tomoto 0.2.2 → 0.3.0

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

@@ -20,8 +20,9 @@ template<typename _LhsScalar, typename _RhsScalar> class level3_blocking;
 template<
   typename Index,
   typename LhsScalar, int LhsStorageOrder, bool ConjugateLhs,
-  typename RhsScalar, int RhsStorageOrder, bool ConjugateRhs>
-struct general_matrix_matrix_product<Index,LhsScalar,LhsStorageOrder,ConjugateLhs,RhsScalar,RhsStorageOrder,ConjugateRhs,RowMajor>
+  typename RhsScalar, int RhsStorageOrder, bool ConjugateRhs,
+  int ResInnerStride>
+struct general_matrix_matrix_product<Index,LhsScalar,LhsStorageOrder,ConjugateLhs,RhsScalar,RhsStorageOrder,ConjugateRhs,RowMajor,ResInnerStride>
 {
   typedef gebp_traits<RhsScalar,LhsScalar> Traits;
 
@@ -30,7 +31,7 @@ struct general_matrix_matrix_product<Index,LhsScalar,LhsStorageOrder,ConjugateLh
     Index rows, Index cols, Index depth,
     const LhsScalar* lhs, Index lhsStride,
     const RhsScalar* rhs, Index rhsStride,
-    ResScalar* res, Index resStride,
+    ResScalar* res, Index resIncr, Index resStride,
     ResScalar alpha,
     level3_blocking<RhsScalar,LhsScalar>& blocking,
     GemmParallelInfo<Index>* info = 0)
@@ -39,8 +40,8 @@ struct general_matrix_matrix_product<Index,LhsScalar,LhsStorageOrder,ConjugateLh
     general_matrix_matrix_product<Index,
       RhsScalar, RhsStorageOrder==RowMajor ? ColMajor : RowMajor, ConjugateRhs,
       LhsScalar, LhsStorageOrder==RowMajor ? ColMajor : RowMajor, ConjugateLhs,
-      ColMajor>
-    ::run(cols,rows,depth,rhs,rhsStride,lhs,lhsStride,res,resStride,alpha,blocking,info);
+      ColMajor,ResInnerStride>
+    ::run(cols,rows,depth,rhs,rhsStride,lhs,lhsStride,res,resIncr,resStride,alpha,blocking,info);
   }
 };
 
@@ -49,8 +50,9 @@ struct general_matrix_matrix_product<Index,LhsScalar,LhsStorageOrder,ConjugateLh
 template<
   typename Index,
   typename LhsScalar, int LhsStorageOrder, bool ConjugateLhs,
-  typename RhsScalar, int RhsStorageOrder, bool ConjugateRhs>
-struct general_matrix_matrix_product<Index,LhsScalar,LhsStorageOrder,ConjugateLhs,RhsScalar,RhsStorageOrder,ConjugateRhs,ColMajor>
+  typename RhsScalar, int RhsStorageOrder, bool ConjugateRhs,
+  int ResInnerStride>
+struct general_matrix_matrix_product<Index,LhsScalar,LhsStorageOrder,ConjugateLhs,RhsScalar,RhsStorageOrder,ConjugateRhs,ColMajor,ResInnerStride>
 {
 
