd(x2, q, p4f_beta_2);
-  q = pmadd(x2, q, p4f_beta_0);
-
-  // Divide the numerator by the denominator.
-  return pdiv(p, q);
+ptanh<Packet4f>(const Packet4f& x) {
+  return internal::generic_fast_tanh_float(x);
 }
 
 } // end namespace internal

Eigen/src/Core/arch/SSE/PacketMath.h

@@ -28,7 +28,7 @@ namespace internal {
 #endif
 #endif
 
-#if (defined EIGEN_VECTORIZE_AVX) && EIGEN_COMP_GNUC_STRICT && (__GXX_ABI_VERSION < 1004)
+#if (defined EIGEN_VECTORIZE_AVX) && (EIGEN_COMP_GNUC_STRICT || EIGEN_COMP_MINGW) && (__GXX_ABI_VERSION < 1004)
 // With GCC's default ABI version, a __m128 or __m256 are the same types and therefore we cannot
 // have overloads for both types without linking error.
 // One solution is to increase ABI version using -fabi-version=4 (or greater).
@@ -162,6 +162,11 @@ template<> struct unpacket_traits<Packet4f> { typedef float  type; enum {size=4,
 template<> struct unpacket_traits<Packet2d> { typedef double type; enum {size=2, alignment=Aligned16}; typedef Packet2d half; };
 template<> struct unpacket_traits<Packet4i> { typedef int    type; enum {size=4, alignment=Aligned16}; typedef Packet4i half; };
 
+#ifndef EIGEN_VECTORIZE_AVX
+template<> struct scalar_div_cost<float,true> { enum { value = 7 }; };
+template<> struct scalar_div_cost<double,true> { enum { value = 8 }; };
+#endif
+
 #if EIGEN_COMP_MSVC==1500
 // Workaround MSVC 9 internal compiler error.
 // TODO: It has been detected with win64 builds (amd64), so let's check whether it also happens in 32bits+SSE mode
@@ -499,30 +504,13 @@ template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs)
 {
   return _mm_hadd_ps(_mm_hadd_ps(vecs[0], vecs[1]),_mm_hadd_ps(vecs[2], vecs[3]));
 }
+
 template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs)
 {
   return _mm_hadd_pd(vecs[0], vecs[1]);
 }
 
-template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a)
-{
-  Packet4f tmp0 = _mm_hadd_ps(a,a);
-  return pfirst<Packet4f>(_mm_hadd_ps(tmp0, tmp0));
-}
-
-template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a) { return pfirst<Packet2d>(_mm_hadd_pd(a, a)); }
 #else
-// SSE2 versions
-template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a)
-{
-  Packet4f tmp = _mm_add_ps(a, _mm_movehl_ps(a,a));
-  return pfirst<Packet4f>(_mm_add_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1)));
-}
-template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a)
-{
-  return pfirst<Packet2d>(_mm_add_sd(a, _mm_unpackhi_pd(a,a)));
-}
-
 template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs)
 {
   Packet4f tmp0, tmp1, tmp2;
@@ -543,6 +531,29 @@ template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs)
 }
 #endif  // SSE3
 
+template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a)
+{
+  // Disable SSE3 _mm_hadd_pd that is extremely slow on all existing Intel's architectures
+  // (from Nehalem to Haswell)
+// #ifdef EIGEN_VECTORIZE_SSE3
+//   Packet4f tmp = _mm_add_ps(a, vec4f_swizzle1(a,2,3,2,3));
+//   return pfirst<Packet4f>(_mm_hadd_ps(tmp, tmp));
+// #else
+  Packet4f tmp = _mm_add_ps(a, _mm_movehl_ps(a,a));
+  return pfirst<Packet4f>(_mm_add_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1)));
+// #endif
+}
+
+template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a)
+{
+  // Disable SSE3 _mm_hadd_pd that is extremely slow on all existing Intel's architectures
+  // (from Nehalem to Haswell)
+// #ifdef EIGEN_VECTORIZE_SSE3
+//   return pfirst<Packet2d>(_mm_hadd_pd(a, a));
+// #else
+  return pfirst<Packet2d>(_mm_add_sd(a, _mm_unpackhi_pd(a,a)));
+// #endif
+}
 
 #ifdef EIGEN_VECTORIZE_SSSE3
 template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs)
@@ -813,6 +824,54 @@ template<> EIGEN_STRONG_INLINE Packet2d pblend(const Selector<2>& ifPacket, cons
 #endif
 }
 
+template<> EIGEN_STRONG_INLINE Packet4f pinsertfirst(const Packet4f& a, float b)
+{
+#ifdef EIGEN_VECTORIZE_SSE4_1
+  return _mm_blend_ps(a,pset1<Packet4f>(b),1);
+#else
+  return _mm_move_ss(a, _mm_load_ss(&b));
+#endif
+}
+
+template<> EIGEN_STRONG_INLINE Packet2d pinsertfirst(const Packet2d& a, double b)
+{
+#ifdef EIGEN_VECTORIZE_SSE4_1
+  return _mm_blend_pd(a,pset1<Packet2d>(b),1);
+#else
+  return _mm_move_sd(a, _mm_load_sd(&b));
+#endif
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f pinsertlast(const Packet4f& a, float b)
+{
+#ifdef EIGEN_VECTORIZE_SSE4_1
+  return _mm_blend_ps(a,pset1<Packet4f>(b),(1<<3));
+#else
+  const Packet4f mask = _mm_castsi128_ps(_mm_setr_epi32(0x0,0x0,0x0,0xFFFFFFFF));
+  return _mm_or_ps(_mm_andnot_ps(mask, a), _mm_and_ps(mask, pset1<Packet4f>(b)));
+#endif
+}
+
+template<> EIGEN_STRONG_INLINE Packet2d pinsertlast(const Packet2d& a, double b)
+{
+#ifdef EIGEN_VECTORIZE_SSE4_1
+  return _mm_blend_pd(a,pset1<Packet2d>(b),(1<<1));
+#else
+  const Packet2d mask = _mm_castsi128_pd(_mm_setr_epi32(0x0,0x0,0xFFFFFFFF,0xFFFFFFFF));
+  return _mm_or_pd(_mm_andnot_pd(mask, a), _mm_and_pd(mask, pset1<Packet2d>(b)));
+#endif
+}
+
+// Scalar path for pmadd with FMA to ensure consistency with vectorized path.
+#ifdef __FMA__
+template<> EIGEN_STRONG_INLINE float pmadd(const float& a, const float& b, const float& c) {
+  return ::fmaf(a,b,c);
+}
+template<> EIGEN_STRONG_INLINE double pmadd(const double& a, const double& b, const double& c) {
+  return ::fma(a,b,c);
+}
+#endif
+
 } // end namespace internal
 
 } // end namespace Eigen

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

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

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

@@ -2,6 +2,7 @@
 // for linear algebra.
 //
 // Copyright (C) 2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2016 Konstantinos Margaritis <markos@freevec.org>
 //
 // 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
@@ -24,13 +25,48 @@ struct Packet1cd
   Packet2d v;
 };
 
+struct Packet2cf
+{
+  EIGEN_STRONG_INLINE Packet2cf() {}
+  EIGEN_STRONG_INLINE explicit Packet2cf(const Packet4f& a) : v(a) {}
+  union {
+    Packet4f v;
+    Packet1cd cd[2];
+  };
+};
+
+template<> struct packet_traits<std::complex<float> >  : default_packet_traits
+{
+  typedef Packet2cf type;
+  typedef Packet2cf half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size = 2,
+    HasHalfPacket = 0,
+
+    HasAdd    = 1,
+    HasSub    = 1,
+    HasMul    = 1,
+    HasDiv    = 1,
+    HasNegate = 1,
+    HasAbs    = 0,
+    HasAbs2   = 0,
+    HasMin    = 0,
+    HasMax    = 0,
+    HasBlend  = 1,
+    HasSetLinear = 0
+  };
+};
+
+
 template<> struct packet_traits<std::complex<double> >  : default_packet_traits
 {
   typedef Packet1cd type;
   typedef Packet1cd half;
   enum {
     Vectorizable = 1,
-    AlignedOnScalar = 0,
+    AlignedOnScalar = 1,
     size = 1,
     HasHalfPacket = 0,
 
@@ -47,20 +83,68 @@ template<> struct packet_traits<std::complex<double> >  : default_packet_traits
   };
 };
 
+template<> struct unpacket_traits<Packet2cf> { typedef std::complex<float>  type; enum {size=2, alignment=Aligned16}; typedef Packet2cf half; };
 template<> struct unpacket_traits<Packet1cd> { typedef std::complex<double> type; enum {size=1, alignment=Aligned16}; typedef Packet1cd half; };
 
+/* Forward declaration */
+EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet2cf,2>& kernel);
+
+template<> EIGEN_STRONG_INLINE Packet2cf pload <Packet2cf>(const std::complex<float>* from)  { EIGEN_DEBUG_ALIGNED_LOAD return Packet2cf(pload<Packet4f>((const float*)from)); }
 template<> EIGEN_STRONG_INLINE Packet1cd pload <Packet1cd>(const std::complex<double>* from) { EIGEN_DEBUG_ALIGNED_LOAD return Packet1cd(pload<Packet2d>((const double*)from)); }
+template<> EIGEN_STRONG_INLINE Packet2cf ploadu<Packet2cf>(const std::complex<float>* from)  { EIGEN_DEBUG_UNALIGNED_LOAD return Packet2cf(ploadu<Packet4f>((const float*)from)); }
 template<> EIGEN_STRONG_INLINE Packet1cd ploadu<Packet1cd>(const std::complex<double>* from) { EIGEN_DEBUG_UNALIGNED_LOAD return Packet1cd(ploadu<Packet2d>((const double*)from)); }
+template<> EIGEN_STRONG_INLINE void pstore <std::complex<float> >(std::complex<float> *     to, const Packet2cf& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((float*)to, from.v); }
 template<> EIGEN_STRONG_INLINE void pstore <std::complex<double> >(std::complex<double> *   to, const Packet1cd& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((double*)to, from.v); }
+template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<float> >(std::complex<float> *     to, const Packet2cf& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((float*)to, from.v); }
 template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<double> >(std::complex<double> *   to, const Packet1cd& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((double*)to, from.v); }
 
 template<> EIGEN_STRONG_INLINE Packet1cd pset1<Packet1cd>(const std::complex<double>&  from)
 { /* here we really have to use unaligned loads :( */ return ploadu<Packet1cd>(&from); }
 
+template<> EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(const std::complex<float>&  from)
+{
+  Packet2cf res;
+  res.cd[0] = Packet1cd(vec_ld2f((const float *)&from));
+  res.cd[1] = res.cd[0];
+  return res;
+}
+template<> EIGEN_DEVICE_FUNC inline Packet2cf pgather<std::complex<float>, Packet2cf>(const std::complex<float>* from, Index stride)
+{
+  std::complex<float> EIGEN_ALIGN16 af[2];
+  af[0] = from[0*stride];
+  af[1] = from[1*stride];
+  return pload<Packet2cf>(af);
+}
+template<> EIGEN_DEVICE_FUNC inline Packet1cd pgather<std::complex<double>, Packet1cd>(const std::complex<double>* from, Index stride EIGEN_UNUSED)
+{
+  return pload<Packet1cd>(from);
+}
+template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<float>, Packet2cf>(std::complex<float>* to, const Packet2cf& from, Index stride)
+{
+  std::complex<float> EIGEN_ALIGN16 af[2];
+  pstore<std::complex<float> >((std::complex<float> *) af, from);
+  to[0*stride] = af[0];
+  to[1*stride] = af[1];
+}
+template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<double>, Packet1cd>(std::complex<double>* to, const Packet1cd& from, Index stride EIGEN_UNUSED)
+{
+  pstore<std::complex<double> >(to, from);
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cf padd<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(padd<Packet4f>(a.v, b.v)); }
 template<> EIGEN_STRONG_INLINE Packet1cd padd<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(a.v + b.v); }
+template<> EIGEN_STRONG_INLINE Packet2cf psub<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(psub<Packet4f>(a.v, b.v)); }
 template<> EIGEN_STRONG_INLINE Packet1cd psub<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(a.v - b.v); }
 template<> EIGEN_STRONG_INLINE Packet1cd pnegate(const Packet1cd& a) { return Packet1cd(pnegate(Packet2d(a.v))); }
+template<> EIGEN_STRONG_INLINE Packet2cf pnegate(const Packet2cf& a) { return Packet2cf(pnegate(Packet4f(a.v))); }
 template<> EIGEN_STRONG_INLINE Packet1cd pconj(const Packet1cd& a) { return Packet1cd((Packet2d)vec_xor((Packet2d)a.v, (Packet2d)p2ul_CONJ_XOR2)); }
+template<> EIGEN_STRONG_INLINE Packet2cf pconj(const Packet2cf& a)
+{
+  Packet2cf res;
+  res.v.v4f[0] = pconj(Packet1cd(reinterpret_cast<Packet2d>(a.v.v4f[0]))).v;
+  res.v.v4f[1] = pconj(Packet1cd(reinterpret_cast<Packet2d>(a.v.v4f[1]))).v;
+  return res;
+}
 
 template<> EIGEN_STRONG_INLINE Packet1cd pmul<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
 {
@@ -79,43 +163,90 @@ template<> EIGEN_STRONG_INLINE Packet1cd pmul<Packet1cd>(const Packet1cd& a, con
 
   return Packet1cd(v1 + v2);
 }
-
-template<> EIGEN_STRONG_INLINE Packet1cd pand   <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_and(a.v,b.v)); }
-template<> EIGEN_STRONG_INLINE Packet1cd por    <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_or(a.v,b.v)); }
-template<> EIGEN_STRONG_INLINE Packet1cd pxor   <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_xor(a.v,b.v)); }
-template<> EIGEN_STRONG_INLINE Packet1cd pandnot<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_and(a.v, vec_nor(b.v,b.v))); }
-
-template<> EIGEN_STRONG_INLINE Packet1cd ploaddup<Packet1cd>(const std::complex<double>*     from)
+template<> EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
 {
-  return pset1<Packet1cd>(*from);
+  Packet2cf res;
+  res.v.v4f[0] = pmul(Packet1cd(reinterpret_cast<Packet2d>(a.v.v4f[0])), Packet1cd(reinterpret_cast<Packet2d>(b.v.v4f[0]))).v;
+  res.v.v4f[1] = pmul(Packet1cd(reinterpret_cast<Packet2d>(a.v.v4f[1])), Packet1cd(reinterpret_cast<Packet2d>(b.v.v4f[1]))).v;
+  return res;
 }
 
