amd/blis
1
0
Fork 0
mirror of https://github.com/amd/blis.git synced 2026-04-19 23:28:52 +00:00
Code Issues Packages Projects Releases Wiki Activity

Add packing support M edge cases in DGEMM 24xk pack kernel

Browse Source 
Previously, the DGEMM implementation used `dscalv` for cases
where the M dimension of matrix A is not in multiple of 24,
resulting in a ~40% performance drop.

This commit introduces a specialized edge cases in pack kernel
to optimize performance for these cases.

The new packing support significantly improves the performance.

- Removed reliance on `dscalv` for edge cases, addressing the
  performance bottleneck.

AMD-Internal: [CPUPL-6677]

Change-Id: I150d13eb536d84f8eb439d7f4a77a04a0d0e6d60
This commit is contained in:
harsdave
2025-04-03 00:16:52 +05:30
parent 8557e2f7b9
commit cd83fc38b5
2 changed files with 10278 additions and 789 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10952
kernels/zen4/1m/bli_packm_zen4_asm_d24xk.c
View File

File diff suppressed because it is too large Load Diff

115
kernels/zen4/3/bli_gemm_tiny_adapter_zen4.c
View File

@@ -67,51 +67,52 @@
* @note
* N = 0 case never occurs.
*/
#define CALL_KERNEL\
if(N >= 8)\
{\
avx512kern_fp[8]( conja,\
conjb,\
M,\
N,\
K,\
(double *)alpha,\
(a_local + (0 * rs_a) + (0 * cs_a)), /*A matrix offset*/\
rs_a,\
cs_a,\
(b_local + (0 * cs_b) + (0 * rs_b)), /*B matrix offset*/\
rs_b,\
cs_b,\
(double *)beta,\
(c_local + 0 * cs_c + 0 * rs_c), /*C matrix offset*/\
rs_c,\
cs_c,\
&aux,\
NULL\
);\
}\
else\
{\
avx512kern_fp[N]( conja,\
conjb,\
M,\
N,\
K,\
(double *)alpha,\
(a_local + (0 * rs_a) + (0 * cs_a)), /*A matrix offset*/\
rs_a,\
cs_a,\
(b_local + (0 * cs_b) + (0 * rs_b)), /*B matrix offset*/\
rs_b,\
cs_b,\
(double *)beta,\
(c_local + 0 * cs_c + 0 * rs_c), /*C matrix offset*/\
rs_c,\
cs_c,\
&aux,\
NULL\
);\
#define CALL_KERNEL \
if(N >= 8) \
{ \
avx512kern_fp[8]( conja, \
conjb, \
M, \
N, \
K, \
(double *)alpha, \
(a_local + (0 * rs_a) + (0 * cs_a)), /*A matrix offset*/ \
rs_a, \
cs_a, \
(b_local + (0 * cs_b) + (0 * rs_b)), /*B matrix offset*/ \
rs_b, \
cs_b, \
(double *)beta, \
(c_local + 0 * cs_c + 0 * rs_c), /*C matrix offset*/ \
rs_c, \
cs_c, \
&aux, \
NULL \
); \
} \
else \
{ \
avx512kern_fp[N]( conja, \
conjb, \
M, \
N, \
K, \
(double *)alpha, \
(a_local + (0 * rs_a) + (0 * cs_a)), /*A matrix offset*/ \
rs_a, \
cs_a, \
(b_local + (0 * cs_b) + (0 * rs_b)), /*B matrix offset*/ \
rs_b, \
cs_b, \
(double *)beta, \
(c_local + 0 * cs_c + 0 * rs_c), /*C matrix offset*/ \
rs_c, \
cs_c, \
&aux, \
NULL \
); \
}
/**
* @brief bli_dgemmsup_placeholder
*
@@ -333,9 +334,33 @@ err_t bli_dgemm_tiny_24x8
ps_a_use = (24 * k);
bli_auxinfo_set_ps_a( ps_a_use, &aux );
/**
* CALL_KERNEL makes actual call to micro kernel,
* which is bli_dgemmsup_rv_zen4_asm_24x8m_new and the family of
* it based on value of N dimension.
* Arguments passed to it are as follows.
* conja whether A matrix is conjugate
conjb whether B matrix is conjugate
M M dimension
N N dimension
K K dimension
(double *)alpha Pointer to alpha value
(a_local + (0 * rs_a) + (0 * cs_a)), A matrix offset
rs_a row stride of A matrix
cs_a column stride of A matrix
(b_local + (0 * cs_b) + (0 * rs_b)), B matrix offset
rs_b row stride of B matrix
cs_b column stride of C matrix
(double *)beta pointer to Beta value
(c_local + 0 * cs_c + 0 * rs_c), C matrix offset
rs_c row stride of C matrix
cs_c column stride of C matrix
&aux Aux structure which carries additional info
NULL we do not use context in tiny path.
*/
CALL_KERNEL
//Return the allocated memory back to small block allocator
//Return the allocated memory back to small block allocator
bli_pba_release(&rntm, &local_mem_buf_A_s);
}
else