494 lines
12 KiB
C++
494 lines
12 KiB
C++
// Copyright (C) 2013 National ICT Australia (NICTA)
|
|
//
|
|
// This Source Code Form is subject to the terms of the Mozilla Public
|
|
// License, v. 2.0. If a copy of the MPL was not distributed with this
|
|
// file, You can obtain one at http://mozilla.org/MPL/2.0/.
|
|
// -------------------------------------------------------------------
|
|
//
|
|
// Written by Conrad Sanderson - http://conradsanderson.id.au
|
|
|
|
|
|
//! \addtogroup syrk
|
|
//! @{
|
|
|
|
|
|
|
|
class syrk_helper
|
|
{
|
|
public:
|
|
|
|
template<typename eT>
|
|
inline
|
|
static
|
|
void
|
|
inplace_copy_upper_tri_to_lower_tri(Mat<eT>& C)
|
|
{
|
|
// under the assumption that C is a square matrix
|
|
|
|
const uword N = C.n_rows;
|
|
|
|
for(uword k=0; k < N; ++k)
|
|
{
|
|
eT* colmem = C.colptr(k);
|
|
|
|
uword i, j;
|
|
for(i=(k+1), j=(k+2); j < N; i+=2, j+=2)
|
|
{
|
|
const eT tmp_i = C.at(k,i);
|
|
const eT tmp_j = C.at(k,j);
|
|
|
|
colmem[i] = tmp_i;
|
|
colmem[j] = tmp_j;
|
|
}
|
|
|
|
if(i < N)
|
|
{
|
|
colmem[i] = C.at(k,i);
|
|
}
|
|
}
|
|
}
|
|
};
|
|
|
|
|
|
|
|
//! partial emulation of BLAS function syrk(), specialised for A being a vector
|
|
template<const bool do_trans_A=false, const bool use_alpha=false, const bool use_beta=false>
|
|
class syrk_vec
|
|
{
|
|
public:
|
|
|
|
template<typename eT, typename TA>
|
|
arma_hot
|
|
inline
|
|
static
|
|
void
|
|
apply
|
|
(
|
|
Mat<eT>& C,
|
|
const TA& A,
|
|
const eT alpha = eT(1),
|
|
const eT beta = eT(0)
|
|
)
|
|
{
|
|
arma_extra_debug_sigprint();
|
|
|
|
const uword A_n1 = (do_trans_A == false) ? A.n_rows : A.n_cols;
|
|
const uword A_n2 = (do_trans_A == false) ? A.n_cols : A.n_rows;
|
|
|
|
const eT* A_mem = A.memptr();
|
|
|
|
if(A_n1 == 1)
|
|
{
|
|
const eT acc1 = op_dot::direct_dot(A_n2, A_mem, A_mem);
|
|
|
|
if( (use_alpha == false) && (use_beta == false) ) { C[0] = acc1; }
|
|
else if( (use_alpha == true ) && (use_beta == false) ) { C[0] = alpha*acc1; }
|
|
else if( (use_alpha == false) && (use_beta == true ) ) { C[0] = acc1 + beta*C[0]; }
|
|
else if( (use_alpha == true ) && (use_beta == true ) ) { C[0] = alpha*acc1 + beta*C[0]; }
|
|
}
|
|
else
|
|
for(uword k=0; k < A_n1; ++k)
|
|
{
|
|
const eT A_k = A_mem[k];
|
|
|
|
uword i,j;
|
|
for(i=(k), j=(k+1); j < A_n1; i+=2, j+=2)
|
|
{
|
|
const eT acc1 = A_k * A_mem[i];
|
|
const eT acc2 = A_k * A_mem[j];
|
|
|
|
if( (use_alpha == false) && (use_beta == false) )
|
|
{
|
|
C.at(k, i) = acc1;
|
|
C.at(k, j) = acc2;
|
|
|
|
C.at(i, k) = acc1;
|
|
C.at(j, k) = acc2;
|
|
}
|
|
else
|
|
if( (use_alpha == true ) && (use_beta == false) )
|
|
{
|
|
const eT val1 = alpha*acc1;
|
|
const eT val2 = alpha*acc2;
|
|
|
|
C.at(k, i) = val1;
|
|
C.at(k, j) = val2;
|
|
