ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-04 13:41:24 +00:00
Code Issues Packages Projects Releases Wiki Activity

[CK_Tile] Enable PreshuffleB for 2d block scale Gemm (#3298)

Browse Source 
* formatted

* formatted

* formatting

* formatting

* formatting

* [CK TILE GEMM] Refactor block_scale_gemm examples

- Split cpp file to reduce building time
- Support multiple GemmConfig

* [CK TILE GEMM] Refactor block_scale_gemm examples

- Update Readme

* enable prefill shapes

* [CK TILE GEMM] Refactor block_scale_gemm examples

- Add support for rowcol and tensor GEMM operations

* [CK TILE GEMM] Refactor block_scale_gemm examples

- Update README

* adding preshuffle quant as new parameter and its associated new files

* remove debugging statements

* adding test

* enable preshuffle quant with permuteN

* updating readme and correcponding gemmconfigs

* updating cmake file

* fixing CI failures for grouped quant gemm

* debugging permuteN

* debugging

* debugging PermuteN

* initial commit

* resolving merge conflicts

* adding test cases

* fixing bq tensor calculation

---------

Co-authored-by: Cong Ma <congma13@amd.com>
Co-authored-by: Thomas Ning <Thomas.Ning@amd.com>
This commit is contained in:
Khushbu Agarwal
2025-12-05 09:57:52 -08:00
committed by GitHub
parent 608232ce82
commit 6b1bceca7b
7 changed files with 257 additions and 36 deletions
Download Patch File Download Diff File Expand all files Collapse all files