+template<> EIGEN_STRONG_INLINE Packet1cd pand   <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_and(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cf pand   <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(pand<Packet4f>(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet1cd por    <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_or(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cf por    <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(por<Packet4f>(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet1cd pxor   <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_xor(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cf pxor   <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(pxor<Packet4f>(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet1cd pandnot<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_and(a.v, vec_nor(b.v,b.v))); }
+template<> EIGEN_STRONG_INLINE Packet2cf pandnot<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(pandnot<Packet4f>(a.v,b.v)); }
+
+template<> EIGEN_STRONG_INLINE Packet1cd ploaddup<Packet1cd>(const std::complex<double>*     from) {  return pset1<Packet1cd>(*from); }
+template<> EIGEN_STRONG_INLINE Packet2cf ploaddup<Packet2cf>(const std::complex<float>*      from) {  return pset1<Packet2cf>(*from); }
+
+template<> EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::complex<float> *     addr) { EIGEN_ZVECTOR_PREFETCH(addr); }
 template<> EIGEN_STRONG_INLINE void prefetch<std::complex<double> >(const std::complex<double> *   addr) { EIGEN_ZVECTOR_PREFETCH(addr); }
 
 template<> EIGEN_STRONG_INLINE std::complex<double>  pfirst<Packet1cd>(const Packet1cd& a)
 {
-  std::complex<double> EIGEN_ALIGN16 res[2];
-  pstore<std::complex<double> >(res, a);
+  std::complex<double> EIGEN_ALIGN16 res;
+  pstore<std::complex<double> >(&res, a);
+
+  return res;
+}
+template<> EIGEN_STRONG_INLINE std::complex<float>  pfirst<Packet2cf>(const Packet2cf& a)
+{
+  std::complex<float> EIGEN_ALIGN16 res[2];
+  pstore<std::complex<float> >(res, a);
 
   return res[0];
 }
 
 template<> EIGEN_STRONG_INLINE Packet1cd preverse(const Packet1cd& a) { return a; }
+template<> EIGEN_STRONG_INLINE Packet2cf preverse(const Packet2cf& a)
+{
+  Packet2cf res;
+  res.cd[0] = a.cd[1];
+  res.cd[1] = a.cd[0];
+  return res;
+}
 
 template<> EIGEN_STRONG_INLINE std::complex<double> predux<Packet1cd>(const Packet1cd& a)
 {
   return pfirst(a);
 }
+template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(const Packet2cf& a)
+{
+  std::complex<float> res;
+  Packet1cd b = padd<Packet1cd>(a.cd[0], a.cd[1]);
+  vec_st2f(b.v, (float*)&res);
+  return res;
+}
 
 template<> EIGEN_STRONG_INLINE Packet1cd preduxp<Packet1cd>(const Packet1cd* vecs)
 {
   return vecs[0];
 }
+template<> EIGEN_STRONG_INLINE Packet2cf preduxp<Packet2cf>(const Packet2cf* vecs)
+{
+  PacketBlock<Packet2cf,2> transpose;
+  transpose.packet[0] = vecs[0];
+  transpose.packet[1] = vecs[1];
+  ptranspose(transpose);
+
+  return padd<Packet2cf>(transpose.packet[0], transpose.packet[1]);
+} 
 
 template<> EIGEN_STRONG_INLINE std::complex<double> predux_mul<Packet1cd>(const Packet1cd& a)
 {
   return pfirst(a);
 }
+template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(const Packet2cf& a)
+{
+  std::complex<float> res;
+  Packet1cd b = pmul<Packet1cd>(a.cd[0], a.cd[1]);
+  vec_st2f(b.v, (float*)&res);
+  return res;
+}
 
 template<int Offset>
 struct palign_impl<Offset,Packet1cd>
@@ -127,6 +258,18 @@ struct palign_impl<Offset,Packet1cd>
   }
 };
 
+template<int Offset>
+struct palign_impl<Offset,Packet2cf>
+{
+  static EIGEN_STRONG_INLINE void run(Packet2cf& first, const Packet2cf& second)
+  {
+    if (Offset == 1) {
+      first.cd[0] = first.cd[1];
+      first.cd[1] = second.cd[0];
+    }
+  }
+};
+
 template<> struct conj_helper<Packet1cd, Packet1cd, false,true>
 {
   EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const
@@ -160,6 +303,39 @@ template<> struct conj_helper<Packet1cd, Packet1cd, true,true>
   }
 };
 
+template<> struct conj_helper<Packet2cf, Packet2cf, false,true>
+{
+  EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const
+  {
+    return internal::pmul(a, pconj(b));
+  }
+};
+
+template<> struct conj_helper<Packet2cf, Packet2cf, true,false>
+{
+  EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const
+  {
+    return internal::pmul(pconj(a), b);
+  }
+};
+
+template<> struct conj_helper<Packet2cf, Packet2cf, true,true>
+{
+  EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const
+  {
+    return pconj(internal::pmul(a, b));
+  }
+};
+
 template<> EIGEN_STRONG_INLINE Packet1cd pdiv<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
 {
   // TODO optimize it for AltiVec
@@ -168,17 +344,49 @@ template<> EIGEN_STRONG_INLINE Packet1cd pdiv<Packet1cd>(const Packet1cd& a, con
   return Packet1cd(pdiv(res.v, s + vec_perm(s, s, p16uc_REVERSE64)));
 }
 
+template<> EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
+{
+  // TODO optimize it for AltiVec
+  Packet2cf res;
+  res.cd[0] = pdiv<Packet1cd>(a.cd[0], b.cd[0]);
+  res.cd[1] = pdiv<Packet1cd>(a.cd[1], b.cd[1]);
+  return res;
+}
+
 EIGEN_STRONG_INLINE Packet1cd pcplxflip/*<Packet1cd>*/(const Packet1cd& x)
 {
   return Packet1cd(preverse(Packet2d(x.v)));
 }
 
+EIGEN_STRONG_INLINE Packet2cf pcplxflip/*<Packet2cf>*/(const Packet2cf& x)
+{
+  Packet2cf res;
+  res.cd[0] = pcplxflip(x.cd[0]);
+  res.cd[1] = pcplxflip(x.cd[1]);
+  return res;
+}
+
 EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet1cd,2>& kernel)
 {
   Packet2d tmp = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_HI);
   kernel.packet[1].v = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_LO);
   kernel.packet[0].v = tmp;
 }
+
+EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet2cf,2>& kernel)
+{
+  Packet1cd tmp = kernel.packet[0].cd[1];
+  kernel.packet[0].cd[1] = kernel.packet[1].cd[0];
+  kernel.packet[1].cd[0] = tmp;
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cf pblend(const Selector<2>& ifPacket, const Packet2cf& thenPacket, const Packet2cf& elsePacket) {
+  Packet2cf result;
+  const Selector<4> ifPacket4 = { ifPacket.select[0], ifPacket.select[0], ifPacket.select[1], ifPacket.select[1] };
+  result.v = pblend<Packet4f>(ifPacket4, thenPacket.v, elsePacket.v);
+  return result;
+}
+
 } // end namespace internal
 
 } // end namespace Eigen

Eigen/src/Core/arch/ZVector/MathFunctions.h

@@ -3,6 +3,7 @@
 //
 // Copyright (C) 2007 Julien Pommier
 // Copyright (C) 2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2016 Konstantinos Margaritis <markos@freevec.org>
 //
 // 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
@@ -19,32 +20,32 @@ namespace Eigen {
 
 namespace internal {
 
+static _EIGEN_DECLARE_CONST_Packet2d(1 , 1.0);
+static _EIGEN_DECLARE_CONST_Packet2d(2 , 2.0);
+static _EIGEN_DECLARE_CONST_Packet2d(half, 0.5);
+
+static _EIGEN_DECLARE_CONST_Packet2d(exp_hi,  709.437);
+static _EIGEN_DECLARE_CONST_Packet2d(exp_lo, -709.436139303);
+
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_LOG2EF, 1.4426950408889634073599);
+
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p0, 1.26177193074810590878e-4);
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p1, 3.02994407707441961300e-2);
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p2, 9.99999999999999999910e-1);
+
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q0, 3.00198505138664455042e-6);
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q1, 2.52448340349684104192e-3);
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q2, 2.27265548208155028766e-1);
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q3, 2.00000000000000000009e0);
+
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C1, 0.693145751953125);
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C2, 1.42860682030941723212e-6);
+
 template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
 Packet2d pexp<Packet2d>(const Packet2d& _x)
 {
   Packet2d x = _x;
 
-  _EIGEN_DECLARE_CONST_Packet2d(1 , 1.0);
-  _EIGEN_DECLARE_CONST_Packet2d(2 , 2.0);
-  _EIGEN_DECLARE_CONST_Packet2d(half, 0.5);
-
-  _EIGEN_DECLARE_CONST_Packet2d(exp_hi,  709.437);
-  _EIGEN_DECLARE_CONST_Packet2d(exp_lo, -709.436139303);
-
-  _EIGEN_DECLARE_CONST_Packet2d(cephes_LOG2EF, 1.4426950408889634073599);
-
-  _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p0, 1.26177193074810590878e-4);
-  _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p1, 3.02994407707441961300e-2);
-  _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p2, 9.99999999999999999910e-1);
-
-  _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q0, 3.00198505138664455042e-6);
-  _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q1, 2.52448340349684104192e-3);
-  _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q2, 2.27265548208155028766e-1);
-  _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q3, 2.00000000000000000009e0);
-
-  _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C1, 0.693145751953125);
-  _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C2, 1.42860682030941723212e-6);
-
   Packet2d tmp, fx;
   Packet2l emm0;
 
@@ -91,18 +92,44 @@ Packet2d pexp<Packet2d>(const Packet2d& _x)
                  isnumber_mask);
 }
 
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet4f pexp<Packet4f>(const Packet4f& x)
+{
+  Packet4f res;
+  res.v4f[0] = pexp<Packet2d>(x.v4f[0]);
+  res.v4f[1] = pexp<Packet2d>(x.v4f[1]);
+  return res;
+}
+
 template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
 Packet2d psqrt<Packet2d>(const Packet2d& x)
 {
   return  __builtin_s390_vfsqdb(x);
 }
 
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet4f psqrt<Packet4f>(const Packet4f& x)
+{
+  Packet4f res;
+  res.v4f[0] = psqrt<Packet2d>(x.v4f[0]);
+  res.v4f[1] = psqrt<Packet2d>(x.v4f[1]);
+  return res;
+}
+
 template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
 Packet2d prsqrt<Packet2d>(const Packet2d& x) {
   // Unfortunately we can't use the much faster mm_rqsrt_pd since it only provides an approximation.
   return pset1<Packet2d>(1.0) / psqrt<Packet2d>(x);
 }
 
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet4f prsqrt<Packet4f>(const Packet4f& x) {
+  Packet4f res;
+  res.v4f[0] = prsqrt<Packet2d>(x.v4f[0]);
+  res.v4f[1] = prsqrt<Packet2d>(x.v4f[1]);
+  return res;
+}
+
 }  // end namespace internal
 
 }  // end namespace Eigen

Eigen/src/Core/arch/ZVector/PacketMath.h

@@ -28,9 +28,8 @@ namespace internal {
 #define EIGEN_HAS_SINGLE_INSTRUCTION_CJMADD
 #endif
 
-// NOTE Altivec has 32 registers, but Eigen only accepts a value of 8 or 16
 #ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS
-#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS  32
+#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS  16
 #endif
 
 typedef __vector int                 Packet4i;
@@ -42,6 +41,10 @@ typedef __vector double              Packet2d;
 typedef __vector unsigned long long  Packet2ul;
 typedef __vector long long           Packet2l;
 
+typedef struct {
+	Packet2d  v4f[2];
+} Packet4f;
+
 typedef union {
   int32_t   i[4];
   uint32_t ui[4];
@@ -88,6 +91,7 @@ static Packet2d p2d_ONE = { 1.0, 1.0 };
 static Packet2d p2d_ZERO_ = { -0.0, -0.0 };
 
 static Packet4i p4i_COUNTDOWN = { 0, 1, 2, 3 };
+static Packet4f p4f_COUNTDOWN = { 0.0, 1.0, 2.0, 3.0 };
 static Packet2d p2d_COUNTDOWN = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet16uc>(p2d_ZERO), reinterpret_cast<Packet16uc>(p2d_ONE), 8));
 
 static Packet16uc p16uc_PSET64_HI = { 0,1,2,3, 4,5,6,7, 0,1,2,3, 4,5,6,7 };
@@ -96,7 +100,7 @@ static Packet16uc p16uc_DUPLICATE32_HI = { 0,1,2,3, 0,1,2,3, 4,5,6,7, 4,5,6,7 };
 // Mask alignment
 #define _EIGEN_MASK_ALIGNMENT	0xfffffffffffffff0
 
-#define _EIGEN_ALIGNED_PTR(x)	((ptrdiff_t)(x) & _EIGEN_MASK_ALIGNMENT)
+#define _EIGEN_ALIGNED_PTR(x)	((std::ptrdiff_t)(x) & _EIGEN_MASK_ALIGNMENT)
 
 // Handle endianness properly while loading constants
 // Define global static constants:
@@ -132,13 +136,11 @@ template<> struct packet_traits<int>    : default_packet_traits
   typedef Packet4i type;
   typedef Packet4i half;
   enum {
-    // FIXME check the Has*
     Vectorizable = 1,
     AlignedOnScalar = 1,
     size = 4,
     HasHalfPacket = 0,
 
-    // FIXME check the Has*
     HasAdd  = 1,
     HasSub  = 1,
     HasMul  = 1,
@@ -147,6 +149,37 @@ template<> struct packet_traits<int>    : default_packet_traits
   };
 };
 
