tomoto 0.2.2 → 0.3.0

data/vendor/eigen/Eigen/src/Core/products/TriangularMatrixMatrix.h

@@ -45,22 +45,24 @@ template <typename Scalar, typename Index,
           int Mode, bool LhsIsTriangular,
           int LhsStorageOrder, bool ConjugateLhs,
           int RhsStorageOrder, bool ConjugateRhs,
-          int ResStorageOrder, int Version = Specialized>
+          int ResStorageOrder, int ResInnerStride,
+          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 Version>
+          int RhsStorageOrder, bool ConjugateRhs,
+          int ResInnerStride, int Version>
 struct product_triangular_matrix_matrix<Scalar,Index,Mode,LhsIsTriangular,
                                            LhsStorageOrder,ConjugateLhs,
-                                           RhsStorageOrder,ConjugateRhs,RowMajor,Version>
+                                           RhsStorageOrder,ConjugateRhs,RowMajor,ResInnerStride,Version>
 {
   static EIGEN_STRONG_INLINE void run(
     Index rows, Index cols, Index depth,
     const Scalar* lhs, Index lhsStride,
     const Scalar* rhs, Index rhsStride,
-    Scalar* res,       Index resStride,
+    Scalar* res,       Index resIncr, Index resStride,
     const Scalar& alpha, level3_blocking<Scalar,Scalar>& blocking)
   {
     product_triangular_matrix_matrix<Scalar, Index,
@@ -70,18 +72,19 @@ struct product_triangular_matrix_matrix<Scalar,Index,Mode,LhsIsTriangular,
       ConjugateRhs,
       LhsStorageOrder==RowMajor ? ColMajor : RowMajor,
       ConjugateLhs,
-      ColMajor>
-      ::run(cols, rows, depth, rhs, rhsStride, lhs, lhsStride, res, resStride, alpha, blocking);
+      ColMajor, ResInnerStride>
+      ::run(cols, rows, depth, rhs, rhsStride, lhs, lhsStride, res, resIncr, resStride, alpha, blocking);
   }
 };
 
 // implements col-major += alpha * op(triangular) * op(general)
 template <typename Scalar, typename Index, int Mode,
           int LhsStorageOrder, bool ConjugateLhs,
-          int RhsStorageOrder, bool ConjugateRhs, int Version>
+          int RhsStorageOrder, bool ConjugateRhs,
+          int ResInnerStride, int Version>
 struct product_triangular_matrix_matrix<Scalar,Index,Mode,true,
                                            LhsStorageOrder,ConjugateLhs,
-                                           RhsStorageOrder,ConjugateRhs,ColMajor,Version>
+                                           RhsStorageOrder,ConjugateRhs,ColMajor,ResInnerStride,Version>
 {
   
   typedef gebp_traits<Scalar,Scalar> Traits;
@@ -95,20 +98,21 @@ struct product_triangular_matrix_matrix<Scalar,Index,Mode,true,
     Index _rows, Index _cols, Index _depth,
     const Scalar* _lhs, Index lhsStride,
     const Scalar* _rhs, Index rhsStride,
-    Scalar* res,        Index resStride,
+    Scalar* res,        Index resIncr, Index resStride,
     const Scalar& alpha, level3_blocking<Scalar,Scalar>& blocking);
 };
 
 template <typename Scalar, typename Index, int Mode,
           int LhsStorageOrder, bool ConjugateLhs,
-          int RhsStorageOrder, bool ConjugateRhs, int Version>
+          int RhsStorageOrder, bool ConjugateRhs,
+          int ResInnerStride, int Version>
 EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,true,
                                                         LhsStorageOrder,ConjugateLhs,
-                                                        RhsStorageOrder,ConjugateRhs,ColMajor,Version>::run(
+                                                        RhsStorageOrder,ConjugateRhs,ColMajor,ResInnerStride,Version>::run(
     Index _rows, Index _cols, Index _depth,
     const Scalar* _lhs, Index lhsStride,
     const Scalar* _rhs, Index rhsStride,
-    Scalar* _res,        Index resStride,
+    Scalar* _res,       Index resIncr, Index resStride,
     const Scalar& alpha, level3_blocking<Scalar,Scalar>& blocking)
   {
     // strip zeros
@@ -119,10 +123,10 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,true,
     
