bump

- Add support for Upper or Lower inputs.
- Add supports for sparse RHS
- Remove transposed cases, remove ordering method interface
- Add full access to PARDISO parameters

.hgeol

@@ -1,3 +1,8 @@
-[patterns]
-**.* = native
-eigen_autoexp_part.dat = CRLF
+[patterns]
+scripts/*.in = LF
+debug/msvc/*.dat = CRLF
+unsupported/test/mpreal/*.* = CRLF
+** = native
+
+[repository]
+native = LF

CMakeLists.txt

@@ -64,6 +64,10 @@ set(CMAKE_MODULE_PATH ${PROJECT_SOURCE_DIR}/cmake)
 
 find_package(StandardMathLibrary)
 
+
+set(EIGEN_TEST_CUSTOM_LINKER_FLAGS  "" CACHE STRING "Additional linker flags when linking unit tests.")
+set(EIGEN_TEST_CUSTOM_CXX_FLAGS     "" CACHE STRING "Additional compiler flags when compiling unit tests.")
+
 set(EIGEN_STANDARD_LIBRARIES_TO_LINK_TO "")
 
 if(NOT STANDARD_MATH_LIBRARY_FOUND)
@@ -103,6 +107,8 @@ endif()
 
 add_definitions("-DEIGEN_PERMANENTLY_DISABLE_STUPID_WARNINGS")
 
+set(EIGEN_TEST_MAX_SIZE "320" CACHE STRING "Maximal matrix/vector size, default is 320")
+
 if(CMAKE_COMPILER_IS_GNUCXX)
   set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wnon-virtual-dtor -Wno-long-long -ansi -Wundef -Wcast-align -Wchar-subscripts -Wall -W -Wpointer-arith -Wwrite-strings -Wformat-security -fexceptions -fno-check-new -fno-common -fstrict-aliasing")
   set(CMAKE_CXX_FLAGS_DEBUG "-g3")
@@ -301,44 +307,9 @@ add_subdirectory(Eigen)
 
 add_subdirectory(doc EXCLUDE_FROM_ALL)
 
-add_custom_target(buildtests)
-add_custom_target(check COMMAND "ctest")
-add_dependencies(check buildtests)
-
-# CMake/Ctest does not allow us to change the build command,
-# so we have to workaround by directly editing the generated DartConfiguration.tcl file
-# save CMAKE_MAKE_PROGRAM
-set(CMAKE_MAKE_PROGRAM_SAVE ${CMAKE_MAKE_PROGRAM})
-# and set a fake one
-set(CMAKE_MAKE_PROGRAM "@EIGEN_MAKECOMMAND_PLACEHOLDER@")
-
-include(CTest)
+include(EigenConfigureTesting)
+# fixme, not sure this line is still needed:
 enable_testing() # must be called from the root CMakeLists, see man page
-include(EigenTesting)
-ei_init_testing()
-
-# overwrite default DartConfiguration.tcl
-# The worarounds are different for each version of the MSVC IDE
-if(MSVC_IDE)
-  if(MSVC_VERSION EQUAL 1600) # MSVC 2010
-    set(EIGEN_MAKECOMMAND_PLACEHOLDER "${CMAKE_MAKE_PROGRAM_SAVE} buildtests.vcxproj /p:Configuration=\${CTEST_CONFIGURATION_TYPE} \n # ")
-  else() # MSVC 2008 (TODO check MSVC 2005)
-    set(EIGEN_MAKECOMMAND_PLACEHOLDER "${CMAKE_MAKE_PROGRAM_SAVE} /project buildtests")
-  endif()
-else()
-  # for make and nmake
-  set(EIGEN_MAKECOMMAND_PLACEHOLDER "${CMAKE_MAKE_PROGRAM_SAVE} buildtests")
-endif()
-
-configure_file(${CMAKE_BINARY_DIR}/DartConfiguration.tcl ${CMAKE_BINARY_DIR}/DartConfiguration.tcl)
-# restore default CMAKE_MAKE_PROGRAM
-set(CMAKE_MAKE_PROGRAM ${CMAKE_MAKE_PROGRAM_SAVE})
-# un-set temporary variables so that it is like they never existed.
-# CMake 2.6.3 introduces the more logical unset() syntax for this.
-set(CMAKE_MAKE_PROGRAM_SAVE)
-set(EIGEN_MAKECOMMAND_PLACEHOLDER)
-
-configure_file(${CMAKE_SOURCE_DIR}/CTestCustom.cmake.in ${CMAKE_BINARY_DIR}/CTestCustom.cmake)
 
 
 if(EIGEN_LEAVE_TEST_IN_ALL_TARGET)
@@ -347,15 +318,13 @@ else()
   add_subdirectory(test EXCLUDE_FROM_ALL)
 endif()
 
-if(NOT MSVC)
-  if(EIGEN_LEAVE_TEST_IN_ALL_TARGET)
-    add_subdirectory(blas)
-    add_subdirectory(lapack)
-  else()
-    add_subdirectory(blas EXCLUDE_FROM_ALL)
-    add_subdirectory(lapack EXCLUDE_FROM_ALL)
-  endif()
-endif(NOT MSVC)
+if(EIGEN_LEAVE_TEST_IN_ALL_TARGET)
+  add_subdirectory(blas)
+  add_subdirectory(lapack)
+else()
+  add_subdirectory(blas EXCLUDE_FROM_ALL)
+  add_subdirectory(lapack EXCLUDE_FROM_ALL)
+endif()
 
 add_subdirectory(unsupported)
 

COPYING.BSD

@@ -0,0 +1,26 @@
+/*
+ Copyright (c) 2011, Intel Corporation. All rights reserved.
+
+ Redistribution and use in source and binary forms, with or without modification,
+ are permitted provided that the following conditions are met:
+
+ * Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+ * Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+ * Neither the name of Intel Corporation nor the names of its contributors may
+   be used to endorse or promote products derived from this software without
+   specific prior written permission.
+
+ THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+ WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+ DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+ (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+ LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/

Eigen/Cholesky

@@ -24,6 +24,9 @@ namespace Eigen {
 #include "src/misc/Solve.h"
 #include "src/Cholesky/LLT.h"
 #include "src/Cholesky/LDLT.h"
+#ifdef EIGEN_USE_LAPACKE
+#include "src/Cholesky/LLT_MKL.h"
+#endif
 
 } // namespace Eigen
 

Eigen/CholmodSupport

@@ -0,0 +1,34 @@
+#ifndef EIGEN_CHOLMODSUPPORT_MODULE_H
+#define EIGEN_CHOLMODSUPPORT_MODULE_H
+
+#include "SparseCore"
+
+#include "src/Core/util/DisableStupidWarnings.h"
+
+extern "C" {
+  #include <cholmod.h>
+}
+
+namespace Eigen {
+
+/** \ingroup Support_modules
+  * \defgroup CholmodSupport_Module CholmodSupport module
+  *
+  *
+  * \code
+  * #include <Eigen/CholmodSupport>
+  * \endcode
+  */
+
+#include "src/misc/Solve.h"
+#include "src/misc/SparseSolve.h"
+
+#include "src/CholmodSupport/CholmodSupport.h"
+
+
+} // namespace Eigen
+
+#include "src/Core/util/ReenableStupidWarnings.h"
+
+#endif // EIGEN_CHOLMODSUPPORT_MODULE_H
+

Eigen/Core

@@ -34,6 +34,10 @@
 // defined e.g. EIGEN_DONT_ALIGN) so it needs to be done before we do anything with vectorization.
 #include "src/Core/util/Macros.h"
 
+// this include file manages BLAS and MKL related macros
+// and inclusion of their respective header files
+#include "src/Core/util/MKL_support.h"
+
 // if alignment is disabled, then disable vectorization. Note: EIGEN_ALIGN is the proper check, it takes into
 // account both the user's will (EIGEN_DONT_ALIGN) and our own platform checks
 #if !EIGEN_ALIGN
@@ -175,9 +179,6 @@
   #include <new>
 #endif
 
-// defined in bits/termios.h
-#undef B0
-
 /** \brief Namespace containing all symbols from the %Eigen library. */
 namespace Eigen {
 
@@ -247,6 +248,10 @@ using std::ptrdiff_t;
   * \endcode
   */
 
+/** \defgroup Support_modules Support modules [category]
+  * Category of modules which add support for external libraries.
+  */
+
 #include "src/Core/util/Constants.h"
 #include "src/Core/util/ForwardDeclarations.h"
 #include "src/Core/util/Meta.h"
@@ -318,7 +323,7 @@ using std::ptrdiff_t;
 #include "src/Core/CommaInitializer.h"
 #include "src/Core/Flagged.h"
 #include "src/Core/ProductBase.h"
-#include "src/Core/Product.h"
+#include "src/Core/GeneralProduct.h"
 #include "src/Core/TriangularMatrix.h"
 #include "src/Core/SelfAdjointView.h"
 #include "src/Core/SolveTriangular.h"
@@ -347,6 +352,21 @@ using std::ptrdiff_t;
 #include "src/Core/ArrayBase.h"
 #include "src/Core/ArrayWrapper.h"
 
+#ifdef EIGEN_USE_BLAS
+#include "src/Core/products/GeneralMatrixMatrix_MKL.h"
+#include "src/Core/products/GeneralMatrixVector_MKL.h"
+#include "src/Core/products/GeneralMatrixMatrixTriangular_MKL.h"
+#include "src/Core/products/SelfadjointMatrixMatrix_MKL.h"
+#include "src/Core/products/SelfadjointMatrixVector_MKL.h"
+#include "src/Core/products/TriangularMatrixMatrix_MKL.h"
+#include "src/Core/products/TriangularMatrixVector_MKL.h"
+#include "src/Core/products/TriangularSolverMatrix_MKL.h"
+#endif // EIGEN_USE_BLAS
+
+#ifdef EIGEN_USE_MKL_VML
+#include "src/Core/Assign_MKL.h"
+#endif
+
 } // namespace Eigen
 
 #include "src/Core/GlobalFunctions.h"

Eigen/Eigen2Support

@@ -33,7 +33,8 @@
 
 namespace Eigen {
 
-/** \defgroup Eigen2Support_Module Eigen2 support module
+/** \ingroup Support_modules
+  * \defgroup Eigen2Support_Module Eigen2 support module
   * This module provides a couple of deprecated functions improving the compatibility with Eigen2.
   *
   * To use it, define EIGEN2_SUPPORT before including any Eigen header
@@ -63,6 +64,24 @@ namespace Eigen {
 // Eigen2 used to include iostream
 #include<iostream>
 
+#define EIGEN_USING_MATRIX_TYPEDEFS_FOR_TYPE_AND_SIZE(TypeSuffix, SizeSuffix) \
+using Eigen::Matrix##SizeSuffix##TypeSuffix; \
+using Eigen::Vector##SizeSuffix##TypeSuffix; \
+using Eigen::RowVector##SizeSuffix##TypeSuffix;
+
+#define EIGEN_USING_MATRIX_TYPEDEFS_FOR_TYPE(TypeSuffix) \
+EIGEN_USING_MATRIX_TYPEDEFS_FOR_TYPE_AND_SIZE(TypeSuffix, 2) \
+EIGEN_USING_MATRIX_TYPEDEFS_FOR_TYPE_AND_SIZE(TypeSuffix, 3) \
+EIGEN_USING_MATRIX_TYPEDEFS_FOR_TYPE_AND_SIZE(TypeSuffix, 4) \
+EIGEN_USING_MATRIX_TYPEDEFS_FOR_TYPE_AND_SIZE(TypeSuffix, X) \
+
+#define EIGEN_USING_MATRIX_TYPEDEFS \
+EIGEN_USING_MATRIX_TYPEDEFS_FOR_TYPE(i) \
+EIGEN_USING_MATRIX_TYPEDEFS_FOR_TYPE(f) \
+EIGEN_USING_MATRIX_TYPEDEFS_FOR_TYPE(d) \
+EIGEN_USING_MATRIX_TYPEDEFS_FOR_TYPE(cf) \
+EIGEN_USING_MATRIX_TYPEDEFS_FOR_TYPE(cd)
+
 #define USING_PART_OF_NAMESPACE_EIGEN \
 EIGEN_USING_MATRIX_TYPEDEFS \
 using Eigen::Matrix; \

Eigen/Eigenvalues

@@ -9,6 +9,7 @@
 #include "Jacobi"
 #include "Householder"
 #include "LU"
+#include "Geometry"
 
 namespace Eigen {
 
@@ -35,6 +36,11 @@ namespace Eigen {
 #include "src/Eigenvalues/ComplexSchur.h"
 #include "src/Eigenvalues/ComplexEigenSolver.h"
 #include "src/Eigenvalues/MatrixBaseEigenvalues.h"
+#ifdef EIGEN_USE_LAPACKE
+#include "src/Eigenvalues/RealSchur_MKL.h"
+#include "src/Eigenvalues/ComplexSchur_MKL.h"
+#include "src/Eigenvalues/SelfAdjointEigenSolver_MKL.h"
+#endif
 
 } // namespace Eigen
 

Eigen/IterativeLinearSolvers

@@ -0,0 +1,37 @@
+#ifndef EIGEN_ITERATIVELINEARSOLVERS_MODULE_H
+#define EIGEN_ITERATIVELINEARSOLVERS_MODULE_H
+
+#include "SparseCore"
+
+#include "src/Core/util/DisableStupidWarnings.h"
+
+namespace Eigen {
+
+/** \ingroup Sparse_modules
+  * \defgroup IterativeLinearSolvers_Module IterativeLinearSolvers module
+  *
+  * This module currently provides iterative methods to solve problems of the form \c A \c x = \c b, where \c A is a squared matrix, usually very large and sparse.
+  * Those solvers are accessible via the following classes:
+  *  - ConjugateGradient for selfadjoint (hermitian) matrices,
+  *  - BiCGSTAB for general square matrices.
+  *
+  * Such problems can also be solved using the direct sparse decomposition modules: SparseCholesky, CholmodSupport, UmfPackSupport, SuperLUSupport.
+  *
+  * \code
+  * #include <Eigen/IterativeLinearSolvers>
+  * \endcode
+  */
+
+#include "src/misc/Solve.h"
+#include "src/misc/SparseSolve.h"
+
+#include "src/IterativeLinearSolvers/IterativeSolverBase.h"
+#include "src/IterativeLinearSolvers/BasicPreconditioners.h"
+#include "src/IterativeLinearSolvers/ConjugateGradient.h"
+#include "src/IterativeLinearSolvers/BiCGSTAB.h"
+
+} // namespace Eigen
+
+#include "src/Core/util/ReenableStupidWarnings.h"
+
+#endif // EIGEN_ITERATIVELINEARSOLVERS_MODULE_H

Eigen/LU

@@ -23,6 +23,9 @@ namespace Eigen {
 #include "src/misc/Image.h"
 #include "src/LU/FullPivLU.h"
 #include "src/LU/PartialPivLU.h"
+#ifdef EIGEN_USE_LAPACKE
+#include "src/LU/PartialPivLU_MKL.h"
+#endif
 #include "src/LU/Determinant.h"
 #include "src/LU/Inverse.h"
 

Eigen/OrderingMethods

@@ -0,0 +1,27 @@
+#ifndef EIGEN_ORDERINGMETHODS_MODULE_H
+#define EIGEN_ORDERINGMETHODS_MODULE_H
+
+#include "SparseCore"
+
+#include "src/Core/util/DisableStupidWarnings.h"
+
+namespace Eigen {
+
+/** \ingroup Sparse_modules
+  * \defgroup OrderingMethods_Module OrderingMethods module
+  *
+  * This module is currently for internal use only.
+  *
+  *
+  * \code
+  * #include <Eigen/OrderingMethods>
+  * \endcode
+  */
+
+#include "src/OrderingMethods/Amd.h"
+
+} // namespace Eigen
+
+#include "src/Core/util/ReenableStupidWarnings.h"
+
+#endif // EIGEN_ORDERINGMETHODS_MODULE_H

Eigen/PARDISOSupport

@@ -0,0 +1,30 @@
+#ifndef EIGEN_PARDISOSUPPORT_MODULE_H
+#define EIGEN_PARDISOSUPPORT_MODULE_H
+
+#include "SparseCore"
+
+#include "src/Core/util/DisableStupidWarnings.h"
+
+#include <mkl_pardiso.h>
+
+#include <unsupported/Eigen/SparseExtra>
+
+namespace Eigen {
+
+/** \ingroup Support_modules
+  * \defgroup PARDISOSupport_Module PARDISOSupport module
+  *
+  * This module brings support for the Intel(R) MKL PARDISO direct sparse solvers
+  *
+  * \code
+  * #include <Eigen/PARDISOSupport>
+  * \endcode
+  */
+
+#include "src/PARDISOSupport/PARDISOSupport.h"
+
+} // namespace Eigen
+
+#include "src/Core/util/ReenableStupidWarnings.h"
+
+#endif // EIGEN_PARDISOSUPPORT_MODULE_H

Eigen/QR

@@ -28,6 +28,10 @@ namespace Eigen {
 #include "src/QR/HouseholderQR.h"
 #include "src/QR/FullPivHouseholderQR.h"
 #include "src/QR/ColPivHouseholderQR.h"
+#ifdef EIGEN_USE_LAPACKE
+#include "src/QR/HouseholderQR_MKL.h"
+#include "src/QR/ColPivHouseholderQR_MKL.h"
+#endif
 
 #ifdef EIGEN2_SUPPORT
 #include "src/Eigen2Support/QR.h"

Eigen/SVD

@@ -13,9 +13,9 @@ namespace Eigen {
   *
   *
   *
-  * This module provides SVD decomposition for (currently) real matrices.
+  * This module provides SVD decomposition for matrices (both real and complex).
   * This decomposition is accessible via the following MatrixBase method:
-  *  - MatrixBase::svd()
+  *  - MatrixBase::jacobiSvd()
   *
   * \code
   * #include <Eigen/SVD>
@@ -24,6 +24,9 @@ namespace Eigen {
 
 #include "src/misc/Solve.h"
 #include "src/SVD/JacobiSVD.h"
+#if defined(EIGEN_USE_LAPACKE) && !defined(EIGEN_USE_LAPACKE_STRICT)
+#include "src/SVD/JacobiSVD_MKL.h"
+#endif
 #include "src/SVD/UpperBidiagonalization.h"
 
 #ifdef EIGEN2_SUPPORT

Eigen/Sparse

@@ -1,69 +1,27 @@
 #ifndef EIGEN_SPARSE_MODULE_H
 #define EIGEN_SPARSE_MODULE_H
 
-#include "Core"
-
-#include "src/Core/util/DisableStupidWarnings.h"
-
-#include <vector>
-#include <map>
-#include <cstdlib>
-#include <cstring>
-#include <algorithm>
-
-#ifdef EIGEN2_SUPPORT
-#define EIGEN_YES_I_KNOW_SPARSE_MODULE_IS_NOT_STABLE_YET
-#endif
-
-#ifndef EIGEN_YES_I_KNOW_SPARSE_MODULE_IS_NOT_STABLE_YET
-#error The sparse module API is not stable yet. To use it anyway, please define the EIGEN_YES_I_KNOW_SPARSE_MODULE_IS_NOT_STABLE_YET preprocessor token.
-#endif
-
 namespace Eigen {
 
-/** \defgroup Sparse_Module Sparse module
+/** \defgroup Sparse_modules Sparse modules
   *
-  *
-  *
-  * See the \ref TutorialSparse "Sparse tutorial"
+  * Meta-module including all related modules:
+  * - SparseCore
+  * - OrderingMethods
+  * - SparseCholesky
+  * - IterativeLinearSolvers
   *
   * \code
   * #include <Eigen/Sparse>
   * \endcode
   */
 
-/** The type used to identify a general sparse storage. */
-struct Sparse {};
-
-#include "src/Sparse/SparseUtil.h"
-#include "src/Sparse/SparseMatrixBase.h"
-#include "src/Sparse/CompressedStorage.h"
-#include "src/Sparse/AmbiVector.h"
-#include "src/Sparse/SparseMatrix.h"
-#include "src/Sparse/DynamicSparseMatrix.h"
-#include "src/Sparse/MappedSparseMatrix.h"
-#include "src/Sparse/SparseVector.h"
-#include "src/Sparse/CoreIterators.h"
-#include "src/Sparse/SparseBlock.h"
-#include "src/Sparse/SparseTranspose.h"
-#include "src/Sparse/SparseCwiseUnaryOp.h"
-#include "src/Sparse/SparseCwiseBinaryOp.h"
-#include "src/Sparse/SparseDot.h"
-#include "src/Sparse/SparseAssign.h"
-#include "src/Sparse/SparseRedux.h"
-#include "src/Sparse/SparseFuzzy.h"
-#include "src/Sparse/SparseProduct.h"
-#include "src/Sparse/SparseSparseProduct.h"
-#include "src/Sparse/SparseDenseProduct.h"
-#include "src/Sparse/SparseDiagonalProduct.h"
-#include "src/Sparse/SparseTriangularView.h"
-#include "src/Sparse/SparseSelfAdjointView.h"
-#include "src/Sparse/TriangularSolver.h"
-#include "src/Sparse/SparseView.h"
-
 } // namespace Eigen
 
-#include "src/Core/util/ReenableStupidWarnings.h"
+#include "SparseCore"
+#include "OrderingMethods"
+#include "SparseCholesky"
+#include "IterativeLinearSolvers"
 
 #endif // EIGEN_SPARSE_MODULE_H
 

Eigen/SparseCholesky

@@ -0,0 +1,34 @@
+#ifndef EIGEN_SPARSECHOLESKY_MODULE_H
+#define EIGEN_SPARSECHOLESKY_MODULE_H
+
+#include "SparseCore"
+
+#include "src/Core/util/DisableStupidWarnings.h"
+
+namespace Eigen {
+
+/** \ingroup Sparse_modules
+  * \defgroup SparseCholesky_Module SparseCholesky module
+  *
+  * This module currently provides two variants of the direct sparse Cholesky decomposition for selfadjoint (hermitian) matrices.
+  * Those decompositions are accessible via the following classes:
+  *  - SimplicialLLt,
+  *  - SimplicialLDLt
+  *
+  * Such problems can also be solved using the ConjugateGradient solver from the IterativeLinearSolvers module.
+  *
+  * \code
+  * #include <Eigen/SparseCholesky>
+  * \endcode
+  */
+
+#include "src/misc/Solve.h"
+#include "src/misc/SparseSolve.h"
+
+#include "src/SparseCholesky/SimplicialCholesky.h"
+
+} // namespace Eigen
+
+#include "src/Core/util/ReenableStupidWarnings.h"
+
+#endif // EIGEN_SPARSECHOLESKY_MODULE_H

Eigen/SparseCore

@@ -0,0 +1,65 @@
+#ifndef EIGEN_SPARSECORE_MODULE_H
+#define EIGEN_SPARSECORE_MODULE_H
+
+#include "Core"
+
+#include "src/Core/util/DisableStupidWarnings.h"
+
+#include <vector>
+#include <map>
+#include <cstdlib>
+#include <cstring>
+#include <algorithm>
+
+namespace Eigen {
+
+/** \ingroup Sparse_modules
+  * \defgroup SparseCore_Module SparseCore module
+  *
+  * This module provides a sparse matrix representation, and basic associatd matrix manipulations
+  * and operations.
+  *
+  * See the \ref TutorialSparse "Sparse tutorial"
+  *
+  * \code
+  * #include <Eigen/SparseCore>
+  * \endcode
+  *
+  * This module depends on: Core.
+  */
+
+/** The type used to identify a general sparse storage. */
+struct Sparse {};
+
+#include "src/SparseCore/SparseUtil.h"
+#include "src/SparseCore/SparseMatrixBase.h"
+#include "src/SparseCore/CompressedStorage.h"
+#include "src/SparseCore/AmbiVector.h"
+#include "src/SparseCore/SparseMatrix.h"
+#include "src/SparseCore/MappedSparseMatrix.h"
+#include "src/SparseCore/SparseVector.h"
+#include "src/SparseCore/CoreIterators.h"
+#include "src/SparseCore/SparseBlock.h"
+#include "src/SparseCore/SparseTranspose.h"
+#include "src/SparseCore/SparseCwiseUnaryOp.h"
+#include "src/SparseCore/SparseCwiseBinaryOp.h"
+#include "src/SparseCore/SparseDot.h"
+#include "src/SparseCore/SparseAssign.h"
+#include "src/SparseCore/SparseRedux.h"
+#include "src/SparseCore/SparseFuzzy.h"
+#include "src/SparseCore/ConservativeSparseSparseProduct.h"
+#include "src/SparseCore/SparseSparseProductWithPruning.h"
+#include "src/SparseCore/SparseProduct.h"
+#include "src/SparseCore/SparseDenseProduct.h"
+#include "src/SparseCore/SparseDiagonalProduct.h"
+#include "src/SparseCore/SparseTriangularView.h"
+#include "src/SparseCore/SparseSelfAdjointView.h"
+#include "src/SparseCore/TriangularSolver.h"
+#include "src/SparseCore/SparseView.h"
+
+} // namespace Eigen
+
+#include "src/Core/util/ReenableStupidWarnings.h"
+
+#endif // EIGEN_SPARSECORE_MODULE_H
+

Eigen/SuperLUSupport

@@ -0,0 +1,53 @@
+#ifndef EIGEN_SUPERLUSUPPORT_MODULE_H
+#define EIGEN_SUPERLUSUPPORT_MODULE_H
+
+#include "SparseCore"
+
+#include "src/Core/util/DisableStupidWarnings.h"
+
+#ifdef EMPTY
+#define EIGEN_EMPTY_WAS_ALREADY_DEFINED
+#endif
+
+typedef int int_t;
+#include <slu_Cnames.h>
+#include <supermatrix.h>
+#include <slu_util.h>
+
+// slu_util.h defines a preprocessor token named EMPTY which is really polluting,
+// so we remove it in favor of a SUPERLU_EMPTY token.
+// If EMPTY was already, defined then we don't undef it.
+
+#if defined(EIGEN_EMPTY_WAS_ALREADY_DEFINED)
+# undef EIGEN_EMPTY_WAS_ALREADY_DEFINED
+#elif defined(EMPTY)
+# undef EMPTY
+#endif
+
+#define SUPERLU_EMPTY (-1)
+
+namespace Eigen { struct SluMatrix; }
+
+namespace Eigen {
+
+/** \ingroup Support_modules
+  * \defgroup SuperLUSupport_Module SuperLUSupport module
+  *
+  * \warning When including this module, you have to use SUPERLU_EMPTY instead of EMPTY which is no longer defined because it is too polluting.
+  *
+  * \code
+  * #include <Eigen/SuperLUSupport>
+  * \endcode
+  */
+
+#include "src/misc/Solve.h"
+#include "src/misc/SparseSolve.h"
+
+#include "src/SuperLUSupport/SuperLUSupport.h"
+
+
+} // namespace Eigen
+
+#include "src/Core/util/ReenableStupidWarnings.h"
+
+#endif // EIGEN_SUPERLUSUPPORT_MODULE_H

Eigen/UmfPackSupport

@@ -0,0 +1,34 @@
+#ifndef EIGEN_UMFPACKSUPPORT_MODULE_H
+#define EIGEN_UMFPACKSUPPORT_MODULE_H
+
+#include "SparseCore"
+
+#include "src/Core/util/DisableStupidWarnings.h"
+
+extern "C" {
+#include <umfpack.h>
+}
+
+namespace Eigen {
+
+/** \ingroup Support_modules
+  * \defgroup UmfPackSupport_Module UmfPackSupport module
+  *
+  *
+  *
+  *
+  * \code
+  * #include <Eigen/UmfPackSupport>
+  * \endcode
+  */
+
+#include "src/misc/Solve.h"
+#include "src/misc/SparseSolve.h"
+
+#include "src/UmfPackSupport/UmfPackSupport.h"
+
+} // namespace Eigen
+
+#include "src/Core/util/ReenableStupidWarnings.h"
+
+#endif // EIGEN_UMFPACKSUPPORT_MODULE_H

Eigen/src/Cholesky/LDLT.h

@@ -1,9 +1,10 @@
 // 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) 2008-2011 Gael Guennebaud <gael.guennebaud@inria.fr>
 // Copyright (C) 2009 Keir Mierle <mierle@gmail.com>
 // Copyright (C) 2009 Benoit Jacob <jacob.benoit.1@gmail.com>
+// Copyright (C) 2011 Timothy E. Holy <tim.holy@gmail.com >
 //
 // Eigen is free software; you can redistribute it and/or
 // modify it under the terms of the GNU Lesser General Public
@@ -31,13 +32,15 @@ namespace internal {
 template<typename MatrixType, int UpLo> struct LDLT_Traits;
 }
 
-/** \ingroup cholesky_Module
+/** \ingroup Cholesky_Module
   *
   * \class LDLT
   *
   * \brief Robust Cholesky decomposition of a matrix with pivoting
   *
   * \param MatrixType the type of the matrix of which to compute the LDL^T Cholesky decomposition
+  * \param UpLo the triangular part that will be used for the decompositon: Lower (default) or Upper.
+  *             The other triangular part won't be read.
   *
   * Perform a robust Cholesky decomposition of a positive semidefinite or negative semidefinite
   * matrix \f$ A \f$ such that \f$ A =  P^TLDL^*P \f$, where P is a permutation matrix, L
@@ -48,14 +51,10 @@ template<typename MatrixType, int UpLo> struct LDLT_Traits;
   * on D also stabilizes the computation.
   *
   * Remember that Cholesky decompositions are not rank-revealing. Also, do not use a Cholesky
-	* decomposition to determine whether a system of equations has a solution.
+  * decomposition to determine whether a system of equations has a solution.
   *
   * \sa MatrixBase::ldlt(), class LLT
   */
- /* THIS PART OF THE DOX IS CURRENTLY DISABLED BECAUSE INACCURATE BECAUSE OF BUG IN THE DECOMPOSITION CODE
-  * Note that during the decomposition, only the upper triangular part of A is considered. Therefore,
-  * the strict lower part does not have to store correct values.
-  */
 template<typename _MatrixType, int _UpLo> class LDLT
 {
   public:
@@ -98,6 +97,11 @@ template<typename _MatrixType, int _UpLo> class LDLT
         m_isInitialized(false)
     {}
 
+    /** \brief Constructor with decomposition
+      *
+      * This calculates the decomposition for the input \a matrix.
+      * \sa LDLT(Index size)
+      */
     LDLT(const MatrixType& matrix)
       : m_matrix(matrix.rows(), matrix.cols()),
         m_transpositions(matrix.rows()),
@@ -107,6 +111,14 @@ template<typename _MatrixType, int _UpLo> class LDLT
       compute(matrix);
     }
 
+    /** Clear any existing decomposition
+     * \sa rankUpdate(w,sigma)
+     */
+    void setZero()
+    {
+      m_isInitialized = false;
+    }
+
     /** \returns a view of the upper triangular matrix U */
     inline typename Traits::MatrixU matrixU() const
     {
@@ -130,14 +142,14 @@ template<typename _MatrixType, int _UpLo> class LDLT
     }
 
     /** \returns the coefficients of the diagonal matrix D */
-    inline Diagonal<const MatrixType> vectorD(void) const
+    inline Diagonal<const MatrixType> vectorD() const
     {
       eigen_assert(m_isInitialized && "LDLT is not initialized.");
       return m_matrix.diagonal();
     }
 
     /** \returns true if the matrix is positive (semidefinite) */
-    inline bool isPositive(void) const
+    inline bool isPositive() const
     {
       eigen_assert(m_isInitialized && "LDLT is not initialized.");
       return m_sign == 1;
@@ -196,6 +208,9 @@ template<typename _MatrixType, int _UpLo> class LDLT
 
     LDLT& compute(const MatrixType& matrix);
 
+    template <typename Derived>
+    LDLT& rankUpdate(const MatrixBase<Derived>& w,RealScalar alpha=1);
+
     /** \returns the internal LDLT decomposition matrix
       *
       * TODO: document the storage layout
@@ -211,6 +226,17 @@ template<typename _MatrixType, int _UpLo> class LDLT
     inline Index rows() const { return m_matrix.rows(); }
     inline Index cols() const { return m_matrix.cols(); }
 
+    /** \brief Reports whether previous computation was successful.
+      *
+      * \returns \c Success if computation was succesful,
+      *          \c NumericalIssue if the matrix.appears to be negative.
+      */
+    ComputationInfo info() const
+    {
+      eigen_assert(m_isInitialized && "LDLT is not initialized.");
+      return Success;
+    }
+
   protected:
 
     /** \internal
@@ -249,7 +275,7 @@ template<> struct ldlt_inplace<Lower>
       return true;
     }
 
-    RealScalar cutoff = 0, biggest_in_corner;
+    RealScalar cutoff(0), biggest_in_corner;
 
     for (Index k = 0; k < size; ++k)
     {
@@ -317,6 +343,61 @@ template<> struct ldlt_inplace<Lower>
 
     return true;
   }
+
+  // Reference for the algorithm: Davis and Hager, "Multiple Rank
+  // Modifications of a Sparse Cholesky Factorization" (Algorithm 1)
+  // Trivial rearrangements of their computations (Timothy E. Holy)
+  // allow their algorithm to work for rank-1 updates even if the
+  // original matrix is not of full rank.
+  // Here only rank-1 updates are implemented, to reduce the
+  // requirement for intermediate storage and improve accuracy
+  template<typename MatrixType, typename WDerived>
+  static bool updateInPlace(MatrixType& mat, MatrixBase<WDerived>& w, typename MatrixType::RealScalar sigma=1)
+  {
+    using internal::isfinite;
+    typedef typename MatrixType::Scalar Scalar;
+    typedef typename MatrixType::RealScalar RealScalar;
+    typedef typename MatrixType::Index Index;
+
+    const Index size = mat.rows();
+    eigen_assert(mat.cols() == size && w.size()==size);
+
+    RealScalar alpha = 1;
+
+    // Apply the update
+    for (Index j = 0; j < size; j++)
+    {
+      // Check for termination due to an original decomposition of low-rank
+      if (!isfinite(alpha))
+        break;
+
+      // Update the diagonal terms
+      RealScalar dj = real(mat.coeff(j,j));
+      Scalar wj = w.coeff(j);
+      RealScalar swj2 = sigma*abs2(wj);
+      RealScalar gamma = dj*alpha + swj2;
+
+      mat.coeffRef(j,j) += swj2/alpha;
+      alpha += swj2/dj;
+
+
+      // Update the terms of L
+      Index rs = size-j-1;
+      w.tail(rs) -= wj * mat.col(j).tail(rs);
+      if(gamma != 0)
+        mat.col(j).tail(rs) += (sigma*conj(wj)/gamma)*w.tail(rs);
+    }
+    return true;
+  }
+
+  template<typename MatrixType, typename TranspositionType, typename Workspace, typename WType>
+  static bool update(MatrixType& mat, const TranspositionType& transpositions, Workspace& tmp, const WType& w, typename MatrixType::RealScalar sigma=1)
+  {
+    // Apply the permutation to the input w
+    tmp = transpositions * w;
+
+    return ldlt_inplace<Lower>::updateInPlace(mat,tmp,sigma);
+  }
 };
 
 template<> struct ldlt_inplace<Upper>
@@ -327,22 +408,29 @@ template<> struct ldlt_inplace<Upper>
     Transpose<MatrixType> matt(mat);
     return ldlt_inplace<Lower>::unblocked(matt, transpositions, temp, sign);
   }
+
+  template<typename MatrixType, typename TranspositionType, typename Workspace, typename WType>
+  static EIGEN_STRONG_INLINE bool update(MatrixType& mat, TranspositionType& transpositions, Workspace& tmp, WType& w, typename MatrixType::RealScalar sigma=1)
+  {
+    Transpose<MatrixType> matt(mat);
+    return ldlt_inplace<Lower>::update(matt, transpositions, tmp, w.conjugate(), sigma);
+  }
 };
 
 template<typename MatrixType> struct LDLT_Traits<MatrixType,Lower>
 {
-  typedef TriangularView<MatrixType, UnitLower> MatrixL;
-  typedef TriangularView<typename MatrixType::AdjointReturnType, UnitUpper> MatrixU;
-  inline static MatrixL getL(const MatrixType& m) { return m; }
-  inline static MatrixU getU(const MatrixType& m) { return m.adjoint(); }
+  typedef TriangularView<const MatrixType, UnitLower> MatrixL;
+  typedef TriangularView<const typename MatrixType::AdjointReturnType, UnitUpper> MatrixU;
+  static inline MatrixL getL(const MatrixType& m) { return m; }
+  static inline MatrixU getU(const MatrixType& m) { return m.adjoint(); }
 };
 
 template<typename MatrixType> struct LDLT_Traits<MatrixType,Upper>
 {
-  typedef TriangularView<typename MatrixType::AdjointReturnType, UnitLower> MatrixL;
-  typedef TriangularView<MatrixType, UnitUpper> MatrixU;
-  inline static MatrixL getL(const MatrixType& m) { return m.adjoint(); }
-  inline static MatrixU getU(const MatrixType& m) { return m; }
+  typedef TriangularView<const typename MatrixType::AdjointReturnType, UnitLower> MatrixL;
+  typedef TriangularView<const MatrixType, UnitUpper> MatrixU;
+  static inline MatrixL getL(const MatrixType& m) { return m.adjoint(); }
+  static inline MatrixU getU(const MatrixType& m) { return m; }
 };
 
 } // end namespace internal
@@ -367,6 +455,37 @@ LDLT<MatrixType,_UpLo>& LDLT<MatrixType,_UpLo>::compute(const MatrixType& a)
   return *this;
 }
 
+/** Update the LDLT decomposition:  given A = L D L^T, efficiently compute the decomposition of A + sigma w w^T.
+ * \param w a vector to be incorporated into the decomposition.
+ * \param sigma a scalar, +1 for updates and -1 for "downdates," which correspond to removing previously-added column vectors. Optional; default value is +1.
+ * \sa setZero()
+  */
+template<typename MatrixType, int _UpLo>
+template<typename Derived>
+LDLT<MatrixType,_UpLo>& LDLT<MatrixType,_UpLo>::rankUpdate(const MatrixBase<Derived>& w,typename NumTraits<typename MatrixType::Scalar>::Real sigma)
+{
+  const Index size = w.rows();
+  if (m_isInitialized)
+  {
+    eigen_assert(m_matrix.rows()==size);
+  }
+  else
+  {    
+    m_matrix.resize(size,size);
+    m_matrix.setZero();
+    m_transpositions.resize(size);
+    for (Index i = 0; i < size; i++)
+      m_transpositions.coeffRef(i) = i;
+    m_temporary.resize(size);
+    m_sign = sigma;
+    m_isInitialized = true;
+  }
+
+  internal::ldlt_inplace<UpLo>::update(m_matrix, m_transpositions, m_temporary, w, sigma);
+
+  return *this;
+}
+
 namespace internal {
 template<typename _MatrixType, int _UpLo, typename Rhs>
 struct solve_retval<LDLT<_MatrixType,_UpLo>, Rhs>

Eigen/src/Cholesky/LLT.h

@@ -29,13 +29,15 @@ namespace internal{
 template<typename MatrixType, int UpLo> struct LLT_Traits;
 }
 
-/** \ingroup cholesky_Module
+/** \ingroup Cholesky_Module
   *
   * \class LLT
   *
   * \brief Standard Cholesky decomposition (LL^T) of a matrix and associated features
   *
   * \param MatrixType the type of the matrix of which we are computing the LL^T Cholesky decomposition
+  * \param UpLo the triangular part that will be used for the decompositon: Lower (default) or Upper.
+  *             The other triangular part won't be read.
   *
   * This class performs a LL^T Cholesky decomposition of a symmetric, positive definite
   * matrix A such that A = LL^* = U^*U, where L is lower triangular.
@@ -49,6 +51,9 @@ template<typename MatrixType, int UpLo> struct LLT_Traits;
   * use LDLT instead for the semidefinite case. Also, do not use a Cholesky decomposition to determine whether a system of equations
   * has a solution.
   *
+  * Example: \include LLT_example.cpp
+  * Output: \verbinclude LLT_example.out
+  *    
   * \sa MatrixBase::llt(), class LDLT
   */
  /* HEY THIS DOX IS DISABLED BECAUSE THERE's A BUG EITHER HERE OR IN LDLT ABOUT THAT (OR BOTH)
@@ -178,6 +183,9 @@ template<typename _MatrixType, int _UpLo> class LLT
     inline Index rows() const { return m_matrix.rows(); }
     inline Index cols() const { return m_matrix.cols(); }
 
+    template<typename VectorType>
+    LLT& rankUpdate(const VectorType& vec, const RealScalar& sigma = 1);
+
   protected:
     /** \internal
       * Used to compute and store L
@@ -190,16 +198,15 @@ template<typename _MatrixType, int _UpLo> class LLT
 
 namespace internal {
 
-template<int UpLo> struct llt_inplace;
+template<typename Scalar, int UpLo> struct llt_inplace;
 
-template<> struct llt_inplace<Lower>
+template<typename Scalar> struct llt_inplace<Scalar, Lower>
 {
+  typedef typename NumTraits<Scalar>::Real RealScalar;
   template<typename MatrixType>
   static typename MatrixType::Index unblocked(MatrixType& mat)
   {
     typedef typename MatrixType::Index Index;
-    typedef typename MatrixType::Scalar Scalar;
-    typedef typename MatrixType::RealScalar RealScalar;
     
     eigen_assert(mat.rows()==mat.cols());
     const Index size = mat.rows();
@@ -254,55 +261,133 @@ template<> struct llt_inplace<Lower>
     }
     return -1;
   }
+
+  template<typename MatrixType, typename VectorType>
+  static typename MatrixType::Index rankUpdate(MatrixType& mat, const VectorType& vec, const RealScalar& sigma)
+  {
+    typedef typename MatrixType::Index Index;
+    typedef typename MatrixType::ColXpr ColXpr;
+    typedef typename internal::remove_all<ColXpr>::type ColXprCleaned;
+    typedef typename ColXprCleaned::SegmentReturnType ColXprSegment;
+    typedef Matrix<Scalar,Dynamic,1> TempVectorType;
+    typedef typename TempVectorType::SegmentReturnType TempVecSegment;
+
+    int n = mat.cols();
+    eigen_assert(mat.rows()==n && vec.size()==n);
+
+    TempVectorType temp;
+
+    if(sigma>0)
+    {
+      // This version is based on Givens rotations.
+      // It is faster than the other one below, but only works for updates,
+      // i.e., for sigma > 0
+      temp = sqrt(sigma) * vec;
+
+      for(int i=0; i<n; ++i)
+      {
+        JacobiRotation<Scalar> g;
+        g.makeGivens(mat(i,i), -temp(i), &mat(i,i));
+
+        int rs = n-i-1;
+        if(rs>0)
+        {
+          ColXprSegment x(mat.col(i).tail(rs));
+          TempVecSegment y(temp.tail(rs));
+          apply_rotation_in_the_plane(x, y, g);
+        }
+      }
+    }
+    else
+    {
+      temp = vec;
+      RealScalar beta = 1;
+      for(int j=0; j<n; ++j)
+      {
+        RealScalar Ljj = real(mat.coeff(j,j));
+        RealScalar dj = abs2(Ljj);
+        Scalar wj = temp.coeff(j);
+        RealScalar swj2 = sigma*abs2(wj);
+        RealScalar gamma = dj*beta + swj2;
+
+        RealScalar x = dj + swj2/beta;
+        if (x<=RealScalar(0))
+          return j;
+        RealScalar nLjj = sqrt(x);
+        mat.coeffRef(j,j) = nLjj;
+        beta += swj2/dj;
+
+        // Update the terms of L
+        Index rs = n-j-1;
+        if(rs)
+        {
+          temp.tail(rs) -= (wj/Ljj) * mat.col(j).tail(rs);
+          if(gamma != 0)
+            mat.col(j).tail(rs) = (nLjj/Ljj) * mat.col(j).tail(rs) + (nLjj * sigma*conj(wj)/gamma)*temp.tail(rs);
+        }
+      }
+    }
+    return -1;
+  }
 };
 
-template<> struct llt_inplace<Upper>
+template<typename Scalar> struct llt_inplace<Scalar, Upper>
 {
+  typedef typename NumTraits<Scalar>::Real RealScalar;
+
   template<typename MatrixType>
   static EIGEN_STRONG_INLINE typename MatrixType::Index unblocked(MatrixType& mat)
   {
     Transpose<MatrixType> matt(mat);
-    return llt_inplace<Lower>::unblocked(matt);
+    return llt_inplace<Scalar, Lower>::unblocked(matt);
   }
   template<typename MatrixType>
   static EIGEN_STRONG_INLINE typename MatrixType::Index blocked(MatrixType& mat)
   {
     Transpose<MatrixType> matt(mat);
-    return llt_inplace<Lower>::blocked(matt);
+    return llt_inplace<Scalar, Lower>::blocked(matt);
+  }
+  template<typename MatrixType, typename VectorType>
+  static typename MatrixType::Index rankUpdate(MatrixType& mat, const VectorType& vec, const RealScalar& sigma)
+  {
+    Transpose<MatrixType> matt(mat);
+    return llt_inplace<Scalar, Lower>::rankUpdate(matt, vec.conjugate(), sigma);
   }
 };
 
 template<typename MatrixType> struct LLT_Traits<MatrixType,Lower>
 {
-  typedef TriangularView<MatrixType, Lower> MatrixL;
-  typedef TriangularView<typename MatrixType::AdjointReturnType, Upper> MatrixU;
-  inline static MatrixL getL(const MatrixType& m) { return m; }
-  inline static MatrixU getU(const MatrixType& m) { return m.adjoint(); }
+  typedef TriangularView<const MatrixType, Lower> MatrixL;
+  typedef TriangularView<const typename MatrixType::AdjointReturnType, Upper> MatrixU;
+  static inline MatrixL getL(const MatrixType& m) { return m; }
+  static inline MatrixU getU(const MatrixType& m) { return m.adjoint(); }
   static bool inplace_decomposition(MatrixType& m)
-  { return llt_inplace<Lower>::blocked(m)==-1; }
+  { return llt_inplace<typename MatrixType::Scalar, Lower>::blocked(m)==-1; }
 };
 
 template<typename MatrixType> struct LLT_Traits<MatrixType,Upper>
 {
-  typedef TriangularView<typename MatrixType::AdjointReturnType, Lower> MatrixL;
-  typedef TriangularView<MatrixType, Upper> MatrixU;
-  inline static MatrixL getL(const MatrixType& m) { return m.adjoint(); }
-  inline static MatrixU getU(const MatrixType& m) { return m; }
+  typedef TriangularView<const typename MatrixType::AdjointReturnType, Lower> MatrixL;
+  typedef TriangularView<const MatrixType, Upper> MatrixU;
+  static inline MatrixL getL(const MatrixType& m) { return m.adjoint(); }
+  static inline MatrixU getU(const MatrixType& m) { return m; }
   static bool inplace_decomposition(MatrixType& m)
-  { return llt_inplace<Upper>::blocked(m)==-1; }
+  { return llt_inplace<typename MatrixType::Scalar, Upper>::blocked(m)==-1; }
 };
 
 } // end namespace internal
 
 /** Computes / recomputes the Cholesky decomposition A = LL^* = U^*U of \a matrix
-  *
   *
   * \returns a reference to *this
+  *
+  * Example: \include TutorialLinAlgComputeTwice.cpp
+  * Output: \verbinclude TutorialLinAlgComputeTwice.out
   */
 template<typename MatrixType, int _UpLo>
 LLT<MatrixType,_UpLo>& LLT<MatrixType,_UpLo>::compute(const MatrixType& a)
 {
-  assert(a.rows()==a.cols());
+  eigen_assert(a.rows()==a.cols());
   const Index size = a.rows();
   m_matrix.resize(size, size);
   m_matrix = a;
@@ -314,6 +399,26 @@ LLT<MatrixType,_UpLo>& LLT<MatrixType,_UpLo>::compute(const MatrixType& a)
   return *this;
 }
 
+/** Performs a rank one update (or dowdate) of the current decomposition.
+  * If A = LL^* before the rank one update,
+  * then after it we have LL^* = A + sigma * v v^* where \a v must be a vector
+  * of same dimension.
+  */
+template<typename MatrixType, int _UpLo>
+template<typename VectorType>
+LLT<MatrixType,_UpLo>& LLT<MatrixType,_UpLo>::rankUpdate(const VectorType& v, const RealScalar& sigma)
+{
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(VectorType);
+  eigen_assert(v.size()==m_matrix.cols());
+  eigen_assert(m_isInitialized);
+  if(internal::llt_inplace<typename MatrixType::Scalar, UpLo>::rankUpdate(m_matrix,v,sigma)>=0)
+    m_info = NumericalIssue;
+  else
+    m_info = Success;
+
+  return *this;
+}
+    
 namespace internal {
 template<typename _MatrixType, int UpLo, typename Rhs>
 struct solve_retval<LLT<_MatrixType, UpLo>, Rhs>
@@ -384,3 +489,4 @@ SelfAdjointView<MatrixType, UpLo>::llt() const
 }
 
 #endif // EIGEN_LLT_H
+

Eigen/src/Cholesky/LLT_MKL.h

@@ -0,0 +1,123 @@
+/*
+ Copyright (c) 2011, Intel Corporation. All rights reserved.
+
+ Redistribution and use in source and binary forms, with or without modification,
+ are permitted provided that the following conditions are met:
+
+ * Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+ * Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+ * Neither the name of Intel Corporation nor the names of its contributors may
+   be used to endorse or promote products derived from this software without
+   specific prior written permission.
+
+ THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+ WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+ DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+ (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+ LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+ ********************************************************************************
+ *   Content : Eigen bindings to Intel(R) MKL
+ *     LLt decomposition based on LAPACKE_?potrf function.
+ ********************************************************************************
+*/
+
+#ifndef EIGEN_LLT_MKL_H
+#define EIGEN_LLT_MKL_H
+
+#include "Eigen/src/Core/util/MKL_support.h"
+#include <iostream>
+
+namespace internal {
+
+template<typename Scalar> struct mkl_llt;
+
+#define EIGEN_MKL_LLT(EIGTYPE, MKLTYPE, MKLPREFIX) \
+template<> struct mkl_llt<EIGTYPE> \
+{ \
+  template<typename MatrixType> \
+  static inline typename MatrixType::Index potrf(MatrixType& m, char uplo) \
+  { \
+    lapack_int matrix_order; \
+    lapack_int size, lda, info, StorageOrder; \
+    EIGTYPE* a; \
+    eigen_assert(m.rows()==m.cols()); \
+/* Set up parameters for ?potrf */ \
+    size = m.rows(); \
+    StorageOrder = MatrixType::Flags&RowMajorBit?RowMajor:ColMajor; \
+    matrix_order = StorageOrder==RowMajor ? LAPACK_ROW_MAJOR : LAPACK_COL_MAJOR; \
+    a = &(m.coeffRef(0,0)); \
+    lda = m.outerStride(); \
+\
+    info = LAPACKE_##MKLPREFIX##potrf( matrix_order, uplo, size, (MKLTYPE*)a, lda ); \
+    info = (info==0) ? Success : NumericalIssue; \
+    return info; \
+  } \
+}; \
+template<> struct llt_inplace<EIGTYPE, Lower> \
+{ \
+  template<typename MatrixType> \
+  static typename MatrixType::Index blocked(MatrixType& m) \
+  { \
+    return mkl_llt<EIGTYPE>::potrf(m, 'L'); \
+  } \
+  template<typename MatrixType, typename VectorType> \
+  static void rankUpdate(MatrixType& mat, const VectorType& vec) \
+  { \
+    typedef typename MatrixType::ColXpr ColXpr; \
+    typedef typename internal::remove_all<ColXpr>::type ColXprCleaned; \
+    typedef typename ColXprCleaned::SegmentReturnType ColXprSegment; \
+    typedef typename MatrixType::Scalar Scalar; \
+    typedef Matrix<Scalar,Dynamic,1> TempVectorType; \
+    typedef typename TempVectorType::SegmentReturnType TempVecSegment; \
+\
+    int n = mat.cols(); \
+    eigen_assert(mat.rows()==n && vec.size()==n); \
+    TempVectorType temp(vec); \
+\
+    for(int i=0; i<n; ++i) \
+    { \
+      JacobiRotation<Scalar> g; \
+      g.makeGivens(mat(i,i), -temp(i), &mat(i,i)); \
+\
+      int rs = n-i-1; \
+      if(rs>0) \
+      { \
+        ColXprSegment x(mat.col(i).tail(rs)); \
+        TempVecSegment y(temp.tail(rs)); \
+        apply_rotation_in_the_plane(x, y, g); \
+      } \
+    } \
+  } \
+}; \
+template<> struct llt_inplace<EIGTYPE, Upper> \
+{ \
+  template<typename MatrixType> \
+  static typename MatrixType::Index blocked(MatrixType& m) \
+  { \
+    return mkl_llt<EIGTYPE>::potrf(m, 'U'); \
+  } \
+  template<typename MatrixType, typename VectorType> \
+  static void rankUpdate(MatrixType& mat, const VectorType& vec) \
+  { \
+    Transpose<MatrixType> matt(mat); \
+    return llt_inplace<EIGTYPE, Lower>::rankUpdate(matt, vec.conjugate()); \
+  } \
+};
+
+EIGEN_MKL_LLT(double, double, d)
+EIGEN_MKL_LLT(float, float, s)
+EIGEN_MKL_LLT(dcomplex, MKL_Complex16, z)
+EIGEN_MKL_LLT(scomplex, MKL_Complex8, c)
+
+}
+
+#endif // EIGEN_LLT_MKL_H

Eigen/src/CholmodSupport/CMakeLists.txt

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

@@ -69,11 +69,20 @@ cholmod_sparse viewAsCholmod(SparseMatrix<_Scalar,_Options,_Index>& mat)
   res.nzmax   = mat.nonZeros();
   res.nrow    = mat.rows();;
   res.ncol    = mat.cols();
-  res.p       = mat._outerIndexPtr();
-  res.i       = mat._innerIndexPtr();
-  res.x       = mat._valuePtr();
+  res.p       = mat.outerIndexPtr();
+  res.i       = mat.innerIndexPtr();
+  res.x       = mat.valuePtr();
   res.sorted  = 1;
-  res.packed  = 1;
+  if(mat.isCompressed())
+  {
+    res.packed  = 1;
+  }
+  else
+  {
+    res.packed  = 0;
+    res.nz = mat.innerNonZeroPtr();
+  }
+
   res.dtype   = 0;
   res.stype   = -1;
   
@@ -149,7 +158,9 @@ enum CholmodMode {
   CholmodAuto, CholmodSimplicialLLt, CholmodSupernodalLLt, CholmodLDLt
 };
 
-/** \brief A Cholesky factorization and solver based on Cholmod
+/** \ingroup CholmodSupport_Module
+  * \class CholmodDecomposition
+  * \brief A Cholesky factorization and solver based on Cholmod
   *
   * This class allows to solve for A.X = B sparse linear problems via a LL^T or LDL^T Cholesky factorization
   * using the Cholmod library. The sparse matrix A must be selfajoint and positive definite. The vectors or matrices
@@ -159,6 +170,9 @@ enum CholmodMode {
   * \tparam _UpLo the triangular part that will be used for the computations. It can be Lower
   *               or Upper. Default is Lower.
   *
+  * This class supports all kind of SparseMatrix<>: row or column major; upper, lower, or both; isCompressed() or unisCompressed().
+  *
+  * \sa \ref TutorialSparseDirectSolvers
   */
 template<typename _MatrixType, int _UpLo = Lower>
 class CholmodDecomposition

Eigen/src/Core/ArrayBase.h

@@ -159,7 +159,7 @@ template<typename Derived> class ArrayBase
     /** \returns an \link MatrixBase Matrix \endlink expression of this array
       * \sa MatrixBase::array() */
     MatrixWrapper<Derived> matrix() { return derived(); }
-    const MatrixWrapper<Derived> matrix() const { return derived(); }
+    const MatrixWrapper<const Derived> matrix() const { return derived(); }
 
 //     template<typename Dest>
 //     inline void evalTo(Dest& dst) const { dst = matrix(); }
@@ -174,10 +174,10 @@ template<typename Derived> class ArrayBase
   protected:
     // mixing arrays and matrices is not legal
     template<typename OtherDerived> Derived& operator+=(const MatrixBase<OtherDerived>& )
-    {EIGEN_STATIC_ASSERT(sizeof(typename OtherDerived::Scalar)==-1,YOU_CANNOT_MIX_ARRAYS_AND_MATRICES);}
+    {EIGEN_STATIC_ASSERT(std::ptrdiff_t(sizeof(typename OtherDerived::Scalar))==-1,YOU_CANNOT_MIX_ARRAYS_AND_MATRICES); return *this;}
     // mixing arrays and matrices is not legal
     template<typename OtherDerived> Derived& operator-=(const MatrixBase<OtherDerived>& )
-    {EIGEN_STATIC_ASSERT(sizeof(typename OtherDerived::Scalar)==-1,YOU_CANNOT_MIX_ARRAYS_AND_MATRICES);}
+    {EIGEN_STATIC_ASSERT(std::ptrdiff_t(sizeof(typename OtherDerived::Scalar))==-1,YOU_CANNOT_MIX_ARRAYS_AND_MATRICES); return *this;}
 };
 
 /** replaces \c *this by \c *this - \a other.

