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 gemv
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 gemv_emul_tinysq
max@0:   {
max@0:   public:
max@0:   
max@0:   
max@0:   template<const uword row, const uword col>
max@0:   struct pos
max@0:     {
max@0:     static const uword n2 = (do_trans_A == false) ? (row + col*2) : (col + row*2);
max@0:     static const uword n3 = (do_trans_A == false) ? (row + col*3) : (col + row*3);
max@0:     static const uword n4 = (do_trans_A == false) ? (row + col*4) : (col + row*4);
max@0:     };
max@0:   
max@0:   
max@0:   
max@0:   template<typename eT, const uword i>
max@0:   arma_hot
max@0:   arma_inline
max@0:   static
max@0:   void
max@0:   assign(eT* y, const eT acc, const eT alpha, const eT beta)
max@0:     {
max@0:     if(use_beta == false)
max@0:       {
max@0:       y[i] = (use_alpha == false) ? acc : alpha*acc;
max@0:       }
max@0:     else
max@0:       {
max@0:       const eT tmp = y[i];
max@0:       
max@0:       y[i] = beta*tmp + ( (use_alpha == false) ? acc : alpha*acc );
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( eT* y, const Mat<eT>& A, const eT* x, const eT alpha = eT(1), const eT beta = eT(0) )
max@0:     {
max@0:     arma_extra_debug_sigprint();
max@0:     
max@0:     const eT*  Am = A.memptr();
max@0:     
max@0:     switch(A.n_rows)
max@0:       {
max@0:       case 1:
max@0:         {
max@0:         const eT acc = Am[0] * x[0];
max@0:         
max@0:         assign<eT, 0>(y, acc, alpha, beta);
max@0:         }
max@0:         break;
max@0:       
max@0:       
max@0:       case 2:
max@0:         {
max@0:         const eT x0 = x[0];
max@0:         const eT x1 = x[1];
max@0:         
max@0:         const eT acc0 = Am[pos<0,0>::n2]*x0 + Am[pos<0,1>::n2]*x1;
max@0:         const eT acc1 = Am[pos<1,0>::n2]*x0 + Am[pos<1,1>::n2]*x1;
max@0:         
max@0:         assign<eT, 0>(y, acc0, alpha, beta);
max@0:         assign<eT, 1>(y, acc1, alpha, beta);
max@0:         }
max@0:         break;
max@0:       
max@0:         
max@0:       case 3:
max@0:         {
max@0:         const eT x0 = x[0];
max@0:         const eT x1 = x[1];
max@0:         const eT x2 = x[2];
max@0:         
max@0:         const eT acc0 = Am[pos<0,0>::n3]*x0 + Am[pos<0,1>::n3]*x1 + Am[pos<0,2>::n3]*x2;
max@0:         const eT acc1 = Am[pos<1,0>::n3]*x0 + Am[pos<1,1>::n3]*x1 + Am[pos<1,2>::n3]*x2;
max@0:         const eT acc2 = Am[pos<2,0>::n3]*x0 + Am[pos<2,1>::n3]*x1 + Am[pos<2,2>::n3]*x2;
max@0:         
max@0:         assign<eT, 0>(y, acc0, alpha, beta);
max@0:         assign<eT, 1>(y, acc1, alpha, beta);
max@0:         assign<eT, 2>(y, acc2, alpha, beta);
max@0:         }
max@0:         break;
max@0:       
max@0:       
max@0:       case 4:
max@0:         {
max@0:         const eT x0 = x[0];
max@0:         const eT x1 = x[1];
max@0:         const eT x2 = x[2];
max@0:         const eT x3 = x[3];
max@0:         
max@0:         const eT acc0 = Am[pos<0,0>::n4]*x0 + Am[pos<0,1>::n4]*x1 + Am[pos<0,2>::n4]*x2 + Am[pos<0,3>::n4]*x3;
max@0:         const eT acc1 = Am[pos<1,0>::n4]*x0 + Am[pos<1,1>::n4]*x1 + Am[pos<1,2>::n4]*x2 + Am[pos<1,3>::n4]*x3;
max@0:         const eT acc2 = Am[pos<2,0>::n4]*x0 + Am[pos<2,1>::n4]*x1 + Am[pos<2,2>::n4]*x2 + Am[pos<2,3>::n4]*x3;
max@0:         const eT acc3 = Am[pos<3,0>::n4]*x0 + Am[pos<3,1>::n4]*x1 + Am[pos<3,2>::n4]*x2 + Am[pos<3,3>::n4]*x3;
max@0:         
max@0:         assign<eT, 0>(y, acc0, alpha, beta);
max@0:         assign<eT, 1>(y, acc1, alpha, beta);
max@0:         assign<eT, 2>(y, acc2, alpha, beta);
max@0:         assign<eT, 3>(y, acc3, alpha, beta);
max@0:         }
max@0:         break;
max@0:       
max@0:       
max@0:       default:
max@0:         ;
max@0:       }
max@0:     }
max@0:     
max@0:   };
max@0: 
max@0: 
max@0: 
max@0: //! \brief
max@0: //! Partial emulation of ATLAS/BLAS gemv().