     typedef const_blas_data_mapper<Scalar, Index, LhsStorageOrder> LhsMapper;
     typedef const_blas_data_mapper<Scalar, Index, RhsStorageOrder> RhsMapper;
-    typedef blas_data_mapper<typename Traits::ResScalar, Index, ColMajor> ResMapper;
+    typedef blas_data_mapper<typename Traits::ResScalar, Index, ColMajor, Unaligned, ResInnerStride> ResMapper;
     LhsMapper lhs(_lhs,lhsStride);
     RhsMapper rhs(_rhs,rhsStride);
-    ResMapper res(_res, resStride);
+    ResMapper res(_res, resStride, resIncr);
 
     Index kc = blocking.kc();                   // cache block size along the K direction
     Index mc = (std::min)(rows,blocking.mc());  // cache block size along the M direction
@@ -151,7 +155,7 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,true,
       triangularBuffer.diagonal().setOnes();
 
     gebp_kernel<Scalar, Scalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp_kernel;
-    gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs;
+    gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, typename Traits::LhsPacket4Packing, LhsStorageOrder> pack_lhs;
     gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr,RhsStorageOrder> pack_rhs;
 
     for(Index k2=IsLower ? depth : 0;
@@ -222,7 +226,7 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,true,
         for(Index i2=start; i2<end; i2+=mc)
         {
           const Index actual_mc = (std::min)(i2+mc,end)-i2;
-          gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr,Traits::LhsProgress, LhsStorageOrder,false>()
+          gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr,Traits::LhsProgress, typename Traits::LhsPacket4Packing, LhsStorageOrder,false>()
             (blockA, lhs.getSubMapper(i2, actual_k2), actual_kc, actual_mc);
 
           gebp_kernel(res.getSubMapper(i2, 0), blockA, blockB, actual_mc,
@@ -235,10 +239,11 @@ EIGEN_DONT_INLINE void 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 Version>
+          int RhsStorageOrder, bool ConjugateRhs,
+          int ResInnerStride, int Version>
 struct product_triangular_matrix_matrix<Scalar,Index,Mode,false,
                                         LhsStorageOrder,ConjugateLhs,
-                                        RhsStorageOrder,ConjugateRhs,ColMajor,Version>
+                                        RhsStorageOrder,ConjugateRhs,ColMajor,ResInnerStride,Version>
 {
   typedef gebp_traits<Scalar,Scalar> Traits;
   enum {
@@ -251,20 +256,21 @@ struct product_triangular_matrix_matrix<Scalar,Index,Mode,false,
     Index _rows, Index _cols, Index _depth,
     const Scalar* _lhs, Index lhsStride,
     const Scalar* _rhs, Index rhsStride,
-    Scalar* res,        Index resStride,
+    Scalar* res,        Index resIncr, Index resStride,
     const Scalar& alpha, level3_blocking<Scalar,Scalar>& blocking);
 };
 
 template <typename Scalar, typename Index, int Mode,
           int LhsStorageOrder, bool ConjugateLhs,
-          int RhsStorageOrder, bool ConjugateRhs, int Version>
+          int RhsStorageOrder, bool ConjugateRhs,
+          int ResInnerStride, int Version>
 EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,false,
                                                         LhsStorageOrder,ConjugateLhs,
-                                                        RhsStorageOrder,ConjugateRhs,ColMajor,Version>::run(
+                                                        RhsStorageOrder,ConjugateRhs,ColMajor,ResInnerStride,Version>::run(
     Index _rows, Index _cols, Index _depth,
     const Scalar* _lhs, Index lhsStride,
     const Scalar* _rhs, Index rhsStride,
-    Scalar* _res,        Index resStride,
+    Scalar* _res,       Index resIncr, Index resStride,
     const Scalar& alpha, level3_blocking<Scalar,Scalar>& blocking)
   {
     const Index PacketBytes = packet_traits<Scalar>::size*sizeof(Scalar);
@@ -276,10 +282,10 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,false,
     
     typedef const_blas_data_mapper<Scalar, Index, LhsStorageOrder> LhsMapper;
     typedef const_blas_data_mapper<Scalar, Index, RhsStorageOrder> RhsMapper;
-    typedef blas_data_mapper<typename Traits::ResScalar, Index, ColMajor> ResMapper;
+    typedef blas_data_mapper<typename Traits::ResScalar, Index, ColMajor, Unaligned, ResInnerStride> ResMapper;
     LhsMapper lhs(_lhs,lhsStride);
     RhsMapper rhs(_rhs,rhsStride);
-    ResMapper res(_res, resStride);
+    ResMapper res(_res, resStride, resIncr);
 