|
|
C.at(i, k) = val1;
|
|
C.at(j, k) = val2;
|
|
}
|
|
else
|
|
if( (use_alpha == false) && (use_beta == true) )
|
|
{
|
|
C.at(k, i) = acc1 + beta*C.at(k, i);
|
|
C.at(k, j) = acc2 + beta*C.at(k, j);
|
|
|
|
if(i != k) { C.at(i, k) = acc1 + beta*C.at(i, k); }
|
|
C.at(j, k) = acc2 + beta*C.at(j, k);
|
|
}
|
|
else
|
|
if( (use_alpha == true ) && (use_beta == true) )
|
|
{
|
|
const eT val1 = alpha*acc1;
|
|
const eT val2 = alpha*acc2;
|
|
|
|
C.at(k, i) = val1 + beta*C.at(k, i);
|
|
C.at(k, j) = val2 + beta*C.at(k, j);
|
|
|
|
if(i != k) { C.at(i, k) = val1 + beta*C.at(i, k); }
|
|
C.at(j, k) = val2 + beta*C.at(j, k);
|
|
}
|
|
}
|
|
|
|
if(i < A_n1)
|
|
{
|
|
const eT acc1 = A_k * A_mem[i];
|
|
|
|
if( (use_alpha == false) && (use_beta == false) )
|
|
{
|
|
C.at(k, i) = acc1;
|
|
C.at(i, k) = acc1;
|
|
}
|
|
else
|
|
if( (use_alpha == true) && (use_beta == false) )
|
|
{
|
|
const eT val1 = alpha*acc1;
|
|
|
|
C.at(k, i) = val1;
|
|
C.at(i, k) = val1;
|
|
}
|
|
else
|
|
if( (use_alpha == false) && (use_beta == true) )
|
|
{
|
|
C.at(k, i) = acc1 + beta*C.at(k, i);
|
|
if(i != k) { C.at(i, k) = acc1 + beta*C.at(i, k); }
|
|
}
|
|
else
|
|
if( (use_alpha == true) && (use_beta == true) )
|
|
{
|
|
const eT val1 = alpha*acc1;
|
|
|
|
C.at(k, i) = val1 + beta*C.at(k, i);
|
|
if(i != k) { C.at(i, k) = val1 + beta*C.at(i, k); }
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
};
|
|
|
|
|
|
|
|
//! partial emulation of BLAS function syrk()
|
|
template<const bool do_trans_A=false, const bool use_alpha=false, const bool use_beta=false>
|
|
class syrk_emul
|
|
{
|
|
public:
|
|
|
|
template<typename eT, typename TA>
|
|
arma_hot
|
|
inline
|
|
static
|
|
void
|
|
apply
|
|
(
|
|
Mat<eT>& C,
|
|
const TA& A,
|
|
const eT alpha = eT(1),
|
|
const eT beta = eT(0)
|
|
)
|
|
{
|
|
arma_extra_debug_sigprint();
|
|
|
|
// do_trans_A == false -> C = alpha * A * A^T + beta*C
|
|
// do_trans_A == true -> C = alpha * A^T * A + beta*C
|
|
|
|
if(do_trans_A == false)
|
|
{
|
|
Mat<eT> AA;
|
|
|
|
op_strans::apply_mat_noalias(AA, A);
|
|
|
|
syrk_emul<true, use_alpha, use_beta>::apply(C, AA, alpha, beta);
|
|
}
|
|
else
|
|
if(do_trans_A == true)
|
|
{
|
|
const uword A_n_rows = A.n_rows;
|
|
const uword A_n_cols = A.n_cols;
|
|
|
|
for(uword col_A=0; col_A < A_n_cols; ++col_A)
|
|
{
|
|
// col_A is interpreted as row_A when storing the results in matrix C
|
|
|
|
const eT* A_coldata = A.colptr(col_A);
|
|
|
|
for(uword k=col_A; k < A_n_cols; ++k)
|
|
{
|
|
const eT acc = op_dot::direct_dot_arma(A_n_rows, A_coldata, A.colptr(k));
|
|
|
|
if( (use_alpha == false) && (use_beta == false) )
|
|
{
|
|
C.at(col_A, k) = acc;
|
|
C.at(k, col_A) = acc;
|
|
}
|
|
else
|
|
if( (use_alpha == true ) && (use_beta == false) )
|
|
{
|