 typedef gebp_traits<LhsScalar,RhsScalar> Traits;
@@ -59,23 +61,23 @@ typedef typename ScalarBinaryOpTraits<LhsScalar, RhsScalar>::ReturnType ResScala
 static void run(Index rows, Index cols, Index depth,
   const LhsScalar* _lhs, Index lhsStride,
   const RhsScalar* _rhs, Index rhsStride,
-  ResScalar* _res, Index resStride,
+  ResScalar* _res, Index resIncr, Index resStride,
   ResScalar alpha,
   level3_blocking<LhsScalar,RhsScalar>& blocking,
   GemmParallelInfo<Index>* info = 0)
 {
   typedef const_blas_data_mapper<LhsScalar, Index, LhsStorageOrder> LhsMapper;
   typedef const_blas_data_mapper<RhsScalar, Index, RhsStorageOrder> RhsMapper;
-  typedef blas_data_mapper<typename Traits::ResScalar, Index, ColMajor> ResMapper;
-  LhsMapper lhs(_lhs,lhsStride);
-  RhsMapper rhs(_rhs,rhsStride);
-  ResMapper res(_res, resStride);
+  typedef blas_data_mapper<typename Traits::ResScalar, Index, ColMajor,Unaligned,ResInnerStride> ResMapper;
+  LhsMapper lhs(_lhs, lhsStride);
+  RhsMapper rhs(_rhs, rhsStride);
+  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
   Index nc = (std::min)(cols,blocking.nc());  // cache block size along the N direction
 
-  gemm_pack_lhs<LhsScalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs;
+  gemm_pack_lhs<LhsScalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, typename Traits::LhsPacket4Packing, LhsStorageOrder> pack_lhs;
   gemm_pack_rhs<RhsScalar, Index, RhsMapper, Traits::nr, RhsStorageOrder> pack_rhs;
   gebp_kernel<LhsScalar, RhsScalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp;
 
@@ -108,7 +110,7 @@ static void run(Index rows, Index cols, Index depth,
       // i.e., we test that info[tid].users equals 0.
       // Then, we set info[tid].users to the number of threads to mark that all other threads are going to use it.
       while(info[tid].users!=0) {}
-      info[tid].users += threads;
+      info[tid].users = threads;
 
       pack_lhs(blockA+info[tid].lhs_start*actual_kc, lhs.getSubMapper(info[tid].lhs_start,k), actual_kc, info[tid].lhs_length);
 
@@ -146,7 +148,9 @@ static void run(Index rows, Index cols, Index depth,
       // Release all the sub blocks A'_i of A' for the current thread,
       // i.e., we simply decrement the number of users by 1
       for(Index i=0; i<threads; ++i)
+#if !EIGEN_HAS_CXX11_ATOMIC
         #pragma omp atomic
+#endif
         info[i].users -= 1;
     }
   }
@@ -226,7 +230,7 @@ struct gemm_functor
     Gemm::run(rows, cols, m_lhs.cols(),
               &m_lhs.coeffRef(row,0), m_lhs.outerStride(),
               &m_rhs.coeffRef(0,col), m_rhs.outerStride(),
-              (Scalar*)&(m_dest.coeffRef(row,col)), m_dest.outerStride(),
+              (Scalar*)&(m_dest.coeffRef(row,col)), m_dest.innerStride(), m_dest.outerStride(),
               m_actualAlpha, m_blocking, info);
   }
 
