max@0: // Copyright (C) 2008-2011 NICTA (www.nicta.com.au)
max@0: // Copyright (C) 2008-2011 Conrad Sanderson
max@0: // 
max@0: // This file is part of the Armadillo C++ library.
max@0: // It is provided without any warranty of fitness
max@0: // for any purpose. You can redistribute this file
max@0: // and/or modify it under the terms of the GNU
max@0: // Lesser General Public License (LGPL) as published
max@0: // by the Free Software Foundation, either version 3
max@0: // of the License or (at your option) any later version.
max@0: // (see http://www.opensource.org/licenses for more info)
max@0: 
max@0: 
max@0: //! \addtogroup gemm
max@0: //! @{
max@0: 
max@0: 
max@0: 
max@0: //! for tiny square matrices, size <= 4x4
max@0: template<const bool do_trans_A=false, const bool use_alpha=false, const bool use_beta=false>
max@0: class gemm_emul_tinysq
max@0:   {
max@0:   public:
max@0:   
max@0:   
max@0:   template<typename eT>
max@0:   arma_hot
max@0:   inline
max@0:   static
max@0:   void
max@0:   apply
max@0:     (
max@0:           Mat<eT>& C,
max@0:     const Mat<eT>& A,
max@0:     const Mat<eT>& B,
max@0:     const eT alpha = eT(1),
max@0:     const eT beta  = eT(0)
max@0:     )
max@0:     {
max@0:     arma_extra_debug_sigprint();
max@0:     
max@0:     switch(A.n_rows)
max@0:       {
max@0:       case 4:
max@0:         gemv_emul_tinysq<do_trans_A, use_alpha, use_beta>::apply( C.colptr(3), A, B.colptr(3), alpha, beta );
max@0:         
max@0:       case 3:
max@0:         gemv_emul_tinysq<do_trans_A, use_alpha, use_beta>::apply( C.colptr(2), A, B.colptr(2), alpha, beta );
max@0:         
max@0:       case 2:
max@0:         gemv_emul_tinysq<do_trans_A, use_alpha, use_beta>::apply( C.colptr(1), A, B.colptr(1), alpha, beta );
max@0:         
max@0:       case 1:
max@0:         gemv_emul_tinysq<do_trans_A, use_alpha, use_beta>::apply( C.colptr(0), A, B.colptr(0), alpha, beta );
max@0:         
max@0:       default:
max@0:         ;
max@0:       }
max@0:     }
max@0:   
max@0:   };
max@0: 
max@0: 
max@0: 
max@0: template<const bool do_trans_A=false, const bool do_trans_B=false, const bool use_alpha=false, const bool use_beta=false>
max@0: class gemm_emul_large
max@0:   {
max@0:   public:
max@0:   
max@0:   template<typename eT>
max@0:   arma_hot
max@0:   inline
max@0:   static
max@0:   void
max@0:   apply
max@0:     (
max@0:           Mat<eT>& C,
max@0:     const Mat<eT>& A,
max@0:     const Mat<eT>& B,
max@0:     const eT alpha = eT(1),
max@0:     const eT beta  = eT(0)
max@0:     )
max@0:     {
max@0:     arma_extra_debug_sigprint();
max@0: 
max@0:     const uword A_n_rows = A.n_rows;
max@0:     const uword A_n_cols = A.n_cols;
max@0:     
max@0:     const uword B_n_rows = B.n_rows;
max@0:     const uword B_n_cols = B.n_cols;
max@0:     
max@0:     if( (do_trans_A == false) && (do_trans_B == false) )
max@0:       {
max@0:       arma_aligned podarray<eT> tmp(A_n_cols);
max@0:       eT* A_rowdata = tmp.memptr();
max@0:       
max@0:       for(uword row_A=0; row_A < A_n_rows; ++row_A)
max@0:         {
max@0:         tmp.copy_row(A, row_A);
max@0:         
max@0:         for(uword col_B=0; col_B < B_n_cols; ++col_B)
max@0:           {
max@0:           const eT acc = op_dot::direct_dot_arma(B_n_rows, A_rowdata, B.colptr(col_B));
max@0:           
max@0:           if( (use_alpha == false) && (use_beta == false) )
max@0:             {
max@0:             C.at(row_A,col_B) = acc;
max@0:             }
max@0:           else
max@0:           if( (use_alpha == true) && (use_beta == false) )
max@0:             {
max@0:             C.at(row_A,col_B) = alpha * acc;