Eigen/src/Core/ArrayWrapper.h

@@ -61,7 +61,7 @@ class ArrayWrapper : public ArrayBase<ArrayWrapper<ExpressionType> >
 
     typedef typename internal::nested<ExpressionType>::type NestedExpressionType;
 
-    inline ArrayWrapper(const ExpressionType& matrix) : m_expression(matrix) {}
+    inline ArrayWrapper(ExpressionType& matrix) : m_expression(matrix) {}
 
     inline Index rows() const { return m_expression.rows(); }
     inline Index cols() const { return m_expression.cols(); }
@@ -71,7 +71,7 @@ class ArrayWrapper : public ArrayBase<ArrayWrapper<ExpressionType> >
     inline ScalarWithConstIfNotLvalue* data() { return m_expression.data(); }
     inline const Scalar* data() const { return m_expression.data(); }
 
-    inline const CoeffReturnType coeff(Index row, Index col) const
+    inline CoeffReturnType coeff(Index row, Index col) const
     {
       return m_expression.coeff(row, col);
     }
@@ -86,7 +86,7 @@ class ArrayWrapper : public ArrayBase<ArrayWrapper<ExpressionType> >
       return m_expression.const_cast_derived().coeffRef(row, col);
     }
 
-    inline const CoeffReturnType coeff(Index index) const
+    inline CoeffReturnType coeff(Index index) const
     {
       return m_expression.coeff(index);
     }
@@ -128,8 +128,14 @@ class ArrayWrapper : public ArrayBase<ArrayWrapper<ExpressionType> >
     template<typename Dest>
     inline void evalTo(Dest& dst) const { dst = m_expression; }
 
+    const typename internal::remove_all<NestedExpressionType>::type& 
+    nestedExpression() const 
+    {
+      return m_expression;
+    }
+
   protected:
-    const NestedExpressionType m_expression;
+    NestedExpressionType m_expression;
 };
 
 /** \class MatrixWrapper
@@ -168,7 +174,7 @@ class MatrixWrapper : public MatrixBase<MatrixWrapper<ExpressionType> >
 
     typedef typename internal::nested<ExpressionType>::type NestedExpressionType;
 
-    inline MatrixWrapper(const ExpressionType& matrix) : m_expression(matrix) {}
+    inline MatrixWrapper(ExpressionType& matrix) : m_expression(matrix) {}
 
     inline Index rows() const { return m_expression.rows(); }
     inline Index cols() const { return m_expression.cols(); }
@@ -178,7 +184,7 @@ class MatrixWrapper : public MatrixBase<MatrixWrapper<ExpressionType> >
     inline ScalarWithConstIfNotLvalue* data() { return m_expression.data(); }
     inline const Scalar* data() const { return m_expression.data(); }
 
-    inline const CoeffReturnType coeff(Index row, Index col) const
+    inline CoeffReturnType coeff(Index row, Index col) const
     {
       return m_expression.coeff(row, col);
     }
@@ -193,7 +199,7 @@ class MatrixWrapper : public MatrixBase<MatrixWrapper<ExpressionType> >
       return m_expression.derived().coeffRef(row, col);
     }
 
-    inline const CoeffReturnType coeff(Index index) const
+    inline CoeffReturnType coeff(Index index) const
     {
       return m_expression.coeff(index);
     }
@@ -232,8 +238,14 @@ class MatrixWrapper : public MatrixBase<MatrixWrapper<ExpressionType> >
       m_expression.const_cast_derived().template writePacket<LoadMode>(index, x);
     }
 
+    const typename internal::remove_all<NestedExpressionType>::type& 
+    nestedExpression() const 
+    {
+      return m_expression;
+    }
+
   protected:
-    const NestedExpressionType m_expression;
+    NestedExpressionType m_expression;
 };
 
 #endif // EIGEN_ARRAYWRAPPER_H

Eigen/src/Core/Assign.h

@@ -152,7 +152,7 @@ struct assign_DefaultTraversal_CompleteUnrolling
     inner = Index % Derived1::InnerSizeAtCompileTime
   };
 
-  EIGEN_STRONG_INLINE static void run(Derived1 &dst, const Derived2 &src)
+  static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src)
   {
     dst.copyCoeffByOuterInner(outer, inner, src);
     assign_DefaultTraversal_CompleteUnrolling<Derived1, Derived2, Index+1, Stop>::run(dst, src);
@@ -162,13 +162,13 @@ struct assign_DefaultTraversal_CompleteUnrolling
 template<typename Derived1, typename Derived2, int Stop>
 struct assign_DefaultTraversal_CompleteUnrolling<Derived1, Derived2, Stop, Stop>
 {
-  EIGEN_STRONG_INLINE static void run(Derived1 &, const Derived2 &) {}
+  static EIGEN_STRONG_INLINE void run(Derived1 &, const Derived2 &) {}
 };
 
 template<typename Derived1, typename Derived2, int Index, int Stop>
 struct assign_DefaultTraversal_InnerUnrolling
 {
-  EIGEN_STRONG_INLINE static void run(Derived1 &dst, const Derived2 &src, int outer)
+  static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src, int outer)
   {
     dst.copyCoeffByOuterInner(outer, Index, src);
     assign_DefaultTraversal_InnerUnrolling<Derived1, Derived2, Index+1, Stop>::run(dst, src, outer);
@@ -178,7 +178,7 @@ struct assign_DefaultTraversal_InnerUnrolling
 template<typename Derived1, typename Derived2, int Stop>
 struct assign_DefaultTraversal_InnerUnrolling<Derived1, Derived2, Stop, Stop>
 {
-  EIGEN_STRONG_INLINE static void run(Derived1 &, const Derived2 &, int) {}
+  static EIGEN_STRONG_INLINE void run(Derived1 &, const Derived2 &, int) {}
 };
 
 /***********************
@@ -188,7 +188,7 @@ struct assign_DefaultTraversal_InnerUnrolling<Derived1, Derived2, Stop, Stop>
 template<typename Derived1, typename Derived2, int Index, int Stop>
 struct assign_LinearTraversal_CompleteUnrolling
 {
-  EIGEN_STRONG_INLINE static void run(Derived1 &dst, const Derived2 &src)
+  static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src)
   {
     dst.copyCoeff(Index, src);
     assign_LinearTraversal_CompleteUnrolling<Derived1, Derived2, Index+1, Stop>::run(dst, src);
@@ -198,7 +198,7 @@ struct assign_LinearTraversal_CompleteUnrolling
 template<typename Derived1, typename Derived2, int Stop>
 struct assign_LinearTraversal_CompleteUnrolling<Derived1, Derived2, Stop, Stop>
 {
-  EIGEN_STRONG_INLINE static void run(Derived1 &, const Derived2 &) {}
+  static EIGEN_STRONG_INLINE void run(Derived1 &, const Derived2 &) {}
 };
 
 /**************************
@@ -214,7 +214,7 @@ struct assign_innervec_CompleteUnrolling
     JointAlignment = assign_traits<Derived1,Derived2>::JointAlignment
   };
 
-  EIGEN_STRONG_INLINE static void run(Derived1 &dst, const Derived2 &src)
+  static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src)
   {
     dst.template copyPacketByOuterInner<Derived2, Aligned, JointAlignment>(outer, inner, src);
     assign_innervec_CompleteUnrolling<Derived1, Derived2,
@@ -225,13 +225,13 @@ struct assign_innervec_CompleteUnrolling
 template<typename Derived1, typename Derived2, int Stop>
 struct assign_innervec_CompleteUnrolling<Derived1, Derived2, Stop, Stop>
 {
-  EIGEN_STRONG_INLINE static void run(Derived1 &, const Derived2 &) {}
+  static EIGEN_STRONG_INLINE void run(Derived1 &, const Derived2 &) {}
 };
 
 template<typename Derived1, typename Derived2, int Index, int Stop>
 struct assign_innervec_InnerUnrolling
 {
-  EIGEN_STRONG_INLINE static void run(Derived1 &dst, const Derived2 &src, int outer)
+  static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src, int outer)
   {
     dst.template copyPacketByOuterInner<Derived2, Aligned, Aligned>(outer, Index, src);
     assign_innervec_InnerUnrolling<Derived1, Derived2,
@@ -242,7 +242,7 @@ struct assign_innervec_InnerUnrolling
 template<typename Derived1, typename Derived2, int Stop>
 struct assign_innervec_InnerUnrolling<Derived1, Derived2, Stop, Stop>
 {
-  EIGEN_STRONG_INLINE static void run(Derived1 &, const Derived2 &, int) {}
+  static EIGEN_STRONG_INLINE void run(Derived1 &, const Derived2 &, int) {}
 };
 
 /***************************************************************************
@@ -251,24 +251,25 @@ struct assign_innervec_InnerUnrolling<Derived1, Derived2, Stop, Stop>
 
 template<typename Derived1, typename Derived2,
          int Traversal = assign_traits<Derived1, Derived2>::Traversal,
-         int Unrolling = assign_traits<Derived1, Derived2>::Unrolling>
+         int Unrolling = assign_traits<Derived1, Derived2>::Unrolling,
+         int Version = Specialized>
 struct assign_impl;
 
 /************************
 *** Default traversal ***
 ************************/
 
-template<typename Derived1, typename Derived2, int Unrolling>
-struct assign_impl<Derived1, Derived2, InvalidTraversal, Unrolling>
+template<typename Derived1, typename Derived2, int Unrolling, int Version>
+struct assign_impl<Derived1, Derived2, InvalidTraversal, Unrolling, Version>
 {
-  inline static void run(Derived1 &, const Derived2 &) { }
+  static inline void run(Derived1 &, const Derived2 &) { }
 };
 
-template<typename Derived1, typename Derived2>
-struct assign_impl<Derived1, Derived2, DefaultTraversal, NoUnrolling>
+template<typename Derived1, typename Derived2, int Version>
+struct assign_impl<Derived1, Derived2, DefaultTraversal, NoUnrolling, Version>
 {
   typedef typename Derived1::Index Index;
-  inline static void run(Derived1 &dst, const Derived2 &src)
+  static inline void run(Derived1 &dst, const Derived2 &src)
   {
     const Index innerSize = dst.innerSize();
     const Index outerSize = dst.outerSize();
@@ -278,21 +279,21 @@ struct assign_impl<Derived1, Derived2, DefaultTraversal, NoUnrolling>
   }
 };
 
-template<typename Derived1, typename Derived2>
-struct assign_impl<Derived1, Derived2, DefaultTraversal, CompleteUnrolling>
+template<typename Derived1, typename Derived2, int Version>
+struct assign_impl<Derived1, Derived2, DefaultTraversal, CompleteUnrolling, Version>
 {
-  EIGEN_STRONG_INLINE static void run(Derived1 &dst, const Derived2 &src)
+  static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src)
   {
     assign_DefaultTraversal_CompleteUnrolling<Derived1, Derived2, 0, Derived1::SizeAtCompileTime>
       ::run(dst, src);
   }
 };
 
-template<typename Derived1, typename Derived2>
-struct assign_impl<Derived1, Derived2, DefaultTraversal, InnerUnrolling>
+template<typename Derived1, typename Derived2, int Version>
+struct assign_impl<Derived1, Derived2, DefaultTraversal, InnerUnrolling, Version>
 {
   typedef typename Derived1::Index Index;
-  EIGEN_STRONG_INLINE static void run(Derived1 &dst, const Derived2 &src)
+  static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src)
   {
     const Index outerSize = dst.outerSize();
     for(Index outer = 0; outer < outerSize; ++outer)
@@ -305,11 +306,11 @@ struct assign_impl<Derived1, Derived2, DefaultTraversal, InnerUnrolling>
 *** Linear traversal ***
 ***********************/
 
-template<typename Derived1, typename Derived2>
-struct assign_impl<Derived1, Derived2, LinearTraversal, NoUnrolling>
+template<typename Derived1, typename Derived2, int Version>
+struct assign_impl<Derived1, Derived2, LinearTraversal, NoUnrolling, Version>
 {
   typedef typename Derived1::Index Index;
-  inline static void run(Derived1 &dst, const Derived2 &src)
+  static inline void run(Derived1 &dst, const Derived2 &src)
   {
     const Index size = dst.size();
     for(Index i = 0; i < size; ++i)
@@ -317,10 +318,10 @@ struct assign_impl<Derived1, Derived2, LinearTraversal, NoUnrolling>
   }
 };
 
-template<typename Derived1, typename Derived2>
-struct assign_impl<Derived1, Derived2, LinearTraversal, CompleteUnrolling>
+template<typename Derived1, typename Derived2, int Version>
+struct assign_impl<Derived1, Derived2, LinearTraversal, CompleteUnrolling, Version>
 {
-  EIGEN_STRONG_INLINE static void run(Derived1 &dst, const Derived2 &src)
+  static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src)
   {
     assign_LinearTraversal_CompleteUnrolling<Derived1, Derived2, 0, Derived1::SizeAtCompileTime>
       ::run(dst, src);
@@ -331,11 +332,11 @@ struct assign_impl<Derived1, Derived2, LinearTraversal, CompleteUnrolling>
 *** Inner vectorization ***
 **************************/
 
-template<typename Derived1, typename Derived2>
-struct assign_impl<Derived1, Derived2, InnerVectorizedTraversal, NoUnrolling>
+template<typename Derived1, typename Derived2, int Version>
+struct assign_impl<Derived1, Derived2, InnerVectorizedTraversal, NoUnrolling, Version>
 {
   typedef typename Derived1::Index Index;
-  inline static void run(Derived1 &dst, const Derived2 &src)
+  static inline void run(Derived1 &dst, const Derived2 &src)
   {
     const Index innerSize = dst.innerSize();
     const Index outerSize = dst.outerSize();
@@ -346,21 +347,21 @@ struct assign_impl<Derived1, Derived2, InnerVectorizedTraversal, NoUnrolling>
   }
 };
 
-template<typename Derived1, typename Derived2>
-struct assign_impl<Derived1, Derived2, InnerVectorizedTraversal, CompleteUnrolling>
+template<typename Derived1, typename Derived2, int Version>
+struct assign_impl<Derived1, Derived2, InnerVectorizedTraversal, CompleteUnrolling, Version>
 {
-  EIGEN_STRONG_INLINE static void run(Derived1 &dst, const Derived2 &src)
+  static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src)
   {
     assign_innervec_CompleteUnrolling<Derived1, Derived2, 0, Derived1::SizeAtCompileTime>
       ::run(dst, src);
   }
 };
 
-template<typename Derived1, typename Derived2>
-struct assign_impl<Derived1, Derived2, InnerVectorizedTraversal, InnerUnrolling>
+template<typename Derived1, typename Derived2, int Version>
+struct assign_impl<Derived1, Derived2, InnerVectorizedTraversal, InnerUnrolling, Version>
 {
   typedef typename Derived1::Index Index;
-  EIGEN_STRONG_INLINE static void run(Derived1 &dst, const Derived2 &src)
+  static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src)
   {
     const Index outerSize = dst.outerSize();
     for(Index outer = 0; outer < outerSize; ++outer)
@@ -398,11 +399,11 @@ struct unaligned_assign_impl<false>
   }
 };
 
-template<typename Derived1, typename Derived2>
-struct assign_impl<Derived1, Derived2, LinearVectorizedTraversal, NoUnrolling>
+template<typename Derived1, typename Derived2, int Version>
+struct assign_impl<Derived1, Derived2, LinearVectorizedTraversal, NoUnrolling, Version>
 {
   typedef typename Derived1::Index Index;
-  EIGEN_STRONG_INLINE static void run(Derived1 &dst, const Derived2 &src)
+  static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src)
   {
     const Index size = dst.size();
     typedef packet_traits<typename Derived1::Scalar> PacketTraits;
@@ -412,7 +413,7 @@ struct assign_impl<Derived1, Derived2, LinearVectorizedTraversal, NoUnrolling>
       srcAlignment = assign_traits<Derived1,Derived2>::JointAlignment
     };
     const Index alignedStart = assign_traits<Derived1,Derived2>::DstIsAligned ? 0
-                             : first_aligned(&dst.coeffRef(0), size);
+                             : internal::first_aligned(&dst.coeffRef(0), size);
     const Index alignedEnd = alignedStart + ((size-alignedStart)/packetSize)*packetSize;
 
     unaligned_assign_impl<assign_traits<Derived1,Derived2>::DstIsAligned!=0>::run(src,dst,0,alignedStart);
@@ -426,11 +427,11 @@ struct assign_impl<Derived1, Derived2, LinearVectorizedTraversal, NoUnrolling>
   }
 };
 
-template<typename Derived1, typename Derived2>
-struct assign_impl<Derived1, Derived2, LinearVectorizedTraversal, CompleteUnrolling>
+template<typename Derived1, typename Derived2, int Version>
+struct assign_impl<Derived1, Derived2, LinearVectorizedTraversal, CompleteUnrolling, Version>
 {
   typedef typename Derived1::Index Index;
-  EIGEN_STRONG_INLINE static void run(Derived1 &dst, const Derived2 &src)
+  static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src)
   {
     enum { size = Derived1::SizeAtCompileTime,
            packetSize = packet_traits<typename Derived1::Scalar>::size,
@@ -445,11 +446,11 @@ struct assign_impl<Derived1, Derived2, LinearVectorizedTraversal, CompleteUnroll
 *** Slice vectorization ***
 ***************************/
 
-template<typename Derived1, typename Derived2>
-struct assign_impl<Derived1, Derived2, SliceVectorizedTraversal, NoUnrolling>
+template<typename Derived1, typename Derived2, int Version>
+struct assign_impl<Derived1, Derived2, SliceVectorizedTraversal, NoUnrolling, Version>
 {
   typedef typename Derived1::Index Index;
-  inline static void run(Derived1 &dst, const Derived2 &src)
+  static inline void run(Derived1 &dst, const Derived2 &src)
   {
     typedef packet_traits<typename Derived1::Scalar> PacketTraits;
     enum {
@@ -463,7 +464,7 @@ struct assign_impl<Derived1, Derived2, SliceVectorizedTraversal, NoUnrolling>
     const Index outerSize = dst.outerSize();
     const Index alignedStep = alignable ? (packetSize - dst.outerStride() % packetSize) & packetAlignedMask : 0;
     Index alignedStart = ((!alignable) || assign_traits<Derived1,Derived2>::DstIsAligned) ? 0
-                       : first_aligned(&dst.coeffRef(0,0), innerSize);
+                       : internal::first_aligned(&dst.coeffRef(0,0), innerSize);
 
     for(Index outer = 0; outer < outerSize; ++outer)
     {
@@ -531,19 +532,19 @@ struct assign_selector;
 
 template<typename Derived, typename OtherDerived>
 struct assign_selector<Derived,OtherDerived,false,false> {
-  EIGEN_STRONG_INLINE static Derived& run(Derived& dst, const OtherDerived& other) { return dst.lazyAssign(other.derived()); }
+  static EIGEN_STRONG_INLINE Derived& run(Derived& dst, const OtherDerived& other) { return dst.lazyAssign(other.derived()); }
 };
 template<typename Derived, typename OtherDerived>
 struct assign_selector<Derived,OtherDerived,true,false> {
-  EIGEN_STRONG_INLINE static Derived& run(Derived& dst, const OtherDerived& other) { return dst.lazyAssign(other.eval()); }
+  static EIGEN_STRONG_INLINE Derived& run(Derived& dst, const OtherDerived& other) { return dst.lazyAssign(other.eval()); }
 };
 template<typename Derived, typename OtherDerived>
 struct assign_selector<Derived,OtherDerived,false,true> {
-  EIGEN_STRONG_INLINE static Derived& run(Derived& dst, const OtherDerived& other) { return dst.lazyAssign(other.transpose()); }
+  static EIGEN_STRONG_INLINE Derived& run(Derived& dst, const OtherDerived& other) { return dst.lazyAssign(other.transpose()); }
 };
 template<typename Derived, typename OtherDerived>
 struct assign_selector<Derived,OtherDerived,true,true> {
-  EIGEN_STRONG_INLINE static Derived& run(Derived& dst, const OtherDerived& other) { return dst.lazyAssign(other.transpose().eval()); }
+  static EIGEN_STRONG_INLINE Derived& run(Derived& dst, const OtherDerived& other) { return dst.lazyAssign(other.transpose().eval()); }
 };
 
 } // end namespace internal

Eigen/src/Core/Assign_MKL.h

@@ -0,0 +1,217 @@
+/*
+ Copyright (c) 2011, Intel Corporation. All rights reserved.
+
+ Redistribution and use in source and binary forms, with or without modification,
+ are permitted provided that the following conditions are met:
+
+ * Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+ * Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+ * Neither the name of Intel Corporation nor the names of its contributors may
+   be used to endorse or promote products derived from this software without
+   specific prior written permission.
+
+ THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+ WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+ DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+ (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+ LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+ ********************************************************************************
+ *   Content : Eigen bindings to Intel(R) MKL
+ *   MKL VML support for coefficient-wise unary Eigen expressions like a=b.sin()
+ ********************************************************************************
+*/
+
+#ifndef EIGEN_ASSIGN_VML_H
+#define EIGEN_ASSIGN_VML_H
+
+namespace internal {
+
+template<typename Op> struct vml_call
+{ enum { IsSupported = 0 }; };
+
+template<typename Dst, typename Src, typename UnaryOp>
+class vml_assign_traits
+{
+  private:
+    enum {
+      DstHasDirectAccess = Dst::Flags & DirectAccessBit,
+      SrcHasDirectAccess = Src::Flags & DirectAccessBit,
+
+      StorageOrdersAgree = (int(Dst::IsRowMajor) == int(Src::IsRowMajor)),
+      InnerSize = int(Dst::IsVectorAtCompileTime) ? int(Dst::SizeAtCompileTime)
+                : int(Dst::Flags)&RowMajorBit ? int(Dst::ColsAtCompileTime)
+                : int(Dst::RowsAtCompileTime),
+      InnerMaxSize  = int(Dst::IsVectorAtCompileTime) ? int(Dst::MaxSizeAtCompileTime)
+                    : int(Dst::Flags)&RowMajorBit ? int(Dst::MaxColsAtCompileTime)
+                    : int(Dst::MaxRowsAtCompileTime),
+      MaxSizeAtCompileTime = Dst::SizeAtCompileTime,
+
+      MightEnableVml =  vml_call<UnaryOp>::IsSupported && StorageOrdersAgree && DstHasDirectAccess && SrcHasDirectAccess
+                     && Src::InnerStrideAtCompileTime==1 && Dst::InnerStrideAtCompileTime==1,
+      MightLinearize = MightEnableVml && (int(Dst::Flags) & int(Src::Flags) & LinearAccessBit),
+      VmlSize = MightLinearize ? MaxSizeAtCompileTime : InnerMaxSize,
+      LargeEnough = VmlSize==Dynamic || VmlSize>=EIGEN_MKL_VML_THRESHOLD,
+      MayEnableVml = MightEnableVml && LargeEnough,
+      MayLinearize = MayEnableVml && MightLinearize
+    };
+  public:
+    enum {
+      Traversal = MayLinearize ? LinearVectorizedTraversal
+                : MayEnableVml ? InnerVectorizedTraversal
+                : DefaultTraversal
+    };
+};
+
+template<typename Derived1, typename Derived2, typename UnaryOp, int Traversal, int Unrolling,
+         int VmlTraversal = vml_assign_traits<Derived1, Derived2, UnaryOp>::Traversal >
+struct vml_assign_impl
+  : assign_impl<Derived1, Eigen::CwiseUnaryOp<UnaryOp, Derived2>,Traversal,Unrolling,BuiltIn>
+{
+};
+
+template<typename Derived1, typename Derived2, typename UnaryOp, int Traversal, int Unrolling>
+struct vml_assign_impl<Derived1, Derived2, UnaryOp, Traversal, Unrolling, InnerVectorizedTraversal>
+{
+  typedef typename Derived1::Scalar Scalar;
+  typedef typename Derived1::Index Index;
+  static inline void run(Derived1& dst, const CwiseUnaryOp<UnaryOp, Derived2>& src)
+  {
+    // in case we want to (or have to) skip VML at runtime we can call:
+    // assign_impl<Derived1,Eigen::CwiseUnaryOp<UnaryOp, Derived2>,Traversal,Unrolling,BuiltIn>::run(dst,src);
+    const Index innerSize = dst.innerSize();
+    const Index outerSize = dst.outerSize();
+    for(Index outer = 0; outer < outerSize; ++outer) {
+      const Scalar *src_ptr = src.IsRowMajor ?  &(src.nestedExpression().coeffRef(outer,0)) :
+                                                &(src.nestedExpression().coeffRef(0, outer));
+      Scalar *dst_ptr = dst.IsRowMajor ? &(dst.coeffRef(outer,0)) : &(dst.coeffRef(0, outer));
+      vml_call<UnaryOp>::run(src.functor(), innerSize, src_ptr, dst_ptr );
+    }
+  }
+};
+
+template<typename Derived1, typename Derived2, typename UnaryOp, int Traversal, int Unrolling>
+struct vml_assign_impl<Derived1, Derived2, UnaryOp, Traversal, Unrolling, LinearVectorizedTraversal>
+{
+  static inline void run(Derived1& dst, const CwiseUnaryOp<UnaryOp, Derived2>& src)
+  {
+    // in case we want to (or have to) skip VML at runtime we can call:
+    // assign_impl<Derived1,Eigen::CwiseUnaryOp<UnaryOp, Derived2>,Traversal,Unrolling,BuiltIn>::run(dst,src);
+    vml_call<UnaryOp>::run(src.functor(), dst.size(), src.nestedExpression().data(), dst.data() );
+  }
+};
+
+// Macroses
+
+#define EIGEN_MKL_VML_SPECIALIZE_ASSIGN(TRAVERSAL,UNROLLING) \
+  template<typename Derived1, typename Derived2, typename UnaryOp> \
+  struct assign_impl<Derived1, Eigen::CwiseUnaryOp<UnaryOp, Derived2>, TRAVERSAL, UNROLLING, Specialized>  {  \
+    static inline void run(Derived1 &dst, const Eigen::CwiseUnaryOp<UnaryOp, Derived2> &src) { \
+      vml_assign_impl<Derived1,Derived2,UnaryOp,TRAVERSAL,UNROLLING>::run(dst, src); \
+    } \
+  };
+
+EIGEN_MKL_VML_SPECIALIZE_ASSIGN(DefaultTraversal,NoUnrolling)
+EIGEN_MKL_VML_SPECIALIZE_ASSIGN(DefaultTraversal,CompleteUnrolling)
+EIGEN_MKL_VML_SPECIALIZE_ASSIGN(DefaultTraversal,InnerUnrolling)
+EIGEN_MKL_VML_SPECIALIZE_ASSIGN(LinearTraversal,NoUnrolling)
+EIGEN_MKL_VML_SPECIALIZE_ASSIGN(LinearTraversal,CompleteUnrolling)
+EIGEN_MKL_VML_SPECIALIZE_ASSIGN(InnerVectorizedTraversal,NoUnrolling)
+EIGEN_MKL_VML_SPECIALIZE_ASSIGN(InnerVectorizedTraversal,CompleteUnrolling)
+EIGEN_MKL_VML_SPECIALIZE_ASSIGN(InnerVectorizedTraversal,InnerUnrolling)
+EIGEN_MKL_VML_SPECIALIZE_ASSIGN(LinearVectorizedTraversal,CompleteUnrolling)
+EIGEN_MKL_VML_SPECIALIZE_ASSIGN(LinearVectorizedTraversal,NoUnrolling)
+EIGEN_MKL_VML_SPECIALIZE_ASSIGN(SliceVectorizedTraversal,NoUnrolling)
+
+
+#if !defined (EIGEN_FAST_MATH) || (EIGEN_FAST_MATH != 1)
+#define  EIGEN_MKL_VML_MODE VML_HA
+#else
+#define  EIGEN_MKL_VML_MODE VML_LA
+#endif
+
+#define EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, VMLOP, EIGENTYPE, VMLTYPE)     \
+  template<> struct vml_call< scalar_##EIGENOP##_op<EIGENTYPE> > {               \
+    enum { IsSupported = 1 };                                                    \
+    static inline void run( const scalar_##EIGENOP##_op<EIGENTYPE>& func,        \
+                            int size, const EIGENTYPE* src, EIGENTYPE* dst) {    \
+      VMLOP(size, (const VMLTYPE*)src, (VMLTYPE*)dst);                           \
+    }                                                                            \
+  };
+
+#define EIGEN_MKL_VML_DECLARE_UNARY_CALL_LA(EIGENOP, VMLOP, EIGENTYPE, VMLTYPE)  \
+  template<> struct vml_call< scalar_##EIGENOP##_op<EIGENTYPE> > {               \
+    enum { IsSupported = 1 };                                                    \
+    static inline void run( const scalar_##EIGENOP##_op<EIGENTYPE>& func,        \
+                            int size, const EIGENTYPE* src, EIGENTYPE* dst) {    \
+      MKL_INT64 vmlMode = EIGEN_MKL_VML_MODE;                                    \
+      VMLOP(size, (const VMLTYPE*)src, (VMLTYPE*)dst, vmlMode);                  \
+    }                                                                            \
+  };
+
+#define EIGEN_MKL_VML_DECLARE_POW_CALL(EIGENOP, VMLOP, EIGENTYPE, VMLTYPE)       \
+  template<> struct vml_call< scalar_##EIGENOP##_op<EIGENTYPE> > {               \
+    enum { IsSupported = 1 };                                                    \
+    static inline void run( const scalar_##EIGENOP##_op<EIGENTYPE>& func,        \
+                          int size, const EIGENTYPE* src, EIGENTYPE* dst) {      \
+      EIGENTYPE exponent = func.m_exponent;                                      \
+      MKL_INT64 vmlMode = EIGEN_MKL_VML_MODE;                                    \
+      VMLOP(&size, (const VMLTYPE*)src, (const VMLTYPE*)&exponent,               \
+                        (VMLTYPE*)dst, &vmlMode);                                \
+    }                                                                            \
+  };
+
+#define EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(EIGENOP, VMLOP)                   \
+  EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, vs##VMLOP, float, float)             \
+  EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, vd##VMLOP, double, double)
+
+#define EIGEN_MKL_VML_DECLARE_UNARY_CALLS_COMPLEX(EIGENOP, VMLOP)                \
+  EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, vc##VMLOP, scomplex, MKL_Complex8)   \
+  EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, vz##VMLOP, dcomplex, MKL_Complex16)
+
+#define EIGEN_MKL_VML_DECLARE_UNARY_CALLS(EIGENOP, VMLOP)                        \
+  EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(EIGENOP, VMLOP)                         \
+  EIGEN_MKL_VML_DECLARE_UNARY_CALLS_COMPLEX(EIGENOP, VMLOP)
+
+
+#define EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL_LA(EIGENOP, VMLOP)                \
+  EIGEN_MKL_VML_DECLARE_UNARY_CALL_LA(EIGENOP, vms##VMLOP, float, float)         \
+  EIGEN_MKL_VML_DECLARE_UNARY_CALL_LA(EIGENOP, vmd##VMLOP, double, double)
+
+#define EIGEN_MKL_VML_DECLARE_UNARY_CALLS_COMPLEX_LA(EIGENOP, VMLOP)             \
+  EIGEN_MKL_VML_DECLARE_UNARY_CALL_LA(EIGENOP, vmc##VMLOP, scomplex, MKL_Complex8)  \
+  EIGEN_MKL_VML_DECLARE_UNARY_CALL_LA(EIGENOP, vmz##VMLOP, dcomplex, MKL_Complex16)
+
+#define EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(EIGENOP, VMLOP)                     \
+  EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL_LA(EIGENOP, VMLOP)                      \
+  EIGEN_MKL_VML_DECLARE_UNARY_CALLS_COMPLEX_LA(EIGENOP, VMLOP)
+
+
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(sin,  Sin)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(asin, Asin)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(cos,  Cos)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(acos, Acos)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(tan,  Tan)
+//EIGEN_MKL_VML_DECLARE_UNARY_CALLS(abs,  Abs)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(exp,  Exp)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(log,  Ln)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(sqrt, Sqrt)
+
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(square, Sqr)
+
+EIGEN_MKL_VML_DECLARE_POW_CALL(pow, vmspowx_, float, float)
+EIGEN_MKL_VML_DECLARE_POW_CALL(pow, vmdpowx_, double, double)
+EIGEN_MKL_VML_DECLARE_POW_CALL(pow, vmcpowx_, scomplex, MKL_Complex8)
+EIGEN_MKL_VML_DECLARE_POW_CALL(pow, vmzpowx_, dcomplex, MKL_Complex16)
+
+} // end namespace internal
+
+#endif // EIGEN_ASSIGN_VML_H

Eigen/src/Core/Block.h

@@ -94,7 +94,7 @@ struct traits<Block<XprType, BlockRows, BlockCols, InnerPanel, HasDirectAccess>
     MaskPacketAccessBit = (InnerSize == Dynamic || (InnerSize % packet_traits<Scalar>::size) == 0)
                        && (InnerStrideAtCompileTime == 1)
                         ? PacketAccessBit : 0,
-    MaskAlignedBit = (InnerPanel && (OuterStrideAtCompileTime!=Dynamic) && (((OuterStrideAtCompileTime * sizeof(Scalar)) % 16) == 0)) ? AlignedBit : 0,
+    MaskAlignedBit = (InnerPanel && (OuterStrideAtCompileTime!=Dynamic) && (((OuterStrideAtCompileTime * int(sizeof(Scalar))) % 16) == 0)) ? AlignedBit : 0,
     FlagsLinearAccessBit = (RowsAtCompileTime == 1 || ColsAtCompileTime == 1) ? LinearAccessBit : 0,
     FlagsLvalueBit = is_lvalue<XprType>::value ? LvalueBit : 0,
     FlagsRowMajorBit = IsRowMajor ? RowMajorBit : 0,
@@ -242,6 +242,21 @@ template<typename XprType, int BlockRows, int BlockCols, bool InnerPanel, bool H
     inline Index outerStride() const;
     #endif
 
+    const typename internal::remove_all<typename XprType::Nested>::type& nestedExpression() const 
+    { 
+      return m_xpr; 
+    }
+      
+    Index startRow() const 
+    { 
+      return m_startRow.value(); 
+    }
+      
+    Index startCol() const 
+    { 
+      return m_startCol.value(); 
+    }
+
   protected:
 
     const typename XprType::Nested m_xpr;
@@ -304,6 +319,11 @@ class Block<XprType,BlockRows,BlockCols, InnerPanel,true>
       init();
     }
 
+    const typename internal::remove_all<typename XprType::Nested>::type& nestedExpression() const 
+    { 
+      return m_xpr; 
+    }
+      
     /** \sa MapBase::innerStride() */
     inline Index innerStride() const
     {
@@ -341,7 +361,7 @@ class Block<XprType,BlockRows,BlockCols, InnerPanel,true>
                     : m_xpr.innerStride();
     }
 
-    const typename XprType::Nested m_xpr;
+    typename XprType::Nested m_xpr;
     int m_outerStride;
 };
 

Eigen/src/Core/BooleanRedux.h

@@ -35,7 +35,7 @@ struct all_unroller
     row = (UnrollCount-1) % Derived::RowsAtCompileTime
   };
 
-  inline static bool run(const Derived &mat)
+  static inline bool run(const Derived &mat)
   {
     return all_unroller<Derived, UnrollCount-1>::run(mat) && mat.coeff(row, col);
   }
@@ -44,13 +44,13 @@ struct all_unroller
 template<typename Derived>
 struct all_unroller<Derived, 1>
 {
-  inline static bool run(const Derived &mat) { return mat.coeff(0, 0); }
+  static inline bool run(const Derived &mat) { return mat.coeff(0, 0); }
 };
 
 template<typename Derived>
 struct all_unroller<Derived, Dynamic>
 {
-  inline static bool run(const Derived &) { return false; }
+  static inline bool run(const Derived &) { return false; }
 };
 
 template<typename Derived, int UnrollCount>
@@ -61,7 +61,7 @@ struct any_unroller
     row = (UnrollCount-1) % Derived::RowsAtCompileTime
   };
 
-  inline static bool run(const Derived &mat)
+  static inline bool run(const Derived &mat)
   {
     return any_unroller<Derived, UnrollCount-1>::run(mat) || mat.coeff(row, col);
   }
@@ -70,13 +70,13 @@ struct any_unroller
 template<typename Derived>
 struct any_unroller<Derived, 1>
 {
-  inline static bool run(const Derived &mat) { return mat.coeff(0, 0); }
+  static inline bool run(const Derived &mat) { return mat.coeff(0, 0); }
 };
 
 template<typename Derived>
 struct any_unroller<Derived, Dynamic>
 {
-  inline static bool run(const Derived &) { return false; }
+  static inline bool run(const Derived &) { return false; }
 };
 
 } // end namespace internal

Eigen/src/Core/CwiseBinaryOp.h

@@ -167,8 +167,8 @@ class CwiseBinaryOp : internal::no_assignment_operator,
     const BinaryOp& functor() const { return m_functor; }
 
   protected:
-    const LhsNested m_lhs;
-    const RhsNested m_rhs;
+    LhsNested m_lhs;
+    RhsNested m_rhs;
     const BinaryOp m_functor;
 };
 

Eigen/src/Core/CwiseNullaryOp.h

@@ -101,6 +101,9 @@ class CwiseNullaryOp : internal::no_assignment_operator,
       return m_functor.packetOp(index);
     }
 
+    /** \returns the functor representing the nullary operation */
+    const NullaryOp& functor() const { return m_functor; }
+
   protected:
     const internal::variable_if_dynamic<Index, RowsAtCompileTime> m_rows;
     const internal::variable_if_dynamic<Index, ColsAtCompileTime> m_cols;

Eigen/src/Core/CwiseUnaryOp.h

@@ -95,7 +95,7 @@ class CwiseUnaryOp : internal::no_assignment_operator,
     nestedExpression() { return m_xpr.const_cast_derived(); }
 
   protected:
-    const typename XprType::Nested m_xpr;
+    typename XprType::Nested m_xpr;
     const UnaryOp m_functor;
 };
 

Eigen/src/Core/CwiseUnaryView.h

@@ -97,7 +97,7 @@ class CwiseUnaryView : internal::no_assignment_operator,
 
   protected:
     // FIXME changed from MatrixType::Nested because of a weird compilation error with sun CC
-    const typename internal::nested<MatrixType>::type m_matrix;
+    typename internal::nested<MatrixType>::type m_matrix;
     ViewOp m_functor;
 };
 

Eigen/src/Core/DenseBase.h

@@ -376,12 +376,13 @@ template<typename Derived> class DenseBase
     inline Derived& operator*=(const Scalar& other);
     inline Derived& operator/=(const Scalar& other);
 
+    typedef typename internal::add_const_on_value_type<typename internal::eval<Derived>::type>::type EvalReturnType;
     /** \returns the matrix or vector obtained by evaluating this expression.
       *
       * Notice that in the case of a plain matrix or vector (not an expression) this function just returns
       * a const reference, in order to avoid a useless copy.
       */
-    EIGEN_STRONG_INLINE const typename internal::eval<Derived>::type eval() const
+    EIGEN_STRONG_INLINE EvalReturnType eval() const
     {
       // Even though MSVC does not honor strong inlining when the return type
       // is a dynamic matrix, we desperately need strong inlining for fixed

Eigen/src/Core/DenseCoeffsBase.h

@@ -710,16 +710,16 @@ namespace internal {
 template<typename Derived, bool JustReturnZero>
 struct first_aligned_impl
 {
-  inline static typename Derived::Index run(const Derived&)
+  static inline typename Derived::Index run(const Derived&)
   { return 0; }
 };
 
 template<typename Derived>
 struct first_aligned_impl<Derived, false>
 {
-  inline static typename Derived::Index run(const Derived& m)
+  static inline typename Derived::Index run(const Derived& m)
   {
-    return first_aligned(&m.const_cast_derived().coeffRef(0,0), m.size());
+    return internal::first_aligned(&m.const_cast_derived().coeffRef(0,0), m.size());
   }
 };
 
@@ -729,7 +729,7 @@ struct first_aligned_impl<Derived, false>
   * documentation.
   */
 template<typename Derived>
-inline static typename Derived::Index first_aligned(const Derived& m)
+static inline typename Derived::Index first_aligned(const Derived& m)
 {
   return first_aligned_impl
           <Derived, (Derived::Flags & AlignedBit) || !(Derived::Flags & DirectAccessBit)>

Eigen/src/Core/DenseStorage.h

@@ -104,8 +104,8 @@ template<typename T, int Size, int _Rows, int _Cols, int _Options> class DenseSt
       : m_data(internal::constructor_without_unaligned_array_assert()) {}
     inline DenseStorage(DenseIndex,DenseIndex,DenseIndex) {}
     inline void swap(DenseStorage& other) { std::swap(m_data,other.m_data); }
-    inline static DenseIndex rows(void) {return _Rows;}
-    inline static DenseIndex cols(void) {return _Cols;}
+    static inline DenseIndex rows(void) {return _Rows;}
+    static inline DenseIndex cols(void) {return _Cols;}
     inline void conservativeResize(DenseIndex,DenseIndex,DenseIndex) {}
     inline void resize(DenseIndex,DenseIndex,DenseIndex) {}
     inline const T *data() const { return m_data.array; }
@@ -120,14 +120,24 @@ template<typename T, int _Rows, int _Cols, int _Options> class DenseStorage<T, 0
     inline DenseStorage(internal::constructor_without_unaligned_array_assert) {}
     inline DenseStorage(DenseIndex,DenseIndex,DenseIndex) {}
     inline void swap(DenseStorage& ) {}
-    inline static DenseIndex rows(void) {return _Rows;}
-    inline static DenseIndex cols(void) {return _Cols;}
+    static inline DenseIndex rows(void) {return _Rows;}
+    static inline DenseIndex cols(void) {return _Cols;}
     inline void conservativeResize(DenseIndex,DenseIndex,DenseIndex) {}
     inline void resize(DenseIndex,DenseIndex,DenseIndex) {}
     inline const T *data() const { return 0; }
     inline T *data() { return 0; }
 };
 
+// more specializations for null matrices; these are necessary to resolve ambiguities
+template<typename T, int _Options> class DenseStorage<T, 0, Dynamic, Dynamic, _Options>
+: public DenseStorage<T, 0, 0, 0, _Options> { };
+
+template<typename T, int _Rows, int _Options> class DenseStorage<T, 0, _Rows, Dynamic, _Options>
+: public DenseStorage<T, 0, 0, 0, _Options> { };
+
+template<typename T, int _Cols, int _Options> class DenseStorage<T, 0, Dynamic, _Cols, _Options>
+: public DenseStorage<T, 0, 0, 0, _Options> { };
+
 // dynamic-size matrix with fixed-size storage
 template<typename T, int Size, int _Options> class DenseStorage<T, Size, Dynamic, Dynamic, _Options>
 {
@@ -241,7 +251,7 @@ template<typename T, int _Rows, int _Options> class DenseStorage<T, Dynamic, _Ro
     { EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN }
     inline ~DenseStorage() { internal::conditional_aligned_delete_auto<T,(_Options&DontAlign)==0>(m_data, _Rows*m_cols); }
     inline void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_cols,other.m_cols); }
-    inline static DenseIndex rows(void) {return _Rows;}
+    static inline DenseIndex rows(void) {return _Rows;}
     inline DenseIndex cols(void) const {return m_cols;}
     inline void conservativeResize(DenseIndex size, DenseIndex, DenseIndex cols)
     {
@@ -278,7 +288,7 @@ template<typename T, int _Cols, int _Options> class DenseStorage<T, Dynamic, Dyn
     inline ~DenseStorage() { internal::conditional_aligned_delete_auto<T,(_Options&DontAlign)==0>(m_data, _Cols*m_rows); }
     inline void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_rows,other.m_rows); }
     inline DenseIndex rows(void) const {return m_rows;}
-    inline static DenseIndex cols(void) {return _Cols;}
+    static inline DenseIndex cols(void) {return _Cols;}
     inline void conservativeResize(DenseIndex size, DenseIndex rows, DenseIndex)
     {
       m_data = internal::conditional_aligned_realloc_new_auto<T,(_Options&DontAlign)==0>(m_data, size, m_rows*_Cols);

Eigen/src/Core/Diagonal.h

@@ -2,6 +2,7 @@
 // for linear algebra.
 //
 // Copyright (C) 2007-2009 Benoit Jacob <jacob.benoit.1@gmail.com>
+// Copyright (C) 2009-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
 //
 // Eigen is free software; you can redistribute it and/or
 // modify it under the terms of the GNU Lesser General Public
@@ -101,6 +102,15 @@ template<typename MatrixType, int DiagIndex> class Diagonal
       return 0;
     }
 
+    typedef typename internal::conditional<
+                       internal::is_lvalue<MatrixType>::value,
+                       Scalar,
+                       const Scalar
+                     >::type ScalarWithConstIfNotLvalue;
+
+    inline ScalarWithConstIfNotLvalue* data() { return &(m_matrix.const_cast_derived().coeffRef(rowOffset(), colOffset())); }
+    inline const Scalar* data() const { return &(m_matrix.const_cast_derived().coeffRef(rowOffset(), colOffset())); }
+
     inline Scalar& coeffRef(Index row, Index)
     {
       EIGEN_STATIC_ASSERT_LVALUE(MatrixType)
@@ -133,8 +143,19 @@ template<typename MatrixType, int DiagIndex> class Diagonal
       return m_matrix.coeff(index+rowOffset(), index+colOffset());
     }
 
+    const typename internal::remove_all<typename MatrixType::Nested>::type& 
+    nestedExpression() const 
+    {
+      return m_matrix;
+    }
+
+    int index() const
+    {
+      return m_index.value();
+    }
+
   protected:
-    const typename MatrixType::Nested m_matrix;
+    typename MatrixType::Nested m_matrix;
     const internal::variable_if_dynamic<Index, DiagIndex> m_index;
 
   private:

Eigen/src/Core/DiagonalMatrix.h

@@ -72,7 +72,7 @@ class DiagonalBase : public EigenBase<Derived>
     const DiagonalProduct<MatrixDerived, Derived, OnTheLeft>
     operator*(const MatrixBase<MatrixDerived> &matrix) const;
 
-    inline const DiagonalWrapper<CwiseUnaryOp<internal::scalar_inverse_op<Scalar>, const DiagonalVectorType> >
+    inline const DiagonalWrapper<const CwiseUnaryOp<internal::scalar_inverse_op<Scalar>, const DiagonalVectorType> >
     inverse() const
     {
       return diagonal().cwiseInverse();
@@ -251,13 +251,13 @@ class DiagonalWrapper
     #endif
 
     /** Constructor from expression of diagonal coefficients to wrap. */
-    inline DiagonalWrapper(const DiagonalVectorType& diagonal) : m_diagonal(diagonal) {}
+    inline DiagonalWrapper(DiagonalVectorType& diagonal) : m_diagonal(diagonal) {}
 
     /** \returns a const reference to the wrapped expression of diagonal coefficients. */
     const DiagonalVectorType& diagonal() const { return m_diagonal; }
 
   protected:
-    const typename DiagonalVectorType::Nested m_diagonal;
+    typename DiagonalVectorType::Nested m_diagonal;
 };
 
 /** \returns a pseudo-expression of a diagonal matrix with *this as vector of diagonal coefficients

Eigen/src/Core/DiagonalProduct.h

@@ -107,8 +107,8 @@ class DiagonalProduct : internal::no_assignment_operator,
                      m_diagonal.diagonal().template packet<DiagonalVectorPacketLoadMode>(id));
     }
 
-    const typename MatrixType::Nested m_matrix;
-    const typename DiagonalType::Nested m_diagonal;
+    typename MatrixType::Nested m_matrix;
+    typename DiagonalType::Nested m_diagonal;
 };
 
 /** \returns the diagonal matrix product of \c *this by the diagonal matrix \a diagonal.