     Index kc = blocking.kc();                   // cache block size along the K direction
     Index mc = (std::min)(rows,blocking.mc());  // cache block size along the M direction
@@ -299,7 +305,7 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,false,
       triangularBuffer.diagonal().setOnes();
 
     gebp_kernel<Scalar, Scalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp_kernel;
-    gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs;
+    gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, typename Traits::LhsPacket4Packing, LhsStorageOrder> pack_lhs;
     gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr,RhsStorageOrder> pack_rhs;
     gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr,RhsStorageOrder,false,true> pack_rhs_panel;
 
@@ -433,12 +439,12 @@ struct triangular_product_impl<Mode,LhsIsTriangular,Lhs,false,Rhs,false>
       Mode, LhsIsTriangular,
       (internal::traits<ActualLhsTypeCleaned>::Flags&RowMajorBit) ? RowMajor : ColMajor, LhsBlasTraits::NeedToConjugate,
       (internal::traits<ActualRhsTypeCleaned>::Flags&RowMajorBit) ? RowMajor : ColMajor, RhsBlasTraits::NeedToConjugate,
-      (internal::traits<Dest          >::Flags&RowMajorBit) ? RowMajor : ColMajor>
+      (internal::traits<Dest          >::Flags&RowMajorBit) ? RowMajor : ColMajor, Dest::InnerStrideAtCompileTime>
       ::run(
         stripedRows, stripedCols, stripedDepth,   // sizes
         &lhs.coeffRef(0,0), lhs.outerStride(),    // lhs info
         &rhs.coeffRef(0,0), rhs.outerStride(),    // rhs info
-        &dst.coeffRef(0,0), dst.outerStride(),    // result info
+        &dst.coeffRef(0,0), dst.innerStride(), dst.outerStride(),    // result info
         actualAlpha, blocking
       );
 

data/vendor/eigen/Eigen/src/Core/products/TriangularMatrixMatrix_BLAS.h

@@ -46,7 +46,7 @@ template <typename Scalar, typename Index,
 struct product_triangular_matrix_matrix_trmm :
        product_triangular_matrix_matrix<Scalar,Index,Mode,
           LhsIsTriangular,LhsStorageOrder,ConjugateLhs,
-          RhsStorageOrder, ConjugateRhs, ResStorageOrder, BuiltIn> {};
+          RhsStorageOrder, ConjugateRhs, ResStorageOrder, 1, BuiltIn> {};
 
 
 // try to go to BLAS specialization
@@ -55,13 +55,15 @@ 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> { \
+           LhsStorageOrder,ConjugateLhs, RhsStorageOrder,ConjugateRhs,ColMajor,1,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, level3_blocking<Scalar,Scalar>& blocking) { \
+    const Scalar* _rhs, Index rhsStride, Scalar* res, Index resIncr, Index resStride, Scalar alpha, level3_blocking<Scalar,Scalar>& blocking) { \
+      EIGEN_ONLY_USED_FOR_DEBUG(resIncr); \
+      eigen_assert(resIncr == 1); \
       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, blocking); \
+          _rows, _cols, _depth, _lhs, lhsStride, _rhs, rhsStride, res, resStride, alpha, blocking); \
   } \
 };
 
