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

Performance fix for FP32 GEMV (#47)

Browse Source 
Details:
- In FP32 GEMM interface, mtag_b is being set to PACK by default.
  This is leading to packing of B matrix even though packing is not
  absolutely required leading to perf regression.
- Setting mtag_b to PACK only if it is absolutely necessary to pack B matrix
  modified check conditions before packing appropriately.

AMD-Internal - [SWLCSG-3575]
This commit is contained in:
Vankadari, Meghana
2025-06-10 14:54:01 +05:30
committed by GitHub
parent 49ae7db89a
commit 8968973c2d
3 changed files with 36 additions and 25 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
addon/aocl_gemm/aocl_gemm_f32f32f32of32.c
View File

@@ -173,15 +173,11 @@ AOCL_GEMM_MATMUL(float,float,float,float,f32f32f32of32)
goto err_hndl;
}
// By default enable packing for B matrix. Before the 5 loop, based on
// the input dimensions, the smart threading logic will adjust it
// (disable/enable) accordingly.
if ( ( is_row_major == TRUE ) && ( mtag_b == UNPACKED ) )
if( ( is_row_major == TRUE ) && ( bli_is_trans(blis_transb ) ) && ( mtag_b == UNPACKED ) )
{
mtag_b = PACK;
}
// Inputs swapped in column major, A becomes B from kernel point of view.
else if ( ( is_column_major == TRUE ) && ( mtag_a == UNPACKED ) )
else if( ( is_column_major == TRUE ) && ( bli_is_trans(blis_transa ) ) && ( mtag_a == UNPACKED ) )
{
mtag_a = PACK;
}

35
addon/aocl_gemm/frame/f32f32f32/lpgemm_f32f32f32.c
View File

@@ -154,6 +154,12 @@ 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;
@@ -189,7 +195,7 @@ LPGEMV(float, float, float, f32f32f32of32)
packa_fp = packa_mr8_f32f32f32of32_col_major;
#endif
// Pack B matrix if rs_b > 1
if( ( mtag_b == PACK ) && ( rs_b != 1 ) )
if( (should_pack_B == TRUE) || ( rs_b != 1 ) )
{
mem_b_size_req = sizeof( float ) * k;
@@ -227,7 +233,7 @@ 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( mtag_a == PACK && cs_a != 1 )
if( (should_pack_A == TRUE) || ( cs_a != 1 ) )
{
mem_a_size_req = sizeof(float) * mc0 * k;
lpgemm_alloc_mem_panel
@@ -258,11 +264,11 @@ LPGEMV(float, float, float, f32f32f32of32)
&post_ops_attr
);
}
if ( ( mtag_a == PACK ) && ( bli_mem_is_alloc( &mem_a ) ) )
if ( ( (should_pack_A == TRUE) || ( cs_a != 1 ) ) && ( bli_mem_is_alloc( &mem_a ) ) )
{
bli_pba_release( rntm, &mem_a );
}
if ( ( mtag_b == PACK ) && ( bli_mem_is_alloc( &mem_b ) ) )
if ( ( (should_pack_B == TRUE) || ( rs_b != 1 ) ) && ( bli_mem_is_alloc( &mem_b ) ) )
{
bli_pba_release( rntm, &mem_b );
}
@@ -294,7 +300,7 @@ 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 ( mtag_a == PACK )
if ( (should_pack_A == TRUE) || ( cs_a != 1 ) )
{
mem_a_size_req = sizeof( float ) * k;
@@ -340,7 +346,7 @@ LPGEMV(float, float, float, f32f32f32of32)
rs_b_use = NR;
cs_b_use = 1;
}
else if (mtag_b == PACK)
else if ( (should_pack_B == TRUE) || ( mtag_b == PACK ) )
{
// nc0 needs to be a multiple of 16 since this gives maximum
// vectorization. Packing B always results in buffers with width
@@ -406,10 +412,15 @@ LPGEMV(float, float, float, f32f32f32of32)
} // jc loop
// Release pack buffers.
if ( ( mtag_b == PACK ) && ( bli_mem_is_alloc( &mem_b ) ) )
if ( ( (should_pack_B == TRUE) || ( 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 ) ) )
{
bli_pba_release( rntm, &mem_a );
}
}
}
@@ -472,7 +483,7 @@ LPGEMM_5LOOP(float, float, float, f32f32f32of32)
siz_t mem_a_size_req = 0;
// Check if packing of B is required.
bool should_pack_B = bli_rntm_pack_b( rntm ) || ( rs_b == 1 );
bool should_pack_B = bli_rntm_pack_b( rntm );
// Pack buffer for B.
float* pack_b_buffer_f32f32f32of32;
@@ -519,8 +530,8 @@ LPGEMM_5LOOP(float, float, float, f32f32f32of32)
bool invoke_rd = FALSE;
if( ( lpgemm_get_enabled_arch() != BLIS_ARCH_ZEN3) &&
( ( n < 48 ) || ( m < 16 ) ) && ( rs_b == 1 ) && ( mtag_b == PACK ) &&
( mtag_a == UNPACKED ) )
( ( n < 48 ) || ( m < 16 ) ) && ( rs_b == 1 ) &&
( mtag_a == UNPACKED ) && ( mtag_b == PACK ) )
{
invoke_rd = TRUE;
mtag_b = UNPACKED;
@@ -558,7 +569,7 @@ 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 == 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 );
@@ -744,7 +755,7 @@ LPGEMM_5LOOP(float, float, float, f32f32f32of32)
}
// Release pack buffers.
if ( ( mtag_b == PACK ) && ( should_pack_B == TRUE ) )
if ( ( mtag_b == PACK ) || ( should_pack_B == TRUE ) )
{
// All threads in work group should wait till B matrix usage is
// completed by the participating threads.

