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

Enabled optional packing of B in sgemm sup

Browse Source 
Details:
- Enabling packing of B helping in performance in sgemm when
  all m,n,k dimensions are above 240 irrespective of the lda alignment.
- We may extend this optional enablement further for other skinny types
  and incase of multithread scenarios.

Change-Id: Icb2a21e458cdcb0f8fdce373d8d0860c51be8d21
This commit is contained in:
Nallani Bhaskar
2021-06-25 14:05:55 +05:30
parent fe3384b3c6
commit 650005e6fe
1 changed files with 31 additions and 18 deletions
Download Patch File Download Diff File Expand all files Collapse all files

49
frame/3/bli_l3_sup_int.c
View File

@@ -52,6 +52,7 @@ err_t bli_gemmsup_int
const num_t dt = bli_obj_dt( c );
const dim_t m = bli_obj_length( c );
const dim_t n = bli_obj_width( c );
const dim_t k = bli_obj_width( a );
const dim_t MR = bli_cntx_get_blksz_def_dt( dt, BLIS_MR, cntx );
const dim_t NR = bli_cntx_get_blksz_def_dt( dt, BLIS_NR, cntx );
const bool auto_factor = bli_rntm_auto_factor( rntm );
@@ -84,14 +85,13 @@ err_t bli_gemmsup_int
return BLIS_FAILURE;
}
if ( is_rrr_rrc_rcr_crr )
{
// This branch handles:
// - rrr rrc rcr crr for row-preferential kernels
// - rcc crc ccr ccc for column-preferential kernels
// - Currently only row-preferential kernels are only supported.
// - rrr rrc rcr crr for row-preferential kernels
// - rcc crc ccr ccc for column-preferential kernels
// - Currently only row-preferential kernels are only supported.
// calculate number of micropanels in m and n dimensions and
// recalculate the automatic thread factorization based on these number of micropanels
const dim_t mu = m / MR;
@@ -101,18 +101,24 @@ err_t bli_gemmsup_int
// with a new factorization based on the matrix dimensions in units
// of micropanels.
if ( auto_factor )
{
{
// In the block-panel algorithm, the m dimension is parallelized
// with ic_nt and the n dimension is parallelized with jc_nt.
bli_thread_partition_2x2( n_threads, mu, nu, &ic_new, &jc_new );
// Update the ways of parallelism for the jc and ic loops, and then
// update the current thread's root thrinfo_t node according to the
// new ways of parallelism value for the jc loop.
bli_rntm_set_ways_only( jc_new, 1, ic_new, 1, 1, rntm );
bli_l3_sup_thrinfo_update_root( rntm, thread );
}
}
/*Enable packing for B matrix for higher sizes*/
if(bli_is_float(dt) && (n_threads==1)) {
if((m > 240) && (k > 240) && (n > 240))
bli_rntm_set_pack_b( 1, rntm );
}
bli_gemmsup_ref_var2m( BLIS_NO_TRANSPOSE,
alpha, a, b, beta, c,
stor_id, cntx, rntm, thread );
@@ -120,28 +126,35 @@ err_t bli_gemmsup_int
else
{
// This branch handles:
// - rrr rrc rcr crr for column-preferential kernels
// - rcc crc ccr ccc for row-preferential kernels
// - Currently only row-preferential kernels are only supported.
// - rrr rrc rcr crr for column-preferential kernels
// - rcc crc ccr ccc for row-preferential kernels
// - Currently only row-preferential kernels are only supported.
const dim_t mu = n / MR; // the n becomes m after a transposition
const dim_t nu = m / NR; // the m becomes n after a transposition
if ( auto_factor )
{
{
// In the block-panel algorithm, the m dimension is parallelized
// with ic_nt and the n dimension is parallelized with jc_nt.
bli_thread_partition_2x2( n_threads, mu, nu, &ic_new, &jc_new );
// Update the ways of parallelism for the jc and ic loops, and then
// update the current thread's root thrinfo_t node according to the
// new ways of parallelism value for the jc loop.
bli_rntm_set_ways_only( jc_new, 1, ic_new, 1, 1, rntm );
bli_l3_sup_thrinfo_update_root( rntm, thread );
}
}
/* Enable packing for B matrix for higher sizes. Note that pack A
* becomes pack B inside var2m because this is transpose case*/
if(bli_is_float(dt) && (n_threads==1)) {
if((m > 240) && (k > 240) && (n > 240))
bli_rntm_set_pack_a( 1, rntm );
}
bli_gemmsup_ref_var2m( BLIS_TRANSPOSE,
alpha, a, b, beta, c,
stor_id, cntx, rntm, thread );
alpha, a, b, beta, c,
stor_id, cntx, rntm, thread );
}
AOCL_DTL_TRACE_EXIT(AOCL_DTL_LEVEL_TRACE_4);