|
const eT val = alpha*acc;
|
|
|
|
C.at(col_A, k) = val;
|
|
C.at(k, col_A) = val;
|
|
}
|
|
else
|
|
if( (use_alpha == false) && (use_beta == true ) )
|
|
{
|
|
C.at(col_A, k) = acc + beta*C.at(col_A, k);
|
|
if(col_A != k) { C.at(k, col_A) = acc + beta*C.at(k, col_A); }
|
|
}
|
|
else
|
|
if( (use_alpha == true ) && (use_beta == true ) )
|
|
{
|
|
const eT val = alpha*acc;
|
|
|
|
C.at(col_A, k) = val + beta*C.at(col_A, k);
|
|
if(col_A != k) { C.at(k, col_A) = val + beta*C.at(k, col_A); }
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
};
|
|
|
|
|
|
|
|
template<const bool do_trans_A=false, const bool use_alpha=false, const bool use_beta=false>
|
|
class syrk
|
|
{
|
|
public:
|
|
|
|
template<typename eT, typename TA>
|
|
inline
|
|
static
|
|
void
|
|
apply_blas_type( Mat<eT>& C, const TA& A, const eT alpha = eT(1), const eT beta = eT(0) )
|
|
{
|
|
arma_extra_debug_sigprint();
|
|
|
|
if(A.is_vec())
|
|
{
|
|
// work around poor handling of vectors by syrk() in ATLAS 3.8.4 and standard BLAS
|
|
|
|
syrk_vec<do_trans_A, use_alpha, use_beta>::apply(C,A,alpha,beta);
|
|
|
|
return;
|
|
}
|
|
|
|
const uword threshold = (is_cx<eT>::yes ? 16u : 48u);
|
|
|
|
if( A.n_elem <= threshold )
|
|
{
|
|
syrk_emul<do_trans_A, use_alpha, use_beta>::apply(C,A,alpha,beta);
|
|
}
|
|
else
|
|
{
|
|
#if defined(ARMA_USE_ATLAS)
|
|
{
|
|
if(use_beta == true)
|
|
{
|
|
// use a temporary matrix, as we can't assume that matrix C is already symmetric
|
|
Mat<eT> D(C.n_rows, C.n_cols);
|
|
|
|
syrk<do_trans_A, use_alpha, false>::apply_blas_type(D,A,alpha);
|
|
|
|
// NOTE: assuming beta=1; this is okay for now, as currently glue_times only uses beta=1
|
|
arrayops::inplace_plus(C.memptr(), D.memptr(), C.n_elem);
|
|
|
|
return;
|
|
}
|
|
|
|
atlas::cblas_syrk<eT>
|
|
(
|
|
atlas::CblasColMajor,
|
|
atlas::CblasUpper,
|
|
(do_trans_A) ? atlas::CblasTrans : atlas::CblasNoTrans,
|
|
C.n_cols,
|
|
(do_trans_A) ? A.n_rows : A.n_cols,
|
|
(use_alpha) ? alpha : eT(1),
|
|
A.mem,
|
|
(do_trans_A) ? A.n_rows : C.n_cols,
|
|
(use_beta) ? beta : eT(0),
|
|
C.memptr(),
|
|
C.n_cols
|
|
);
|
|
|
|
syrk_helper::inplace_copy_upper_tri_to_lower_tri(C);
|
|
}
|
|
#elif defined(ARMA_USE_BLAS)
|
|
{
|
|
if(use_beta == true)
|
|
{
|
|
// use a temporary matrix, as we can't assume that matrix C is already symmetric
|
|
Mat<eT> D(C.n_rows, C.n_cols);
|
|
|
|
syrk<do_trans_A, use_alpha, false>::apply_blas_type(D,A,alpha);
|
|
|
|
// NOTE: assuming beta=1; this is okay for now, as currently glue_times only uses beta=1
|
|
arrayops::inplace_plus(C.memptr(), D.memptr(), C.n_elem);
|
|
|
|
return;
|
|
}
|
|
|
|
arma_extra_debug_print("blas::syrk()");
|
|
|
|
const char uplo = 'U';
|
|
|
|
const char trans_A = (do_trans_A) ? 'T' : 'N';
|
|
|
|
