amd/blis
1
0
Fork 0
mirror of https://github.com/amd/blis.git synced 2026-04-20 15:48:50 +00:00
Code Issues Packages Projects Releases Wiki Activity

Removed unnecessary pack checks in FP32 GEMV (#54)

Browse Source 
Details:
- In FP32 GEMM, when threading is disabled, rntm_pack_a and rntm_pack_b
  were set to true by default. This leads to perf regression for smaller
  sizes. Modified FP32 interface API to not overwrite the packA and
  packB variables in rntm structure.
- In FP32 GEMV, Removed the decision making code based on mtag_A/B
  and should_pack_A/B for packing. Matrices will be packed only
  if the storage format of the matrices doesn't match the storage
  format required by the kernel.
- Changed the control flow of checking the value of mtag to whether
  matrix is "reordered" or "to-be-packed" or "unpacked". checking
  for "reorder" first, followed by "pack". This will ensure that
  packing doesn't happen when the matrix is already reordered even
  though user forces packing by setting "BLIS_PACK_A/B"
-Modified python script to generate testcases based on block sizes

AMD-Internal: SWLCSG-3527
This commit is contained in:
Vankadari, Meghana
2025-06-16 12:34:11 +05:30
committed by GitHub
parent 1847a1e8c6
commit 8649cdc14b
4 changed files with 165 additions and 73 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
addon/aocl_gemm/aocl_gemm_f32f32f32of32.c
View File

@@ -285,10 +285,6 @@ AOCL_GEMM_MATMUL(float,float,float,float,f32f32f32of32)
);
}
#else
// Setting pack A and B by default for non open mp case.
bli_rntm_set_pack_a( 1, &rntm_g );
bli_rntm_set_pack_b( 1, &rntm_g );
// Swapping inputs to induce row major computation for column major inputs.
if ( is_column_major == TRUE )
{

54
addon/aocl_gemm/frame/f32f32f32/lpgemm_f32f32f32.c
View File

@@ -124,6 +124,10 @@ typedef void (*lpgemv_a_pack_ft)
LPGEMV(float, float, float, f32f32f32of32)
{
/* Ignoring mtag_a/b and should_pack_A/B for now .
Matrices are packed only when the storage format is not supported by the kernel.
*/
const float* a_use = (float*)a;
inc_t rs_a_use = rs_a;
inc_t cs_a_use = cs_a;
@@ -154,12 +158,6 @@ LPGEMV(float, float, float, f32f32f32of32)
if (c_downscale < F32) post_ops_attr.buf_downscale = c;
else post_ops_attr.buf_downscale = NULL;
// Should_pack_A/B is set either by the user through env variable
// or by the smart threading logic based on work distribution.
// Storage format of the matrices doesn't affect should_pack_A/B.
bool should_pack_B = bli_rntm_pack_b( rntm );
bool should_pack_A = bli_rntm_pack_a( rntm );
// Generate thrinfo objects for jc and ic loops from lpgemm_thrinfo_t.
thrinfo_t thread_jc;
thrinfo_t thread_ic;
@@ -195,7 +193,7 @@ LPGEMV(float, float, float, f32f32f32of32)
packa_fp = packa_mr8_f32f32f32of32_col_major;
#endif
// Pack B matrix if rs_b > 1
if( (should_pack_B == TRUE) || ( rs_b != 1 ) )
if( rs_b != 1 )
{
mem_b_size_req = sizeof( float ) * k;
@@ -233,7 +231,7 @@ LPGEMV(float, float, float, f32f32f32of32)
post_ops_attr.post_op_c_i = ic;
// To-Do: pack A case needs to be handled for AVX2 case.
if( (should_pack_A == TRUE) || ( cs_a != 1 ) )
if( cs_a != 1 )
{
mem_a_size_req = sizeof(float) * mc0 * k;
lpgemm_alloc_mem_panel
@@ -264,11 +262,11 @@ LPGEMV(float, float, float, f32f32f32of32)
&post_ops_attr
);
}
if ( ( (should_pack_A == TRUE) || ( cs_a != 1 ) ) && ( bli_mem_is_alloc( &mem_a ) ) )
if ( ( cs_a != 1 ) && ( bli_mem_is_alloc( &mem_a ) ) )
{
bli_pba_release( rntm, &mem_a );
}
if ( ( (should_pack_B == TRUE) || ( rs_b != 1 ) ) && ( bli_mem_is_alloc( &mem_b ) ) )
if ( ( rs_b != 1 ) && ( bli_mem_is_alloc( &mem_b ) ) )
{
bli_pba_release( rntm, &mem_b );
}
@@ -300,7 +298,7 @@ LPGEMV(float, float, float, f32f32f32of32)
thread_jc.work_id = thread->tid;
bli_thread_range_sub(&thread_jc, n, NR, FALSE, &jc_start, &jc_end);
if ( (should_pack_A == TRUE) || ( cs_a != 1 ) )
if ( cs_a != 1 )
{
mem_a_size_req = sizeof( float ) * k;
@@ -346,7 +344,7 @@ LPGEMV(float, float, float, f32f32f32of32)
rs_b_use = NR;
cs_b_use = 1;
}
else if ( (should_pack_B == TRUE) || ( mtag_b == PACK ) )
else if ( mtag_b == PACK )
{
// nc0 needs to be a multiple of 16 since this gives maximum
// vectorization. Packing B always results in buffers with width
@@ -412,12 +410,12 @@ LPGEMV(float, float, float, f32f32f32of32)
} // jc loop
// Release pack buffers.
if ( ( (should_pack_B == TRUE) || ( mtag_b == PACK ) ) && ( bli_mem_is_alloc( &mem_b ) ) )
if ( ( mtag_b == PACK ) && ( bli_mem_is_alloc( &mem_b ) ) )
{
bli_pba_release( rntm, &mem_b );
}
if ( ( (should_pack_A == TRUE) || ( cs_a != 1 ) ) && ( bli_mem_is_alloc( &mem_a ) ) )
if ( ( cs_a != 1 ) && ( bli_mem_is_alloc( &mem_a ) ) )
{
bli_pba_release( rntm, &mem_a );
}
@@ -569,7 +567,20 @@ LPGEMM_5LOOP(float, float, float, f32f32f32of32)
is_last_k = ( ( pc + KC ) >= k ) ? ( TRUE ) : ( FALSE );
post_ops_attr.is_last_k = is_last_k;
if ( ( mtag_b == PACK ) || ( should_pack_B == TRUE ) )
if ( mtag_b == REORDERED )
{
// In multi-threaded scenarios, an extra offset into a given
// packed B panel is required, since the jc loop split can
// result in per thread start offset inside the panel, instead
// of panel boundaries.
b_use = b + ( jc_cur_loop * k ) +
( n_sub_updated * pc ) + ( jc_cur_loop_rem * kc0 );
rs_b_use = NR;
cs_b_use = 1;
ps_b_use = kc0;
}
else if ( ( mtag_b == PACK ) || ( should_pack_B == TRUE ) )
{
// Pack B chunks are based on jc work id.
dim_t jc_work_id = bli_thread_work_id( &thread_jc );
@@ -649,19 +660,6 @@ LPGEMM_5LOOP(float, float, float, f32f32f32of32)
);
b_use = pack_b_buffer_f32f32f32of32;
}
else if ( mtag_b == REORDERED )
{
// In multi-threaded scenarios, an extra offset into a given
// packed B panel is required, since the jc loop split can
// result in per thread start offset inside the panel, instead
// of panel boundaries.
b_use = b + ( jc_cur_loop * k ) +
( n_sub_updated * pc ) + ( jc_cur_loop_rem * kc0 );
rs_b_use = NR;
cs_b_use = 1;
ps_b_use = kc0;
}
else
{
b_use = b + ( pc * rs_b ) + ( jc * cs_b );