Eigen/src/Core/Dot.h

@@ -176,7 +176,7 @@ template<typename Derived, int p>
 struct lpNorm_selector
 {
   typedef typename NumTraits<typename traits<Derived>::Scalar>::Real RealScalar;
-  inline static RealScalar run(const MatrixBase<Derived>& m)
+  static inline RealScalar run(const MatrixBase<Derived>& m)
   {
     return pow(m.cwiseAbs().array().pow(p).sum(), RealScalar(1)/p);
   }
@@ -185,7 +185,7 @@ struct lpNorm_selector
 template<typename Derived>
 struct lpNorm_selector<Derived, 1>
 {
-  inline static typename NumTraits<typename traits<Derived>::Scalar>::Real run(const MatrixBase<Derived>& m)
+  static inline typename NumTraits<typename traits<Derived>::Scalar>::Real run(const MatrixBase<Derived>& m)
   {
     return m.cwiseAbs().sum();
   }
@@ -194,7 +194,7 @@ struct lpNorm_selector<Derived, 1>
 template<typename Derived>
 struct lpNorm_selector<Derived, 2>
 {
-  inline static typename NumTraits<typename traits<Derived>::Scalar>::Real run(const MatrixBase<Derived>& m)
+  static inline typename NumTraits<typename traits<Derived>::Scalar>::Real run(const MatrixBase<Derived>& m)
   {
     return m.norm();
   }
@@ -203,7 +203,7 @@ struct lpNorm_selector<Derived, 2>
 template<typename Derived>
 struct lpNorm_selector<Derived, Infinity>
 {
-  inline static typename NumTraits<typename traits<Derived>::Scalar>::Real run(const MatrixBase<Derived>& m)
+  static inline typename NumTraits<typename traits<Derived>::Scalar>::Real run(const MatrixBase<Derived>& m)
   {
     return m.cwiseAbs().maxCoeff();
   }

Eigen/src/Core/Functors.h

@@ -220,6 +220,38 @@ struct functor_traits<scalar_quotient_op<Scalar> > {
   };
 };
 
+/** \internal
+  * \brief Template functor to compute the and of two booleans
+  *
+  * \sa class CwiseBinaryOp, ArrayBase::operator&&
+  */
+struct scalar_boolean_and_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_boolean_and_op)
+  EIGEN_STRONG_INLINE bool operator() (const bool& a, const bool& b) const { return a && b; }
+};
+template<> struct functor_traits<scalar_boolean_and_op> {
+  enum {
+    Cost = NumTraits<bool>::AddCost,
+    PacketAccess = false
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute the or of two booleans
+  *
+  * \sa class CwiseBinaryOp, ArrayBase::operator||
+  */
+struct scalar_boolean_or_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_boolean_or_op)
+  EIGEN_STRONG_INLINE bool operator() (const bool& a, const bool& b) const { return a || b; }
+};
+template<> struct functor_traits<scalar_boolean_or_op> {
+  enum {
+    Cost = NumTraits<bool>::AddCost,
+    PacketAccess = false
+  };
+};
+
 // unary functors:
 
 /** \internal

Eigen/src/Core/Fuzzy.h

@@ -35,8 +35,8 @@ struct isApprox_selector
   static bool run(const Derived& x, const OtherDerived& y, typename Derived::RealScalar prec)
   {
     using std::min;
-    const typename internal::nested<Derived,2>::type nested(x);
-    const typename internal::nested<OtherDerived,2>::type otherNested(y);
+    typename internal::nested<Derived,2>::type nested(x);
+    typename internal::nested<OtherDerived,2>::type otherNested(y);
     return (nested - otherNested).cwiseAbs2().sum() <= prec * prec * (min)(nested.cwiseAbs2().sum(), otherNested.cwiseAbs2().sum());
   }
 };

Eigen/src/Core/GeneralProduct.h

@@ -0,0 +1,624 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
+// Copyright (C) 2008-2011 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// Eigen is free software; you can redistribute it and/or
+// modify it under the terms of the GNU Lesser General Public
+// License as published by the Free Software Foundation; either
+// version 3 of the License, or (at your option) any later version.
+//
+// Alternatively, you can redistribute it and/or
+// modify it under the terms of the GNU General Public License as
+// published by the Free Software Foundation; either version 2 of
+// the License, or (at your option) any later version.
+//
+// Eigen is distributed in the hope that it will be useful, but WITHOUT ANY
+// WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+// FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public License or the
+// GNU General Public License for more details.
+//
+// You should have received a copy of the GNU Lesser General Public
+// License and a copy of the GNU General Public License along with
+// Eigen. If not, see <http://www.gnu.org/licenses/>.
+
+#ifndef EIGEN_GENERAL_PRODUCT_H
+#define EIGEN_GENERAL_PRODUCT_H
+
+/** \class GeneralProduct
+  * \ingroup Core_Module
+  *
+  * \brief Expression of the product of two general matrices or vectors
+  *
+  * \param LhsNested the type used to store the left-hand side
+  * \param RhsNested the type used to store the right-hand side
+  * \param ProductMode the type of the product
+  *
+  * This class represents an expression of the product of two general matrices.
+  * We call a general matrix, a dense matrix with full storage. For instance,
+  * This excludes triangular, selfadjoint, and sparse matrices.
+  * It is the return type of the operator* between general matrices. Its template
+  * arguments are determined automatically by ProductReturnType. Therefore,
+  * GeneralProduct should never be used direclty. To determine the result type of a
+  * function which involves a matrix product, use ProductReturnType::Type.
+  *
+  * \sa ProductReturnType, MatrixBase::operator*(const MatrixBase<OtherDerived>&)
+  */
+template<typename Lhs, typename Rhs, int ProductType = internal::product_type<Lhs,Rhs>::value>
+class GeneralProduct;
+
+enum {
+  Large = 2,
+  Small = 3
+};
+
+namespace internal {
+
+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,
+         value = is_large  ? Large
+               : Size == 1 ? 1
+                           : Small
+  };
+};
+
+template<typename Lhs, typename Rhs> struct product_type
+{
+  typedef typename remove_all<Lhs>::type _Lhs;
+  typedef typename remove_all<Rhs>::type _Rhs;
+  enum {
+    MaxRows  = _Lhs::MaxRowsAtCompileTime,
+    Rows  = _Lhs::RowsAtCompileTime,
+    MaxCols  = _Rhs::MaxColsAtCompileTime,
+    Cols  = _Rhs::ColsAtCompileTime,
+    MaxDepth = EIGEN_SIZE_MIN_PREFER_FIXED(_Lhs::MaxColsAtCompileTime,
+                                           _Rhs::MaxRowsAtCompileTime),
+    Depth = EIGEN_SIZE_MIN_PREFER_FIXED(_Lhs::ColsAtCompileTime,
+                                        _Rhs::RowsAtCompileTime),
+    LargeThreshold = EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD
+  };
+
+  // the splitting into different lines of code here, introducing the _select enums and the typedef below,
+  // is to work around an internal compiler error with gcc 4.1 and 4.2.
+private:
+  enum {
+    rows_select = product_size_category<Rows,MaxRows>::value,
+    cols_select = product_size_category<Cols,MaxCols>::value,
+    depth_select = product_size_category<Depth,MaxDepth>::value
+  };
+  typedef product_type_selector<rows_select, cols_select, depth_select> selector;
+
+public:
+  enum {
+    value = selector::ret
+  };
+#ifdef EIGEN_DEBUG_PRODUCT
+  static void debug()
+  {
+      EIGEN_DEBUG_VAR(Rows);
+      EIGEN_DEBUG_VAR(Cols);
+      EIGEN_DEBUG_VAR(Depth);
+      EIGEN_DEBUG_VAR(rows_select);
+      EIGEN_DEBUG_VAR(cols_select);
+      EIGEN_DEBUG_VAR(depth_select);
+      EIGEN_DEBUG_VAR(value);
+  }
+#endif
+};
+
+
+/* The following allows to select the kind of product at compile time
+ * based on the three dimensions of the product.
+ * This is a compile time mapping from {1,Small,Large}^3 -> {product types} */
+// FIXME I'm not sure the current mapping is the ideal one.
+template<int M, int N>  struct product_type_selector<M,N,1>              { enum { ret = OuterProduct }; };
+template<int Depth>     struct product_type_selector<1,    1,    Depth>  { enum { ret = InnerProduct }; };
+template<>              struct product_type_selector<1,    1,    1>      { enum { ret = InnerProduct }; };
+template<>              struct product_type_selector<Small,1,    Small>  { enum { ret = CoeffBasedProductMode }; };
+template<>              struct product_type_selector<1,    Small,Small>  { enum { ret = CoeffBasedProductMode }; };
+template<>              struct product_type_selector<Small,Small,Small>  { enum { ret = CoeffBasedProductMode }; };
+template<>              struct product_type_selector<Small, Small, 1>    { enum { ret = LazyCoeffBasedProductMode }; };
+template<>              struct product_type_selector<Small, Large, 1>    { enum { ret = LazyCoeffBasedProductMode }; };
+template<>              struct product_type_selector<Large, Small, 1>    { enum { ret = LazyCoeffBasedProductMode }; };
+template<>              struct product_type_selector<1,    Large,Small>  { enum { ret = CoeffBasedProductMode }; };
+template<>              struct product_type_selector<1,    Large,Large>  { enum { ret = GemvProduct }; };
+template<>              struct product_type_selector<1,    Small,Large>  { enum { ret = CoeffBasedProductMode }; };
+template<>              struct product_type_selector<Large,1,    Small>  { enum { ret = CoeffBasedProductMode }; };
+template<>              struct product_type_selector<Large,1,    Large>  { enum { ret = GemvProduct }; };
+template<>              struct product_type_selector<Small,1,    Large>  { enum { ret = CoeffBasedProductMode }; };
+template<>              struct product_type_selector<Small,Small,Large>  { enum { ret = GemmProduct }; };
+template<>              struct product_type_selector<Large,Small,Large>  { enum { ret = GemmProduct }; };
+template<>              struct product_type_selector<Small,Large,Large>  { enum { ret = GemmProduct }; };
+template<>              struct product_type_selector<Large,Large,Large>  { enum { ret = GemmProduct }; };
+template<>              struct product_type_selector<Large,Small,Small>  { enum { ret = GemmProduct }; };
+template<>              struct product_type_selector<Small,Large,Small>  { enum { ret = GemmProduct }; };
+template<>              struct product_type_selector<Large,Large,Small>  { enum { ret = GemmProduct }; };
+
+} // end namespace internal
+
+/** \class ProductReturnType
+  * \ingroup Core_Module
+  *
+  * \brief Helper class to get the correct and optimized returned type of operator*
+  *
+  * \param Lhs the type of the left-hand side
+  * \param Rhs the type of the right-hand side
+  * \param ProductMode the type of the product (determined automatically by internal::product_mode)
+  *
+  * This class defines the typename Type representing the optimized product expression
+  * between two matrix expressions. In practice, using ProductReturnType<Lhs,Rhs>::Type
+  * is the recommended way to define the result type of a function returning an expression
+  * which involve a matrix product. The class Product should never be
+  * used directly.
+  *
+  * \sa class Product, MatrixBase::operator*(const MatrixBase<OtherDerived>&)
+  */
+template<typename Lhs, typename Rhs, int ProductType>
+struct ProductReturnType
+{
+  // TODO use the nested type to reduce instanciations ????
+//   typedef typename internal::nested<Lhs,Rhs::ColsAtCompileTime>::type LhsNested;
+//   typedef typename internal::nested<Rhs,Lhs::RowsAtCompileTime>::type RhsNested;
+
+  typedef GeneralProduct<Lhs/*Nested*/, Rhs/*Nested*/, ProductType> Type;
+};
+
+template<typename Lhs, typename Rhs>
+struct ProductReturnType<Lhs,Rhs,CoeffBasedProductMode>
+{
+  typedef typename internal::nested<Lhs, Rhs::ColsAtCompileTime, typename internal::plain_matrix_type<Lhs>::type >::type LhsNested;
+  typedef typename internal::nested<Rhs, Lhs::RowsAtCompileTime, typename internal::plain_matrix_type<Rhs>::type >::type RhsNested;
+  typedef CoeffBasedProduct<LhsNested, RhsNested, EvalBeforeAssigningBit | EvalBeforeNestingBit> Type;
+};
+
+template<typename Lhs, typename Rhs>
+struct ProductReturnType<Lhs,Rhs,LazyCoeffBasedProductMode>
+{
+  typedef typename internal::nested<Lhs, Rhs::ColsAtCompileTime, typename internal::plain_matrix_type<Lhs>::type >::type LhsNested;
+  typedef typename internal::nested<Rhs, Lhs::RowsAtCompileTime, typename internal::plain_matrix_type<Rhs>::type >::type RhsNested;
+  typedef CoeffBasedProduct<LhsNested, RhsNested, NestByRefBit> Type;
+};
+
+// this is a workaround for sun CC
+template<typename Lhs, typename Rhs>
+struct LazyProductReturnType : public ProductReturnType<Lhs,Rhs,LazyCoeffBasedProductMode>
+{};
+
+/***********************************************************************
+*  Implementation of Inner Vector Vector Product
+***********************************************************************/
+
+// FIXME : maybe the "inner product" could return a Scalar
+// instead of a 1x1 matrix ??
+// Pro: more natural for the user
+// Cons: this could be a problem if in a meta unrolled algorithm a matrix-matrix
+// product ends up to a row-vector times col-vector product... To tackle this use
+// case, we could have a specialization for Block<MatrixType,1,1> with: operator=(Scalar x);
+
+namespace internal {
+
+template<typename Lhs, typename Rhs>
+struct traits<GeneralProduct<Lhs,Rhs,InnerProduct> >
+ : traits<Matrix<typename scalar_product_traits<typename Lhs::Scalar, typename Rhs::Scalar>::ReturnType,1,1> >
+{};
+
+}
+
+template<typename Lhs, typename Rhs>
+class GeneralProduct<Lhs, Rhs, InnerProduct>
+  : internal::no_assignment_operator,
+    public Matrix<typename internal::scalar_product_traits<typename Lhs::Scalar, typename Rhs::Scalar>::ReturnType,1,1>
+{
+    typedef Matrix<typename internal::scalar_product_traits<typename Lhs::Scalar, typename Rhs::Scalar>::ReturnType,1,1> Base;
+  public:
+    GeneralProduct(const Lhs& lhs, const Rhs& rhs)
+    {
+      EIGEN_STATIC_ASSERT((internal::is_same<typename Lhs::RealScalar, typename Rhs::RealScalar>::value),
+        YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY)
+
+      Base::coeffRef(0,0) = (lhs.transpose().cwiseProduct(rhs)).sum();
+    }
+
+    /** Convertion to scalar */
+    operator const typename Base::Scalar() const {
+      return Base::coeff(0,0);
+    }
+};
+
+/***********************************************************************
+*  Implementation of Outer Vector Vector Product
+***********************************************************************/
+
+namespace internal {
+template<int StorageOrder> struct outer_product_selector;
+
+template<typename Lhs, typename Rhs>
+struct traits<GeneralProduct<Lhs,Rhs,OuterProduct> >
+ : traits<ProductBase<GeneralProduct<Lhs,Rhs,OuterProduct>, Lhs, Rhs> >
+{};
+
+}
+
+template<typename Lhs, typename Rhs>
+class GeneralProduct<Lhs, Rhs, OuterProduct>
+  : public ProductBase<GeneralProduct<Lhs,Rhs,OuterProduct>, Lhs, Rhs>
+{
+  public:
+    EIGEN_PRODUCT_PUBLIC_INTERFACE(GeneralProduct)
+
+    GeneralProduct(const Lhs& lhs, const Rhs& rhs) : Base(lhs,rhs)
+    {
+      EIGEN_STATIC_ASSERT((internal::is_same<typename Lhs::RealScalar, typename Rhs::RealScalar>::value),
+        YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY)
+    }
+
+    template<typename Dest> void scaleAndAddTo(Dest& dest, Scalar alpha) const
+    {
+      internal::outer_product_selector<(int(Dest::Flags)&RowMajorBit) ? RowMajor : ColMajor>::run(*this, dest, alpha);
+    }
+};
+
+namespace internal {
+
+template<> struct outer_product_selector<ColMajor> {
+  template<typename ProductType, typename Dest>
+  static EIGEN_DONT_INLINE void run(const ProductType& prod, Dest& dest, typename ProductType::Scalar alpha) {
+    typedef typename Dest::Index Index;
+    // FIXME make sure lhs is sequentially stored
+    // FIXME not very good if rhs is real and lhs complex while alpha is real too
+    const Index cols = dest.cols();
+    for (Index j=0; j<cols; ++j)
+      dest.col(j) += (alpha * prod.rhs().coeff(j)) * prod.lhs();
+  }
+};
+
+template<> struct outer_product_selector<RowMajor> {
+  template<typename ProductType, typename Dest>
+  static EIGEN_DONT_INLINE void run(const ProductType& prod, Dest& dest, typename ProductType::Scalar alpha) {
+    typedef typename Dest::Index Index;
+    // FIXME make sure rhs is sequentially stored
+    // FIXME not very good if lhs is real and rhs complex while alpha is real too
+    const Index rows = dest.rows();
+    for (Index i=0; i<rows; ++i)
+      dest.row(i) += (alpha * prod.lhs().coeff(i)) * prod.rhs();
+  }
+};
+
+} // end namespace internal
+
+/***********************************************************************
+*  Implementation of General Matrix Vector Product
+***********************************************************************/
+
+/*  According to the shape/flags of the matrix we have to distinghish 3 different cases:
+ *   1 - the matrix is col-major, BLAS compatible and M is large => call fast BLAS-like colmajor routine
+ *   2 - the matrix is row-major, BLAS compatible and N is large => call fast BLAS-like rowmajor routine
+ *   3 - all other cases are handled using a simple loop along the outer-storage direction.
+ *  Therefore we need a lower level meta selector.
+ *  Furthermore, if the matrix is the rhs, then the product has to be transposed.
+ */
+namespace internal {
+
+template<typename Lhs, typename Rhs>
+struct traits<GeneralProduct<Lhs,Rhs,GemvProduct> >
+ : traits<ProductBase<GeneralProduct<Lhs,Rhs,GemvProduct>, Lhs, Rhs> >
+{};
+
+template<int Side, int StorageOrder, bool BlasCompatible>
+struct gemv_selector;
+
+} // end namespace internal
+
+template<typename Lhs, typename Rhs>
+class GeneralProduct<Lhs, Rhs, GemvProduct>
+  : public ProductBase<GeneralProduct<Lhs,Rhs,GemvProduct>, Lhs, Rhs>
+{
+  public:
+    EIGEN_PRODUCT_PUBLIC_INTERFACE(GeneralProduct)
+
+    typedef typename Lhs::Scalar LhsScalar;
+    typedef typename Rhs::Scalar RhsScalar;
+
+    GeneralProduct(const Lhs& lhs, const Rhs& rhs) : Base(lhs,rhs)
+    {
+//       EIGEN_STATIC_ASSERT((internal::is_same<typename Lhs::Scalar, typename Rhs::Scalar>::value),
+//         YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY)
+    }
+
+    enum { Side = Lhs::IsVectorAtCompileTime ? OnTheLeft : OnTheRight };
+    typedef typename internal::conditional<int(Side)==OnTheRight,_LhsNested,_RhsNested>::type MatrixType;
+
+    template<typename Dest> void scaleAndAddTo(Dest& dst, Scalar alpha) const
+    {
+      eigen_assert(m_lhs.rows() == dst.rows() && m_rhs.cols() == dst.cols());
+      internal::gemv_selector<Side,(int(MatrixType::Flags)&RowMajorBit) ? RowMajor : ColMajor,
+                       bool(internal::blas_traits<MatrixType>::HasUsableDirectAccess)>::run(*this, dst, alpha);
+    }
+};
+
+namespace internal {
+
+// The vector is on the left => transposition
+template<int StorageOrder, bool BlasCompatible>
+struct gemv_selector<OnTheLeft,StorageOrder,BlasCompatible>
+{
+  template<typename ProductType, typename Dest>
+  static void run(const ProductType& prod, Dest& dest, typename ProductType::Scalar alpha)
+  {
+    Transpose<Dest> destT(dest);
+    enum { OtherStorageOrder = StorageOrder == RowMajor ? ColMajor : RowMajor };
+    gemv_selector<OnTheRight,OtherStorageOrder,BlasCompatible>
+      ::run(GeneralProduct<Transpose<const typename ProductType::_RhsNested>,Transpose<const typename ProductType::_LhsNested>, GemvProduct>
+        (prod.rhs().transpose(), prod.lhs().transpose()), destT, alpha);
+  }
+};
+
+template<typename Scalar,int Size,int MaxSize,bool Cond> struct gemv_static_vector_if;
+
+template<typename Scalar,int Size,int MaxSize>
+struct gemv_static_vector_if<Scalar,Size,MaxSize,false>
+{
+  EIGEN_STRONG_INLINE  Scalar* data() { eigen_internal_assert(false && "should never be called"); return 0; }
+};
+
+template<typename Scalar,int Size>
+struct gemv_static_vector_if<Scalar,Size,Dynamic,true>
+{
+  EIGEN_STRONG_INLINE Scalar* data() { return 0; }
+};
+
+template<typename Scalar,int Size,int MaxSize>
+struct gemv_static_vector_if<Scalar,Size,MaxSize,true>
+{
+  #if EIGEN_ALIGN_STATICALLY
+  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;
+  EIGEN_STRONG_INLINE Scalar* data() {
+    return ForceAlignment
+            ? reinterpret_cast<Scalar*>((reinterpret_cast<size_t>(m_data.array) & ~(size_t(15))) + 16)
+            : m_data.array;
+  }
+  #endif
+};
+
+template<> struct gemv_selector<OnTheRight,ColMajor,true>
+{
+  template<typename ProductType, typename Dest>
+  static inline void run(const ProductType& prod, Dest& dest, typename ProductType::Scalar alpha)
+  {
+    typedef typename ProductType::Index Index;
+    typedef typename ProductType::LhsScalar   LhsScalar;
+    typedef typename ProductType::RhsScalar   RhsScalar;
+    typedef typename ProductType::Scalar      ResScalar;
+    typedef typename ProductType::RealScalar  RealScalar;
+    typedef typename ProductType::ActualLhsType ActualLhsType;
+    typedef typename ProductType::ActualRhsType ActualRhsType;
+    typedef typename ProductType::LhsBlasTraits LhsBlasTraits;
+    typedef typename ProductType::RhsBlasTraits RhsBlasTraits;
+    typedef Map<Matrix<ResScalar,Dynamic,1>, Aligned> MappedDest;
+
+    ActualLhsType actualLhs = LhsBlasTraits::extract(prod.lhs());
+    ActualRhsType actualRhs = RhsBlasTraits::extract(prod.rhs());
+
+    ResScalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(prod.lhs())
+                                  * RhsBlasTraits::extractScalarFactor(prod.rhs());
+
+    enum {
+      // FIXME find a way to allow an inner stride on the result if packet_traits<Scalar>::size==1
+      // on, the other hand it is good for the cache to pack the vector anyways...
+      EvalToDestAtCompileTime = Dest::InnerStrideAtCompileTime==1,
+      ComplexByReal = (NumTraits<LhsScalar>::IsComplex) && (!NumTraits<RhsScalar>::IsComplex),
+      MightCannotUseDest = (Dest::InnerStrideAtCompileTime!=1) || ComplexByReal
+    };
+
+    gemv_static_vector_if<ResScalar,Dest::SizeAtCompileTime,Dest::MaxSizeAtCompileTime,MightCannotUseDest> static_dest;
+
+    bool alphaIsCompatible = (!ComplexByReal) || (imag(actualAlpha)==RealScalar(0));
+    bool evalToDest = EvalToDestAtCompileTime && alphaIsCompatible;
+    
+    RhsScalar compatibleAlpha = get_factor<ResScalar,RhsScalar>::run(actualAlpha);
+
+    ei_declare_aligned_stack_constructed_variable(ResScalar,actualDestPtr,dest.size(),
+                                                  evalToDest ? dest.data() : static_dest.data());
+    
+    if(!evalToDest)
+    {
+      #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN
+      int size = dest.size();
+      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,ColMajor,LhsBlasTraits::NeedToConjugate,RhsScalar,RhsBlasTraits::NeedToConjugate>::run(
+        actualLhs.rows(), actualLhs.cols(),
+        actualLhs.data(), actualLhs.outerStride(),
+        actualRhs.data(), actualRhs.innerStride(),
+        actualDestPtr, 1,
+        compatibleAlpha);
+
+    if (!evalToDest)
+    {
+      if(!alphaIsCompatible)
+        dest += actualAlpha * MappedDest(actualDestPtr, dest.size());
+      else
+        dest = MappedDest(actualDestPtr, dest.size());
+    }
+  }
+};
+
+template<> struct gemv_selector<OnTheRight,RowMajor,true>
+{
+  template<typename ProductType, typename Dest>
+  static void run(const ProductType& prod, Dest& dest, typename ProductType::Scalar alpha)
+  {
+    typedef typename ProductType::LhsScalar LhsScalar;
+    typedef typename ProductType::RhsScalar RhsScalar;
+    typedef typename ProductType::Scalar    ResScalar;
+    typedef typename ProductType::Index Index;
+    typedef typename ProductType::ActualLhsType ActualLhsType;
+    typedef typename ProductType::ActualRhsType ActualRhsType;
+    typedef typename ProductType::_ActualRhsType _ActualRhsType;
+    typedef typename ProductType::LhsBlasTraits LhsBlasTraits;
+    typedef typename ProductType::RhsBlasTraits RhsBlasTraits;
+
+    typename add_const<ActualLhsType>::type actualLhs = LhsBlasTraits::extract(prod.lhs());
+    typename add_const<ActualRhsType>::type actualRhs = RhsBlasTraits::extract(prod.rhs());
+
+    ResScalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(prod.lhs())
+                                  * RhsBlasTraits::extractScalarFactor(prod.rhs());
+
+    enum {
+      // FIXME find a way to allow an inner stride on the result if packet_traits<Scalar>::size==1
+      // on, the other hand it is good for the cache to pack the vector anyways...
+      DirectlyUseRhs = _ActualRhsType::InnerStrideAtCompileTime==1
+    };
+
+    gemv_static_vector_if<RhsScalar,_ActualRhsType::SizeAtCompileTime,_ActualRhsType::MaxSizeAtCompileTime,!DirectlyUseRhs> static_rhs;
+
+    ei_declare_aligned_stack_constructed_variable(RhsScalar,actualRhsPtr,actualRhs.size(),
+        DirectlyUseRhs ? const_cast<RhsScalar*>(actualRhs.data()) : static_rhs.data());
+
+    if(!DirectlyUseRhs)
+    {
+      #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN
+      int size = actualRhs.size();
+      EIGEN_DENSE_STORAGE_CTOR_PLUGIN
+      #endif
+      Map<typename _ActualRhsType::PlainObject>(actualRhsPtr, actualRhs.size()) = actualRhs;
+    }
+
+    general_matrix_vector_product
+      <Index,LhsScalar,RowMajor,LhsBlasTraits::NeedToConjugate,RhsScalar,RhsBlasTraits::NeedToConjugate>::run(
+        actualLhs.rows(), actualLhs.cols(),
+        actualLhs.data(), actualLhs.outerStride(),
+        actualRhsPtr, 1,
+        dest.data(), dest.innerStride(),
+        actualAlpha);
+  }
+};
+
+template<> struct gemv_selector<OnTheRight,ColMajor,false>
+{
+  template<typename ProductType, typename Dest>
+  static void run(const ProductType& prod, Dest& dest, typename ProductType::Scalar alpha)
+  {
+    typedef typename Dest::Index Index;
+    // TODO makes sure dest is sequentially stored in memory, otherwise use a temp
+    const Index size = prod.rhs().rows();
+    for(Index k=0; k<size; ++k)
+      dest += (alpha*prod.rhs().coeff(k)) * prod.lhs().col(k);
+  }
+};
+
+template<> struct gemv_selector<OnTheRight,RowMajor,false>
+{
+  template<typename ProductType, typename Dest>
+  static void run(const ProductType& prod, Dest& dest, typename ProductType::Scalar alpha)
+  {
+    typedef typename Dest::Index Index;
+    // TODO makes sure rhs is sequentially stored in memory, otherwise use a temp
+    const Index rows = prod.rows();
+    for(Index i=0; i<rows; ++i)
+      dest.coeffRef(i) += alpha * (prod.lhs().row(i).cwiseProduct(prod.rhs().transpose())).sum();
+  }
+};
+
+} // end namespace internal
+
+/***************************************************************************
+* Implementation of matrix base methods
+***************************************************************************/
+
+/** \returns the matrix product of \c *this and \a other.
+  *
+  * \note If instead of the matrix product you want the coefficient-wise product, see Cwise::operator*().
+  *
+  * \sa lazyProduct(), operator*=(const MatrixBase&), Cwise::operator*()
+  */
+template<typename Derived>
+template<typename OtherDerived>
+inline const typename ProductReturnType<Derived, OtherDerived>::Type
+MatrixBase<Derived>::operator*(const MatrixBase<OtherDerived> &other) const
+{
+  // A note regarding the function declaration: In MSVC, this function will sometimes
+  // not be inlined since DenseStorage is an unwindable object for dynamic
+  // matrices and product types are holding a member to store the result.
+  // Thus it does not help tagging this function with EIGEN_STRONG_INLINE.
+  enum {
+    ProductIsValid =  Derived::ColsAtCompileTime==Dynamic
+                   || OtherDerived::RowsAtCompileTime==Dynamic
+                   || int(Derived::ColsAtCompileTime)==int(OtherDerived::RowsAtCompileTime),
+    AreVectors = Derived::IsVectorAtCompileTime && OtherDerived::IsVectorAtCompileTime,
+    SameSizes = EIGEN_PREDICATE_SAME_MATRIX_SIZE(Derived,OtherDerived)
+  };
+  // note to the lost user:
+  //    * for a dot product use: v1.dot(v2)
+  //    * for a coeff-wise product use: v1.cwiseProduct(v2)
+  EIGEN_STATIC_ASSERT(ProductIsValid || !(AreVectors && SameSizes),
+    INVALID_VECTOR_VECTOR_PRODUCT__IF_YOU_WANTED_A_DOT_OR_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTIONS)
+  EIGEN_STATIC_ASSERT(ProductIsValid || !(SameSizes && !AreVectors),
+    INVALID_MATRIX_PRODUCT__IF_YOU_WANTED_A_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTION)
+  EIGEN_STATIC_ASSERT(ProductIsValid || SameSizes, INVALID_MATRIX_PRODUCT)
+#ifdef EIGEN_DEBUG_PRODUCT
+  internal::product_type<Derived,OtherDerived>::debug();
+#endif
+  return typename ProductReturnType<Derived,OtherDerived>::Type(derived(), other.derived());
+}
+
+/** \returns an expression of the matrix product of \c *this and \a other without implicit evaluation.
+  *
+  * The returned product will behave like any other expressions: the coefficients of the product will be
+  * computed once at a time as requested. This might be useful in some extremely rare cases when only
+  * a small and no coherent fraction of the result's coefficients have to be computed.
+  *
+  * \warning This version of the matrix product can be much much slower. So use it only if you know
+  * what you are doing and that you measured a true speed improvement.
+  *
+  * \sa operator*(const MatrixBase&)
+  */
+template<typename Derived>
+template<typename OtherDerived>
+const typename LazyProductReturnType<Derived,OtherDerived>::Type
+MatrixBase<Derived>::lazyProduct(const MatrixBase<OtherDerived> &other) const
+{
+  enum {
+    ProductIsValid =  Derived::ColsAtCompileTime==Dynamic
+                   || OtherDerived::RowsAtCompileTime==Dynamic
+                   || int(Derived::ColsAtCompileTime)==int(OtherDerived::RowsAtCompileTime),
+    AreVectors = Derived::IsVectorAtCompileTime && OtherDerived::IsVectorAtCompileTime,
+    SameSizes = EIGEN_PREDICATE_SAME_MATRIX_SIZE(Derived,OtherDerived)
+  };
+  // note to the lost user:
+  //    * for a dot product use: v1.dot(v2)
+  //    * for a coeff-wise product use: v1.cwiseProduct(v2)
+  EIGEN_STATIC_ASSERT(ProductIsValid || !(AreVectors && SameSizes),
+    INVALID_VECTOR_VECTOR_PRODUCT__IF_YOU_WANTED_A_DOT_OR_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTIONS)
+  EIGEN_STATIC_ASSERT(ProductIsValid || !(SameSizes && !AreVectors),
+    INVALID_MATRIX_PRODUCT__IF_YOU_WANTED_A_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTION)
+  EIGEN_STATIC_ASSERT(ProductIsValid || SameSizes, INVALID_MATRIX_PRODUCT)
+
+  return typename LazyProductReturnType<Derived,OtherDerived>::Type(derived(), other.derived());
+}
+
+#endif // EIGEN_PRODUCT_H

Eigen/src/Core/GenericPacketMath.h

@@ -312,7 +312,7 @@ template<int Offset,typename PacketType>
 struct palign_impl
 {
   // by default data are aligned, so there is nothing to be done :)
-  inline static void run(PacketType&, const PacketType&) {}
+  static inline void run(PacketType&, const PacketType&) {}
 };
 
 /** \internal update \a first using the concatenation of the \a Offset last elements

Eigen/src/Core/IO.h

@@ -171,7 +171,7 @@ std::ostream & print_matrix(std::ostream & s, const Derived& _m, const IOFormat&
     return s;
   }
   
-  const typename Derived::Nested m = _m;
+  typename Derived::Nested m = _m;
   typedef typename Derived::Scalar Scalar;
   typedef typename Derived::Index Index;
 

Eigen/src/Core/MathFunctions.h

@@ -309,8 +309,7 @@ struct abs2_impl<std::complex<RealScalar> >
 {
   static inline RealScalar run(const std::complex<RealScalar>& x)
   {
-    using std::norm;
-    return norm(x);
+    return real(x)*real(x) + imag(x)*imag(x);
   }
 };
 
@@ -553,7 +552,7 @@ struct pow_default_impl<Scalar, true>
 {
   static inline Scalar run(Scalar x, Scalar y)
   {
-    Scalar res = 1;
+    Scalar res(1);
     eigen_assert(!NumTraits<Scalar>::IsSigned || y >= 0);
     if(y & 1) res *= x;
     y >>= 1;
@@ -838,6 +837,17 @@ template<> struct scalar_fuzzy_impl<bool>
   
 };
 
+/****************************************************************************
+* Special functions                                                          *
+****************************************************************************/
+
+// std::isfinite is non standard, so let's define our own version,
+// even though it is not very efficient.
+template<typename T> bool isfinite(const T& x)
+{
+  return x<NumTraits<T>::highest() && x>NumTraits<T>::lowest();
+}
+
 } // end namespace internal
 
 #endif // EIGEN_MATHFUNCTIONS_H

Eigen/src/Core/Matrix.h

@@ -411,25 +411,6 @@ EIGEN_MAKE_TYPEDEFS_ALL_SIZES(std::complex<double>, cd)
 
 #undef EIGEN_MAKE_TYPEDEFS_ALL_SIZES
 #undef EIGEN_MAKE_TYPEDEFS
-
-#undef EIGEN_MAKE_TYPEDEFS_LARGE
-
-#define EIGEN_USING_MATRIX_TYPEDEFS_FOR_TYPE_AND_SIZE(TypeSuffix, SizeSuffix) \
-using Eigen::Matrix##SizeSuffix##TypeSuffix; \
-using Eigen::Vector##SizeSuffix##TypeSuffix; \
-using Eigen::RowVector##SizeSuffix##TypeSuffix;
-
-#define EIGEN_USING_MATRIX_TYPEDEFS_FOR_TYPE(TypeSuffix) \
-EIGEN_USING_MATRIX_TYPEDEFS_FOR_TYPE_AND_SIZE(TypeSuffix, 2) \
-EIGEN_USING_MATRIX_TYPEDEFS_FOR_TYPE_AND_SIZE(TypeSuffix, 3) \
-EIGEN_USING_MATRIX_TYPEDEFS_FOR_TYPE_AND_SIZE(TypeSuffix, 4) \
-EIGEN_USING_MATRIX_TYPEDEFS_FOR_TYPE_AND_SIZE(TypeSuffix, X) \
-
-#define EIGEN_USING_MATRIX_TYPEDEFS \
-EIGEN_USING_MATRIX_TYPEDEFS_FOR_TYPE(i) \
-EIGEN_USING_MATRIX_TYPEDEFS_FOR_TYPE(f) \
-EIGEN_USING_MATRIX_TYPEDEFS_FOR_TYPE(d) \
-EIGEN_USING_MATRIX_TYPEDEFS_FOR_TYPE(cf) \
-EIGEN_USING_MATRIX_TYPEDEFS_FOR_TYPE(cd)
+#undef EIGEN_MAKE_FIXED_TYPEDEFS
 
 #endif // EIGEN_MATRIX_H

Eigen/src/Core/MatrixBase.h

@@ -250,8 +250,7 @@ template<typename Derived> class MatrixBase
     
     // huuuge hack. make Eigen2's matrix.part<Diagonal>() work in eigen3. Problem: Diagonal is now a class template instead
     // of an integer constant. Solution: overload the part() method template wrt template parameters list.
-    // Note: replacing next line by "template<template<typename T, int n> class U>" produces a mysterious error C2082 in MSVC.
-    template<template<typename, int> class U>
+    template<template<typename T, int n> class U>
     const DiagonalWrapper<ConstDiagonalReturnType> part() const
     { return diagonal().asDiagonal(); }
     #endif // EIGEN2_SUPPORT
@@ -331,7 +330,7 @@ template<typename Derived> class MatrixBase
     /** \returns an \link ArrayBase Array \endlink expression of this matrix
       * \sa ArrayBase::matrix() */
     ArrayWrapper<Derived> array() { return derived(); }
-    const ArrayWrapper<Derived> array() const { return derived(); }
+    const ArrayWrapper<const Derived> array() const { return derived(); }
 
 /////////// LU module ///////////
 
@@ -466,6 +465,8 @@ template<typename Derived> class MatrixBase
     const MatrixFunctionReturnValue<Derived> sinh() const;
     const MatrixFunctionReturnValue<Derived> cos() const;
     const MatrixFunctionReturnValue<Derived> sin() const;
+    const MatrixSquareRootReturnValue<Derived> sqrt() const;
+    const MatrixLogarithmReturnValue<Derived> log() const;
 
 #ifdef EIGEN2_SUPPORT
     template<typename ProductDerived, typename Lhs, typename Rhs>
@@ -512,10 +513,10 @@ template<typename Derived> class MatrixBase
   protected:
     // mixing arrays and matrices is not legal
     template<typename OtherDerived> Derived& operator+=(const ArrayBase<OtherDerived>& )
-    {EIGEN_STATIC_ASSERT(sizeof(typename OtherDerived::Scalar)==-1,YOU_CANNOT_MIX_ARRAYS_AND_MATRICES);}
+    {EIGEN_STATIC_ASSERT(std::ptrdiff_t(sizeof(typename OtherDerived::Scalar))==-1,YOU_CANNOT_MIX_ARRAYS_AND_MATRICES); return *this;}
     // mixing arrays and matrices is not legal
     template<typename OtherDerived> Derived& operator-=(const ArrayBase<OtherDerived>& )
-    {EIGEN_STATIC_ASSERT(sizeof(typename OtherDerived::Scalar)==-1,YOU_CANNOT_MIX_ARRAYS_AND_MATRICES);}
+    {EIGEN_STATIC_ASSERT(std::ptrdiff_t(sizeof(typename OtherDerived::Scalar))==-1,YOU_CANNOT_MIX_ARRAYS_AND_MATRICES); return *this;}
 };
 
 #endif // EIGEN_MATRIXBASE_H

Eigen/src/Core/NumTraits.h

@@ -81,14 +81,14 @@ template<typename T> struct GenericNumTraits
                    >::type NonInteger;
   typedef T Nested;
 
-  inline static Real epsilon() { return std::numeric_limits<T>::epsilon(); }
-  inline static Real dummy_precision()
+  static inline Real epsilon() { return std::numeric_limits<T>::epsilon(); }
+  static inline Real dummy_precision()
   {
     // make sure to override this for floating-point types
     return Real(0);
   }
-  inline static T highest() { return (std::numeric_limits<T>::max)(); }
-  inline static T lowest()  { return IsInteger ? (std::numeric_limits<T>::min)() : (-(std::numeric_limits<T>::max)()); }
+  static inline T highest() { return (std::numeric_limits<T>::max)(); }
+  static inline T lowest()  { return IsInteger ? (std::numeric_limits<T>::min)() : (-(std::numeric_limits<T>::max)()); }
   
 #ifdef EIGEN2_SUPPORT
   enum {
@@ -104,12 +104,12 @@ template<typename T> struct NumTraits : GenericNumTraits<T>
 template<> struct NumTraits<float>
   : GenericNumTraits<float>
 {
-  inline static float dummy_precision() { return 1e-5f; }
+  static inline float dummy_precision() { return 1e-5f; }
 };
 
 template<> struct NumTraits<double> : GenericNumTraits<double>
 {
-  inline static double dummy_precision() { return 1e-12; }
+  static inline double dummy_precision() { return 1e-12; }
 };
 
 template<> struct NumTraits<long double>
@@ -130,8 +130,8 @@ template<typename _Real> struct NumTraits<std::complex<_Real> >
     MulCost = 4 * NumTraits<Real>::MulCost + 2 * NumTraits<Real>::AddCost
   };
 
-  inline static Real epsilon() { return NumTraits<Real>::epsilon(); }
-  inline static Real dummy_precision() { return NumTraits<Real>::dummy_precision(); }
+  static inline Real epsilon() { return NumTraits<Real>::epsilon(); }
+  static inline Real dummy_precision() { return NumTraits<Real>::dummy_precision(); }
 };
 
 template<typename Scalar, int Rows, int Cols, int Options, int MaxRows, int MaxCols>

Eigen/src/Core/PermutationMatrix.h

@@ -511,7 +511,7 @@ class PermutationWrapper : public PermutationBase<PermutationWrapper<_IndicesTyp
 
   protected:
 
-    const typename IndicesType::Nested m_indices;
+    typename IndicesType::Nested m_indices;
 };
 
 /** \returns the matrix with the permutation applied to the columns.
@@ -608,7 +608,7 @@ struct permut_matrix_product_retval
 
   protected:
     const PermutationType& m_permutation;
-    const typename MatrixType::Nested m_matrix;
+    typename MatrixType::Nested m_matrix;
 };
 
 /* Template partial specialization for transposed/inverse permutations */

Eigen/src/Core/PlainObjectBase.h

@@ -53,7 +53,7 @@ template<typename MatrixTypeA, typename MatrixTypeB, bool SwapPointers> struct m
 
 } // end namespace internal
 
-/**
+/** \class PlainObjectBase
   * \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
@@ -61,8 +61,29 @@ template<typename MatrixTypeA, typename MatrixTypeB, bool SwapPointers> struct m
   *
   * \sa \ref TopicClassHierarchy
   */
+#ifdef EIGEN_PARSED_BY_DOXYGEN
+namespace internal {
+
+// this is a warkaround to doxygen not being able to understand the inheritence logic
+// when it is hidden by the dense_xpr_base helper struct.
+template<typename Derived> struct dense_xpr_base_dispatcher_for_doxygen;// : public MatrixBase<Derived> {};
+/** This class is just a workaround for Doxygen and it does not not actually exist. */
+template<typename _Scalar, int _Rows, int _Cols, int _Options, int _MaxRows, int _MaxCols>
+struct dense_xpr_base_dispatcher_for_doxygen<Matrix<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> >
+    : public MatrixBase<Matrix<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> > {};
+/** 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_for_doxygen<Array<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> >
+    : public ArrayBase<Array<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> > {};
+
+} // namespace internal
+
+template<typename Derived>
+class PlainObjectBase : public internal::dense_xpr_base_dispatcher_for_doxygen<Derived>
+#else
 template<typename Derived>
 class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
+#endif
 {
   public:
     enum { Options = internal::traits<Derived>::Options };
@@ -443,68 +464,68 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       * \see class Map
       */
     //@{
-    inline static ConstMapType Map(const Scalar* data)
+    static inline ConstMapType Map(const Scalar* data)
     { return ConstMapType(data); }
-    inline static MapType Map(Scalar* data)
+    static inline MapType Map(Scalar* data)
     { return MapType(data); }
-    inline static ConstMapType Map(const Scalar* data, Index size)
+    static inline ConstMapType Map(const Scalar* data, Index size)
     { return ConstMapType(data, size); }
-    inline static MapType Map(Scalar* data, Index size)
+    static inline MapType Map(Scalar* data, Index size)
     { return MapType(data, size); }
-    inline static ConstMapType Map(const Scalar* data, Index rows, Index cols)
+    static inline ConstMapType Map(const Scalar* data, Index rows, Index cols)
     { return ConstMapType(data, rows, cols); }
-    inline static MapType Map(Scalar* data, Index rows, Index cols)
+    static inline MapType Map(Scalar* data, Index rows, Index cols)
     { return MapType(data, rows, cols); }
 
