FLA_Error FLA_Trmm_lut( FLA_Diag diag, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, fla_trmm_t* cntl )
{
FLA_Error r_val = FLA_SUCCESS;
if ( FLA_Cntl_variant( cntl ) == FLA_SUBPROBLEM )
{
r_val = FLA_Trmm_lut_task( diag, alpha, A, B, cntl );
}
#ifdef FLA_ENABLE_NON_CRITICAL_CODE
else if ( FLA_Cntl_variant( cntl ) == FLA_BLOCKED_VARIANT1 )
{
r_val = FLA_Trmm_lut_blk_var1( diag, alpha, A, B, cntl );
}
#endif
else if ( FLA_Cntl_variant( cntl ) == FLA_BLOCKED_VARIANT2 )
{
r_val = FLA_Trmm_lut_blk_var2( diag, alpha, A, B, cntl );
}
else if ( FLA_Cntl_variant( cntl ) == FLA_BLOCKED_VARIANT3 )
{
r_val = FLA_Trmm_lut_blk_var3( diag, alpha, A, B, cntl );
}
#ifdef FLA_ENABLE_NON_CRITICAL_CODE
else if ( FLA_Cntl_variant( cntl ) == FLA_BLOCKED_VARIANT4 )
{
r_val = FLA_Trmm_lut_blk_var4( diag, alpha, A, B, cntl );
}
#endif
#ifdef FLA_ENABLE_NON_CRITICAL_CODE
else if ( FLA_Cntl_variant( cntl ) == FLA_UNBLOCKED_VARIANT1 )
{
r_val = FLA_Trmm_lut_unb_var1( diag, alpha, A, B );
}
else if ( FLA_Cntl_variant( cntl ) == FLA_UNBLOCKED_VARIANT2 )
{
r_val = FLA_Trmm_lut_unb_var2( diag, alpha, A, B );
}
else if ( FLA_Cntl_variant( cntl ) == FLA_UNBLOCKED_VARIANT3 )
{
r_val = FLA_Trmm_lut_unb_var3( diag, alpha, A, B );
}
else if ( FLA_Cntl_variant( cntl ) == FLA_UNBLOCKED_VARIANT4 )
{
r_val = FLA_Trmm_lut_unb_var4( diag, alpha, A, B );
}
#endif
else
{
r_val = FLA_Check_error_code( FLA_NOT_YET_IMPLEMENTED );
}
return r_val;
}
void time_Trmm_lut(
int variant, int type, int nrepeats, int n, int nb_alg,
FLA_Obj A, FLA_Obj B, FLA_Obj C, FLA_Obj C_ref,
double *dtime, double *diff, double *gflops )
{
int
irep;
double
dtime_old = 1.0e9;
FLA_Obj
C_old;
fla_blocksize_t*
bp;
fla_gemm_t*
cntl_gemm_blas;
fla_trmm_t*
cntl_trmm_blas;
fla_trmm_t*
cntl_trmm_var;
bp = FLA_Blocksize_create( nb_alg, nb_alg, nb_alg, nb_alg );
cntl_gemm_blas = FLA_Cntl_gemm_obj_create( FLA_FLAT, FLA_SUBPROBLEM, NULL, NULL );
cntl_trmm_blas = FLA_Cntl_trmm_obj_create( FLA_FLAT, FLA_SUBPROBLEM, NULL, NULL, NULL );
cntl_trmm_var = FLA_Cntl_trmm_obj_create( FLA_FLAT, variant, bp, cntl_trmm_blas, cntl_gemm_blas );
FLA_Obj_create_conf_to( FLA_NO_TRANSPOSE, C, &C_old );
FLA_Copy_external( C, C_old );
for ( irep = 0 ; irep < nrepeats; irep++ )
{
FLA_Copy_external( C_old, C );
*dtime = FLA_Clock();
switch( variant ){
case 0:
// Time reference implementation
REF_Trmm( FLA_LEFT, FLA_UPPER_TRIANGULAR,
FLA_TRANSPOSE, FLA_NONUNIT_DIAG,
FLA_ONE, A, C );
break;
case 1:{
// Time variant 1
switch( type ){
case FLA_ALG_UNBLOCKED:
FLA_Trmm_lut_unb_var1( FLA_NONUNIT_DIAG, FLA_ONE, A, C );
break;
case FLA_ALG_BLOCKED:
FLA_Trmm_lut_blk_var1( FLA_NONUNIT_DIAG, FLA_ONE, A, C, cntl_trmm_var );
break;
default:
printf("trouble\n");
}
break;
}
case 2:{
// Time variant 2
switch( type ){
case FLA_ALG_UNBLOCKED:
FLA_Trmm_lut_unb_var2( FLA_NONUNIT_DIAG, FLA_ONE, A, C );
break;
case FLA_ALG_BLOCKED:
FLA_Trmm_lut_blk_var2( FLA_NONUNIT_DIAG, FLA_ONE, A, C, cntl_trmm_var );
break;
default:
printf("trouble\n");
}
break;
}
case 3:{
// Time variant 3
switch( type ){
case FLA_ALG_UNBLOCKED:
FLA_Trmm_lut_unb_var3( FLA_NONUNIT_DIAG, FLA_ONE, A, C );
break;
case FLA_ALG_BLOCKED:
FLA_Trmm_lut_blk_var3( FLA_NONUNIT_DIAG, FLA_ONE, A, C, cntl_trmm_var );
break;
default:
printf("trouble\n");
}
break;
}
case 4:{
// Time variant 4
switch( type ){
case FLA_ALG_UNBLOCKED:
FLA_Trmm_lut_unb_var4( FLA_NONUNIT_DIAG, FLA_ONE, A, C );
break;
case FLA_ALG_BLOCKED:
FLA_Trmm_lut_blk_var4( FLA_NONUNIT_DIAG, FLA_ONE, A, C, cntl_trmm_var );
break;
default:
printf("trouble\n");
}
break;
}
}
*dtime = FLA_Clock() - *dtime;
dtime_old = min( *dtime, dtime_old );
}
FLA_Cntl_obj_free( cntl_trmm_var );
FLA_Cntl_obj_free( cntl_trmm_blas );
FLA_Cntl_obj_free( cntl_gemm_blas );
FLA_Blocksize_free( bp );
if ( variant == 0 )
{
FLA_Copy_external( C, C_ref );
*diff = 0.0;
}
else
{
*diff = FLA_Max_elemwise_diff( C, C_ref );
}
*gflops = 1.0 *
FLA_Obj_length( C ) *
FLA_Obj_width( C ) *
FLA_Obj_width( A ) /
dtime_old /
1.0e9;
*dtime = dtime_old;
FLA_Copy_external( C_old, C );
FLA_Obj_free( &C_old );
}