Chris@49: // Copyright (C) 2008-2013 NICTA (www.nicta.com.au)
Chris@49: // Copyright (C) 2008-2013 Conrad Sanderson
Chris@49: // 
Chris@49: // This Source Code Form is subject to the terms of the Mozilla Public
Chris@49: // License, v. 2.0. If a copy of the MPL was not distributed with this
Chris@49: // file, You can obtain one at http://mozilla.org/MPL/2.0/.
Chris@49: 
Chris@49: 
Chris@49: //! \addtogroup gemm
Chris@49: //! @{
Chris@49: 
Chris@49: 
Chris@49: 
Chris@49: //! for tiny square matrices, size <= 4x4
Chris@49: template<const bool do_trans_A=false, const bool use_alpha=false, const bool use_beta=false>
Chris@49: class gemm_emul_tinysq
Chris@49:   {
Chris@49:   public:
Chris@49:   
Chris@49:   
Chris@49:   template<typename eT, typename TA, typename TB>
Chris@49:   arma_hot
Chris@49:   inline
Chris@49:   static
Chris@49:   void
Chris@49:   apply
Chris@49:     (
Chris@49:           Mat<eT>& C,
Chris@49:     const TA&      A,
Chris@49:     const TB&      B,
Chris@49:     const eT alpha = eT(1),
Chris@49:     const eT beta  = eT(0)
Chris@49:     )
Chris@49:     {
Chris@49:     arma_extra_debug_sigprint();
Chris@49:     
Chris@49:     switch(A.n_rows)
Chris@49:       {
Chris@49:       case 4:
Chris@49:         gemv_emul_tinysq<do_trans_A, use_alpha, use_beta>::apply( C.colptr(3), A, B.colptr(3), alpha, beta );
Chris@49:         
Chris@49:       case 3:
Chris@49:         gemv_emul_tinysq<do_trans_A, use_alpha, use_beta>::apply( C.colptr(2), A, B.colptr(2), alpha, beta );
Chris@49:         
Chris@49:       case 2:
Chris@49:         gemv_emul_tinysq<do_trans_A, use_alpha, use_beta>::apply( C.colptr(1), A, B.colptr(1), alpha, beta );
Chris@49:         
Chris@49:       case 1:
Chris@49:         gemv_emul_tinysq<do_trans_A, use_alpha, use_beta>::apply( C.colptr(0), A, B.colptr(0), alpha, beta );
Chris@49:         
Chris@49:       default:
Chris@49:         ;
Chris@49:       }
Chris@49:     }
Chris@49:   
Chris@49:   };
Chris@49: 
Chris@49: 
Chris@49: 
Chris@49: template<const bool do_trans_A=false, const bool do_trans_B=false, const bool use_alpha=false, const bool use_beta=false>
Chris@49: class gemm_emul_large
Chris@49:   {
Chris@49:   public:
Chris@49:   
Chris@49:   template<typename eT, typename TA, typename TB>
Chris@49:   arma_hot
Chris@49:   inline
Chris@49:   static
Chris@49:   void
Chris@49:   apply
Chris@49:     (
Chris@49:           Mat<eT>& C,
Chris@49:     const TA&      A,
Chris@49:     const TB&      B,
Chris@49:     const eT alpha = eT(1),
Chris@49:     const eT beta  = eT(0)
Chris@49:     )
Chris@49:     {
Chris@49:     arma_extra_debug_sigprint();
Chris@49: 
Chris@49:     const uword A_n_rows = A.n_rows;
Chris@49:     const uword A_n_cols = A.n_cols;
Chris@49:     
Chris@49:     const uword B_n_rows = B.n_rows;
Chris@49:     const uword B_n_cols = B.n_cols;
Chris@49:     
Chris@49:     if( (do_trans_A == false) && (do_trans_B == false) )
Chris@49:       {
Chris@49:       arma_aligned podarray<eT> tmp(A_n_cols);
Chris@49:       
Chris@49:       eT* A_rowdata = tmp.memptr();
Chris@49:       
Chris@49:       for(uword row_A=0; row_A < A_n_rows; ++row_A)
Chris@49:         {