@@ -427,8 +431,14 @@ struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,GemmProduct>
   template<typename Dst>
   static void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
   {
-    if((rhs.rows()+dst.rows()+dst.cols())<20 && rhs.rows()>0)
-      lazyproduct::evalTo(dst, lhs, rhs);
+    // See http://eigen.tuxfamily.org/bz/show_bug.cgi?id=404 for a discussion and helper program
+    // to determine the following heuristic.
+    // EIGEN_GEMM_TO_COEFFBASED_THRESHOLD is typically defined to 20 in GeneralProduct.h,
+    // unless it has been specialized by the user or for a given architecture.
+    // Note that the condition rhs.rows()>0 was required because lazy product is (was?) not happy with empty inputs.
+    // I'm not sure it is still required.
+    if((rhs.rows()+dst.rows()+dst.cols())<EIGEN_GEMM_TO_COEFFBASED_THRESHOLD && rhs.rows()>0)
+      lazyproduct::eval_dynamic(dst, lhs, rhs, internal::assign_op<typename Dst::Scalar,Scalar>());
     else
     {
       dst.setZero();
@@ -439,8 +449,8 @@ struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,GemmProduct>
   template<typename Dst>
   static void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
   {
-    if((rhs.rows()+dst.rows()+dst.cols())<20 && rhs.rows()>0)
-      lazyproduct::addTo(dst, lhs, rhs);
+    if((rhs.rows()+dst.rows()+dst.cols())<EIGEN_GEMM_TO_COEFFBASED_THRESHOLD && rhs.rows()>0)
+      lazyproduct::eval_dynamic(dst, lhs, rhs, internal::add_assign_op<typename Dst::Scalar,Scalar>());
     else
       scaleAndAddTo(dst,lhs, rhs, Scalar(1));
   }
@@ -448,8 +458,8 @@ struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,GemmProduct>
   template<typename Dst>
   static void subTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
   {
-    if((rhs.rows()+dst.rows()+dst.cols())<20 && rhs.rows()>0)
-      lazyproduct::subTo(dst, lhs, rhs);
+    if((rhs.rows()+dst.rows()+dst.cols())<EIGEN_GEMM_TO_COEFFBASED_THRESHOLD && rhs.rows()>0)
+      lazyproduct::eval_dynamic(dst, lhs, rhs, internal::sub_assign_op<typename Dst::Scalar,Scalar>());
     else
       scaleAndAddTo(dst, lhs, rhs, Scalar(-1));
   }
@@ -461,11 +471,25 @@ struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,GemmProduct>
     if(a_lhs.cols()==0 || a_lhs.rows()==0 || a_rhs.cols()==0)
       return;
 
+    if (dst.cols() == 1)
+    {
+      // Fallback to GEMV if either the lhs or rhs is a runtime vector
+      typename Dest::ColXpr dst_vec(dst.col(0));
+      return internal::generic_product_impl<Lhs,typename Rhs::ConstColXpr,DenseShape,DenseShape,GemvProduct>
+        ::scaleAndAddTo(dst_vec, a_lhs, a_rhs.col(0), alpha);
+    }
+    else if (dst.rows() == 1)
+    {
+      // Fallback to GEMV if either the lhs or rhs is a runtime vector
+      typename Dest::RowXpr dst_vec(dst.row(0));
+      return internal::generic_product_impl<typename Lhs::ConstRowXpr,Rhs,DenseShape,DenseShape,GemvProduct>
+        ::scaleAndAddTo(dst_vec, a_lhs.row(0), a_rhs, alpha);
+    }
+
     typename internal::add_const_on_value_type<ActualLhsType>::type lhs = LhsBlasTraits::extract(a_lhs);
     typename internal::add_const_on_value_type<ActualRhsType>::type rhs = RhsBlasTraits::extract(a_rhs);
 
-    Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(a_lhs)
-                               * RhsBlasTraits::extractScalarFactor(a_rhs);
+    Scalar actualAlpha = combine_scalar_factors(alpha, a_lhs, a_rhs);
 
     typedef internal::gemm_blocking_space<(Dest::Flags&RowMajorBit) ? RowMajor : ColMajor,LhsScalar,RhsScalar,
             Dest::MaxRowsAtCompileTime,Dest::MaxColsAtCompileTime,MaxDepthAtCompileTime> BlockingType;
@@ -476,7 +500,8 @@ struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,GemmProduct>
         Index,
         LhsScalar, (ActualLhsTypeCleaned::Flags&RowMajorBit) ? RowMajor : ColMajor, bool(LhsBlasTraits::NeedToConjugate),
         RhsScalar, (ActualRhsTypeCleaned::Flags&RowMajorBit) ? RowMajor : ColMajor, bool(RhsBlasTraits::NeedToConjugate),
-        (Dest::Flags&RowMajorBit) ? RowMajor : ColMajor>,
+        (Dest::Flags&RowMajorBit) ? RowMajor : ColMajor,
+        Dest::InnerStrideAtCompileTime>,
       ActualLhsTypeCleaned, ActualRhsTypeCleaned, Dest, BlockingType> GemmFunctor;
 