@@ -115,8 +117,8 @@ struct product_triangular_matrix_matrix_trmm<EIGTYPE,Index,Mode,true, \
      if (((nthr==1) && (((std::max)(rows,depth)-diagSize)/(double)diagSize < 0.5))) { \
      /* Most likely no benefit to call TRMM or GEMM from BLAS */ \
        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, blocking); \
+       LhsStorageOrder,ConjugateLhs, RhsStorageOrder, ConjugateRhs, ColMajor, 1, BuiltIn>::run( \
+           _rows, _cols, _depth, _lhs, lhsStride, _rhs, rhsStride, res, 1, resStride, alpha, blocking); \
      /*std::cout << "TRMM_L: A is not square! Go to Eigen TRMM implementation!\n";*/ \
      } else { \
      /* Make sense to call GEMM */ \
@@ -124,8 +126,8 @@ struct product_triangular_matrix_matrix_trmm<EIGTYPE,Index,Mode,true, \
        MatrixLhs aa_tmp=lhsMap.template triangularView<Mode>(); \
        BlasIndex aStride = convert_index<BlasIndex>(aa_tmp.outerStride()); \
        gemm_blocking_space<ColMajor,EIGTYPE,EIGTYPE,Dynamic,Dynamic,Dynamic> gemm_blocking(_rows,_cols,_depth, 1, true); \
-       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, gemm_blocking, 0); \
+       general_matrix_matrix_product<Index,EIGTYPE,LhsStorageOrder,ConjugateLhs,EIGTYPE,RhsStorageOrder,ConjugateRhs,ColMajor,1>::run( \
+       rows, cols, depth, aa_tmp.data(), aStride, _rhs, rhsStride, res, 1, resStride, alpha, gemm_blocking, 0); \
 \
      /*std::cout << "TRMM_L: A is not square! Go to BLAS GEMM implementation! " << nthr<<" \n";*/ \
      } \
@@ -232,8 +234,8 @@ struct product_triangular_matrix_matrix_trmm<EIGTYPE,Index,Mode,false, \
      if ((nthr==1) && (((std::max)(cols,depth)-diagSize)/(double)diagSize < 0.5)) { \
      /* Most likely no benefit to call TRMM or GEMM from BLAS*/ \
        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, blocking); \
+       LhsStorageOrder,ConjugateLhs, RhsStorageOrder, ConjugateRhs, ColMajor, 1, BuiltIn>::run( \
+           _rows, _cols, _depth, _lhs, lhsStride, _rhs, rhsStride, res, 1, resStride, alpha, blocking); \
        /*std::cout << "TRMM_R: A is not square! Go to Eigen TRMM implementation!\n";*/ \
      } else { \
      /* Make sense to call GEMM */ \
@@ -241,8 +243,8 @@ struct product_triangular_matrix_matrix_trmm<EIGTYPE,Index,Mode,false, \
        MatrixRhs aa_tmp=rhsMap.template triangularView<Mode>(); \
        BlasIndex aStride = convert_index<BlasIndex>(aa_tmp.outerStride()); \
        gemm_blocking_space<ColMajor,EIGTYPE,EIGTYPE,Dynamic,Dynamic,Dynamic> gemm_blocking(_rows,_cols,_depth, 1, true); \
-       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, gemm_blocking, 0); \
+       general_matrix_matrix_product<Index,EIGTYPE,LhsStorageOrder,ConjugateLhs,EIGTYPE,RhsStorageOrder,ConjugateRhs,ColMajor,1>::run( \
+       rows, cols, depth, _lhs, lhsStride, aa_tmp.data(), aStride, res, 1, resStride, alpha, gemm_blocking, 0); \
 \
      /*std::cout << "TRMM_R: A is not square! Go to BLAS GEMM implementation! " << nthr<<" \n";*/ \
      } \