-    inline static ConstAlignedMapType MapAligned(const Scalar* data)
+    static inline ConstAlignedMapType MapAligned(const Scalar* data)
     { return ConstAlignedMapType(data); }
-    inline static AlignedMapType MapAligned(Scalar* data)
+    static inline AlignedMapType MapAligned(Scalar* data)
     { return AlignedMapType(data); }
-    inline static ConstAlignedMapType MapAligned(const Scalar* data, Index size)
+    static inline ConstAlignedMapType MapAligned(const Scalar* data, Index size)
     { return ConstAlignedMapType(data, size); }
-    inline static AlignedMapType MapAligned(Scalar* data, Index size)
+    static inline AlignedMapType MapAligned(Scalar* data, Index size)
     { return AlignedMapType(data, size); }
-    inline static ConstAlignedMapType MapAligned(const Scalar* data, Index rows, Index cols)
+    static inline ConstAlignedMapType MapAligned(const Scalar* data, Index rows, Index cols)
     { return ConstAlignedMapType(data, rows, cols); }
-    inline static AlignedMapType MapAligned(Scalar* data, Index rows, Index cols)
+    static inline AlignedMapType MapAligned(Scalar* data, Index rows, Index cols)
     { return AlignedMapType(data, rows, cols); }
 
     template<int Outer, int Inner>
-    inline static typename StridedConstMapType<Stride<Outer, Inner> >::type Map(const Scalar* data, const Stride<Outer, Inner>& stride)
+    static inline typename StridedConstMapType<Stride<Outer, Inner> >::type Map(const Scalar* data, const Stride<Outer, Inner>& stride)
     { return typename StridedConstMapType<Stride<Outer, Inner> >::type(data, stride); }
     template<int Outer, int Inner>
-    inline static typename StridedMapType<Stride<Outer, Inner> >::type Map(Scalar* data, const Stride<Outer, Inner>& stride)
+    static inline typename StridedMapType<Stride<Outer, Inner> >::type Map(Scalar* data, const Stride<Outer, Inner>& stride)
     { return typename StridedMapType<Stride<Outer, Inner> >::type(data, stride); }
     template<int Outer, int Inner>
-    inline static typename StridedConstMapType<Stride<Outer, Inner> >::type Map(const Scalar* data, Index size, const Stride<Outer, Inner>& stride)
+    static inline typename StridedConstMapType<Stride<Outer, Inner> >::type Map(const Scalar* data, Index size, const Stride<Outer, Inner>& stride)
     { return typename StridedConstMapType<Stride<Outer, Inner> >::type(data, size, stride); }
     template<int Outer, int Inner>
-    inline static typename StridedMapType<Stride<Outer, Inner> >::type Map(Scalar* data, Index size, const Stride<Outer, Inner>& stride)
+    static inline typename StridedMapType<Stride<Outer, Inner> >::type Map(Scalar* data, Index size, const Stride<Outer, Inner>& stride)
     { return typename StridedMapType<Stride<Outer, Inner> >::type(data, size, stride); }
     template<int Outer, int Inner>
-    inline static typename StridedConstMapType<Stride<Outer, Inner> >::type Map(const Scalar* data, Index rows, Index cols, const Stride<Outer, Inner>& stride)
+    static inline typename StridedConstMapType<Stride<Outer, Inner> >::type Map(const Scalar* data, Index rows, Index cols, const Stride<Outer, Inner>& stride)
     { return typename StridedConstMapType<Stride<Outer, Inner> >::type(data, rows, cols, stride); }
     template<int Outer, int Inner>
-    inline static typename StridedMapType<Stride<Outer, Inner> >::type Map(Scalar* data, Index rows, Index cols, const Stride<Outer, Inner>& stride)
+    static inline typename StridedMapType<Stride<Outer, Inner> >::type Map(Scalar* data, Index rows, Index cols, const Stride<Outer, Inner>& stride)
     { return typename StridedMapType<Stride<Outer, Inner> >::type(data, rows, cols, stride); }
 
     template<int Outer, int Inner>
-    inline static typename StridedConstAlignedMapType<Stride<Outer, Inner> >::type MapAligned(const Scalar* data, const Stride<Outer, Inner>& stride)
+    static inline typename StridedConstAlignedMapType<Stride<Outer, Inner> >::type MapAligned(const Scalar* data, const Stride<Outer, Inner>& stride)
     { return typename StridedConstAlignedMapType<Stride<Outer, Inner> >::type(data, stride); }
     template<int Outer, int Inner>
-    inline static typename StridedAlignedMapType<Stride<Outer, Inner> >::type MapAligned(Scalar* data, const Stride<Outer, Inner>& stride)
+    static inline typename StridedAlignedMapType<Stride<Outer, Inner> >::type MapAligned(Scalar* data, const Stride<Outer, Inner>& stride)
     { return typename StridedAlignedMapType<Stride<Outer, Inner> >::type(data, stride); }
     template<int Outer, int Inner>
-    inline static typename StridedConstAlignedMapType<Stride<Outer, Inner> >::type MapAligned(const Scalar* data, Index size, const Stride<Outer, Inner>& stride)
+    static inline typename StridedConstAlignedMapType<Stride<Outer, Inner> >::type MapAligned(const Scalar* data, Index size, const Stride<Outer, Inner>& stride)
     { return typename StridedConstAlignedMapType<Stride<Outer, Inner> >::type(data, size, stride); }
     template<int Outer, int Inner>
-    inline static typename StridedAlignedMapType<Stride<Outer, Inner> >::type MapAligned(Scalar* data, Index size, const Stride<Outer, Inner>& stride)
+    static inline typename StridedAlignedMapType<Stride<Outer, Inner> >::type MapAligned(Scalar* data, Index size, const Stride<Outer, Inner>& stride)
     { return typename StridedAlignedMapType<Stride<Outer, Inner> >::type(data, size, stride); }
     template<int Outer, int Inner>
-    inline static typename StridedConstAlignedMapType<Stride<Outer, Inner> >::type MapAligned(const Scalar* data, Index rows, Index cols, const Stride<Outer, Inner>& stride)
+    static inline typename StridedConstAlignedMapType<Stride<Outer, Inner> >::type MapAligned(const Scalar* data, Index rows, Index cols, const Stride<Outer, Inner>& stride)
     { return typename StridedConstAlignedMapType<Stride<Outer, Inner> >::type(data, rows, cols, stride); }
     template<int Outer, int Inner>
-    inline static typename StridedAlignedMapType<Stride<Outer, Inner> >::type MapAligned(Scalar* data, Index rows, Index cols, const Stride<Outer, Inner>& stride)
+    static inline typename StridedAlignedMapType<Stride<Outer, Inner> >::type MapAligned(Scalar* data, Index rows, Index cols, const Stride<Outer, Inner>& stride)
     { return typename StridedAlignedMapType<Stride<Outer, Inner> >::type(data, rows, cols, stride); }
     //@}
 
@@ -594,6 +615,9 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
     template<typename T0, typename T1>
     EIGEN_STRONG_INLINE void _init2(Index rows, Index cols, typename internal::enable_if<Base::SizeAtCompileTime!=2,T0>::type* = 0)
     {
+      EIGEN_STATIC_ASSERT(bool(NumTraits<T0>::IsInteger) &&
+                          bool(NumTraits<T1>::IsInteger),
+                          FLOATING_POINT_ARGUMENT_PASSED__INTEGER_WAS_EXPECTED)
       eigen_assert(rows >= 0 && (RowsAtCompileTime == Dynamic || RowsAtCompileTime == rows)
              && cols >= 0 && (ColsAtCompileTime == Dynamic || ColsAtCompileTime == cols));
       internal::check_rows_cols_for_overflow(rows, cols);      
@@ -623,7 +647,7 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
 
   public:
 #ifndef EIGEN_PARSED_BY_DOXYGEN
-    EIGEN_STRONG_INLINE static void _check_template_params()
+    static EIGEN_STRONG_INLINE void _check_template_params()
     {
       EIGEN_STATIC_ASSERT((EIGEN_IMPLIES(MaxRowsAtCompileTime==1 && MaxColsAtCompileTime!=1, (Options&RowMajor)==RowMajor)
                         && EIGEN_IMPLIES(MaxColsAtCompileTime==1 && MaxRowsAtCompileTime!=1, (Options&RowMajor)==0)

Eigen/src/Core/Product.h

@@ -1,8 +1,7 @@
 // This file is part of Eigen, a lightweight C++ template library
 // for linear algebra.
 //
-// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
-// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2008-2011 Gael Guennebaud <gael.guennebaud@inria.fr>
 //
 // Eigen is free software; you can redistribute it and/or
 // modify it under the terms of the GNU Lesser General Public
@@ -26,600 +25,89 @@
 #ifndef EIGEN_PRODUCT_H
 #define EIGEN_PRODUCT_H
 
-/** \class GeneralProduct
+template<typename Lhs, typename Rhs> class Product;
+template<typename Lhs, typename Rhs, typename StorageKind> class ProductImpl;
+
+/** \class Product
   * \ingroup Core_Module
   *
-  * \brief Expression of the product of two general matrices or vectors
+  * \brief Expression of the product of two arbitrary matrices or vectors
   *
-  * \param LhsNested the type used to store the left-hand side
-  * \param RhsNested the type used to store the right-hand side
-  * \param ProductMode the type of the product
+  * \param Lhs the type of the left-hand side expression
+  * \param Rhs the type of the right-hand side expression
   *
-  * This class represents an expression of the product of two general matrices.
-  * We call a general matrix, a dense matrix with full storage. For instance,
-  * This excludes triangular, selfadjoint, and sparse matrices.
-  * It is the return type of the operator* between general matrices. Its template
-  * arguments are determined automatically by ProductReturnType. Therefore,
-  * GeneralProduct should never be used direclty. To determine the result type of a
-  * function which involves a matrix product, use ProductReturnType::Type.
+  * This class represents an expression of the product of two arbitrary matrices.
   *
-  * \sa ProductReturnType, MatrixBase::operator*(const MatrixBase<OtherDerived>&)
   */
-template<typename Lhs, typename Rhs, int ProductType = internal::product_type<Lhs,Rhs>::value>
-class GeneralProduct;
-
-enum {
-  Large = 2,
-  Small = 3
-};
 
 namespace internal {
-
-template<int Rows, int Cols, int Depth> struct product_type_selector;
-
-template<int Size, int MaxSize> struct product_size_category
+template<typename Lhs, typename Rhs>
+struct traits<Product<Lhs, Rhs> >
 {
-  enum { is_large = MaxSize == Dynamic ||
-                    Size >= EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD,
-         value = is_large  ? Large
-               : Size == 1 ? 1
-                           : Small
+  typedef MatrixXpr XprKind;
+  typedef typename remove_all<Lhs>::type LhsCleaned;
+  typedef typename remove_all<Rhs>::type RhsCleaned;
+  typedef typename scalar_product_traits<typename traits<LhsCleaned>::Scalar, typename traits<RhsCleaned>::Scalar>::ReturnType Scalar;
+  typedef typename promote_storage_type<typename traits<LhsCleaned>::StorageKind,
+                                        typename traits<RhsCleaned>::StorageKind>::ret StorageKind;
+  typedef typename promote_index_type<typename traits<LhsCleaned>::Index,
+                                      typename traits<RhsCleaned>::Index>::type Index;
+  enum {
+    RowsAtCompileTime = LhsCleaned::RowsAtCompileTime,
+    ColsAtCompileTime = RhsCleaned::ColsAtCompileTime,
+    MaxRowsAtCompileTime = LhsCleaned::MaxRowsAtCompileTime,
+    MaxColsAtCompileTime = RhsCleaned::MaxColsAtCompileTime,
+    Flags = (MaxRowsAtCompileTime==1 ? RowMajorBit : 0), // TODO should be no storage order
+    CoeffReadCost = 0 // TODO CoeffReadCost should not be part of the expression traits
   };
 };
-
-template<typename Lhs, typename Rhs> struct product_type
-{
-  typedef typename remove_all<Lhs>::type _Lhs;
-  typedef typename remove_all<Rhs>::type _Rhs;
-  enum {
-    MaxRows  = _Lhs::MaxRowsAtCompileTime,
-    Rows  = _Lhs::RowsAtCompileTime,
-    MaxCols  = _Rhs::MaxColsAtCompileTime,
-    Cols  = _Rhs::ColsAtCompileTime,
-    MaxDepth = EIGEN_SIZE_MIN_PREFER_FIXED(_Lhs::MaxColsAtCompileTime,
-                                           _Rhs::MaxRowsAtCompileTime),
-    Depth = EIGEN_SIZE_MIN_PREFER_FIXED(_Lhs::ColsAtCompileTime,
-                                        _Rhs::RowsAtCompileTime),
-    LargeThreshold = EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD
-  };
-
-  // the splitting into different lines of code here, introducing the _select enums and the typedef below,
-  // is to work around an internal compiler error with gcc 4.1 and 4.2.
-private:
-  enum {
-    rows_select = product_size_category<Rows,MaxRows>::value,
-    cols_select = product_size_category<Cols,MaxCols>::value,
-    depth_select = product_size_category<Depth,MaxDepth>::value
-  };
-  typedef product_type_selector<rows_select, cols_select, depth_select> selector;
-
-public:
-  enum {
-    value = selector::ret
-  };
-#ifdef EIGEN_DEBUG_PRODUCT
-  static void debug()
-  {
-      EIGEN_DEBUG_VAR(Rows);
-      EIGEN_DEBUG_VAR(Cols);
-      EIGEN_DEBUG_VAR(Depth);
-      EIGEN_DEBUG_VAR(rows_select);
-      EIGEN_DEBUG_VAR(cols_select);
-      EIGEN_DEBUG_VAR(depth_select);
-      EIGEN_DEBUG_VAR(value);
-  }
-#endif
-};
-
-
-/* The following allows to select the kind of product at compile time
- * based on the three dimensions of the product.
- * This is a compile time mapping from {1,Small,Large}^3 -> {product types} */
-// FIXME I'm not sure the current mapping is the ideal one.
-template<int M, int N>  struct product_type_selector<M,N,1>              { enum { ret = OuterProduct }; };
-template<int Depth>     struct product_type_selector<1,    1,    Depth>  { enum { ret = InnerProduct }; };
-template<>              struct product_type_selector<1,    1,    1>      { enum { ret = InnerProduct }; };
-template<>              struct product_type_selector<Small,1,    Small>  { enum { ret = CoeffBasedProductMode }; };
-template<>              struct product_type_selector<1,    Small,Small>  { enum { ret = CoeffBasedProductMode }; };
-template<>              struct product_type_selector<Small,Small,Small>  { enum { ret = CoeffBasedProductMode }; };
-template<>              struct product_type_selector<Small, Small, 1>    { enum { ret = LazyCoeffBasedProductMode }; };
-template<>              struct product_type_selector<Small, Large, 1>    { enum { ret = LazyCoeffBasedProductMode }; };
-template<>              struct product_type_selector<Large, Small, 1>    { enum { ret = LazyCoeffBasedProductMode }; };
-template<>              struct product_type_selector<1,    Large,Small>  { enum { ret = CoeffBasedProductMode }; };
-template<>              struct product_type_selector<1,    Large,Large>  { enum { ret = GemvProduct }; };
-template<>              struct product_type_selector<1,    Small,Large>  { enum { ret = CoeffBasedProductMode }; };
-template<>              struct product_type_selector<Large,1,    Small>  { enum { ret = CoeffBasedProductMode }; };
-template<>              struct product_type_selector<Large,1,    Large>  { enum { ret = GemvProduct }; };
-template<>              struct product_type_selector<Small,1,    Large>  { enum { ret = CoeffBasedProductMode }; };
-template<>              struct product_type_selector<Small,Small,Large>  { enum { ret = GemmProduct }; };
-template<>              struct product_type_selector<Large,Small,Large>  { enum { ret = GemmProduct }; };
-template<>              struct product_type_selector<Small,Large,Large>  { enum { ret = GemmProduct }; };
-template<>              struct product_type_selector<Large,Large,Large>  { enum { ret = GemmProduct }; };
-template<>              struct product_type_selector<Large,Small,Small>  { enum { ret = GemmProduct }; };
-template<>              struct product_type_selector<Small,Large,Small>  { enum { ret = GemmProduct }; };
-template<>              struct product_type_selector<Large,Large,Small>  { enum { ret = GemmProduct }; };
-
 } // end namespace internal
 
-/** \class ProductReturnType
-  * \ingroup Core_Module
-  *
-  * \brief Helper class to get the correct and optimized returned type of operator*
-  *
-  * \param Lhs the type of the left-hand side
-  * \param Rhs the type of the right-hand side
-  * \param ProductMode the type of the product (determined automatically by internal::product_mode)
-  *
-  * This class defines the typename Type representing the optimized product expression
-  * between two matrix expressions. In practice, using ProductReturnType<Lhs,Rhs>::Type
-  * is the recommended way to define the result type of a function returning an expression
-  * which involve a matrix product. The class Product should never be
-  * used directly.
-  *
-  * \sa class Product, MatrixBase::operator*(const MatrixBase<OtherDerived>&)
-  */
-template<typename Lhs, typename Rhs, int ProductType>
-struct ProductReturnType
-{
-  // TODO use the nested type to reduce instanciations ????
-//   typedef typename internal::nested<Lhs,Rhs::ColsAtCompileTime>::type LhsNested;
-//   typedef typename internal::nested<Rhs,Lhs::RowsAtCompileTime>::type RhsNested;
-
-  typedef GeneralProduct<Lhs/*Nested*/, Rhs/*Nested*/, ProductType> Type;
-};
 
 template<typename Lhs, typename Rhs>
-struct ProductReturnType<Lhs,Rhs,CoeffBasedProductMode>
-{
-  typedef typename internal::nested<Lhs, Rhs::ColsAtCompileTime, typename internal::plain_matrix_type<Lhs>::type >::type LhsNested;
-  typedef typename internal::nested<Rhs, Lhs::RowsAtCompileTime, typename internal::plain_matrix_type<Rhs>::type >::type RhsNested;
-  typedef CoeffBasedProduct<LhsNested, RhsNested, EvalBeforeAssigningBit | EvalBeforeNestingBit> Type;
-};
-
-template<typename Lhs, typename Rhs>
-struct ProductReturnType<Lhs,Rhs,LazyCoeffBasedProductMode>
-{
-  typedef typename internal::nested<Lhs, Rhs::ColsAtCompileTime, typename internal::plain_matrix_type<Lhs>::type >::type LhsNested;
-  typedef typename internal::nested<Rhs, Lhs::RowsAtCompileTime, typename internal::plain_matrix_type<Rhs>::type >::type RhsNested;
-  typedef CoeffBasedProduct<LhsNested, RhsNested, NestByRefBit> Type;
-};
-
-// this is a workaround for sun CC
-template<typename Lhs, typename Rhs>
-struct LazyProductReturnType : public ProductReturnType<Lhs,Rhs,LazyCoeffBasedProductMode>
-{};
-
-/***********************************************************************
-*  Implementation of Inner Vector Vector Product
-***********************************************************************/
-
-// FIXME : maybe the "inner product" could return a Scalar
-// instead of a 1x1 matrix ??
-// Pro: more natural for the user
-// Cons: this could be a problem if in a meta unrolled algorithm a matrix-matrix
-// product ends up to a row-vector times col-vector product... To tackle this use
-// case, we could have a specialization for Block<MatrixType,1,1> with: operator=(Scalar x);
-
-namespace internal {
-
-template<typename Lhs, typename Rhs>
-struct traits<GeneralProduct<Lhs,Rhs,InnerProduct> >
- : traits<Matrix<typename scalar_product_traits<typename Lhs::Scalar, typename Rhs::Scalar>::ReturnType,1,1> >
-{};
-
-}
-
-template<typename Lhs, typename Rhs>
-class GeneralProduct<Lhs, Rhs, InnerProduct>
-  : internal::no_assignment_operator,
-    public Matrix<typename internal::scalar_product_traits<typename Lhs::Scalar, typename Rhs::Scalar>::ReturnType,1,1>
-{
-    typedef Matrix<typename internal::scalar_product_traits<typename Lhs::Scalar, typename Rhs::Scalar>::ReturnType,1,1> Base;
-  public:
-    GeneralProduct(const Lhs& lhs, const Rhs& rhs)
-    {
-      EIGEN_STATIC_ASSERT((internal::is_same<typename Lhs::RealScalar, typename Rhs::RealScalar>::value),
-        YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY)
-
-      Base::coeffRef(0,0) = (lhs.transpose().cwiseProduct(rhs)).sum();
-    }
-
-    /** Convertion to scalar */
-    operator const typename Base::Scalar() const {
-      return Base::coeff(0,0);
-    }
-};
-
-/***********************************************************************
-*  Implementation of Outer Vector Vector Product
-***********************************************************************/
-
-namespace internal {
-template<int StorageOrder> struct outer_product_selector;
-
-template<typename Lhs, typename Rhs>
-struct traits<GeneralProduct<Lhs,Rhs,OuterProduct> >
- : traits<ProductBase<GeneralProduct<Lhs,Rhs,OuterProduct>, Lhs, Rhs> >
-{};
-
-}
-
-template<typename Lhs, typename Rhs>
-class GeneralProduct<Lhs, Rhs, OuterProduct>
-  : public ProductBase<GeneralProduct<Lhs,Rhs,OuterProduct>, Lhs, Rhs>
+class Product : public ProductImpl<Lhs,Rhs,typename internal::promote_storage_type<typename internal::traits<Lhs>::StorageKind,
+                                                                            typename internal::traits<Rhs>::StorageKind>::ret>
 {
   public:
-    EIGEN_PRODUCT_PUBLIC_INTERFACE(GeneralProduct)
-
-    GeneralProduct(const Lhs& lhs, const Rhs& rhs) : Base(lhs,rhs)
-    {
-      EIGEN_STATIC_ASSERT((internal::is_same<typename Lhs::RealScalar, typename Rhs::RealScalar>::value),
-        YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY)
-    }
-
-    template<typename Dest> void scaleAndAddTo(Dest& dest, Scalar alpha) const
-    {
-      internal::outer_product_selector<(int(Dest::Flags)&RowMajorBit) ? RowMajor : ColMajor>::run(*this, dest, alpha);
-    }
-};
-
-namespace internal {
-
-template<> struct outer_product_selector<ColMajor> {
-  template<typename ProductType, typename Dest>
-  static EIGEN_DONT_INLINE void run(const ProductType& prod, Dest& dest, typename ProductType::Scalar alpha) {
-    typedef typename Dest::Index Index;
-    // FIXME make sure lhs is sequentially stored
-    // FIXME not very good if rhs is real and lhs complex while alpha is real too
-    const Index cols = dest.cols();
-    for (Index j=0; j<cols; ++j)
-      dest.col(j) += (alpha * prod.rhs().coeff(j)) * prod.lhs();
-  }
-};
-
-template<> struct outer_product_selector<RowMajor> {
-  template<typename ProductType, typename Dest>
-  static EIGEN_DONT_INLINE void run(const ProductType& prod, Dest& dest, typename ProductType::Scalar alpha) {
-    typedef typename Dest::Index Index;
-    // FIXME make sure rhs is sequentially stored
-    // FIXME not very good if lhs is real and rhs complex while alpha is real too
-    const Index rows = dest.rows();
-    for (Index i=0; i<rows; ++i)
-      dest.row(i) += (alpha * prod.lhs().coeff(i)) * prod.rhs();
-  }
-};
-
-} // end namespace internal
-
-/***********************************************************************
-*  Implementation of General Matrix Vector Product
-***********************************************************************/
-
-/*  According to the shape/flags of the matrix we have to distinghish 3 different cases:
- *   1 - the matrix is col-major, BLAS compatible and M is large => call fast BLAS-like colmajor routine
- *   2 - the matrix is row-major, BLAS compatible and N is large => call fast BLAS-like rowmajor routine
- *   3 - all other cases are handled using a simple loop along the outer-storage direction.
- *  Therefore we need a lower level meta selector.
- *  Furthermore, if the matrix is the rhs, then the product has to be transposed.
- */
-namespace internal {
-
-template<typename Lhs, typename Rhs>
-struct traits<GeneralProduct<Lhs,Rhs,GemvProduct> >
- : traits<ProductBase<GeneralProduct<Lhs,Rhs,GemvProduct>, Lhs, Rhs> >
-{};
-
-template<int Side, int StorageOrder, bool BlasCompatible>
-struct gemv_selector;
-
-} // end namespace internal
-
-template<typename Lhs, typename Rhs>
-class GeneralProduct<Lhs, Rhs, GemvProduct>
-  : public ProductBase<GeneralProduct<Lhs,Rhs,GemvProduct>, Lhs, Rhs>
-{
-  public:
-    EIGEN_PRODUCT_PUBLIC_INTERFACE(GeneralProduct)
-
-    typedef typename Lhs::Scalar LhsScalar;
-    typedef typename Rhs::Scalar RhsScalar;
-
-    GeneralProduct(const Lhs& lhs, const Rhs& rhs) : Base(lhs,rhs)
-    {
-//       EIGEN_STATIC_ASSERT((internal::is_same<typename Lhs::Scalar, typename Rhs::Scalar>::value),
-//         YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY)
-    }
-
-    enum { Side = Lhs::IsVectorAtCompileTime ? OnTheLeft : OnTheRight };
-    typedef typename internal::conditional<int(Side)==OnTheRight,_LhsNested,_RhsNested>::type MatrixType;
-
-    template<typename Dest> void scaleAndAddTo(Dest& dst, Scalar alpha) const
-    {
-      eigen_assert(m_lhs.rows() == dst.rows() && m_rhs.cols() == dst.cols());
-      internal::gemv_selector<Side,(int(MatrixType::Flags)&RowMajorBit) ? RowMajor : ColMajor,
-                       bool(internal::blas_traits<MatrixType>::HasUsableDirectAccess)>::run(*this, dst, alpha);
-    }
-};
-
-namespace internal {
-
-// The vector is on the left => transposition
-template<int StorageOrder, bool BlasCompatible>
-struct gemv_selector<OnTheLeft,StorageOrder,BlasCompatible>
-{
-  template<typename ProductType, typename Dest>
-  static void run(const ProductType& prod, Dest& dest, typename ProductType::Scalar alpha)
-  {
-    Transpose<Dest> destT(dest);
-    enum { OtherStorageOrder = StorageOrder == RowMajor ? ColMajor : RowMajor };
-    gemv_selector<OnTheRight,OtherStorageOrder,BlasCompatible>
-      ::run(GeneralProduct<Transpose<const typename ProductType::_RhsNested>,Transpose<const typename ProductType::_LhsNested>, GemvProduct>
-        (prod.rhs().transpose(), prod.lhs().transpose()), destT, alpha);
-  }
-};
-
-template<typename Scalar,int Size,int MaxSize,bool Cond> struct gemv_static_vector_if;
-
-template<typename Scalar,int Size,int MaxSize>
-struct gemv_static_vector_if<Scalar,Size,MaxSize,false>
-{
-  EIGEN_STRONG_INLINE  Scalar* data() { eigen_internal_assert(false && "should never be called"); return 0; }
-};
-
-template<typename Scalar,int Size>
-struct gemv_static_vector_if<Scalar,Size,Dynamic,true>
-{
-  EIGEN_STRONG_INLINE Scalar* data() { return 0; }
-};
-
-template<typename Scalar,int Size,int MaxSize>
-struct gemv_static_vector_if<Scalar,Size,MaxSize,true>
-{
-  #if EIGEN_ALIGN_STATICALLY
-  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;
-  EIGEN_STRONG_INLINE Scalar* data() {
-    return ForceAlignment
-            ? reinterpret_cast<Scalar*>((reinterpret_cast<size_t>(m_data.array) & ~(size_t(15))) + 16)
-            : m_data.array;
-  }
-  #endif
-};
-
-template<> struct gemv_selector<OnTheRight,ColMajor,true>
-{
-  template<typename ProductType, typename Dest>
-  static inline void run(const ProductType& prod, Dest& dest, typename ProductType::Scalar alpha)
-  {
-    typedef typename ProductType::Index Index;
-    typedef typename ProductType::LhsScalar   LhsScalar;
-    typedef typename ProductType::RhsScalar   RhsScalar;
-    typedef typename ProductType::Scalar      ResScalar;
-    typedef typename ProductType::RealScalar  RealScalar;
-    typedef typename ProductType::ActualLhsType ActualLhsType;
-    typedef typename ProductType::ActualRhsType ActualRhsType;
-    typedef typename ProductType::LhsBlasTraits LhsBlasTraits;
-    typedef typename ProductType::RhsBlasTraits RhsBlasTraits;
-    typedef Map<Matrix<ResScalar,Dynamic,1>, Aligned> MappedDest;
-
-    const ActualLhsType actualLhs = LhsBlasTraits::extract(prod.lhs());
-    const ActualRhsType actualRhs = RhsBlasTraits::extract(prod.rhs());
-
-    ResScalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(prod.lhs())
-                                  * RhsBlasTraits::extractScalarFactor(prod.rhs());
-
-    enum {
-      // FIXME find a way to allow an inner stride on the result if packet_traits<Scalar>::size==1
-      // on, the other hand it is good for the cache to pack the vector anyways...
-      EvalToDestAtCompileTime = Dest::InnerStrideAtCompileTime==1,
-      ComplexByReal = (NumTraits<LhsScalar>::IsComplex) && (!NumTraits<RhsScalar>::IsComplex),
-      MightCannotUseDest = (Dest::InnerStrideAtCompileTime!=1) || ComplexByReal
-    };
-
-    gemv_static_vector_if<ResScalar,Dest::SizeAtCompileTime,Dest::MaxSizeAtCompileTime,MightCannotUseDest> static_dest;
-
-    // this is written like this (i.e., with a ?:) to workaround an ICE with ICC 12
-    bool alphaIsCompatible = (!ComplexByReal) ? true : (imag(actualAlpha)==RealScalar(0));
-    bool evalToDest = EvalToDestAtCompileTime && alphaIsCompatible;
     
-    RhsScalar compatibleAlpha = get_factor<ResScalar,RhsScalar>::run(actualAlpha);
+    typedef typename ProductImpl<
+        Lhs, Rhs,
+        typename internal::promote_storage_type<typename Lhs::StorageKind,
+                                                typename Rhs::StorageKind>::ret>::Base Base;
+    EIGEN_GENERIC_PUBLIC_INTERFACE(Product)
 
-    ei_declare_aligned_stack_constructed_variable(ResScalar,actualDestPtr,dest.size(),
-                                                  evalToDest ? dest.data() : static_dest.data());
-    
-    if(!evalToDest)
+    typedef typename Lhs::Nested LhsNested;
+    typedef typename Rhs::Nested RhsNested;
+    typedef typename internal::remove_all<LhsNested>::type LhsNestedCleaned;
+    typedef typename internal::remove_all<RhsNested>::type RhsNestedCleaned;
+
+    Product(const Lhs& lhs, const Rhs& rhs) : m_lhs(lhs), m_rhs(rhs)
     {
-      #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN
-      int 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;
+      eigen_assert(lhs.cols() == rhs.rows()
+        && "invalid matrix product"
+        && "if you wanted a coeff-wise or a dot product use the respective explicit functions");
     }
 
-    general_matrix_vector_product
-      <Index,LhsScalar,ColMajor,LhsBlasTraits::NeedToConjugate,RhsScalar,RhsBlasTraits::NeedToConjugate>::run(
-        actualLhs.rows(), actualLhs.cols(),
-        &actualLhs.coeffRef(0,0), actualLhs.outerStride(),
-        actualRhs.data(), actualRhs.innerStride(),
-        actualDestPtr, 1,
-        compatibleAlpha);
+    inline Index rows() const { return m_lhs.rows(); }
+    inline Index cols() const { return m_rhs.cols(); }
 
-    if (!evalToDest)
-    {
-      if(!alphaIsCompatible)
-        dest += actualAlpha * MappedDest(actualDestPtr, dest.size());
-      else
-        dest = MappedDest(actualDestPtr, dest.size());
-    }
-  }
+    const LhsNestedCleaned& lhs() const { return m_lhs; }
+    const RhsNestedCleaned& rhs() const { return m_rhs; }
+
+  protected:
+
+    const LhsNested m_lhs;
+    const RhsNested m_rhs;
 };
 
-template<> struct gemv_selector<OnTheRight,RowMajor,true>
+template<typename Lhs, typename Rhs>
+class ProductImpl<Lhs,Rhs,Dense> : public internal::dense_xpr_base<Product<Lhs,Rhs> >::type
 {
-  template<typename ProductType, typename Dest>
-  static void run(const ProductType& prod, Dest& dest, typename ProductType::Scalar alpha)
-  {
-    typedef typename ProductType::LhsScalar LhsScalar;
-    typedef typename ProductType::RhsScalar RhsScalar;
-    typedef typename ProductType::Scalar    ResScalar;
-    typedef typename ProductType::Index Index;
-    typedef typename ProductType::ActualLhsType ActualLhsType;
-    typedef typename ProductType::ActualRhsType ActualRhsType;
-    typedef typename ProductType::_ActualRhsType _ActualRhsType;
-    typedef typename ProductType::LhsBlasTraits LhsBlasTraits;
-    typedef typename ProductType::RhsBlasTraits RhsBlasTraits;
+    typedef Product<Lhs, Rhs> Derived;
+  public:
 
-    typename add_const<ActualLhsType>::type actualLhs = LhsBlasTraits::extract(prod.lhs());
-    typename add_const<ActualRhsType>::type actualRhs = RhsBlasTraits::extract(prod.rhs());
-
-    ResScalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(prod.lhs())
-                                  * RhsBlasTraits::extractScalarFactor(prod.rhs());
-
-    enum {
-      // FIXME find a way to allow an inner stride on the result if packet_traits<Scalar>::size==1
-      // on, the other hand it is good for the cache to pack the vector anyways...
-      DirectlyUseRhs = _ActualRhsType::InnerStrideAtCompileTime==1
-    };
-
-    gemv_static_vector_if<RhsScalar,_ActualRhsType::SizeAtCompileTime,_ActualRhsType::MaxSizeAtCompileTime,!DirectlyUseRhs> static_rhs;
-
-    ei_declare_aligned_stack_constructed_variable(RhsScalar,actualRhsPtr,actualRhs.size(),
-        DirectlyUseRhs ? const_cast<RhsScalar*>(actualRhs.data()) : static_rhs.data());
-
-    if(!DirectlyUseRhs)
-    {
-      #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN
-      int size = actualRhs.size();
-      EIGEN_DENSE_STORAGE_CTOR_PLUGIN
-      #endif
-      Map<typename _ActualRhsType::PlainObject>(actualRhsPtr, actualRhs.size()) = actualRhs;
-    }
-
-    general_matrix_vector_product
-      <Index,LhsScalar,RowMajor,LhsBlasTraits::NeedToConjugate,RhsScalar,RhsBlasTraits::NeedToConjugate>::run(
-        actualLhs.rows(), actualLhs.cols(),
-        &actualLhs.coeffRef(0,0), actualLhs.outerStride(),
-        actualRhsPtr, 1,
-        &dest.coeffRef(0,0), dest.innerStride(),
-        actualAlpha);
-  }
+    typedef typename internal::dense_xpr_base<Product<Lhs, Rhs> >::type Base;
+    EIGEN_DENSE_PUBLIC_INTERFACE(Derived)
 };
 
-template<> struct gemv_selector<OnTheRight,ColMajor,false>
-{
-  template<typename ProductType, typename Dest>
-  static void run(const ProductType& prod, Dest& dest, typename ProductType::Scalar alpha)
-  {
-    typedef typename Dest::Index Index;
-    // TODO makes sure dest is sequentially stored in memory, otherwise use a temp
-    const Index size = prod.rhs().rows();
-    for(Index k=0; k<size; ++k)
-      dest += (alpha*prod.rhs().coeff(k)) * prod.lhs().col(k);
-  }
-};
-
-template<> struct gemv_selector<OnTheRight,RowMajor,false>
-{
-  template<typename ProductType, typename Dest>
-  static void run(const ProductType& prod, Dest& dest, typename ProductType::Scalar alpha)
-  {
-    typedef typename Dest::Index Index;
-    // TODO makes sure rhs is sequentially stored in memory, otherwise use a temp
-    const Index rows = prod.rows();
-    for(Index i=0; i<rows; ++i)
-      dest.coeffRef(i) += alpha * (prod.lhs().row(i).cwiseProduct(prod.rhs().transpose())).sum();
-  }
-};
-
-} // end namespace internal
-
-/***************************************************************************
-* Implementation of matrix base methods
-***************************************************************************/
-
-/** \returns the matrix product of \c *this and \a other.
-  *
-  * \note If instead of the matrix product you want the coefficient-wise product, see Cwise::operator*().
-  *
-  * \sa lazyProduct(), operator*=(const MatrixBase&), Cwise::operator*()
-  */
-template<typename Derived>
-template<typename OtherDerived>
-inline const typename ProductReturnType<Derived,OtherDerived>::Type
-MatrixBase<Derived>::operator*(const MatrixBase<OtherDerived> &other) const
-{
-  // A note regarding the function declaration: In MSVC, this function will sometimes
-  // not be inlined since DenseStorage is an unwindable object for dynamic
-  // matrices and product types are holding a member to store the result.
-  // Thus it does not help tagging this function with EIGEN_STRONG_INLINE.
-  enum {
-    ProductIsValid =  Derived::ColsAtCompileTime==Dynamic
-                   || OtherDerived::RowsAtCompileTime==Dynamic
-                   || int(Derived::ColsAtCompileTime)==int(OtherDerived::RowsAtCompileTime),
-    AreVectors = Derived::IsVectorAtCompileTime && OtherDerived::IsVectorAtCompileTime,
-    SameSizes = EIGEN_PREDICATE_SAME_MATRIX_SIZE(Derived,OtherDerived)
-  };
-  // note to the lost user:
-  //    * for a dot product use: v1.dot(v2)
-  //    * for a coeff-wise product use: v1.cwiseProduct(v2)
-  EIGEN_STATIC_ASSERT(ProductIsValid || !(AreVectors && SameSizes),
-    INVALID_VECTOR_VECTOR_PRODUCT__IF_YOU_WANTED_A_DOT_OR_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTIONS)
-  EIGEN_STATIC_ASSERT(ProductIsValid || !(SameSizes && !AreVectors),
-    INVALID_MATRIX_PRODUCT__IF_YOU_WANTED_A_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTION)
-  EIGEN_STATIC_ASSERT(ProductIsValid || SameSizes, INVALID_MATRIX_PRODUCT)
-#ifdef EIGEN_DEBUG_PRODUCT
-  internal::product_type<Derived,OtherDerived>::debug();
-#endif
-  return typename ProductReturnType<Derived,OtherDerived>::Type(derived(), other.derived());
-}
-
-/** \returns an expression of the matrix product of \c *this and \a other without implicit evaluation.
-  *
-  * The returned product will behave like any other expressions: the coefficients of the product will be
-  * computed once at a time as requested. This might be useful in some extremely rare cases when only
-  * a small and no coherent fraction of the result's coefficients have to be computed.
-  *
-  * \warning This version of the matrix product can be much much slower. So use it only if you know
-  * what you are doing and that you measured a true speed improvement.
-  *
-  * \sa operator*(const MatrixBase&)
-  */
-template<typename Derived>
-template<typename OtherDerived>
-const typename LazyProductReturnType<Derived,OtherDerived>::Type
-MatrixBase<Derived>::lazyProduct(const MatrixBase<OtherDerived> &other) const
-{
-  enum {
-    ProductIsValid =  Derived::ColsAtCompileTime==Dynamic
-                   || OtherDerived::RowsAtCompileTime==Dynamic
-                   || int(Derived::ColsAtCompileTime)==int(OtherDerived::RowsAtCompileTime),
-    AreVectors = Derived::IsVectorAtCompileTime && OtherDerived::IsVectorAtCompileTime,
-    SameSizes = EIGEN_PREDICATE_SAME_MATRIX_SIZE(Derived,OtherDerived)
-  };
-  // note to the lost user:
-  //    * for a dot product use: v1.dot(v2)
-  //    * for a coeff-wise product use: v1.cwiseProduct(v2)
-  EIGEN_STATIC_ASSERT(ProductIsValid || !(AreVectors && SameSizes),
-    INVALID_VECTOR_VECTOR_PRODUCT__IF_YOU_WANTED_A_DOT_OR_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTIONS)
-  EIGEN_STATIC_ASSERT(ProductIsValid || !(SameSizes && !AreVectors),
-    INVALID_MATRIX_PRODUCT__IF_YOU_WANTED_A_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTION)
-  EIGEN_STATIC_ASSERT(ProductIsValid || SameSizes, INVALID_MATRIX_PRODUCT)
-
-  return typename LazyProductReturnType<Derived,OtherDerived>::Type(derived(), other.derived());
-}
-
 #endif // EIGEN_PRODUCT_H

Eigen/src/Core/ProductBase.h

@@ -115,10 +115,10 @@ class ProductBase : public MatrixBase<Derived>
     inline void evalTo(Dest& dst) const { dst.setZero(); scaleAndAddTo(dst,Scalar(1)); }
 
     template<typename Dest>
-    inline void addTo(Dest& dst) const { scaleAndAddTo(dst,1); }
+    inline void addTo(Dest& dst) const { scaleAndAddTo(dst,Scalar(1)); }
 
     template<typename Dest>
-    inline void subTo(Dest& dst) const { scaleAndAddTo(dst,-1); }
+    inline void subTo(Dest& dst) const { scaleAndAddTo(dst,Scalar(-1)); }
 
     template<typename Dest>
     inline void scaleAndAddTo(Dest& dst,Scalar alpha) const { derived().scaleAndAddTo(dst,alpha); }
@@ -179,8 +179,8 @@ class ProductBase : public MatrixBase<Derived>
 
   protected:
 
-    const LhsNested m_lhs;
-    const RhsNested m_rhs;
+    LhsNested m_lhs;
+    RhsNested m_rhs;
 
     mutable PlainObject m_result;
 };

Eigen/src/Core/Redux.h

@@ -95,7 +95,7 @@ struct redux_novec_unroller
 
   typedef typename Derived::Scalar Scalar;
 
-  EIGEN_STRONG_INLINE static Scalar run(const Derived &mat, const Func& func)
+  static EIGEN_STRONG_INLINE Scalar run(const Derived &mat, const Func& func)
   {
     return func(redux_novec_unroller<Func, Derived, Start, HalfLength>::run(mat,func),
                 redux_novec_unroller<Func, Derived, Start+HalfLength, Length-HalfLength>::run(mat,func));
@@ -112,7 +112,7 @@ struct redux_novec_unroller<Func, Derived, Start, 1>
 
   typedef typename Derived::Scalar Scalar;
 
-  EIGEN_STRONG_INLINE static Scalar run(const Derived &mat, const Func&)
+  static EIGEN_STRONG_INLINE Scalar run(const Derived &mat, const Func&)
   {
     return mat.coeffByOuterInner(outer, inner);
   }
@@ -125,7 +125,7 @@ template<typename Func, typename Derived, int Start>
 struct redux_novec_unroller<Func, Derived, Start, 0>
 {
   typedef typename Derived::Scalar Scalar;
-  EIGEN_STRONG_INLINE static Scalar run(const Derived&, const Func&) { return Scalar(); }
+  static EIGEN_STRONG_INLINE Scalar run(const Derived&, const Func&) { return Scalar(); }
 };
 
 /*** vectorization ***/
@@ -141,7 +141,7 @@ struct redux_vec_unroller
   typedef typename Derived::Scalar Scalar;
   typedef typename packet_traits<Scalar>::type PacketScalar;
 
-  EIGEN_STRONG_INLINE static PacketScalar run(const Derived &mat, const Func& func)
+  static EIGEN_STRONG_INLINE PacketScalar run(const Derived &mat, const Func& func)
   {
     return func.packetOp(
             redux_vec_unroller<Func, Derived, Start, HalfLength>::run(mat,func),
@@ -162,7 +162,7 @@ struct redux_vec_unroller<Func, Derived, Start, 1>
   typedef typename Derived::Scalar Scalar;
   typedef typename packet_traits<Scalar>::type PacketScalar;
 
-  EIGEN_STRONG_INLINE static PacketScalar run(const Derived &mat, const Func&)
+  static EIGEN_STRONG_INLINE PacketScalar run(const Derived &mat, const Func&)
   {
     return mat.template packetByOuterInner<alignment>(outer, inner);
   }
@@ -214,20 +214,33 @@ struct redux_impl<Func, Derived, LinearVectorizedTraversal, NoUnrolling>
     const Index size = mat.size();
     eigen_assert(size && "you are using an empty matrix");
     const Index packetSize = packet_traits<Scalar>::size;
-    const Index alignedStart = first_aligned(mat);
+    const Index alignedStart = internal::first_aligned(mat);
     enum {
       alignment = bool(Derived::Flags & DirectAccessBit) || bool(Derived::Flags & AlignedBit)
                 ? Aligned : Unaligned
     };
-    const Index alignedSize = ((size-alignedStart)/packetSize)*packetSize;
-    const Index alignedEnd = alignedStart + alignedSize;
+    const Index alignedSize2 = ((size-alignedStart)/(2*packetSize))*(2*packetSize);
+    const Index alignedSize = ((size-alignedStart)/(packetSize))*(packetSize);
+    const Index alignedEnd2 = alignedStart + alignedSize2;
+    const Index alignedEnd  = alignedStart + alignedSize;
     Scalar res;
     if(alignedSize)
     {
-      PacketScalar packet_res = mat.template packet<alignment>(alignedStart);
-      for(Index index = alignedStart + packetSize; index < alignedEnd; index += packetSize)
-        packet_res = func.packetOp(packet_res, mat.template packet<alignment>(index));
-      res = func.predux(packet_res);
+      PacketScalar packet_res0 = mat.template packet<alignment>(alignedStart);
+      if(alignedSize>packetSize) // we have at least two packets to partly unroll the loop
+      {
+        PacketScalar packet_res1 = mat.template packet<alignment>(alignedStart+packetSize);
+        for(Index index = alignedStart + 2*packetSize; index < alignedEnd2; index += 2*packetSize)
+        {
+          packet_res0 = func.packetOp(packet_res0, mat.template packet<alignment>(index));
+          packet_res1 = func.packetOp(packet_res1, mat.template packet<alignment>(index+packetSize));
+        }
+
+        packet_res0 = func.packetOp(packet_res0,packet_res1);
+        if(alignedEnd>alignedEnd2)
+          packet_res0 = func.packetOp(packet_res0, mat.template packet<alignment>(alignedEnd2));
+      }
+      res = func.predux(packet_res0);
 
       for(Index index = 0; index < alignedStart; ++index)
         res = func(res,mat.coeff(index));
@@ -296,7 +309,7 @@ struct redux_impl<Func, Derived, LinearVectorizedTraversal, CompleteUnrolling>
     Size = Derived::SizeAtCompileTime,
     VectorizedSize = (Size / PacketSize) * PacketSize
   };
-  EIGEN_STRONG_INLINE static Scalar run(const Derived& mat, const Func& func)
+  static EIGEN_STRONG_INLINE Scalar run(const Derived& mat, const Func& func)
   {
     eigen_assert(mat.rows()>0 && mat.cols()>0 && "you are using an empty matrix");
     Scalar res = func.predux(redux_vec_unroller<Func, Derived, 0, Size / PacketSize>::run(mat,func));

Eigen/src/Core/Replicate.h

@@ -122,9 +122,13 @@ template<typename MatrixType,int RowFactor,int ColFactor> class Replicate
       return m_matrix.template packet<LoadMode>(actual_row, actual_col);
     }
 
+    const typename internal::remove_all<typename MatrixType::Nested>::type& nestedExpression() const 
+    { 
+      return m_matrix; 
+    }
 
   protected:
-    const typename MatrixType::Nested m_matrix;
+    typename MatrixType::Nested m_matrix;
     const internal::variable_if_dynamic<Index, RowFactor> m_rowFactor;
     const internal::variable_if_dynamic<Index, ColFactor> m_colFactor;
 };

Eigen/src/Core/Reverse.h

@@ -183,8 +183,14 @@ template<typename MatrixType, int Direction> class Reverse
       m_matrix.const_cast_derived().template writePacket<LoadMode>(m_matrix.size() - index - PacketSize, internal::preverse(x));
     }
 
+    const typename internal::remove_all<typename MatrixType::Nested>::type& 
+    nestedExpression() const 
+    {
+      return m_matrix;
+    }
+
   protected:
-    const typename MatrixType::Nested m_matrix;
+    typename MatrixType::Nested m_matrix;
 };
 
 /** \returns an expression of the reverse of *this.

