void FLA_Herk_cntl_init()
{
// Set blocksizes with default values for conventional storage.
fla_herk_var2_bsize = FLA_Query_blocksizes( FLA_DIMENSION_MIN );
fla_herk_var5_bsize = FLA_Query_blocksizes( FLA_DIMENSION_MIN );
// Create a control tree that assumes A is a b x b block.
fla_herk_cntl_blas = FLA_Cntl_herk_obj_create( FLA_FLAT,
FLA_SUBPROBLEM,
NULL,
NULL,
NULL,
NULL );
// Create a control tree that assumes A * A' forms an inner panel product.
fla_herk_cntl_ip = FLA_Cntl_herk_obj_create( FLA_FLAT,
FLA_BLOCKED_VARIANT5,
fla_herk_var5_bsize,
fla_scalr_cntl_blas,
fla_herk_cntl_blas,
NULL );
// Create a control tree that assumes A * A' forms an outer panel product.
fla_herk_cntl_op = FLA_Cntl_herk_obj_create( FLA_FLAT,
FLA_BLOCKED_VARIANT2,
fla_herk_var2_bsize,
fla_scalr_cntl_blas,
fla_herk_cntl_blas,
fla_gemm_cntl_blas );
// Create a control tree that assumes A is large.
fla_herk_cntl_mm = FLA_Cntl_herk_obj_create( FLA_FLAT,
FLA_BLOCKED_VARIANT5,
fla_herk_var5_bsize,
fla_scalr_cntl_blas,
fla_herk_cntl_op,
NULL );
}
void FLASH_Herk_cntl_init()
{
// Set herk blocksize for hierarchical storage.
flash_herk_bsize = FLA_Blocksize_create( 1, 1, 1, 1 );
// Create a control tree that assumes A is a b x b block.
flash_herk_cntl_blas = FLA_Cntl_herk_obj_create( FLA_HIER,
FLA_SUBPROBLEM,
NULL,
NULL,
NULL,
NULL );
// Create a control tree that assumes A * A' forms an inner panel product.
flash_herk_cntl_ip = FLA_Cntl_herk_obj_create( FLA_HIER,
FLA_BLOCKED_VARIANT5,
flash_herk_bsize,
flash_scalr_cntl,
flash_herk_cntl_blas,
NULL );
// Create a control tree that assumes A * A' forms an outer panel product.
flash_herk_cntl_op = FLA_Cntl_herk_obj_create( FLA_HIER,
FLA_BLOCKED_VARIANT2,
flash_herk_bsize,
flash_scalr_cntl,
flash_herk_cntl_blas,
flash_gemm_cntl_pb_bb );
// Create a control tree that assumes A is large.
flash_herk_cntl_mm = FLA_Cntl_herk_obj_create( FLA_HIER,
FLA_BLOCKED_VARIANT5,
flash_herk_bsize,
flash_scalr_cntl,
flash_herk_cntl_op,
NULL );
}
void time_Chol_u(
int variant, int type, int nrepeats, int n, int nb_alg,
FLA_Obj A, FLA_Obj b, FLA_Obj b_orig, FLA_Obj norm,
double *dtime, double *diff, double *gflops )
{
int
irep;
double
dtime_save = 1.0e9;
FLA_Obj
A_save, b_save, b_orig_save;
fla_blocksize_t*
bp;
fla_chol_t*
cntl_chol_var;
fla_chol_t*
cntl_chol_unb;
fla_syrk_t*
cntl_syrk_blas;
fla_herk_t*
cntl_herk_blas;
fla_trsm_t*
cntl_trsm_blas;
fla_gemm_t*
cntl_gemm_blas;
/*
if( type == FLA_ALG_UNBLOCKED && n > 400 )
{
*gflops = 0.0;
*diff = 0.0;
return;
}
*/
bp = FLA_Blocksize_create( nb_alg, nb_alg, nb_alg, nb_alg );
cntl_chol_unb = FLA_Cntl_chol_obj_create( FLA_FLAT, FLA_UNB_OPT_VARIANT2, NULL, NULL, NULL, NULL, NULL, NULL );
cntl_syrk_blas = FLA_Cntl_syrk_obj_create( FLA_FLAT, FLA_SUBPROBLEM, NULL, NULL, NULL );
cntl_herk_blas = FLA_Cntl_herk_obj_create( FLA_FLAT, FLA_SUBPROBLEM, NULL, NULL, NULL );
cntl_trsm_blas = FLA_Cntl_trsm_obj_create( FLA_FLAT, FLA_SUBPROBLEM, NULL, NULL, NULL );