max@0: //! 'y' is assumed to have been set to the correct size (i.e. taking into account the transpose)
max@0: 
max@0: template<const bool do_trans_A=false, const bool use_alpha=false, const bool use_beta=false>
max@0: class gemv_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( eT* y, const Mat<eT>& A, const eT* x, const eT alpha = eT(1), const eT beta = eT(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:     if(do_trans_A == false)
max@0:       {
max@0:       if(A_n_rows == 1)
max@0:         {
max@0:         const eT acc = op_dot::direct_dot_arma(A_n_cols, A.mem, x);
max@0:         
max@0:         if( (use_alpha == false) && (use_beta == false) )
max@0:           {
max@0:           y[0] = acc;
max@0:           }
max@0:         else
max@0:         if( (use_alpha == true) && (use_beta == false) )
max@0:           {
max@0:           y[0] = alpha * acc;
max@0:           }
max@0:         else
max@0:         if( (use_alpha == false) && (use_beta == true) )
max@0:           {
max@0:           y[0] = acc + beta*y[0];
max@0:           }
max@0:         else
max@0:         if( (use_alpha == true) && (use_beta == true) )
max@0:           {
max@0:           y[0] = alpha*acc + beta*y[0];
max@0:           }
max@0:         }
max@0:       else
max@0:       for(uword row=0; row < A_n_rows; ++row)
max@0:         {
max@0:         eT acc = eT(0);
max@0:         
max@0:         for(uword i=0; i < A_n_cols; ++i)
max@0:           {
max@0:           acc += A.at(row,i) * x[i];
max@0:           }
max@0:         
max@0:         if( (use_alpha == false) && (use_beta == false) )
max@0:           {
max@0:           y[row] = acc;
max@0:           }
max@0:         else
max@0:         if( (use_alpha == true) && (use_beta == false) )
max@0:           {
max@0:           y[row] = alpha * acc;
max@0:           }
max@0:         else
max@0:         if( (use_alpha == false) && (use_beta == true) )
max@0:           {
max@0:           y[row] = acc + beta*y[row];
max@0:           }
max@0:         else
max@0:         if( (use_alpha == true) && (use_beta == true) )
max@0:           {
max@0:           y[row] = alpha*acc + beta*y[row];
max@0:           }
max@0:         }
max@0:       }
max@0:     else
max@0:     if(do_trans_A == true)
max@0:       {
max@0:       for(uword col=0; col < A_n_cols; ++col)
max@0:         {
max@0:         // col is interpreted as row when storing the results in 'y'
max@0:         
max@0:         
max@0:         // const eT* A_coldata = A.colptr(col);
max@0:         // 
max@0:         // eT acc = eT(0);
max@0:         // for(uword row=0; row < A_n_rows; ++row)
max@0:         //   {
max@0:         //   acc += A_coldata[row] * x[row];
max@0:         //   }
max@0:         
max@0:         const eT acc = op_dot::direct_dot_arma(A_n_rows, A.colptr(col), x);
max@0:         
max@0:         if( (use_alpha == false) && (use_beta == false) )
max@0:           {
max@0:           y[col] = acc;
max@0:           }
max@0:         else
max@0:         if( (use_alpha == true) && (use_beta == false) )
max@0:           {
max@0:           y[col] = alpha * acc;
max@0:           }
max@0:         else
max@0:         if( (use_alpha == false) && (use_beta == true) )
max@0:           {
max@0:           y[col] = acc + beta*y[col];
max@0:           }
max@0:         else
max@0:         if( (use_alpha == true) && (use_beta == true) )
max@0:           {
max@0:           y[col] = alpha*acc + beta*y[col];
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 use_alpha=false, const bool use_beta=false>
max@0: class gemv_emul
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( eT* y, const Mat<eT>& A, const eT* x, const eT alpha = eT(1), const eT beta = eT(0), const typename arma_not_cx<eT>::result* junk = 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:     if( (A_n_rows <= 4) && (A_n_rows == A_n_cols) )
max@0:       {
max@0:       gemv_emul_tinysq<do_trans_A, use_alpha, use_beta>::apply(y, A, x, alpha, beta);
max@0:       }
max@0:     else
max@0:       {
max@0:       gemv_emul_large<do_trans_A, use_alpha, use_beta>::apply(y, A, x, 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( eT* y, const Mat<eT>& A, const eT* x, const eT alpha = eT(1), const eT beta = eT(0), const typename arma_cx_only<eT>::result* junk = 0 )
max@0:     {
max@0:     arma_extra_debug_sigprint();
max@0:     
max@0:     Mat<eT> tmp_A;
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:     const Mat<eT>& AA = (do_trans_A == false) ? A : tmp_A;
max@0:     
max@0:     const uword AA_n_rows = AA.n_rows;
max@0:     const uword AA_n_cols = AA.n_cols;
max@0:     
max@0:     if( (AA_n_rows <= 4) && (AA_n_rows == AA_n_cols) )
max@0:       {
max@0:       gemv_emul_tinysq<false, use_alpha, use_beta>::apply(y, AA, x, alpha, beta);
max@0:       }
max@0:     else
max@0:       {
max@0:       gemv_emul_large<false, use_alpha, use_beta>::apply(y, AA, x, alpha, beta);
max@0:       }
max@0:     }
max@0:   };
max@0: 
max@0: 
max@0: 
max@0: //! \brief
max@0: //! Wrapper for ATLAS/BLAS gemv function, using template arguments to control the arguments passed to gemv.
