ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-11 00:40:09 +00:00
Code Issues Packages Projects Releases Wiki Activity

[Navi3x] Multiple issue fix (#612)

Browse Source 
* Change gridwise gemm mD blockwise gemm to naive

* RRR Gemm fix

* Fix RCR gemm bug

* Isolate wmma instructions

* Update amd_inline_asm.hpp

* Update amd_wmma.hpp

* Update amd_wmma.hpp

* fix syntax and update Jenkinsfile

---------

Co-authored-by: zjing14 <zhangjing14@gmail.com>
Co-authored-by: Illia Silin <98187287+illsilin@users.noreply.github.com>
Co-authored-by: illsilin <Illia.Silin@amd.com>
This commit is contained in:
Haocong WANG
2023-03-11 07:04:28 +08:00
committed by GitHub
parent 76fcdc60e9
commit 087e310589
6 changed files with 90 additions and 26 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
include/ck/utility/amd_inline_asm.hpp
View File

@@ -220,8 +220,8 @@ amd_assembly_outer_product_1x2(int8x4_t a, int8x4_t b0, int8x4_t b1, int32_t& c0
"0"(c0),
"1"(c1));
#else
c0 = __builtin_amdgcn_sdot4(bit_cast<int32_t>(a), bit_cast<int32_t>(b0), c0, false);
c1 = __builtin_amdgcn_sdot4(bit_cast<int32_t>(a), bit_cast<int32_t>(b1), c1, false);
c0 = __builtin_amdgcn_sdot4(bit_cast<int32_t>(a), bit_cast<int32_t>(b0), c0, false);
c1 = __builtin_amdgcn_sdot4(bit_cast<int32_t>(a), bit_cast<int32_t>(b1), c1, false);
#endif
}
@@ -257,10 +257,10 @@ __device__ void amd_assembly_outer_product_1x4(int8x4_t a,
"2"(c2),
"3"(c3));
#else
c0 = __builtin_amdgcn_sdot4(bit_cast<int32_t>(a), bit_cast<int32_t>(b0), c0, false);
c1 = __builtin_amdgcn_sdot4(bit_cast<int32_t>(a), bit_cast<int32_t>(b1), c1, false);
c2 = __builtin_amdgcn_sdot4(bit_cast<int32_t>(a), bit_cast<int32_t>(b2), c2, false);
c3 = __builtin_amdgcn_sdot4(bit_cast<int32_t>(a), bit_cast<int32_t>(b3), c3, false);
c0 = __builtin_amdgcn_sdot4(bit_cast<int32_t>(a), bit_cast<int32_t>(b0), c0, false);
c1 = __builtin_amdgcn_sdot4(bit_cast<int32_t>(a), bit_cast<int32_t>(b1), c1, false);
c2 = __builtin_amdgcn_sdot4(bit_cast<int32_t>(a), bit_cast<int32_t>(b2), c2, false);
c3 = __builtin_amdgcn_sdot4(bit_cast<int32_t>(a), bit_cast<int32_t>(b3), c3, false);
#endif
}
@@ -358,7 +358,13 @@ __device__ void amd_assembly_outer_product_1x4(int8x16_t a,
// Ranged input operand
__device__ void amd_assembly_wmma_f32_16x16x16_f16_w32(half16_t a, half16_t b, float8_t& c)
{
#if defined(__gfx11__)
asm volatile("v_wmma_f32_16x16x16_f16 %0, %1, %2, %0" : "=v"(c) : "v"(a), "v"(b), "0"(c));
#else
ignore = a;
ignore = b;
ignore = c;
#endif
}
} // namespace ck