max@0:             }
max@0:           else
max@0:           if( (use_alpha == false) && (use_beta == true) )
max@0:             {
max@0:             C.at(row_A,col_B) = acc + beta*C.at(row_A,col_B);
max@0:             }
max@0:           else
max@0:           if( (use_alpha == true) && (use_beta == true) )
max@0:             {
max@0:             C.at(row_A,col_B) = alpha*acc + beta*C.at(row_A,col_B);
max@0:             }
max@0:           
max@0:           }
max@0:         }
max@0:       }
max@0:     else
max@0:     if( (do_trans_A == true) && (do_trans_B == false) )
max@0:       {
max@0:       for(uword col_A=0; col_A < A_n_cols; ++col_A)
max@0:         {
max@0:         // col_A is interpreted as row_A when storing the results in matrix C
max@0:         
max@0:         const eT* A_coldata = A.colptr(col_A);
max@0:         
max@0:         for(uword col_B=0; col_B < B_n_cols; ++col_B)
max@0:           {
max@0:           const eT acc = op_dot::direct_dot_arma(B_n_rows, A_coldata, B.colptr(col_B));
max@0:           
max@0:           if( (use_alpha == false) && (use_beta == false) )
max@0:             {
max@0:             C.at(col_A,col_B) = acc;
max@0:             }
max@0:           else
max@0:           if( (use_alpha == true) && (use_beta == false) )
max@0:             {
max@0:             C.at(col_A,col_B) = alpha * acc;
max@0:             }
max@0:           else
max@0:           if( (use_alpha == false) && (use_beta == true) )
max@0:             {
max@0:             C.at(col_A,col_B) = acc + beta*C.at(col_A,col_B);
max@0:             }
max@0:           else
max@0:           if( (use_alpha == true) && (use_beta == true) )
max@0:             {
max@0:             C.at(col_A,col_B) = alpha*acc + beta*C.at(col_A,col_B);
max@0:             }
max@0:           
max@0:           }
max@0:         }
max@0:       }
max@0:     else
max@0:     if( (do_trans_A == false) && (do_trans_B == true) )
max@0:       {
max@0:       Mat<eT> BB;
max@0:       op_strans::apply_noalias(BB, B);
max@0:       
max@0:       gemm_emul_large<false, false, use_alpha, use_beta>::apply(C, A, BB, alpha, beta);
max@0:       }
max@0:     else
max@0:     if( (do_trans_A == true) && (do_trans_B == true) )
max@0:       {
max@0:       // mat B_tmp = trans(B);
max@0:       // dgemm_arma<true, false,  use_alpha, use_beta>::apply(C, A, B_tmp, alpha, beta);
max@0:       
max@0:       
max@0:       // By using the trans(A)*trans(B) = trans(B*A) equivalency,
max@0:       // transpose operations are not needed
max@0:       
max@0:       arma_aligned podarray<eT> tmp(B.n_cols);
max@0:       eT* B_rowdata = tmp.memptr();
max@0:       
max@0:       for(uword row_B=0; row_B < B_n_rows; ++row_B)
max@0:         {
max@0:         tmp.copy_row(B, row_B);
max@0:         
max@0:         for(uword col_A=0; col_A < A_n_cols; ++col_A)
max@0:           {
max@0:           const eT acc = op_dot::direct_dot_arma(A_n_rows, B_rowdata, A.colptr(col_A));
max@0:           
max@0:           if( (use_alpha == false) && (use_beta == false) )
max@0:             {
max@0:             C.at(col_A,row_B) = acc;
max@0:             }
max@0:           else
max@0:           if( (use_alpha == true) && (use_beta == false) )
max@0:             {
max@0:             C.at(col_A,row_B) = alpha * acc;
max@0:             }
max@0:           else
max@0:           if( (use_alpha == false) && (use_beta == true) )
max@0:             {
max@0:             C.at(col_A,row_B) = acc + beta*C.at(col_A,row_B);
max@0:             }
max@0:           else
max@0:           if( (use_alpha == true) && (use_beta == true) )
max@0:             {