cntl_gemm_blas = FLA_Cntl_gemm_obj_create( FLA_FLAT, FLA_SUBPROBLEM, NULL, NULL );
cntl_chol_var = FLA_Cntl_chol_obj_create( FLA_FLAT, variant, bp,
cntl_chol_unb,
cntl_syrk_blas,
cntl_herk_blas,
cntl_trsm_blas,
cntl_gemm_blas );
FLA_Obj_create_conf_to( FLA_NO_TRANSPOSE, A, &A_save );
FLA_Obj_create_conf_to( FLA_NO_TRANSPOSE, b, &b_save );
FLA_Obj_create_conf_to( FLA_NO_TRANSPOSE, b_orig, &b_orig_save );
FLA_Copy_external( A, A_save );
FLA_Copy_external( b, b_save );
FLA_Copy_external( b_orig, b_orig_save );
for ( irep = 0 ; irep < nrepeats; irep++ ){
FLA_Copy_external( A_save, A );
*dtime = FLA_Clock();
switch( variant ){
case 0:
REF_Chol_u( A );
break;
case 1:{
// Time variant 1
switch( type ){
case FLA_ALG_UNBLOCKED:
FLA_Chol_u_unb_var1( A );
break;
case FLA_ALG_UNB_OPT:
FLA_Chol_u_opt_var1( A );
break;
case FLA_ALG_BLOCKED:
FLA_Chol_u_blk_var1( A, cntl_chol_var );
break;
default:
printf("trouble\n");
}
break;
}
case 2:{
// Time variant 2
switch( type ){
case FLA_ALG_UNBLOCKED:
FLA_Chol_u_unb_var2( A );
break;
case FLA_ALG_UNB_OPT:
FLA_Chol_u_opt_var2( A );
break;
case FLA_ALG_BLOCKED:
FLA_Chol_u_blk_var2( A, cntl_chol_var );
break;
default:
printf("trouble\n");
}
break;
}
case 3:{
// Time variant 3
switch( type ){
case FLA_ALG_UNBLOCKED:
FLA_Chol_u_unb_var3( A );
break;
case FLA_ALG_UNB_OPT:
FLA_Chol_u_opt_var3( A );
break;
case FLA_ALG_BLOCKED:
FLA_Chol_u_blk_var3( A, cntl_chol_var );
break;
default:
printf("trouble\n");
}
break;
}
}
*dtime = FLA_Clock() - *dtime;
dtime_save = min( *dtime, dtime_save );
}
FLA_Cntl_obj_free( cntl_chol_var );
FLA_Cntl_obj_free( cntl_chol_unb );
FLA_Cntl_obj_free( cntl_syrk_blas );
FLA_Cntl_obj_free( cntl_herk_blas );
FLA_Cntl_obj_free( cntl_trsm_blas );
FLA_Cntl_obj_free( cntl_gemm_blas );
FLA_Blocksize_free( bp );
if ( type == FLA_ALG_REFERENCE )
{
FLA_Trsv_external( FLA_UPPER_TRIANGULAR, FLA_CONJ_TRANSPOSE,
FLA_UNIT_DIAG, A, b );
FLA_Trsv_external( FLA_UPPER_TRIANGULAR, FLA_NO_TRANSPOSE,
FLA_NONUNIT_DIAG, A, b );
FLA_Hemv_external( FLA_UPPER_TRIANGULAR,
FLA_ONE, A_save, b, FLA_MINUS_ONE, b_orig );
FLA_Nrm2_external( b_orig, norm );
FLA_Copy_object_to_buffer( FLA_NO_TRANSPOSE, 0, 0, norm,
1, 1, diff, 1, 1 );
}
else
{
FLA_Trsv_external( FLA_UPPER_TRIANGULAR, FLA_CONJ_TRANSPOSE,
FLA_UNIT_DIAG, A, b );
FLA_Trsv_external( FLA_UPPER_TRIANGULAR, FLA_NO_TRANSPOSE,
FLA_NONUNIT_DIAG, A, b );
FLA_Hemv_external( FLA_UPPER_TRIANGULAR,
FLA_ONE, A_save, b, FLA_MINUS_ONE, b_orig );
FLA_Nrm2_external( b_orig, norm );
FLA_Copy_object_to_buffer( FLA_NO_TRANSPOSE, 0, 0, norm,
1, 1, diff, 1, 1 );
}
*gflops = 1.0 / 3.0 *
FLA_Obj_length( A ) *
FLA_Obj_length( A ) *
FLA_Obj_length( A ) /
dtime_save / 1e9;
if ( FLA_Obj_is_complex( A ) )
*gflops *= 4.0;
*dtime = dtime_save;
FLA_Copy_external( A_save, A );
FLA_Copy_external( b_save, b );
FLA_Copy_external( b_orig_save, b_orig );
FLA_Obj_free( &A_save );
FLA_Obj_free( &b_save );
FLA_Obj_free( &b_orig_save );
}