Chris@49:         //tmp.copy_row(A, row_A);
Chris@49:         const eT acc0 = op_dot::dot_and_copy_row(A_rowdata, A, row_A, B.colptr(0), A_n_cols);
Chris@49:         
Chris@49:         if( (use_alpha == false) && (use_beta == false) )
Chris@49:           {
Chris@49:           C.at(row_A,0) = acc0;
Chris@49:           }
Chris@49:         else
Chris@49:         if( (use_alpha == true) && (use_beta == false) )
Chris@49:           {
Chris@49:           C.at(row_A,0) = alpha * acc0;
Chris@49:           }
Chris@49:         else
Chris@49:         if( (use_alpha == false) && (use_beta == true) )
Chris@49:           {
Chris@49:           C.at(row_A,0) = acc0 + beta*C.at(row_A,0);
Chris@49:           }
Chris@49:         else
Chris@49:         if( (use_alpha == true) && (use_beta == true) )
Chris@49:           {
Chris@49:           C.at(row_A,0) = alpha*acc0 + beta*C.at(row_A,0);
Chris@49:           }
Chris@49: 
Chris@49:         //for(uword col_B=0; col_B < B_n_cols; ++col_B)
Chris@49:         for(uword col_B=1; col_B < B_n_cols; ++col_B)
Chris@49:           {
Chris@49:           const eT acc = op_dot::direct_dot_arma(B_n_rows, A_rowdata, B.colptr(col_B));
Chris@49:           
Chris@49:           if( (use_alpha == false) && (use_beta == false) )
Chris@49:             {
Chris@49:             C.at(row_A,col_B) = acc;
Chris@49:             }
Chris@49:           else
Chris@49:           if( (use_alpha == true) && (use_beta == false) )
Chris@49:             {
Chris@49:             C.at(row_A,col_B) = alpha * acc;
Chris@49:             }
Chris@49:           else
Chris@49:           if( (use_alpha == false) && (use_beta == true) )
Chris@49:             {
Chris@49:             C.at(row_A,col_B) = acc + beta*C.at(row_A,col_B);
Chris@49:             }
Chris@49:           else
Chris@49:           if( (use_alpha == true) && (use_beta == true) )
Chris@49:             {
Chris@49:             C.at(row_A,col_B) = alpha*acc + beta*C.at(row_A,col_B);
Chris@49:             }
Chris@49:           
Chris@49:           }
Chris@49:         }
Chris@49:       }
Chris@49:     else
Chris@49:     if( (do_trans_A == true) && (do_trans_B == false) )
Chris@49:       {
Chris@49:       for(uword col_A=0; col_A < A_n_cols; ++col_A)
Chris@49:         {
Chris@49:         // col_A is interpreted as row_A when storing the results in matrix C
Chris@49:         
Chris@49:         const eT* A_coldata = A.colptr(col_A);
Chris@49:         
Chris@49:         for(uword col_B=0; col_B < B_n_cols; ++col_B)
Chris@49:           {
Chris@49:           const eT acc = op_dot::direct_dot_arma(B_n_rows, A_coldata, B.colptr(col_B));
Chris@49:           
Chris@49:           if( (use_alpha == false) && (use_beta == false) )
Chris@49:             {
Chris@49:             C.at(col_A,col_B) = acc;
Chris@49:             }
Chris@49:           else
Chris@49:           if( (use_alpha == true) && (use_beta == false) )
Chris@49:             {
Chris@49:             C.at(col_A,col_B) = alpha * acc;
Chris@49:             }
Chris@49:           else
Chris@49:           if( (use_alpha == false) && (use_beta == true) )
Chris@49:             {
