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 op_strans
max@0: //! @{
max@0: 
max@0: 
max@0: 
max@0: //! for tiny square matrices (size <= 4x4)
max@0: template<typename eT>
max@0: inline
max@0: void
max@0: op_strans::apply_noalias_tinysq(Mat<eT>& out, const Mat<eT>& A)
max@0:   {
max@0:   const eT* Am   = A.memptr();
max@0:         eT* outm = out.memptr();
max@0:   
max@0:   switch(A.n_rows)
max@0:     {
max@0:     case 1:
max@0:       {
max@0:       outm[0] = Am[0];
max@0:       }
max@0:       break;
max@0:       
max@0:     case 2:
max@0:       {
max@0:       outm[pos<false,0,0>::n2] = Am[pos<true,0,0>::n2];
max@0:       outm[pos<false,1,0>::n2] = Am[pos<true,1,0>::n2];
max@0:       
max@0:       outm[pos<false,0,1>::n2] = Am[pos<true,0,1>::n2];
max@0:       outm[pos<false,1,1>::n2] = Am[pos<true,1,1>::n2];
max@0:       }
max@0:       break;
max@0:     
max@0:     case 3:
max@0:       {
max@0:       outm[pos<false,0,0>::n3] = Am[pos<true,0,0>::n3];
max@0:       outm[pos<false,1,0>::n3] = Am[pos<true,1,0>::n3];
max@0:       outm[pos<false,2,0>::n3] = Am[pos<true,2,0>::n3];
max@0:       
max@0:       outm[pos<false,0,1>::n3] = Am[pos<true,0,1>::n3];
max@0:       outm[pos<false,1,1>::n3] = Am[pos<true,1,1>::n3];
max@0:       outm[pos<false,2,1>::n3] = Am[pos<true,2,1>::n3];
max@0:       
max@0:       outm[pos<false,0,2>::n3] = Am[pos<true,0,2>::n3];
max@0:       outm[pos<false,1,2>::n3] = Am[pos<true,1,2>::n3];
max@0:       outm[pos<false,2,2>::n3] = Am[pos<true,2,2>::n3];
max@0:       }
max@0:       break;
max@0:     
max@0:     case 4:
max@0:       {
max@0:       outm[pos<false,0,0>::n4] = Am[pos<true,0,0>::n4];
max@0:       outm[pos<false,1,0>::n4] = Am[pos<true,1,0>::n4];
max@0:       outm[pos<false,2,0>::n4] = Am[pos<true,2,0>::n4];
max@0:       outm[pos<false,3,0>::n4] = Am[pos<true,3,0>::n4];
max@0:       
max@0:       outm[pos<false,0,1>::n4] = Am[pos<true,0,1>::n4];
max@0:       outm[pos<false,1,1>::n4] = Am[pos<true,1,1>::n4];
max@0:       outm[pos<false,2,1>::n4] = Am[pos<true,2,1>::n4];
max@0:       outm[pos<false,3,1>::n4] = Am[pos<true,3,1>::n4];
max@0:       
max@0:       outm[pos<false,0,2>::n4] = Am[pos<true,0,2>::n4];
max@0:       outm[pos<false,1,2>::n4] = Am[pos<true,1,2>::n4];
max@0:       outm[pos<false,2,2>::n4] = Am[pos<true,2,2>::n4];
max@0:       outm[pos<false,3,2>::n4] = Am[pos<true,3,2>::n4];
max@0:       
max@0:       outm[pos<false,0,3>::n4] = Am[pos<true,0,3>::n4];
max@0:       outm[pos<false,1,3>::n4] = Am[pos<true,1,3>::n4];
max@0:       outm[pos<false,2,3>::n4] = Am[pos<true,2,3>::n4];
max@0:       outm[pos<false,3,3>::n4] = Am[pos<true,3,3>::n4];
max@0:       }
max@0:       break;
max@0:     
max@0:     default:
max@0:       ;
max@0:     }
max@0:   
max@0:   }
max@0: 
max@0: 
max@0: 
max@0: //! Immediate transpose of a dense matrix
max@0: template<typename eT>
max@0: inline
max@0: void
max@0: op_strans::apply_noalias(Mat<eT>& out, const Mat<eT>& A)
max@0:   {
max@0:   arma_extra_debug_sigprint();
max@0:   
max@0:   const uword A_n_cols = A.n_cols;
max@0:   const uword A_n_rows = A.n_rows;
max@0:   
max@0:   out.set_size(A_n_cols, A_n_rows);
max@0:   
max@0:   if( (A_n_cols == 1) || (A_n_rows == 1) )
max@0:     {
max@0:     arrayops::copy( out.memptr(), A.mem, A.n_elem );
max@0:     }
max@0:   else
max@0:     {
max@0:     if( (A_n_rows <= 4) && (A_n_rows == A_n_cols) )
max@0:       {
max@0:       op_strans::apply_noalias_tinysq(out, A);
max@0:       }
max@0:     else
max@0:       {
max@0:       for(uword k=0; k < A_n_cols; ++k)
max@0:         {
