ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-04-20 06:49:15 +00:00
Code Issues Packages Projects Releases Wiki Activity

[CK_TILE] ABQuant New Preshuffle (#3638)

Browse Source 
* Refactor

* Gemm quant improvement

* Change preshuffle

* Fix

* Fix grouped gemm ut

* Fix

---------

Co-authored-by: Thomas Ning <Thomas.Ning@amd.com>
This commit is contained in:
Yi DING
2026-01-28 15:46:49 +08:00
committed by GitHub
parent 91e32f305f
commit 8e3d84aba3
32 changed files with 182 additions and 213 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
include/ck_tile/ops/gemm_quant/kernel/gemm_quant_kernel.hpp
View File

@@ -693,13 +693,13 @@ struct QuantGemmKernel
{
if constexpr(PreshuffleB)
{
index_t kFlatK =
GemmPipeline::flatKPerWarp *
(k_size / GemmPipeline::BlockGemmShape::WarpTile::at(number<2>{}));
index_t kFlatN = kargs.N * kargs.K / kFlatK;
constexpr auto warp_k = GemmPipeline::BlockGemmShape::WarpTile::at(I2);
index_t kFlatKSplit = GemmPipeline::flatKPerWarp * (k_size / warp_k);
index_t kFlatK = GemmPipeline::flatKPerWarp * (kargs.K / warp_k);
index_t kFlatN = kargs.N * kargs.K / kFlatK;
return make_naive_tensor_view<address_space_enum::global>(
b_ptr,
make_tuple(kFlatN, kFlatK),
make_tuple(kFlatN, kFlatKSplit),
make_tuple(kFlatK, 1),
number<GemmPipeline::GetVectorSizeB()>{},
number<1>{});