data/vendor/eigen/Eigen/src/Core/products/TriangularSolverMatrix.h

@@ -15,48 +15,48 @@ namespace Eigen {
 namespace internal {
 
 // if the rhs is row major, let's transpose the product
-template <typename Scalar, typename Index, int Side, int Mode, bool Conjugate, int TriStorageOrder>
-struct triangular_solve_matrix<Scalar,Index,Side,Mode,Conjugate,TriStorageOrder,RowMajor>
+template <typename Scalar, typename Index, int Side, int Mode, bool Conjugate, int TriStorageOrder, int OtherInnerStride>
+struct triangular_solve_matrix<Scalar,Index,Side,Mode,Conjugate,TriStorageOrder,RowMajor,OtherInnerStride>
 {
   static void run(
     Index size, Index cols,
     const Scalar*  tri, Index triStride,
-    Scalar* _other, Index otherStride,
+    Scalar* _other, Index otherIncr, Index otherStride,
     level3_blocking<Scalar,Scalar>& blocking)
   {
     triangular_solve_matrix<
       Scalar, Index, Side==OnTheLeft?OnTheRight:OnTheLeft,
       (Mode&UnitDiag) | ((Mode&Upper) ? Lower : Upper),
       NumTraits<Scalar>::IsComplex && Conjugate,
-      TriStorageOrder==RowMajor ? ColMajor : RowMajor, ColMajor>
-      ::run(size, cols, tri, triStride, _other, otherStride, blocking);
+      TriStorageOrder==RowMajor ? ColMajor : RowMajor, ColMajor, OtherInnerStride>
+      ::run(size, cols, tri, triStride, _other, otherIncr, otherStride, blocking);
   }
 };
 
 /* Optimized triangular solver with multiple right hand side and the triangular matrix on the left
  */
-template <typename Scalar, typename Index, int Mode, bool Conjugate, int TriStorageOrder>
-struct triangular_solve_matrix<Scalar,Index,OnTheLeft,Mode,Conjugate,TriStorageOrder,ColMajor>
+template <typename Scalar, typename Index, int Mode, bool Conjugate, int TriStorageOrder,int OtherInnerStride>
+struct triangular_solve_matrix<Scalar,Index,OnTheLeft,Mode,Conjugate,TriStorageOrder,ColMajor,OtherInnerStride>
 {
   static EIGEN_DONT_INLINE void run(
     Index size, Index otherSize,
     const Scalar* _tri, Index triStride,
-    Scalar* _other, Index otherStride,
+    Scalar* _other, Index otherIncr, Index otherStride,
     level3_blocking<Scalar,Scalar>& blocking);
 };
-template <typename Scalar, typename Index, int Mode, bool Conjugate, int TriStorageOrder>
-EIGEN_DONT_INLINE void triangular_solve_matrix<Scalar,Index,OnTheLeft,Mode,Conjugate,TriStorageOrder,ColMajor>::run(
+template <typename Scalar, typename Index, int Mode, bool Conjugate, int TriStorageOrder, int OtherInnerStride>
+EIGEN_DONT_INLINE void triangular_solve_matrix<Scalar,Index,OnTheLeft,Mode,Conjugate,TriStorageOrder,ColMajor,OtherInnerStride>::run(
     Index size, Index otherSize,
     const Scalar* _tri, Index triStride,
-    Scalar* _other, Index otherStride,
+    Scalar* _other, Index otherIncr, Index otherStride,
     level3_blocking<Scalar,Scalar>& blocking)
   {
     Index cols = otherSize;
 
     typedef const_blas_data_mapper<Scalar, Index, TriStorageOrder> TriMapper;
-    typedef blas_data_mapper<Scalar, Index, ColMajor> OtherMapper;
+    typedef blas_data_mapper<Scalar, Index, ColMajor, Unaligned, OtherInnerStride> OtherMapper;
     TriMapper tri(_tri, triStride);
-    OtherMapper other(_other, otherStride);
+    OtherMapper other(_other, otherStride, otherIncr);
 
     typedef gebp_traits<Scalar,Scalar> Traits;
 
@@ -76,7 +76,7 @@ EIGEN_DONT_INLINE void triangular_solve_matrix<Scalar,Index,OnTheLeft,Mode,Conju
 
     conj_if<Conjugate> conj;
     gebp_kernel<Scalar, Scalar, Index, OtherMapper, Traits::mr, Traits::nr, Conjugate, false> gebp_kernel;
-    gemm_pack_lhs<Scalar, Index, TriMapper, Traits::mr, Traits::LhsProgress, TriStorageOrder> pack_lhs;
+    gemm_pack_lhs<Scalar, Index, TriMapper, Traits::mr, Traits::LhsProgress, typename Traits::LhsPacket4Packing, TriStorageOrder> pack_lhs;
     gemm_pack_rhs<Scalar, Index, OtherMapper, Traits::nr, ColMajor, false, true> pack_rhs;
 
