503 lines
11 KiB
C++
503 lines
11 KiB
C++
// Copyright 2008-2016 Conrad Sanderson (http://conradsanderson.id.au)
|
|
// Copyright 2008-2016 National ICT Australia (NICTA)
|
|
//
|
|
// Licensed under the Apache License, Version 2.0 (the "License");
|
|
// you may not use this file except in compliance with the License.
|
|
// You may obtain a copy of the License at
|
|
// http://www.apache.org/licenses/LICENSE-2.0
|
|
//
|
|
// Unless required by applicable law or agreed to in writing, software
|
|
// distributed under the License is distributed on an "AS IS" BASIS,
|
|
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
// See the License for the specific language governing permissions and
|
|
// limitations under the License.
|
|
// ------------------------------------------------------------------------
|
|
|
|
|
|
//! \addtogroup op_strans
|
|
//! @{
|
|
|
|
|
|
|
|
//! for tiny square matrices (size <= 4x4)
|
|
template<typename eT, typename TA>
|
|
arma_cold
|
|
inline
|
|
void
|
|
op_strans::apply_mat_noalias_tinysq(Mat<eT>& out, const TA& A)
|
|
{
|
|
const eT* Am = A.memptr();
|
|
eT* outm = out.memptr();
|
|
|
|
switch(A.n_rows)
|
|
{
|
|
case 1:
|
|
{
|
|
outm[0] = Am[0];
|
|
}
|
|
break;
|
|
|
|
case 2:
|
|
{
|
|
outm[pos<false,0,0>::n2] = Am[pos<true,0,0>::n2];
|
|
outm[pos<false,1,0>::n2] = Am[pos<true,1,0>::n2];
|
|
|
|
outm[pos<false,0,1>::n2] = Am[pos<true,0,1>::n2];
|
|
outm[pos<false,1,1>::n2] = Am[pos<true,1,1>::n2];
|
|
}
|
|
break;
|
|
|
|
case 3:
|
|
{
|
|
outm[pos<false,0,0>::n3] = Am[pos<true,0,0>::n3];
|
|
outm[pos<false,1,0>::n3] = Am[pos<true,1,0>::n3];
|
|
outm[pos<false,2,0>::n3] = Am[pos<true,2,0>::n3];
|
|
|
|
outm[pos<false,0,1>::n3] = Am[pos<true,0,1>::n3];
|
|
outm[pos<false,1,1>::n3] = Am[pos<true,1,1>::n3];
|
|
outm[pos<false,2,1>::n3] = Am[pos<true,2,1>::n3];
|
|
|
|
outm[pos<false,0,2>::n3] = Am[pos<true,0,2>::n3];
|
|
outm[pos<false,1,2>::n3] = Am[pos<true,1,2>::n3];
|
|
outm[pos<false,2,2>::n3] = Am[pos<true,2,2>::n3];
|
|
}
|
|
break;
|
|
|
|
case 4:
|
|
{
|
|
outm[pos<false,0,0>::n4] = Am[pos<true,0,0>::n4];
|
|
outm[pos<false,1,0>::n4] = Am[pos<true,1,0>::n4];
|
|
outm[pos<false,2,0>::n4] = Am[pos<true,2,0>::n4];
|
|
outm[pos<false,3,0>::n4] = Am[pos<true,3,0>::n4];
|
|
|
|
outm[pos<false,0,1>::n4] = Am[pos<true,0,1>::n4];
|
|
outm[pos<false,1,1>::n4] = Am[pos<true,1,1>::n4];
|
|
outm[pos<false,2,1>::n4] = Am[pos<true,2,1>::n4];
|
|
outm[pos<false,3,1>::n4] = Am[pos<true,3,1>::n4];
|
|
|
|
outm[pos<false,0,2>::n4] = Am[pos<true,0,2>::n4];
|
|
outm[pos<false,1,2>::n4] = Am[pos<true,1,2>::n4];
|
|
outm[pos<false,2,2>::n4] = Am[pos<true,2,2>::n4];
|
|
outm[pos<false,3,2>::n4] = Am[pos<true,3,2>::n4];
|
|
|
|
outm[pos<false,0,3>::n4] = Am[pos<true,0,3>::n4];
|
|
outm[pos<false,1,3>::n4] = Am[pos<true,1,3>::n4];
|
|
outm[pos<false,2,3>::n4] = Am[pos<true,2,3>::n4];