max@0:             C.at(col_A,row_B) = alpha*acc + beta*C.at(col_A,row_B);
max@0:             }
max@0:           
max@0:           }
max@0:         }
max@0:       
max@0:       }
max@0:     }
max@0:   
max@0:   };
max@0:     
max@0:   
max@0:   
max@0: template<const bool do_trans_A=false, const bool do_trans_B=false, const bool use_alpha=false, const bool use_beta=false>
max@0: class gemm_emul
max@0:   {
max@0:   public:
max@0:   
max@0:   
max@0:   template<typename eT>
max@0:   arma_hot
max@0:   inline
max@0:   static
max@0:   void
max@0:   apply
max@0:     (
max@0:           Mat<eT>& C,
max@0:     const Mat<eT>& A,
max@0:     const Mat<eT>& B,
max@0:     const eT alpha = eT(1),
max@0:     const eT beta  = eT(0),
max@0:     const typename arma_not_cx<eT>::result* junk = 0
max@0:     )
max@0:     {
max@0:     arma_extra_debug_sigprint();
max@0:     arma_ignore(junk);
max@0:     
max@0:     const uword A_n_rows = A.n_rows;
max@0:     const uword A_n_cols = A.n_cols;
max@0:     
max@0:     const uword B_n_rows = B.n_rows;
max@0:     const uword B_n_cols = B.n_cols;
max@0:     
max@0:     if( (A_n_rows <= 4) && (A_n_rows == A_n_cols) && (A_n_rows == B_n_rows) && (B_n_rows == B_n_cols) )
max@0:       {
max@0:       if(do_trans_B == false)
max@0:         {
max@0:         gemm_emul_tinysq<do_trans_A, use_alpha, use_beta>::apply(C, A, B, alpha, beta);
max@0:         }
max@0:       else
max@0:         {
max@0:         Mat<eT> BB(A_n_rows, A_n_rows);
max@0:         op_strans::apply_noalias_tinysq(BB, B);
max@0:         
max@0:         gemm_emul_tinysq<do_trans_A, use_alpha, use_beta>::apply(C, A, BB, alpha, beta);
max@0:         }
max@0:       }
max@0:     else
max@0:       {
max@0:       gemm_emul_large<do_trans_A, do_trans_B, use_alpha, use_beta>::apply(C, A, B, alpha, beta);
max@0:       }
max@0:     }
max@0:   
max@0: 
max@0: 
max@0:   template<typename eT>
max@0:   arma_hot
max@0:   inline
max@0:   static
max@0:   void
max@0:   apply
max@0:     (
max@0:           Mat<eT>& C,
max@0:     const Mat<eT>& A,
max@0:     const Mat<eT>& B,
max@0:     const eT alpha = eT(1),
max@0:     const eT beta  = eT(0),
max@0:     const typename arma_cx_only<eT>::result* junk = 0
max@0:     )
max@0:     {
max@0:     arma_extra_debug_sigprint();
max@0:     arma_ignore(junk);
max@0:     
max@0:     // "better than nothing" handling of hermitian transposes for complex number matrices
max@0:     
max@0:     Mat<eT> tmp_A;
max@0:     Mat<eT> tmp_B;
max@0:     
max@0:     if(do_trans_A)
max@0:       {
max@0:       op_htrans::apply_noalias(tmp_A, A);
max@0:       }
max@0:     
max@0:     if(do_trans_B)
max@0:       {
max@0:       op_htrans::apply_noalias(tmp_B, B);
max@0:       }
max@0:     
max@0:     const Mat<eT>& AA = (do_trans_A == false) ? A : tmp_A;
max@0:     const Mat<eT>& BB = (do_trans_B == false) ? B : tmp_B;
max@0:     
max@0:     const uword A_n_rows = AA.n_rows;
max@0:     const uword A_n_cols = AA.n_cols;
max@0:     
max@0:     const uword B_n_rows = BB.n_rows;
max@0:     const uword B_n_cols = BB.n_cols;
max@0:     
max@0:     if( (A_n_rows <= 4) && (A_n_rows == A_n_cols) && (A_n_rows == B_n_rows) && (B_n_rows == B_n_cols) )
max@0:       {
max@0:       gemm_emul_tinysq<false, use_alpha, use_beta>::apply(C, AA, BB, alpha, beta);
max@0:       }
max@0:     else
max@0:       {
max@0:       gemm_emul_large<false, false, use_alpha, use_beta>::apply(C, AA, BB, alpha, beta);
max@0:       }
max@0:     }
max@0: 
max@0:   };
max@0: 
max@0: 
max@0: 
max@0: //! \brief
max@0: //! Wrapper for ATLAS/BLAS dgemm function, using template arguments to control the arguments passed to dgemm.
