amd/blis
1
0
Fork 0
mirror of https://github.com/amd/blis.git synced 2026-05-11 17:50:00 +00:00
Code Issues Packages Projects Releases Wiki Activity

Enable 1m only if real domain ukr is not reference.

Browse Source 
Details:
- Previously, BLIS would automatically enable use of the 1m method
  for a given precision if the complex domain microkernel was a
  reference kernel. This commit adds an additional constraint so that
  1m is only enabled if the corresponding real domain microkernel is
  NOT reference. That is, BLIS now forgos use of 1m if both the real and
  complex domain kernels are reference implementations. Note that this
  does not prevent 1m from being enabled manually under those
  conditions; it only means that 1m will not be enabled automatically
  at initialization-time.
This commit is contained in:
Field G. Van Zee
2021-01-04 16:16:32 -06:00
parent ed50c94738
commit c3ed2cbb9f
1 changed files with 16 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

25
frame/ind/bli_ind.c
View File

@@ -49,20 +49,27 @@ static char* bli_ind_impl_str[BLIS_NUM_IND_METHODS] =
void bli_ind_init( void )
{
// Enable the default induced method (1m) if one or both complex domain
// gemm micro-kernels are unoptimized in the native context.
// NOTE: Instead of calling bli_gks_query_cntx(), we call
// bli_gks_query_cntx_noinit() to avoid the call to bli_init_once().
cntx_t* cntx = bli_gks_query_cntx_noinit();
bool c_is_ref = bli_gks_cntx_l3_nat_ukr_is_ref
( BLIS_SCOMPLEX, BLIS_GEMM_UKR, cntx );
bool z_is_ref = bli_gks_cntx_l3_nat_ukr_is_ref
( BLIS_DCOMPLEX, BLIS_GEMM_UKR, cntx );
// For each precision, enable the default induced method (1m) if both of
// the following conditions are met:
// - the complex domain kernel is the (unoptimized) reference kernel
// - the real domain kernel is NOT the (unoptimized) reference kernel
// The second condition means that BLIS will not bother to use an induced
// method if the real domain kernel upon which that induced method would
// rely is unoptimized. (There is no point in using an induced method based
// on a reference real domain kernel when we can use the reference complex
// domain kernel instead.)
if ( c_is_ref ) bli_ind_enable_dt( BLIS_1M, BLIS_SCOMPLEX );
if ( z_is_ref ) bli_ind_enable_dt( BLIS_1M, BLIS_DCOMPLEX );
bool s_is_ref = bli_gks_cntx_l3_nat_ukr_is_ref( BLIS_FLOAT, BLIS_GEMM_UKR, cntx );
bool d_is_ref = bli_gks_cntx_l3_nat_ukr_is_ref( BLIS_DOUBLE, BLIS_GEMM_UKR, cntx );
bool c_is_ref = bli_gks_cntx_l3_nat_ukr_is_ref( BLIS_SCOMPLEX, BLIS_GEMM_UKR, cntx );
bool z_is_ref = bli_gks_cntx_l3_nat_ukr_is_ref( BLIS_DCOMPLEX, BLIS_GEMM_UKR, cntx );
if ( c_is_ref && !s_is_ref ) bli_ind_enable_dt( BLIS_1M, BLIS_SCOMPLEX );
if ( z_is_ref && !d_is_ref ) bli_ind_enable_dt( BLIS_1M, BLIS_DCOMPLEX );
}
void bli_ind_finalize( void )