|
|
outm[pos<false,3,3>::n4] = Am[pos<true,3,3>::n4];
|
|
}
|
|
break;
|
|
|
|
default:
|
|
;
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
template<typename eT>
|
|
arma_hot
|
|
inline
|
|
void
|
|
op_strans::block_worker(eT* Y, const eT* X, const uword X_n_rows, const uword Y_n_rows, const uword n_rows, const uword n_cols)
|
|
{
|
|
for(uword row = 0; row < n_rows; ++row)
|
|
{
|
|
const uword Y_offset = row * Y_n_rows;
|
|
|
|
for(uword col = 0; col < n_cols; ++col)
|
|
{
|
|
const uword X_offset = col * X_n_rows;
|
|
|
|
Y[col + Y_offset] = X[row + X_offset];
|
|
}
|
|
}
|
|
}
|
|
|
|
|
|
|
|
template<typename eT>
|
|
arma_hot
|
|
inline
|
|
void
|
|
op_strans::apply_mat_noalias_large(Mat<eT>& out, const Mat<eT>& A)
|
|
{
|
|
arma_extra_debug_sigprint();
|
|
|
|
const uword n_rows = A.n_rows;
|
|
const uword n_cols = A.n_cols;
|
|
|
|
const uword block_size = 64;
|
|
|
|
const uword n_rows_base = block_size * (n_rows / block_size);
|
|
const uword n_cols_base = block_size * (n_cols / block_size);
|
|
|
|
const uword n_rows_extra = n_rows - n_rows_base;
|
|
const uword n_cols_extra = n_cols - n_cols_base;
|
|
|
|
const eT* X = A.memptr();
|
|
eT* Y = out.memptr();
|
|
|
|
for(uword row = 0; row < n_rows_base; row += block_size)
|
|
{
|
|
const uword Y_offset = row * n_cols;
|
|
|
|
for(uword col = 0; col < n_cols_base; col += block_size)
|
|
{
|
|
const uword X_offset = col * n_rows;
|
|
|
|
op_strans::block_worker(&Y[col + Y_offset], &X[row + X_offset], n_rows, n_cols, block_size, block_size);
|
|
}
|
|
|
|
const uword X_offset = n_cols_base * n_rows;
|
|
|
|
op_strans::block_worker(&Y[n_cols_base + Y_offset], &X[row + X_offset], n_rows, n_cols, block_size, n_cols_extra);
|
|
}
|
|
|
|
if(n_rows_extra == 0) { return; }
|
|
|
|
const uword Y_offset = n_rows_base * n_cols;
|
|
|
|
for(uword col = 0; col < n_cols_base; col += block_size)
|
|
{
|
|
const uword X_offset = col * n_rows;
|
|
|
|
op_strans::block_worker(&Y[col + Y_offset], &X[n_rows_base + X_offset], n_rows, n_cols, n_rows_extra, block_size);
|
|
}
|
|
|
|
const uword X_offset = n_cols_base * n_rows;
|
|
|
|
op_strans::block_worker(&Y[n_cols_base + Y_offset], &X[n_rows_base + X_offset], n_rows, n_cols, n_rows_extra, n_cols_extra);
|
|
}
|
|
|
|
|
|
|
|
//! Immediate transpose of a dense matrix
|
|
template<typename eT, typename TA>
|
|
arma_hot
|
|
inline
|
|
void
|
|
op_strans::apply_mat_noalias(Mat<eT>& out, const TA& A)
|
|
{
|
|
arma_extra_debug_sigprint();
|
|
|
|
const uword A_n_cols = A.n_cols;
|
|
const uword A_n_rows = A.n_rows;
|
|
|
|
out.set_size(A_n_cols, A_n_rows);
|
|
|
|
if( (TA::is_row) || (TA::is_col) || (A_n_cols == 1) || (A_n_rows == 1) )
|
|
{
|
|
arrayops::copy( out.memptr(), A.memptr(), A.n_elem );
|
|
}
|
|
else
|
|
{
|
|
if( (A_n_rows <= 4) && (A_n_rows == A_n_cols) )
|