     BlockingType blocking(dst.rows(), dst.cols(), lhs.cols(), 1, true);

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

@@ -25,51 +25,54 @@ namespace internal {
 **********************************************************************/
 
 // forward declarations (defined at the end of this file)
-template<typename LhsScalar, typename RhsScalar, typename Index, int mr, int nr, bool ConjLhs, bool ConjRhs, int UpLo>
+template<typename LhsScalar, typename RhsScalar, typename Index, int mr, int nr, bool ConjLhs, bool ConjRhs, int ResInnerStride, int UpLo>
 struct tribb_kernel;
   
 /* Optimized matrix-matrix product evaluating only one triangular half */
 template <typename Index,
           typename LhsScalar, int LhsStorageOrder, bool ConjugateLhs,
           typename RhsScalar, int RhsStorageOrder, bool ConjugateRhs,
-                              int ResStorageOrder, int  UpLo, int Version = Specialized>
+                              int ResStorageOrder, int ResInnerStride, 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, int Version>
-struct general_matrix_matrix_triangular_product<Index,LhsScalar,LhsStorageOrder,ConjugateLhs,RhsScalar,RhsStorageOrder,ConjugateRhs,RowMajor,UpLo,Version>
+                          typename RhsScalar, int RhsStorageOrder, bool ConjugateRhs,
+                          int ResInnerStride, int  UpLo, int Version>
+struct general_matrix_matrix_triangular_product<Index,LhsScalar,LhsStorageOrder,ConjugateLhs,RhsScalar,RhsStorageOrder,ConjugateRhs,RowMajor,ResInnerStride,UpLo,Version>
 {
   typedef typename ScalarBinaryOpTraits<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,
+                                      const RhsScalar* rhs, Index rhsStride, ResScalar* res, Index resIncr, Index resStride,
                                       const ResScalar& alpha, level3_blocking<RhsScalar,LhsScalar>& blocking)
   {
     general_matrix_matrix_triangular_product<Index,
         RhsScalar, RhsStorageOrder==RowMajor ? ColMajor : RowMajor, ConjugateRhs,
         LhsScalar, LhsStorageOrder==RowMajor ? ColMajor : RowMajor, ConjugateLhs,
-        ColMajor, UpLo==Lower?Upper:Lower>
-      ::run(size,depth,rhs,rhsStride,lhs,lhsStride,res,resStride,alpha,blocking);
+        ColMajor, ResInnerStride, UpLo==Lower?Upper:Lower>
+      ::run(size,depth,rhs,rhsStride,lhs,lhsStride,res,resIncr,resStride,alpha,blocking);
   }
 };
 
 template <typename Index, typename LhsScalar, int LhsStorageOrder, bool ConjugateLhs,
-                          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>
+                          typename RhsScalar, int RhsStorageOrder, bool ConjugateRhs,
+                          int ResInnerStride, int  UpLo, int Version>
+struct general_matrix_matrix_triangular_product<Index,LhsScalar,LhsStorageOrder,ConjugateLhs,RhsScalar,RhsStorageOrder,ConjugateRhs,ColMajor,ResInnerStride,UpLo,Version>
 {
   typedef typename ScalarBinaryOpTraits<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,
+                                      const RhsScalar* _rhs, Index rhsStride,
+                                      ResScalar* _res, Index resIncr, Index resStride,
                                       const ResScalar& alpha, level3_blocking<LhsScalar,RhsScalar>& blocking)
   {
     typedef gebp_traits<LhsScalar,RhsScalar> Traits;
 
     typedef const_blas_data_mapper<LhsScalar, Index, LhsStorageOrder> LhsMapper;
     typedef const_blas_data_mapper<RhsScalar, 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();
     Index mc = (std::min)(size,blocking.mc());
@@ -84,10 +87,10 @@ struct general_matrix_matrix_triangular_product<Index,LhsScalar,LhsStorageOrder,
     ei_declare_aligned_stack_constructed_variable(LhsScalar, blockA, sizeA, blocking.blockA());
     ei_declare_aligned_stack_constructed_variable(RhsScalar, blockB, sizeB, blocking.blockB());
 
