mirror of
https://github.com/amd/blis.git
synced 2026-05-04 06:21:12 +00:00
3df4aac2d2
- A matrix by default isn't expected to be packed for a normal row-stored case. Hence the packing implementation is incomplete. - But if the user explicitly enables packing, interface wasn't handling the condition appropriately leading to data overwriting inside the incomplete pack kernels, thereby leading to accuracy failure. - As a fix, updated the interface to set the explicit PACK A to UNPACKED and proceed with GEMM in cases where transpose of A is not necessary. - Updated the batch gemm input file with additional test cases covering all the APIs. Bug Fixes: - Fixed implementation logic for column major inputs with post-ops to be disabled in S8 batch mat-mul. With the existing implementation, column-major inputs wouldn't be executed in case of of32/os32 inputs. - Fixed the Scale/ZP calculation in bench foru8s8s32ou8 condition, which was leading to accuracy failures. [AMD-Internal: CPUPL-7283 ]
29 lines
1007 B
Plaintext
29 lines
1007 B
Plaintext
f32f32f32of32:group_count=2
|
|
group_size=4
|
|
r n n n n 6 64 128 128 64 64 bias=bf16,relu,swish
|
|
group_size=3
|
|
r n t n n 78 9810 1229 1229 9810 9810 matrix_add=bf16,matrix_mul=f32
|
|
s8s8s32obf16:group_count=1
|
|
group_size=5
|
|
r n n n r 67 21 1823 1823 21 21 scale=vector,zp=scalar,relu,clip
|
|
f32f32f32of32:group_count=1
|
|
group_size=7
|
|
r n t n n 43 2240 1553 1553 1553 2240 scale=vector,zp=scalar,relu,clip
|
|
bf16bf16f32obf16:group_count=1
|
|
group_size=6
|
|
r n n n r 79 2676 1995 1995 2676 2676 bias=na,swish
|
|
bf16bf16f32of32:group_count=1
|
|
group_size=6
|
|
r t n n r 143 1943 730 143 1943 1943 bias=na,clip
|
|
bf16s4f32of32:group_count=1
|
|
group_size=6
|
|
r t n n r 79 1177 1968 79 1177 1177 scale=vector,zp=scalar,relu,clip
|
|
bf16s4f32obf16:group_count=1
|
|
group_size=6
|
|
r n n n r 17 2714 468 468 2714 2714 scale=vector,zp=vector,bias=na
|
|
s8s8s32obf16:group_count=1
|
|
group_size=4
|
|
r n n n n 43 2240 1553 1553 2240 2240 scale=vector,zp=scalar,relu,clip
|
|
*:group_count=1
|
|
group_size=3
|
|
r t t n n 92 1479 589 92 589 1479 scale=vector,zp=vector,bias=na,clip |