+template<> struct packet_traits<float> : default_packet_traits
+{
+  typedef Packet4f type;
+  typedef Packet4f half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size=4,
+    HasHalfPacket = 0,
+
+    HasAdd  = 1,
+    HasSub  = 1,
+    HasMul  = 1,
+    HasDiv  = 1,
+    HasMin  = 1,
+    HasMax  = 1,
+    HasAbs  = 1,
+    HasSin  = 0,
+    HasCos  = 0,
+    HasLog  = 0,
+    HasExp  = 1,
+    HasSqrt = 1,
+    HasRsqrt = 1,
+    HasRound = 1,
+    HasFloor = 1,
+    HasCeil = 1,
+    HasNegate = 1,
+    HasBlend = 1
+  };
+};
+
 template<> struct packet_traits<double> : default_packet_traits
 {
   typedef Packet2d type;
@@ -157,7 +190,6 @@ template<> struct packet_traits<double> : default_packet_traits
     size=2,
     HasHalfPacket = 1,
 
-    // FIXME check the Has*
     HasAdd  = 1,
     HasSub  = 1,
     HasMul  = 1,
@@ -180,8 +212,12 @@ template<> struct packet_traits<double> : default_packet_traits
 };
 
 template<> struct unpacket_traits<Packet4i> { typedef int    type; enum {size=4, alignment=Aligned16}; typedef Packet4i half; };
+template<> struct unpacket_traits<Packet4f> { typedef float  type; enum {size=4, alignment=Aligned16}; typedef Packet4f half; };
 template<> struct unpacket_traits<Packet2d> { typedef double type; enum {size=2, alignment=Aligned16}; typedef Packet2d half; };
 
+/* Forward declaration */
+EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet4f,4>& kernel);
+ 
 inline std::ostream & operator <<(std::ostream & s, const Packet4i & v)
 {
   Packet vt;
@@ -222,6 +258,32 @@ inline std::ostream & operator <<(std::ostream & s, const Packet2d & v)
   return s;
 }
 
+/* Helper function to simulate a vec_splat_packet4f
+ */
+template<int element> EIGEN_STRONG_INLINE Packet4f vec_splat_packet4f(const Packet4f&   from)
+{
+  Packet4f splat;
+  switch (element) {
+  case 0:
+    splat.v4f[0] = vec_splat(from.v4f[0], 0);
+    splat.v4f[1] = splat.v4f[0];
+    break;
+  case 1:
+    splat.v4f[0] = vec_splat(from.v4f[0], 1);
+    splat.v4f[1] = splat.v4f[0];
+    break;
+  case 2:
+    splat.v4f[0] = vec_splat(from.v4f[1], 0);
+    splat.v4f[1] = splat.v4f[0];
+    break;
+  case 3:
+    splat.v4f[0] = vec_splat(from.v4f[1], 1);
+    splat.v4f[1] = splat.v4f[0];
+    break;
+  }
+  return splat;
+}
+
 template<int Offset>
 struct palign_impl<Offset,Packet4i>
 {
@@ -238,6 +300,31 @@ struct palign_impl<Offset,Packet4i>
   }
 };
 
+/* This is a tricky one, we have to translate float alignment to vector elements of sizeof double
+ */
+template<int Offset>
+struct palign_impl<Offset,Packet4f>
+{
+  static EIGEN_STRONG_INLINE void run(Packet4f& first, const Packet4f& second)
+  {
+    switch (Offset % 4) {
+    case 1:
+      first.v4f[0] = vec_sld(first.v4f[0], first.v4f[1], 8);
+      first.v4f[1] = vec_sld(first.v4f[1], second.v4f[0], 8);
+      break;
+    case 2:
+      first.v4f[0] = first.v4f[1];
+      first.v4f[1] = second.v4f[0];
+      break;
+    case 3:
+      first.v4f[0] = vec_sld(first.v4f[1],  second.v4f[0], 8);
+      first.v4f[1] = vec_sld(second.v4f[0], second.v4f[1], 8);
+      break;
+    }
+  }
+};
+
+
 template<int Offset>
 struct palign_impl<Offset,Packet2d>
 {
@@ -257,6 +344,16 @@ template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int*     from)
   return vfrom->v4i;
 }
 
+template<> EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(const float*   from)
+{
+  // FIXME: No intrinsic yet
+  EIGEN_DEBUG_ALIGNED_LOAD
+  Packet4f vfrom;
+  vfrom.v4f[0] = vec_ld2f(&from[0]);
+  vfrom.v4f[1] = vec_ld2f(&from[2]);
+  return vfrom;
+}
+
 template<> EIGEN_STRONG_INLINE Packet2d pload<Packet2d>(const double* from)
 {
   // FIXME: No intrinsic yet
@@ -275,6 +372,15 @@ template<> EIGEN_STRONG_INLINE void pstore<int>(int*       to, const Packet4i& f
   vto->v4i = from;
 }
 
+template<> EIGEN_STRONG_INLINE void pstore<float>(float*   to, const Packet4f& from)
+{
+  // FIXME: No intrinsic yet
+  EIGEN_DEBUG_ALIGNED_STORE
+  vec_st2f(from.v4f[0], &to[0]);
+  vec_st2f(from.v4f[1], &to[2]);
+}
+
+
 template<> EIGEN_STRONG_INLINE void pstore<double>(double*   to, const Packet2d& from)
 {
   // FIXME: No intrinsic yet
@@ -288,10 +394,16 @@ template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int&    from)
 {
   return vec_splats(from);
 }
-
 template<> EIGEN_STRONG_INLINE Packet2d pset1<Packet2d>(const double& from) {
   return vec_splats(from);
 }
+template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float&    from)
+{
+  Packet4f to;
+  to.v4f[0] = pset1<Packet2d>(static_cast<const double&>(from));
+  to.v4f[1] = to.v4f[0];
+  return to;
+}
 
 template<> EIGEN_STRONG_INLINE void
 pbroadcast4<Packet4i>(const int *a,
@@ -304,6 +416,17 @@ pbroadcast4<Packet4i>(const int *a,
   a3 = vec_splat(a3, 3);
 }
 
+template<> EIGEN_STRONG_INLINE void
+pbroadcast4<Packet4f>(const float *a,
+                      Packet4f& a0, Packet4f& a1, Packet4f& a2, Packet4f& a3)
+{
+  a3 = pload<Packet4f>(a);
+  a0 = vec_splat_packet4f<0>(a3);
+  a1 = vec_splat_packet4f<1>(a3);
+  a2 = vec_splat_packet4f<2>(a3);
+  a3 = vec_splat_packet4f<3>(a3);
+}
+
 template<> EIGEN_STRONG_INLINE void
 pbroadcast4<Packet2d>(const double *a,
                       Packet2d& a0, Packet2d& a1, Packet2d& a2, Packet2d& a3)
@@ -326,6 +449,16 @@ template<> EIGEN_DEVICE_FUNC inline Packet4i pgather<int, Packet4i>(const int* f
  return pload<Packet4i>(ai);
 }
 
+template<> EIGEN_DEVICE_FUNC inline Packet4f pgather<float, Packet4f>(const float* from, Index stride)
+{
+  float EIGEN_ALIGN16 ai[4];
+  ai[0] = from[0*stride];
+  ai[1] = from[1*stride];
+  ai[2] = from[2*stride];
+  ai[3] = from[3*stride];
+ return pload<Packet4f>(ai);
+}
+
 template<> EIGEN_DEVICE_FUNC inline Packet2d pgather<double, Packet2d>(const double* from, Index stride)
 {
   double EIGEN_ALIGN16 af[2];
@@ -344,6 +477,16 @@ template<> EIGEN_DEVICE_FUNC inline void pscatter<int, Packet4i>(int* to, const
   to[3*stride] = ai[3];
 }
 
+template<> EIGEN_DEVICE_FUNC inline void pscatter<float, Packet4f>(float* to, const Packet4f& from, Index stride)
+{
+  float EIGEN_ALIGN16 ai[4];
+  pstore<float>((float *)ai, from);
+  to[0*stride] = ai[0];
+  to[1*stride] = ai[1];
+  to[2*stride] = ai[2];
+  to[3*stride] = ai[3];
+}
+
 template<> EIGEN_DEVICE_FUNC inline void pscatter<double, Packet2d>(double* to, const Packet2d& from, Index stride)
 {
   double EIGEN_ALIGN16 af[2];
@@ -353,52 +496,160 @@ template<> EIGEN_DEVICE_FUNC inline void pscatter<double, Packet2d>(double* to,
 }
 
 template<> EIGEN_STRONG_INLINE Packet4i padd<Packet4i>(const Packet4i& a, const Packet4i& b) { return (a + b); }
+template<> EIGEN_STRONG_INLINE Packet4f padd<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f c;
+  c.v4f[0] = a.v4f[0] + b.v4f[0];
+  c.v4f[1] = a.v4f[1] + b.v4f[1];
+  return c;
+}
 template<> EIGEN_STRONG_INLINE Packet2d padd<Packet2d>(const Packet2d& a, const Packet2d& b) { return (a + b); }
 
 template<> EIGEN_STRONG_INLINE Packet4i psub<Packet4i>(const Packet4i& a, const Packet4i& b) { return (a - b); }
+template<> EIGEN_STRONG_INLINE Packet4f psub<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f c;
+  c.v4f[0] = a.v4f[0] - b.v4f[0];
+  c.v4f[1] = a.v4f[1] - b.v4f[1];
+  return c;
+}
 template<> EIGEN_STRONG_INLINE Packet2d psub<Packet2d>(const Packet2d& a, const Packet2d& b) { return (a - b); }
 
 template<> EIGEN_STRONG_INLINE Packet4i pmul<Packet4i>(const Packet4i& a, const Packet4i& b) { return (a * b); }
+template<> EIGEN_STRONG_INLINE Packet4f pmul<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f c;
+  c.v4f[0] = a.v4f[0] * b.v4f[0];
+  c.v4f[1] = a.v4f[1] * b.v4f[1];
+  return c;
+}
 template<> EIGEN_STRONG_INLINE Packet2d pmul<Packet2d>(const Packet2d& a, const Packet2d& b) { return (a * b); }
 
 template<> EIGEN_STRONG_INLINE Packet4i pdiv<Packet4i>(const Packet4i& a, const Packet4i& b) { return (a / b); }
+template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f c;
+  c.v4f[0] = a.v4f[0] / b.v4f[0];
+  c.v4f[1] = a.v4f[1] / b.v4f[1];
+  return c;
+}
 template<> EIGEN_STRONG_INLINE Packet2d pdiv<Packet2d>(const Packet2d& a, const Packet2d& b) { return (a / b); }
 
 template<> EIGEN_STRONG_INLINE Packet4i pnegate(const Packet4i& a) { return (-a); }
+template<> EIGEN_STRONG_INLINE Packet4f pnegate(const Packet4f& a)
+{
+  Packet4f c;
+  c.v4f[0] = -a.v4f[0];
+  c.v4f[1] = -a.v4f[1];
+  return c;
+}
 template<> EIGEN_STRONG_INLINE Packet2d pnegate(const Packet2d& a) { return (-a); }
 
 template<> EIGEN_STRONG_INLINE Packet4i pconj(const Packet4i& a) { return a; }
+template<> EIGEN_STRONG_INLINE Packet4f pconj(const Packet4f& a) { return a; }
 template<> EIGEN_STRONG_INLINE Packet2d pconj(const Packet2d& a) { return a; }
 
 template<> EIGEN_STRONG_INLINE Packet4i pmadd(const Packet4i& a, const Packet4i& b, const Packet4i& c) { return padd<Packet4i>(pmul<Packet4i>(a, b), c); }
+template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c)
+{
+  Packet4f res;
+  res.v4f[0] = vec_madd(a.v4f[0], b.v4f[0], c.v4f[0]);
+  res.v4f[1] = vec_madd(a.v4f[1], b.v4f[1], c.v4f[1]);
+  return res;
+}
 template<> EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d& b, const Packet2d& c) { return vec_madd(a, b, c); }
 
 template<> EIGEN_STRONG_INLINE Packet4i plset<Packet4i>(const int& a)    { return padd<Packet4i>(pset1<Packet4i>(a), p4i_COUNTDOWN); }
+template<> EIGEN_STRONG_INLINE Packet4f plset<Packet4f>(const float& a)  { return padd<Packet4f>(pset1<Packet4f>(a), p4f_COUNTDOWN); }
 template<> EIGEN_STRONG_INLINE Packet2d plset<Packet2d>(const double& a) { return padd<Packet2d>(pset1<Packet2d>(a), p2d_COUNTDOWN); }
 
 template<> EIGEN_STRONG_INLINE Packet4i pmin<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_min(a, b); }
 template<> EIGEN_STRONG_INLINE Packet2d pmin<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_min(a, b); }
+template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f res;
+  res.v4f[0] = pmin(a.v4f[0], b.v4f[0]);
+  res.v4f[1] = pmin(a.v4f[1], b.v4f[1]);
+  return res;
+}
 
 template<> EIGEN_STRONG_INLINE Packet4i pmax<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_max(a, b); }
 template<> EIGEN_STRONG_INLINE Packet2d pmax<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_max(a, b); }
+template<> EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f res;
+  res.v4f[0] = pmax(a.v4f[0], b.v4f[0]);
+  res.v4f[1] = pmax(a.v4f[1], b.v4f[1]);
+  return res;
+}
 
 template<> EIGEN_STRONG_INLINE Packet4i pand<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_and(a, b); }
 template<> EIGEN_STRONG_INLINE Packet2d pand<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_and(a, b); }
+template<> EIGEN_STRONG_INLINE Packet4f pand<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f res;
+  res.v4f[0] = pand(a.v4f[0], b.v4f[0]);
+  res.v4f[1] = pand(a.v4f[1], b.v4f[1]);
+  return res;
+}
 
 template<> EIGEN_STRONG_INLINE Packet4i por<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_or(a, b); }
 template<> EIGEN_STRONG_INLINE Packet2d por<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_or(a, b); }
+template<> EIGEN_STRONG_INLINE Packet4f por<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f res;
+  res.v4f[0] = pand(a.v4f[0], b.v4f[0]);
+  res.v4f[1] = pand(a.v4f[1], b.v4f[1]);
+  return res;
+}
 
 template<> EIGEN_STRONG_INLINE Packet4i pxor<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_xor(a, b); }
 template<> EIGEN_STRONG_INLINE Packet2d pxor<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_xor(a, b); }
+template<> EIGEN_STRONG_INLINE Packet4f pxor<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f res;
+  res.v4f[0] = pand(a.v4f[0], b.v4f[0]);
+  res.v4f[1] = pand(a.v4f[1], b.v4f[1]);
+  return res;
+}
 
 template<> EIGEN_STRONG_INLINE Packet4i pandnot<Packet4i>(const Packet4i& a, const Packet4i& b) { return pand<Packet4i>(a, vec_nor(b, b)); }
 template<> EIGEN_STRONG_INLINE Packet2d pandnot<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_and(a, vec_nor(b, b)); }
+template<> EIGEN_STRONG_INLINE Packet4f pandnot<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f res;
+  res.v4f[0] = pandnot(a.v4f[0], b.v4f[0]);
+  res.v4f[1] = pandnot(a.v4f[1], b.v4f[1]);
+  return res;
+}
 