Chris@49:             C.at(col_A,col_B) = acc + beta*C.at(col_A,col_B);
Chris@49:             }
Chris@49:           else
Chris@49:           if( (use_alpha == true) && (use_beta == true) )
Chris@49:             {
Chris@49:             C.at(col_A,col_B) = alpha*acc + beta*C.at(col_A,col_B);
Chris@49:             }
Chris@49:           
Chris@49:           }
Chris@49:         }
Chris@49:       }
Chris@49:     else
Chris@49:     if( (do_trans_A == false) && (do_trans_B == true) )
Chris@49:       {
Chris@49:       Mat<eT> BB;
Chris@49:       op_strans::apply_noalias(BB, B);
Chris@49:       
Chris@49:       gemm_emul_large<false, false, use_alpha, use_beta>::apply(C, A, BB, alpha, beta);
Chris@49:       }
Chris@49:     else
Chris@49:     if( (do_trans_A == true) && (do_trans_B == true) )
Chris@49:       {
Chris@49:       // mat B_tmp = trans(B);
Chris@49:       // dgemm_arma<true, false,  use_alpha, use_beta>::apply(C, A, B_tmp, alpha, beta);
Chris@49:       
Chris@49:       
Chris@49:       // By using the trans(A)*trans(B) = trans(B*A) equivalency,
Chris@49:       // transpose operations are not needed
Chris@49:       
Chris@49:       arma_aligned podarray<eT> tmp(B.n_cols);
Chris@49:       eT* B_rowdata = tmp.memptr();
Chris@49:       
Chris@49:       for(uword row_B=0; row_B < B_n_rows; ++row_B)
Chris@49:         {
Chris@49:         tmp.copy_row(B, row_B);
Chris@49:         
Chris@49:         for(uword col_A=0; col_A < A_n_cols; ++col_A)
Chris@49:           {
Chris@49:           const eT acc = op_dot::direct_dot_arma(A_n_rows, B_rowdata, A.colptr(col_A));
Chris@49:           
Chris@49:           if( (use_alpha == false) && (use_beta == false) )
Chris@49:             {
Chris@49:             C.at(col_A,row_B) = acc;
Chris@49:             }
Chris@49:           else
Chris@49:           if( (use_alpha == true) && (use_beta == false) )
Chris@49:             {
Chris@49:             C.at(col_A,row_B) = alpha * acc;
Chris@49:             }
Chris@49:           else
Chris@49:           if( (use_alpha == false) && (use_beta == true) )
Chris@49:             {
Chris@49:             C.at(col_A,row_B) = acc + beta*C.at(col_A,row_B);
Chris@49:             }
Chris@49:           else
Chris@49:           if( (use_alpha == true) && (use_beta == true) )
Chris@49:             {
Chris@49:             C.at(col_A,row_B) = alpha*acc + beta*C.at(col_A,row_B);
Chris@49:             }
Chris@49:           
Chris@49:           }
Chris@49:         }
Chris@49:       
Chris@49:       }
Chris@49:     }
Chris@49:   
Chris@49:   };
Chris@49:   
Chris@49: 
Chris@49: 
Chris@49: template<const bool do_trans_A=false, const bool do_trans_B=false, const bool use_alpha=false, const bool use_beta=false>
Chris@49: class gemm_emul
Chris@49:   {
Chris@49:   public:
Chris@49:   
Chris@49:   
Chris@49:   template<typename eT, typename TA, typename TB>
Chris@49:   arma_hot
Chris@49:   inline
Chris@49:   static
Chris@49:   void
Chris@49:   apply
Chris@49:     (
Chris@49:           Mat<eT>& C,
Chris@49:     const TA&      A,
Chris@49:     const TB&      B,
Chris@49:     const eT alpha = eT(1),
Chris@49:     const eT beta  = eT(0),
Chris@49:     const typename arma_not_cx<eT>::result* junk = 0
