ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-02 04:31:25 +00:00
Code Issues Packages Projects Releases Wiki Activity

[CK_TILE] Refine fp8 support in flatmm (#2239)

Browse Source 
* [CK_TILE] Refine fp8 in flatmm

1. Replace USING_MFMA_16x16x32 & USING_MFMA_16x16x32 with constexpr
2. Add an additional const check to avoid build error in HotLoopScheduler
3. Refine shuffleb to support both tile 32x32 and 16x16
4. Support command option -init
5. Move Gemm warp defintion to a separate struct

* fix clang format

* fix clang format

* keep default bhavior unchanged (warp tile = 16x16)

* fix tile engine build error

* fix a typo in codegen_utils.py

* address review comments

* address review comments

---------

Co-authored-by: Thomas Ning <Thomas.Ning@amd.com>
This commit is contained in:
linqunAMD
2025-06-25 16:07:45 +08:00
committed by GitHub
parent 50fad03524
commit 37e1a27537
10 changed files with 313 additions and 198 deletions
Download Patch File Download Diff File Expand all files Collapse all files

44
example/ck_tile/18_flatmm/flatmm_basic.cpp
View File

@@ -17,12 +17,12 @@ template <typename ADataType,
typename BDataType,
typename AccDataType,
typename CDataType,
typename FlatmmConfig,
typename ALayout,
typename BLayout,
typename CLayout>
float flatmm_calc(const ck_tile::FlatmmHostArgs& args, const ck_tile::stream_config& s)
{
using FlatmmConfig = FlatmmConfig<ADataType>;
using CodegenFlatmmShape = ck_tile::TileFlatmmShape<
ck_tile::sequence<FlatmmConfig::M_Tile, FlatmmConfig::N_Tile, FlatmmConfig::K_Tile>,
ck_tile::sequence<FlatmmConfig::M_Warp, FlatmmConfig::N_Warp, FlatmmConfig::K_Warp>,
@@ -32,18 +32,20 @@ float flatmm_calc(const ck_tile::FlatmmHostArgs& args, const ck_tile::stream_con
using TilePartitioner = ck_tile::GemmTile1DPartitioner<CodegenFlatmmShape>;
using CodegenGemmTraits = ck_tile::TileGemmTraits<FlatmmConfig::kPadM,
using CodegenGemmTraits = ck_tile::TileGemmTraits<FlatmmConfig::kPadM,
FlatmmConfig::kPadN,
FlatmmConfig::kPadK,
ALayout,
BLayout,
CLayout>;
using CodegenPipelineProblem = ck_tile::GemmPipelineProblem<ADataType,
BDataType,
AccDataType,
CodegenFlatmmShape,
CodegenGemmTraits>;
const auto Run = [&](const auto memory_operation_) {
const auto Run = [&](const auto memory_operation_) {
constexpr auto memory_operation = memory_operation_.value;
using GemmEpilogue = ck_tile::CShuffleEpilogue<
@@ -151,6 +153,7 @@ float flatmm_calc(const ck_tile::FlatmmHostArgs& args, const ck_tile::stream_con
}
}
template <template <typename PreType> typename FlatmmConfig>
int run_flatmm_example(int argc, char* argv[])
{
auto [result, arg_parser] = create_args(argc, argv);
@@ -163,24 +166,27 @@ int run_flatmm_example(int argc, char* argv[])
std::string data_type = arg_parser.get_str("prec");
std::string a_layout = arg_parser.get_str("a_layout");
std::string b_layout = arg_parser.get_str("b_layout");
if(a_layout == "R" && b_layout == "C")
{
if(data_type == "fp16")
{
run_flatmm_example_with_layouts<ck_tile::half_t>(argc, argv, Row{}, Col{}, Row{});
run_flatmm_example_with_layouts<ck_tile::half_t, FlatmmConfig<ck_tile::half_t>>(
argc, argv, Row{}, Col{}, Row{});
}
else if(data_type == "bf16")
{
run_flatmm_example_with_layouts<ck_tile::bf16_t>(argc, argv, Row{}, Col{}, Row{});
run_flatmm_example_with_layouts<ck_tile::bf16_t, FlatmmConfig<ck_tile::bf16_t>>(
argc, argv, Row{}, Col{}, Row{});
}
else if(data_type == "fp8")
{
run_flatmm_example_with_layouts<ck_tile::fp8_t>(argc, argv, Row{}, Col{}, Row{});
run_flatmm_example_with_layouts<ck_tile::fp8_t, FlatmmConfig<ck_tile::fp8_t>>(
argc, argv, Row{}, Col{}, Row{});
}
else if(data_type == "bf8")
{
run_flatmm_example_with_layouts<ck_tile::bf8_t>(argc, argv, Row{}, Col{}, Row{});
run_flatmm_example_with_layouts<ck_tile::bf8_t, FlatmmConfig<ck_tile::bf8_t>>(
argc, argv, Row{}, Col{}, Row{});
}
else
{
@@ -196,9 +202,29 @@ int run_flatmm_example(int argc, char* argv[])
int main(int argc, char* argv[])
{
auto [result, arg_parser] = create_args(argc, argv);
if(!result)
return EXIT_FAILURE;
try
{
return !run_flatmm_example(argc, argv);
int warp_tile = arg_parser.get_int("warp_tile");
if(warp_tile == 0)
{
return !run_flatmm_example<FlatmmConfig16>(argc, argv);
}
else if(warp_tile == 1)
{
return !run_flatmm_example<FlatmmConfig32>(argc, argv);
}
else if(warp_tile == 2)
{
return !run_flatmm_example<FlatmmConfig16_950>(argc, argv);
}
else
{
return !run_flatmm_example<FlatmmConfig32_950>(argc, argv);
}
}
catch(const std::runtime_error& e)
{