-    gemm_pack_lhs<LhsScalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs;
+    gemm_pack_lhs<LhsScalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, typename Traits::LhsPacket4Packing, LhsStorageOrder> pack_lhs;
     gemm_pack_rhs<RhsScalar, Index, RhsMapper, Traits::nr, RhsStorageOrder> pack_rhs;
     gebp_kernel<LhsScalar, RhsScalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp;
-    tribb_kernel<LhsScalar, RhsScalar, Index, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs, UpLo> sybb;
+    tribb_kernel<LhsScalar, RhsScalar, Index, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs, ResInnerStride, UpLo> sybb;
 
     for(Index k2=0; k2<depth; k2+=kc)
     {
@@ -110,8 +113,7 @@ struct general_matrix_matrix_triangular_product<Index,LhsScalar,LhsStorageOrder,
           gebp(res.getSubMapper(i2, 0), blockA, blockB, actual_mc, actual_kc,
                (std::min)(size,i2), alpha, -1, -1, 0, 0);
 
-
-        sybb(_res+resStride*i2 + i2, resStride, blockA, blockB + actual_kc*i2, actual_mc, actual_kc, alpha);
+        sybb(_res+resStride*i2 + resIncr*i2, resIncr, resStride, blockA, blockB + actual_kc*i2, actual_mc, actual_kc, alpha);
 
         if (UpLo==Upper)
         {
@@ -133,7 +135,7 @@ struct general_matrix_matrix_triangular_product<Index,LhsScalar,LhsStorageOrder,
 //   while the triangular block overlapping the diagonal is evaluated into a
 //   small temporary buffer which is then accumulated into the result using a
 //   triangular traversal.
-template<typename LhsScalar, typename RhsScalar, typename Index, int mr, int nr, bool ConjLhs, bool ConjRhs, int UpLo>
+template<typename LhsScalar, typename RhsScalar, typename Index, int mr, int nr, bool ConjLhs, bool ConjRhs, int ResInnerStride, int UpLo>
 struct tribb_kernel
 {
   typedef gebp_traits<LhsScalar,RhsScalar,ConjLhs,ConjRhs> Traits;
@@ -142,11 +144,13 @@ struct tribb_kernel
   enum {
     BlockSize  = meta_least_common_multiple<EIGEN_PLAIN_ENUM_MAX(mr,nr),EIGEN_PLAIN_ENUM_MIN(mr,nr)>::ret
   };
-  void operator()(ResScalar* _res, Index resStride, const LhsScalar* blockA, const RhsScalar* blockB, Index size, Index depth, const ResScalar& alpha)
+  void operator()(ResScalar* _res, Index resIncr, Index resStride, const LhsScalar* blockA, const RhsScalar* blockB, Index size, Index depth, const ResScalar& alpha)
   {
-    typedef blas_data_mapper<ResScalar, Index, ColMajor> ResMapper;
-    ResMapper res(_res, resStride);
-    gebp_kernel<LhsScalar, RhsScalar, Index, ResMapper, mr, nr, ConjLhs, ConjRhs> gebp_kernel;
+    typedef blas_data_mapper<ResScalar, Index, ColMajor, Unaligned, ResInnerStride> ResMapper;
+    typedef blas_data_mapper<ResScalar, Index, ColMajor, Unaligned> BufferMapper;
+    ResMapper res(_res, resStride, resIncr);
+    gebp_kernel<LhsScalar, RhsScalar, Index, ResMapper, mr, nr, ConjLhs, ConjRhs> gebp_kernel1;
+    gebp_kernel<LhsScalar, RhsScalar, Index, BufferMapper, mr, nr, ConjLhs, ConjRhs> gebp_kernel2;
 
     Matrix<ResScalar,BlockSize,BlockSize,ColMajor> buffer((internal::constructor_without_unaligned_array_assert()));
 