+template<> EIGEN_STRONG_INLINE Packet4f pround<Packet4f>(const Packet4f& a)
+{
+  Packet4f res;
+  res.v4f[0] = vec_round(a.v4f[0]);
+  res.v4f[1] = vec_round(a.v4f[1]);
+  return res;
+}
 template<> EIGEN_STRONG_INLINE Packet2d pround<Packet2d>(const Packet2d& a) { return vec_round(a); }
+template<> EIGEN_STRONG_INLINE Packet4f pceil<Packet4f>(const  Packet4f& a)
+{
+  Packet4f res;
+  res.v4f[0] = vec_ceil(a.v4f[0]);
+  res.v4f[1] = vec_ceil(a.v4f[1]);
+  return res;
+}
 template<> EIGEN_STRONG_INLINE Packet2d pceil<Packet2d>(const  Packet2d& a) { return vec_ceil(a); }
+template<> EIGEN_STRONG_INLINE Packet4f pfloor<Packet4f>(const Packet4f& a)
+{
+  Packet4f res;
+  res.v4f[0] = vec_floor(a.v4f[0]);
+  res.v4f[1] = vec_floor(a.v4f[1]);
+  return res;
+}
 template<> EIGEN_STRONG_INLINE Packet2d pfloor<Packet2d>(const Packet2d& a) { return vec_floor(a); }
 
 template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int*       from) { return pload<Packet4i>(from); }
+template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float*     from) { return pload<Packet4f>(from); }
 template<> EIGEN_STRONG_INLINE Packet2d ploadu<Packet2d>(const double*    from) { return pload<Packet2d>(from); }
 
 
@@ -408,6 +659,14 @@ template<> EIGEN_STRONG_INLINE Packet4i ploaddup<Packet4i>(const int*     from)
   return vec_perm(p, p, p16uc_DUPLICATE32_HI);
 }
 
+template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float*    from)
+{
+  Packet4f p = pload<Packet4f>(from);
+  p.v4f[1] = vec_splat(p.v4f[0], 1);
+  p.v4f[0] = vec_splat(p.v4f[0], 0);
+  return p;
+}
+
 template<> EIGEN_STRONG_INLINE Packet2d ploaddup<Packet2d>(const double*   from)
 {
   Packet2d p = pload<Packet2d>(from);
@@ -415,12 +674,15 @@ template<> EIGEN_STRONG_INLINE Packet2d ploaddup<Packet2d>(const double*   from)
 }
 
 template<> EIGEN_STRONG_INLINE void pstoreu<int>(int*        to, const Packet4i& from) { pstore<int>(to, from); }
+template<> EIGEN_STRONG_INLINE void pstoreu<float>(float*    to, const Packet4f& from) { pstore<float>(to, from); }
 template<> EIGEN_STRONG_INLINE void pstoreu<double>(double*  to, const Packet2d& from) { pstore<double>(to, from); }
 
 template<> EIGEN_STRONG_INLINE void prefetch<int>(const int*       addr) { EIGEN_ZVECTOR_PREFETCH(addr); }
+template<> EIGEN_STRONG_INLINE void prefetch<float>(const float*   addr) { EIGEN_ZVECTOR_PREFETCH(addr); }
 template<> EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) { EIGEN_ZVECTOR_PREFETCH(addr); }
 
 template<> EIGEN_STRONG_INLINE int    pfirst<Packet4i>(const Packet4i& a) { int    EIGEN_ALIGN16 x[4]; pstore(x, a); return x[0]; }
+template<> EIGEN_STRONG_INLINE float  pfirst<Packet4f>(const Packet4f& a) { float  EIGEN_ALIGN16 x[2]; vec_st2f(a.v4f[0], &x[0]); return x[0]; }
 template<> EIGEN_STRONG_INLINE double pfirst<Packet2d>(const Packet2d& a) { double EIGEN_ALIGN16 x[2]; pstore(x, a); return x[0]; }
 
 template<> EIGEN_STRONG_INLINE Packet4i preverse(const Packet4i& a)
@@ -433,8 +695,23 @@ template<> EIGEN_STRONG_INLINE Packet2d preverse(const Packet2d& a)
   return reinterpret_cast<Packet2d>(vec_perm(reinterpret_cast<Packet16uc>(a), reinterpret_cast<Packet16uc>(a), p16uc_REVERSE64));
 }
 
-template<> EIGEN_STRONG_INLINE Packet4i pabs(const Packet4i& a) { return vec_abs(a); }
-template<> EIGEN_STRONG_INLINE Packet2d pabs(const Packet2d& a) { return vec_abs(a); }
+template<> EIGEN_STRONG_INLINE Packet4f preverse(const Packet4f& a)
+{
+  Packet4f rev;
+  rev.v4f[0] = preverse<Packet2d>(a.v4f[1]);
+  rev.v4f[1] = preverse<Packet2d>(a.v4f[0]);
+  return rev;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4i pabs<Packet4i>(const Packet4i& a) { return vec_abs(a); }
+template<> EIGEN_STRONG_INLINE Packet2d pabs<Packet2d>(const Packet2d& a) { return vec_abs(a); }
+template<> EIGEN_STRONG_INLINE Packet4f pabs<Packet4f>(const Packet4f& a)
+{
+  Packet4f res;
+  res.v4f[0] = pabs(a.v4f[0]);
+  res.v4f[1] = pabs(a.v4f[1]);
+  return res;
+}
 
 template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a)
 {
@@ -453,6 +730,13 @@ template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a)
   sum = padd<Packet2d>(a, b);
   return pfirst(sum);
 }
+template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a)
+{
+  Packet2d sum;
+  sum = padd<Packet2d>(a.v4f[0], a.v4f[1]);
+  double first = predux<Packet2d>(sum);
+  return static_cast<float>(first);
+}
 
 template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs)
 {
@@ -493,6 +777,21 @@ template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs)
   return sum;
 }
 
+template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs)
+{
+  PacketBlock<Packet4f,4> transpose;
+  transpose.packet[0] = vecs[0];
+  transpose.packet[1] = vecs[1];
+  transpose.packet[2] = vecs[2];
+  transpose.packet[3] = vecs[3];
+  ptranspose(transpose);
+
+  Packet4f sum = padd(transpose.packet[0], transpose.packet[1]);
+  sum = padd(sum, transpose.packet[2]);
+  sum = padd(sum, transpose.packet[3]);
+  return sum;
+}
+
 // Other reduction functions:
 // mul
 template<> EIGEN_STRONG_INLINE int predux_mul<Packet4i>(const Packet4i& a)
@@ -507,6 +806,12 @@ template<> EIGEN_STRONG_INLINE double predux_mul<Packet2d>(const Packet2d& a)
   return pfirst(pmul(a, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(a), reinterpret_cast<Packet4i>(a), 8))));
 }
 
+template<> EIGEN_STRONG_INLINE float predux_mul<Packet4f>(const Packet4f& a)
+{
+  // Return predux_mul<Packet2d> of the subvectors product
+  return static_cast<float>(pfirst(predux_mul(pmul(a.v4f[0], a.v4f[1]))));
+}
+
 // min
 template<> EIGEN_STRONG_INLINE int predux_min<Packet4i>(const Packet4i& a)
 {
@@ -521,6 +826,14 @@ template<> EIGEN_STRONG_INLINE double predux_min<Packet2d>(const Packet2d& a)
   return pfirst(pmin<Packet2d>(a, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(a), reinterpret_cast<Packet4i>(a), 8))));
 }
 
+template<> EIGEN_STRONG_INLINE float predux_min<Packet4f>(const Packet4f& a)
+{
+  Packet2d b, res;
+  b   = pmin<Packet2d>(a.v4f[0], a.v4f[1]);
+  res = pmin<Packet2d>(b, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(b), reinterpret_cast<Packet4i>(b), 8)));
+  return static_cast<float>(pfirst(res));
+}
+
 // max
 template<> EIGEN_STRONG_INLINE int predux_max<Packet4i>(const Packet4i& a)
 {
@@ -536,6 +849,14 @@ template<> EIGEN_STRONG_INLINE double predux_max<Packet2d>(const Packet2d& a)
   return pfirst(pmax<Packet2d>(a, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(a), reinterpret_cast<Packet4i>(a), 8))));
 }
 
+template<> EIGEN_STRONG_INLINE float predux_max<Packet4f>(const Packet4f& a)
+{
+  Packet2d b, res;
+  b   = pmax<Packet2d>(a.v4f[0], a.v4f[1]);
+  res = pmax<Packet2d>(b, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(b), reinterpret_cast<Packet4i>(b), 8)));
+  return static_cast<float>(pfirst(res));
+}
+
 EIGEN_DEVICE_FUNC inline void
 ptranspose(PacketBlock<Packet4i,4>& kernel) {
   Packet4i t0 = vec_mergeh(kernel.packet[0], kernel.packet[2]);
@@ -556,12 +877,61 @@ ptranspose(PacketBlock<Packet2d,2>& kernel) {
   kernel.packet[1] = t1;
 }
 
+/* Split the Packet4f PacketBlock into 4 Packet2d PacketBlocks and transpose each one
+ */
+EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<Packet4f,4>& kernel) {
+  PacketBlock<Packet2d,2> t0,t1,t2,t3;
+  // copy top-left 2x2 Packet2d block
+  t0.packet[0] = kernel.packet[0].v4f[0];
+  t0.packet[1] = kernel.packet[1].v4f[0];
+
+  // copy top-right 2x2 Packet2d block
+  t1.packet[0] = kernel.packet[0].v4f[1];
+  t1.packet[1] = kernel.packet[1].v4f[1];
+
+  // copy bottom-left 2x2 Packet2d block
+  t2.packet[0] = kernel.packet[2].v4f[0];
+  t2.packet[1] = kernel.packet[3].v4f[0];
+
+  // copy bottom-right 2x2 Packet2d block
+  t3.packet[0] = kernel.packet[2].v4f[1];
+  t3.packet[1] = kernel.packet[3].v4f[1];
+
+  // Transpose all 2x2 blocks
+  ptranspose(t0);
+  ptranspose(t1);
+  ptranspose(t2);
+  ptranspose(t3);
+
+  // Copy back transposed blocks, but exchange t1 and t2 due to transposition
+  kernel.packet[0].v4f[0] = t0.packet[0];
+  kernel.packet[0].v4f[1] = t2.packet[0];
+  kernel.packet[1].v4f[0] = t0.packet[1];
+  kernel.packet[1].v4f[1] = t2.packet[1];
+  kernel.packet[2].v4f[0] = t1.packet[0];
+  kernel.packet[2].v4f[1] = t3.packet[0];
+  kernel.packet[3].v4f[0] = t1.packet[1];
+  kernel.packet[3].v4f[1] = t3.packet[1];
+}
+
 template<> EIGEN_STRONG_INLINE Packet4i pblend(const Selector<4>& ifPacket, const Packet4i& thenPacket, const Packet4i& elsePacket) {
   Packet4ui select = { ifPacket.select[0], ifPacket.select[1], ifPacket.select[2], ifPacket.select[3] };
   Packet4ui mask = vec_cmpeq(select, reinterpret_cast<Packet4ui>(p4i_ONE));
   return vec_sel(elsePacket, thenPacket, mask);
 }
 
+template<> EIGEN_STRONG_INLINE Packet4f pblend(const Selector<4>& ifPacket, const Packet4f& thenPacket, const Packet4f& elsePacket) {
+  Packet2ul select_hi = { ifPacket.select[0], ifPacket.select[1] };
+  Packet2ul select_lo = { ifPacket.select[2], ifPacket.select[3] };
+  Packet2ul mask_hi = vec_cmpeq(select_hi, reinterpret_cast<Packet2ul>(p2l_ONE));
+  Packet2ul mask_lo = vec_cmpeq(select_lo, reinterpret_cast<Packet2ul>(p2l_ONE));
+  Packet4f result;
+  result.v4f[0] = vec_sel(elsePacket.v4f[0], thenPacket.v4f[0], mask_hi);
+  result.v4f[1] = vec_sel(elsePacket.v4f[1], thenPacket.v4f[1], mask_lo);
+  return result;
+}
+
 template<> EIGEN_STRONG_INLINE Packet2d pblend(const Selector<2>& ifPacket, const Packet2d& thenPacket, const Packet2d& elsePacket) {
   Packet2ul select = { ifPacket.select[0], ifPacket.select[1] };
   Packet2ul mask = vec_cmpeq(select, reinterpret_cast<Packet2ul>(p2l_ONE));

Eigen/src/Core/functors/AssignmentFunctors.h

@@ -28,7 +28,7 @@ template<typename DstScalar,typename SrcScalar> struct assign_op {
   { internal::pstoret<DstScalar,Packet,Alignment>(a,b); }
 };
 
-// Empty overload for void type (used by PermutationMatrix
+// Empty overload for void type (used by PermutationMatrix)
 template<typename DstScalar> struct assign_op<DstScalar,void> {};
 
 template<typename DstScalar,typename SrcScalar>

Eigen/src/Core/functors/BinaryFunctors.h

@@ -266,7 +266,7 @@ struct scalar_hypot_op<Scalar,Scalar> : binary_op_base<Scalar,Scalar>
 //   typedef typename NumTraits<Scalar>::Real result_type;
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& _x, const Scalar& _y) const
   {
-    using std::sqrt;
+    EIGEN_USING_STD_MATH(sqrt)
     Scalar p, qp;
     if(_x>_y)
     {
@@ -287,7 +287,7 @@ struct functor_traits<scalar_hypot_op<Scalar,Scalar> > {
   {
     Cost = 3 * NumTraits<Scalar>::AddCost +
            2 * NumTraits<Scalar>::MulCost +
-           2 * NumTraits<Scalar>::template Div<false>::Cost,
+           2 * scalar_div_cost<Scalar,false>::value,
     PacketAccess = false
   };
 };
@@ -375,7 +375,7 @@ struct functor_traits<scalar_quotient_op<LhsScalar,RhsScalar> > {
   typedef typename scalar_quotient_op<LhsScalar,RhsScalar>::result_type result_type;
   enum {
     PacketAccess = is_same<LhsScalar,RhsScalar>::value && packet_traits<LhsScalar>::HasDiv && packet_traits<RhsScalar>::HasDiv,
-    Cost = NumTraits<result_type>::template Div<PacketAccess>::Cost
+    Cost = scalar_div_cost<result_type,PacketAccess>::value
   };
 };
 