|
{
|
|
op_strans::apply_mat_noalias_tinysq(out, A);
|
|
}
|
|
else
|
|
if( (A_n_rows >= 512) && (A_n_cols >= 512) )
|
|
{
|
|
op_strans::apply_mat_noalias_large(out, A);
|
|
}
|
|
else
|
|
{
|
|
eT* outptr = out.memptr();
|
|
|
|
for(uword k=0; k < A_n_rows; ++k)
|
|
{
|
|
const eT* Aptr = &(A.at(k,0));
|
|
|
|
uword j;
|
|
for(j=1; j < A_n_cols; j+=2)
|
|
{
|
|
const eT tmp_i = (*Aptr); Aptr += A_n_rows;
|
|
const eT tmp_j = (*Aptr); Aptr += A_n_rows;
|
|
|
|
(*outptr) = tmp_i; outptr++;
|
|
(*outptr) = tmp_j; outptr++;
|
|
}
|
|
|
|
if((j-1) < A_n_cols)
|
|
{
|
|
(*outptr) = (*Aptr); outptr++;;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
|
|
|
|
template<typename eT>
|
|
arma_hot
|
|
inline
|
|
void
|
|
op_strans::apply_mat_inplace(Mat<eT>& out)
|
|
{
|
|
arma_extra_debug_sigprint();
|
|
|
|
const uword n_rows = out.n_rows;
|
|
const uword n_cols = out.n_cols;
|
|
|
|
if(n_rows == n_cols)
|
|
{
|
|
arma_extra_debug_print("op_strans::apply(): doing in-place transpose of a square matrix");
|
|
|
|
const uword N = n_rows;
|
|
|
|
for(uword k=0; k < N; ++k)
|
|
{
|
|
eT* colptr = &(out.at(k,k));
|
|
eT* rowptr = colptr;
|
|
|
|
colptr++;
|
|
rowptr += N;
|
|
|
|
uword j;
|
|
|
|
for(j=(k+2); j < N; j+=2)
|
|
{
|
|
std::swap( (*rowptr), (*colptr) ); rowptr += N; colptr++;
|
|
std::swap( (*rowptr), (*colptr) ); rowptr += N; colptr++;
|
|
}
|
|
|
|
if((j-1) < N)
|
|
{
|
|
std::swap( (*rowptr), (*colptr) );
|
|
}
|
|
}
|
|
}
|
|
else
|
|
{
|
|
Mat<eT> tmp;
|
|
|
|
op_strans::apply_mat_noalias(tmp, out);
|
|
|
|
out.steal_mem(tmp);
|
|
}
|
|
}
|
|
|
|
|
|
|
|
template<typename eT, typename TA>
|
|
arma_hot
|
|
inline
|
|
void
|
|
op_strans::apply_mat(Mat<eT>& out, const TA& A)
|
|
{
|
|
arma_extra_debug_sigprint();
|
|
|
|
if(&out != &A)
|
|
{
|
|
op_strans::apply_mat_noalias(out, A);
|
|
}
|
|
else
|
|
{
|
|
op_strans::apply_mat_inplace(out);
|
|
}
|
|
}
|
|
|
|
|
|
|
|
template<typename T1>
|
|
arma_hot
|
|
inline
|
|
void
|
|
op_strans::apply_proxy(Mat<typename T1::elem_type>& out, const T1& X)
|
|
{
|
|
arma_extra_debug_sigprint();
|
|
|
|
typedef typename T1::elem_type eT;
|
|
|
|
const Proxy<T1> P(X);
|
|
|
|
const uword n_rows = P.get_n_rows();
|
|
const uword n_cols = P.get_n_cols();
|
|
|
|
const bool is_alias = P.is_alias(out);
|
|
|
|
if( (resolves_to_vector<T1>::yes) && (Proxy<T1>::use_at == false) )
|
|
{
|
|
if(is_alias == false)
|
|
{
|
|
out.set_size(n_cols, n_rows);
|
|
|
|
eT* out_mem = out.memptr();
|
|
|
|
const uword n_elem = P.get_n_elem();
|
|
|
|
typename Proxy<T1>::ea_type Pea = P.get_ea();
|
|
|
|
uword i,j;
|
|
for(i=0, j=1; j < n_elem; i+=2, j+=2)
|
|
{
|
|
const eT tmp_i = Pea[i];
|
|
const eT tmp_j = Pea[j];
|
|
|
|
out_mem[i] = tmp_i;
|
|
out_mem[j] = tmp_j;
|
|
}
|
|
|
|
if(i < n_elem)
|
|
{
|
|
out_mem[i] = Pea[i];
|
|
}
|
|
}
|
|
else // aliasing
|
