merge with default branch

1 files changed, 7 insertions, 11 deletions

diff --git a/Eigen/src/Core/products/TriangularMatrixMatrix.h b/Eigen/src/Core/products/TriangularMatrixMatrix.h
index 3f0618410..d93277cb8 100644
--- a/Eigen/src/Core/products/TriangularMatrixMatrix.h
+++ b/Eigen/src/Core/products/TriangularMatrixMatrix.h
@@ -125,11 +125,9 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,true,
 
     std::size_t sizeA = kc*mc;
     std::size_t sizeB = kc*cols;
-    std::size_t sizeW = kc*Traits::WorkSpaceFactor;
 
     ei_declare_aligned_stack_constructed_variable(Scalar, blockA, sizeA, blocking.blockA());
     ei_declare_aligned_stack_constructed_variable(Scalar, blockB, sizeB, blocking.blockB());
-    ei_declare_aligned_stack_constructed_variable(Scalar, blockW, sizeW, blocking.blockW());
 
     Matrix<Scalar,SmallPanelWidth,SmallPanelWidth,LhsStorageOrder> triangularBuffer;
     triangularBuffer.setZero();
@@ -187,7 +185,7 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,true,
           pack_lhs(blockA, triangularBuffer.data(), triangularBuffer.outerStride(), actualPanelWidth, actualPanelWidth);
 
           gebp_kernel(res+startBlock, resStride, blockA, blockB, actualPanelWidth, actualPanelWidth, cols, alpha,
-                      actualPanelWidth, actual_kc, 0, blockBOffset, blockW);
+                      actualPanelWidth, actual_kc, 0, blockBOffset);
 
           // GEBP with remaining micro panel
           if (lengthTarget>0)
@@ -197,7 +195,7 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,true,
             pack_lhs(blockA, &lhs(startTarget,startBlock), lhsStride, actualPanelWidth, lengthTarget);
 
             gebp_kernel(res+startTarget, resStride, blockA, blockB, lengthTarget, actualPanelWidth, cols, alpha,
-                        actualPanelWidth, actual_kc, 0, blockBOffset, blockW);
+                        actualPanelWidth, actual_kc, 0, blockBOffset);
           }
         }
       }
@@ -211,7 +209,7 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,true,
           gemm_pack_lhs<Scalar, Index, Traits::mr,Traits::LhsProgress, LhsStorageOrder,false>()
             (blockA, &lhs(i2, actual_k2), lhsStride, actual_kc, actual_mc);
 
-          gebp_kernel(res+i2, resStride, blockA, blockB, actual_mc, actual_kc, cols, alpha, -1, -1, 0, 0, blockW);
+          gebp_kernel(res+i2, resStride, blockA, blockB, actual_mc, actual_kc, cols, alpha, -1, -1, 0, 0);
         }
       }
     }
@@ -265,12 +263,10 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,false,
     Index mc = (std::min)(rows,blocking.mc());  // cache block size along the M direction
 
     std::size_t sizeA = kc*mc;
-    std::size_t sizeB = kc*cols;
-    std::size_t sizeW = kc*Traits::WorkSpaceFactor;
+    std::size_t sizeB = kc*cols+EIGEN_ALIGN_BYTES/sizeof(Scalar);
 
     ei_declare_aligned_stack_constructed_variable(Scalar, blockA, sizeA, blocking.blockA());
     ei_declare_aligned_stack_constructed_variable(Scalar, blockB, sizeB, blocking.blockB());
-    ei_declare_aligned_stack_constructed_variable(Scalar, blockW, sizeW, blocking.blockW());
 
     Matrix<Scalar,SmallPanelWidth,SmallPanelWidth,RhsStorageOrder> triangularBuffer;
     triangularBuffer.setZero();
@@ -304,6 +300,7 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,false,
       Index ts = (IsLower && actual_k2>=cols) ? 0 : actual_kc;
 
       Scalar* geb = blockB+ts*ts;
+      geb = geb + internal::first_aligned(geb,EIGEN_ALIGN_BYTES/sizeof(Scalar));
 
       pack_rhs(geb, &rhs(actual_k2,IsLower ? 0 : k2), rhsStride, actual_kc, rs);
 
@@ -357,14 +354,13 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,false,
                         actual_mc, panelLength, actualPanelWidth,
                         alpha,
                         actual_kc, actual_kc,  // strides
-                        blockOffset, blockOffset,// offsets
-                        blockW); // workspace
+                        blockOffset, blockOffset);// offsets
           }
         }
         gebp_kernel(res+i2+(IsLower ? 0 : k2)*resStride, resStride,
                     blockA, geb, actual_mc, actual_kc, rs,
                     alpha,
-                    -1, -1, 0, 0, blockW);
+                    -1, -1, 0, 0);
       }
     }
   }