max@0:         uword i, j;
max@0:         
max@0:         const eT* colptr = A.colptr(k);
max@0:         
max@0:         for(i=0, j=1; j < A_n_rows; i+=2, j+=2)
max@0:           {
max@0:           const eT tmp_i = colptr[i];
max@0:           const eT tmp_j = colptr[j];
max@0:           
max@0:           out.at(k, i) = tmp_i;
max@0:           out.at(k, j) = tmp_j;
max@0:           }
max@0:         
max@0:         if(i < A_n_rows)
max@0:           {
max@0:           out.at(k, i) = colptr[i];
max@0:           }
max@0:         }
max@0:       }
max@0:     }
max@0:   }
max@0: 
max@0: 
max@0: 
max@0: template<typename eT>
max@0: inline
max@0: void
max@0: op_strans::apply(Mat<eT>& out, const Mat<eT>& A)
max@0:   {
max@0:   arma_extra_debug_sigprint();
max@0:   
max@0:   if(&out != &A)
max@0:     {
max@0:     op_strans::apply_noalias(out, A);
max@0:     }
max@0:   else
max@0:     {
max@0:     const uword n_rows = out.n_rows;
max@0:     const uword n_cols = out.n_cols;
max@0:       
max@0:     if(n_rows == n_cols)
max@0:       {
max@0:       arma_extra_debug_print("op_strans::apply(): doing in-place transpose of a square matrix");
max@0:       
max@0:       const uword N = n_rows;
max@0:       
max@0:       for(uword k=0; k < N; ++k)
max@0:         {
max@0:         eT* colptr = out.colptr(k);
max@0:         
max@0:         uword i,j;
max@0:         
max@0:         for(i=(k+1), j=(k+2); j < N; i+=2, j+=2)
max@0:           {
max@0:           std::swap(out.at(k,i), colptr[i]);
max@0:           std::swap(out.at(k,j), colptr[j]);
max@0:           }
max@0:         
max@0:         if(i < N)
max@0:           {
max@0:           std::swap(out.at(k,i), colptr[i]);
max@0:           }
max@0:         }
max@0:       }
max@0:     else
max@0:       {
max@0:       Mat<eT> tmp;
max@0:       op_strans::apply_noalias(tmp, A);
max@0:       
max@0:       out.steal_mem(tmp);
max@0:       }
max@0:     }
max@0:   }
max@0: 
max@0: 
max@0: 
max@0: template<typename T1>
max@0: inline
max@0: void
max@0: op_strans::apply(Mat<typename T1::elem_type>& out, const Op<T1,op_strans>& in)
max@0:   {
max@0:   arma_extra_debug_sigprint();
max@0:   
max@0:   typedef typename T1::elem_type eT;
max@0:   
max@0:   const unwrap<T1>   tmp(in.m);
max@0:   const Mat<eT>& A = tmp.M;
max@0:   
max@0:   op_strans::apply(out, A);
max@0:   }
max@0: 
max@0: 
max@0: 
max@0: // inline void op_strans::apply_inplace(mat &X)
max@0: //   {
max@0: //   arma_extra_debug_sigprint();
max@0: //   
max@0: //   if((X.n_rows == 1) || (X.n_cols == 1))
max@0: //     {
max@0: //     const uword old_n_rows = X.n_rows;
max@0: //     access::rw(X.n_rows) = X.n_cols;
max@0: //     access::rw(X.n_cols) = old_n_rows;
max@0: //     }
max@0: //   else
max@0: //   if(X.n_rows == X.n_cols)
max@0: //     {
max@0: //     for(uword col=0; col < X.n_cols; ++col)
max@0: //       {
max@0: //       double* X_coldata = X.colptr(col);
max@0: //       
max@0: //       for(uword row=(col+1); row < X.n_rows; ++row)
max@0: //         {
max@0: //         std::swap( A.at(col,row), A_coldata[row] );
max@0: //         }
max@0: //       }
max@0: //     }
max@0: //   else
max@0: //     {
max@0: //     mat tmp = trans(X);
max@0: //     
max@0: //     if(X.mem != X.mem_local)
max@0: //       {
max@0: //       double* old_mem = X.memptr();
max@0: //       access::rw(X.mem) = tmp.memptr();
max@0: //       access::rw(tmp.mem) = old_mem;
max@0: //       }
max@0: //     else
max@0: //       {
max@0: //       X = tmp;
max@0: //       }
max@0: //     }
max@0: //   
max@0: //   }
max@0: 
max@0: 
max@0: 
max@0: 
max@0: //! @}