Eigen/src/Core/Select.h

@@ -101,10 +101,25 @@ class Select : internal::no_assignment_operator,
         return m_else.coeff(i);
     }
 
+    const ConditionMatrixType& conditionMatrix() const
+    {
+      return m_condition;
+    }
+
+    const ThenMatrixType& thenMatrix() const
+    {
+      return m_then;
+    }
+
+    const ElseMatrixType& elseMatrix() const
+    {
+      return m_else;
+    }
+
   protected:
-    const typename ConditionMatrixType::Nested m_condition;
-    const typename ThenMatrixType::Nested m_then;
-    const typename ElseMatrixType::Nested m_else;
+    typename ConditionMatrixType::Nested m_condition;
+    typename ThenMatrixType::Nested m_then;
+    typename ElseMatrixType::Nested m_else;
 };
 
 

Eigen/src/Core/SelfAdjointView.h

@@ -82,7 +82,7 @@ template<typename MatrixType, unsigned int UpLo> class SelfAdjointView
     };
     typedef typename MatrixType::PlainObject PlainObject;
 
-    inline SelfAdjointView(const MatrixType& matrix) : m_matrix(matrix)
+    inline SelfAdjointView(MatrixType& matrix) : m_matrix(matrix)
     {}
 
     inline Index rows() const { return m_matrix.rows(); }
@@ -199,7 +199,7 @@ template<typename MatrixType, unsigned int UpLo> class SelfAdjointView
     #endif
 
   protected:
-    const MatrixTypeNested m_matrix;
+    MatrixTypeNested m_matrix;
 };
 
 
@@ -222,7 +222,7 @@ struct triangular_assignment_selector<Derived1, Derived2, (SelfAdjoint|Upper), U
     row = (UnrollCount-1) % Derived1::RowsAtCompileTime
   };
 
-  inline static void run(Derived1 &dst, const Derived2 &src)
+  static inline void run(Derived1 &dst, const Derived2 &src)
   {
     triangular_assignment_selector<Derived1, Derived2, (SelfAdjoint|Upper), UnrollCount-1, ClearOpposite>::run(dst, src);
 
@@ -236,7 +236,7 @@ struct triangular_assignment_selector<Derived1, Derived2, (SelfAdjoint|Upper), U
 template<typename Derived1, typename Derived2, bool ClearOpposite>
 struct triangular_assignment_selector<Derived1, Derived2, SelfAdjoint|Upper, 0, ClearOpposite>
 {
-  inline static void run(Derived1 &, const Derived2 &) {}
+  static inline void run(Derived1 &, const Derived2 &) {}
 };
 
 template<typename Derived1, typename Derived2, int UnrollCount, bool ClearOpposite>
@@ -247,7 +247,7 @@ struct triangular_assignment_selector<Derived1, Derived2, (SelfAdjoint|Lower), U
     row = (UnrollCount-1) % Derived1::RowsAtCompileTime
   };
 
-  inline static void run(Derived1 &dst, const Derived2 &src)
+  static inline void run(Derived1 &dst, const Derived2 &src)
   {
     triangular_assignment_selector<Derived1, Derived2, (SelfAdjoint|Lower), UnrollCount-1, ClearOpposite>::run(dst, src);
 
@@ -261,14 +261,14 @@ struct triangular_assignment_selector<Derived1, Derived2, (SelfAdjoint|Lower), U
 template<typename Derived1, typename Derived2, bool ClearOpposite>
 struct triangular_assignment_selector<Derived1, Derived2, SelfAdjoint|Lower, 0, ClearOpposite>
 {
-  inline static void run(Derived1 &, const Derived2 &) {}
+  static inline void run(Derived1 &, const Derived2 &) {}
 };
 
 template<typename Derived1, typename Derived2, bool ClearOpposite>
 struct triangular_assignment_selector<Derived1, Derived2, SelfAdjoint|Upper, Dynamic, ClearOpposite>
 {
   typedef typename Derived1::Index Index;
-  inline static void run(Derived1 &dst, const Derived2 &src)
+  static inline void run(Derived1 &dst, const Derived2 &src)
   {
     for(Index j = 0; j < dst.cols(); ++j)
     {
@@ -285,7 +285,7 @@ struct triangular_assignment_selector<Derived1, Derived2, SelfAdjoint|Upper, Dyn
 template<typename Derived1, typename Derived2, bool ClearOpposite>
 struct triangular_assignment_selector<Derived1, Derived2, SelfAdjoint|Lower, Dynamic, ClearOpposite>
 {
-  inline static void run(Derived1 &dst, const Derived2 &src)
+  static inline void run(Derived1 &dst, const Derived2 &src)
   {
   typedef typename Derived1::Index Index;
     for(Index i = 0; i < dst.rows(); ++i)

Eigen/src/Core/SelfCwiseBinaryOp.h

@@ -163,6 +163,16 @@ template<typename BinaryOp, typename Lhs, typename Rhs> class SelfCwiseBinaryOp
       return Base::operator=(rhs);
     }
 
+    Lhs& expression() const 
+    { 
+      return m_matrix;
+    }
+
+    const BinaryOp& functor() const 
+    { 
+      return m_functor;
+    }
+
   protected:
     Lhs& m_matrix;
     const BinaryOp& m_functor;

Eigen/src/Core/SolveTriangular.h

@@ -100,7 +100,7 @@ struct triangular_solver_selector<Lhs,Rhs,Side,Mode,NoUnrolling,Dynamic>
   typedef typename LhsProductTraits::DirectLinearAccessType ActualLhsType;
   static void run(const Lhs& lhs, Rhs& rhs)
   {
-    const ActualLhsType actualLhs = LhsProductTraits::extract(lhs);
+    typename internal::add_const_on_value_type<ActualLhsType>::type actualLhs = LhsProductTraits::extract(lhs);
     triangular_solve_matrix<Scalar,Index,Side,Mode,LhsProductTraits::NeedToConjugate,(int(Lhs::Flags) & RowMajorBit) ? RowMajor : ColMajor,
                                (Rhs::Flags&RowMajorBit) ? RowMajor : ColMajor>
       ::run(lhs.rows(), Side==OnTheLeft? rhs.cols() : rhs.rows(), &actualLhs.coeffRef(0,0), actualLhs.outerStride(), &rhs.coeffRef(0,0), rhs.outerStride());
@@ -177,10 +177,8 @@ template<int Side, typename OtherDerived>
 void TriangularView<MatrixType,Mode>::solveInPlace(const MatrixBase<OtherDerived>& _other) const
 {
   OtherDerived& other = _other.const_cast_derived();
-  eigen_assert(cols() == rows());
-  eigen_assert( (Side==OnTheLeft && cols() == other.rows()) || (Side==OnTheRight && cols() == other.cols()) );
-  eigen_assert(!(Mode & ZeroDiag));
-  eigen_assert((Mode & (Upper|Lower)) != 0);
+  eigen_assert( cols() == rows() && ((Side==OnTheLeft && cols() == other.rows()) || (Side==OnTheRight && cols() == other.cols())) );
+  eigen_assert((!(Mode & ZeroDiag)) && bool(Mode & (Upper|Lower)));
 
   enum { copy = internal::traits<OtherDerived>::Flags & RowMajorBit  && OtherDerived::IsVectorAtCompileTime };
   typedef typename internal::conditional<copy,
@@ -255,7 +253,7 @@ template<int Side, typename TriangularType, typename Rhs> struct triangular_solv
 
   protected:
     const TriangularType& m_triangularMatrix;
-    const typename Rhs::Nested m_rhs;
+    typename Rhs::Nested m_rhs;
 };
 
 } // namespace internal

Eigen/src/Core/StableNorm.h

@@ -58,9 +58,9 @@ MatrixBase<Derived>::stableNorm() const
 {
   using std::min;
   const Index blockSize = 4096;
-  RealScalar scale = 0;
-  RealScalar invScale = 1;
-  RealScalar ssq = 0; // sum of square
+  RealScalar scale(0);
+  RealScalar invScale(1);
+  RealScalar ssq(0); // sum of square
   enum {
     Alignment = (int(Flags)&DirectAccessBit) || (int(Flags)&AlignedBit) ? 1 : 0
   };

Eigen/src/Core/Swap.h

@@ -52,6 +52,15 @@ template<typename ExpressionType> class SwapWrapper
     inline Index cols() const { return m_expression.cols(); }
     inline Index outerStride() const { return m_expression.outerStride(); }
     inline Index innerStride() const { return m_expression.innerStride(); }
+    
+    typedef typename internal::conditional<
+                       internal::is_lvalue<ExpressionType>::value,
+                       Scalar,
+                       const Scalar
+                     >::type ScalarWithConstIfNotLvalue;
+                     
+    inline ScalarWithConstIfNotLvalue* data() { return m_expression.data(); }
+    inline const Scalar* data() const { return m_expression.data(); }
 
     inline Scalar& coeffRef(Index row, Index col)
     {
@@ -119,6 +128,8 @@ template<typename ExpressionType> class SwapWrapper
       _other.template writePacket<LoadMode>(index, tmp);
     }
 
+    ExpressionType& expression() const { return m_expression; }
+
   protected:
     ExpressionType& m_expression;
 };

Eigen/src/Core/Transpose.h

@@ -91,7 +91,7 @@ template<typename MatrixType> class Transpose
     nestedExpression() { return m_matrix.const_cast_derived(); }
 
   protected:
-    const typename MatrixType::Nested m_matrix;
+    typename MatrixType::Nested m_matrix;
 };
 
 namespace internal {
@@ -152,12 +152,12 @@ template<typename MatrixType> class TransposeImpl<MatrixType,Dense>
       return derived().nestedExpression().coeffRef(index);
     }
 
-    inline const CoeffReturnType coeff(Index row, Index col) const
+    inline CoeffReturnType coeff(Index row, Index col) const
     {
       return derived().nestedExpression().coeff(col, row);
     }
 
-    inline const CoeffReturnType coeff(Index index) const
+    inline CoeffReturnType coeff(Index index) const
     {
       return derived().nestedExpression().coeff(index);
     }

Eigen/src/Core/Transpositions.h

@@ -404,7 +404,7 @@ struct transposition_matrix_product_retval
 
   protected:
     const TranspositionType& m_transpositions;
-    const typename MatrixType::Nested m_matrix;
+    typename MatrixType::Nested m_matrix;
 };
 
 } // end namespace internal

Eigen/src/Core/TriangularMatrix.h

@@ -273,11 +273,8 @@ template<typename _MatrixType, unsigned int _Mode> class TriangularView
     inline const TriangularView<MatrixConjugateReturnType,Mode> conjugate() const
     { return m_matrix.conjugate(); }
 
-    /** \sa MatrixBase::adjoint() */
-    inline TriangularView<typename MatrixType::AdjointReturnType,TransposeMode> adjoint()
-    { return m_matrix.adjoint(); }
     /** \sa MatrixBase::adjoint() const */
-    inline const TriangularView<typename MatrixType::AdjointReturnType,TransposeMode> adjoint() const
+    inline const TriangularView<const typename MatrixType::AdjointReturnType,TransposeMode> adjoint() const
     { return m_matrix.adjoint(); }
 
     /** \sa MatrixBase::transpose() */
@@ -288,11 +285,13 @@ template<typename _MatrixType, unsigned int _Mode> class TriangularView
     }
     /** \sa MatrixBase::transpose() const */
     inline const TriangularView<Transpose<MatrixType>,TransposeMode> transpose() const
-    { return m_matrix.transpose(); }
+    {
+      return m_matrix.transpose();
+    }
 
     /** Efficient triangular matrix times vector/matrix product */
     template<typename OtherDerived>
-    TriangularProduct<Mode,true,MatrixType,false,OtherDerived,OtherDerived::IsVectorAtCompileTime>
+    TriangularProduct<Mode,true,MatrixType,false,OtherDerived, OtherDerived::IsVectorAtCompileTime>
     operator*(const MatrixBase<OtherDerived>& rhs) const
     {
       return TriangularProduct
@@ -375,7 +374,8 @@ template<typename _MatrixType, unsigned int _Mode> class TriangularView
     template<typename OtherDerived>
     void swap(MatrixBase<OtherDerived> const & other)
     {
-      TriangularView<SwapWrapper<MatrixType>,Mode>(const_cast<MatrixType&>(m_matrix)).lazyAssign(other.derived());
+      SwapWrapper<MatrixType> swaper(const_cast<MatrixType&>(m_matrix));
+      TriangularView<SwapWrapper<MatrixType>,Mode>(swaper).lazyAssign(other.derived());
     }
 
     Scalar determinant() const
@@ -433,7 +433,7 @@ template<typename _MatrixType, unsigned int _Mode> class TriangularView
     template<typename ProductDerived, typename Lhs, typename Rhs>
     EIGEN_STRONG_INLINE TriangularView& assignProduct(const ProductBase<ProductDerived, Lhs,Rhs>& prod, const Scalar& alpha);
 
-    const MatrixTypeNested m_matrix;
+    MatrixTypeNested m_matrix;
 };
 
 /***************************************************************************
@@ -452,7 +452,7 @@ struct triangular_assignment_selector
   
   typedef typename Derived1::Scalar Scalar;
 
-  inline static void run(Derived1 &dst, const Derived2 &src)
+  static inline void run(Derived1 &dst, const Derived2 &src)
   {
     triangular_assignment_selector<Derived1, Derived2, Mode, UnrollCount-1, ClearOpposite>::run(dst, src);
 
@@ -480,7 +480,7 @@ struct triangular_assignment_selector
 template<typename Derived1, typename Derived2, unsigned int Mode, bool ClearOpposite>
 struct triangular_assignment_selector<Derived1, Derived2, Mode, 0, ClearOpposite>
 {
-  inline static void run(Derived1 &, const Derived2 &) {}
+  static inline void run(Derived1 &, const Derived2 &) {}
 };
 
 template<typename Derived1, typename Derived2, bool ClearOpposite>
@@ -488,7 +488,7 @@ struct triangular_assignment_selector<Derived1, Derived2, Upper, Dynamic, ClearO
 {
   typedef typename Derived1::Index Index;
   typedef typename Derived1::Scalar Scalar;
-  inline static void run(Derived1 &dst, const Derived2 &src)
+  static inline void run(Derived1 &dst, const Derived2 &src)
   {
     for(Index j = 0; j < dst.cols(); ++j)
     {
@@ -506,7 +506,7 @@ template<typename Derived1, typename Derived2, bool ClearOpposite>
 struct triangular_assignment_selector<Derived1, Derived2, Lower, Dynamic, ClearOpposite>
 {
   typedef typename Derived1::Index Index;
-  inline static void run(Derived1 &dst, const Derived2 &src)
+  static inline void run(Derived1 &dst, const Derived2 &src)
   {
     for(Index j = 0; j < dst.cols(); ++j)
     {
@@ -524,7 +524,7 @@ template<typename Derived1, typename Derived2, bool ClearOpposite>
 struct triangular_assignment_selector<Derived1, Derived2, StrictlyUpper, Dynamic, ClearOpposite>
 {
   typedef typename Derived1::Index Index;
-  inline static void run(Derived1 &dst, const Derived2 &src)
+  static inline void run(Derived1 &dst, const Derived2 &src)
   {
     for(Index j = 0; j < dst.cols(); ++j)
     {
@@ -542,7 +542,7 @@ template<typename Derived1, typename Derived2, bool ClearOpposite>
 struct triangular_assignment_selector<Derived1, Derived2, StrictlyLower, Dynamic, ClearOpposite>
 {
   typedef typename Derived1::Index Index;
-  inline static void run(Derived1 &dst, const Derived2 &src)
+  static inline void run(Derived1 &dst, const Derived2 &src)
   {
     for(Index j = 0; j < dst.cols(); ++j)
     {
@@ -560,7 +560,7 @@ template<typename Derived1, typename Derived2, bool ClearOpposite>
 struct triangular_assignment_selector<Derived1, Derived2, UnitUpper, Dynamic, ClearOpposite>
 {
   typedef typename Derived1::Index Index;
-  inline static void run(Derived1 &dst, const Derived2 &src)
+  static inline void run(Derived1 &dst, const Derived2 &src)
   {
     for(Index j = 0; j < dst.cols(); ++j)
     {
@@ -580,7 +580,7 @@ template<typename Derived1, typename Derived2, bool ClearOpposite>
 struct triangular_assignment_selector<Derived1, Derived2, UnitLower, Dynamic, ClearOpposite>
 {
   typedef typename Derived1::Index Index;
-  inline static void run(Derived1 &dst, const Derived2 &src)
+  static inline void run(Derived1 &dst, const Derived2 &src)
   {
     for(Index j = 0; j < dst.cols(); ++j)
     {

Eigen/src/Core/VectorwiseOp.h

@@ -110,7 +110,7 @@ class PartialReduxExpr : internal::no_assignment_operator,
     }
 
   protected:
-    const MatrixTypeNested m_matrix;
+    MatrixTypeNested m_matrix;
     const MemberOp m_functor;
 };
 
@@ -237,7 +237,10 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
     typename ExtendedType<OtherDerived>::Type
     extendedTo(const DenseBase<OtherDerived>& other) const
     {
-      EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived);
+      EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(Direction==Vertical, OtherDerived::MaxColsAtCompileTime==1),
+                          YOU_PASSED_A_ROW_VECTOR_BUT_A_COLUMN_VECTOR_WAS_EXPECTED)
+      EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(Direction==Horizontal, OtherDerived::MaxRowsAtCompileTime==1),
+                          YOU_PASSED_A_COLUMN_VECTOR_BUT_A_ROW_VECTOR_WAS_EXPECTED)
       return typename ExtendedType<OtherDerived>::Type
                       (other.derived(),
                        Direction==Vertical   ? 1 : m_matrix.rows(),
@@ -418,10 +421,9 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
     ExpressionType& operator=(const DenseBase<OtherDerived>& other)
     {
       EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
+      EIGEN_STATIC_ASSERT_SAME_XPR_KIND(ExpressionType, OtherDerived)
       //eigen_assert((m_matrix.isNull()) == (other.isNull())); FIXME
-      for(Index j=0; j<subVectors(); ++j)
-        subVector(j) = other;
-      return const_cast<ExpressionType&>(m_matrix);
+      return const_cast<ExpressionType&>(m_matrix = extendedTo(other.derived()));
     }
 
     /** Adds the vector \a other to each subvector of \c *this */
@@ -429,9 +431,8 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
     ExpressionType& operator+=(const DenseBase<OtherDerived>& other)
     {
       EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
-      for(Index j=0; j<subVectors(); ++j)
-        subVector(j) += other.derived();
-      return const_cast<ExpressionType&>(m_matrix);
+      EIGEN_STATIC_ASSERT_SAME_XPR_KIND(ExpressionType, OtherDerived)
+      return const_cast<ExpressionType&>(m_matrix += extendedTo(other.derived()));
     }
 
     /** Substracts the vector \a other to each subvector of \c *this */
@@ -439,8 +440,29 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
     ExpressionType& operator-=(const DenseBase<OtherDerived>& other)
     {
       EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
-      for(Index j=0; j<subVectors(); ++j)
-        subVector(j) -= other.derived();
+      EIGEN_STATIC_ASSERT_SAME_XPR_KIND(ExpressionType, OtherDerived)
+      return const_cast<ExpressionType&>(m_matrix -= extendedTo(other.derived()));
+    }
+
+    /** Multiples each subvector of \c *this by the vector \a other */
+    template<typename OtherDerived>
+    ExpressionType& operator*=(const DenseBase<OtherDerived>& other)
+    {
+      EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
+      EIGEN_STATIC_ASSERT_ARRAYXPR(ExpressionType)
+      EIGEN_STATIC_ASSERT_SAME_XPR_KIND(ExpressionType, OtherDerived)
+      m_matrix *= extendedTo(other.derived());
+      return const_cast<ExpressionType&>(m_matrix);
+    }
+
+    /** Divides each subvector of \c *this by the vector \a other */
+    template<typename OtherDerived>
+    ExpressionType& operator/=(const DenseBase<OtherDerived>& other)
+    {
+      EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
+      EIGEN_STATIC_ASSERT_ARRAYXPR(ExpressionType)
+      EIGEN_STATIC_ASSERT_SAME_XPR_KIND(ExpressionType, OtherDerived)
+      m_matrix /= extendedTo(other.derived());
       return const_cast<ExpressionType&>(m_matrix);
     }
 
@@ -451,7 +473,8 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
                   const typename ExtendedType<OtherDerived>::Type>
     operator+(const DenseBase<OtherDerived>& other) const
     {
-      EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived);
+      EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
+      EIGEN_STATIC_ASSERT_SAME_XPR_KIND(ExpressionType, OtherDerived)
       return m_matrix + extendedTo(other.derived());
     }
 
@@ -462,10 +485,39 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
                   const typename ExtendedType<OtherDerived>::Type>
     operator-(const DenseBase<OtherDerived>& other) const
     {
-      EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived);
+      EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
+      EIGEN_STATIC_ASSERT_SAME_XPR_KIND(ExpressionType, OtherDerived)
       return m_matrix - extendedTo(other.derived());
     }
 
+    /** Returns the expression where each subvector is the product of the vector \a other
+      * by the corresponding subvector of \c *this */
+    template<typename OtherDerived> EIGEN_STRONG_INLINE
+    CwiseBinaryOp<internal::scalar_product_op<Scalar>,
+                  const ExpressionTypeNestedCleaned,
+                  const typename ExtendedType<OtherDerived>::Type>
+    operator*(const DenseBase<OtherDerived>& other) const
+    {
+      EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
+      EIGEN_STATIC_ASSERT_ARRAYXPR(ExpressionType)
+      EIGEN_STATIC_ASSERT_SAME_XPR_KIND(ExpressionType, OtherDerived)
+      return m_matrix * extendedTo(other.derived());
+    }
+
+    /** Returns the expression where each subvector is the quotient of the corresponding
+      * subvector of \c *this by the vector \a other */
+    template<typename OtherDerived>
+    CwiseBinaryOp<internal::scalar_quotient_op<Scalar>,
+                  const ExpressionTypeNestedCleaned,
+                  const typename ExtendedType<OtherDerived>::Type>
+    operator/(const DenseBase<OtherDerived>& other) const
+    {
+      EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
+      EIGEN_STATIC_ASSERT_ARRAYXPR(ExpressionType)
+      EIGEN_STATIC_ASSERT_SAME_XPR_KIND(ExpressionType, OtherDerived)
+      return m_matrix / extendedTo(other.derived());
+    }
+
 /////////// Geometry module ///////////
 
     #if EIGEN2_SUPPORT_STAGE > STAGE20_RESOLVE_API_CONFLICTS
@@ -509,7 +561,7 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
   * Example: \include MatrixBase_colwise.cpp
   * Output: \verbinclude MatrixBase_colwise.out
   *
-  * \sa rowwise(), class VectorwiseOp
+  * \sa rowwise(), class VectorwiseOp, \ref TutorialReductionsVisitorsBroadcasting
   */
 template<typename Derived>
 inline const typename DenseBase<Derived>::ConstColwiseReturnType
@@ -520,7 +572,7 @@ DenseBase<Derived>::colwise() const
 
 /** \returns a writable VectorwiseOp wrapper of *this providing additional partial reduction operations
   *
-  * \sa rowwise(), class VectorwiseOp
+  * \sa rowwise(), class VectorwiseOp, \ref TutorialReductionsVisitorsBroadcasting
   */
 template<typename Derived>
 inline typename DenseBase<Derived>::ColwiseReturnType
@@ -534,7 +586,7 @@ DenseBase<Derived>::colwise()
   * Example: \include MatrixBase_rowwise.cpp
   * Output: \verbinclude MatrixBase_rowwise.out
   *
-  * \sa colwise(), class VectorwiseOp
+  * \sa colwise(), class VectorwiseOp, \ref TutorialReductionsVisitorsBroadcasting
   */
 template<typename Derived>
 inline const typename DenseBase<Derived>::ConstRowwiseReturnType
@@ -545,7 +597,7 @@ DenseBase<Derived>::rowwise() const
 
 /** \returns a writable VectorwiseOp wrapper of *this providing additional partial reduction operations
   *
-  * \sa colwise(), class VectorwiseOp
+  * \sa colwise(), class VectorwiseOp, \ref TutorialReductionsVisitorsBroadcasting
   */
 template<typename Derived>
 inline typename DenseBase<Derived>::RowwiseReturnType

Eigen/src/Core/Visitor.h

@@ -35,7 +35,7 @@ struct visitor_impl
     row = (UnrollCount-1) % Derived::RowsAtCompileTime
   };
 
-  inline static void run(const Derived &mat, Visitor& visitor)
+  static inline void run(const Derived &mat, Visitor& visitor)
   {
     visitor_impl<Visitor, Derived, UnrollCount-1>::run(mat, visitor);
     visitor(mat.coeff(row, col), row, col);
@@ -45,7 +45,7 @@ struct visitor_impl
 template<typename Visitor, typename Derived>
 struct visitor_impl<Visitor, Derived, 1>
 {
-  inline static void run(const Derived &mat, Visitor& visitor)
+  static inline void run(const Derived &mat, Visitor& visitor)
   {
     return visitor.init(mat.coeff(0, 0), 0, 0);
   }
@@ -55,7 +55,7 @@ template<typename Visitor, typename Derived>
 struct visitor_impl<Visitor, Derived, Dynamic>
 {
   typedef typename Derived::Index Index;
-  inline static void run(const Derived& mat, Visitor& visitor)
+  static inline void run(const Derived& mat, Visitor& visitor)
   {
     visitor.init(mat.coeff(0,0), 0, 0);
     for(Index i = 1; i < mat.rows(); ++i)

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

@@ -168,7 +168,7 @@ template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(const P
 template<int Offset>
 struct palign_impl<Offset,Packet2cf>
 {
-  EIGEN_STRONG_INLINE static void run(Packet2cf& first, const Packet2cf& second)
+  static EIGEN_STRONG_INLINE void run(Packet2cf& first, const Packet2cf& second)
   {
     if (Offset==1)
     {

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

@@ -487,7 +487,7 @@ template<> EIGEN_STRONG_INLINE int predux_max<Packet4i>(const Packet4i& a)
 template<int Offset>
 struct palign_impl<Offset,Packet4f>
 {
-  EIGEN_STRONG_INLINE static void run(Packet4f& first, const Packet4f& second)
+  static EIGEN_STRONG_INLINE void run(Packet4f& first, const Packet4f& second)
   {
     if (Offset!=0)
       first = vec_sld(first, second, Offset*4);
@@ -497,7 +497,7 @@ struct palign_impl<Offset,Packet4f>
 template<int Offset>
 struct palign_impl<Offset,Packet4i>
 {
-  EIGEN_STRONG_INLINE static void run(Packet4i& first, const Packet4i& second)
+  static EIGEN_STRONG_INLINE void run(Packet4i& first, const Packet4i& second)
   {
     if (Offset!=0)
       first = vec_sld(first, second, Offset*4);

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

@@ -102,7 +102,7 @@ template<> EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(const std::complex<flo
   Packet2cf res;
   #if EIGEN_GNUC_AT_MOST(4,2)
   // workaround annoying "may be used uninitialized in this function" warning with gcc 4.2
-  res.v = _mm_loadl_pi(_mm_set1_ps(0.0f), (const __m64*)&from);
+  res.v = _mm_loadl_pi(_mm_set1_ps(0.0f), reinterpret_cast<const __m64*>(&from));
   #else
   res.v = _mm_loadl_pi(res.v, (const __m64*)&from);
   #endif
@@ -151,7 +151,7 @@ template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(const P
 template<int Offset>
 struct palign_impl<Offset,Packet2cf>
 {
-  EIGEN_STRONG_INLINE static void run(Packet2cf& first, const Packet2cf& second)
+  static EIGEN_STRONG_INLINE void run(Packet2cf& first, const Packet2cf& second)
   {
     if (Offset==1)
     {
@@ -350,7 +350,7 @@ template<> EIGEN_STRONG_INLINE std::complex<double> predux_mul<Packet1cd>(const
 template<int Offset>
 struct palign_impl<Offset,Packet1cd>
 {
-  EIGEN_STRONG_INLINE static void run(Packet1cd& /*first*/, const Packet1cd& /*second*/)
+  static EIGEN_STRONG_INLINE void run(Packet1cd& /*first*/, const Packet1cd& /*second*/)
   {
     // FIXME is it sure we never have to align a Packet1cd?
     // Even though a std::complex<double> has 16 bytes, it is not necessarily aligned on a 16 bytes boundary...

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

@@ -110,9 +110,18 @@ template<> struct unpacket_traits<Packet4f> { typedef float  type; enum {size=4}
 template<> struct unpacket_traits<Packet2d> { typedef double type; enum {size=2}; };
 template<> struct unpacket_traits<Packet4i> { typedef int    type; enum {size=4}; };
 
+#if defined(_MSC_VER) && (_MSC_VER==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
+// TODO: let's check whether there does not exist a better fix, like adding a pset0() function. (it crashed on pset1(0)).
+template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float&  from) { return _mm_set_ps(from,from,from,from); }
+template<> EIGEN_STRONG_INLINE Packet2d pset1<Packet2d>(const double& from) { return _mm_set_pd(from,from); }
+template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int&    from) { return _mm_set_epi32(from,from,from,from); }
+#else
 template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float&  from) { return _mm_set1_ps(from); }
 template<> EIGEN_STRONG_INLINE Packet2d pset1<Packet2d>(const double& from) { return _mm_set1_pd(from); }
 template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int&    from) { return _mm_set1_epi32(from); }
+#endif
 
 template<> EIGEN_STRONG_INLINE Packet4f plset<float>(const float& a) { return _mm_add_ps(pset1<Packet4f>(a), _mm_set_ps(3,2,1,0)); }
 template<> EIGEN_STRONG_INLINE Packet2d plset<double>(const double& a) { return _mm_add_pd(pset1<Packet2d>(a),_mm_set_pd(1,0)); }
@@ -282,7 +291,7 @@ template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from)
 
 template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float*   from)
 {
-  return vec4f_swizzle1(_mm_castpd_ps(_mm_load_sd((const double*)from)), 0, 0, 1, 1);
+  return vec4f_swizzle1(_mm_castpd_ps(_mm_load_sd(reinterpret_cast<const double*>(from))), 0, 0, 1, 1);
 }
 template<> EIGEN_STRONG_INLINE Packet2d ploaddup<Packet2d>(const double*  from)
 { return pset1<Packet2d>(from[0]); }
@@ -302,8 +311,8 @@ template<> EIGEN_STRONG_INLINE void pstoreu<double>(double* to, const Packet2d&
   _mm_storel_pd((to), from);
   _mm_storeh_pd((to+1), from);
 }
-template<> EIGEN_STRONG_INLINE void pstoreu<float>(float*  to, const Packet4f& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((double*)to, _mm_castps_pd(from)); }
-template<> EIGEN_STRONG_INLINE void pstoreu<int>(int*      to, const Packet4i& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((double*)to, _mm_castsi128_pd(from)); }
+template<> EIGEN_STRONG_INLINE void pstoreu<float>(float*  to, const Packet4f& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(reinterpret_cast<double*>(to), _mm_castps_pd(from)); }
+template<> EIGEN_STRONG_INLINE void pstoreu<int>(int*      to, const Packet4i& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(reinterpret_cast<double*>(to), _mm_castsi128_pd(from)); }
 
 // some compilers might be tempted to perform multiple moves instead of using a vector path.
 template<> EIGEN_STRONG_INLINE void pstore1<Packet4f>(float* to, const float& a)
@@ -541,7 +550,7 @@ template<> EIGEN_STRONG_INLINE int predux_max<Packet4i>(const Packet4i& a)
 template<int Offset>
 struct palign_impl<Offset,Packet4f>
 {
-  EIGEN_STRONG_INLINE static void run(Packet4f& first, const Packet4f& second)
+  static EIGEN_STRONG_INLINE void run(Packet4f& first, const Packet4f& second)
   {
     if (Offset!=0)
       first = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(second), _mm_castps_si128(first), Offset*4));
@@ -551,7 +560,7 @@ struct palign_impl<Offset,Packet4f>
 template<int Offset>
 struct palign_impl<Offset,Packet4i>
 {
-  EIGEN_STRONG_INLINE static void run(Packet4i& first, const Packet4i& second)
+  static EIGEN_STRONG_INLINE void run(Packet4i& first, const Packet4i& second)
   {
     if (Offset!=0)
       first = _mm_alignr_epi8(second,first, Offset*4);
@@ -561,7 +570,7 @@ struct palign_impl<Offset,Packet4i>
 template<int Offset>
 struct palign_impl<Offset,Packet2d>
 {
-  EIGEN_STRONG_INLINE static void run(Packet2d& first, const Packet2d& second)
+  static EIGEN_STRONG_INLINE void run(Packet2d& first, const Packet2d& second)
   {
     if (Offset==1)
       first = _mm_castsi128_pd(_mm_alignr_epi8(_mm_castpd_si128(second), _mm_castpd_si128(first), 8));
@@ -572,7 +581,7 @@ struct palign_impl<Offset,Packet2d>
 template<int Offset>
 struct palign_impl<Offset,Packet4f>
 {
-  EIGEN_STRONG_INLINE static void run(Packet4f& first, const Packet4f& second)
+  static EIGEN_STRONG_INLINE void run(Packet4f& first, const Packet4f& second)
   {
     if (Offset==1)
     {
@@ -595,7 +604,7 @@ struct palign_impl<Offset,Packet4f>
 template<int Offset>
 struct palign_impl<Offset,Packet4i>
 {
-  EIGEN_STRONG_INLINE static void run(Packet4i& first, const Packet4i& second)
+  static EIGEN_STRONG_INLINE void run(Packet4i& first, const Packet4i& second)
   {
     if (Offset==1)
     {
@@ -618,7 +627,7 @@ struct palign_impl<Offset,Packet4i>
 template<int Offset>
 struct palign_impl<Offset,Packet2d>
 {
-  EIGEN_STRONG_INLINE static void run(Packet2d& first, const Packet2d& second)
+  static EIGEN_STRONG_INLINE void run(Packet2d& first, const Packet2d& second)
   {
     if (Offset==1)
     {

Eigen/src/Core/products/CoeffBasedProduct.h

@@ -224,8 +224,8 @@ class CoeffBasedProduct
     { return reinterpret_cast<const LazyCoeffBasedProductType&>(*this).diagonal(index); }
 
   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;
 
     mutable PlainObject m_result;
 };
@@ -252,7 +252,7 @@ template<int UnrollingIndex, typename Lhs, typename Rhs, typename RetScalar>
 struct product_coeff_impl<DefaultTraversal, UnrollingIndex, Lhs, Rhs, RetScalar>
 {
   typedef typename Lhs::Index Index;
-  EIGEN_STRONG_INLINE static void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, RetScalar &res)
+  static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, RetScalar &res)
   {
     product_coeff_impl<DefaultTraversal, UnrollingIndex-1, Lhs, Rhs, RetScalar>::run(row, col, lhs, rhs, res);
     res += lhs.coeff(row, UnrollingIndex) * rhs.coeff(UnrollingIndex, col);
@@ -263,7 +263,7 @@ template<typename Lhs, typename Rhs, typename RetScalar>
 struct product_coeff_impl<DefaultTraversal, 0, Lhs, Rhs, RetScalar>
 {
   typedef typename Lhs::Index Index;
-  EIGEN_STRONG_INLINE static void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, RetScalar &res)
+  static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, RetScalar &res)
   {
     res = lhs.coeff(row, 0) * rhs.coeff(0, col);
   }
@@ -273,7 +273,7 @@ template<typename Lhs, typename Rhs, typename RetScalar>
 struct product_coeff_impl<DefaultTraversal, Dynamic, Lhs, Rhs, RetScalar>
 {
   typedef typename Lhs::Index Index;
-  EIGEN_STRONG_INLINE static void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, RetScalar& res)
+  static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, RetScalar& res)
   {
     eigen_assert(lhs.cols()>0 && "you are using a non initialized matrix");
     res = lhs.coeff(row, 0) * rhs.coeff(0, col);
@@ -291,7 +291,7 @@ struct product_coeff_vectorized_unroller
 {
   typedef typename Lhs::Index Index;
   enum { PacketSize = packet_traits<typename Lhs::Scalar>::size };
-  EIGEN_STRONG_INLINE static void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, typename Lhs::PacketScalar &pres)
+  static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, typename Lhs::PacketScalar &pres)
   {
     product_coeff_vectorized_unroller<UnrollingIndex-PacketSize, Lhs, Rhs, Packet>::run(row, col, lhs, rhs, pres);
     pres = padd(pres, pmul( lhs.template packet<Aligned>(row, UnrollingIndex) , rhs.template packet<Aligned>(UnrollingIndex, col) ));
@@ -302,7 +302,7 @@ template<typename Lhs, typename Rhs, typename Packet>
 struct product_coeff_vectorized_unroller<0, Lhs, Rhs, Packet>
 {
   typedef typename Lhs::Index Index;
-  EIGEN_STRONG_INLINE static void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, typename Lhs::PacketScalar &pres)
+  static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, typename Lhs::PacketScalar &pres)
   {
     pres = pmul(lhs.template packet<Aligned>(row, 0) , rhs.template packet<Aligned>(0, col));
   }
@@ -314,7 +314,7 @@ struct product_coeff_impl<InnerVectorizedTraversal, UnrollingIndex, Lhs, Rhs, Re
   typedef typename Lhs::PacketScalar Packet;
   typedef typename Lhs::Index Index;
   enum { PacketSize = packet_traits<typename Lhs::Scalar>::size };
-  EIGEN_STRONG_INLINE static void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, RetScalar &res)
+  static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, RetScalar &res)
   {
     Packet pres;
     product_coeff_vectorized_unroller<UnrollingIndex+1-PacketSize, Lhs, Rhs, Packet>::run(row, col, lhs, rhs, pres);
@@ -327,7 +327,7 @@ template<typename Lhs, typename Rhs, int LhsRows = Lhs::RowsAtCompileTime, int R
 struct product_coeff_vectorized_dyn_selector
 {
   typedef typename Lhs::Index Index;
-  EIGEN_STRONG_INLINE static void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, typename Lhs::Scalar &res)
+  static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, typename Lhs::Scalar &res)
   {
     res = lhs.row(row).transpose().cwiseProduct(rhs.col(col)).sum();
   }
@@ -339,7 +339,7 @@ template<typename Lhs, typename Rhs, int RhsCols>
 struct product_coeff_vectorized_dyn_selector<Lhs,Rhs,1,RhsCols>
 {
   typedef typename Lhs::Index Index;
-  EIGEN_STRONG_INLINE static void run(Index /*row*/, Index col, const Lhs& lhs, const Rhs& rhs, typename Lhs::Scalar &res)
+  static EIGEN_STRONG_INLINE void run(Index /*row*/, Index col, const Lhs& lhs, const Rhs& rhs, typename Lhs::Scalar &res)
   {
     res = lhs.transpose().cwiseProduct(rhs.col(col)).sum();
   }
@@ -349,7 +349,7 @@ template<typename Lhs, typename Rhs, int LhsRows>
 struct product_coeff_vectorized_dyn_selector<Lhs,Rhs,LhsRows,1>
 {
   typedef typename Lhs::Index Index;
-  EIGEN_STRONG_INLINE static void run(Index row, Index /*col*/, const Lhs& lhs, const Rhs& rhs, typename Lhs::Scalar &res)
+  static EIGEN_STRONG_INLINE void run(Index row, Index /*col*/, const Lhs& lhs, const Rhs& rhs, typename Lhs::Scalar &res)
   {
     res = lhs.row(row).transpose().cwiseProduct(rhs).sum();
   }
@@ -359,7 +359,7 @@ template<typename Lhs, typename Rhs>
 struct product_coeff_vectorized_dyn_selector<Lhs,Rhs,1,1>
 {
   typedef typename Lhs::Index Index;
-  EIGEN_STRONG_INLINE static void run(Index /*row*/, Index /*col*/, const Lhs& lhs, const Rhs& rhs, typename Lhs::Scalar &res)
+  static EIGEN_STRONG_INLINE void run(Index /*row*/, Index /*col*/, const Lhs& lhs, const Rhs& rhs, typename Lhs::Scalar &res)
   {
     res = lhs.transpose().cwiseProduct(rhs).sum();
   }
@@ -369,7 +369,7 @@ template<typename Lhs, typename Rhs, typename RetScalar>
 struct product_coeff_impl<InnerVectorizedTraversal, Dynamic, Lhs, Rhs, RetScalar>
 {
   typedef typename Lhs::Index Index;
-  EIGEN_STRONG_INLINE static void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, typename Lhs::Scalar &res)
+  static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, typename Lhs::Scalar &res)
   {
     product_coeff_vectorized_dyn_selector<Lhs,Rhs>::run(row, col, lhs, rhs, res);
   }
@@ -383,7 +383,7 @@ template<int UnrollingIndex, typename Lhs, typename Rhs, typename Packet, int Lo
 struct product_packet_impl<RowMajor, UnrollingIndex, Lhs, Rhs, Packet, LoadMode>
 {
   typedef typename Lhs::Index Index;
-  EIGEN_STRONG_INLINE static void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Packet &res)
+  static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Packet &res)
   {
     product_packet_impl<RowMajor, UnrollingIndex-1, Lhs, Rhs, Packet, LoadMode>::run(row, col, lhs, rhs, res);
     res =  pmadd(pset1<Packet>(lhs.coeff(row, UnrollingIndex)), rhs.template packet<LoadMode>(UnrollingIndex, col), res);
@@ -394,7 +394,7 @@ template<int UnrollingIndex, typename Lhs, typename Rhs, typename Packet, int Lo
 struct product_packet_impl<ColMajor, UnrollingIndex, Lhs, Rhs, Packet, LoadMode>
 {
   typedef typename Lhs::Index Index;
-  EIGEN_STRONG_INLINE static void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Packet &res)
+  static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Packet &res)
   {
     product_packet_impl<ColMajor, UnrollingIndex-1, Lhs, Rhs, Packet, LoadMode>::run(row, col, lhs, rhs, res);
     res =  pmadd(lhs.template packet<LoadMode>(row, UnrollingIndex), pset1<Packet>(rhs.coeff(UnrollingIndex, col)), res);
@@ -405,7 +405,7 @@ template<typename Lhs, typename Rhs, typename Packet, int LoadMode>
 struct product_packet_impl<RowMajor, 0, Lhs, Rhs, Packet, LoadMode>
 {
   typedef typename Lhs::Index Index;
-  EIGEN_STRONG_INLINE static void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Packet &res)
+  static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Packet &res)
   {
     res = pmul(pset1<Packet>(lhs.coeff(row, 0)),rhs.template packet<LoadMode>(0, col));
   }
@@ -415,7 +415,7 @@ template<typename Lhs, typename Rhs, typename Packet, int LoadMode>
 struct product_packet_impl<ColMajor, 0, Lhs, Rhs, Packet, LoadMode>
 {
   typedef typename Lhs::Index Index;
-  EIGEN_STRONG_INLINE static void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Packet &res)
+  static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Packet &res)
   {
     res = pmul(lhs.template packet<LoadMode>(row, 0), pset1<Packet>(rhs.coeff(0, col)));
   }
@@ -425,7 +425,7 @@ template<typename Lhs, typename Rhs, typename Packet, int LoadMode>
 struct product_packet_impl<RowMajor, Dynamic, Lhs, Rhs, Packet, LoadMode>
 {
   typedef typename Lhs::Index Index;
-  EIGEN_STRONG_INLINE static void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Packet& res)
+  static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Packet& res)
   {
     eigen_assert(lhs.cols()>0 && "you are using a non initialized matrix");
     res = pmul(pset1<Packet>(lhs.coeff(row, 0)),rhs.template packet<LoadMode>(0, col));
@@ -438,7 +438,7 @@ template<typename Lhs, typename Rhs, typename Packet, int LoadMode>
 struct product_packet_impl<ColMajor, Dynamic, Lhs, Rhs, Packet, LoadMode>
 {
   typedef typename Lhs::Index Index;
-  EIGEN_STRONG_INLINE static void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Packet& res)
+  static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Packet& res)
   {
     eigen_assert(lhs.cols()>0 && "you are using a non initialized matrix");
     res = pmul(lhs.template packet<LoadMode>(row, 0), pset1<Packet>(rhs.coeff(0, col)));

Eigen/src/Core/products/GeneralBlockPanelKernel.h

@@ -30,19 +30,18 @@ namespace internal {
 template<typename _LhsScalar, typename _RhsScalar, bool _ConjLhs=false, bool _ConjRhs=false>
 class gebp_traits;
 
+
+/** \internal \returns b if a<=0, and returns a otherwise. */
+inline std::ptrdiff_t manage_caching_sizes_helper(std::ptrdiff_t a, std::ptrdiff_t b)
+{
+  return a<=0 ? b : a;
+}
+
 /** \internal */
 inline void manage_caching_sizes(Action action, std::ptrdiff_t* l1=0, std::ptrdiff_t* l2=0)
 {
-  static std::ptrdiff_t m_l1CacheSize = 0;
-  static std::ptrdiff_t m_l2CacheSize = 0;
-  if(m_l1CacheSize==0)
-  {
-    m_l1CacheSize = queryL1CacheSize();
-    m_l2CacheSize = queryTopLevelCacheSize();
-
-    if(m_l1CacheSize<=0) m_l1CacheSize = 8 * 1024;
-    if(m_l2CacheSize<=0) m_l2CacheSize = 1 * 1024 * 1024;
-  }
+  static std::ptrdiff_t m_l1CacheSize = manage_caching_sizes_helper(queryL1CacheSize(),8 * 1024);
+  static std::ptrdiff_t m_l2CacheSize = manage_caching_sizes_helper(queryTopLevelCacheSize(),1*1024*1024);
 
   if(action==SetAction)
   {
@@ -536,7 +535,7 @@ struct gebp_kernel
     ResPacketSize = Traits::ResPacketSize
   };
 
-  EIGEN_FLATTEN_ATTRIB
+  EIGEN_DONT_INLINE EIGEN_FLATTEN_ATTRIB
   void operator()(ResScalar* res, Index resStride, const LhsScalar* blockA, const RhsScalar* blockB, Index rows, Index depth, Index cols, ResScalar alpha,
                   Index strideA=-1, Index strideB=-1, Index offsetA=0, Index offsetB=0, RhsScalar* unpackedB = 0)
   {
@@ -598,64 +597,64 @@ struct gebp_kernel
           if(nr==2)
           {
             LhsPacket A0, A1;
-            RhsPacket B0;
+            RhsPacket B_0;
             RhsPacket T0;
             
 EIGEN_ASM_COMMENT("mybegin2");
             traits.loadLhs(&blA[0*LhsProgress], A0);
             traits.loadLhs(&blA[1*LhsProgress], A1);
-            traits.loadRhs(&blB[0*RhsProgress], B0);
-            traits.madd(A0,B0,C0,T0);
-            traits.madd(A1,B0,C4,B0);
-            traits.loadRhs(&blB[1*RhsProgress], B0);
-            traits.madd(A0,B0,C1,T0);
-            traits.madd(A1,B0,C5,B0);
+            traits.loadRhs(&blB[0*RhsProgress], B_0);
+            traits.madd(A0,B_0,C0,T0);
+            traits.madd(A1,B_0,C4,B_0);
+            traits.loadRhs(&blB[1*RhsProgress], B_0);
+            traits.madd(A0,B_0,C1,T0);
+            traits.madd(A1,B_0,C5,B_0);
 
             traits.loadLhs(&blA[2*LhsProgress], A0);
             traits.loadLhs(&blA[3*LhsProgress], A1);
-            traits.loadRhs(&blB[2*RhsProgress], B0);
-            traits.madd(A0,B0,C0,T0);
-            traits.madd(A1,B0,C4,B0);
-            traits.loadRhs(&blB[3*RhsProgress], B0);
-            traits.madd(A0,B0,C1,T0);
-            traits.madd(A1,B0,C5,B0);
+            traits.loadRhs(&blB[2*RhsProgress], B_0);
+            traits.madd(A0,B_0,C0,T0);
+            traits.madd(A1,B_0,C4,B_0);
+            traits.loadRhs(&blB[3*RhsProgress], B_0);
+            traits.madd(A0,B_0,C1,T0);
+            traits.madd(A1,B_0,C5,B_0);
 
             traits.loadLhs(&blA[4*LhsProgress], A0);
             traits.loadLhs(&blA[5*LhsProgress], A1);
-            traits.loadRhs(&blB[4*RhsProgress], B0);
-            traits.madd(A0,B0,C0,T0);
-            traits.madd(A1,B0,C4,B0);
-            traits.loadRhs(&blB[5*RhsProgress], B0);
-            traits.madd(A0,B0,C1,T0);
-            traits.madd(A1,B0,C5,B0);
+            traits.loadRhs(&blB[4*RhsProgress], B_0);
+            traits.madd(A0,B_0,C0,T0);
+            traits.madd(A1,B_0,C4,B_0);
+            traits.loadRhs(&blB[5*RhsProgress], B_0);
+            traits.madd(A0,B_0,C1,T0);
+            traits.madd(A1,B_0,C5,B_0);
 
             traits.loadLhs(&blA[6*LhsProgress], A0);
             traits.loadLhs(&blA[7*LhsProgress], A1);
-            traits.loadRhs(&blB[6*RhsProgress], B0);
-            traits.madd(A0,B0,C0,T0);
-            traits.madd(A1,B0,C4,B0);
-            traits.loadRhs(&blB[7*RhsProgress], B0);
-            traits.madd(A0,B0,C1,T0);
-            traits.madd(A1,B0,C5,B0);
+            traits.loadRhs(&blB[6*RhsProgress], B_0);
+            traits.madd(A0,B_0,C0,T0);
+            traits.madd(A1,B_0,C4,B_0);
+            traits.loadRhs(&blB[7*RhsProgress], B_0);
+            traits.madd(A0,B_0,C1,T0);
+            traits.madd(A1,B_0,C5,B_0);
 EIGEN_ASM_COMMENT("myend");
           }
           else
           {
 EIGEN_ASM_COMMENT("mybegin4");
             LhsPacket A0, A1;
-            RhsPacket B0, B1, B2, B3;
+            RhsPacket B_0, B1, B2, B3;
             RhsPacket T0;
             
             traits.loadLhs(&blA[0*LhsProgress], A0);
             traits.loadLhs(&blA[1*LhsProgress], A1);
-            traits.loadRhs(&blB[0*RhsProgress], B0);
+            traits.loadRhs(&blB[0*RhsProgress], B_0);
             traits.loadRhs(&blB[1*RhsProgress], B1);
 
-            traits.madd(A0,B0,C0,T0);
+            traits.madd(A0,B_0,C0,T0);
             traits.loadRhs(&blB[2*RhsProgress], B2);
-            traits.madd(A1,B0,C4,B0);
+            traits.madd(A1,B_0,C4,B_0);
             traits.loadRhs(&blB[3*RhsProgress], B3);
-            traits.loadRhs(&blB[4*RhsProgress], B0);
+            traits.loadRhs(&blB[4*RhsProgress], B_0);
             traits.madd(A0,B1,C1,T0);
             traits.madd(A1,B1,C5,B1);
             traits.loadRhs(&blB[5*RhsProgress], B1);
@@ -667,9 +666,9 @@ EIGEN_ASM_COMMENT("mybegin4");
             traits.madd(A1,B3,C7,B3);
             traits.loadLhs(&blA[3*LhsProgress], A1);
             traits.loadRhs(&blB[7*RhsProgress], B3);
-            traits.madd(A0,B0,C0,T0);
-            traits.madd(A1,B0,C4,B0);
-            traits.loadRhs(&blB[8*RhsProgress], B0);
+            traits.madd(A0,B_0,C0,T0);
+            traits.madd(A1,B_0,C4,B_0);
+            traits.loadRhs(&blB[8*RhsProgress], B_0);
             traits.madd(A0,B1,C1,T0);
             traits.madd(A1,B1,C5,B1);
             traits.loadRhs(&blB[9*RhsProgress], B1);
@@ -682,9 +681,9 @@ EIGEN_ASM_COMMENT("mybegin4");
             traits.loadLhs(&blA[5*LhsProgress], A1);
             traits.loadRhs(&blB[11*RhsProgress], B3);
 