     // the goal here is to subdivise the Rhs panels such that we keep some cache
@@ -128,19 +128,21 @@ EIGEN_DONT_INLINE void triangular_solve_matrix<Scalar,Index,OnTheLeft,Mode,Conju
               {
                 Scalar b(0);
                 const Scalar* l = &tri(i,s);
-                Scalar* r = &other(s,j);
+                typename OtherMapper::LinearMapper r = other.getLinearMapper(s,j);
                 for (Index i3=0; i3<k; ++i3)
-                  b += conj(l[i3]) * r[i3];
+                  b += conj(l[i3]) * r(i3);
 
                 other(i,j) = (other(i,j) - b)*a;
               }
               else
               {
-                Scalar b = (other(i,j) *= a);
-                Scalar* r = &other(s,j);
-                const Scalar* l = &tri(s,i);
+                Scalar& otherij = other(i,j);
+                otherij *= a;
+                Scalar b = otherij;
+                typename OtherMapper::LinearMapper r = other.getLinearMapper(s,j);
+                typename TriMapper::LinearMapper l = tri.getLinearMapper(s,i);
                 for (Index i3=0;i3<rs;++i3)
-                  r[i3] -= b * conj(l[i3]);
+                  r(i3) -= b * conj(l(i3));
               }
             }
           }
@@ -185,28 +187,28 @@ EIGEN_DONT_INLINE void triangular_solve_matrix<Scalar,Index,OnTheLeft,Mode,Conju
 
 /* Optimized triangular solver with multiple left hand sides and the triangular matrix on the right
  */
-template <typename Scalar, typename Index, int Mode, bool Conjugate, int TriStorageOrder>
-struct triangular_solve_matrix<Scalar,Index,OnTheRight,Mode,Conjugate,TriStorageOrder,ColMajor>
+template <typename Scalar, typename Index, int Mode, bool Conjugate, int TriStorageOrder, int OtherInnerStride>
+struct triangular_solve_matrix<Scalar,Index,OnTheRight,Mode,Conjugate,TriStorageOrder,ColMajor,OtherInnerStride>
 {
   static EIGEN_DONT_INLINE void run(
     Index size, Index otherSize,
     const Scalar* _tri, Index triStride,
-    Scalar* _other, Index otherStride,
+    Scalar* _other, Index otherIncr, Index otherStride,
     level3_blocking<Scalar,Scalar>& blocking);
 };
-template <typename Scalar, typename Index, int Mode, bool Conjugate, int TriStorageOrder>
-EIGEN_DONT_INLINE void triangular_solve_matrix<Scalar,Index,OnTheRight,Mode,Conjugate,TriStorageOrder,ColMajor>::run(
+template <typename Scalar, typename Index, int Mode, bool Conjugate, int TriStorageOrder, int OtherInnerStride>
+EIGEN_DONT_INLINE void triangular_solve_matrix<Scalar,Index,OnTheRight,Mode,Conjugate,TriStorageOrder,ColMajor,OtherInnerStride>::run(
     Index size, Index otherSize,
     const Scalar* _tri, Index triStride,
-    Scalar* _other, Index otherStride,
+    Scalar* _other, Index otherIncr, Index otherStride,
     level3_blocking<Scalar,Scalar>& blocking)
   {
     Index rows = otherSize;
     typedef typename NumTraits<Scalar>::Real RealScalar;
 
-    typedef blas_data_mapper<Scalar, Index, ColMajor> LhsMapper;
+    typedef blas_data_mapper<Scalar, Index, ColMajor, Unaligned, OtherInnerStride> LhsMapper;
     typedef const_blas_data_mapper<Scalar, Index, TriStorageOrder> RhsMapper;
-    LhsMapper lhs(_other, otherStride);
+    LhsMapper lhs(_other, otherStride, otherIncr);
     RhsMapper rhs(_tri, triStride);
 
     typedef gebp_traits<Scalar,Scalar> Traits;
@@ -229,7 +231,7 @@ EIGEN_DONT_INLINE void triangular_solve_matrix<Scalar,Index,OnTheRight,Mode,Conj
     gebp_kernel<Scalar, Scalar, Index, LhsMapper, Traits::mr, Traits::nr, false, Conjugate> gebp_kernel;
     gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr, RhsStorageOrder> pack_rhs;
     gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr, RhsStorageOrder,false,true> pack_rhs_panel;
-    gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, ColMajor, false, true> pack_lhs_panel;
+    gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, typename Traits::LhsPacket4Packing, ColMajor, false, true> pack_lhs_panel;
 