Chris@49:     )
Chris@49:     {
Chris@49:     arma_extra_debug_sigprint();
Chris@49:     arma_ignore(junk);
Chris@49:     
Chris@49:     const uword A_n_rows = A.n_rows;
Chris@49:     const uword A_n_cols = A.n_cols;
Chris@49:     
Chris@49:     const uword B_n_rows = B.n_rows;
Chris@49:     const uword B_n_cols = B.n_cols;
Chris@49:     
Chris@49:     if( (A_n_rows <= 4) && (A_n_rows == A_n_cols) && (A_n_rows == B_n_rows) && (B_n_rows == B_n_cols) )
Chris@49:       {
Chris@49:       if(do_trans_B == false)
Chris@49:         {
Chris@49:         gemm_emul_tinysq<do_trans_A, use_alpha, use_beta>::apply(C, A, B, alpha, beta);
Chris@49:         }
Chris@49:       else
Chris@49:         {
Chris@49:         Mat<eT> BB(A_n_rows, A_n_rows);
Chris@49:         op_strans::apply_noalias_tinysq(BB, B);
Chris@49:         
Chris@49:         gemm_emul_tinysq<do_trans_A, use_alpha, use_beta>::apply(C, A, BB, alpha, beta);
Chris@49:         }
Chris@49:       }
Chris@49:     else
Chris@49:       {
Chris@49:       gemm_emul_large<do_trans_A, do_trans_B, use_alpha, use_beta>::apply(C, A, B, alpha, beta);
Chris@49:       }
Chris@49:     }
Chris@49:   
Chris@49:   
Chris@49:   
Chris@49:   template<typename eT>
Chris@49:   arma_hot
Chris@49:   inline
Chris@49:   static
Chris@49:   void
Chris@49:   apply
Chris@49:     (
Chris@49:           Mat<eT>& C,
Chris@49:     const Mat<eT>& A,
Chris@49:     const Mat<eT>& B,
Chris@49:     const eT alpha = eT(1),
Chris@49:     const eT beta  = eT(0),
Chris@49:     const typename arma_cx_only<eT>::result* junk = 0
Chris@49:     )
Chris@49:     {
Chris@49:     arma_extra_debug_sigprint();
Chris@49:     arma_ignore(junk);
Chris@49:     
Chris@49:     // "better than nothing" handling of hermitian transposes for complex number matrices
Chris@49:     
Chris@49:     Mat<eT> tmp_A;
Chris@49:     Mat<eT> tmp_B;
Chris@49:     
Chris@49:     if(do_trans_A)
Chris@49:       {
Chris@49:       op_htrans::apply_noalias(tmp_A, A);
Chris@49:       }
Chris@49:     
Chris@49:     if(do_trans_B)
Chris@49:       {
Chris@49:       op_htrans::apply_noalias(tmp_B, B);
Chris@49:       }
Chris@49:     
Chris@49:     const Mat<eT>& AA = (do_trans_A == false) ? A : tmp_A;
Chris@49:     const Mat<eT>& BB = (do_trans_B == false) ? B : tmp_B;
Chris@49:     
Chris@49:     const uword A_n_rows = AA.n_rows;
Chris@49:     const uword A_n_cols = AA.n_cols;
Chris@49:     
Chris@49:     const uword B_n_rows = BB.n_rows;
Chris@49:     const uword B_n_cols = BB.n_cols;
Chris@49:     
Chris@49:     if( (A_n_rows <= 4) && (A_n_rows == A_n_cols) && (A_n_rows == B_n_rows) && (B_n_rows == B_n_cols) )
Chris@49:       {
Chris@49:       gemm_emul_tinysq<false, use_alpha, use_beta>::apply(C, AA, BB, alpha, beta);
Chris@49:       }
Chris@49:     else
Chris@49:       {
Chris@49:       gemm_emul_large<false, false, use_alpha, use_beta>::apply(C, AA, BB, alpha, beta);
Chris@49:       }
Chris@49:     }
Chris@49: 
Chris@49:   };
Chris@49: 
Chris@49: 
Chris@49: 
Chris@49: //! \brief
Chris@49: //! Wrapper for ATLAS/BLAS dgemm function, using template arguments to control the arguments passed to dgemm.