-            traits.madd(A0,B0,C0,T0);
-            traits.madd(A1,B0,C4,B0);
-            traits.loadRhs(&blB[12*RhsProgress], B0);
+            traits.madd(A0,B_0,C0,T0);
+            traits.madd(A1,B_0,C4,B_0);
+            traits.loadRhs(&blB[12*RhsProgress], B_0);
             traits.madd(A0,B1,C1,T0);
             traits.madd(A1,B1,C5,B1);
             traits.loadRhs(&blB[13*RhsProgress], B1);
@@ -696,8 +695,8 @@ EIGEN_ASM_COMMENT("mybegin4");
             traits.madd(A1,B3,C7,B3);
             traits.loadLhs(&blA[7*LhsProgress], A1);
             traits.loadRhs(&blB[15*RhsProgress], B3);
-            traits.madd(A0,B0,C0,T0);
-            traits.madd(A1,B0,C4,B0);
+            traits.madd(A0,B_0,C0,T0);
+            traits.madd(A1,B_0,C4,B_0);
             traits.madd(A0,B1,C1,T0);
             traits.madd(A1,B1,C5,B1);
             traits.madd(A0,B2,C2,T0);
@@ -715,32 +714,32 @@ EIGEN_ASM_COMMENT("mybegin4");
           if(nr==2)
           {
             LhsPacket A0, A1;
-            RhsPacket B0;
+            RhsPacket B_0;
             RhsPacket T0;
 
             traits.loadLhs(&blA[0*LhsProgress], A0);
             traits.loadLhs(&blA[1*LhsProgress], A1);
-            traits.loadRhs(&blB[0*RhsProgress], B0);
-            traits.madd(A0,B0,C0,T0);
-            traits.madd(A1,B0,C4,B0);
-            traits.loadRhs(&blB[1*RhsProgress], B0);
-            traits.madd(A0,B0,C1,T0);
-            traits.madd(A1,B0,C5,B0);
+            traits.loadRhs(&blB[0*RhsProgress], B_0);
+            traits.madd(A0,B_0,C0,T0);
+            traits.madd(A1,B_0,C4,B_0);
+            traits.loadRhs(&blB[1*RhsProgress], B_0);
+            traits.madd(A0,B_0,C1,T0);
+            traits.madd(A1,B_0,C5,B_0);
           }
           else
           {
             LhsPacket A0, A1;
-            RhsPacket B0, B1, B2, B3;
+            RhsPacket B_0, B1, B2, B3;
             RhsPacket T0;
 
             traits.loadLhs(&blA[0*LhsProgress], A0);
             traits.loadLhs(&blA[1*LhsProgress], A1);
-            traits.loadRhs(&blB[0*RhsProgress], B0);
+            traits.loadRhs(&blB[0*RhsProgress], B_0);
             traits.loadRhs(&blB[1*RhsProgress], B1);
 
-            traits.madd(A0,B0,C0,T0);
+            traits.madd(A0,B_0,C0,T0);
             traits.loadRhs(&blB[2*RhsProgress], B2);
-            traits.madd(A1,B0,C4,B0);
+            traits.madd(A1,B_0,C4,B_0);
             traits.loadRhs(&blB[3*RhsProgress], B3);
             traits.madd(A0,B1,C1,T0);
             traits.madd(A1,B1,C5,B1);
@@ -827,42 +826,42 @@ EIGEN_ASM_COMMENT("mybegin4");
           if(nr==2)
           {
             LhsPacket A0;
-            RhsPacket B0, B1;
+            RhsPacket B_0, B1;
 
             traits.loadLhs(&blA[0*LhsProgress], A0);
-            traits.loadRhs(&blB[0*RhsProgress], B0);
+            traits.loadRhs(&blB[0*RhsProgress], B_0);
             traits.loadRhs(&blB[1*RhsProgress], B1);
-            traits.madd(A0,B0,C0,B0);
-            traits.loadRhs(&blB[2*RhsProgress], B0);
+            traits.madd(A0,B_0,C0,B_0);
+            traits.loadRhs(&blB[2*RhsProgress], B_0);
             traits.madd(A0,B1,C1,B1);
             traits.loadLhs(&blA[1*LhsProgress], A0);
             traits.loadRhs(&blB[3*RhsProgress], B1);
-            traits.madd(A0,B0,C0,B0);
-            traits.loadRhs(&blB[4*RhsProgress], B0);
+            traits.madd(A0,B_0,C0,B_0);
+            traits.loadRhs(&blB[4*RhsProgress], B_0);
             traits.madd(A0,B1,C1,B1);
             traits.loadLhs(&blA[2*LhsProgress], A0);
             traits.loadRhs(&blB[5*RhsProgress], B1);
-            traits.madd(A0,B0,C0,B0);
-            traits.loadRhs(&blB[6*RhsProgress], B0);
+            traits.madd(A0,B_0,C0,B_0);
+            traits.loadRhs(&blB[6*RhsProgress], B_0);
             traits.madd(A0,B1,C1,B1);
             traits.loadLhs(&blA[3*LhsProgress], A0);
             traits.loadRhs(&blB[7*RhsProgress], B1);
-            traits.madd(A0,B0,C0,B0);
+            traits.madd(A0,B_0,C0,B_0);
             traits.madd(A0,B1,C1,B1);
           }
           else
           {
             LhsPacket A0;
-            RhsPacket B0, B1, B2, B3;
+            RhsPacket B_0, B1, B2, B3;
 
             traits.loadLhs(&blA[0*LhsProgress], A0);
-            traits.loadRhs(&blB[0*RhsProgress], B0);
+            traits.loadRhs(&blB[0*RhsProgress], B_0);
             traits.loadRhs(&blB[1*RhsProgress], B1);
 
-            traits.madd(A0,B0,C0,B0);
+            traits.madd(A0,B_0,C0,B_0);
             traits.loadRhs(&blB[2*RhsProgress], B2);
             traits.loadRhs(&blB[3*RhsProgress], B3);
-            traits.loadRhs(&blB[4*RhsProgress], B0);
+            traits.loadRhs(&blB[4*RhsProgress], B_0);
             traits.madd(A0,B1,C1,B1);
             traits.loadRhs(&blB[5*RhsProgress], B1);
             traits.madd(A0,B2,C2,B2);
@@ -870,8 +869,8 @@ EIGEN_ASM_COMMENT("mybegin4");
             traits.madd(A0,B3,C3,B3);
             traits.loadLhs(&blA[1*LhsProgress], A0);
             traits.loadRhs(&blB[7*RhsProgress], B3);
-            traits.madd(A0,B0,C0,B0);
-            traits.loadRhs(&blB[8*RhsProgress], B0);
+            traits.madd(A0,B_0,C0,B_0);
+            traits.loadRhs(&blB[8*RhsProgress], B_0);
             traits.madd(A0,B1,C1,B1);
             traits.loadRhs(&blB[9*RhsProgress], B1);
             traits.madd(A0,B2,C2,B2);
@@ -880,8 +879,8 @@ EIGEN_ASM_COMMENT("mybegin4");
             traits.loadLhs(&blA[2*LhsProgress], A0);
             traits.loadRhs(&blB[11*RhsProgress], B3);
 
-            traits.madd(A0,B0,C0,B0);
-            traits.loadRhs(&blB[12*RhsProgress], B0);
+            traits.madd(A0,B_0,C0,B_0);
+            traits.loadRhs(&blB[12*RhsProgress], B_0);
             traits.madd(A0,B1,C1,B1);
             traits.loadRhs(&blB[13*RhsProgress], B1);
             traits.madd(A0,B2,C2,B2);
@@ -890,7 +889,7 @@ EIGEN_ASM_COMMENT("mybegin4");
 
             traits.loadLhs(&blA[3*LhsProgress], A0);
             traits.loadRhs(&blB[15*RhsProgress], B3);
-            traits.madd(A0,B0,C0,B0);
+            traits.madd(A0,B_0,C0,B_0);
             traits.madd(A0,B1,C1,B1);
             traits.madd(A0,B2,C2,B2);
             traits.madd(A0,B3,C3,B3);
@@ -905,26 +904,26 @@ EIGEN_ASM_COMMENT("mybegin4");
           if(nr==2)
           {
             LhsPacket A0;
-            RhsPacket B0, B1;
+            RhsPacket B_0, B1;
 
             traits.loadLhs(&blA[0*LhsProgress], A0);
-            traits.loadRhs(&blB[0*RhsProgress], B0);
+            traits.loadRhs(&blB[0*RhsProgress], B_0);
             traits.loadRhs(&blB[1*RhsProgress], B1);
-            traits.madd(A0,B0,C0,B0);
+            traits.madd(A0,B_0,C0,B_0);
             traits.madd(A0,B1,C1,B1);
           }
           else
           {
             LhsPacket A0;
-            RhsPacket B0, B1, B2, B3;
+            RhsPacket B_0, B1, B2, B3;
 
             traits.loadLhs(&blA[0*LhsProgress], A0);
-            traits.loadRhs(&blB[0*RhsProgress], B0);
+            traits.loadRhs(&blB[0*RhsProgress], B_0);
             traits.loadRhs(&blB[1*RhsProgress], B1);
             traits.loadRhs(&blB[2*RhsProgress], B2);
             traits.loadRhs(&blB[3*RhsProgress], B3);
 
-            traits.madd(A0,B0,C0,B0);
+            traits.madd(A0,B_0,C0,B_0);
             traits.madd(A0,B1,C1,B1);
             traits.madd(A0,B2,C2,B2);
             traits.madd(A0,B3,C3,B3);
@@ -971,26 +970,26 @@ EIGEN_ASM_COMMENT("mybegin4");
           if(nr==2)
           {
             LhsScalar A0;
-            RhsScalar B0, B1;
+            RhsScalar B_0, B1;
 
             A0 = blA[k];
-            B0 = blB[0];
+            B_0 = blB[0];
             B1 = blB[1];
-            MADD(cj,A0,B0,C0,B0);
+            MADD(cj,A0,B_0,C0,B_0);
             MADD(cj,A0,B1,C1,B1);
           }
           else
           {
             LhsScalar A0;
-            RhsScalar B0, B1, B2, B3;
+            RhsScalar B_0, B1, B2, B3;
 
             A0 = blA[k];
-            B0 = blB[0];
+            B_0 = blB[0];
             B1 = blB[1];
             B2 = blB[2];
             B3 = blB[3];
 
-            MADD(cj,A0,B0,C0,B0);
+            MADD(cj,A0,B_0,C0,B_0);
             MADD(cj,A0,B1,C1,B1);
             MADD(cj,A0,B2,C2,B2);
             MADD(cj,A0,B3,C3,B3);
@@ -1027,14 +1026,14 @@ EIGEN_ASM_COMMENT("mybegin4");
         for(Index k=0; k<depth; k++)
         {
           LhsPacket A0, A1;
-          RhsPacket B0;
+          RhsPacket B_0;
           RhsPacket T0;
 
           traits.loadLhs(&blA[0*LhsProgress], A0);
           traits.loadLhs(&blA[1*LhsProgress], A1);
-          traits.loadRhs(&blB[0*RhsProgress], B0);
-          traits.madd(A0,B0,C0,T0);
-          traits.madd(A1,B0,C4,B0);
+          traits.loadRhs(&blB[0*RhsProgress], B_0);
+          traits.madd(A0,B_0,C0,T0);
+          traits.madd(A1,B_0,C4,B_0);
 
           blB += RhsProgress;
           blA += 2*LhsProgress;
@@ -1066,10 +1065,10 @@ EIGEN_ASM_COMMENT("mybegin4");
         for(Index k=0; k<depth; k++)
         {
           LhsPacket A0;
-          RhsPacket B0;
+          RhsPacket B_0;
           traits.loadLhs(blA, A0);
-          traits.loadRhs(blB, B0);
-          traits.madd(A0, B0, C0, B0);
+          traits.loadRhs(blB, B_0);
+          traits.madd(A0, B_0, C0, B_0);
           blB += RhsProgress;
           blA += LhsProgress;
         }
@@ -1091,8 +1090,8 @@ EIGEN_ASM_COMMENT("mybegin4");
         for(Index k=0; k<depth; k++)
         {
           LhsScalar A0 = blA[k];
-          RhsScalar B0 = blB[k];
-          MADD(cj, A0, B0, C0, B0);
+          RhsScalar B_0 = blB[k];
+          MADD(cj, A0, B_0, C0, B_0);
         }
         res[(j2+0)*resStride + i] += alpha*C0;
       }
@@ -1103,7 +1102,7 @@ EIGEN_ASM_COMMENT("mybegin4");
 #undef CJMADD
 
 // pack a block of the lhs
-// The travesal is as follow (mr==4):
+// The traversal is as follow (mr==4):
 //   0  4  8 12 ...
 //   1  5  9 13 ...
 //   2  6 10 14 ...
@@ -1119,11 +1118,15 @@ EIGEN_ASM_COMMENT("mybegin4");
 template<typename Scalar, typename Index, int Pack1, int Pack2, int StorageOrder, bool Conjugate, bool PanelMode>
 struct gemm_pack_lhs
 {
-  void operator()(Scalar* blockA, const Scalar* EIGEN_RESTRICT _lhs, Index lhsStride, Index depth, Index rows,
+  EIGEN_DONT_INLINE void operator()(Scalar* blockA, const Scalar* EIGEN_RESTRICT _lhs, Index lhsStride, Index depth, Index rows,
                   Index stride=0, Index offset=0)
   {
-//     enum { PacketSize = packet_traits<Scalar>::size };
+    typedef typename packet_traits<Scalar>::type Packet;
+    enum { PacketSize = packet_traits<Scalar>::size };
+
+    EIGEN_ASM_COMMENT("EIGEN PRODUCT PACK LHS");
     eigen_assert(((!PanelMode) && stride==0 && offset==0) || (PanelMode && stride>=depth && offset<=stride));
+    eigen_assert( (StorageOrder==RowMajor) || ((Pack1%PacketSize)==0 && Pack1<=4*PacketSize) );
     conj_if<NumTraits<Scalar>::IsComplex && Conjugate> cj;
     const_blas_data_mapper<Scalar, Index, StorageOrder> lhs(_lhs,lhsStride);
     Index count = 0;
@@ -1131,9 +1134,44 @@ struct gemm_pack_lhs
     for(Index i=0; i<peeled_mc; i+=Pack1)
     {
       if(PanelMode) count += Pack1 * offset;
-      for(Index k=0; k<depth; k++)
-        for(Index w=0; w<Pack1; w++)
-          blockA[count++] = cj(lhs(i+w, k));
+
+      if(StorageOrder==ColMajor)
+      {
+        for(Index k=0; k<depth; k++)
+        {
+          Packet A, B, C, D;
+          if(Pack1>=1*PacketSize) A = ploadu<Packet>(&lhs(i+0*PacketSize, k));
+          if(Pack1>=2*PacketSize) B = ploadu<Packet>(&lhs(i+1*PacketSize, k));
+          if(Pack1>=3*PacketSize) C = ploadu<Packet>(&lhs(i+2*PacketSize, k));
+          if(Pack1>=4*PacketSize) D = ploadu<Packet>(&lhs(i+3*PacketSize, k));
+          if(Pack1>=1*PacketSize) { pstore(blockA+count, cj.pconj(A)); count+=PacketSize; }
+          if(Pack1>=2*PacketSize) { pstore(blockA+count, cj.pconj(B)); count+=PacketSize; }
+          if(Pack1>=3*PacketSize) { pstore(blockA+count, cj.pconj(C)); count+=PacketSize; }
+          if(Pack1>=4*PacketSize) { pstore(blockA+count, cj.pconj(D)); count+=PacketSize; }
+        }
+      }
+      else
+      {
+        for(Index k=0; k<depth; k++)
+        {
+          // TODO add a vectorized transpose here
+          Index w=0;
+          for(; w<Pack1-3; w+=4)
+          {
+            Scalar a(cj(lhs(i+w+0, k))),
+                   b(cj(lhs(i+w+1, k))),
+                   c(cj(lhs(i+w+2, k))),
+                   d(cj(lhs(i+w+3, k)));
+            blockA[count++] = a;
+            blockA[count++] = b;
+            blockA[count++] = c;
+            blockA[count++] = d;
+          }
+          if(Pack1%4)
+            for(;w<Pack1;++w)
+              blockA[count++] = cj(lhs(i+w, k));
+        }
+      }
       if(PanelMode) count += Pack1 * (stride-offset-depth);
     }
     if(rows-peeled_mc>=Pack2)
@@ -1167,9 +1205,10 @@ struct gemm_pack_rhs<Scalar, Index, nr, ColMajor, Conjugate, PanelMode>
 {
   typedef typename packet_traits<Scalar>::type Packet;
   enum { PacketSize = packet_traits<Scalar>::size };
-  void operator()(Scalar* blockB, const Scalar* rhs, Index rhsStride, Index depth, Index cols,
+  EIGEN_DONT_INLINE void operator()(Scalar* blockB, const Scalar* rhs, Index rhsStride, Index depth, Index cols,
                   Index stride=0, Index offset=0)
   {
+    EIGEN_ASM_COMMENT("EIGEN PRODUCT PACK RHS COLMAJOR");
     eigen_assert(((!PanelMode) && stride==0 && offset==0) || (PanelMode && stride>=depth && offset<=stride));
     conj_if<NumTraits<Scalar>::IsComplex && Conjugate> cj;
     Index packet_cols = (cols/nr) * nr;
@@ -1214,9 +1253,10 @@ template<typename Scalar, typename Index, int nr, bool Conjugate, bool PanelMode
 struct gemm_pack_rhs<Scalar, Index, nr, RowMajor, Conjugate, PanelMode>
 {
   enum { PacketSize = packet_traits<Scalar>::size };
-  void operator()(Scalar* blockB, const Scalar* rhs, Index rhsStride, Index depth, Index cols,
+  EIGEN_DONT_INLINE void operator()(Scalar* blockB, const Scalar* rhs, Index rhsStride, Index depth, Index cols,
                   Index stride=0, Index offset=0)
   {
+    EIGEN_ASM_COMMENT("EIGEN PRODUCT PACK RHS ROWMAJOR");
     eigen_assert(((!PanelMode) && stride==0 && offset==0) || (PanelMode && stride>=depth && offset<=stride));
     conj_if<NumTraits<Scalar>::IsComplex && Conjugate> cj;
     Index packet_cols = (cols/nr) * nr;

Eigen/src/Core/products/GeneralMatrixMatrix.h

@@ -412,8 +412,8 @@ class GeneralProduct<Lhs, Rhs, GemmProduct>
     {
       eigen_assert(dst.rows()==m_lhs.rows() && dst.cols()==m_rhs.cols());
 
-      const ActualLhsType lhs = LhsBlasTraits::extract(m_lhs);
-      const ActualRhsType rhs = RhsBlasTraits::extract(m_rhs);
+      typename internal::add_const_on_value_type<ActualLhsType>::type lhs = LhsBlasTraits::extract(m_lhs);
+      typename internal::add_const_on_value_type<ActualRhsType>::type rhs = RhsBlasTraits::extract(m_rhs);
 
       Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(m_lhs)
                                  * RhsBlasTraits::extractScalarFactor(m_rhs);

Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h

@@ -42,14 +42,14 @@ struct tribb_kernel;
 template <typename Index,
           typename LhsScalar, int LhsStorageOrder, bool ConjugateLhs,
           typename RhsScalar, int RhsStorageOrder, bool ConjugateRhs,
-                              int ResStorageOrder, int  UpLo>
+                              int ResStorageOrder, int  UpLo, int Version = Specialized>
 struct general_matrix_matrix_triangular_product;
 
 // as usual if the result is row major => we transpose the product
 template <typename Index, typename LhsScalar, int LhsStorageOrder, bool ConjugateLhs,
-                          typename RhsScalar, int RhsStorageOrder, bool ConjugateRhs, int  UpLo>
-struct general_matrix_matrix_triangular_product<Index,LhsScalar,LhsStorageOrder,ConjugateLhs,RhsScalar,RhsStorageOrder,ConjugateRhs,RowMajor,UpLo>
-{  
+                          typename RhsScalar, int RhsStorageOrder, bool ConjugateRhs, int  UpLo, int Version>
+struct general_matrix_matrix_triangular_product<Index,LhsScalar,LhsStorageOrder,ConjugateLhs,RhsScalar,RhsStorageOrder,ConjugateRhs,RowMajor,UpLo,Version>
+{
   typedef typename scalar_product_traits<LhsScalar, RhsScalar>::ReturnType ResScalar;
   static EIGEN_STRONG_INLINE void run(Index size, Index depth,const LhsScalar* lhs, Index lhsStride,
                                       const RhsScalar* rhs, Index rhsStride, ResScalar* res, Index resStride, ResScalar alpha)
@@ -63,8 +63,8 @@ struct general_matrix_matrix_triangular_product<Index,LhsScalar,LhsStorageOrder,
 };
 
 template <typename Index, typename LhsScalar, int LhsStorageOrder, bool ConjugateLhs,
-                          typename RhsScalar, int RhsStorageOrder, bool ConjugateRhs, int  UpLo>
-struct general_matrix_matrix_triangular_product<Index,LhsScalar,LhsStorageOrder,ConjugateLhs,RhsScalar,RhsStorageOrder,ConjugateRhs,ColMajor,UpLo>
+                          typename RhsScalar, int RhsStorageOrder, bool ConjugateRhs, int  UpLo, int Version>
+struct general_matrix_matrix_triangular_product<Index,LhsScalar,LhsStorageOrder,ConjugateLhs,RhsScalar,RhsStorageOrder,ConjugateRhs,ColMajor,UpLo,Version>
 {
   typedef typename scalar_product_traits<LhsScalar, RhsScalar>::ReturnType ResScalar;
   static EIGEN_STRONG_INLINE void run(Index size, Index depth,const LhsScalar* _lhs, Index lhsStride,
@@ -201,13 +201,13 @@ TriangularView<MatrixType,UpLo>& TriangularView<MatrixType,UpLo>::assignProduct(
   typedef internal::blas_traits<Lhs> LhsBlasTraits;
   typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhs;
   typedef typename internal::remove_all<ActualLhs>::type _ActualLhs;
-  const ActualLhs actualLhs = LhsBlasTraits::extract(prod.lhs());
+  typename internal::add_const_on_value_type<ActualLhs>::type actualLhs = LhsBlasTraits::extract(prod.lhs());
   
   typedef typename internal::remove_all<typename ProductDerived::RhsNested>::type Rhs;
   typedef internal::blas_traits<Rhs> RhsBlasTraits;
   typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhs;
   typedef typename internal::remove_all<ActualRhs>::type _ActualRhs;
-  const ActualRhs actualRhs = RhsBlasTraits::extract(prod.rhs());
+  typename internal::add_const_on_value_type<ActualRhs>::type actualRhs = RhsBlasTraits::extract(prod.rhs());
 
   typename ProductDerived::Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(prod.lhs().derived()) * RhsBlasTraits::extractScalarFactor(prod.rhs().derived());
 

Eigen/src/Core/products/GeneralMatrixMatrixTriangular_MKL.h

@@ -0,0 +1,142 @@
+/*
+ Copyright (c) 2011, Intel Corporation. All rights reserved.
+
+ Redistribution and use in source and binary forms, with or without modification,
+ are permitted provided that the following conditions are met:
+
+ * Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+ * Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+ * Neither the name of Intel Corporation nor the names of its contributors may
+   be used to endorse or promote products derived from this software without
+   specific prior written permission.
+
+ THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+ WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+ DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+ (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+ LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+ ********************************************************************************
+ *   Content : Eigen bindings to Intel(R) MKL
+ *   Level 3 BLAS SYRK/HERK implementation.
+ ********************************************************************************
+*/
+
+#ifndef EIGEN_GENERAL_MATRIX_MATRIX_TRIANGULAR_MKL_H
+#define EIGEN_GENERAL_MATRIX_MATRIX_TRIANGULAR_MKL_H
+
+namespace internal {
+
+template <typename Index, typename Scalar, int AStorageOrder, bool ConjugateA, int ResStorageOrder, int  UpLo>
+struct general_matrix_matrix_rankupdate :
+       general_matrix_matrix_triangular_product<
+         Index,Scalar,AStorageOrder,ConjugateA,Scalar,AStorageOrder,ConjugateA,ResStorageOrder,UpLo,BuiltIn> {};
+
+
+// try to go to BLAS specialization
+#define EIGEN_MKL_RANKUPDATE_SPECIALIZE(Scalar) \
+template <typename Index, int LhsStorageOrder, bool ConjugateLhs, \
+                          int RhsStorageOrder, bool ConjugateRhs, int  UpLo> \
+struct general_matrix_matrix_triangular_product<Index,Scalar,LhsStorageOrder,ConjugateLhs, \
+               Scalar,RhsStorageOrder,ConjugateRhs,ColMajor,UpLo,Specialized> { \
+  static EIGEN_STRONG_INLINE void run(Index size, Index depth,const Scalar* lhs, Index lhsStride, \
+                          const Scalar* rhs, Index rhsStride, Scalar* res, Index resStride, Scalar alpha) \
+  { \
+    if (lhs==rhs) { \
+      general_matrix_matrix_rankupdate<Index,Scalar,LhsStorageOrder,ConjugateLhs,ColMajor,UpLo> \
+      ::run(size,depth,lhs,lhsStride,rhs,rhsStride,res,resStride,alpha); \
+    } else { \
+      general_matrix_matrix_triangular_product<Index, \
+        Scalar, LhsStorageOrder, ConjugateLhs, \
+        Scalar, RhsStorageOrder, ConjugateRhs, \
+        ColMajor, UpLo, BuiltIn> \
+      ::run(size,depth,lhs,lhsStride,rhs,rhsStride,res,resStride,alpha); \
+    } \
+  } \
+};
+
+EIGEN_MKL_RANKUPDATE_SPECIALIZE(double)
+//EIGEN_MKL_RANKUPDATE_SPECIALIZE(dcomplex)
+EIGEN_MKL_RANKUPDATE_SPECIALIZE(float)
+//EIGEN_MKL_RANKUPDATE_SPECIALIZE(scomplex)
+
+// SYRK for float/double
+#define EIGEN_MKL_RANKUPDATE_R(EIGTYPE, MKLTYPE, MKLFUNC) \
+template <typename Index, int AStorageOrder, bool ConjugateA, int  UpLo> \
+struct general_matrix_matrix_rankupdate<Index,EIGTYPE,AStorageOrder,ConjugateA,ColMajor,UpLo> { \
+  enum { \
+    IsLower = (UpLo&Lower) == Lower, \
+    LowUp = IsLower ? Lower : Upper, \
+    conjA = ((AStorageOrder==ColMajor) && ConjugateA) ? 1 : 0 \
+  }; \
+  static EIGEN_STRONG_INLINE void run(Index size, Index depth,const EIGTYPE* lhs, Index lhsStride, \
+                          const EIGTYPE* rhs, Index rhsStride, EIGTYPE* res, Index resStride, EIGTYPE alpha) \
+  { \
+  /* typedef Matrix<EIGTYPE, Dynamic, Dynamic, RhsStorageOrder> MatrixRhs;*/ \
+\
+   MKL_INT lda=lhsStride, ldc=resStride, n=size, k=depth; \
+   char uplo=(IsLower) ? 'L' : 'U', trans=(AStorageOrder==RowMajor) ? 'T':'N'; \
+   MKLTYPE alpha_, beta_; \
+\
+/* Set alpha_ & beta_ */ \
+   assign_scalar_eig2mkl<MKLTYPE, EIGTYPE>(alpha_, alpha); \
+   assign_scalar_eig2mkl<MKLTYPE, EIGTYPE>(beta_, EIGTYPE(1)); \
+   MKLFUNC(&uplo, &trans, &n, &k, &alpha_, lhs, &lda, &beta_, res, &ldc); \
+  } \
+};
+
+// HERK for complex data
+#define EIGEN_MKL_RANKUPDATE_C(EIGTYPE, MKLTYPE, RTYPE, MKLFUNC) \
+template <typename Index, int AStorageOrder, bool ConjugateA, int  UpLo> \
+struct general_matrix_matrix_rankupdate<Index,EIGTYPE,AStorageOrder,ConjugateA,ColMajor,UpLo> { \
+  enum { \
+    IsLower = (UpLo&Lower) == Lower, \
+    LowUp = IsLower ? Lower : Upper, \
+    conjA = (((AStorageOrder==ColMajor) && ConjugateA) || ((AStorageOrder==RowMajor) && !ConjugateA)) ? 1 : 0 \
+  }; \
+  static EIGEN_STRONG_INLINE void run(Index size, Index depth,const EIGTYPE* lhs, Index lhsStride, \
+                          const EIGTYPE* rhs, Index rhsStride, EIGTYPE* res, Index resStride, EIGTYPE alpha) \
+  { \
+   typedef Matrix<EIGTYPE, Dynamic, Dynamic, AStorageOrder> MatrixType; \
+\
+   MKL_INT lda=lhsStride, ldc=resStride, n=size, k=depth; \
+   char uplo=(IsLower) ? 'L' : 'U', trans=(AStorageOrder==RowMajor) ? 'C':'N'; \
+   RTYPE alpha_, beta_; \
+   const EIGTYPE* a_ptr; \
+\
+/* Set alpha_ & beta_ */ \
+/*   assign_scalar_eig2mkl<MKLTYPE, EIGTYPE>(alpha_, alpha); */\
+/*   assign_scalar_eig2mkl<MKLTYPE, EIGTYPE>(beta_, EIGTYPE(1));*/ \
+   alpha_ = alpha.real(); \
+   beta_ = 1.0; \
+/* Copy with conjugation in some cases*/ \
+   MatrixType a; \
+   if (conjA) { \
+     Map<const MatrixType, 0, OuterStride<> > mapA(lhs,n,k,OuterStride<>(lhsStride)); \
+     a = mapA.conjugate(); \
+     lda = a.outerStride(); \
+     a_ptr = a.data(); \
+   } else a_ptr=lhs; \
+   MKLFUNC(&uplo, &trans, &n, &k, &alpha_, (MKLTYPE*)a_ptr, &lda, &beta_, (MKLTYPE*)res, &ldc); \
+  } \
+};
+
+
+EIGEN_MKL_RANKUPDATE_R(double, double, dsyrk)
+EIGEN_MKL_RANKUPDATE_R(float,  float,  ssyrk)
+
+//EIGEN_MKL_RANKUPDATE_C(dcomplex, MKL_Complex16, double, zherk)
+//EIGEN_MKL_RANKUPDATE_C(scomplex, MKL_Complex8,  double, cherk)
+
+
+} // end namespace internal
+
+#endif // EIGEN_GENERAL_MATRIX_MATRIX_TRIANGULAR_MKL_H

Eigen/src/Core/products/GeneralMatrixMatrix_MKL.h

@@ -0,0 +1,114 @@
+/*
+ Copyright (c) 2011, Intel Corporation. All rights reserved.
+
+ Redistribution and use in source and binary forms, with or without modification,
+ are permitted provided that the following conditions are met:
+
+ * Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+ * Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+ * Neither the name of Intel Corporation nor the names of its contributors may
+   be used to endorse or promote products derived from this software without
+   specific prior written permission.
+
+ THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+ WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+ DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+ (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+ LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+ ********************************************************************************
+ *   Content : Eigen bindings to Intel(R) MKL
+ *   General matrix-matrix product functionality based on ?GEMM.
+ ********************************************************************************
+*/
+
+#ifndef EIGEN_GENERAL_MATRIX_MATRIX_MKL_H
+#define EIGEN_GENERAL_MATRIX_MATRIX_MKL_H
+
+namespace internal {
+
+/**********************************************************************
+* This file implements general matrix-matrix multiplication using BLAS
+* gemm function via partial specialization of
+* general_matrix_matrix_product::run(..) method for float, double,
+* std::complex<float> and std::complex<double> types
+**********************************************************************/
+
+// gemm specialization
+
+#define GEMM_SPECIALIZATION(EIGTYPE, EIGPREFIX, MKLTYPE, MKLPREFIX) \
+template< \
+  typename Index, \
+  int LhsStorageOrder, bool ConjugateLhs, \
+  int RhsStorageOrder, bool ConjugateRhs> \
+struct general_matrix_matrix_product<Index,EIGTYPE,LhsStorageOrder,ConjugateLhs,EIGTYPE,RhsStorageOrder,ConjugateRhs,ColMajor> \
+{ \
+static void run(Index rows, Index cols, Index depth, \
+  const EIGTYPE* _lhs, Index lhsStride, \
+  const EIGTYPE* _rhs, Index rhsStride, \
+  EIGTYPE* res, Index resStride, \
+  EIGTYPE alpha, \
+  level3_blocking<EIGTYPE, EIGTYPE>& blocking, \
+  GemmParallelInfo<Index>* info = 0) \
+{ \
+  using std::conj; \
+\
+  char transa, transb; \
+  MKL_INT m, n, k, lda, ldb, ldc; \
+  const EIGTYPE *a, *b; \
+  MKLTYPE alpha_, beta_; \
+  MatrixX##EIGPREFIX a_tmp, b_tmp; \
+  EIGTYPE myone(1);\
+\
+/* Set transpose options */ \
+  transa = (LhsStorageOrder==RowMajor) ? ((ConjugateLhs) ? 'C' : 'T') : 'N'; \
+  transb = (RhsStorageOrder==RowMajor) ? ((ConjugateRhs) ? 'C' : 'T') : 'N'; \
+\
+/* Set m, n, k */ \
+  m = (MKL_INT)rows;  \
+  n = (MKL_INT)cols;  \
+  k = (MKL_INT)depth; \
+\
+/* Set alpha_ & beta_ */ \
+  assign_scalar_eig2mkl(alpha_, alpha); \
+  assign_scalar_eig2mkl(beta_, myone); \
+\
+/* Set lda, ldb, ldc */ \
+  lda = (MKL_INT)lhsStride; \
+  ldb = (MKL_INT)rhsStride; \
+  ldc = (MKL_INT)resStride; \
+\
+/* Set a, b, c */ \
+  if ((LhsStorageOrder==ColMajor) && (ConjugateLhs)) { \
+    Map<const MatrixX##EIGPREFIX, 0, OuterStride<> > lhs(_lhs,m,k,OuterStride<>(lhsStride)); \
+    a_tmp = lhs.conjugate(); \
+    a = a_tmp.data(); \
+    lda = a_tmp.outerStride(); \
+  } else a = _lhs; \
+\
+  if ((RhsStorageOrder==ColMajor) && (ConjugateRhs)) { \
+    Map<const MatrixX##EIGPREFIX, 0, OuterStride<> > rhs(_rhs,k,n,OuterStride<>(rhsStride)); \
+    b_tmp = rhs.conjugate(); \
+    b = b_tmp.data(); \
+    ldb = b_tmp.outerStride(); \
+  } else b = _rhs; \
+\
+  MKLPREFIX##gemm(&transa, &transb, &m, &n, &k, &alpha_, (const MKLTYPE*)a, &lda, (const MKLTYPE*)b, &ldb, &beta_, (MKLTYPE*)res, &ldc); \
+}};
+
+GEMM_SPECIALIZATION(double,   d,  double,        d)
+GEMM_SPECIALIZATION(float,    f,  float,         s)
+GEMM_SPECIALIZATION(dcomplex, cd, MKL_Complex16, z)
+GEMM_SPECIALIZATION(scomplex, cf, MKL_Complex8,  c)
+
+} //end of namespase
+
+#endif // EIGEN_GENERAL_MATRIX_MATRIX_MKL_H

Eigen/src/Core/products/GeneralMatrixVector.h

@@ -40,8 +40,8 @@ namespace internal {
  *  |cplx |real |cplx | invalid, the caller has to do tmp: = A * B; C += alpha*tmp
  *  |cplx |real |real | optimal case, vectorization possible via real-cplx mul
  */
-template<typename Index, typename LhsScalar, bool ConjugateLhs, typename RhsScalar, bool ConjugateRhs>
-struct general_matrix_vector_product<Index,LhsScalar,ColMajor,ConjugateLhs,RhsScalar,ConjugateRhs>
+template<typename Index, typename LhsScalar, bool ConjugateLhs, typename RhsScalar, bool ConjugateRhs, int Version>
+struct general_matrix_vector_product<Index,LhsScalar,ColMajor,ConjugateLhs,RhsScalar,ConjugateRhs,Version>
 {
 typedef typename scalar_product_traits<LhsScalar, RhsScalar>::ReturnType ResScalar;
 
@@ -99,7 +99,7 @@ EIGEN_DONT_INLINE static void run(
   
   // How many coeffs of the result do we have to skip to be aligned.
   // Here we assume data are at least aligned on the base scalar type.
-  Index alignedStart = first_aligned(res,size);
+  Index alignedStart = internal::first_aligned(res,size);
   Index alignedSize = ResPacketSize>1 ? alignedStart + ((size-alignedStart) & ~ResPacketAlignedMask) : 0;
   const Index peeledSize  = peels>1 ? alignedStart + ((alignedSize-alignedStart) & ~PeelAlignedMask) : alignedStart;
 
@@ -109,7 +109,7 @@ EIGEN_DONT_INLINE static void run(
                        : FirstAligned;
 
   // we cannot assume the first element is aligned because of sub-matrices
-  const Index lhsAlignmentOffset = first_aligned(lhs,size);
+  const Index lhsAlignmentOffset = internal::first_aligned(lhs,size);
 
   // find how many columns do we have to skip to be aligned with the result (if possible)
   Index skipColumns = 0;
@@ -296,8 +296,8 @@ EIGEN_DONT_INLINE static void run(
  *  - alpha is always a complex (or converted to a complex)
  *  - no vectorization
  */
-template<typename Index, typename LhsScalar, bool ConjugateLhs, typename RhsScalar, bool ConjugateRhs>
-struct general_matrix_vector_product<Index,LhsScalar,RowMajor,ConjugateLhs,RhsScalar,ConjugateRhs>
+template<typename Index, typename LhsScalar, bool ConjugateLhs, typename RhsScalar, bool ConjugateRhs, int Version>
+struct general_matrix_vector_product<Index,LhsScalar,RowMajor,ConjugateLhs,RhsScalar,ConjugateRhs,Version>
 {
 typedef typename scalar_product_traits<LhsScalar, RhsScalar>::ReturnType ResScalar;
 
@@ -351,7 +351,7 @@ EIGEN_DONT_INLINE static void run(
   // How many coeffs of the result do we have to skip to be aligned.
   // Here we assume data are at least aligned on the base scalar type
   // if that's not the case then vectorization is discarded, see below.
-  Index alignedStart = first_aligned(rhs, depth);
+  Index alignedStart = internal::first_aligned(rhs, depth);
   Index alignedSize = RhsPacketSize>1 ? alignedStart + ((depth-alignedStart) & ~RhsPacketAlignedMask) : 0;
   const Index peeledSize  = peels>1 ? alignedStart + ((alignedSize-alignedStart) & ~PeelAlignedMask) : alignedStart;
 
@@ -361,7 +361,7 @@ EIGEN_DONT_INLINE static void run(
                          : FirstAligned;
 
   // we cannot assume the first element is aligned because of sub-matrices
-  const Index lhsAlignmentOffset = first_aligned(lhs,depth);
+  const Index lhsAlignmentOffset = internal::first_aligned(lhs,depth);
 
   // find how many rows do we have to skip to be aligned with rhs (if possible)
   Index skipRows = 0;

Eigen/src/Core/products/GeneralMatrixVector_MKL.h

@@ -0,0 +1,127 @@
+/*
+ Copyright (c) 2011, Intel Corporation. All rights reserved.
+
+ Redistribution and use in source and binary forms, with or without modification,
+ are permitted provided that the following conditions are met:
+
+ * Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+ * Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+ * Neither the name of Intel Corporation nor the names of its contributors may
+   be used to endorse or promote products derived from this software without
+   specific prior written permission.
+
+ THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+ WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+ DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+ (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+ LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+ ********************************************************************************
+ *   Content : Eigen bindings to Intel(R) MKL
+ *   General matrix-vector product functionality based on ?GEMV.
+ ********************************************************************************
+*/
+
+#ifndef EIGEN_GENERAL_MATRIX_VECTOR_MKL_H
+#define EIGEN_GENERAL_MATRIX_VECTOR_MKL_H
+
+namespace internal {
+
+/**********************************************************************
+* This file implements general matrix-vector multiplication using BLAS
+* gemv function via partial specialization of
+* general_matrix_vector_product::run(..) method for float, double,
+* std::complex<float> and std::complex<double> types
+**********************************************************************/
+
+// gemv specialization
+
+template<typename Index, typename LhsScalar, int LhsStorageOrder, bool ConjugateLhs, typename RhsScalar, bool ConjugateRhs>
+struct general_matrix_vector_product_gemv :
+  general_matrix_vector_product<Index,LhsScalar,LhsStorageOrder,ConjugateLhs,RhsScalar,ConjugateRhs,BuiltIn> {};
+
+#define EIGEN_MKL_GEMV_SPECIALIZE(Scalar) \
+template<typename Index, bool ConjugateLhs, bool ConjugateRhs> \
+struct general_matrix_vector_product<Index,Scalar,ColMajor,ConjugateLhs,Scalar,ConjugateRhs,Specialized> { \
+static EIGEN_DONT_INLINE void run( \
+  Index rows, Index cols, \
+  const Scalar* lhs, Index lhsStride, \
+  const Scalar* rhs, Index rhsIncr, \
+  Scalar* res, Index resIncr, Scalar alpha) \
+{ \
+  if (ConjugateLhs) { \
+    general_matrix_vector_product<Index,Scalar,ColMajor,ConjugateLhs,Scalar,ConjugateRhs,BuiltIn>::run( \
+      rows, cols, lhs, lhsStride, rhs, rhsIncr, res, resIncr, alpha); \
+  } else { \
+    general_matrix_vector_product_gemv<Index,Scalar,ColMajor,ConjugateLhs,Scalar,ConjugateRhs>::run( \
+      rows, cols, lhs, lhsStride, rhs, rhsIncr, res, resIncr, alpha); \
+  } \
+} \
+}; \
+template<typename Index, bool ConjugateLhs, bool ConjugateRhs> \
+struct general_matrix_vector_product<Index,Scalar,RowMajor,ConjugateLhs,Scalar,ConjugateRhs,Specialized> { \
+static EIGEN_DONT_INLINE void run( \
+  Index rows, Index cols, \
+  const Scalar* lhs, Index lhsStride, \
+  const Scalar* rhs, Index rhsIncr, \
+  Scalar* res, Index resIncr, Scalar alpha) \
+{ \
+    general_matrix_vector_product_gemv<Index,Scalar,RowMajor,ConjugateLhs,Scalar,ConjugateRhs>::run( \
+      rows, cols, lhs, lhsStride, rhs, rhsIncr, res, resIncr, alpha); \
+} \
+}; \
+
+EIGEN_MKL_GEMV_SPECIALIZE(double)
+EIGEN_MKL_GEMV_SPECIALIZE(float)
+EIGEN_MKL_GEMV_SPECIALIZE(dcomplex)
+EIGEN_MKL_GEMV_SPECIALIZE(scomplex)
+
+#define EIGEN_MKL_GEMV_SPECIALIZATION(EIGTYPE,MKLTYPE,MKLPREFIX) \
+template<typename Index, int LhsStorageOrder, bool ConjugateLhs, bool ConjugateRhs> \
+struct general_matrix_vector_product_gemv<Index,EIGTYPE,LhsStorageOrder,ConjugateLhs,EIGTYPE,ConjugateRhs> \
+{ \
+typedef Matrix<EIGTYPE,Dynamic,1,ColMajor> GEMVVector;\
+\
+static EIGEN_DONT_INLINE void run( \
+  Index rows, Index cols, \
+  const EIGTYPE* lhs, Index lhsStride, \
+  const EIGTYPE* rhs, Index rhsIncr, \
+  EIGTYPE* res, Index resIncr, EIGTYPE alpha) \
+{ \
+  MKL_INT m=rows, n=cols, lda=lhsStride, incx=rhsIncr, incy=resIncr; \
+  MKLTYPE alpha_, beta_; \
+  const EIGTYPE *x_ptr, myone(1); \
+  char trans=(LhsStorageOrder==ColMajor) ? 'N' : (ConjugateLhs) ? 'C' : 'T'; \
+  if (LhsStorageOrder==RowMajor) { \
+    m=cols; \
+    n=rows; \
+  }\
+  assign_scalar_eig2mkl(alpha_, alpha); \
+  assign_scalar_eig2mkl(beta_, myone); \
+  GEMVVector x_tmp; \
+  if (ConjugateRhs) { \
+    Map<const GEMVVector, 0, InnerStride<> > map_x(rhs,cols,1,InnerStride<>(incx)); \
+    x_tmp=map_x.conjugate(); \
+    x_ptr=x_tmp.data(); \
+    incx=1; \
+  } else x_ptr=rhs; \
+  MKLPREFIX##gemv(&trans, &m, &n, &alpha_, (const MKLTYPE*)lhs, &lda, (const MKLTYPE*)x_ptr, &incx, &beta_, (MKLTYPE*)res, &incy); \
+}\
+};
+
+EIGEN_MKL_GEMV_SPECIALIZATION(double,   double,        d)
+EIGEN_MKL_GEMV_SPECIALIZATION(float,    float,         s)
+EIGEN_MKL_GEMV_SPECIALIZATION(dcomplex, MKL_Complex16, z)
+EIGEN_MKL_GEMV_SPECIALIZATION(scomplex, MKL_Complex8,  c)
+
+} //end of namespase
+
+#endif // EIGEN_GENERAL_MATRIX_VECTOR_MKL_H

Eigen/src/Core/products/Parallelizer.h

@@ -85,7 +85,9 @@ template<typename Index> struct GemmParallelInfo
 template<bool Condition, typename Functor, typename Index>
 void parallelize_gemm(const Functor& func, Index rows, Index cols, bool transpose)
 {
-#ifndef EIGEN_HAS_OPENMP
+  // TODO when EIGEN_USE_BLAS is defined,
+  // we should still enable OMP for other scalar types
+#if !(defined (EIGEN_HAS_OPENMP)) || defined (EIGEN_USE_BLAS)
   // FIXME the transpose variable is only needed to properly split
   // the matrix product when multithreading is enabled. This is a temporary
   // fix to support row-major destination matrices. This whole

Eigen/src/Core/products/SelfadjointMatrixMatrix.h

@@ -400,8 +400,8 @@ struct SelfadjointProductMatrix<Lhs,LhsMode,false,Rhs,RhsMode,false>
   {
     eigen_assert(dst.rows()==m_lhs.rows() && dst.cols()==m_rhs.cols());
 
-    const ActualLhsType lhs = LhsBlasTraits::extract(m_lhs);
-    const ActualRhsType rhs = RhsBlasTraits::extract(m_rhs);
+    typename internal::add_const_on_value_type<ActualLhsType>::type lhs = LhsBlasTraits::extract(m_lhs);
+    typename internal::add_const_on_value_type<ActualRhsType>::type rhs = RhsBlasTraits::extract(m_rhs);
 
     Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(m_lhs)
                                * RhsBlasTraits::extractScalarFactor(m_rhs);