18
addon/aocl_gemm/frame/f32f32f32/lpgemm_f32f32f32_tiny.c
View File

@@ -141,6 +141,7 @@ LPGEMV_TINY(float, float, float, f32f32f32of32)
float* pack_b_buffer_f32f32f32of32 = NULL;
err_t err = BLIS_SUCCESS;
if(n == 1)
{
dim_t MR;
@@ -168,8 +169,9 @@ LPGEMV_TINY(float, float, float, f32f32f32of32)
ker_fp = lpgemv_n_one_f32f32f32of32_avx2;
packa_fp = packa_mr8_f32f32f32of32_col_major;
#endif
// Pack B matrix if rs_b > 1
if( ( mtag_b == PACK ) && ( rs_b != 1 ) )
// Pack B matrix if rs_b > 1, ignoring the mtag_b here.
// For tiny sizes, it is better to pack B if it affects output accuracy.
if( ( rs_b != 1 ) )
{
siz_t mem_b_size_req = sizeof( float ) * k;
pack_b_buffer_f32f32f32of32 =
@@ -185,7 +187,8 @@ LPGEMV_TINY(float, float, float, f32f32f32of32)
cs_b_use = 1;
}
if( ( mtag_a == PACK ) && ( cs_a != 1 ) )
// For tiny sizes, it is better to pack A if it affects output accuracy.
if( ( cs_a != 1 ) )
{
siz_t mem_a_size_req = sizeof(float) * m * k;
pack_a_buffer_f32f32f32of32 =
@@ -248,7 +251,8 @@ LPGEMV_TINY(float, float, float, f32f32f32of32)
#else
ker_fp = lpgemv_m_one_f32f32f32of32_avx2;
#endif
if( mtag_a == PACK && cs_a != 1)
// For tiny sizes, it is better to pack A if it affects output accuracy.
if( ( cs_a != 1 ) )
{
siz_t mem_a_size_req = sizeof( float ) * k;
pack_a_buffer_f32f32f32of32 =
@@ -262,7 +266,7 @@ LPGEMV_TINY(float, float, float, f32f32f32of32)
cs_a_use = 1;
}
if (mtag_b == PACK)
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;
@@ -324,7 +328,7 @@ LPGEMV_TINY(float, float, float, f32f32f32of32)
LPGEMM_TINY(float,float,float,f32f32f32of32)
{
// Handle using LPGEMV when m or/and n equal to 1
// Handle using LPGEMV when m or/and n equal to 1
if ( ( ( ( m == 1 ) || ( n == 1 ) ) ) && ( ( bli_cpuid_is_avx512_supported() == TRUE ) ||
( bli_cpuid_is_avx2fma3_supported() == TRUE ) ) )
{
@@ -384,7 +388,7 @@ 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 ) && ( rs_b == 1 ) )
if ( ( mtag_b == PACK ) )
{
dim_t nc0_updated = make_multiple_of_n( n, NR );
mem_b_size_req = sizeof( float ) * nc0_updated * k;