Eigen/src/Core/functors/CMakeLists.txt

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

Eigen/src/Core/functors/NullaryFunctors.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) 2008-2016 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
@@ -18,10 +18,9 @@ template<typename Scalar>
 struct scalar_constant_op {
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE scalar_constant_op(const scalar_constant_op& other) : m_other(other.m_other) { }
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE scalar_constant_op(const Scalar& other) : m_other(other) { }
-  template<typename Index>
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (Index, Index = 0) const { return m_other; }
-  template<typename Index, typename PacketType>
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const PacketType packetOp(Index, Index = 0) const { return internal::pset1<PacketType>(m_other); }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() () const { return m_other; }
+  template<typename PacketType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const PacketType packetOp() const { return internal::pset1<PacketType>(m_other); }
   const Scalar m_other;
 };
 template<typename Scalar>
@@ -31,94 +30,85 @@ struct functor_traits<scalar_constant_op<Scalar> >
 
 template<typename Scalar> struct scalar_identity_op {
   EIGEN_EMPTY_STRUCT_CTOR(scalar_identity_op)
-  template<typename Index>
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (Index row, Index col) const { return row==col ? Scalar(1) : Scalar(0); }
+  template<typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (IndexType row, IndexType 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, typename Packet, bool RandomAccess, bool IsInteger> 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, typename Packet>
-struct linspaced_op_impl<Scalar,Packet,/*RandomAccess*/false,/*IsInteger*/false>
-{
-  linspaced_op_impl(const Scalar& low, const Scalar& high, Index num_steps) :
-    m_low(low), m_step(num_steps==1 ? Scalar() : (high-low)/Scalar(num_steps-1)),
-    m_packetStep(pset1<Packet>(unpacket_traits<Packet>::size*m_step)),
-    m_base(padd(pset1<Packet>(low), pmul(pset1<Packet>(m_step),plset<Packet>(-unpacket_traits<Packet>::size)))) {}
-
-  template<typename Index>
-  EIGEN_DEVICE_FUNC 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_DEVICE_FUNC 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, typename Packet>
-struct linspaced_op_impl<Scalar,Packet,/*RandomAccess*/true,/*IsInteger*/false>
-{
-  linspaced_op_impl(const Scalar& low, const Scalar& high, Index num_steps) :
-    m_low(low), m_step(num_steps==1 ? Scalar() : (high-low)/Scalar(num_steps-1)),
-    m_lowPacket(pset1<Packet>(m_low)), m_stepPacket(pset1<Packet>(m_step)), m_interPacket(plset<Packet>(0)) {}
-
-  template<typename Index>
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (Index i) const { return m_low+i*m_step; }
-
-  template<typename Index>
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(Index i) const
-  { return internal::padd(m_lowPacket, pmul(m_stepPacket, padd(pset1<Packet>(Scalar(i)),m_interPacket))); }
-
-  const Scalar m_low;
-  const Scalar m_step;
-  const Packet m_lowPacket;
-  const Packet m_stepPacket;
-  const Packet m_interPacket;
-};
+template <typename Scalar, typename Packet, bool IsInteger> struct linspaced_op_impl;
 
 template <typename Scalar, typename Packet>
-struct linspaced_op_impl<Scalar,Packet,/*RandomAccess*/true,/*IsInteger*/true>
+struct linspaced_op_impl<Scalar,Packet,/*IsInteger*/false>
 {
   linspaced_op_impl(const Scalar& low, const Scalar& high, Index num_steps) :
-    m_low(low), m_length(high-low), m_divisor(convert_index<Scalar>(num_steps==1?1:num_steps-1)), m_interPacket(plset<Packet>(0))
+    m_low(low), m_high(high), m_size1(num_steps==1 ? 1 : num_steps-1), m_step(num_steps==1 ? Scalar() : (high-low)/Scalar(num_steps-1)),
+    m_interPacket(plset<Packet>(0)),
+    m_flip(numext::abs(high)<numext::abs(low))
   {}
 
-  template<typename Index>
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
-  const Scalar operator() (Index i) const {
-    return m_low + (m_length*Scalar(i))/m_divisor;
+  template<typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (IndexType i) const {
+    if(m_flip)
+      return (i==0)? m_low : (m_high - (m_size1-i)*m_step);
+    else
+      return (i==m_size1)? m_high : (m_low + i*m_step);
   }
 
-  template<typename Index>
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
-  const Packet packetOp(Index i) const {
-    return internal::padd(pset1<Packet>(m_low), pdiv(pmul(pset1<Packet>(m_length), padd(pset1<Packet>(Scalar(i)),m_interPacket)),
-                                                     pset1<Packet>(m_divisor))); }
+  template<typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(IndexType i) const
+  {
+    // Principle:
+    // [low, ..., low] + ( [step, ..., step] * ( [i, ..., i] + [0, ..., size] ) )
+    if(m_flip)
+    {
+      Packet pi = padd(pset1<Packet>(Scalar(i-m_size1)),m_interPacket);
+      Packet res = padd(pset1<Packet>(m_high), pmul(pset1<Packet>(m_step), pi));
+      if(i==0)
+        res = pinsertfirst(res, m_low);
+      return res;
+    }
+    else
+    {
+      Packet pi = padd(pset1<Packet>(Scalar(i)),m_interPacket);
+      Packet res = padd(pset1<Packet>(m_low), pmul(pset1<Packet>(m_step), pi));
+      if(i==m_size1-unpacket_traits<Packet>::size+1)
+        res = pinsertlast(res, m_high);
+      return res;
+    }
+  }
 
   const Scalar m_low;
-  const Scalar m_length;
-  const Scalar  m_divisor;
+  const Scalar m_high;
+  const Index m_size1;
+  const Scalar m_step;
   const Packet m_interPacket;
+  const bool m_flip;
+};
+
+template <typename Scalar, typename Packet>
+struct linspaced_op_impl<Scalar,Packet,/*IsInteger*/true>
+{
+  linspaced_op_impl(const Scalar& low, const Scalar& high, Index num_steps) :
+    m_low(low),
+    m_multiplier((high-low)/convert_index<Scalar>(num_steps<=1 ? 1 : num_steps-1)),
+    m_divisor(convert_index<Scalar>((high>=low?num_steps:-num_steps)+(high-low))/((numext::abs(high-low)+1)==0?1:(numext::abs(high-low)+1))),
+    m_use_divisor(num_steps>1 && (numext::abs(high-low)+1)<num_steps)
+  {}
+
+  template<typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  const Scalar operator() (IndexType i) const
+  {
+    if(m_use_divisor) return m_low + convert_index<Scalar>(i)/m_divisor;
+    else              return m_low + convert_index<Scalar>(i)*m_multiplier;
+  }
+
+  const Scalar m_low;
+  const Scalar m_multiplier;
+  const Scalar m_divisor;
+  const bool m_use_divisor;
 };
 
 // ----- Linspace functor ----------------------------------------------------------------
@@ -126,60 +116,71 @@ struct linspaced_op_impl<Scalar,Packet,/*RandomAccess*/true,/*IsInteger*/true>
 // 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, typename PacketType, bool RandomAccess = true> struct linspaced_op;
-template <typename Scalar, typename PacketType, bool RandomAccess> struct functor_traits< linspaced_op<Scalar,PacketType,RandomAccess> >
+template <typename Scalar, typename PacketType> struct linspaced_op;
+template <typename Scalar, typename PacketType> struct functor_traits< linspaced_op<Scalar,PacketType> >
 {
   enum
   {
     Cost = 1,
-    PacketAccess =   packet_traits<Scalar>::HasSetLinear
-                  && ((!NumTraits<Scalar>::IsInteger) || packet_traits<Scalar>::HasDiv),
+    PacketAccess =   (!NumTraits<Scalar>::IsInteger) && packet_traits<Scalar>::HasSetLinear && packet_traits<Scalar>::HasBlend,
+                  /*&& ((!NumTraits<Scalar>::IsInteger) || packet_traits<Scalar>::HasDiv),*/ // <- vectorization for integer is currently disabled
     IsRepeatable = true
   };
 };
-template <typename Scalar, typename PacketType, bool RandomAccess> struct linspaced_op
+template <typename Scalar, typename PacketType> struct linspaced_op
 {
   linspaced_op(const Scalar& low, const Scalar& high, Index num_steps)
     : impl((num_steps==1 ? high : low),high,num_steps)
   {}
 
-  template<typename Index>
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (Index i) const { return impl(i); }
+  template<typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (IndexType 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_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (Index row, Index col) const 
-  {
-    eigen_assert(col==0 || row==0);
-    return impl(col + row);
-  }
+  template<typename Packet,typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(IndexType i) const { return impl.packetOp(i); }
 
-  template<typename Index, typename Packet>
-  EIGEN_DEVICE_FUNC 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, typename Packet>
-  EIGEN_DEVICE_FUNC 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.
-  // As long as we don't have a Bresenham-like implementation for linear-access and integer types,
-  // we have to by-pass RandomAccess for integer types. See bug 698.
-  const linspaced_op_impl<Scalar,PacketType,(NumTraits<Scalar>::IsInteger?true:RandomAccess),NumTraits<Scalar>::IsInteger> impl;
+  // This proxy object handles the actual required temporaries and the different
+  // implementations (integer vs. floating point).
+  const linspaced_op_impl<Scalar,PacketType,NumTraits<Scalar>::IsInteger> 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.
-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 }; };
+// Linear access is automatically determined from the operator() prototypes available for the given functor.
+// If it exposes an operator()(i,j), then we assume the i and j coefficients are required independently
+// and linear access is not possible. In all other cases, linear access is enabled.
+// Users should not have to deal with this structure.
+template<typename Functor> struct functor_has_linear_access { enum { ret = !has_binary_operator<Functor>::value }; };
+
+// For unreliable compilers, let's specialize the has_*ary_operator
+// helpers so that at least built-in nullary functors work fine.
+#if !( (EIGEN_COMP_MSVC>1600) || (EIGEN_GNUC_AT_LEAST(4,8)) || (EIGEN_COMP_ICC>=1600))
+template<typename Scalar,typename IndexType>
+struct has_nullary_operator<scalar_constant_op<Scalar>,IndexType> { enum { value = 1}; };
+template<typename Scalar,typename IndexType>
+struct has_unary_operator<scalar_constant_op<Scalar>,IndexType> { enum { value = 0}; };
+template<typename Scalar,typename IndexType>
+struct has_binary_operator<scalar_constant_op<Scalar>,IndexType> { enum { value = 0}; };
+
+template<typename Scalar,typename IndexType>
+struct has_nullary_operator<scalar_identity_op<Scalar>,IndexType> { enum { value = 0}; };
+template<typename Scalar,typename IndexType>
+struct has_unary_operator<scalar_identity_op<Scalar>,IndexType> { enum { value = 0}; };
+template<typename Scalar,typename IndexType>
+struct has_binary_operator<scalar_identity_op<Scalar>,IndexType> { enum { value = 1}; };
+
+template<typename Scalar, typename PacketType,typename IndexType>
+struct has_nullary_operator<linspaced_op<Scalar,PacketType>,IndexType> { enum { value = 0}; };
+template<typename Scalar, typename PacketType,typename IndexType>
+struct has_unary_operator<linspaced_op<Scalar,PacketType>,IndexType> { enum { value = 1}; };
+template<typename Scalar, typename PacketType,typename IndexType>
+struct has_binary_operator<linspaced_op<Scalar,PacketType>,IndexType> { enum { value = 0}; };
+
+template<typename Scalar,typename IndexType>
+struct has_nullary_operator<scalar_random_op<Scalar>,IndexType> { enum { value = 1}; };
+template<typename Scalar,typename IndexType>
+struct has_unary_operator<scalar_random_op<Scalar>,IndexType> { enum { value = 0}; };
+template<typename Scalar,typename IndexType>
+struct has_binary_operator<scalar_random_op<Scalar>,IndexType> { enum { value = 0}; };
+#endif
 
 } // end namespace internal
 

Eigen/src/Core/functors/StlFunctors.h

@@ -72,7 +72,7 @@ template<typename T>
 struct functor_traits<std::not_equal_to<T> >
 { enum { Cost = 1, PacketAccess = false }; };
 
-#if(__cplusplus < 201103L)
+#if (__cplusplus < 201103L) && (EIGEN_COMP_MSVC <= 1900)
 // std::binder* are deprecated since c++11 and will be removed in c++17
 template<typename T>
 struct functor_traits<std::binder2nd<T> >

Eigen/src/Core/functors/UnaryFunctors.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) 2008-2016 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
@@ -248,7 +248,7 @@ struct functor_traits<scalar_exp_op<Scalar> > {
      // double: 7 pmadd, 5 pmul, 3 padd/psub, 1 div,  13 other
      : (14 * NumTraits<Scalar>::AddCost +
         6 * NumTraits<Scalar>::MulCost +
-        NumTraits<Scalar>::template Div<packet_traits<Scalar>::HasDiv>::Cost))
+        scalar_div_cost<Scalar,packet_traits<Scalar>::HasDiv>::value))
 #else
     Cost =
     (sizeof(Scalar) == 4
@@ -257,7 +257,7 @@ struct functor_traits<scalar_exp_op<Scalar> > {
      // double: 7 pmadd, 5 pmul, 3 padd/psub, 1 div,  13 other
      : (23 * NumTraits<Scalar>::AddCost +
         12 * NumTraits<Scalar>::MulCost +
-        NumTraits<Scalar>::template Div<packet_traits<Scalar>::HasDiv>::Cost))
+        scalar_div_cost<Scalar,packet_traits<Scalar>::HasDiv>::value))
 #endif
   };
 };
@@ -321,7 +321,7 @@ struct functor_traits<scalar_log1p_op<Scalar> > {
   */
 template<typename Scalar> struct scalar_log10_op {
   EIGEN_EMPTY_STRUCT_CTOR(scalar_log10_op)
-  EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { using std::log10; return log10(a); }
+  EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { EIGEN_USING_STD_MATH(log10) return log10(a); }
   template <typename Packet>
   EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::plog10(a); }
 };
@@ -491,85 +491,40 @@ struct functor_traits<scalar_atan_op<Scalar> >
   };
 };
 