Eigen/src/Core/products/SelfadjointMatrixMatrix_MKL.h

@@ -0,0 +1,291 @@
+/*
+ Copyright (c) 2011, Intel Corporation. All rights reserved.
+
+ Redistribution and use in source and binary forms, with or without modification,
+ are permitted provided that the following conditions are met:
+
+ * Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+ * Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+ * Neither the name of Intel Corporation nor the names of its contributors may
+   be used to endorse or promote products derived from this software without
+   specific prior written permission.
+
+ THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+ WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+ DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+ (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+ LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+ ********************************************************************************
+ *   Content : Eigen bindings to Intel(R) MKL
+ *   Self adjoint matrix * matrix product functionality based on ?SYMM/?HEMM.
+ ********************************************************************************
+*/
+
+#ifndef EIGEN_SELFADJOINT_MATRIX_MATRIX_MKL_H
+#define EIGEN_SELFADJOINT_MATRIX_MATRIX_MKL_H
+
+namespace internal {
+
+
+/* Optimized selfadjoint matrix * matrix (?SYMM/?HEMM) product */
+
+#define EIGEN_MKL_SYMM_L(EIGTYPE, MKLTYPE, EIGPREFIX, MKLPREFIX) \
+template <typename Index, \
+          int LhsStorageOrder, bool ConjugateLhs, \
+          int RhsStorageOrder, bool ConjugateRhs> \
+struct product_selfadjoint_matrix<EIGTYPE,Index,LhsStorageOrder,true,ConjugateLhs,RhsStorageOrder,false,ConjugateRhs,ColMajor> \
+{\
+\
+  static EIGEN_DONT_INLINE void run( \
+    Index rows, Index cols, \
+    const EIGTYPE* _lhs, Index lhsStride, \
+    const EIGTYPE* _rhs, Index rhsStride, \
+    EIGTYPE* res,        Index resStride, \
+    EIGTYPE alpha) \
+  { \
+    char side='L', uplo='L'; \
+    MKL_INT m, n, lda, ldb, ldc; \
+    const EIGTYPE *a, *b; \
+    MKLTYPE alpha_, beta_; \
+    MatrixX##EIGPREFIX b_tmp; \
+    EIGTYPE myone(1);\
+\
+/* Set transpose options */ \
+/* Set m, n, k */ \
+    m = (MKL_INT)rows;  \
+    n = (MKL_INT)cols;  \
+\
+/* Set alpha_ & beta_ */ \
+    assign_scalar_eig2mkl(alpha_, alpha); \
+    assign_scalar_eig2mkl(beta_, myone); \
+\
+/* Set lda, ldb, ldc */ \
+    lda = (MKL_INT)lhsStride; \
+    ldb = (MKL_INT)rhsStride; \
+    ldc = (MKL_INT)resStride; \
+\
+/* Set a, b, c */ \
+    if (LhsStorageOrder==RowMajor) uplo='U'; \
+    a = _lhs; \
+\
+    if (RhsStorageOrder==RowMajor) { \
+      Map<const MatrixX##EIGPREFIX, 0, OuterStride<> > rhs(_rhs,n,m,OuterStride<>(rhsStride)); \
+      b_tmp = rhs.adjoint(); \
+      b = b_tmp.data(); \
+      ldb = b_tmp.outerStride(); \
+    } else b = _rhs; \
+\
+    MKLPREFIX##symm(&side, &uplo, &m, &n, &alpha_, (const MKLTYPE*)a, &lda, (const MKLTYPE*)b, &ldb, &beta_, (MKLTYPE*)res, &ldc); \
+\
+  } \
+};
+
+
+#define EIGEN_MKL_HEMM_L(EIGTYPE, MKLTYPE, EIGPREFIX, MKLPREFIX) \
+template <typename Index, \
+          int LhsStorageOrder, bool ConjugateLhs, \
+          int RhsStorageOrder, bool ConjugateRhs> \
+struct product_selfadjoint_matrix<EIGTYPE,Index,LhsStorageOrder,true,ConjugateLhs,RhsStorageOrder,false,ConjugateRhs,ColMajor> \
+{\
+  static EIGEN_DONT_INLINE void run( \
+    Index rows, Index cols, \
+    const EIGTYPE* _lhs, Index lhsStride, \
+    const EIGTYPE* _rhs, Index rhsStride, \
+    EIGTYPE* res,        Index resStride, \
+    EIGTYPE alpha) \
+  { \
+    char side='L', uplo='L'; \
+    MKL_INT m, n, lda, ldb, ldc; \
+    const EIGTYPE *a, *b; \
+    MKLTYPE alpha_, beta_; \
+    MatrixX##EIGPREFIX b_tmp; \
+    Matrix<EIGTYPE, Dynamic, Dynamic, LhsStorageOrder> a_tmp; \
+    EIGTYPE myone(1); \
+\
+/* Set transpose options */ \
+/* Set m, n, k */ \
+    m = (MKL_INT)rows; \
+    n = (MKL_INT)cols; \
+\
+/* Set alpha_ & beta_ */ \
+    assign_scalar_eig2mkl(alpha_, alpha); \
+    assign_scalar_eig2mkl(beta_, myone); \
+\
+/* Set lda, ldb, ldc */ \
+    lda = (MKL_INT)lhsStride; \
+    ldb = (MKL_INT)rhsStride; \
+    ldc = (MKL_INT)resStride; \
+\
+/* Set a, b, c */ \
+    if (((LhsStorageOrder==ColMajor) && ConjugateLhs) || ((LhsStorageOrder==RowMajor) && (!ConjugateLhs))) { \
+      Map<const Matrix<EIGTYPE, Dynamic, Dynamic, LhsStorageOrder>, 0, OuterStride<> > lhs(_lhs,m,m,OuterStride<>(lhsStride)); \
+      a_tmp = lhs.conjugate(); \
+      a = a_tmp.data(); \
+      lda = a_tmp.outerStride(); \
+    } else a = _lhs; \
+    if (LhsStorageOrder==RowMajor) uplo='U'; \
+\
+    if (RhsStorageOrder==ColMajor && (!ConjugateRhs)) { \
+       b = _rhs; } \
+    else { \
+      if (RhsStorageOrder==ColMajor && ConjugateRhs) { \
+        Map<const MatrixX##EIGPREFIX, 0, OuterStride<> > rhs(_rhs,m,n,OuterStride<>(rhsStride)); \
+        b_tmp = rhs.conjugate(); \
+      } else \
+      if (ConjugateRhs) { \
+        Map<const MatrixX##EIGPREFIX, 0, OuterStride<> > rhs(_rhs,n,m,OuterStride<>(rhsStride)); \
+        b_tmp = rhs.adjoint(); \
+      } else { \
+        Map<const MatrixX##EIGPREFIX, 0, OuterStride<> > rhs(_rhs,n,m,OuterStride<>(rhsStride)); \
+        b_tmp = rhs.transpose(); \
+      } \
+      b = b_tmp.data(); \
+      ldb = b_tmp.outerStride(); \
+    } \
+\
+    MKLPREFIX##hemm(&side, &uplo, &m, &n, &alpha_, (const MKLTYPE*)a, &lda, (const MKLTYPE*)b, &ldb, &beta_, (MKLTYPE*)res, &ldc); \
+\
+  } \
+};
+
+EIGEN_MKL_SYMM_L(double, double, d, d)
+EIGEN_MKL_SYMM_L(float, float, f, s)
+EIGEN_MKL_HEMM_L(dcomplex, MKL_Complex16, cd, z)
+EIGEN_MKL_HEMM_L(scomplex, MKL_Complex8, cf, c)
+
+
+/* Optimized matrix * selfadjoint matrix (?SYMM/?HEMM) product */
+
+#define EIGEN_MKL_SYMM_R(EIGTYPE, MKLTYPE, EIGPREFIX, MKLPREFIX) \
+template <typename Index, \
+          int LhsStorageOrder, bool ConjugateLhs, \
+          int RhsStorageOrder, bool ConjugateRhs> \
+struct product_selfadjoint_matrix<EIGTYPE,Index,LhsStorageOrder,false,ConjugateLhs,RhsStorageOrder,true,ConjugateRhs,ColMajor> \
+{\
+\
+  static EIGEN_DONT_INLINE void run( \
+    Index rows, Index cols, \
+    const EIGTYPE* _lhs, Index lhsStride, \
+    const EIGTYPE* _rhs, Index rhsStride, \
+    EIGTYPE* res,        Index resStride, \
+    EIGTYPE alpha) \
+  { \
+    char side='R', uplo='L'; \
+    MKL_INT m, n, lda, ldb, ldc; \
+    const EIGTYPE *a, *b; \
+    MKLTYPE alpha_, beta_; \
+    MatrixX##EIGPREFIX b_tmp; \
+    EIGTYPE myone(1);\
+\
+/* Set m, n, k */ \
+    m = (MKL_INT)rows;  \
+    n = (MKL_INT)cols;  \
+\
+/* Set alpha_ & beta_ */ \
+    assign_scalar_eig2mkl(alpha_, alpha); \
+    assign_scalar_eig2mkl(beta_, myone); \
+\
+/* Set lda, ldb, ldc */ \
+    lda = (MKL_INT)rhsStride; \
+    ldb = (MKL_INT)lhsStride; \
+    ldc = (MKL_INT)resStride; \
+\
+/* Set a, b, c */ \
+    if (RhsStorageOrder==RowMajor) uplo='U'; \
+    a = _rhs; \
+\
+    if (LhsStorageOrder==RowMajor) { \
+      Map<const MatrixX##EIGPREFIX, 0, OuterStride<> > lhs(_lhs,n,m,OuterStride<>(rhsStride)); \
+      b_tmp = lhs.adjoint(); \
+      b = b_tmp.data(); \
+      ldb = b_tmp.outerStride(); \
+    } else b = _lhs; \
+\
+    MKLPREFIX##symm(&side, &uplo, &m, &n, &alpha_, (const MKLTYPE*)a, &lda, (const MKLTYPE*)b, &ldb, &beta_, (MKLTYPE*)res, &ldc); \
+\
+  } \
+};
+
+
+#define EIGEN_MKL_HEMM_R(EIGTYPE, MKLTYPE, EIGPREFIX, MKLPREFIX) \
+template <typename Index, \
+          int LhsStorageOrder, bool ConjugateLhs, \
+          int RhsStorageOrder, bool ConjugateRhs> \
+struct product_selfadjoint_matrix<EIGTYPE,Index,LhsStorageOrder,false,ConjugateLhs,RhsStorageOrder,true,ConjugateRhs,ColMajor> \
+{\
+  static EIGEN_DONT_INLINE void run( \
+    Index rows, Index cols, \
+    const EIGTYPE* _lhs, Index lhsStride, \
+    const EIGTYPE* _rhs, Index rhsStride, \
+    EIGTYPE* res,        Index resStride, \
+    EIGTYPE alpha) \
+  { \
+    char side='R', uplo='L'; \
+    MKL_INT m, n, lda, ldb, ldc; \
+    const EIGTYPE *a, *b; \
+    MKLTYPE alpha_, beta_; \
+    MatrixX##EIGPREFIX b_tmp; \
+    Matrix<EIGTYPE, Dynamic, Dynamic, RhsStorageOrder> a_tmp; \
+    EIGTYPE myone(1); \
+\
+/* Set m, n, k */ \
+    m = (MKL_INT)rows; \
+    n = (MKL_INT)cols; \
+\
+/* Set alpha_ & beta_ */ \
+    assign_scalar_eig2mkl(alpha_, alpha); \
+    assign_scalar_eig2mkl(beta_, myone); \
+\
+/* Set lda, ldb, ldc */ \
+    lda = (MKL_INT)rhsStride; \
+    ldb = (MKL_INT)lhsStride; \
+    ldc = (MKL_INT)resStride; \
+\
+/* Set a, b, c */ \
+    if (((RhsStorageOrder==ColMajor) && ConjugateRhs) || ((RhsStorageOrder==RowMajor) && (!ConjugateRhs))) { \
+      Map<const Matrix<EIGTYPE, Dynamic, Dynamic, RhsStorageOrder>, 0, OuterStride<> > rhs(_rhs,n,n,OuterStride<>(rhsStride)); \
+      a_tmp = rhs.conjugate(); \
+      a = a_tmp.data(); \
+      lda = a_tmp.outerStride(); \
+    } else a = _rhs; \
+    if (RhsStorageOrder==RowMajor) uplo='U'; \
+\
+    if (LhsStorageOrder==ColMajor && (!ConjugateLhs)) { \
+       b = _lhs; } \
+    else { \
+      if (LhsStorageOrder==ColMajor && ConjugateLhs) { \
+        Map<const MatrixX##EIGPREFIX, 0, OuterStride<> > lhs(_lhs,m,n,OuterStride<>(lhsStride)); \
+        b_tmp = lhs.conjugate(); \
+      } else \
+      if (ConjugateLhs) { \
+        Map<const MatrixX##EIGPREFIX, 0, OuterStride<> > lhs(_lhs,n,m,OuterStride<>(lhsStride)); \
+        b_tmp = lhs.adjoint(); \
+      } else { \
+        Map<const MatrixX##EIGPREFIX, 0, OuterStride<> > lhs(_lhs,n,m,OuterStride<>(lhsStride)); \
+        b_tmp = lhs.transpose(); \
+      } \
+      b = b_tmp.data(); \
+      ldb = b_tmp.outerStride(); \
+    } \
+\
+    MKLPREFIX##hemm(&side, &uplo, &m, &n, &alpha_, (const MKLTYPE*)a, &lda, (const MKLTYPE*)b, &ldb, &beta_, (MKLTYPE*)res, &ldc); \
+  } \
+};
+
+EIGEN_MKL_SYMM_R(double, double, d, d)
+EIGEN_MKL_SYMM_R(float, float, f, s)
+EIGEN_MKL_HEMM_R(dcomplex, MKL_Complex16, cd, z)
+EIGEN_MKL_HEMM_R(scomplex, MKL_Complex8, cf, c)
+
+} // end namespace internal
+
+#endif // EIGEN_SELFADJOINT_MATRIX_MATRIX_MKL_H

Eigen/src/Core/products/SelfadjointMatrixVector.h

@@ -32,8 +32,15 @@ namespace internal {
  * the number of load/stores of the result by a factor 2 and to reduce
  * the instruction dependency.
  */
-template<typename Scalar, typename Index, int StorageOrder, int UpLo, bool ConjugateLhs, bool ConjugateRhs>
-static EIGEN_DONT_INLINE void product_selfadjoint_vector(
+
+template<typename Scalar, typename Index, int StorageOrder, int UpLo, bool ConjugateLhs, bool ConjugateRhs, int Version=Specialized>
+struct selfadjoint_matrix_vector_product;
+
+template<typename Scalar, typename Index, int StorageOrder, int UpLo, bool ConjugateLhs, bool ConjugateRhs, int Version>
+struct selfadjoint_matrix_vector_product
+
+{
+static EIGEN_DONT_INLINE void run(
   Index size,
   const Scalar*  lhs, Index lhsStride,
   const Scalar* _rhs, Index rhsIncr,
@@ -85,14 +92,14 @@ static EIGEN_DONT_INLINE void product_selfadjoint_vector(
     Scalar t1 = cjAlpha * rhs[j+1];
     Packet ptmp1 = pset1<Packet>(t1);
 
-    Scalar t2 = 0;
+    Scalar t2(0);
     Packet ptmp2 = pset1<Packet>(t2);
-    Scalar t3 = 0;
+    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) + first_aligned(&res[starti], endi-starti);
+    size_t alignedStart = (starti) + internal::first_aligned(&res[starti], endi-starti);
     size_t alignedEnd = alignedStart + ((endi-alignedStart)/(PacketSize))*(PacketSize);
 
     // TODO make sure this product is a real * complex and that the rhs is properly conjugated if needed
@@ -148,7 +155,7 @@ static EIGEN_DONT_INLINE void product_selfadjoint_vector(
     register const Scalar* EIGEN_RESTRICT A0 = lhs + j*lhsStride;
 
     Scalar t1 = cjAlpha * rhs[j];
-    Scalar t2 = 0;
+    Scalar t2(0);
     // TODO make sure this product is a real * complex and that the rhs is properly conjugated if needed
     res[j] += cjd.pmul(internal::real(A0[j]), t1);
     for (Index i=FirstTriangular ? 0 : j+1; i<(FirstTriangular ? j : size); i++)
@@ -159,6 +166,7 @@ static EIGEN_DONT_INLINE void product_selfadjoint_vector(
     res[j] += alpha * t2;
   }
 }
+};
 
 } // end namespace internal 
 
@@ -193,8 +201,8 @@ struct SelfadjointProductMatrix<Lhs,LhsMode,false,Rhs,0,true>
     
     eigen_assert(dest.rows()==m_lhs.rows() && dest.cols()==m_rhs.cols());
 
-    const ActualLhsType lhs = LhsBlasTraits::extract(m_lhs);
-    const ActualRhsType rhs = RhsBlasTraits::extract(m_rhs);
+    typename internal::add_const_on_value_type<ActualLhsType>::type lhs = LhsBlasTraits::extract(m_lhs);
+    typename internal::add_const_on_value_type<ActualRhsType>::type rhs = RhsBlasTraits::extract(m_rhs);
 
     Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(m_lhs)
                                * RhsBlasTraits::extractScalarFactor(m_rhs);
@@ -232,7 +240,7 @@ struct SelfadjointProductMatrix<Lhs,LhsMode,false,Rhs,0,true>
     }
       
       
-    internal::product_selfadjoint_vector<Scalar, Index, (internal::traits<_ActualLhsType>::Flags&RowMajorBit) ? RowMajor : ColMajor, int(LhsUpLo), bool(LhsBlasTraits::NeedToConjugate), bool(RhsBlasTraits::NeedToConjugate)>
+    internal::selfadjoint_matrix_vector_product<Scalar, Index, (internal::traits<_ActualLhsType>::Flags&RowMajorBit) ? RowMajor : ColMajor, int(LhsUpLo), bool(LhsBlasTraits::NeedToConjugate), bool(RhsBlasTraits::NeedToConjugate)>::run
       (
         lhs.rows(),                             // size
         &lhs.coeffRef(0,0),  lhs.outerStride(), // lhs info

Eigen/src/Core/products/SelfadjointMatrixVector_MKL.h

@@ -0,0 +1,110 @@
+/*
+ Copyright (c) 2011, Intel Corporation. All rights reserved.
+
+ Redistribution and use in source and binary forms, with or without modification,
+ are permitted provided that the following conditions are met:
+
+ * Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+ * Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+ * Neither the name of Intel Corporation nor the names of its contributors may
+   be used to endorse or promote products derived from this software without
+   specific prior written permission.
+
+ THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+ WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+ DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+ (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+ LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+ ********************************************************************************
+ *   Content : Eigen bindings to Intel(R) MKL
+ *   Selfadjoint matrix-vector product functionality based on ?SYMV/HEMV.
+ ********************************************************************************
+*/
+
+#ifndef EIGEN_SELFADJOINT_MATRIX_VECTOR_MKL_H
+#define EIGEN_SELFADJOINT_MATRIX_VECTOR_MKL_H
+
+namespace internal {
+
+/**********************************************************************
+* This file implements selfadjoint matrix-vector multiplication using BLAS
+**********************************************************************/
+
+// symv/hemv specialization
+
+template<typename Scalar, typename Index, int StorageOrder, int UpLo, bool ConjugateLhs, bool ConjugateRhs>
+struct selfadjoint_matrix_vector_product_symv :
+  selfadjoint_matrix_vector_product<Scalar,Index,StorageOrder,UpLo,ConjugateLhs,ConjugateRhs,BuiltIn> {};
+
+#define EIGEN_MKL_SYMV_SPECIALIZE(Scalar) \
+template<typename Index, int StorageOrder, int UpLo, bool ConjugateLhs, bool ConjugateRhs> \
+struct selfadjoint_matrix_vector_product<Scalar,Index,StorageOrder,UpLo,ConjugateLhs,ConjugateRhs,Specialized> { \
+static EIGEN_DONT_INLINE void run( \
+  Index size, const Scalar*  lhs, Index lhsStride, \
+  const Scalar* _rhs, Index rhsIncr, Scalar* res, Scalar alpha) { \
+    enum {\
+      IsColMajor = StorageOrder==ColMajor \
+    }; \
+    if (IsColMajor == ConjugateLhs) {\
+      selfadjoint_matrix_vector_product<Scalar,Index,StorageOrder,UpLo,ConjugateLhs,ConjugateRhs,BuiltIn>::run( \
+        size, lhs, lhsStride, _rhs, rhsIncr, res, alpha);  \
+    } else {\
+      selfadjoint_matrix_vector_product_symv<Scalar,Index,StorageOrder,UpLo,ConjugateLhs,ConjugateRhs>::run( \
+        size, lhs, lhsStride, _rhs, rhsIncr, res, alpha);  \
+    }\
+  } \
+}; \
+
+EIGEN_MKL_SYMV_SPECIALIZE(double)
+EIGEN_MKL_SYMV_SPECIALIZE(float)
+EIGEN_MKL_SYMV_SPECIALIZE(dcomplex)
+EIGEN_MKL_SYMV_SPECIALIZE(scomplex)
+
+#define EIGEN_MKL_SYMV_SPECIALIZATION(EIGTYPE,MKLTYPE,MKLFUNC) \
+template<typename Index, int StorageOrder, int UpLo, bool ConjugateLhs, bool ConjugateRhs> \
+struct selfadjoint_matrix_vector_product_symv<EIGTYPE,Index,StorageOrder,UpLo,ConjugateLhs,ConjugateRhs> \
+{ \
+typedef Matrix<EIGTYPE,Dynamic,1,ColMajor> SYMVVector;\
+\
+static EIGEN_DONT_INLINE void run( \
+Index size, const EIGTYPE*  lhs, Index lhsStride, \
+const EIGTYPE* _rhs, Index rhsIncr, EIGTYPE* res, EIGTYPE alpha) \
+{ \
+  enum {\
+    IsRowMajor = StorageOrder==RowMajor ? 1 : 0, \
+    IsLower = UpLo == Lower ? 1 : 0, \
+  }; \
+  MKL_INT n=size, lda=lhsStride, incx=rhsIncr, incy=1; \
+  MKLTYPE alpha_, beta_; \
+  const EIGTYPE *x_ptr, myone(1); \
+  char uplo=(IsRowMajor) ? (IsLower ? 'U' : 'L') : (IsLower ? 'L' : 'U'); \
+  assign_scalar_eig2mkl(alpha_, alpha); \
+  assign_scalar_eig2mkl(beta_, myone); \
+  SYMVVector x_tmp; \
+  if (ConjugateRhs) { \
+    Map<const SYMVVector, 0, InnerStride<> > map_x(_rhs,size,1,InnerStride<>(incx)); \
+    x_tmp=map_x.conjugate(); \
+    x_ptr=x_tmp.data(); \
+    incx=1; \
+  } else x_ptr=_rhs; \
+  MKLFUNC(&uplo, &n, &alpha_, (const MKLTYPE*)lhs, &lda, (const MKLTYPE*)x_ptr, &incx, &beta_, (MKLTYPE*)res, &incy); \
+}\
+};
+
+EIGEN_MKL_SYMV_SPECIALIZATION(double,   double,        dsymv)
+EIGEN_MKL_SYMV_SPECIALIZATION(float,    float,         ssymv)
+EIGEN_MKL_SYMV_SPECIALIZATION(dcomplex, MKL_Complex16, zhemv)
+EIGEN_MKL_SYMV_SPECIALIZATION(scomplex, MKL_Complex8,  chemv)
+
+} //end of namespase
+
+#endif // EIGEN_SELFADJOINT_MATRIX_VECTOR_MKL_H

Eigen/src/Core/products/SelfadjointProduct.h

@@ -72,7 +72,7 @@ struct selfadjoint_product_selector<MatrixType,OtherType,UpLo,true>
     typedef internal::blas_traits<OtherType> OtherBlasTraits;
     typedef typename OtherBlasTraits::DirectLinearAccessType ActualOtherType;
     typedef typename internal::remove_all<ActualOtherType>::type _ActualOtherType;
-    const ActualOtherType actualOther = OtherBlasTraits::extract(other.derived());
+    typename internal::add_const_on_value_type<ActualOtherType>::type actualOther = OtherBlasTraits::extract(other.derived());
 
     Scalar actualAlpha = alpha * OtherBlasTraits::extractScalarFactor(other.derived());
 
@@ -105,12 +105,12 @@ struct selfadjoint_product_selector<MatrixType,OtherType,UpLo,false>
     typedef internal::blas_traits<OtherType> OtherBlasTraits;
     typedef typename OtherBlasTraits::DirectLinearAccessType ActualOtherType;
     typedef typename internal::remove_all<ActualOtherType>::type _ActualOtherType;
-    const ActualOtherType actualOther = OtherBlasTraits::extract(other.derived());
+    typename internal::add_const_on_value_type<ActualOtherType>::type actualOther = OtherBlasTraits::extract(other.derived());
 
     Scalar actualAlpha = alpha * OtherBlasTraits::extractScalarFactor(other.derived());
 
     enum { IsRowMajor = (internal::traits<MatrixType>::Flags&RowMajorBit) ? 1 : 0 };
-    
+
     internal::general_matrix_matrix_triangular_product<Index,
       Scalar, _ActualOtherType::Flags&RowMajorBit ? RowMajor : ColMajor,   OtherBlasTraits::NeedToConjugate  && NumTraits<Scalar>::IsComplex,
       Scalar, _ActualOtherType::Flags&RowMajorBit ? ColMajor : RowMajor, (!OtherBlasTraits::NeedToConjugate) && NumTraits<Scalar>::IsComplex,

Eigen/src/Core/products/SelfadjointRank2Update.h

@@ -76,12 +76,12 @@ SelfAdjointView<MatrixType,UpLo>& SelfAdjointView<MatrixType,UpLo>
   typedef internal::blas_traits<DerivedU> UBlasTraits;
   typedef typename UBlasTraits::DirectLinearAccessType ActualUType;
   typedef typename internal::remove_all<ActualUType>::type _ActualUType;
-  const ActualUType actualU = UBlasTraits::extract(u.derived());
+  typename internal::add_const_on_value_type<ActualUType>::type actualU = UBlasTraits::extract(u.derived());
 
   typedef internal::blas_traits<DerivedV> VBlasTraits;
   typedef typename VBlasTraits::DirectLinearAccessType ActualVType;
   typedef typename internal::remove_all<ActualVType>::type _ActualVType;
-  const ActualVType actualV = VBlasTraits::extract(v.derived());
+  typename internal::add_const_on_value_type<ActualVType>::type actualV = VBlasTraits::extract(v.derived());
 
   // If MatrixType is row major, then we use the routine for lower triangular in the upper triangular case and
   // vice versa, and take the complex conjugate of all coefficients and vector entries.

Eigen/src/Core/products/TriangularMatrixMatrix.h

@@ -58,16 +58,16 @@ template <typename Scalar, typename Index,
           int Mode, bool LhsIsTriangular,
           int LhsStorageOrder, bool ConjugateLhs,
           int RhsStorageOrder, bool ConjugateRhs,
-          int ResStorageOrder>
+          int ResStorageOrder, int Version = Specialized>
 struct product_triangular_matrix_matrix;
 
 template <typename Scalar, typename Index,
           int Mode, bool LhsIsTriangular,
           int LhsStorageOrder, bool ConjugateLhs,
-          int RhsStorageOrder, bool ConjugateRhs>
+          int RhsStorageOrder, bool ConjugateRhs, int Version>
 struct product_triangular_matrix_matrix<Scalar,Index,Mode,LhsIsTriangular,
                                            LhsStorageOrder,ConjugateLhs,
-                                           RhsStorageOrder,ConjugateRhs,RowMajor>
+                                           RhsStorageOrder,ConjugateRhs,RowMajor,Version>
 {
   static EIGEN_STRONG_INLINE void run(
     Index rows, Index cols, Index depth,
@@ -91,15 +91,15 @@ struct product_triangular_matrix_matrix<Scalar,Index,Mode,LhsIsTriangular,
 // implements col-major += alpha * op(triangular) * op(general)
 template <typename Scalar, typename Index, int Mode,
           int LhsStorageOrder, bool ConjugateLhs,
-          int RhsStorageOrder, bool ConjugateRhs>
+          int RhsStorageOrder, bool ConjugateRhs, int Version>
 struct product_triangular_matrix_matrix<Scalar,Index,Mode,true,
                                            LhsStorageOrder,ConjugateLhs,
-                                           RhsStorageOrder,ConjugateRhs,ColMajor>
+                                           RhsStorageOrder,ConjugateRhs,ColMajor,Version>
 {
   
   typedef gebp_traits<Scalar,Scalar> Traits;
   enum {
-    SmallPanelWidth   = EIGEN_PLAIN_ENUM_MAX(Traits::mr,Traits::nr),
+    SmallPanelWidth   = 2 * EIGEN_PLAIN_ENUM_MAX(Traits::mr,Traits::nr),
     IsLower = (Mode&Lower) == Lower,
     SetDiag = (Mode&(ZeroDiag|UnitDiag)) ? 0 : 1
   };
@@ -220,10 +220,10 @@ struct product_triangular_matrix_matrix<Scalar,Index,Mode,true,
 // implements col-major += alpha * op(general) * op(triangular)
 template <typename Scalar, typename Index, int Mode,
           int LhsStorageOrder, bool ConjugateLhs,
-          int RhsStorageOrder, bool ConjugateRhs>
+          int RhsStorageOrder, bool ConjugateRhs, int Version>
 struct product_triangular_matrix_matrix<Scalar,Index,Mode,false,
                                            LhsStorageOrder,ConjugateLhs,
-                                           RhsStorageOrder,ConjugateRhs,ColMajor>
+                                           RhsStorageOrder,ConjugateRhs,ColMajor,Version>
 {
   typedef gebp_traits<Scalar,Scalar> Traits;
   enum {
@@ -378,8 +378,8 @@ struct TriangularProduct<Mode,LhsIsTriangular,Lhs,false,Rhs,false>
 
   template<typename Dest> void scaleAndAddTo(Dest& dst, Scalar alpha) const
   {
-    const ActualLhsType lhs = LhsBlasTraits::extract(m_lhs);
-    const ActualRhsType rhs = RhsBlasTraits::extract(m_rhs);
+    typename internal::add_const_on_value_type<ActualLhsType>::type lhs = LhsBlasTraits::extract(m_lhs);
+    typename internal::add_const_on_value_type<ActualRhsType>::type rhs = RhsBlasTraits::extract(m_rhs);
 
     Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(m_lhs)
                                * RhsBlasTraits::extractScalarFactor(m_rhs);

Eigen/src/Core/products/TriangularMatrixMatrix_MKL.h

@@ -0,0 +1,305 @@
+/*
+ Copyright (c) 2011, Intel Corporation. All rights reserved.
+
+ Redistribution and use in source and binary forms, with or without modification,
+ are permitted provided that the following conditions are met:
+
+ * Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+ * Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+ * Neither the name of Intel Corporation nor the names of its contributors may
+   be used to endorse or promote products derived from this software without
+   specific prior written permission.
+
+ THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+ WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+ DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+ (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+ LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+ ********************************************************************************
+ *   Content : Eigen bindings to Intel(R) MKL
+ *   Triangular matrix * matrix product functionality based on ?TRMM.
+ ********************************************************************************
+*/
+
+#ifndef EIGEN_TRIANGULAR_MATRIX_MATRIX_MKL_H
+#define EIGEN_TRIANGULAR_MATRIX_MATRIX_MKL_H
+
+namespace internal {
+
+
+template <typename Scalar, typename Index,
+          int Mode, bool LhsIsTriangular,
+          int LhsStorageOrder, bool ConjugateLhs,
+          int RhsStorageOrder, bool ConjugateRhs,
+          int ResStorageOrder>
+struct product_triangular_matrix_matrix_trmm :
+       product_triangular_matrix_matrix<Scalar,Index,Mode,
+          LhsIsTriangular,LhsStorageOrder,ConjugateLhs,
+          RhsStorageOrder, ConjugateRhs, ResStorageOrder, BuiltIn> {};
+
+
+// try to go to BLAS specialization
+#define EIGEN_MKL_TRMM_SPECIALIZE(Scalar, LhsIsTriangular) \
+template <typename Index, int Mode, \
+          int LhsStorageOrder, bool ConjugateLhs, \
+          int RhsStorageOrder, bool ConjugateRhs> \
+struct product_triangular_matrix_matrix<Scalar,Index, Mode, LhsIsTriangular, \
+           LhsStorageOrder,ConjugateLhs, RhsStorageOrder,ConjugateRhs,ColMajor,Specialized> { \
+  static inline void run(Index _rows, Index _cols, Index _depth, const Scalar* _lhs, Index lhsStride,\
+    const Scalar* _rhs, Index rhsStride, Scalar* res, Index resStride, Scalar alpha) { \
+      product_triangular_matrix_matrix_trmm<Scalar,Index,Mode, \
+        LhsIsTriangular,LhsStorageOrder,ConjugateLhs, \
+        RhsStorageOrder, ConjugateRhs, ColMajor>::run( \
+        _rows, _cols, _depth, _lhs, lhsStride, _rhs, rhsStride, res, resStride, alpha); \
+  } \
+};
+
+EIGEN_MKL_TRMM_SPECIALIZE(double, true)
+EIGEN_MKL_TRMM_SPECIALIZE(double, false)
+EIGEN_MKL_TRMM_SPECIALIZE(dcomplex, true)
+EIGEN_MKL_TRMM_SPECIALIZE(dcomplex, false)
+EIGEN_MKL_TRMM_SPECIALIZE(float, true)
+EIGEN_MKL_TRMM_SPECIALIZE(float, false)
+EIGEN_MKL_TRMM_SPECIALIZE(scomplex, true)
+EIGEN_MKL_TRMM_SPECIALIZE(scomplex, false)
+
+// implements col-major += alpha * op(triangular) * op(general)
+#define EIGEN_MKL_TRMM_L(EIGTYPE, MKLTYPE, EIGPREFIX, MKLPREFIX) \
+template <typename Index, int Mode, \
+          int LhsStorageOrder, bool ConjugateLhs, \
+          int RhsStorageOrder, bool ConjugateRhs> \
+struct product_triangular_matrix_matrix_trmm<EIGTYPE,Index,Mode,true, \
+         LhsStorageOrder,ConjugateLhs,RhsStorageOrder,ConjugateRhs,ColMajor> \
+{ \
+  enum { \
+    IsLower = (Mode&Lower) == Lower, \
+    SetDiag = (Mode&(ZeroDiag|UnitDiag)) ? 0 : 1, \
+    IsUnitDiag  = (Mode&UnitDiag) ? 1 : 0, \
+    IsZeroDiag  = (Mode&ZeroDiag) ? 1 : 0, \
+    LowUp = IsLower ? Lower : Upper, \
+    conjA = ((LhsStorageOrder==ColMajor) && ConjugateLhs) ? 1 : 0 \
+  }; \
+\
+  static EIGEN_DONT_INLINE void run( \
+    Index _rows, Index _cols, Index _depth, \
+    const EIGTYPE* _lhs, Index lhsStride, \
+    const EIGTYPE* _rhs, Index rhsStride, \
+    EIGTYPE* res,        Index resStride, \
+    EIGTYPE alpha) \
+  { \
+   Index diagSize  = (std::min)(_rows,_depth); \
+   Index rows      = IsLower ? _rows : diagSize; \
+   Index depth     = IsLower ? diagSize : _depth; \
+   Index cols      = _cols; \
+\
+   typedef Matrix<EIGTYPE, Dynamic, Dynamic, LhsStorageOrder> MatrixLhs; \
+   typedef Matrix<EIGTYPE, Dynamic, Dynamic, RhsStorageOrder> MatrixRhs; \
+\
+/* Non-square case - doesn't fit to MKL ?TRMM. Fall to default triangular product or call MKL ?GEMM*/ \
+   if (rows != depth) { \
+\
+     int nthr = mkl_domain_get_max_threads(MKL_BLAS); \
+\
+     if (((nthr==1) && (((std::max)(rows,depth)-diagSize)/(double)diagSize < 0.5))) { \
+     /* Most likely no benefit to call TRMM or GEMM from MKL*/ \
+       product_triangular_matrix_matrix<EIGTYPE,Index,Mode,true, \
+       LhsStorageOrder,ConjugateLhs, RhsStorageOrder, ConjugateRhs, ColMajor, BuiltIn>::run( \
+           _rows, _cols, _depth, _lhs, lhsStride, _rhs, rhsStride, res, resStride, alpha); \
+     /*std::cout << "TRMM_L: A is not square! Go to Eigen TRMM implementation!\n";*/ \
+     } else { \
+     /* Make sense to call GEMM */ \
+       Map<const MatrixLhs, 0, OuterStride<> > lhsMap(_lhs,rows,depth,OuterStride<>(lhsStride)); \
+       MatrixLhs aa_tmp=lhsMap.template triangularView<Mode>(); \
+       MKL_INT aStride = aa_tmp.outerStride(); \
+       gemm_blocking_space<ColMajor,EIGTYPE,EIGTYPE,Dynamic,Dynamic,Dynamic> blocking(_rows,_cols,_depth); \
+       general_matrix_matrix_product<Index,EIGTYPE,LhsStorageOrder,ConjugateLhs,EIGTYPE,RhsStorageOrder,ConjugateRhs,ColMajor>::run( \
+       rows, cols, depth, aa_tmp.data(), aStride, _rhs, rhsStride, res, resStride, alpha, blocking); \
+\
+     /*std::cout << "TRMM_L: A is not square! Go to MKL GEMM implementation! " << nthr<<" \n";*/ \
+     } \
+     return; \
+   } \
+   char side = 'L', transa, uplo, diag = 'N'; \
+   EIGTYPE *b; \
+   const EIGTYPE *a; \
+   MKL_INT m, n, k, lda, ldb, ldc; \
+   MKLTYPE alpha_; \
+\
+/* Set alpha_*/ \
+   assign_scalar_eig2mkl<MKLTYPE, EIGTYPE>(alpha_, alpha); \
+\
+/* Set m, n */ \
+   m = (MKL_INT)diagSize; \
+   n = (MKL_INT)cols; \
+\
+/* Set trans */ \
+   transa = (LhsStorageOrder==RowMajor) ? ((ConjugateLhs) ? 'C' : 'T') : 'N'; \
+\
+/* Set b, ldb */ \
+   Map<const MatrixRhs, 0, OuterStride<> > rhs(_rhs,depth,cols,OuterStride<>(rhsStride)); \
+   MatrixX##EIGPREFIX b_tmp; \
+\
+   if (ConjugateRhs) b_tmp = rhs.conjugate(); else b_tmp = rhs; \
+   b = b_tmp.data(); \
+   ldb = b_tmp.outerStride(); \
+\
+/* Set uplo */ \
+   uplo = IsLower ? 'L' : 'U'; \
+   if (LhsStorageOrder==RowMajor) uplo = (uplo == 'L') ? 'U' : 'L'; \
+/* Set a, lda */ \
+   Map<const MatrixLhs, 0, OuterStride<> > lhs(_lhs,rows,depth,OuterStride<>(lhsStride)); \
+   MatrixLhs a_tmp; \
+\
+   if ((conjA!=0) || (SetDiag==0)) { \
+     if (conjA) a_tmp = lhs.conjugate(); else a_tmp = lhs; \
+     if (IsZeroDiag) \
+       a_tmp.diagonal().setZero(); \
+     else if (IsUnitDiag) \
+       a_tmp.diagonal().setOnes();\
+     a = a_tmp.data(); \
+     lda = a_tmp.outerStride(); \
+   } else { \
+     a = _lhs; \
+     lda = lhsStride; \
+   } \
+   /*std::cout << "TRMM_L: A is square! Go to MKL TRMM implementation! \n";*/ \
+/* call ?trmm*/ \
+   MKLPREFIX##trmm(&side, &uplo, &transa, &diag, &m, &n, &alpha_, (const MKLTYPE*)a, &lda, (MKLTYPE*)b, &ldb); \
+\
+/* Add op(a_triangular)*b into res*/ \
+   Map<MatrixX##EIGPREFIX, 0, OuterStride<> > res_tmp(res,rows,cols,OuterStride<>(resStride)); \
+   res_tmp=res_tmp+b_tmp; \
+  } \
+};
+
+EIGEN_MKL_TRMM_L(double, double, d, d)
+EIGEN_MKL_TRMM_L(dcomplex, MKL_Complex16, cd, z)
+EIGEN_MKL_TRMM_L(float, float, f, s)
+EIGEN_MKL_TRMM_L(scomplex, MKL_Complex8, cf, c)
+
+// implements col-major += alpha * op(general) * op(triangular)
+#define EIGEN_MKL_TRMM_R(EIGTYPE, MKLTYPE, EIGPREFIX, MKLPREFIX) \
+template <typename Index, int Mode, \
+          int LhsStorageOrder, bool ConjugateLhs, \
+          int RhsStorageOrder, bool ConjugateRhs> \
+struct product_triangular_matrix_matrix_trmm<EIGTYPE,Index,Mode,false, \
+         LhsStorageOrder,ConjugateLhs,RhsStorageOrder,ConjugateRhs,ColMajor> \
+{ \
+  enum { \
+    IsLower = (Mode&Lower) == Lower, \
+    SetDiag = (Mode&(ZeroDiag|UnitDiag)) ? 0 : 1, \
+    IsUnitDiag  = (Mode&UnitDiag) ? 1 : 0, \
+    IsZeroDiag  = (Mode&ZeroDiag) ? 1 : 0, \
+    LowUp = IsLower ? Lower : Upper, \
+    conjA = ((RhsStorageOrder==ColMajor) && ConjugateRhs) ? 1 : 0 \
+  }; \
+\
+  static EIGEN_DONT_INLINE void run( \
+    Index _rows, Index _cols, Index _depth, \
+    const EIGTYPE* _lhs, Index lhsStride, \
+    const EIGTYPE* _rhs, Index rhsStride, \
+    EIGTYPE* res,        Index resStride, \
+    EIGTYPE alpha) \
+  { \
+   Index diagSize  = (std::min)(_cols,_depth); \
+   Index rows      = _rows; \
+   Index depth     = IsLower ? _depth : diagSize; \
+   Index cols      = IsLower ? diagSize : _cols; \
+\
+   typedef Matrix<EIGTYPE, Dynamic, Dynamic, LhsStorageOrder> MatrixLhs; \
+   typedef Matrix<EIGTYPE, Dynamic, Dynamic, RhsStorageOrder> MatrixRhs; \
+\
+/* Non-square case - doesn't fit to MKL ?TRMM. Fall to default triangular product or call MKL ?GEMM*/ \
+   if (cols != depth) { \
+\
+     int nthr = mkl_domain_get_max_threads(MKL_BLAS); \
+\
+     if ((nthr==1) && (((std::max)(cols,depth)-diagSize)/(double)diagSize < 0.5)) { \
+     /* Most likely no benefit to call TRMM or GEMM from MKL*/ \
+       product_triangular_matrix_matrix<EIGTYPE,Index,Mode,false, \
+       LhsStorageOrder,ConjugateLhs, RhsStorageOrder, ConjugateRhs, ColMajor, BuiltIn>::run( \
+           _rows, _cols, _depth, _lhs, lhsStride, _rhs, rhsStride, res, resStride, alpha); \
+       /*std::cout << "TRMM_R: A is not square! Go to Eigen TRMM implementation!\n";*/ \
+     } else { \
+     /* Make sense to call GEMM */ \
+       Map<const MatrixRhs, 0, OuterStride<> > rhsMap(_rhs,depth,cols, OuterStride<>(rhsStride)); \
+       MatrixRhs aa_tmp=rhsMap.template triangularView<Mode>(); \
+       MKL_INT aStride = aa_tmp.outerStride(); \
+       gemm_blocking_space<ColMajor,EIGTYPE,EIGTYPE,Dynamic,Dynamic,Dynamic> blocking(_rows,_cols,_depth); \
+       general_matrix_matrix_product<Index,EIGTYPE,LhsStorageOrder,ConjugateLhs,EIGTYPE,RhsStorageOrder,ConjugateRhs,ColMajor>::run( \
+       rows, cols, depth, _lhs, lhsStride, aa_tmp.data(), aStride, res, resStride, alpha, blocking); \
+\
+     /*std::cout << "TRMM_R: A is not square! Go to MKL GEMM implementation! " << nthr<<" \n";*/ \
+     } \
+     return; \
+   } \
+   char side = 'R', transa, uplo, diag = 'N'; \
+   EIGTYPE *b; \
+   const EIGTYPE *a; \
+   MKL_INT m, n, k, lda, ldb, ldc; \
+   MKLTYPE alpha_; \
+\
+/* Set alpha_*/ \
+   assign_scalar_eig2mkl<MKLTYPE, EIGTYPE>(alpha_, alpha); \
+\
+/* Set m, n */ \
+   m = (MKL_INT)rows; \
+   n = (MKL_INT)diagSize; \
+\
+/* Set trans */ \
+   transa = (RhsStorageOrder==RowMajor) ? ((ConjugateRhs) ? 'C' : 'T') : 'N'; \
+\
+/* Set b, ldb */ \
+   Map<const MatrixLhs, 0, OuterStride<> > lhs(_lhs,rows,depth,OuterStride<>(lhsStride)); \
+   MatrixX##EIGPREFIX b_tmp; \
+\
+   if (ConjugateLhs) b_tmp = lhs.conjugate(); else b_tmp = lhs; \
+   b = b_tmp.data(); \
+   ldb = b_tmp.outerStride(); \
+\
+/* Set uplo */ \
+   uplo = IsLower ? 'L' : 'U'; \
+   if (RhsStorageOrder==RowMajor) uplo = (uplo == 'L') ? 'U' : 'L'; \
+/* Set a, lda */ \
+   Map<const MatrixRhs, 0, OuterStride<> > rhs(_rhs,depth,cols, OuterStride<>(rhsStride)); \
+   MatrixRhs a_tmp; \
+\
+   if ((conjA!=0) || (SetDiag==0)) { \
+     if (conjA) a_tmp = rhs.conjugate(); else a_tmp = rhs; \
+     if (IsZeroDiag) \
+       a_tmp.diagonal().setZero(); \
+     else if (IsUnitDiag) \
+       a_tmp.diagonal().setOnes();\
+     a = a_tmp.data(); \
+     lda = a_tmp.outerStride(); \
+   } else { \
+     a = _rhs; \
+     lda = rhsStride; \
+   } \
+   /*std::cout << "TRMM_R: A is square! Go to MKL TRMM implementation! \n";*/ \
+/* call ?trmm*/ \
+   MKLPREFIX##trmm(&side, &uplo, &transa, &diag, &m, &n, &alpha_, (const MKLTYPE*)a, &lda, (MKLTYPE*)b, &ldb); \
+\
+/* Add op(a_triangular)*b into res*/ \
+   Map<MatrixX##EIGPREFIX, 0, OuterStride<> > res_tmp(res,rows,cols,OuterStride<>(resStride)); \
+   res_tmp=res_tmp+b_tmp; \
+  } \
+};
+
+EIGEN_MKL_TRMM_R(double, double, d, d)
+EIGEN_MKL_TRMM_R(dcomplex, MKL_Complex16, cd, z)
+EIGEN_MKL_TRMM_R(float, float, f, s)
+EIGEN_MKL_TRMM_R(scomplex, MKL_Complex8, cf, c)
+
+} // end namespace internal
+
+#endif // EIGEN_TRIANGULAR_MATRIX_MATRIX_MKL_H

Eigen/src/Core/products/TriangularMatrixVector.h

@@ -27,21 +27,25 @@
 
 namespace internal {
 
-template<typename Index, int Mode, typename LhsScalar, bool ConjLhs, typename RhsScalar, bool ConjRhs, int StorageOrder>
-struct product_triangular_matrix_vector;
+template<typename Index, int Mode, typename LhsScalar, bool ConjLhs, typename RhsScalar, bool ConjRhs, int StorageOrder, int Version=Specialized>
+struct triangular_matrix_vector_product;
 
-template<typename Index, int Mode, typename LhsScalar, bool ConjLhs, typename RhsScalar, bool ConjRhs>
-struct product_triangular_matrix_vector<Index,Mode,LhsScalar,ConjLhs,RhsScalar,ConjRhs,ColMajor>
+template<typename Index, int Mode, typename LhsScalar, bool ConjLhs, typename RhsScalar, bool ConjRhs, int Version>
+struct triangular_matrix_vector_product<Index,Mode,LhsScalar,ConjLhs,RhsScalar,ConjRhs,ColMajor,Version>
 {
   typedef typename scalar_product_traits<LhsScalar, RhsScalar>::ReturnType ResScalar;
   enum {
     IsLower = ((Mode&Lower)==Lower),
-    HasUnitDiag = (Mode & UnitDiag)==UnitDiag
+    HasUnitDiag = (Mode & UnitDiag)==UnitDiag,
+    HasZeroDiag = (Mode & ZeroDiag)==ZeroDiag
   };
-  static EIGEN_DONT_INLINE  void run(Index rows, Index cols, const LhsScalar* _lhs, Index lhsStride,
+  static EIGEN_DONT_INLINE  void run(Index _rows, Index _cols, const LhsScalar* _lhs, Index lhsStride,
                                      const RhsScalar* _rhs, Index rhsIncr, ResScalar* _res, Index resIncr, ResScalar alpha)
   {
     static const Index PanelWidth = EIGEN_TUNE_TRIANGULAR_PANEL_WIDTH;
+    Index size = (std::min)(_rows,_cols);
+    Index rows = IsLower ? _rows : (std::min)(_rows,_cols);
+    Index cols = IsLower ? (std::min)(_rows,_cols) : _cols;
 
     typedef Map<const Matrix<LhsScalar,Dynamic,Dynamic,ColMajor>, 0, OuterStride<> > LhsMap;
     const LhsMap lhs(_lhs,rows,cols,OuterStride<>(lhsStride));
@@ -54,45 +58,57 @@ struct product_triangular_matrix_vector<Index,Mode,LhsScalar,ConjLhs,RhsScalar,C
     typedef Map<Matrix<ResScalar,Dynamic,1> > ResMap;
     ResMap res(_res,rows);
 
-    for (Index pi=0; pi<cols; pi+=PanelWidth)
+    for (Index pi=0; pi<size; pi+=PanelWidth)
     {
-      Index actualPanelWidth = (std::min)(PanelWidth, cols-pi);
+      Index actualPanelWidth = (std::min)(PanelWidth, size-pi);
       for (Index k=0; k<actualPanelWidth; ++k)
       {
         Index i = pi + k;
-        Index s = IsLower ? (HasUnitDiag ? i+1 : i ) : pi;
+        Index s = IsLower ? ((HasUnitDiag||HasZeroDiag) ? i+1 : i ) : pi;
         Index r = IsLower ? actualPanelWidth-k : k+1;
-        if ((!HasUnitDiag) || (--r)>0)
+        if ((!(HasUnitDiag||HasZeroDiag)) || (--r)>0)
           res.segment(s,r) += (alpha * cjRhs.coeff(i)) * cjLhs.col(i).segment(s,r);
         if (HasUnitDiag)
           res.coeffRef(i) += alpha * cjRhs.coeff(i);
       }
-      Index r = IsLower ? cols - pi - actualPanelWidth : pi;
+      Index r = IsLower ? rows - pi - actualPanelWidth : pi;
       if (r>0)
       {
         Index s = IsLower ? pi+actualPanelWidth : 0;
-        general_matrix_vector_product<Index,LhsScalar,ColMajor,ConjLhs,RhsScalar,ConjRhs>::run(
+        general_matrix_vector_product<Index,LhsScalar,ColMajor,ConjLhs,RhsScalar,ConjRhs,BuiltIn>::run(
             r, actualPanelWidth,
             &lhs.coeffRef(s,pi), lhsStride,
             &rhs.coeffRef(pi), rhsIncr,
             &res.coeffRef(s), resIncr, alpha);
       }
     }
+    if((!IsLower) && cols>size)
+    {
+      general_matrix_vector_product<Index,LhsScalar,ColMajor,ConjLhs,RhsScalar,ConjRhs>::run(
+          rows, cols-size,
+          &lhs.coeffRef(0,size), lhsStride,
+          &rhs.coeffRef(size), rhsIncr,
+          _res, resIncr, alpha);
+    }
   }
 };
 
-template<typename Index, int Mode, typename LhsScalar, bool ConjLhs, typename RhsScalar, bool ConjRhs>
-struct product_triangular_matrix_vector<Index,Mode,LhsScalar,ConjLhs,RhsScalar,ConjRhs,RowMajor>
+template<typename Index, int Mode, typename LhsScalar, bool ConjLhs, typename RhsScalar, bool ConjRhs,int Version>
+struct triangular_matrix_vector_product<Index,Mode,LhsScalar,ConjLhs,RhsScalar,ConjRhs,RowMajor,Version>
 {
   typedef typename scalar_product_traits<LhsScalar, RhsScalar>::ReturnType ResScalar;
   enum {
     IsLower = ((Mode&Lower)==Lower),
-    HasUnitDiag = (Mode & UnitDiag)==UnitDiag
+    HasUnitDiag = (Mode & UnitDiag)==UnitDiag,
+    HasZeroDiag = (Mode & ZeroDiag)==ZeroDiag
   };
-  static void run(Index rows, Index cols, const LhsScalar* _lhs, Index lhsStride,
+  static void run(Index _rows, Index _cols, const LhsScalar* _lhs, Index lhsStride,
                   const RhsScalar* _rhs, Index rhsIncr, ResScalar* _res, Index resIncr, ResScalar alpha)
   {
     static const Index PanelWidth = EIGEN_TUNE_TRIANGULAR_PANEL_WIDTH;
+    Index diagSize = (std::min)(_rows,_cols);
+    Index rows = IsLower ? _rows : diagSize;
+    Index cols = IsLower ? diagSize : _cols;
 
     typedef Map<const Matrix<LhsScalar,Dynamic,Dynamic,RowMajor>, 0, OuterStride<> > LhsMap;
     const LhsMap lhs(_lhs,rows,cols,OuterStride<>(lhsStride));
@@ -105,15 +121,15 @@ struct product_triangular_matrix_vector<Index,Mode,LhsScalar,ConjLhs,RhsScalar,C
     typedef Map<Matrix<ResScalar,Dynamic,1>, 0, InnerStride<> > ResMap;
     ResMap res(_res,rows,InnerStride<>(resIncr));
     