max@0: //! Matrix 'C' is assumed to have been set to the correct size (i.e. taking into account transposes)
max@0: 
max@0: template<const bool do_trans_A=false, const bool do_trans_B=false, const bool use_alpha=false, const bool use_beta=false>
max@0: class gemm
max@0:   {
max@0:   public:
max@0:   
max@0:   template<typename eT>
max@0:   inline
max@0:   static
max@0:   void
max@0:   apply_blas_type( Mat<eT>& C, const Mat<eT>& A, const Mat<eT>& B, const eT alpha = eT(1), const eT beta = eT(0) )
max@0:     {
max@0:     arma_extra_debug_sigprint();
max@0:     
max@0:     if( (A.n_elem <= 48u) && (B.n_elem <= 48u) )
max@0:       {
max@0:       gemm_emul<do_trans_A, do_trans_B, use_alpha, use_beta>::apply(C,A,B,alpha,beta);
max@0:       }
max@0:     else
max@0:       {
max@0:       #if defined(ARMA_USE_ATLAS)
max@0:         {
max@0:         arma_extra_debug_print("atlas::cblas_gemm()");
max@0:         
max@0:         atlas::cblas_gemm<eT>
max@0:           (
max@0:           atlas::CblasColMajor,
max@0:           (do_trans_A) ? ( is_complex<eT>::value ? CblasConjTrans : atlas::CblasTrans ) : atlas::CblasNoTrans,
max@0:           (do_trans_B) ? ( is_complex<eT>::value ? CblasConjTrans : atlas::CblasTrans ) : atlas::CblasNoTrans,
max@0:           C.n_rows,
max@0:           C.n_cols,
max@0:           (do_trans_A) ? A.n_rows : A.n_cols,
max@0:           (use_alpha) ? alpha : eT(1),
max@0:           A.mem,
max@0:           (do_trans_A) ? A.n_rows : C.n_rows,
max@0:           B.mem,
max@0:           (do_trans_B) ? C.n_cols : ( (do_trans_A) ? A.n_rows : A.n_cols ),
max@0:           (use_beta) ? beta : eT(0),
max@0:           C.memptr(),
max@0:           C.n_rows
max@0:           );
max@0:         }
max@0:       #elif defined(ARMA_USE_BLAS)
max@0:         {
max@0:         arma_extra_debug_print("blas::gemm()");
max@0:         
max@0:         const char trans_A = (do_trans_A) ? ( is_complex<eT>::value ? 'C' : 'T' ) : 'N';
max@0:         const char trans_B = (do_trans_B) ? ( is_complex<eT>::value ? 'C' : 'T' ) : 'N';
max@0:         
max@0:         const blas_int m   = C.n_rows;
max@0:         const blas_int n   = C.n_cols;
max@0:         const blas_int k   = (do_trans_A) ? A.n_rows : A.n_cols;
max@0:         
max@0:         const eT local_alpha = (use_alpha) ? alpha : eT(1);
max@0:         
max@0:         const blas_int lda = (do_trans_A) ? k : m;
max@0:         const blas_int ldb = (do_trans_B) ? n : k;
max@0:         
max@0:         const eT local_beta  = (use_beta) ? beta : eT(0);
max@0:         
max@0:         arma_extra_debug_print( arma_boost::format("blas::gemm(): trans_A = %c") % trans_A );
max@0:         arma_extra_debug_print( arma_boost::format("blas::gemm(): trans_B = %c") % trans_B );
max@0:         
max@0:         blas::gemm<eT>
max@0:           (
max@0:           &trans_A,
max@0:           &trans_B,
max@0:           &m,