44
test/ck_tile/gemm_block_scale/test_gemm_quant_bquant_preshuffle.cpp
View File

@@ -19,6 +19,12 @@ using PkInt4 = ck_tile::pk_int4_t;
using BQuantGrouped = std::integral_constant<ck_tile::QuantType, ck_tile::QuantType::BQuantGrouped>;
using GroupSize = ck_tile::QuantGroupShape<ck_tile::sequence<1, 1, 128>>;
// 2d block sizes for BQuant
using GroupSize2D8N = ck_tile::QuantGroupShape<ck_tile::sequence<1, 8, 128>>;
using GroupSize2D16N = ck_tile::QuantGroupShape<ck_tile::sequence<1, 16, 128>>;
using GroupSize2D32N = ck_tile::QuantGroupShape<ck_tile::sequence<1, 32, 128>>;
using GroupSize2D64N = ck_tile::QuantGroupShape<ck_tile::sequence<1, 64, 128>>;
// Type combinations for BQuant tests with PreshuffleB
// Tuple format: <ALayout, BLayout, CLayout, AQLayout, ADataType, BDataType, QDataType, CDataType,
// QuantType, GemmConfig, QuantGroupSize>
@@ -37,7 +43,43 @@ using BPreshuffleBQuantTypes = ::testing::Types<
std::tuple<RowMajor, ColumnMajor, RowMajor, RowMajor, FP8, FP8, float, Half, BQuantGrouped, GemmConfigPreshuffleBPrefillTiledPermuteN, GroupSize>,
std::tuple<RowMajor, ColumnMajor, RowMajor, RowMajor, BF8, BF8, float, Half, BQuantGrouped, GemmConfigPreshuffleBPrefillTiledPermuteN, GroupSize>,
std::tuple<RowMajor, ColumnMajor, RowMajor, RowMajor, FP8, PkInt4, FP8, Half, BQuantGrouped, GemmConfigPreshuffleBPrefillTiledPermuteN, GroupSize>,
std::tuple<RowMajor, ColumnMajor, RowMajor, RowMajor, BF8, PkInt4, BF8, Half, BQuantGrouped, GemmConfigPreshuffleBPrefillTiledPermuteN, GroupSize>
std::tuple<RowMajor, ColumnMajor, RowMajor, RowMajor, BF8, PkInt4, BF8, Half, BQuantGrouped, GemmConfigPreshuffleBPrefillTiledPermuteN, GroupSize>,
// //2d cases with preshuffle B
std::tuple<RowMajor, ColumnMajor, RowMajor, RowMajor, FP8, FP8, float, Half, BQuantGrouped, GemmConfigPreshuffleBDecode, GroupSize2D8N>,
std::tuple<RowMajor, ColumnMajor, RowMajor, RowMajor, BF8, BF8, float, Half, BQuantGrouped, GemmConfigPreshuffleBDecode, GroupSize2D8N>,
std::tuple<RowMajor, ColumnMajor, RowMajor, RowMajor, FP8, PkInt4, FP8, Half, BQuantGrouped, GemmConfigPreshuffleBDecode, GroupSize2D8N>,
std::tuple<RowMajor, ColumnMajor, RowMajor, RowMajor, BF8, PkInt4, BF8, Half, BQuantGrouped, GemmConfigPreshuffleBDecode, GroupSize2D8N>,
std::tuple<RowMajor, ColumnMajor, RowMajor, RowMajor, FP8, FP8, float, Half, BQuantGrouped, GemmConfigPreshuffleBDecode, GroupSize2D16N>,
std::tuple<RowMajor, ColumnMajor, RowMajor, RowMajor, BF8, BF8, float, Half, BQuantGrouped, GemmConfigPreshuffleBDecode, GroupSize2D16N>,
std::tuple<RowMajor, ColumnMajor, RowMajor, RowMajor, FP8, PkInt4, FP8, Half, BQuantGrouped, GemmConfigPreshuffleBDecode, GroupSize2D16N>,
std::tuple<RowMajor, ColumnMajor, RowMajor, RowMajor, BF8, PkInt4, BF8, Half, BQuantGrouped, GemmConfigPreshuffleBDecode, GroupSize2D16N>,
std::tuple<RowMajor, ColumnMajor, RowMajor, RowMajor, FP8, FP8, float, Half, BQuantGrouped, GemmConfigPreshuffleBDecode, GroupSize2D32N>,
std::tuple<RowMajor, ColumnMajor, RowMajor, RowMajor, BF8, BF8, float, Half, BQuantGrouped, GemmConfigPreshuffleBDecode, GroupSize2D32N>,
std::tuple<RowMajor, ColumnMajor, RowMajor, RowMajor, FP8, PkInt4, FP8, Half, BQuantGrouped, GemmConfigPreshuffleBDecode, GroupSize2D32N>,
std::tuple<RowMajor, ColumnMajor, RowMajor, RowMajor, BF8, PkInt4, BF8, Half, BQuantGrouped, GemmConfigPreshuffleBDecode, GroupSize2D32N>,
std::tuple<RowMajor, ColumnMajor, RowMajor, RowMajor, FP8, FP8, float, Half, BQuantGrouped, GemmConfigPreshuffleBDecode, GroupSize2D64N>,
std::tuple<RowMajor, ColumnMajor, RowMajor, RowMajor, BF8, BF8, float, Half, BQuantGrouped, GemmConfigPreshuffleBDecode, GroupSize2D64N>,
std::tuple<RowMajor, ColumnMajor, RowMajor, RowMajor, FP8, PkInt4, FP8, Half, BQuantGrouped, GemmConfigPreshuffleBDecode, GroupSize2D64N>,
std::tuple<RowMajor, ColumnMajor, RowMajor, RowMajor, BF8, PkInt4, BF8, Half, BQuantGrouped, GemmConfigPreshuffleBDecode, GroupSize2D64N>,
std::tuple<RowMajor, ColumnMajor, RowMajor, RowMajor, FP8, FP8, float, Half, BQuantGrouped, GemmConfigPreshuffleBPrefill, GroupSize2D8N>,
std::tuple<RowMajor, ColumnMajor, RowMajor, RowMajor, BF8, BF8, float, Half, BQuantGrouped, GemmConfigPreshuffleBPrefill, GroupSize2D8N>,
std::tuple<RowMajor, ColumnMajor, RowMajor, RowMajor, FP8, PkInt4, FP8, Half, BQuantGrouped, GemmConfigPreshuffleBPrefill, GroupSize2D8N>,
std::tuple<RowMajor, ColumnMajor, RowMajor, RowMajor, BF8, PkInt4, BF8, Half, BQuantGrouped, GemmConfigPreshuffleBPrefill, GroupSize2D8N>,
std::tuple<RowMajor, ColumnMajor, RowMajor, RowMajor, FP8, FP8, float, Half, BQuantGrouped, GemmConfigPreshuffleBPrefill, GroupSize2D16N>,
std::tuple<RowMajor, ColumnMajor, RowMajor, RowMajor, BF8, BF8, float, Half, BQuantGrouped, GemmConfigPreshuffleBPrefill, GroupSize2D16N>,
std::tuple<RowMajor, ColumnMajor, RowMajor, RowMajor, FP8, PkInt4, FP8, Half, BQuantGrouped, GemmConfigPreshuffleBPrefill, GroupSize2D16N>,
std::tuple<RowMajor, ColumnMajor, RowMajor, RowMajor, BF8, PkInt4, BF8, Half, BQuantGrouped, GemmConfigPreshuffleBPrefill, GroupSize2D16N>,
std::tuple<RowMajor, ColumnMajor, RowMajor, RowMajor, FP8, FP8, float, Half, BQuantGrouped, GemmConfigPreshuffleBPrefill, GroupSize2D32N>,
std::tuple<RowMajor, ColumnMajor, RowMajor, RowMajor, BF8, BF8, float, Half, BQuantGrouped, GemmConfigPreshuffleBPrefill, GroupSize2D32N>,
std::tuple<RowMajor, ColumnMajor, RowMajor, RowMajor, FP8, PkInt4, FP8, Half, BQuantGrouped, GemmConfigPreshuffleBPrefill, GroupSize2D32N>,
std::tuple<RowMajor, ColumnMajor, RowMajor, RowMajor, BF8, PkInt4, BF8, Half, BQuantGrouped, GemmConfigPreshuffleBPrefill, GroupSize2D32N>,
std::tuple<RowMajor, ColumnMajor, RowMajor, RowMajor, FP8, FP8, float, Half, BQuantGrouped, GemmConfigPreshuffleBPrefill, GroupSize2D64N>,
std::tuple<RowMajor, ColumnMajor, RowMajor, RowMajor, BF8, BF8, float, Half, BQuantGrouped, GemmConfigPreshuffleBPrefill, GroupSize2D64N>,
std::tuple<RowMajor, ColumnMajor, RowMajor, RowMajor, FP8, PkInt4, FP8, Half, BQuantGrouped, GemmConfigPreshuffleBPrefill, GroupSize2D64N>,
std::tuple<RowMajor, ColumnMajor, RowMajor, RowMajor, BF8, PkInt4, BF8, Half, BQuantGrouped, GemmConfigPreshuffleBPrefill, GroupSize2D64N>
>;
// clang-format on