const blas_int n = blas_int(C.n_cols);
|
|
const blas_int k = (do_trans_A) ? blas_int(A.n_rows) : blas_int(A.n_cols);
|
|
|
|
const eT local_alpha = (use_alpha) ? alpha : eT(1);
|
|
const eT local_beta = (use_beta) ? beta : eT(0);
|
|
|
|
const blas_int lda = (do_trans_A) ? k : n;
|
|
|
|
arma_extra_debug_print( arma_boost::format("blas::syrk(): trans_A = %c") % trans_A );
|
|
|
|
blas::syrk<eT>
|
|
(
|
|
&uplo,
|
|
&trans_A,
|
|
&n,
|
|
&k,
|
|
&local_alpha,
|
|
A.mem,
|
|
&lda,
|
|
&local_beta,
|
|
C.memptr(),
|
|
&n // &ldc
|
|
);
|
|
|
|
syrk_helper::inplace_copy_upper_tri_to_lower_tri(C);
|
|
}
|
|
#else
|
|
{
|
|
syrk_emul<do_trans_A, use_alpha, use_beta>::apply(C,A,alpha,beta);
|
|
}
|
|
#endif
|
|
}
|
|
}
|
|
|
|
|
|
|
|
template<typename eT, typename TA>
|
|
inline
|
|
static
|
|
void
|
|
apply( Mat<eT>& C, const TA& A, const eT alpha = eT(1), const eT beta = eT(0) )
|
|
{
|
|
if(is_cx<eT>::no)
|
|
{
|
|
if(A.is_vec())
|
|
{
|
|
syrk_vec<do_trans_A, use_alpha, use_beta>::apply(C,A,alpha,beta);
|
|
}
|
|
else
|
|
{
|
|
syrk_emul<do_trans_A, use_alpha, use_beta>::apply(C,A,alpha,beta);
|
|
}
|
|
}
|
|
else
|
|
{
|
|
// handling of complex matrix by syrk_emul() is not yet implemented
|
|
return;
|
|
}
|
|
}
|
|
|
|
|
|
|
|
template<typename TA>
|
|
arma_inline
|
|
static
|
|
void
|
|
apply
|
|
(
|
|
Mat<float>& C,
|
|
const TA& A,
|
|
const float alpha = float(1),
|
|
const float beta = float(0)
|
|
)
|
|
{
|
|
syrk<do_trans_A, use_alpha, use_beta>::apply_blas_type(C,A,alpha,beta);
|
|
}
|
|
|
|
|
|
|
|
template<typename TA>
|
|
arma_inline
|
|
static
|
|
void
|
|
apply
|
|
(
|
|
Mat<double>& C,
|
|
const TA& A,
|
|
const double alpha = double(1),
|
|
const double beta = double(0)
|
|
)
|
|
{
|
|
syrk<do_trans_A, use_alpha, use_beta>::apply_blas_type(C,A,alpha,beta);
|
|
}
|
|
|
|
|
|
|
|
template<typename TA>
|
|
arma_inline
|
|
static
|
|
void
|
|
apply
|
|
(
|
|
Mat< std::complex<float> >& C,
|
|
const TA& A,
|
|
const std::complex<float> alpha = std::complex<float>(1),
|
|
const std::complex<float> beta = std::complex<float>(0)
|
|
)
|
|
{
|
|
arma_ignore(C);
|
|
arma_ignore(A);
|
|
arma_ignore(alpha);
|
|
arma_ignore(beta);
|
|
|
|
// handling of complex matrix by syrk() is not yet implemented
|
|
return;
|
|
}
|
|
|
|
|
|
|
|
template<typename TA>
|
|
arma_inline
|
|
static
|
|
void
|
|
apply
|
|
(
|
|
Mat< std::complex<double> >& C,
|
|
const TA& A,
|
|
const std::complex<double> alpha = std::complex<double>(1),
|
|
const std::complex<double> beta = std::complex<double>(0)
|
|
)
|
|
{
|
|
arma_ignore(C);
|
|
arma_ignore(A);
|
|
arma_ignore(alpha);
|
|
arma_ignore(beta);
|
|
|
|
// handling of complex matrix by syrk() is not yet implemented
|
|
return;
|
|
}
|
|
|
|
};
|
|
|
|
|
|
|
|
//! @}
|