30
addon/aocl_gemm/frame/f32f32f32/lpgemm_f32f32f32_tiny.c
View File

@@ -266,7 +266,13 @@ LPGEMV_TINY(float, float, float, f32f32f32of32)
cs_a_use = 1;
}
if ( ( mtag_b == PACK ) )
if ( mtag_b == REORDERED )
{
b_use = ( float* )b;
rs_b_use = NR;
cs_b_use = 1;
}
else if ( ( mtag_b == PACK ) )
{
dim_t nc0_updated = make_multiple_of_n(n, NR);
siz_t mem_b_size_req = sizeof(float) * nc0_updated * k;
@@ -288,12 +294,6 @@ LPGEMV_TINY(float, float, float, f32f32f32of32)
b_use = pack_b_buffer_f32f32f32of32;
}
else if ( mtag_b == REORDERED )
{
b_use = ( float* )b;
rs_b_use = NR;
cs_b_use = 1;
}
else
{
b_use = ( float* )b;
@@ -388,7 +388,14 @@ LPGEMM_TINY(float,float,float,f32f32f32of32)
// Even if the mtag_b is set to PACK, for tiny sizes its better to
// pack only if it affects output accuracy (like column major B),
// else ignore it.
if ( ( mtag_b == PACK ) )
if ( mtag_b == REORDERED )
{
b_use = b;
rs_b_use = NR;
cs_b_use = 1;
ps_b_use = k;
}
else if ( ( mtag_b == PACK ) )
{
dim_t nc0_updated = make_multiple_of_n( n, NR );
mem_b_size_req = sizeof( float ) * nc0_updated * k;
@@ -410,13 +417,6 @@ LPGEMM_TINY(float,float,float,f32f32f32of32)
b_use = pack_b_buffer_f32f32f32of32;
}
else if ( mtag_b == REORDERED )
{
b_use = b;
rs_b_use = NR;
cs_b_use = 1;
ps_b_use = k;
}
else
{
b_use = b;