-    for (Index pi=0; pi<cols; pi+=PanelWidth)
+    for (Index pi=0; pi<diagSize; pi+=PanelWidth)
     {
-      Index actualPanelWidth = (std::min)(PanelWidth, cols-pi);
+      Index actualPanelWidth = (std::min)(PanelWidth, diagSize-pi);
       for (Index k=0; k<actualPanelWidth; ++k)
       {
         Index i = pi + k;
-        Index s = IsLower ? pi  : (HasUnitDiag ? i+1 : i);
+        Index s = IsLower ? pi  : ((HasUnitDiag||HasZeroDiag) ? i+1 : i);
         Index r = IsLower ? k+1 : actualPanelWidth-k;
-        if ((!HasUnitDiag) || (--r)>0)
+        if ((!(HasUnitDiag||HasZeroDiag)) || (--r)>0)
           res.coeffRef(i) += alpha * (cjLhs.row(i).segment(s,r).cwiseProduct(cjRhs.segment(s,r).transpose())).sum();
         if (HasUnitDiag)
           res.coeffRef(i) += alpha * cjRhs.coeff(i);
@@ -122,13 +138,21 @@ struct product_triangular_matrix_vector<Index,Mode,LhsScalar,ConjLhs,RhsScalar,C
       if (r>0)
       {
         Index s = IsLower ? 0 : pi + actualPanelWidth;
-        general_matrix_vector_product<Index,LhsScalar,RowMajor,ConjLhs,RhsScalar,ConjRhs>::run(
+        general_matrix_vector_product<Index,LhsScalar,RowMajor,ConjLhs,RhsScalar,ConjRhs,BuiltIn>::run(
             actualPanelWidth, r,
             &lhs.coeffRef(pi,s), lhsStride,
             &rhs.coeffRef(s), rhsIncr,
             &res.coeffRef(pi), resIncr, alpha);
       }
     }
+    if(IsLower && rows>diagSize)
+    {
+      general_matrix_vector_product<Index,LhsScalar,RowMajor,ConjLhs,RhsScalar,ConjRhs>::run(
+            rows-diagSize, cols,
+            &lhs.coeffRef(diagSize,0), lhsStride,
+            &rhs.coeffRef(0), rhsIncr,
+            &res.coeffRef(diagSize), resIncr, alpha);
+    }
   }
 };
 
@@ -180,7 +204,7 @@ struct TriangularProduct<Mode,false,Lhs,true,Rhs,false>
   {
     eigen_assert(dst.rows()==m_lhs.rows() && dst.cols()==m_rhs.cols());
 
-    typedef TriangularProduct<(Mode & UnitDiag) | ((Mode & Lower) ? Upper : Lower),true,Transpose<const Rhs>,false,Transpose<const Lhs>,true> TriangularProductTranspose;
+    typedef TriangularProduct<(Mode & (UnitDiag|ZeroDiag)) | ((Mode & Lower) ? Upper : Lower),true,Transpose<const Rhs>,false,Transpose<const Lhs>,true> TriangularProductTranspose;
     Transpose<Dest> dstT(dst);
     internal::trmv_selector<(int(internal::traits<Rhs>::Flags)&RowMajorBit) ? ColMajor : RowMajor>::run(
       TriangularProductTranspose(m_rhs.transpose(),m_lhs.transpose()), dstT, alpha);
@@ -208,8 +232,8 @@ template<> struct trmv_selector<ColMajor>
     typedef typename ProductType::RhsBlasTraits RhsBlasTraits;
     typedef Map<Matrix<ResScalar,Dynamic,1>, Aligned> MappedDest;
 
-    const ActualLhsType actualLhs = LhsBlasTraits::extract(prod.lhs());
-    const ActualRhsType actualRhs = RhsBlasTraits::extract(prod.rhs());
+    typename internal::add_const_on_value_type<ActualLhsType>::type actualLhs = LhsBlasTraits::extract(prod.lhs());
+    typename internal::add_const_on_value_type<ActualRhsType>::type actualRhs = RhsBlasTraits::extract(prod.rhs());
 
     ResScalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(prod.lhs())
                                   * RhsBlasTraits::extractScalarFactor(prod.rhs());
@@ -247,7 +271,7 @@ template<> struct trmv_selector<ColMajor>
         MappedDest(actualDestPtr, dest.size()) = dest;
     }
     
-    internal::product_triangular_matrix_vector
+    internal::triangular_matrix_vector_product
       <Index,Mode,
        LhsScalar, LhsBlasTraits::NeedToConjugate,
        RhsScalar, RhsBlasTraits::NeedToConjugate,
@@ -307,7 +331,7 @@ template<> struct trmv_selector<RowMajor>
       Map<typename _ActualRhsType::PlainObject>(actualRhsPtr, actualRhs.size()) = actualRhs;
     }
     
-    internal::product_triangular_matrix_vector
+    internal::triangular_matrix_vector_product
       <Index,Mode,
        LhsScalar, LhsBlasTraits::NeedToConjugate,
        RhsScalar, RhsBlasTraits::NeedToConjugate,

Eigen/src/Core/products/TriangularMatrixVector_MKL.h

@@ -0,0 +1,245 @@
+/*
+ Copyright (c) 2011, Intel Corporation. All rights reserved.
+
+ Redistribution and use in source and binary forms, with or without modification,
+ are permitted provided that the following conditions are met:
+
+ * Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+ * Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+ * Neither the name of Intel Corporation nor the names of its contributors may
+   be used to endorse or promote products derived from this software without
+   specific prior written permission.
+
+ THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+ WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+ DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+ (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+ LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+ ********************************************************************************
+ *   Content : Eigen bindings to Intel(R) MKL
+ *   Triangular matrix-vector product functionality based on ?TRMV.
+ ********************************************************************************
+*/
+
+#ifndef EIGEN_TRIANGULAR_MATRIX_VECTOR_MKL_H
+#define EIGEN_TRIANGULAR_MATRIX_VECTOR_MKL_H
+
+namespace internal {
+
+/**********************************************************************
+* This file implements triangular matrix-vector multiplication using BLAS
+**********************************************************************/
+
+// trmv/hemv specialization
+
+template<typename Index, int Mode, typename LhsScalar, bool ConjLhs, typename RhsScalar, bool ConjRhs, int StorageOrder>
+struct triangular_matrix_vector_product_trmv :
+  triangular_matrix_vector_product<Index,Mode,LhsScalar,ConjLhs,RhsScalar,ConjRhs,StorageOrder,BuiltIn> {};
+
+#define EIGEN_MKL_TRMV_SPECIALIZE(Scalar) \
+template<typename Index, int Mode, bool ConjLhs, bool ConjRhs> \
+struct triangular_matrix_vector_product<Index,Mode,Scalar,ConjLhs,Scalar,ConjRhs,ColMajor,Specialized> { \
+ static EIGEN_DONT_INLINE void run(Index _rows, Index _cols, const Scalar* _lhs, Index lhsStride, \
+                                     const Scalar* _rhs, Index rhsIncr, Scalar* _res, Index resIncr, Scalar alpha) { \
+      triangular_matrix_vector_product_trmv<Index,Mode,Scalar,ConjLhs,Scalar,ConjRhs,ColMajor>::run( \
+        _rows, _cols, _lhs, lhsStride, _rhs, rhsIncr, _res, resIncr, alpha); \
+  } \
+}; \
+template<typename Index, int Mode, bool ConjLhs, bool ConjRhs> \
+struct triangular_matrix_vector_product<Index,Mode,Scalar,ConjLhs,Scalar,ConjRhs,RowMajor,Specialized> { \
+ static EIGEN_DONT_INLINE void run(Index _rows, Index _cols, const Scalar* _lhs, Index lhsStride, \
+                                     const Scalar* _rhs, Index rhsIncr, Scalar* _res, Index resIncr, Scalar alpha) { \
+      triangular_matrix_vector_product_trmv<Index,Mode,Scalar,ConjLhs,Scalar,ConjRhs,RowMajor>::run( \
+        _rows, _cols, _lhs, lhsStride, _rhs, rhsIncr, _res, resIncr, alpha); \
+  } \
+};
+
+EIGEN_MKL_TRMV_SPECIALIZE(double)
+EIGEN_MKL_TRMV_SPECIALIZE(float)
+EIGEN_MKL_TRMV_SPECIALIZE(dcomplex)
+EIGEN_MKL_TRMV_SPECIALIZE(scomplex)
+
+// implements col-major: res += alpha * op(triangular) * vector
+#define EIGEN_MKL_TRMV_CM(EIGTYPE, MKLTYPE, EIGPREFIX, MKLPREFIX) \
+template<typename Index, int Mode, bool ConjLhs, bool ConjRhs> \
+struct triangular_matrix_vector_product_trmv<Index,Mode,EIGTYPE,ConjLhs,EIGTYPE,ConjRhs,ColMajor> { \
+  enum { \
+    IsLower = (Mode&Lower) == Lower, \
+    SetDiag = (Mode&(ZeroDiag|UnitDiag)) ? 0 : 1, \
+    IsUnitDiag  = (Mode&UnitDiag) ? 1 : 0, \
+    IsZeroDiag  = (Mode&ZeroDiag) ? 1 : 0, \
+    LowUp = IsLower ? Lower : Upper \
+  }; \
+ static EIGEN_DONT_INLINE void run(Index _rows, Index _cols, const EIGTYPE* _lhs, Index lhsStride, \
+                             const EIGTYPE* _rhs, Index rhsIncr, EIGTYPE* _res, Index resIncr, EIGTYPE alpha) \
+ { \
+   if (ConjLhs || IsZeroDiag) { \
+     triangular_matrix_vector_product<Index,Mode,EIGTYPE,ConjLhs,EIGTYPE,ConjRhs,ColMajor,BuiltIn>::run( \
+       _rows, _cols, _lhs, lhsStride, _rhs, rhsIncr, _res, resIncr, alpha); \
+     return; \
+   }\
+   Index size = (std::min)(_rows,_cols); \
+   Index rows = IsLower ? _rows : size; \
+   Index cols = IsLower ? size : _cols; \
+\
+   typedef VectorX##EIGPREFIX VectorRhs; \
+   EIGTYPE *x, *y;\
+\
+/* Set x*/ \
+   Map<const VectorRhs, 0, InnerStride<> > rhs(_rhs,cols,InnerStride<>(rhsIncr)); \
+   VectorRhs x_tmp; \
+   if (ConjRhs) x_tmp = rhs.conjugate(); else x_tmp = rhs; \
+   x = x_tmp.data(); \
+\
+/* Square part handling */\
+\
+   char trans, uplo, diag; \
+   MKL_INT m, n, k, lda, incx, incy; \
+   EIGTYPE const *a; \
+   MKLTYPE alpha_, beta_; \
+   assign_scalar_eig2mkl<MKLTYPE, EIGTYPE>(alpha_, alpha); \
+   assign_scalar_eig2mkl<MKLTYPE, EIGTYPE>(beta_, EIGTYPE(1)); \
+\
+/* Set m, n */ \
+   n = (MKL_INT)size; \
+   lda = lhsStride; \
+   incx = 1; \
+   incy = resIncr; \
+\
+/* Set uplo, trans and diag*/ \
+   trans = 'N'; \
+   uplo = IsLower ? 'L' : 'U'; \
+   diag = IsUnitDiag ? 'U' : 'N'; \
+\
+/* call ?TRMV*/ \
+      std::cout << "TRMV: CM\n";\
+   MKLPREFIX##trmv(&uplo, &trans, &diag, &n, (const MKLTYPE*)_lhs, &lda, (MKLTYPE*)x, &incx); \
+\
+/* Add op(a_tr)rhs into res*/ \
+   MKLPREFIX##axpy(&n, &alpha_,(const MKLTYPE*)x, &incx, (MKLTYPE*)_res, &incy); \
+/* Non-square case - doesn't fit to MKL ?TRMV. Fall to default triangular product*/ \
+   if (size<(std::max)(rows,cols)) { \
+     typedef Matrix<EIGTYPE, Dynamic, Dynamic> MatrixLhs; \
+     if (ConjRhs) x_tmp = rhs.conjugate(); else x_tmp = rhs; \
+     x = x_tmp.data(); \
+     if (size<rows) { \
+       y = _res + size*resIncr; \
+       a = _lhs + size; \
+       m = rows-size; \
+       n = size; \
+     } \
+     if (size<cols) { \
+       x += size; \
+       y = _res; \
+       a = _lhs + size*lda; \
+       m = size; \
+       n = cols-size; \
+     } \
+     MKLPREFIX##gemv(&trans, &m, &n, &alpha_, (const MKLTYPE*)a, &lda, (const MKLTYPE*)x, &incx, &beta_, (MKLTYPE*)y, &incy); \
+   } \
+  } \
+};
+
+EIGEN_MKL_TRMV_CM(double, double, d, d)
+EIGEN_MKL_TRMV_CM(dcomplex, MKL_Complex16, cd, z)
+EIGEN_MKL_TRMV_CM(float, float, f, s)
+EIGEN_MKL_TRMV_CM(scomplex, MKL_Complex8, cf, c)
+
+// implements row-major: res += alpha * op(triangular) * vector
+#define EIGEN_MKL_TRMV_RM(EIGTYPE, MKLTYPE, EIGPREFIX, MKLPREFIX) \
+template<typename Index, int Mode, bool ConjLhs, bool ConjRhs> \
+struct triangular_matrix_vector_product_trmv<Index,Mode,EIGTYPE,ConjLhs,EIGTYPE,ConjRhs,RowMajor> { \
+  enum { \
+    IsLower = (Mode&Lower) == Lower, \
+    SetDiag = (Mode&(ZeroDiag|UnitDiag)) ? 0 : 1, \
+    IsUnitDiag  = (Mode&UnitDiag) ? 1 : 0, \
+    IsZeroDiag  = (Mode&ZeroDiag) ? 1 : 0, \
+    LowUp = IsLower ? Lower : Upper \
+  }; \
+ static EIGEN_DONT_INLINE void run(Index _rows, Index _cols, const EIGTYPE* _lhs, Index lhsStride, \
+                             const EIGTYPE* _rhs, Index rhsIncr, EIGTYPE* _res, Index resIncr, EIGTYPE alpha) \
+ { \
+   if (IsZeroDiag) { \
+     triangular_matrix_vector_product<Index,Mode,EIGTYPE,ConjLhs,EIGTYPE,ConjRhs,RowMajor,BuiltIn>::run( \
+       _rows, _cols, _lhs, lhsStride, _rhs, rhsIncr, _res, resIncr, alpha); \
+     return; \
+   }\
+   Index size = (std::min)(_rows,_cols); \
+   Index rows = IsLower ? _rows : size; \
+   Index cols = IsLower ? size : _cols; \
+\
+   typedef VectorX##EIGPREFIX VectorRhs; \
+   EIGTYPE *x, *y;\
+\
+/* Set x*/ \
+   Map<const VectorRhs, 0, InnerStride<> > rhs(_rhs,cols,InnerStride<>(rhsIncr)); \
+   VectorRhs x_tmp; \
+   if (ConjRhs) x_tmp = rhs.conjugate(); else x_tmp = rhs; \
+   x = x_tmp.data(); \
+\
+/* Square part handling */\
+\
+   char trans, uplo, diag; \
+   MKL_INT m, n, k, lda, incx, incy; \
+   EIGTYPE const *a; \
+   MKLTYPE alpha_, beta_; \
+   assign_scalar_eig2mkl<MKLTYPE, EIGTYPE>(alpha_, alpha); \
+   assign_scalar_eig2mkl<MKLTYPE, EIGTYPE>(beta_, EIGTYPE(1)); \
+\
+/* Set m, n */ \
+   n = (MKL_INT)size; \
+   lda = lhsStride; \
+   incx = 1; \
+   incy = resIncr; \
+\
+/* Set uplo, trans and diag*/ \
+   trans = ConjLhs ? 'C' : 'T'; \
+   uplo = IsLower ? 'U' : 'L'; \
+   diag = IsUnitDiag ? 'U' : 'N'; \
+\
+/* call ?TRMV*/ \
+      std::cout << "TRMV: RM\n";\
+   MKLPREFIX##trmv(&uplo, &trans, &diag, &n, (const MKLTYPE*)_lhs, &lda, (MKLTYPE*)x, &incx); \
+\
+/* Add op(a_tr)rhs into res*/ \
+   MKLPREFIX##axpy(&n, &alpha_,(const MKLTYPE*)x, &incx, (MKLTYPE*)_res, &incy); \
+/* Non-square case - doesn't fit to MKL ?TRMV. Fall to default triangular product*/ \
+   if (size<(std::max)(rows,cols)) { \
+     typedef Matrix<EIGTYPE, Dynamic, Dynamic> MatrixLhs; \
+     if (ConjRhs) x_tmp = rhs.conjugate(); else x_tmp = rhs; \
+     x = x_tmp.data(); \
+     if (size<rows) { \
+       y = _res + size*resIncr; \
+       a = _lhs + size*lda; \
+       m = rows-size; \
+       n = size; \
+     } \
+     if (size<cols) { \
+       x += size; \
+       y = _res; \
+       a = _lhs + size; \
+       m = size; \
+       n = cols-size; \
+     } \
+     MKLPREFIX##gemv(&trans, &n, &m, &alpha_, (const MKLTYPE*)a, &lda, (const MKLTYPE*)x, &incx, &beta_, (MKLTYPE*)y, &incy); \
+   } \
+  } \
+};
+
+EIGEN_MKL_TRMV_RM(double, double, d, d)
+EIGEN_MKL_TRMV_RM(dcomplex, MKL_Complex16, cd, z)
+EIGEN_MKL_TRMV_RM(float, float, f, s)
+EIGEN_MKL_TRMV_RM(scomplex, MKL_Complex8, cf, c)
+
+} //end of namespase
+
+#endif // EIGEN_TRIANGULAR_MATRIX_VECTOR_MKL_H

Eigen/src/Core/products/TriangularSolverMatrix.h

@@ -75,12 +75,20 @@ struct triangular_solve_matrix<Scalar,Index,OnTheLeft,Mode,Conjugate,TriStorageO
     ei_declare_aligned_stack_constructed_variable(Scalar, blockA, kc*mc, 0);
     ei_declare_aligned_stack_constructed_variable(Scalar, allocatedBlockB, sizeB, 0);
     Scalar* blockB = allocatedBlockB + sizeW;
+    Scalar* blockW = allocatedBlockB;
 
     conj_if<Conjugate> conj;
     gebp_kernel<Scalar, Scalar, Index, Traits::mr, Traits::nr, Conjugate, false> gebp_kernel;
     gemm_pack_lhs<Scalar, Index, Traits::mr, Traits::LhsProgress, TriStorageOrder> pack_lhs;
     gemm_pack_rhs<Scalar, Index, Traits::nr, ColMajor, false, true> pack_rhs;
 
+    // the goal here is to subdivise the Rhs panels such that we keep some cache
+    // coherence when accessing the rhs elements
+    std::ptrdiff_t l1, l2;
+    manage_caching_sizes(GetAction, &l1, &l2);
+    Index subcols = cols>0 ? l2/(4 * sizeof(Scalar) * otherStride) : 0;
+    subcols = std::max<Index>((subcols/Traits::nr)*Traits::nr, Traits::nr);
+
     for(Index k2=IsLower ? 0 : size;
         IsLower ? k2<size : k2>0;
         IsLower ? k2+=kc : k2-=kc)
@@ -92,16 +100,18 @@ struct triangular_solve_matrix<Scalar,Index,OnTheLeft,Mode,Conjugate,TriStorageO
       // A11 (the triangular part) and A21 the remaining rectangular part.
       // Then the high level algorithm is:
       //  - B = R1                    => general block copy (done during the next step)
-      //  - R1 = L1^-1 B              => tricky part
+      //  - R1 = A11^-1 B             => tricky part
       //  - update B from the new R1  => actually this has to be performed continuously during the above step
-      //  - R2 = L2 * B               => GEPP
+      //  - R2 -= A21 * B             => GEPP
 
-      // The tricky part: compute R1 = L1^-1 B while updating B from R1
-      // The idea is to split L1 into multiple small vertical panels.
-      // Each panel can be split into a small triangular part A1 which is processed without optimization,
-      // and the remaining small part A2 which is processed using gebp with appropriate block strides
+      // The tricky part: compute R1 = A11^-1 B while updating B from R1
+      // The idea is to split A11 into multiple small vertical panels.
+      // Each panel can be split into a small triangular part T1k which is processed without optimization,
+      // and the remaining small part T2k which is processed using gebp with appropriate block strides
+      for(Index j2=0; j2<cols; j2+=subcols)
       {
-        // for each small vertical panels of lhs
+        Index actual_cols = (std::min)(cols-j2,subcols);
+        // for each small vertical panels [T1k^T, T2k^T]^T of lhs
         for (Index k1=0; k1<actual_kc; k1+=SmallPanelWidth)
         {
           Index actualPanelWidth = std::min<Index>(actual_kc-k1, SmallPanelWidth);
@@ -114,11 +124,11 @@ struct triangular_solve_matrix<Scalar,Index,OnTheLeft,Mode,Conjugate,TriStorageO
             Index rs = actualPanelWidth - k - 1; // remaining size
 
             Scalar a = (Mode & UnitDiag) ? Scalar(1) : Scalar(1)/conj(tri(i,i));
-            for (Index j=0; j<cols; ++j)
+            for (Index j=j2; j<j2+actual_cols; ++j)
             {
               if (TriStorageOrder==RowMajor)
               {
-                Scalar b = 0;
+                Scalar b(0);
                 const Scalar* l = &tri(i,s);
                 Scalar* r = &other(s,j);
                 for (Index i3=0; i3<k; ++i3)
@@ -143,7 +153,7 @@ struct triangular_solve_matrix<Scalar,Index,OnTheLeft,Mode,Conjugate,TriStorageO
           Index blockBOffset = IsLower ? k1 : lengthTarget;
 
           // update the respective rows of B from other
-          pack_rhs(blockB, _other+startBlock, otherStride, actualPanelWidth, cols, actual_kc, blockBOffset);
+          pack_rhs(blockB+actual_kc*j2, &other(startBlock,j2), otherStride, actualPanelWidth, actual_cols, actual_kc, blockBOffset);
 
           // GEBP
           if (lengthTarget>0)
@@ -152,13 +162,13 @@ struct triangular_solve_matrix<Scalar,Index,OnTheLeft,Mode,Conjugate,TriStorageO
 
             pack_lhs(blockA, &tri(startTarget,startBlock), triStride, actualPanelWidth, lengthTarget);
 
-            gebp_kernel(_other+startTarget, otherStride, blockA, blockB, lengthTarget, actualPanelWidth, cols, Scalar(-1),
-                        actualPanelWidth, actual_kc, 0, blockBOffset);
+            gebp_kernel(&other(startTarget,j2), otherStride, blockA, blockB+actual_kc*j2, lengthTarget, actualPanelWidth, actual_cols, Scalar(-1),
+                        actualPanelWidth, actual_kc, 0, blockBOffset, blockW);
           }
         }
       }
-
-      // R2 = A2 * B => GEPP
+      
+      // R2 -= A21 * B => GEPP
       {
         Index start = IsLower ? k2+kc : 0;
         Index end   = IsLower ? size : k2-kc;

Eigen/src/Core/products/TriangularSolverMatrix_MKL.h

@@ -0,0 +1,151 @@
+/*
+ Copyright (c) 2011, Intel Corporation. All rights reserved.
+
+ Redistribution and use in source and binary forms, with or without modification,
+ are permitted provided that the following conditions are met:
+
+ * Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+ * Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+ * Neither the name of Intel Corporation nor the names of its contributors may
+   be used to endorse or promote products derived from this software without
+   specific prior written permission.
+
+ THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+ WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+ DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+ (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+ LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+ ********************************************************************************
+ *   Content : Eigen bindings to Intel(R) MKL
+ *   Triangular matrix * matrix product functionality based on ?TRMM.
+ ********************************************************************************
+*/
+
+#ifndef EIGEN_TRIANGULAR_SOLVER_MATRIX_MKL_H
+#define EIGEN_TRIANGULAR_SOLVER_MATRIX_MKL_H
+
+namespace internal {
+
+// implements LeftSide op(triangular)^-1 * general
+#define EIGEN_MKL_TRSM_L(EIGTYPE, MKLTYPE, MKLPREFIX) \
+template <typename Index, int Mode, bool Conjugate, int TriStorageOrder> \
+struct triangular_solve_matrix<EIGTYPE,Index,OnTheLeft,Mode,Conjugate,TriStorageOrder,ColMajor> \
+{ \
+  enum { \
+    IsLower = (Mode&Lower) == Lower, \
+    IsUnitDiag  = (Mode&UnitDiag) ? 1 : 0, \
+    IsZeroDiag  = (Mode&ZeroDiag) ? 1 : 0, \
+    conjA = ((TriStorageOrder==ColMajor) && Conjugate) ? 1 : 0 \
+  }; \
+  static EIGEN_DONT_INLINE void run( \
+      Index size, Index otherSize, \
+      const EIGTYPE* _tri, Index triStride, \
+      EIGTYPE* _other, Index otherStride) \
+  { \
+   MKL_INT m = size, n = otherSize, lda, ldb; \
+   char side = 'L', uplo, diag='N', transa; \
+   /* Set alpha_ */ \
+   MKLTYPE alpha; \
+   EIGTYPE myone(1); \
+   assign_scalar_eig2mkl(alpha, myone); \
+   ldb = otherStride;\
+\
+   const EIGTYPE *a; \
+/* Set trans */ \
+   transa = (TriStorageOrder==RowMajor) ? ((Conjugate) ? 'C' : 'T') : 'N'; \
+/* Set uplo */ \
+   uplo = IsLower ? 'L' : 'U'; \
+   if (TriStorageOrder==RowMajor) uplo = (uplo == 'L') ? 'U' : 'L'; \
+/* Set a, lda */ \
+   typedef Matrix<EIGTYPE, Dynamic, Dynamic, TriStorageOrder> MatrixTri; \
+   Map<const MatrixTri, 0, OuterStride<> > tri(_tri,size,size,OuterStride<>(triStride)); \
+   MatrixTri a_tmp; \
+\
+   if (conjA) { \
+     a_tmp = tri.conjugate(); \
+     a = a_tmp.data(); \
+     lda = a_tmp.outerStride(); \
+   } else { \
+     a = _tri; \
+     lda = triStride; \
+   } \
+   if (IsUnitDiag) diag='U'; \
+/* call ?trsm*/ \
+   MKLPREFIX##trsm(&side, &uplo, &transa, &diag, &m, &n, &alpha, (const MKLTYPE*)a, &lda, (MKLTYPE*)_other, &ldb); \
+ } \
+};
+
+EIGEN_MKL_TRSM_L(double, double, d)
+EIGEN_MKL_TRSM_L(dcomplex, MKL_Complex16, z)
+EIGEN_MKL_TRSM_L(float, float, s)
+EIGEN_MKL_TRSM_L(scomplex, MKL_Complex8, c)
+
+
+// implements RightSide general * op(triangular)^-1
+#define EIGEN_MKL_TRSM_R(EIGTYPE, MKLTYPE, MKLPREFIX) \
+template <typename Index, int Mode, bool Conjugate, int TriStorageOrder> \
+struct triangular_solve_matrix<EIGTYPE,Index,OnTheRight,Mode,Conjugate,TriStorageOrder,ColMajor> \
+{ \
+  enum { \
+    IsLower = (Mode&Lower) == Lower, \
+    IsUnitDiag  = (Mode&UnitDiag) ? 1 : 0, \
+    IsZeroDiag  = (Mode&ZeroDiag) ? 1 : 0, \
+    conjA = ((TriStorageOrder==ColMajor) && Conjugate) ? 1 : 0 \
+  }; \
+  static EIGEN_DONT_INLINE void run( \
+      Index size, Index otherSize, \
+      const EIGTYPE* _tri, Index triStride, \
+      EIGTYPE* _other, Index otherStride) \
+  { \
+   MKL_INT m = otherSize, n = size, lda, ldb; \
+   char side = 'R', uplo, diag='N', transa; \
+   /* Set alpha_ */ \
+   MKLTYPE alpha; \
+   EIGTYPE myone(1); \
+   assign_scalar_eig2mkl(alpha, myone); \
+   ldb = otherStride;\
+\
+   const EIGTYPE *a; \
+/* Set trans */ \
+   transa = (TriStorageOrder==RowMajor) ? ((Conjugate) ? 'C' : 'T') : 'N'; \
+/* Set uplo */ \
+   uplo = IsLower ? 'L' : 'U'; \
+   if (TriStorageOrder==RowMajor) uplo = (uplo == 'L') ? 'U' : 'L'; \
+/* Set a, lda */ \
+   typedef Matrix<EIGTYPE, Dynamic, Dynamic, TriStorageOrder> MatrixTri; \
+   Map<const MatrixTri, 0, OuterStride<> > tri(_tri,size,size,OuterStride<>(triStride)); \
+   MatrixTri a_tmp; \
+\
+   if (conjA) { \
+     a_tmp = tri.conjugate(); \
+     a = a_tmp.data(); \
+     lda = a_tmp.outerStride(); \
+   } else { \
+     a = _tri; \
+     lda = triStride; \
+   } \
+   if (IsUnitDiag) diag='U'; \
+/* call ?trsm*/ \
+   MKLPREFIX##trsm(&side, &uplo, &transa, &diag, &m, &n, &alpha, (const MKLTYPE*)a, &lda, (MKLTYPE*)_other, &ldb); \
+   /*std::cout << "TRMS_L specialization!\n";*/ \
+ } \
+};
+
+EIGEN_MKL_TRSM_R(double, double, d)
+EIGEN_MKL_TRSM_R(dcomplex, MKL_Complex16, z)
+EIGEN_MKL_TRSM_R(float, float, s)
+EIGEN_MKL_TRSM_R(scomplex, MKL_Complex8, c)
+
+
+} // end namespace internal
+
+#endif // EIGEN_TRIANGULAR_SOLVER_MATRIX_MKL_H

Eigen/src/Core/util/BlasUtil.h

@@ -47,7 +47,7 @@ template<
   int ResStorageOrder>
 struct general_matrix_matrix_product;
 
-template<typename Index, typename LhsScalar, int LhsStorageOrder, bool ConjugateLhs, typename RhsScalar, bool ConjugateRhs>
+template<typename Index, typename LhsScalar, int LhsStorageOrder, bool ConjugateLhs, typename RhsScalar, bool ConjugateRhs, int Version=Specialized>
 struct general_matrix_vector_product;
 
 
@@ -56,11 +56,15 @@ template<bool Conjugate> struct conj_if;
 template<> struct conj_if<true> {
   template<typename T>
   inline T operator()(const T& x) { return conj(x); }
+  template<typename T>
+  inline T pconj(const T& x) { return internal::pconj(x); }
 };
 
 template<> struct conj_if<false> {
   template<typename T>
   inline const T& operator()(const T& x) { return x; }
+  template<typename T>
+  inline const T& pconj(const T& x) { return x; }
 };
 
 template<typename Scalar> struct conj_helper<Scalar,Scalar,false,false>
@@ -118,11 +122,11 @@ template<typename RealScalar,bool Conj> struct conj_helper<RealScalar, std::comp
 };
 
 template<typename From,typename To> struct get_factor {
-  EIGEN_STRONG_INLINE static To run(const From& x) { return x; }
+  static EIGEN_STRONG_INLINE To run(const From& x) { return x; }
 };
 
 template<typename Scalar> struct get_factor<Scalar,typename NumTraits<Scalar>::Real> {
-  EIGEN_STRONG_INLINE static typename NumTraits<Scalar>::Real run(const Scalar& x) { return real(x); }
+  static EIGEN_STRONG_INLINE typename NumTraits<Scalar>::Real run(const Scalar& x) { return real(x); }
 };
 
 // Lightweight helper class to access matrix coefficients.
@@ -175,7 +179,7 @@ template<typename XprType> struct blas_traits
     ExtractType,
     typename _ExtractType::PlainObject
     >::type DirectLinearAccessType;
-  static inline const ExtractType extract(const XprType& x) { return x; }
+  static inline ExtractType extract(const XprType& x) { return x; }
   static inline const Scalar extractScalarFactor(const XprType&) { return Scalar(1); }
 };
 
@@ -192,7 +196,7 @@ struct blas_traits<CwiseUnaryOp<scalar_conjugate_op<Scalar>, NestedXpr> >
     IsComplex = NumTraits<Scalar>::IsComplex,
     NeedToConjugate = Base::NeedToConjugate ? 0 : IsComplex
   };
-  static inline const ExtractType extract(const XprType& x) { return Base::extract(x.nestedExpression()); }
+  static inline ExtractType extract(const XprType& x) { return Base::extract(x.nestedExpression()); }
   static inline Scalar extractScalarFactor(const XprType& x) { return conj(Base::extractScalarFactor(x.nestedExpression())); }
 };
 
@@ -204,7 +208,7 @@ struct blas_traits<CwiseUnaryOp<scalar_multiple_op<Scalar>, NestedXpr> >
   typedef blas_traits<NestedXpr> Base;
   typedef CwiseUnaryOp<scalar_multiple_op<Scalar>, NestedXpr> XprType;
   typedef typename Base::ExtractType ExtractType;
-  static inline const ExtractType extract(const XprType& x) { return Base::extract(x.nestedExpression()); }
+  static inline ExtractType extract(const XprType& x) { return Base::extract(x.nestedExpression()); }
   static inline Scalar extractScalarFactor(const XprType& x)
   { return x.functor().m_other * Base::extractScalarFactor(x.nestedExpression()); }
 };
@@ -217,7 +221,7 @@ struct blas_traits<CwiseUnaryOp<scalar_opposite_op<Scalar>, NestedXpr> >
   typedef blas_traits<NestedXpr> Base;
   typedef CwiseUnaryOp<scalar_opposite_op<Scalar>, NestedXpr> XprType;
   typedef typename Base::ExtractType ExtractType;
-  static inline const ExtractType extract(const XprType& x) { return Base::extract(x.nestedExpression()); }
+  static inline ExtractType extract(const XprType& x) { return Base::extract(x.nestedExpression()); }
   static inline Scalar extractScalarFactor(const XprType& x)
   { return - Base::extractScalarFactor(x.nestedExpression()); }
 };
@@ -239,7 +243,7 @@ struct blas_traits<Transpose<NestedXpr> >
   enum {
     IsTransposed = Base::IsTransposed ? 0 : 1
   };
-  static inline const ExtractType extract(const XprType& x) { return Base::extract(x.nestedExpression()); }
+  static inline ExtractType extract(const XprType& x) { return Base::extract(x.nestedExpression()); }
   static inline Scalar extractScalarFactor(const XprType& x) { return Base::extractScalarFactor(x.nestedExpression()); }
 };
 
@@ -252,7 +256,7 @@ template<typename T, bool HasUsableDirectAccess=blas_traits<T>::HasUsableDirectA
 struct extract_data_selector {
   static const typename T::Scalar* run(const T& m)
   {
-    return const_cast<typename T::Scalar*>(&blas_traits<T>::extract(m).coeffRef(0,0)); // FIXME this should be .data()
+    return blas_traits<T>::extract(m).data();
   }
 };
 

Eigen/src/Core/util/Constants.h

@@ -188,7 +188,9 @@ enum {
   /** View matrix as an upper triangular matrix with zeros on the diagonal. */
   StrictlyUpper=ZeroDiag|Upper,
   /** Used in BandMatrix and SelfAdjointView to indicate that the matrix is self-adjoint. */
-  SelfAdjoint=0x10
+  SelfAdjoint=0x10,
+  /** Used to support symmetric, non-selfadjoint, complex matrices. */
+  Symmetric=0x20
 };
 
 /** \ingroup enums
@@ -200,8 +202,6 @@ enum {
   Aligned=1 
 };
 
-enum { ConditionalJumpCost = 5 };
-
 /** \ingroup enums
  * Enum used by DenseBase::corner() in Eigen2 compatibility mode. */
 // FIXME after the corner() API change, this was not needed anymore, except by AlignedBox
@@ -223,8 +223,6 @@ enum DirectionType {
   BothDirections 
 };
 
-enum ProductEvaluationMode { NormalProduct, CacheFriendlyProduct };
-
 /** \internal \ingroup enums
   * Enum to specify how to traverse the entries of a matrix. */
 enum {
@@ -257,6 +255,13 @@ enum {
   CompleteUnrolling
 };
 
+/** \internal \ingroup enums
+  * Enum to specify whether to use the default (built-in) implementation or the specialization. */
+enum {
+  Specialized,
+  BuiltIn
+};
+
 /** \ingroup enums
   * Enum containing possible values for the \p _Options template parameter of
   * Matrix, Array and BandMatrix. */
@@ -375,7 +380,7 @@ enum QRPreconditioners {
 #error The preprocessor symbol 'Success' is defined, possibly by the X11 header file X.h
 #endif
 
-/** \ingroups enums
+/** \ingroup enums
   * Enum for reporting the status of a computation. */
 enum ComputationInfo {
   /** Computation was successful. */
@@ -383,7 +388,10 @@ enum ComputationInfo {
   /** The provided data did not satisfy the prerequisites. */
   NumericalIssue = 1, 
   /** Iterative procedure did not converge. */
-  NoConvergence = 2
+  NoConvergence = 2,
+  /** The inputs are invalid, or the algorithm has been improperly called.
+    * When assertions are enabled, such errors trigger an assert. */
+  InvalidInput = 3
 };
 
 /** \ingroup enums

Eigen/src/Core/util/DisableStupidWarnings.h

@@ -21,15 +21,13 @@
 #elif defined __INTEL_COMPILER
   // 2196 - routine is both "inline" and "noinline" ("noinline" assumed)
   //        ICC 12 generates this warning even without any inline keyword, when defining class methods 'inline' i.e. inside of class body
-  // 2536 - type qualifiers are meaningless here
-  //        ICC 12 generates this warning when a function return type is const qualified, even if that type is a template-parameter-dependent
   //        typedef that may be a reference type.
   // 279  - controlling expression is constant
   //        ICC 12 generates this warning on assert(constant_expression_depending_on_template_params) and frankly this is a legitimate use case.
   #ifndef EIGEN_PERMANENTLY_DISABLE_STUPID_WARNINGS
     #pragma warning push
   #endif
-  #pragma warning disable 2196 2536 279
+  #pragma warning disable 2196 279
 #elif defined __clang__
   // -Wconstant-logical-operand - warning: use of logical && with constant operand; switch to bitwise & or remove constant
   //     this is really a stupid warning as it warns on compile-time expressions involving enums

Eigen/src/Core/util/ForwardDeclarations.h

@@ -133,6 +133,7 @@ template<typename ExpressionType> class WithFormat;
 template<typename MatrixType> struct CommaInitializer;
 template<typename Derived> class ReturnByValue;
 template<typename ExpressionType> class ArrayWrapper;
+template<typename ExpressionType> class MatrixWrapper;
 
 namespace internal {
 template<typename DecompositionType, typename Rhs> struct solve_retval_base;
@@ -282,6 +283,8 @@ template<typename MatrixType,int Direction> class Homogeneous;
 // MatrixFunctions module
 template<typename Derived> struct MatrixExponentialReturnValue;
 template<typename Derived> class MatrixFunctionReturnValue;
+template<typename Derived> class MatrixSquareRootReturnValue;
+template<typename Derived> class MatrixLogarithmReturnValue;
 
 namespace internal {
 template <typename Scalar>

Eigen/src/Core/util/MKL_support.h

@@ -0,0 +1,109 @@
+/*
+ Copyright (c) 2011, Intel Corporation. All rights reserved.
+
+ Redistribution and use in source and binary forms, with or without modification,
+ are permitted provided that the following conditions are met:
+
+ * Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+ * Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+ * Neither the name of Intel Corporation nor the names of its contributors may
+   be used to endorse or promote products derived from this software without
+   specific prior written permission.
+
+ THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+ WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+ DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+ (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+ LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+ ********************************************************************************
+ *   Content : Eigen bindings to Intel(R) MKL
+ *   Include file with common MKL declarations
+ ********************************************************************************
+*/
+
+#ifndef EIGEN_MKL_SUPPORT_H
+#define EIGEN_MKL_SUPPORT_H
+
+#ifdef EIGEN_USE_MKL_ALL
+  #ifndef EIGEN_USE_BLAS
+    #define EIGEN_USE_BLAS
+  #endif
+  #ifndef EIGEN_USE_LAPACKE
+    #define EIGEN_USE_LAPACKE
+  #endif
+  #ifndef EIGEN_USE_MKL_VML
+    #define EIGEN_USE_MKL_VML
+  #endif
+#endif
+
+#ifdef EIGEN_USE_LAPACKE_STRICT
+  #define EIGEN_USE_LAPACKE
+#endif
+
+#if defined(EIGEN_USE_BLAS) || defined(EIGEN_USE_LAPACKE) || defined(EIGEN_USE_MKL_VML)
+  #define EIGEN_USE_MKL
+#endif
+
+#if defined EIGEN_USE_MKL
+
+#include <mkl.h>
+#include <mkl_lapacke.h>
+#define EIGEN_MKL_VML_THRESHOLD 128
+
+namespace Eigen {
+
+typedef std::complex<double> dcomplex;
+typedef std::complex<float>  scomplex;
+
+namespace internal {
+
+template<typename MKLType, typename EigenType>
+static inline void assign_scalar_eig2mkl(MKLType& mklScalar, const EigenType& eigenScalar) {
+  mklScalar=eigenScalar;
+}
+
+template<typename MKLType, typename EigenType>
+static inline void assign_conj_scalar_eig2mkl(MKLType& mklScalar, const EigenType& eigenScalar) {
+  mklScalar=eigenScalar;
+}
+
+template <>
+inline void assign_scalar_eig2mkl<MKL_Complex16,dcomplex>(MKL_Complex16& mklScalar, const dcomplex& eigenScalar) {
+  mklScalar.real=eigenScalar.real();
+  mklScalar.imag=eigenScalar.imag();
+}
+
+template <>
+inline void assign_scalar_eig2mkl<MKL_Complex8,scomplex>(MKL_Complex8& mklScalar, const scomplex& eigenScalar) {
+  mklScalar.real=eigenScalar.real();
+  mklScalar.imag=eigenScalar.imag();
+}
+
+template <>
+inline void assign_conj_scalar_eig2mkl<MKL_Complex16,dcomplex>(MKL_Complex16& mklScalar, const dcomplex& eigenScalar) {
+  mklScalar.real=eigenScalar.real();
+  mklScalar.imag=-eigenScalar.imag();
+}
+
+template <>
+inline void assign_conj_scalar_eig2mkl<MKL_Complex8,scomplex>(MKL_Complex8& mklScalar, const scomplex& eigenScalar) {
+  mklScalar.real=eigenScalar.real();
+  mklScalar.imag=-eigenScalar.imag();
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif
+
+#endif // EIGEN_MKL_SUPPORT_H

Eigen/src/Core/util/Macros.h

@@ -28,7 +28,7 @@
 
 #define EIGEN_WORLD_VERSION 3
 #define EIGEN_MAJOR_VERSION 0
-#define EIGEN_MINOR_VERSION 4
+#define EIGEN_MINOR_VERSION 92
 
 #define EIGEN_VERSION_AT_LEAST(x,y,z) (EIGEN_WORLD_VERSION>x || (EIGEN_WORLD_VERSION>=x && \
                                       (EIGEN_MAJOR_VERSION>y || (EIGEN_MAJOR_VERSION>=y && \
@@ -234,12 +234,16 @@
 #define EIGEN_ONLY_USED_FOR_DEBUG(x)
 #endif
 
-#if (defined __GNUC__)
-#define EIGEN_DEPRECATED __attribute__((deprecated))
-#elif (defined _MSC_VER)
-#define EIGEN_DEPRECATED __declspec(deprecated)
+#ifndef EIGEN_NO_DEPRECATED_WARNING
+  #if (defined __GNUC__)
+    #define EIGEN_DEPRECATED __attribute__((deprecated))
+  #elif (defined _MSC_VER)
+    #define EIGEN_DEPRECATED __declspec(deprecated)
+  #else
+    #define EIGEN_DEPRECATED
+  #endif
 #else
-#define EIGEN_DEPRECATED
+  #define EIGEN_DEPRECATED
 #endif
 
 #if (defined __GNUC__)
@@ -251,7 +255,7 @@
 // Suppresses 'unused variable' warnings.
 #define EIGEN_UNUSED_VARIABLE(var) (void)var;
 
-#if (defined __GNUC__)
+#if !defined(EIGEN_ASM_COMMENT) && (defined __GNUC__)
 #define EIGEN_ASM_COMMENT(X)  asm("#" X)
 #else
 #define EIGEN_ASM_COMMENT(X)

Eigen/src/Core/util/Memory.h

@@ -457,7 +457,7 @@ template<typename T, bool Align> inline void conditional_aligned_delete_auto(T *
   * There is also the variant first_aligned(const MatrixBase&) defined in DenseCoeffsBase.h.
   */
 template<typename Scalar, typename Index>
-inline static Index first_aligned(const Scalar* array, Index size)
+static inline Index first_aligned(const Scalar* array, Index size)
 {
   typedef typename packet_traits<Scalar>::type Packet;
   enum { PacketSize = packet_traits<Scalar>::size,
@@ -483,6 +483,27 @@ inline static Index first_aligned(const Scalar* array, Index size)
   }
 }
 
+
+// std::copy is much slower than memcpy, so let's introduce a smart_copy which
+// use memcpy on trivial types, i.e., on types that does not require an initialization ctor.
+template<typename T, bool UseMemcpy> struct smart_copy_helper;
+
+template<typename T> void smart_copy(const T* start, const T* end, T* target)
+{
+  smart_copy_helper<T,!NumTraits<T>::RequireInitialization>::run(start, end, target);
+}
+
+template<typename T> struct smart_copy_helper<T,true> {
+  static inline void run(const T* start, const T* end, T* target)
+  { memcpy(target, start, std::ptrdiff_t(end)-std::ptrdiff_t(start)); }
+};
+
+template<typename T> struct smart_copy_helper<T,false> {
+  static inline void run(const T* start, const T* end, T* target)
+  { std::copy(start, end, target); }
+};
+
+
 } // end namespace internal
 
 /*****************************************************************************
@@ -538,7 +559,7 @@ template<typename T> class aligned_stack_memory_handler
   * if SIZE is smaller than EIGEN_STACK_ALLOCATION_LIMIT, and if stack allocation is supported by the platform
   * (currently, this is Linux and Visual Studio only). Otherwise the memory is allocated on the heap.
   * The allocated buffer is automatically deleted when exiting the scope of this declaration.
-  * If BUFFER is non nul, then the declared variable is simply an alias for BUFFER, and no allocation/deletion occurs.
+  * If BUFFER is non null, then the declared variable is simply an alias for BUFFER, and no allocation/deletion occurs.
   * Here is an example:
   * \code
   * {
@@ -619,7 +640,7 @@ template<typename T> class aligned_stack_memory_handler
 
 #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(true)
 #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(Scalar,Size) \
-  EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(((Size)!=Eigen::Dynamic) && ((sizeof(Scalar)*(Size))%16==0))
+  EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(bool(((Size)!=Eigen::Dynamic) && ((sizeof(Scalar)*(Size))%16==0)))
 
 /****************************************************************************/
 
@@ -667,24 +688,24 @@ public:
         return &value;
     }
 
-    aligned_allocator() throw()
+    aligned_allocator()
     {
     }
 
-    aligned_allocator( const aligned_allocator& ) throw()
+    aligned_allocator( const aligned_allocator& )
     {
     }
 
     template<class U>
-    aligned_allocator( const aligned_allocator<U>& ) throw()
+    aligned_allocator( const aligned_allocator<U>& )
     {
     }
 
-    ~aligned_allocator() throw()
+    ~aligned_allocator()
     {
     }
 
-    size_type max_size() const throw()
+    size_type max_size() const
     {
         return (std::numeric_limits<size_type>::max)();
     }
@@ -720,19 +741,21 @@ public:
 
 //---------- Cache sizes ----------
 
-#if defined(__GNUC__) && ( defined(__i386__) || defined(__x86_64__) )
-#  if defined(__PIC__) && defined(__i386__)
-     // Case for x86 with PIC
-#    define EIGEN_CPUID(abcd,func,id) \
-       __asm__ __volatile__ ("xchgl %%ebx, %%esi;cpuid; xchgl %%ebx,%%esi": "=a" (abcd[0]), "=S" (abcd[1]), "=c" (abcd[2]), "=d" (abcd[3]) : "a" (func), "c" (id));
-#  else
-     // Case for x86_64 or x86 w/o PIC
-#    define EIGEN_CPUID(abcd,func,id) \
-       __asm__ __volatile__ ("cpuid": "=a" (abcd[0]), "=b" (abcd[1]), "=c" (abcd[2]), "=d" (abcd[3]) : "a" (func), "c" (id) );
-#  endif
-#elif defined(_MSC_VER)
-#  if (_MSC_VER > 1500)
-#    define EIGEN_CPUID(abcd,func,id) __cpuidex((int*)abcd,func,id)
+#if !defined(EIGEN_NO_CPUID)
+#  if defined(__GNUC__) && ( defined(__i386__) || defined(__x86_64__) )
+#    if defined(__PIC__) && defined(__i386__)
+       // Case for x86 with PIC
+#      define EIGEN_CPUID(abcd,func,id) \
+         __asm__ __volatile__ ("xchgl %%ebx, %%esi;cpuid; xchgl %%ebx,%%esi": "=a" (abcd[0]), "=S" (abcd[1]), "=c" (abcd[2]), "=d" (abcd[3]) : "a" (func), "c" (id));
+#    else
+       // Case for x86_64 or x86 w/o PIC
+#      define EIGEN_CPUID(abcd,func,id) \
+         __asm__ __volatile__ ("cpuid": "=a" (abcd[0]), "=b" (abcd[1]), "=c" (abcd[2]), "=d" (abcd[3]) : "a" (func), "c" (id) );
+#    endif
+#  elif defined(_MSC_VER)
+#    if (_MSC_VER > 1500)
+#      define EIGEN_CPUID(abcd,func,id) __cpuidex((int*)abcd,func,id)
+#    endif
 #  endif
 #endif
 
@@ -742,7 +765,7 @@ namespace internal {
 
 inline bool cpuid_is_vendor(int abcd[4], const char* vendor)
 {
-  return abcd[1]==((int*)(vendor))[0] && abcd[3]==((int*)(vendor))[1] && abcd[2]==((int*)(vendor))[2];
+  return abcd[1]==(reinterpret_cast<const int*>(vendor))[0] && abcd[3]==(reinterpret_cast<const int*>(vendor))[1] && abcd[2]==(reinterpret_cast<const int*>(vendor))[2];
 }
 
 inline void queryCacheSizes_intel_direct(int& l1, int& l2, int& l3)