6
test/ck_tile/gemm_block_scale/test_gemm_quant_fixtures.hpp
View File

@@ -433,7 +433,7 @@ class TestCkTileGemmBQuant : public TestCkTileGemmQuantBase<Tuple, TestCkTileGem
ck_tile::HostTensor<BDataType> b_k_n_dev = b_k_n;
if constexpr(PreshuffleB)
{
if constexpr(TiledMMAPermuteN)
if constexpr(TiledMMAPermuteN && QuantGroupSize::kN == 1)
{
printf("PreshuffleB with TiledMMAPermuteN\n");
b_k_n_dev = ck_tile::shuffle_b_permuteN<GemmConfig>(b_k_n);
@@ -451,11 +451,11 @@ class TestCkTileGemmBQuant : public TestCkTileGemmQuantBase<Tuple, TestCkTileGem
b_k_n_dev_buf.ToDevice(b_k_n_dev.data());
if constexpr(PreshuffleB && TiledMMAPermuteN)
if constexpr(PreshuffleB && TiledMMAPermuteN && QuantGroupSize::kN == 1)
{
printf("Preshuffle BQ with TiledMMAPermuteN \n");
ck_tile::HostTensor<QDataType> bq_shuffle_host =
ck_tile::bq_permuteN<GemmConfig>(bq_bqk_bqn);
ck_tile::bq_permuteN<GemmConfig>(bq_bqk_bqn, QuantGroupSize::kN);
bq_bqk_bqn_dev_buf.ToDevice(bq_shuffle_host.data());
}
else if constexpr(GemmConfig::PreshuffleQuant)