@@ -158,31 +162,32 @@ struct tribb_kernel
       const RhsScalar* actual_b = blockB+j*depth;
 
       if(UpLo==Upper)
-        gebp_kernel(res.getSubMapper(0, j), blockA, actual_b, j, depth, actualBlockSize, alpha,
-                    -1, -1, 0, 0);
-
+        gebp_kernel1(res.getSubMapper(0, j), blockA, actual_b, j, depth, actualBlockSize, alpha,
+                     -1, -1, 0, 0);
+      
       // selfadjoint micro block
       {
         Index i = j;
         buffer.setZero();
         // 1 - apply the kernel on the temporary buffer
-        gebp_kernel(ResMapper(buffer.data(), BlockSize), blockA+depth*i, actual_b, actualBlockSize, depth, actualBlockSize, alpha,
-                    -1, -1, 0, 0);
+        gebp_kernel2(BufferMapper(buffer.data(), BlockSize), blockA+depth*i, actual_b, actualBlockSize, depth, actualBlockSize, alpha,
+                     -1, -1, 0, 0);
+
         // 2 - triangular accumulation
         for(Index j1=0; j1<actualBlockSize; ++j1)
         {
-          ResScalar* r = &res(i, j + j1);
+          typename ResMapper::LinearMapper r = res.getLinearMapper(i,j+j1);
           for(Index i1=UpLo==Lower ? j1 : 0;
               UpLo==Lower ? i1<actualBlockSize : i1<=j1; ++i1)
-            r[i1] += buffer(i1,j1);
+            r(i1) += buffer(i1,j1);
         }
       }
 
       if(UpLo==Lower)
       {
         Index i = j+actualBlockSize;
-        gebp_kernel(res.getSubMapper(i, j), blockA+depth*i, actual_b, size-i, 
-                    depth, actualBlockSize, alpha, -1, -1, 0, 0);
+        gebp_kernel1(res.getSubMapper(i, j), blockA+depth*i, actual_b, size-i, 
+                     depth, actualBlockSize, alpha, -1, -1, 0, 0);
       }
     }
   }
@@ -286,23 +291,24 @@ struct general_product_to_triangular_selector<MatrixType,ProductType,UpLo,false>
     internal::general_matrix_matrix_triangular_product<Index,
       typename Lhs::Scalar, LhsIsRowMajor ? RowMajor : ColMajor, LhsBlasTraits::NeedToConjugate,
       typename Rhs::Scalar, RhsIsRowMajor ? RowMajor : ColMajor, RhsBlasTraits::NeedToConjugate,
-      IsRowMajor ? RowMajor : ColMajor, UpLo&(Lower|Upper)>
+      IsRowMajor ? RowMajor : ColMajor, MatrixType::InnerStrideAtCompileTime, UpLo&(Lower|Upper)>
       ::run(size, depth,
             &actualLhs.coeffRef(SkipDiag&&(UpLo&Lower)==Lower ? 1 : 0,0), actualLhs.outerStride(),
             &actualRhs.coeffRef(0,SkipDiag&&(UpLo&Upper)==Upper ? 1 : 0), actualRhs.outerStride(),
-            mat.data() + (SkipDiag ? (bool(IsRowMajor) != ((UpLo&Lower)==Lower) ? 1 : mat.outerStride() ) : 0), mat.outerStride(), actualAlpha, blocking);
+            mat.data() + (SkipDiag ? (bool(IsRowMajor) != ((UpLo&Lower)==Lower) ? mat.innerStride() : mat.outerStride() ) : 0),
+            mat.innerStride(), mat.outerStride(), actualAlpha, blocking);
   }
 };
 