     for(Index k2=IsLower ? size : 0;
         IsLower ? k2>0 : k2<size;
@@ -297,24 +299,24 @@ EIGEN_DONT_INLINE void triangular_solve_matrix<Scalar,Index,OnTheRight,Mode,Conj
             {
               Index j = IsLower ? absolute_j2+actualPanelWidth-k-1 : absolute_j2+k;
 
-              Scalar* r = &lhs(i2,j);
+              typename LhsMapper::LinearMapper r = lhs.getLinearMapper(i2,j);
               for (Index k3=0; k3<k; ++k3)
               {
                 Scalar b = conj(rhs(IsLower ? j+1+k3 : absolute_j2+k3,j));
-                Scalar* a = &lhs(i2,IsLower ? j+1+k3 : absolute_j2+k3);
+                typename LhsMapper::LinearMapper a = lhs.getLinearMapper(i2,IsLower ? j+1+k3 : absolute_j2+k3);
                 for (Index i=0; i<actual_mc; ++i)
-                  r[i] -= a[i] * b;
+                  r(i) -= a(i) * b;
               }
               if((Mode & UnitDiag)==0)
               {
                 Scalar inv_rjj = RealScalar(1)/conj(rhs(j,j));
                 for (Index i=0; i<actual_mc; ++i)
-                  r[i] *= inv_rjj;
+                  r(i) *= inv_rjj;
               }
             }
 
             // pack the just computed part of lhs to A
-            pack_lhs_panel(blockA, LhsMapper(_other+absolute_j2*otherStride+i2, otherStride),
+            pack_lhs_panel(blockA, lhs.getSubMapper(i2,absolute_j2),
                            actualPanelWidth, actual_mc,
                            actual_kc, j2);
           }

data/vendor/eigen/Eigen/src/Core/products/TriangularSolverMatrix_BLAS.h

@@ -40,7 +40,7 @@ namespace internal {
 // implements LeftSide op(triangular)^-1 * general
 #define EIGEN_BLAS_TRSM_L(EIGTYPE, BLASTYPE, BLASFUNC) \
 template <typename Index, int Mode, bool Conjugate, int TriStorageOrder> \
-struct triangular_solve_matrix<EIGTYPE,Index,OnTheLeft,Mode,Conjugate,TriStorageOrder,ColMajor> \
+struct triangular_solve_matrix<EIGTYPE,Index,OnTheLeft,Mode,Conjugate,TriStorageOrder,ColMajor,1> \
 { \
   enum { \
     IsLower = (Mode&Lower) == Lower, \
@@ -51,8 +51,10 @@ struct triangular_solve_matrix<EIGTYPE,Index,OnTheLeft,Mode,Conjugate,TriStorage
   static void run( \
       Index size, Index otherSize, \
       const EIGTYPE* _tri, Index triStride, \
-      EIGTYPE* _other, Index otherStride, level3_blocking<EIGTYPE,EIGTYPE>& /*blocking*/) \
+      EIGTYPE* _other, Index otherIncr, Index otherStride, level3_blocking<EIGTYPE,EIGTYPE>& /*blocking*/) \
   { \
+   EIGEN_ONLY_USED_FOR_DEBUG(otherIncr); \
+   eigen_assert(otherIncr == 1); \
    BlasIndex m = convert_index<BlasIndex>(size), n = convert_index<BlasIndex>(otherSize), lda, ldb; \
    char side = 'L', uplo, diag='N', transa; \
    /* Set alpha_ */ \
@@ -99,7 +101,7 @@ EIGEN_BLAS_TRSM_L(scomplex, float,  ctrsm_)
 // implements RightSide general * op(triangular)^-1
 #define EIGEN_BLAS_TRSM_R(EIGTYPE, BLASTYPE, BLASFUNC) \
 template <typename Index, int Mode, bool Conjugate, int TriStorageOrder> \
-struct triangular_solve_matrix<EIGTYPE,Index,OnTheRight,Mode,Conjugate,TriStorageOrder,ColMajor> \
+struct triangular_solve_matrix<EIGTYPE,Index,OnTheRight,Mode,Conjugate,TriStorageOrder,ColMajor,1> \
 { \
   enum { \
     IsLower = (Mode&Lower) == Lower, \
@@ -110,8 +112,10 @@ struct triangular_solve_matrix<EIGTYPE,Index,OnTheRight,Mode,Conjugate,TriStorag
   static void run( \
       Index size, Index otherSize, \
       const EIGTYPE* _tri, Index triStride, \
-      EIGTYPE* _other, Index otherStride, level3_blocking<EIGTYPE,EIGTYPE>& /*blocking*/) \
+      EIGTYPE* _other, Index otherIncr, Index otherStride, level3_blocking<EIGTYPE,EIGTYPE>& /*blocking*/) \
   { \
+   EIGEN_ONLY_USED_FOR_DEBUG(otherIncr); \
+   eigen_assert(otherIncr == 1); \
    BlasIndex m = convert_index<BlasIndex>(otherSize), n = convert_index<BlasIndex>(size), lda, ldb; \
    char side = 'R', uplo, diag='N', transa; \
    /* Set alpha_ */ \