|
{
|
|
Mat<eT> out2(n_cols, n_rows, arma_nozeros_indicator());
|
|
|
|
eT* out_mem = out2.memptr();
|
|
|
|
const uword n_elem = P.get_n_elem();
|
|
|
|
typename Proxy<T1>::ea_type Pea = P.get_ea();
|
|
|
|
uword i,j;
|
|
for(i=0, j=1; j < n_elem; i+=2, j+=2)
|
|
{
|
|
const eT tmp_i = Pea[i];
|
|
const eT tmp_j = Pea[j];
|
|
|
|
out_mem[i] = tmp_i;
|
|
out_mem[j] = tmp_j;
|
|
}
|
|
|
|
if(i < n_elem)
|
|
{
|
|
out_mem[i] = Pea[i];
|
|
}
|
|
|
|
out.steal_mem(out2);
|
|
}
|
|
}
|
|
else // general matrix transpose
|
|
{
|
|
if(is_alias == false)
|
|
{
|
|
out.set_size(n_cols, n_rows);
|
|
|
|
eT* outptr = out.memptr();
|
|
|
|
for(uword k=0; k < n_rows; ++k)
|
|
{
|
|
uword j;
|
|
for(j=1; j < n_cols; j+=2)
|
|
{
|
|
const uword i = j-1;
|
|
|
|
const eT tmp_i = P.at(k,i);
|
|
const eT tmp_j = P.at(k,j);
|
|
|
|
(*outptr) = tmp_i; outptr++;
|
|
(*outptr) = tmp_j; outptr++;
|
|
}
|
|
|
|
const uword i = j-1;
|
|
|
|
if(i < n_cols)
|
|
{
|
|
(*outptr) = P.at(k,i); outptr++;
|
|
}
|
|
}
|
|
}
|
|
else // aliasing
|
|
{
|
|
Mat<eT> out2(n_cols, n_rows, arma_nozeros_indicator());
|
|
|
|
eT* out2ptr = out2.memptr();
|
|
|
|
for(uword k=0; k < n_rows; ++k)
|
|
{
|
|
uword j;
|
|
for(j=1; j < n_cols; j+=2)
|
|
{
|
|
const uword i = j-1;
|
|
|
|
const eT tmp_i = P.at(k,i);
|
|
const eT tmp_j = P.at(k,j);
|
|
|
|
(*out2ptr) = tmp_i; out2ptr++;
|
|
(*out2ptr) = tmp_j; out2ptr++;
|
|
}
|
|
|
|
const uword i = j-1;
|
|
|
|
if(i < n_cols)
|
|
{
|
|
(*out2ptr) = P.at(k,i); out2ptr++;
|
|
}
|
|
}
|
|
|
|
out.steal_mem(out2);
|
|
}
|
|
}
|
|
}
|
|
|
|
|
|
|
|
template<typename T1>
|
|
arma_hot
|
|
inline
|
|
void
|
|
op_strans::apply_direct(Mat<typename T1::elem_type>& out, const T1& X)
|
|
{
|
|
arma_extra_debug_sigprint();
|
|
|
|
// allow detection of in-place transpose
|
|
if(is_Mat<T1>::value || is_Mat<typename Proxy<T1>::stored_type>::value || (arma_config::openmp && Proxy<T1>::use_mp))
|
|
{
|
|
const unwrap<T1> U(X);
|
|
|
|
op_strans::apply_mat(out, U.M);
|
|
}
|
|
else
|
|
{
|
|
op_strans::apply_proxy(out, X);
|
|
}
|
|
}
|
|
|
|
|
|
|
|
template<typename T1>
|
|
arma_hot
|
|
inline
|
|
void
|
|
op_strans::apply(Mat<typename T1::elem_type>& out, const Op<T1,op_strans>& in)
|
|
{
|
|
arma_extra_debug_sigprint();
|
|
|
|
op_strans::apply_direct(out, in.m);
|
|
}
|
|
|
|
|
|
|
|
//
|
|
//
|
|
//
|
|
|
|
|
|
|
|
template<typename eT>
|
|
inline
|
|
void
|
|
op_strans_cube::apply_noalias(Cube<eT>& out, const Cube<eT>& X)
|
|
{
|
|
out.set_size(X.n_cols, X.n_rows, X.n_slices);
|
|
|
|
for(uword s=0; s < X.n_slices; ++s)
|
|
{
|
|
Mat<eT> out_slice( out.slice_memptr(s), X.n_cols, X.n_rows, false, true );
|
|
|
|
const Mat<eT> X_slice( const_cast<eT*>(X.slice_memptr(s)), X.n_rows, X.n_cols, false, true );
|
|
|
|
op_strans::apply_mat_noalias(out_slice, X_slice);
|
|
}
|
|
}
|
|
|
|
|
|
|
|
//! @}
|