-
 /** \internal
   * \brief Template functor to compute the tanh of a scalar
   * \sa class CwiseUnaryOp, ArrayBase::tanh()
   */
-template<typename Scalar> struct scalar_tanh_op {
+template <typename Scalar>
+struct scalar_tanh_op {
   EIGEN_EMPTY_STRUCT_CTOR(scalar_tanh_op)
-  EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return numext::tanh(a); }
+  EIGEN_DEVICE_FUNC inline const Scalar operator()(const Scalar& a) const { return numext::tanh(a); }
   template <typename Packet>
-  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& _x) const {
-    /** \internal \returns the hyperbolic tan of \a a (coeff-wise)
-        Doesn't do anything fancy, just a 13/6-degree rational interpolant which
-        is accurate up to a couple of ulp in the range [-9, 9], outside of which the
-        fl(tanh(x)) = +/-1. */
-
-    // Clamp the inputs to the range [-9, 9] since anything outside
-    // this range is +/-1.0f in single-precision.
-    const Packet plus_9 = pset1<Packet>(9.0);
-    const Packet minus_9 = pset1<Packet>(-9.0);
-    const Packet x = pmax(minus_9, pmin(plus_9, _x));
-    
-    // The monomial coefficients of the numerator polynomial (odd).
-    const Packet alpha_1 = pset1<Packet>(4.89352455891786e-03);
-    const Packet alpha_3 = pset1<Packet>(6.37261928875436e-04);
-    const Packet alpha_5 = pset1<Packet>(1.48572235717979e-05);
-    const Packet alpha_7 = pset1<Packet>(5.12229709037114e-08);
-    const Packet alpha_9 = pset1<Packet>(-8.60467152213735e-11);
-    const Packet alpha_11 = pset1<Packet>(2.00018790482477e-13);
-    const Packet alpha_13 = pset1<Packet>(-2.76076847742355e-16);
-    
-    // The monomial coefficients of the denominator polynomial (even).
-    const Packet beta_0 = pset1<Packet>(4.89352518554385e-03);
-    const Packet beta_2 = pset1<Packet>(2.26843463243900e-03);
-    const Packet beta_4 = pset1<Packet>(1.18534705686654e-04);
-    const Packet beta_6 = pset1<Packet>(1.19825839466702e-06);
-    
-    // Since the polynomials are odd/even, we need x^2.
-    const Packet x2 = pmul(x, x);
-  
-  // Evaluate the numerator polynomial p.
-    Packet p = pmadd(x2, alpha_13, alpha_11);
-    p = pmadd(x2, p, alpha_9);
-    p = pmadd(x2, p, alpha_7);
-    p = pmadd(x2, p, alpha_5);
-    p = pmadd(x2, p, alpha_3);
-    p = pmadd(x2, p, alpha_1);
-    p = pmul(x, p);
-    
-    // Evaluate the denominator polynomial p.
-    Packet q = pmadd(x2, beta_6, beta_4);
-    q = pmadd(x2, q, beta_2);
-    q = pmadd(x2, q, beta_0);
-    
-    // Divide the numerator by the denominator.
-    return pdiv(p, q);
-  }
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& x) const { return ptanh(x); }
 };
-template<typename Scalar>
-struct functor_traits<scalar_tanh_op<Scalar> >
-{
+
+template <typename Scalar>
+struct functor_traits<scalar_tanh_op<Scalar> > {
   enum {
     PacketAccess = packet_traits<Scalar>::HasTanh,
-    Cost =
-    (PacketAccess
-     // The following numbers are based on the AVX implementation,
+    Cost = ( (EIGEN_FAST_MATH && is_same<Scalar,float>::value)
+// The following numbers are based on the AVX implementation,
 #ifdef EIGEN_VECTORIZE_FMA
-     // Haswell can issue 2 add/mul/madd per cycle.
-     // 9 pmadd, 2 pmul, 1 div, 2 other
-     ? (2 * NumTraits<Scalar>::AddCost + 6 * NumTraits<Scalar>::MulCost +
-     NumTraits<Scalar>::template Div<packet_traits<Scalar>::HasDiv>::Cost)
+                // Haswell can issue 2 add/mul/madd per cycle.
+                // 9 pmadd, 2 pmul, 1 div, 2 other
+                ? (2 * NumTraits<Scalar>::AddCost +
+                   6 * NumTraits<Scalar>::MulCost +
+                   scalar_div_cost<Scalar,packet_traits<Scalar>::HasDiv>::value)
 #else
-     ? (11 * NumTraits<Scalar>::AddCost +
-        11 * NumTraits<Scalar>::MulCost +
-        NumTraits<Scalar>::template Div<packet_traits<Scalar>::HasDiv>::Cost)
+                ? (11 * NumTraits<Scalar>::AddCost +
+                   11 * NumTraits<Scalar>::MulCost +
+                   scalar_div_cost<Scalar,packet_traits<Scalar>::HasDiv>::value)
 #endif
-     // This number assumes a naive implementation of tanh
-     : (6 * NumTraits<Scalar>::AddCost + 3 * NumTraits<Scalar>::MulCost +
-        2 * NumTraits<Scalar>::template Div<packet_traits<Scalar>::HasDiv>::Cost +
-        functor_traits<scalar_exp_op<Scalar> >::Cost))
+                // This number assumes a naive implementation of tanh
+                : (6 * NumTraits<Scalar>::AddCost +
+                   3 * NumTraits<Scalar>::MulCost +
+                   2 * scalar_div_cost<Scalar,packet_traits<Scalar>::HasDiv>::value +
+                   functor_traits<scalar_exp_op<Scalar> >::Cost))
   };
 };
 

Eigen/src/Core/products/CMakeLists.txt

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

Eigen/src/Core/products/GeneralBlockPanelKernel.h

@@ -434,15 +434,16 @@ public:
   template<typename LhsPacketType, typename RhsPacketType, typename AccPacketType>
   EIGEN_STRONG_INLINE void madd(const LhsPacketType& a, const RhsPacketType& b, AccPacketType& c, AccPacketType& tmp) const
   {
+    conj_helper<LhsPacketType,RhsPacketType,ConjLhs,ConjRhs> cj;
     // It would be a lot cleaner to call pmadd all the time. Unfortunately if we
     // let gcc allocate the register in which to store the result of the pmul
     // (in the case where there is no FMA) gcc fails to figure out how to avoid
     // spilling register.
 #ifdef EIGEN_HAS_SINGLE_INSTRUCTION_MADD
     EIGEN_UNUSED_VARIABLE(tmp);
-    c = pmadd(a,b,c);
+    c = cj.pmadd(a,b,c);
 #else
-    tmp = b; tmp = pmul(a,tmp); c = padd(c,tmp);
+    tmp = b; tmp = cj.pmul(a,tmp); c = padd(c,tmp);
 #endif
   }
 
@@ -457,9 +458,6 @@ public:
     r = pmadd(c,alpha,r);
   }
 
-protected:
-//   conj_helper<LhsScalar,RhsScalar,ConjLhs,ConjRhs> cj;
-//   conj_helper<LhsPacket,RhsPacket,ConjLhs,ConjRhs> pcj;
 };
 
 template<typename RealScalar, bool _ConjLhs>
@@ -582,7 +580,7 @@ DoublePacket<Packet> padd(const DoublePacket<Packet> &a, const DoublePacket<Pack
 }
 
 template<typename Packet>
-const DoublePacket<Packet>& predux4(const DoublePacket<Packet> &a)
+const DoublePacket<Packet>& predux_downto4(const DoublePacket<Packet> &a)
 {
   return a;
 }
@@ -974,7 +972,7 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga
               EIGEN_ASM_COMMENT("begin step of gebp micro kernel 3pX4"); \
               EIGEN_ASM_COMMENT("Note: these asm comments work around bug 935!"); \
               internal::prefetch(blA+(3*K+16)*LhsProgress); \
-              if (EIGEN_ARCH_ARM) internal::prefetch(blB+(4*K+16)*RhsProgress); /* Bug 953 */ \
+              if (EIGEN_ARCH_ARM) { internal::prefetch(blB+(4*K+16)*RhsProgress); } /* Bug 953 */ \
               traits.loadLhs(&blA[(0+3*K)*LhsProgress], A0);  \
               traits.loadLhs(&blA[(1+3*K)*LhsProgress], A1);  \
               traits.loadLhs(&blA[(2+3*K)*LhsProgress], A2);  \
@@ -1583,10 +1581,10 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga
             if(SwappedTraits::LhsProgress==8)
             {
               // Special case where we have to first reduce the accumulation register C0
-              typedef typename conditional<SwappedTraits::LhsProgress==8,typename unpacket_traits<SResPacket>::half,SResPacket>::type SResPacketHalf;
-              typedef typename conditional<SwappedTraits::LhsProgress==8,typename unpacket_traits<SLhsPacket>::half,SLhsPacket>::type SLhsPacketHalf;
-              typedef typename conditional<SwappedTraits::LhsProgress==8,typename unpacket_traits<SLhsPacket>::half,SRhsPacket>::type SRhsPacketHalf;
-              typedef typename conditional<SwappedTraits::LhsProgress==8,typename unpacket_traits<SAccPacket>::half,SAccPacket>::type SAccPacketHalf;
+              typedef typename conditional<SwappedTraits::LhsProgress>=8,typename unpacket_traits<SResPacket>::half,SResPacket>::type SResPacketHalf;
+              typedef typename conditional<SwappedTraits::LhsProgress>=8,typename unpacket_traits<SLhsPacket>::half,SLhsPacket>::type SLhsPacketHalf;
+              typedef typename conditional<SwappedTraits::LhsProgress>=8,typename unpacket_traits<SLhsPacket>::half,SRhsPacket>::type SRhsPacketHalf;
+              typedef typename conditional<SwappedTraits::LhsProgress>=8,typename unpacket_traits<SAccPacket>::half,SAccPacket>::type SAccPacketHalf;
 
               SResPacketHalf R = res.template gatherPacket<SResPacketHalf>(i, j2);
               SResPacketHalf alphav = pset1<SResPacketHalf>(alpha);
@@ -1598,13 +1596,13 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga
                 SRhsPacketHalf b0;
                 straits.loadLhsUnaligned(blB, a0);
                 straits.loadRhs(blA, b0);
-                SAccPacketHalf c0 = predux4(C0);
+                SAccPacketHalf c0 = predux_downto4(C0);
                 straits.madd(a0,b0,c0,b0);
                 straits.acc(c0, alphav, R);
               }
               else
               {
-                straits.acc(predux4(C0), alphav, R);
+                straits.acc(predux_downto4(C0), alphav, R);
               }
               res.scatterPacket(i, j2, R);
             }

Eigen/src/Core/products/GeneralMatrixMatrix.h

@@ -10,7 +10,7 @@
 #ifndef EIGEN_GENERAL_MATRIX_MATRIX_H
 #define EIGEN_GENERAL_MATRIX_MATRIX_H
 
-namespace Eigen { 
+namespace Eigen {
 
 namespace internal {
 
@@ -24,7 +24,7 @@ template<
 struct general_matrix_matrix_product<Index,LhsScalar,LhsStorageOrder,ConjugateLhs,RhsScalar,RhsStorageOrder,ConjugateRhs,RowMajor>
 {
   typedef gebp_traits<RhsScalar,LhsScalar> Traits;
-  
+
   typedef typename ScalarBinaryOpTraits<LhsScalar, RhsScalar>::ReturnType ResScalar;
   static EIGEN_STRONG_INLINE void run(
     Index rows, Index cols, Index depth,
@@ -54,7 +54,7 @@ struct general_matrix_matrix_product<Index,LhsScalar,LhsStorageOrder,ConjugateLh
 {
 
 typedef gebp_traits<LhsScalar,RhsScalar> Traits;
-  
+
 typedef typename ScalarBinaryOpTraits<LhsScalar, RhsScalar>::ReturnType ResScalar;
 static void run(Index rows, Index cols, Index depth,
   const LhsScalar* _lhs, Index lhsStride,
@@ -83,15 +83,15 @@ static void run(Index rows, Index cols, Index depth,
   if(info)
   {
     // this is the parallel version!
-    Index tid = omp_get_thread_num();
-    Index threads = omp_get_num_threads();
-    
+    int tid = omp_get_thread_num();
+    int threads = omp_get_num_threads();
+
     LhsScalar* blockA = blocking.blockA();
     eigen_internal_assert(blockA!=0);
-    
+
     std::size_t sizeB = kc*nc;
     ei_declare_aligned_stack_constructed_variable(RhsScalar, blockB, sizeB, 0);
-      
+
     // For each horizontal panel of the rhs, and corresponding vertical panel of the lhs...
     for(Index k=0; k<depth; k+=kc)
     {
@@ -114,11 +114,11 @@ static void run(Index rows, Index cols, Index depth,
 
       // Notify the other threads that the part A'_i is ready to go.
       info[tid].sync = k;
-      
+
       // Computes C_i += A' * B' per A'_i
-      for(Index shift=0; shift<threads; ++shift)
+      for(int shift=0; shift<threads; ++shift)
       {
-        Index i = (tid+shift)%threads;
+        int i = (tid+shift)%threads;
 
         // At this point we have to make sure that A'_i has been updated by the thread i,
         // we use testAndSetOrdered to mimic a volatile access.
@@ -161,7 +161,7 @@ static void run(Index rows, Index cols, Index depth,
 
     ei_declare_aligned_stack_constructed_variable(LhsScalar, blockA, sizeA, blocking.blockA());
     ei_declare_aligned_stack_constructed_variable(RhsScalar, blockB, sizeB, blocking.blockB());
-    
+
     const bool pack_rhs_once = mc!=rows && kc==depth && nc==cols;
 