data/vendor/eigen/Eigen/src/Core/products/TriangularSolverVector.h

@@ -58,7 +58,7 @@ struct triangular_solve_vector<LhsScalar, RhsScalar, Index, OnTheLeft, Mode, Con
       {
         // let's directly call the low level product function because:
         // 1 - it is faster to compile
-        // 2 - it is slighlty faster at runtime
+        // 2 - it is slightly faster at runtime
         Index startRow = IsLower ? pi : pi-actualPanelWidth;
         Index startCol = IsLower ? 0 : pi;
 
@@ -77,7 +77,7 @@ struct triangular_solve_vector<LhsScalar, RhsScalar, Index, OnTheLeft, Mode, Con
         if (k>0)
           rhs[i] -= (cjLhs.row(i).segment(s,k).transpose().cwiseProduct(Map<const Matrix<RhsScalar,Dynamic,1> >(rhs+s,k))).sum();
 
-        if(!(Mode & UnitDiag))
+        if((!(Mode & UnitDiag)) && numext::not_equal_strict(rhs[i],RhsScalar(0)))
           rhs[i] /= cjLhs(i,i);
       }
     }
@@ -114,20 +114,23 @@ struct triangular_solve_vector<LhsScalar, RhsScalar, Index, OnTheLeft, Mode, Con
       for(Index k=0; k<actualPanelWidth; ++k)
       {
         Index i = IsLower ? pi+k : pi-k-1;
-        if(!(Mode & UnitDiag))
-          rhs[i] /= cjLhs.coeff(i,i);
-
-        Index r = actualPanelWidth - k - 1; // remaining size
-        Index s = IsLower ? i+1 : i-r;
-        if (r>0)
-          Map<Matrix<RhsScalar,Dynamic,1> >(rhs+s,r) -= rhs[i] * cjLhs.col(i).segment(s,r);
+        if(numext::not_equal_strict(rhs[i],RhsScalar(0)))
+        {
+          if(!(Mode & UnitDiag))
+            rhs[i] /= cjLhs.coeff(i,i);
+
+          Index r = actualPanelWidth - k - 1; // remaining size
+          Index s = IsLower ? i+1 : i-r;
+          if (r>0)
+            Map<Matrix<RhsScalar,Dynamic,1> >(rhs+s,r) -= rhs[i] * cjLhs.col(i).segment(s,r);
+        }
       }
       Index r = IsLower ? size - endBlock : startBlock; // remaining size
       if (r > 0)
       {
         // let's directly call the low level product function because:
         // 1 - it is faster to compile
-        // 2 - it is slighlty faster at runtime
+        // 2 - it is slightly faster at runtime
         general_matrix_vector_product<Index,LhsScalar,LhsMapper,ColMajor,Conjugate,RhsScalar,RhsMapper,false>::run(
             r, actualPanelWidth,
             LhsMapper(&lhs.coeffRef(endBlock,startBlock), lhsStride),