Chris@49: //! Matrix 'C' is assumed to have been set to the correct size (i.e. taking into account transposes)
Chris@49: 
Chris@49: template<const bool do_trans_A=false, const bool do_trans_B=false, const bool use_alpha=false, const bool use_beta=false>
Chris@49: class gemm
Chris@49:   {
Chris@49:   public:
Chris@49:   
Chris@49:   template<typename eT, typename TA, typename TB>
Chris@49:   inline
Chris@49:   static
Chris@49:   void
Chris@49:   apply_blas_type( Mat<eT>& C, const TA& A, const TB& B, const eT alpha = eT(1), const eT beta = eT(0) )
Chris@49:     {
Chris@49:     arma_extra_debug_sigprint();
Chris@49:     
Chris@49:     const uword threshold = (is_Mat_fixed<TA>::value && is_Mat_fixed<TB>::value)
Chris@49:                             ? (is_complex<eT>::value ? 16u : 64u)
Chris@49:                             : (is_complex<eT>::value ? 16u : 48u);
Chris@49:     
Chris@49:     if( (A.n_elem <= threshold) && (B.n_elem <= threshold) )
Chris@49:       {
Chris@49:       gemm_emul<do_trans_A, do_trans_B, use_alpha, use_beta>::apply(C,A,B,alpha,beta);
Chris@49:       }
Chris@49:     else
Chris@49:       {
Chris@49:       #if defined(ARMA_USE_ATLAS)
Chris@49:         {
Chris@49:         arma_extra_debug_print("atlas::cblas_gemm()");
Chris@49:         
Chris@49:         atlas::cblas_gemm<eT>
Chris@49:           (
Chris@49:           atlas::CblasColMajor,
Chris@49:           (do_trans_A) ? ( is_complex<eT>::value ? CblasConjTrans : atlas::CblasTrans ) : atlas::CblasNoTrans,
Chris@49:           (do_trans_B) ? ( is_complex<eT>::value ? CblasConjTrans : atlas::CblasTrans ) : atlas::CblasNoTrans,
Chris@49:           C.n_rows,
Chris@49:           C.n_cols,
Chris@49:           (do_trans_A) ? A.n_rows : A.n_cols,
Chris@49:           (use_alpha) ? alpha : eT(1),
Chris@49:           A.mem,
Chris@49:           (do_trans_A) ? A.n_rows : C.n_rows,
Chris@49:           B.mem,
Chris@49:           (do_trans_B) ? C.n_cols : ( (do_trans_A) ? A.n_rows : A.n_cols ),
Chris@49:           (use_beta) ? beta : eT(0),
Chris@49:           C.memptr(),
Chris@49:           C.n_rows
Chris@49:           );
Chris@49:         }
Chris@49:       #elif defined(ARMA_USE_BLAS)
Chris@49:         {
Chris@49:         arma_extra_debug_print("blas::gemm()");
Chris@49:         
Chris@49:         const char trans_A = (do_trans_A) ? ( is_complex<eT>::value ? 'C' : 'T' ) : 'N';
Chris@49:         const char trans_B = (do_trans_B) ? ( is_complex<eT>::value ? 'C' : 'T' ) : 'N';
Chris@49:         
Chris@49:         const blas_int m   = C.n_rows;
Chris@49:         const blas_int n   = C.n_cols;
Chris@49:         const blas_int k   = (do_trans_A) ? A.n_rows : A.n_cols;
Chris@49:         
Chris@49:         const eT local_alpha = (use_alpha) ? alpha : eT(1);
Chris@49:         
Chris@49:         const blas_int lda = (do_trans_A) ? k : m;
Chris@49:         const blas_int ldb = (do_trans_B) ? n : k;
Chris@49:         
Chris@49:         const eT local_beta  = (use_beta) ? beta : eT(0);
Chris@49:         
Chris@49:         arma_extra_debug_print( arma_boost::format("blas::gemm(): trans_A = %c") % trans_A );
Chris@49:         arma_extra_debug_print( arma_boost::format("blas::gemm(): trans_B = %c") % trans_B );
Chris@49:         
Chris@49:         blas::gemm<eT>
Chris@49:           (
Chris@49:           &trans_A,
Chris@49:           &trans_B,
Chris@49:           &m,