 template<typename MatrixType, unsigned int UpLo>
 template<typename ProductType>
-TriangularView<MatrixType,UpLo>& TriangularViewImpl<MatrixType,UpLo,Dense>::_assignProduct(const ProductType& prod, const Scalar& alpha, bool beta)
+EIGEN_DEVICE_FUNC TriangularView<MatrixType,UpLo>& TriangularViewImpl<MatrixType,UpLo,Dense>::_assignProduct(const ProductType& prod, const Scalar& alpha, bool beta)
 {
   EIGEN_STATIC_ASSERT((UpLo&UnitDiag)==0, WRITING_TO_TRIANGULAR_PART_WITH_UNIT_DIAGONAL_IS_NOT_SUPPORTED);
   eigen_assert(derived().nestedExpression().rows() == prod.rows() && derived().cols() == prod.cols());
-  
+
   general_product_to_triangular_selector<MatrixType, ProductType, UpLo, internal::traits<ProductType>::InnerSize==1>::run(derived().nestedExpression().const_cast_derived(), prod, alpha, beta);
-  
+
   return derived();
 }
 

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

@@ -37,10 +37,10 @@ namespace Eigen {
 
 namespace internal {
 
-template <typename Index, typename Scalar, int AStorageOrder, bool ConjugateA, int ResStorageOrder, int  UpLo>
+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> {};
+         Index,Scalar,AStorageOrder,ConjugateA,Scalar,AStorageOrder,ConjugateA,ResStorageOrder,1,UpLo,BuiltIn> {};
 
 
 // try to go to BLAS specialization
@@ -48,9 +48,9 @@ struct general_matrix_matrix_rankupdate :
 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> { \
+               Scalar,RhsStorageOrder,ConjugateRhs,ColMajor,1,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, level3_blocking<Scalar, Scalar>& blocking) \
+                          const Scalar* rhs, Index rhsStride, Scalar* res, Index resIncr, Index resStride, Scalar alpha, level3_blocking<Scalar, Scalar>& blocking) \
   { \
     if ( lhs==rhs && ((UpLo&(Lower|Upper))==UpLo) ) { \
       general_matrix_matrix_rankupdate<Index,Scalar,LhsStorageOrder,ConjugateLhs,ColMajor,UpLo> \
@@ -59,8 +59,8 @@ struct general_matrix_matrix_triangular_product<Index,Scalar,LhsStorageOrder,Con
       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,blocking); \
+        ColMajor, 1, UpLo, BuiltIn> \
+      ::run(size,depth,lhs,lhsStride,rhs,rhsStride,res,resIncr,resStride,alpha,blocking); \
     } \
   } \
 };

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

@@ -51,20 +51,22 @@ template< \
   typename Index, \
   int LhsStorageOrder, bool ConjugateLhs, \
   int RhsStorageOrder, bool ConjugateRhs> \
-struct general_matrix_matrix_product<Index,EIGTYPE,LhsStorageOrder,ConjugateLhs,EIGTYPE,RhsStorageOrder,ConjugateRhs,ColMajor> \
+struct general_matrix_matrix_product<Index,EIGTYPE,LhsStorageOrder,ConjugateLhs,EIGTYPE,RhsStorageOrder,ConjugateRhs,ColMajor,1> \
 { \
 typedef gebp_traits<EIGTYPE,EIGTYPE> Traits; \
 \
 static void run(Index rows, Index cols, Index depth, \
   const EIGTYPE* _lhs, Index lhsStride, \
   const EIGTYPE* _rhs, Index rhsStride, \
-  EIGTYPE* res, Index resStride, \
+  EIGTYPE* res, Index resIncr, Index resStride, \
   EIGTYPE alpha, \
   level3_blocking<EIGTYPE, EIGTYPE>& /*blocking*/, \
   GemmParallelInfo<Index>* /*info = 0*/) \
 { \
   using std::conj; \
 \
+  EIGEN_ONLY_USED_FOR_DEBUG(resIncr); \
+  eigen_assert(resIncr == 1); \
   char transa, transb; \
   BlasIndex m, n, k, lda, ldb, ldc; \
   const EIGTYPE *a, *b; \