     // For each horizontal panel of the rhs, and corresponding panel of the lhs...
@@ -172,24 +172,24 @@ static void run(Index rows, Index cols, Index depth,
       for(Index k2=0; k2<depth; k2+=kc)
       {
         const Index actual_kc = (std::min)(k2+kc,depth)-k2;
-        
+
         // OK, here we have selected one horizontal panel of rhs and one vertical panel of lhs.
         // => Pack lhs's panel into a sequential chunk of memory (L2/L3 caching)
         // Note that this panel will be read as many times as the number of blocks in the rhs's
         // horizontal panel which is, in practice, a very low number.
         pack_lhs(blockA, lhs.getSubMapper(i2,k2), actual_kc, actual_mc);
-        
+
         // For each kc x nc block of the rhs's horizontal panel...
         for(Index j2=0; j2<cols; j2+=nc)
         {
           const Index actual_nc = (std::min)(j2+nc,cols)-j2;
-          
+
           // We pack the rhs's block into a sequential chunk of memory (L2 caching)
           // Note that this block will be read a very high number of times, which is equal to the number of
           // micro horizontal panel of the large rhs's panel (e.g., rows/12 times).
           if((!pack_rhs_once) || i2==0)
             pack_rhs(blockB, rhs.getSubMapper(k2,j2), actual_kc, actual_nc);
-          
+
           // Everything is packed, we can now call the panel * block kernel:
           gebp(res.getSubMapper(i2, j2), blockA, blockB, actual_mc, actual_kc, actual_nc, alpha);
         }
@@ -229,7 +229,7 @@ struct gemm_functor
               (Scalar*)&(m_dest.coeffRef(row,col)), m_dest.outerStride(),
               m_actualAlpha, m_blocking, info);
   }
-  
+
   typedef typename Gemm::Traits Traits;
 
   protected:
@@ -313,7 +313,7 @@ class gemm_blocking_space<StorageOrder,_LhsScalar,_RhsScalar,MaxRows, MaxCols, M
       this->m_blockB = reinterpret_cast<RhsScalar*>((internal::UIntPtr(m_staticB) + (EIGEN_DEFAULT_ALIGN_BYTES-1)) & ~std::size_t(EIGEN_DEFAULT_ALIGN_BYTES-1));
 #endif
     }
-    
+
     void initParallel(Index, Index, Index, Index)
     {}
 
@@ -359,14 +359,14 @@ class gemm_blocking_space<StorageOrder,_LhsScalar,_RhsScalar,MaxRows, MaxCols, M
       m_sizeA = this->m_mc * this->m_kc;
       m_sizeB = this->m_kc * this->m_nc;
     }
-    
+
     void initParallel(Index rows, Index cols, Index depth, Index num_threads)
     {
       this->m_mc = Transpose ? cols : rows;
       this->m_nc = Transpose ? rows : cols;
       this->m_kc = depth;
-      
-      eigen_internal_assert(this->m_blockA==0 && this->m_blockB==0);      
+
+      eigen_internal_assert(this->m_blockA==0 && this->m_blockB==0);
       Index m = this->m_mc;
       computeProductBlockingSizes<LhsScalar,RhsScalar,KcFactor>(this->m_kc, m, this->m_nc, num_threads);
       m_sizeA = this->m_mc * this->m_kc;
@@ -401,7 +401,7 @@ class gemm_blocking_space<StorageOrder,_LhsScalar,_RhsScalar,MaxRows, MaxCols, M
 } // end namespace internal
 
 namespace internal {
-  
+
 template<typename Lhs, typename Rhs>
 struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,GemmProduct>
   : generic_product_impl_base<Lhs,Rhs,generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,GemmProduct> >
@@ -409,21 +409,21 @@ struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,GemmProduct>
   typedef typename Product<Lhs,Rhs>::Scalar Scalar;
   typedef typename Lhs::Scalar LhsScalar;
   typedef typename Rhs::Scalar RhsScalar;
-  
+
   typedef internal::blas_traits<Lhs> LhsBlasTraits;
   typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType;
   typedef typename internal::remove_all<ActualLhsType>::type ActualLhsTypeCleaned;
-  
+
   typedef internal::blas_traits<Rhs> RhsBlasTraits;
   typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType;
   typedef typename internal::remove_all<ActualRhsType>::type ActualRhsTypeCleaned;
-  
+
   enum {
     MaxDepthAtCompileTime = EIGEN_SIZE_MIN_PREFER_FIXED(Lhs::MaxColsAtCompileTime,Rhs::MaxRowsAtCompileTime)
   };
-  
+
   typedef generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,CoeffBasedProductMode> lazyproduct;
-  
+
   template<typename Dst>
   static void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
   {
@@ -453,7 +453,7 @@ struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,GemmProduct>
     else
       scaleAndAddTo(dst, lhs, rhs, Scalar(-1));
   }
-  
+
   template<typename Dest>
   static void scaleAndAddTo(Dest& dst, const Lhs& a_lhs, const Rhs& a_rhs, const Scalar& alpha)
   {
@@ -481,7 +481,7 @@ struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,GemmProduct>
 
     BlockingType blocking(dst.rows(), dst.cols(), lhs.cols(), 1, true);
     internal::parallelize_gemm<(Dest::MaxRowsAtCompileTime>32 || Dest::MaxRowsAtCompileTime==Dynamic)>
-                              (GemmFunctor(lhs, rhs, dst, actualAlpha, blocking), a_lhs.rows(), a_rhs.cols(), Dest::Flags&RowMajorBit);
+        (GemmFunctor(lhs, rhs, dst, actualAlpha, blocking), a_lhs.rows(), a_rhs.cols(), a_lhs.cols(), Dest::Flags&RowMajorBit);
   }
 };
 

Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h

@@ -148,7 +148,7 @@ struct tribb_kernel
     ResMapper res(_res, resStride);
     gebp_kernel<LhsScalar, RhsScalar, Index, ResMapper, mr, nr, ConjLhs, ConjRhs> gebp_kernel;
 
-    Matrix<ResScalar,BlockSize,BlockSize,ColMajor> buffer;
+    Matrix<ResScalar,BlockSize,BlockSize,ColMajor> buffer((internal::constructor_without_unaligned_array_assert()));
 
     // let's process the block per panel of actual_mc x BlockSize,
     // again, each is split into three parts, etc.
@@ -199,7 +199,7 @@ struct general_product_to_triangular_selector;
 template<typename MatrixType, typename ProductType, int UpLo>
 struct general_product_to_triangular_selector<MatrixType,ProductType,UpLo,true>
 {
-  static void run(MatrixType& mat, const ProductType& prod, const typename MatrixType::Scalar& alpha)
+  static void run(MatrixType& mat, const ProductType& prod, const typename MatrixType::Scalar& alpha, bool beta)
   {
     typedef typename MatrixType::Scalar Scalar;
     
@@ -217,6 +217,9 @@ struct general_product_to_triangular_selector<MatrixType,ProductType,UpLo,true>
 
     Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(prod.lhs().derived()) * RhsBlasTraits::extractScalarFactor(prod.rhs().derived());
 
+    if(!beta)
+      mat.template triangularView<UpLo>().setZero();
+
     enum {
       StorageOrder = (internal::traits<MatrixType>::Flags&RowMajorBit) ? RowMajor : ColMajor,
       UseLhsDirectly = _ActualLhs::InnerStrideAtCompileTime==1,
@@ -244,7 +247,7 @@ struct general_product_to_triangular_selector<MatrixType,ProductType,UpLo,true>
 template<typename MatrixType, typename ProductType, int UpLo>
 struct general_product_to_triangular_selector<MatrixType,ProductType,UpLo,false>
 {
-  static void run(MatrixType& mat, const ProductType& prod, const typename MatrixType::Scalar& alpha)
+  static void run(MatrixType& mat, const ProductType& prod, const typename MatrixType::Scalar& alpha, bool beta)
   {
     typedef typename internal::remove_all<typename ProductType::LhsNested>::type Lhs;
     typedef internal::blas_traits<Lhs> LhsBlasTraits;
@@ -260,6 +263,9 @@ struct general_product_to_triangular_selector<MatrixType,ProductType,UpLo,false>
 
     typename ProductType::Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(prod.lhs().derived()) * RhsBlasTraits::extractScalarFactor(prod.rhs().derived());
 
+    if(!beta)
+      mat.template triangularView<UpLo>().setZero();
+
     enum {
       IsRowMajor = (internal::traits<MatrixType>::Flags&RowMajorBit) ? 1 : 0,
       LhsIsRowMajor = _ActualLhs::Flags&RowMajorBit ? 1 : 0,
@@ -286,11 +292,11 @@ struct general_product_to_triangular_selector<MatrixType,ProductType,UpLo,false>
 
 template<typename MatrixType, unsigned int UpLo>
 template<typename ProductType>
-TriangularView<MatrixType,UpLo>& TriangularViewImpl<MatrixType,UpLo,Dense>::_assignProduct(const ProductType& prod, const Scalar& alpha)
+TriangularView<MatrixType,UpLo>& TriangularViewImpl<MatrixType,UpLo,Dense>::_assignProduct(const ProductType& prod, const Scalar& alpha, bool beta)
 {
   eigen_assert(derived().nestedExpression().rows() == prod.rows() && derived().cols() == prod.cols());
   
-  general_product_to_triangular_selector<MatrixType, ProductType, UpLo, internal::traits<ProductType>::InnerSize==1>::run(derived().nestedExpression().const_cast_derived(), prod, alpha);
+  general_product_to_triangular_selector<MatrixType, ProductType, UpLo, internal::traits<ProductType>::InnerSize==1>::run(derived().nestedExpression().const_cast_derived(), prod, alpha, beta);
   
   return derived();
 }

Eigen/src/Core/products/GeneralMatrixMatrixTriangular_BLAS.h

@@ -33,7 +33,7 @@
 #ifndef EIGEN_GENERAL_MATRIX_MATRIX_TRIANGULAR_BLAS_H
 #define EIGEN_GENERAL_MATRIX_MATRIX_TRIANGULAR_BLAS_H
 
-namespace Eigen { 
+namespace Eigen {
 
 namespace internal {
 
@@ -86,8 +86,8 @@ struct general_matrix_matrix_rankupdate<Index,EIGTYPE,AStorageOrder,ConjugateA,C
   /* typedef Matrix<EIGTYPE, Dynamic, Dynamic, RhsStorageOrder> MatrixRhs;*/ \
 \
    BlasIndex lda=convert_index<BlasIndex>(lhsStride), ldc=convert_index<BlasIndex>(resStride), n=convert_index<BlasIndex>(size), k=convert_index<BlasIndex>(depth); \
-   char uplo=(IsLower) ? 'L' : 'U', trans=(AStorageOrder==RowMajor) ? 'T':'N'; \
-   EIGTYPE beta; \
+   char uplo=((IsLower) ? 'L' : 'U'), trans=((AStorageOrder==RowMajor) ? 'T':'N'); \
+   EIGTYPE beta(1); \
    BLASFUNC(&uplo, &trans, &n, &k, &numext::real_ref(alpha), lhs, &lda, &numext::real_ref(beta), res, &ldc); \
   } \
 };
@@ -107,7 +107,7 @@ struct general_matrix_matrix_rankupdate<Index,EIGTYPE,AStorageOrder,ConjugateA,C
    typedef Matrix<EIGTYPE, Dynamic, Dynamic, AStorageOrder> MatrixType; \
 \
    BlasIndex lda=convert_index<BlasIndex>(lhsStride), ldc=convert_index<BlasIndex>(resStride), n=convert_index<BlasIndex>(size), k=convert_index<BlasIndex>(depth); \
-   char uplo=(IsLower) ? 'L' : 'U', trans=(AStorageOrder==RowMajor) ? 'C':'N'; \
+   char uplo=((IsLower) ? 'L' : 'U'), trans=((AStorageOrder==RowMajor) ? 'C':'N'); \
    RTYPE alpha_, beta_; \
    const EIGTYPE* a_ptr; \
 \

Eigen/src/Core/products/GeneralMatrixVector.h

@@ -183,8 +183,8 @@ EIGEN_DONT_INLINE void general_matrix_vector_product<Index,LhsScalar,LhsMapper,C
     alignmentPattern = AllAligned;
   }
 
-  const Index offset1 = (FirstAligned && alignmentStep==1?3:1);
-  const Index offset3 = (FirstAligned && alignmentStep==1?1:3);
+  const Index offset1 = (FirstAligned && alignmentStep==1)?3:1;
+  const Index offset3 = (FirstAligned && alignmentStep==1)?1:3;
 
   Index columnBound = ((cols-skipColumns)/columnsAtOnce)*columnsAtOnce + skipColumns;
   for (Index i=skipColumns; i<columnBound; i+=columnsAtOnce)
@@ -457,8 +457,8 @@ EIGEN_DONT_INLINE void general_matrix_vector_product<Index,LhsScalar,LhsMapper,R
     alignmentPattern = AllAligned;
   }
 
-  const Index offset1 = (FirstAligned && alignmentStep==1?3:1);
-  const Index offset3 = (FirstAligned && alignmentStep==1?1:3);
+  const Index offset1 = (FirstAligned && alignmentStep==1)?3:1;
+  const Index offset3 = (FirstAligned && alignmentStep==1)?1:3;
 
   Index rowBound = ((rows-skipRows)/rowsAtOnce)*rowsAtOnce + skipRows;
   for (Index i=skipRows; i<rowBound; i+=rowsAtOnce)

Eigen/src/Core/products/Parallelizer.h

@@ -10,7 +10,7 @@
 #ifndef EIGEN_PARALLELIZER_H
 #define EIGEN_PARALLELIZER_H
 
-namespace Eigen { 
+namespace Eigen {
 
 namespace internal {
 
@@ -75,7 +75,7 @@ template<typename Index> struct GemmParallelInfo
 {
   GemmParallelInfo() : sync(-1), users(0), lhs_start(0), lhs_length(0) {}
 
-  int volatile sync;
+  Index volatile sync;
   int volatile users;
 
   Index lhs_start;
@@ -83,7 +83,7 @@ template<typename Index> struct GemmParallelInfo
 };
 
 template<bool Condition, typename Functor, typename Index>
-void parallelize_gemm(const Functor& func, Index rows, Index cols, bool transpose)
+void parallelize_gemm(const Functor& func, Index rows, Index cols, Index depth, bool transpose)
 {
   // TODO when EIGEN_USE_BLAS is defined,
   // we should still enable OMP for other scalar types
@@ -92,6 +92,7 @@ void parallelize_gemm(const Functor& func, Index rows, Index cols, bool transpos
   // the matrix product when multithreading is enabled. This is a temporary
   // fix to support row-major destination matrices. This whole
   // parallelizer mechanism has to be redisigned anyway.
+  EIGEN_UNUSED_VARIABLE(depth);
   EIGEN_UNUSED_VARIABLE(transpose);
   func(0,rows, 0,cols);
 #else
@@ -103,9 +104,16 @@ void parallelize_gemm(const Functor& func, Index rows, Index cols, bool transpos
   // - the sizes are large enough
 