Chris@49:           &n,
Chris@49:           &k,
Chris@49:           &local_alpha,
Chris@49:           A.mem,
Chris@49:           &lda,
Chris@49:           B.mem,
Chris@49:           &ldb,
Chris@49:           &local_beta,
Chris@49:           C.memptr(),
Chris@49:           &m
Chris@49:           );
Chris@49:         }
Chris@49:       #else
Chris@49:         {
Chris@49:         gemm_emul<do_trans_A, do_trans_B, use_alpha, use_beta>::apply(C,A,B,alpha,beta);
Chris@49:         }
Chris@49:       #endif
Chris@49:       }
Chris@49:     }
Chris@49:   
Chris@49:   
Chris@49:   
Chris@49:   //! immediate multiplication of matrices A and B, storing the result in C
Chris@49:   template<typename eT, typename TA, typename TB>
Chris@49:   inline
Chris@49:   static
Chris@49:   void
Chris@49:   apply( Mat<eT>& C, const TA& A, const TB& B, const eT alpha = eT(1), const eT beta = eT(0) )
Chris@49:     {
Chris@49:     gemm_emul<do_trans_A, do_trans_B, use_alpha, use_beta>::apply(C,A,B,alpha,beta);
Chris@49:     }
Chris@49:   
Chris@49:   
Chris@49:   
Chris@49:   template<typename TA, typename TB>
Chris@49:   arma_inline
Chris@49:   static
Chris@49:   void
Chris@49:   apply
Chris@49:     (
Chris@49:           Mat<float>& C,
Chris@49:     const TA&         A,
Chris@49:     const TB&         B,
Chris@49:     const float alpha = float(1),
Chris@49:     const float beta  = float(0)
Chris@49:     )
Chris@49:     {
Chris@49:     gemm<do_trans_A, do_trans_B, use_alpha, use_beta>::apply_blas_type(C,A,B,alpha,beta);
Chris@49:     }
Chris@49:   
Chris@49:   
Chris@49:   
Chris@49:   template<typename TA, typename TB>
Chris@49:   arma_inline
Chris@49:   static
Chris@49:   void
Chris@49:   apply
Chris@49:     (
Chris@49:           Mat<double>& C,
Chris@49:     const TA&          A,
Chris@49:     const TB&          B,
Chris@49:     const double alpha = double(1),
Chris@49:     const double beta  = double(0)
Chris@49:     )
Chris@49:     {
Chris@49:     gemm<do_trans_A, do_trans_B, use_alpha, use_beta>::apply_blas_type(C,A,B,alpha,beta);
Chris@49:     }
Chris@49:   
Chris@49:   
Chris@49:   
Chris@49:   template<typename TA, typename TB>
Chris@49:   arma_inline
Chris@49:   static
Chris@49:   void
Chris@49:   apply
Chris@49:     (
Chris@49:           Mat< std::complex<float> >& C,
Chris@49:     const TA&                         A,
Chris@49:     const TB&                         B,
Chris@49:     const std::complex<float> alpha = std::complex<float>(1),
Chris@49:     const std::complex<float> beta  = std::complex<float>(0)
Chris@49:     )
Chris@49:     {
Chris@49:     gemm<do_trans_A, do_trans_B, use_alpha, use_beta>::apply_blas_type(C,A,B,alpha,beta);
Chris@49:     }
Chris@49:   
Chris@49:   
Chris@49:   
Chris@49:   template<typename TA, typename TB>
Chris@49:   arma_inline
Chris@49:   static
Chris@49:   void
Chris@49:   apply
Chris@49:     (
Chris@49:           Mat< std::complex<double> >& C,
Chris@49:     const TA&                          A,
Chris@49:     const TB&                          B,
Chris@49:     const std::complex<double> alpha = std::complex<double>(1),
Chris@49:     const std::complex<double> beta  = std::complex<double>(0)
Chris@49:     )
Chris@49:     {
Chris@49:     gemm<do_trans_A, do_trans_B, use_alpha, use_beta>::apply_blas_type(C,A,B,alpha,beta);
Chris@49:     }
Chris@49:   
Chris@49:   };
Chris@49: 
Chris@49: 
Chris@49: 
Chris@49: //! @}