max@0: //! 'y' is assumed to have been set to the correct size (i.e. taking into account the transpose)
max@0: 
max@0: template<const bool do_trans_A=false, const bool use_alpha=false, const bool use_beta=false>
max@0: class gemv
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( eT* y, const Mat<eT>& A, const eT* x, 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 <= 64u)
max@0:       {
max@0:       gemv_emul<do_trans_A, use_alpha, use_beta>::apply(y,A,x,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_gemv()");
max@0:         
max@0:         atlas::cblas_gemv<eT>
max@0:           (
max@0:           atlas::CblasColMajor,
max@0:           (do_trans_A) ? ( is_complex<eT>::value ? CblasConjTrans : atlas::CblasTrans ) : atlas::CblasNoTrans,
max@0:           A.n_rows,
max@0:           A.n_cols,
max@0:           (use_alpha) ? alpha : eT(1),
max@0:           A.mem,
max@0:           A.n_rows,
max@0:           x,
max@0:           1,
max@0:           (use_beta) ? beta : eT(0),
max@0:           y,
max@0:           1
max@0:           );
max@0:         }
max@0:       #elif defined(ARMA_USE_BLAS)
max@0:         {
max@0:         arma_extra_debug_print("blas::gemv()");
max@0:         
max@0:         const char      trans_A     = (do_trans_A) ? ( is_complex<eT>::value ? 'C' : 'T' ) : 'N';
max@0:         const blas_int  m           = A.n_rows;
max@0:         const blas_int  n           = A.n_cols;
max@0:         const eT        local_alpha = (use_alpha) ? alpha : eT(1);
max@0:         //const blas_int  lda         = A.n_rows;
max@0:         const blas_int  inc         = 1;
max@0:         const eT        local_beta  = (use_beta) ? beta : eT(0);
max@0:         
max@0:         arma_extra_debug_print( arma_boost::format("blas::gemv(): trans_A = %c") % trans_A );
max@0:         
max@0:         blas::gemv<eT>
max@0:           (
max@0:           &trans_A,
max@0:           &m,
max@0:           &n,
max@0:           &local_alpha,
max@0:           A.mem,
max@0:           &m,  // lda
max@0:           x,
max@0:           &inc,
max@0:           &local_beta,
max@0:           y,
max@0:           &inc
max@0:           );
max@0:         }
max@0:       #else
max@0:         {
max@0:         gemv_emul<do_trans_A, use_alpha, use_beta>::apply(y,A,x,alpha,beta);
max@0:         }
max@0:       #endif
max@0:       }
max@0:     
max@0:     }
max@0:   
max@0:   
max@0:   
max@0:   template<typename eT>
max@0:   arma_inline
max@0:   static
max@0:   void
max@0:   apply( eT* y, const Mat<eT>& A, const eT* x, const eT alpha = eT(1), const eT beta = eT(0) )
max@0:     {
max@0:     gemv_emul<do_trans_A, use_alpha, use_beta>::apply(y,A,x,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:           float*      y,
max@0:     const Mat<float>& A,
max@0:     const float*      x,
max@0:     const float       alpha = float(1),
max@0:     const float       beta  = float(0)
max@0:     )
max@0:     {
max@0:     gemv<do_trans_A, use_alpha, use_beta>::apply_blas_type(y,A,x,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:           double*      y,
max@0:     const Mat<double>& A,
max@0:     const double*      x,
max@0:     const double       alpha = double(1),
max@0:     const double       beta  = double(0)
max@0:     )
max@0:     {
max@0:     gemv<do_trans_A, use_alpha, use_beta>::apply_blas_type(y,A,x,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:           std::complex<float>*         y,
max@0:     const Mat< std::complex<float > >& A,
max@0:     const std::complex<float>*         x,
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:     gemv<do_trans_A, use_alpha, use_beta>::apply_blas_type(y,A,x,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:           std::complex<double>*        y,
max@0:     const Mat< std::complex<double> >& A,
max@0:     const std::complex<double>*        x,
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:     gemv<do_trans_A, use_alpha, use_beta>::apply_blas_type(y,A,x,alpha,beta);
max@0:     }
max@0: 
max@0: 
max@0:   
max@0:   };
max@0: 
max@0: 
max@0: //! @}