max@0:           &n,
max@0:           &k,
max@0:           &local_alpha,
max@0:           A.mem,
max@0:           &lda,
max@0:           B.mem,
max@0:           &ldb,
max@0:           &local_beta,
max@0:           C.memptr(),
max@0:           &m
max@0:           );
max@0:         }
max@0:       #else
max@0:         {
max@0:         gemm_emul<do_trans_A, do_trans_B, use_alpha, use_beta>::apply(C,A,B,alpha,beta);
max@0:         }
max@0:       #endif
max@0:       }
max@0:     }
max@0:   
max@0:   
max@0:   
max@0:   //! immediate multiplication of matrices A and B, storing the result in C
max@0:   template<typename eT>
max@0:   inline
max@0:   static
max@0:   void
max@0:   apply( Mat<eT>& C, const Mat<eT>& A, const Mat<eT>& B, const eT alpha = eT(1), const eT beta = eT(0) )
max@0:     {
max@0:     gemm_emul<do_trans_A, do_trans_B, use_alpha, use_beta>::apply(C,A,B,alpha,beta);
max@0:     }
max@0:   
max@0:   
max@0:   
max@0:   arma_inline
max@0:   static
max@0:   void
max@0:   apply
max@0:     (
max@0:           Mat<float>& C,
max@0:     const Mat<float>& A,
max@0:     const Mat<float>& B,
max@0:     const float alpha = float(1),
max@0:     const float beta  = float(0)
max@0:     )
max@0:     {
max@0:     gemm<do_trans_A, do_trans_B, use_alpha, use_beta>::apply_blas_type(C,A,B,alpha,beta);
max@0:     }
max@0:   
max@0:   
max@0:   
max@0:   arma_inline
max@0:   static
max@0:   void
max@0:   apply
max@0:     (
max@0:           Mat<double>& C,
max@0:     const Mat<double>& A,
max@0:     const Mat<double>& B,
max@0:     const double alpha = double(1),
max@0:     const double beta  = double(0)
max@0:     )
max@0:     {
max@0:     gemm<do_trans_A, do_trans_B, use_alpha, use_beta>::apply_blas_type(C,A,B,alpha,beta);
max@0:     }
max@0:   
max@0:   
max@0:   
max@0:   arma_inline
max@0:   static
max@0:   void
max@0:   apply
max@0:     (
max@0:           Mat< std::complex<float> >& C,
max@0:     const Mat< std::complex<float> >& A,
max@0:     const Mat< std::complex<float> >& B,
max@0:     const std::complex<float> alpha = std::complex<float>(1),
max@0:     const std::complex<float> beta  = std::complex<float>(0)
max@0:     )
max@0:     {
max@0:     gemm<do_trans_A, do_trans_B, use_alpha, use_beta>::apply_blas_type(C,A,B,alpha,beta);
max@0:     }
max@0:   
max@0:   
max@0:   
max@0:   arma_inline
max@0:   static
max@0:   void
max@0:   apply
max@0:     (
max@0:           Mat< std::complex<double> >& C,
max@0:     const Mat< std::complex<double> >& A,
max@0:     const Mat< std::complex<double> >& B,
max@0:     const std::complex<double> alpha = std::complex<double>(1),
max@0:     const std::complex<double> beta  = std::complex<double>(0)
max@0:     )
max@0:     {
max@0:     gemm<do_trans_A, do_trans_B, use_alpha, use_beta>::apply_blas_type(C,A,B,alpha,beta);
max@0:     }
max@0:   
max@0:   };
max@0: 
max@0: 
max@0: 
max@0: //! @}