   // compute the maximal number of threads from the size of the product:
-  // FIXME this has to be fine tuned
+  // This first heuristic takes into account that the product kernel is fully optimized when working with nr columns at once.
   Index size = transpose ? rows : cols;
-  Index pb_max_threads = std::max<Index>(1,size / 32);
+  Index pb_max_threads = std::max<Index>(1,size / Functor::Traits::nr);
+
+  // compute the maximal number of threads from the total amount of work:
+  double work = static_cast<double>(rows) * static_cast<double>(cols) *
+      static_cast<double>(depth);
+  double kMinTaskSize = 50000;  // FIXME improve this heuristic.
+  pb_max_threads = std::max<Index>(1, std::min<Index>(pb_max_threads, work / kMinTaskSize));
+
   // compute the number of threads we are going to use
   Index threads = std::min<Index>(nbThreads(), pb_max_threads);
 
@@ -120,19 +128,19 @@ void parallelize_gemm(const Functor& func, Index rows, Index cols, bool transpos
 
   if(transpose)
     std::swap(rows,cols);
-  
+
   ei_declare_aligned_stack_constructed_variable(GemmParallelInfo<Index>,info,threads,0);
-  
+
   #pragma omp parallel num_threads(threads)
   {
     Index i = omp_get_thread_num();
     // Note that the actual number of threads might be lower than the number of request ones.
     Index actual_threads = omp_get_num_threads();
-    
+
     Index blockCols = (cols / actual_threads) & ~Index(0x3);
     Index blockRows = (rows / actual_threads);
     blockRows = (blockRows/Functor::Traits::mr)*Functor::Traits::mr;
-  
+
     Index r0 = i*blockRows;
     Index actualBlockRows = (i+1==actual_threads) ? rows-r0 : blockRows;
 

Eigen/src/Core/products/SelfadjointMatrixVector.h

@@ -83,10 +83,10 @@ EIGEN_DONT_INLINE void selfadjoint_matrix_vector_product<Scalar,Index,StorageOrd
     Scalar t3(0);
     Packet ptmp3 = pset1<Packet>(t3);
 
-    size_t starti = FirstTriangular ? 0 : j+2;
-    size_t endi   = FirstTriangular ? j : size;
-    size_t alignedStart = (starti) + internal::first_default_aligned(&res[starti], endi-starti);
-    size_t alignedEnd = alignedStart + ((endi-alignedStart)/(PacketSize))*(PacketSize);
+    Index starti = FirstTriangular ? 0 : j+2;
+    Index endi   = FirstTriangular ? j : size;
+    Index alignedStart = (starti) + internal::first_default_aligned(&res[starti], endi-starti);
+    Index alignedEnd = alignedStart + ((endi-alignedStart)/(PacketSize))*(PacketSize);
 
     res[j]   += cjd.pmul(numext::real(A0[j]), t0);
     res[j+1] += cjd.pmul(numext::real(A1[j+1]), t1);
@@ -101,7 +101,7 @@ EIGEN_DONT_INLINE void selfadjoint_matrix_vector_product<Scalar,Index,StorageOrd
       t2 += cj1.pmul(A0[j+1], rhs[j+1]);
     }
 
-    for (size_t i=starti; i<alignedStart; ++i)
+    for (Index i=starti; i<alignedStart; ++i)
     {
       res[i] += cj0.pmul(A0[i], t0) + cj0.pmul(A1[i],t1);
       t2 += cj1.pmul(A0[i], rhs[i]);
@@ -113,7 +113,7 @@ EIGEN_DONT_INLINE void selfadjoint_matrix_vector_product<Scalar,Index,StorageOrd
     const Scalar* EIGEN_RESTRICT a1It  = A1  + alignedStart;
     const Scalar* EIGEN_RESTRICT rhsIt = rhs + alignedStart;
           Scalar* EIGEN_RESTRICT resIt = res + alignedStart;
-    for (size_t i=alignedStart; i<alignedEnd; i+=PacketSize)
+    for (Index i=alignedStart; i<alignedEnd; i+=PacketSize)
     {
       Packet A0i = ploadu<Packet>(a0It);  a0It  += PacketSize;
       Packet A1i = ploadu<Packet>(a1It);  a1It  += PacketSize;
@@ -125,7 +125,7 @@ EIGEN_DONT_INLINE void selfadjoint_matrix_vector_product<Scalar,Index,StorageOrd
       ptmp3 = pcj1.pmadd(A1i,  Bi, ptmp3);
       pstore(resIt,Xi); resIt += PacketSize;
     }
-    for (size_t i=alignedEnd; i<endi; i++)
+    for (Index i=alignedEnd; i<endi; i++)
     {
       res[i] += cj0.pmul(A0[i], t0) + cj0.pmul(A1[i],t1);
       t2 += cj1.pmul(A0[i], rhs[i]);
@@ -179,7 +179,7 @@ struct selfadjoint_product_impl<Lhs,LhsMode,false,Rhs,0,true>
   {
     typedef typename Dest::Scalar ResScalar;
     typedef typename Rhs::Scalar RhsScalar;
-    typedef Map<Matrix<ResScalar,Dynamic,1>, Aligned> MappedDest;
+    typedef Map<Matrix<ResScalar,Dynamic,1>, EIGEN_PLAIN_ENUM_MIN(AlignedMax,internal::packet_traits<ResScalar>::size)> MappedDest;
     
     eigen_assert(dest.rows()==a_lhs.rows() && dest.cols()==a_rhs.cols());
 

Eigen/src/Core/products/TriangularMatrixMatrix.h

@@ -137,7 +137,7 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,true,
     ei_declare_aligned_stack_constructed_variable(Scalar, blockA, sizeA, blocking.blockA());
     ei_declare_aligned_stack_constructed_variable(Scalar, blockB, sizeB, blocking.blockB());
 
-    Matrix<Scalar,SmallPanelWidth,SmallPanelWidth,LhsStorageOrder> triangularBuffer;
+    Matrix<Scalar,SmallPanelWidth,SmallPanelWidth,LhsStorageOrder> triangularBuffer((internal::constructor_without_unaligned_array_assert()));
     triangularBuffer.setZero();
     if((Mode&ZeroDiag)==ZeroDiag)
       triangularBuffer.diagonal().setZero();
@@ -284,7 +284,7 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,false,
     ei_declare_aligned_stack_constructed_variable(Scalar, blockA, sizeA, blocking.blockA());
     ei_declare_aligned_stack_constructed_variable(Scalar, blockB, sizeB, blocking.blockB());
 
-    Matrix<Scalar,SmallPanelWidth,SmallPanelWidth,RhsStorageOrder> triangularBuffer;
+    Matrix<Scalar,SmallPanelWidth,SmallPanelWidth,RhsStorageOrder> triangularBuffer((internal::constructor_without_unaligned_array_assert()));
     triangularBuffer.setZero();
     if((Mode&ZeroDiag)==ZeroDiag)
       triangularBuffer.diagonal().setZero();

Eigen/src/Core/products/TriangularMatrixVector.h

@@ -216,7 +216,7 @@ template<int Mode> struct trmv_selector<Mode,ColMajor>
     typedef internal::blas_traits<Rhs> RhsBlasTraits;
     typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType;
     
-    typedef Map<Matrix<ResScalar,Dynamic,1>, Aligned> MappedDest;
+    typedef Map<Matrix<ResScalar,Dynamic,1>, EIGEN_PLAIN_ENUM_MIN(AlignedMax,internal::packet_traits<ResScalar>::size)> MappedDest;
 
     typename internal::add_const_on_value_type<ActualLhsType>::type actualLhs = LhsBlasTraits::extract(lhs);
     typename internal::add_const_on_value_type<ActualRhsType>::type actualRhs = RhsBlasTraits::extract(rhs);

Eigen/src/Core/products/TriangularSolverMatrix.h

@@ -183,7 +183,7 @@ EIGEN_DONT_INLINE void triangular_solve_matrix<Scalar,Index,OnTheLeft,Mode,Conju
     }
   }
 
-/* Optimized triangular solver with multiple left hand sides and the trinagular matrix on the right
+/* Optimized triangular solver with multiple left hand sides and the triangular matrix on the right
  */
 template <typename Scalar, typename Index, int Mode, bool Conjugate, int TriStorageOrder>
 struct triangular_solve_matrix<Scalar,Index,OnTheRight,Mode,Conjugate,TriStorageOrder,ColMajor>
@@ -202,6 +202,7 @@ EIGEN_DONT_INLINE void triangular_solve_matrix<Scalar,Index,OnTheRight,Mode,Conj
     level3_blocking<Scalar,Scalar>& blocking)
   {
     Index rows = otherSize;
+    typedef typename NumTraits<Scalar>::Real RealScalar;
 
     typedef blas_data_mapper<Scalar, Index, ColMajor> LhsMapper;
     typedef const_blas_data_mapper<Scalar, Index, TriStorageOrder> RhsMapper;
@@ -306,9 +307,9 @@ EIGEN_DONT_INLINE void triangular_solve_matrix<Scalar,Index,OnTheRight,Mode,Conj
               }
               if((Mode & UnitDiag)==0)
               {
-                Scalar b = conj(rhs(j,j));
+                Scalar inv_rjj = RealScalar(1)/conj(rhs(j,j));
                 for (Index i=0; i<actual_mc; ++i)
-                  r[i] /= b;
+                  r[i] *= inv_rjj;
               }
             }
 

Eigen/src/Core/util/BlasUtil.h

@@ -44,16 +44,29 @@ template<bool Conjugate> struct conj_if;
 
 template<> struct conj_if<true> {
   template<typename T>
-  inline T operator()(const T& x) { return numext::conj(x); }
+  inline T operator()(const T& x) const { return numext::conj(x); }
   template<typename T>
-  inline T pconj(const T& x) { return internal::pconj(x); }
+  inline T pconj(const T& x) const { return internal::pconj(x); }
 };
 
 template<> struct conj_if<false> {
   template<typename T>
-  inline const T& operator()(const T& x) { return x; }
+  inline const T& operator()(const T& x) const { return x; }
   template<typename T>
-  inline const T& pconj(const T& x) { return x; }
+  inline const T& pconj(const T& x) const { return x; }
+};
+
+// Generic implementation for custom complex types.
+template<typename LhsScalar, typename RhsScalar, bool ConjLhs, bool ConjRhs>
+struct conj_helper
+{
+  typedef typename ScalarBinaryOpTraits<LhsScalar,RhsScalar>::ReturnType Scalar;
+
+  EIGEN_STRONG_INLINE Scalar pmadd(const LhsScalar& x, const RhsScalar& y, const Scalar& c) const
+  { return padd(c, pmul(x,y)); }
+
+  EIGEN_STRONG_INLINE Scalar pmul(const LhsScalar& x, const RhsScalar& y) const
+  { return conj_if<ConjLhs>()(x) *  conj_if<ConjRhs>()(y); }
 };
 
 template<typename Scalar> struct conj_helper<Scalar,Scalar,false,false>
@@ -315,6 +328,11 @@ struct blas_traits<CwiseBinaryOp<scalar_product_op<Scalar>, NestedXpr, const Cwi
   static inline Scalar extractScalarFactor(const XprType& x)
   { return Base::extractScalarFactor(x.lhs()) * x.rhs().functor().m_other; }
 };
+template<typename Scalar, typename Plain1, typename Plain2>
+struct blas_traits<CwiseBinaryOp<scalar_product_op<Scalar>, const CwiseNullaryOp<scalar_constant_op<Scalar>,Plain1>,
+                                                            const CwiseNullaryOp<scalar_constant_op<Scalar>,Plain2> > >
+ : blas_traits<CwiseNullaryOp<scalar_constant_op<Scalar>,Plain1> >
+{};
 
 // pop opposite
 template<typename Scalar, typename NestedXpr>

Eigen/src/Core/util/CMakeLists.txt

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

Eigen/src/Core/util/DisableStupidWarnings.h

@@ -14,12 +14,13 @@
   // 4512 - assignment operator could not be generated
   // 4522 - 'class' : multiple assignment operators specified
   // 4700 - uninitialized local variable 'xyz' used
+  // 4714 - function marked as __forceinline not inlined
   // 4717 - 'function' : recursive on all control paths, function will cause runtime stack overflow
   // 4800 - 'type' : forcing value to bool 'true' or 'false' (performance warning)
   #ifndef EIGEN_PERMANENTLY_DISABLE_STUPID_WARNINGS
     #pragma warning( push )
   #endif
-  #pragma warning( disable : 4100 4101 4127 4181 4211 4244 4273 4324 4503 4512 4522 4700 4717 4800)
+  #pragma warning( disable : 4100 4101 4127 4181 4211 4244 4273 4324 4503 4512 4522 4700 4714 4717 4800)
 
 #elif defined __INTEL_COMPILER
   // 2196 - routine is both "inline" and "noinline" ("noinline" assumed)
@@ -56,13 +57,19 @@
   #pragma diag_suppress code_is_unreachable
   // Disable the "dynamic initialization in unreachable code" message
   #pragma diag_suppress initialization_not_reachable
-  // Disable the "calling a __host__ function from a __host__ __device__ function is not allowed" messages (yes, there are 4 of them)
+  // Disable the "invalid error number" message that we get with older versions of nvcc
+  #pragma diag_suppress 1222
+  // Disable the "calling a __host__ function from a __host__ __device__ function is not allowed" messages (yes, there are many of them and they seem to change with every version of the compiler)
+  #pragma diag_suppress 2527
+  #pragma diag_suppress 2529
   #pragma diag_suppress 2651
   #pragma diag_suppress 2653
   #pragma diag_suppress 2668
   #pragma diag_suppress 2669
   #pragma diag_suppress 2670
   #pragma diag_suppress 2671
+  #pragma diag_suppress 2735
+  #pragma diag_suppress 2737
 #endif
 
 #endif // not EIGEN_WARNINGS_DISABLED