Matrix multplication where the matrices have different element types. //! Simple version (no caching). //! Matrix 'C' is assumed to have been set to the correct size (i.e. taking into account transposes). More...
#include <gemm_mixed.hpp>
Static Public Member Functions | |
| template<typename out_eT , typename in_eT1 , typename in_eT2 > | |
| static arma_hot void | apply (Mat< out_eT > &C, const Mat< in_eT1 > &A, const Mat< in_eT2 > &B, const out_eT alpha=out_eT(1), const out_eT beta=out_eT(0)) |
Matrix multplication where the matrices have different element types. //! Simple version (no caching). //! Matrix 'C' is assumed to have been set to the correct size (i.e. taking into account transposes).
Definition at line 213 of file gemm_mixed.hpp.
| static arma_hot void gemm_mixed_simple< do_trans_A, do_trans_B, use_alpha, use_beta >::apply | ( | Mat< out_eT > & | C, | |
| const Mat< in_eT1 > & | A, | |||
| const Mat< in_eT2 > & | B, | |||
| const out_eT | alpha = out_eT(1), |
|||
| const out_eT | beta = out_eT(0) | |||
| ) | [inline, static] |
Definition at line 223 of file gemm_mixed.hpp.
References Mat< eT >::at(), Mat< eT >::colptr(), Mat< eT >::n_cols, and Mat< eT >::n_rows.
{
arma_extra_debug_sigprint();
const u32 A_n_rows = A.n_rows;
const u32 A_n_cols = A.n_cols;
const u32 B_n_rows = B.n_rows;
const u32 B_n_cols = B.n_cols;
if( (do_trans_A == false) && (do_trans_B == false) )
{
for(u32 row_A = 0; row_A < A_n_rows; ++row_A)
{
for(u32 col_B = 0; col_B < B_n_cols; ++col_B)
{
const in_eT2* B_coldata = B.colptr(col_B);
out_eT acc = out_eT(0);
for(u32 i = 0; i < B_n_rows; ++i)
{
const out_eT val1 = upgrade_val<in_eT1,in_eT2>::apply(A.at(row_A,i));
const out_eT val2 = upgrade_val<in_eT1,in_eT2>::apply(B_coldata[i]);
acc += val1 * val2;
//acc += upgrade_val<in_eT1,in_eT2>::apply(A.at(row_A,i)) * upgrade_val<in_eT1,in_eT2>::apply(B_coldata[i]);
}
if( (use_alpha == false) && (use_beta == false) )
{
C.at(row_A,col_B) = acc;
}
else
if( (use_alpha == true) && (use_beta == false) )
{
C.at(row_A,col_B) = alpha * acc;
}
else
if( (use_alpha == false) && (use_beta == true) )
{
C.at(row_A,col_B) = acc + beta*C.at(row_A,col_B);
}
else
if( (use_alpha == true) && (use_beta == true) )
{
C.at(row_A,col_B) = alpha*acc + beta*C.at(row_A,col_B);
}
}
}
}
else
if( (do_trans_A == true) && (do_trans_B == false) )
{
for(u32 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 in_eT1* A_coldata = A.colptr(col_A);
for(u32 col_B=0; col_B < B_n_cols; ++col_B)
{
const in_eT2* B_coldata = B.colptr(col_B);
out_eT acc = out_eT(0);
for(u32 i=0; i < B_n_rows; ++i)
{
acc += upgrade_val<in_eT1,in_eT2>::apply(A_coldata[i]) * upgrade_val<in_eT1,in_eT2>::apply(B_coldata[i]);
}
if( (use_alpha == false) && (use_beta == false) )
{
C.at(col_A,col_B) = acc;
}
else
if( (use_alpha == true) && (use_beta == false) )
{
C.at(col_A,col_B) = alpha * acc;
}
else
if( (use_alpha == false) && (use_beta == true) )
{
C.at(col_A,col_B) = acc + beta*C.at(col_A,col_B);
}
else
if( (use_alpha == true) && (use_beta == true) )
{
C.at(col_A,col_B) = alpha*acc + beta*C.at(col_A,col_B);
}
}
}
}
else
if( (do_trans_A == false) && (do_trans_B == true) )
{
for(u32 row_A = 0; row_A < A_n_rows; ++row_A)
{
for(u32 row_B = 0; row_B < B_n_rows; ++row_B)
{
out_eT acc = out_eT(0);
for(u32 i = 0; i < B_n_cols; ++i)
{
acc += upgrade_val<in_eT1,in_eT2>::apply(A.at(row_A,i)) * upgrade_val<in_eT1,in_eT2>::apply(B.at(row_B,i));
}
if( (use_alpha == false) && (use_beta == false) )
{
C.at(row_A,row_B) = acc;
}
else
if( (use_alpha == true) && (use_beta == false) )
{
C.at(row_A,row_B) = alpha * acc;
}
else
if( (use_alpha == false) && (use_beta == true) )
{
C.at(row_A,row_B) = acc + beta*C.at(row_A,row_B);
}
else
if( (use_alpha == true) && (use_beta == true) )
{
C.at(row_A,row_B) = alpha*acc + beta*C.at(row_A,row_B);
}
}
}
}
else
if( (do_trans_A == true) && (do_trans_B == true) )
{
for(u32 row_B=0; row_B < B_n_rows; ++row_B)
{
for(u32 col_A=0; col_A < A_n_cols; ++col_A)
{
const in_eT1* A_coldata = A.colptr(col_A);
out_eT acc = out_eT(0);
for(u32 i=0; i < A_n_rows; ++i)
{
acc += upgrade_val<in_eT1,in_eT2>::apply(B.at(row_B,i)) * upgrade_val<in_eT1,in_eT2>::apply(A_coldata[i]);
}
if( (use_alpha == false) && (use_beta == false) )
{
C.at(col_A,row_B) = acc;
}
else
if( (use_alpha == true) && (use_beta == false) )
{
C.at(col_A,row_B) = alpha * acc;
}
else
if( (use_alpha == false) && (use_beta == true) )
{
C.at(col_A,row_B) = acc + beta*C.at(col_A,row_B);
}
else
if( (use_alpha == true) && (use_beta == true) )
{
C.at(col_A,row_B) = alpha*acc + beta*C.at(col_A,row_B);
}
}
}
}
}