mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-03-26 18:17:40 +00:00
1a4aa7fd89
ck_tile: add gtest unit tests for MX flatmm (gfx950)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
## Summary
- Add correctness unit tests for the MX-format flatmm kernel
(`example/ck_tile/18_flatmm/mxgemm`) under `test/ck_tile/flatmm/`
- Tests cover all five dtype combinations: FP4×FP4, FP8×FP8, FP6×FP6,
FP8×FP4, FP4×FP8
- Tests cover all four kernel dispatch paths (the `has_hot_loop` ×
`tail_num` product):
- `has_hot_loop=false, tail=ODD` (K=256, num_loop=1)
- `has_hot_loop=false, tail=EVEN` (K=512, num_loop=2)
- `has_hot_loop=true, tail=ODD` (K=768, num_loop=3)
- `has_hot_loop=true, tail=EVEN` (K=1024, num_loop=4)
- Remove unsupported `-split_k` CLI option from
`tile_example_mx_flatmm`; the pre-shuffled B layout is incompatible with
K-splitting and the option silently produced wrong results
## Changes
**New files (`test/ck_tile/flatmm/`):**
- `CMakeLists.txt` — builds 40 kernel instances as a shared OBJECT
library, links into 5 per-dtype test executables; forwards
`-DCK_TILE_USE_OCP_FP8` when `CK_USE_OCP_FP8` is ON
- `test_mx_flatmm_base.hpp` — base test fixture with
`run_test_with_validation(M, N, K, kbatch=1)`
- `test_mx_flatmm_fixtures.hpp` — concrete `TestMXFlatmm` typed test
class and type aliases
- `test_mx_flatmm_fp{4fp4,8fp8,6fp6,8fp4,4fp8}.cpp` — per-dtype
`TYPED_TEST_SUITE` files
**Modified files:**
- `example/ck_tile/18_flatmm/mxgemm/mx_flatmm_arch_traits.hpp` — moved
`preShuffleWeight` here (was in `mx_flatmm.cpp`) so it is includeable by
both the example and the tests
- `example/ck_tile/18_flatmm/mxgemm/mx_flatmm.cpp` / `run_mx_flatmm.inc`
— removed `-split_k` CLI arg, hardcoded `k_batch=1`, fixed `k_split`
formula, updated call sites after `preShuffleWeight` move
- `test/ck_tile/CMakeLists.txt` — added `add_subdirectory(flatmm)`
279 lines
12 KiB
C++
279 lines
12 KiB
C++
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
|
|
// SPDX-License-Identifier: MIT
|
|
|
|
#include <hip/hip_runtime.h>
|
|
|
|
#include <cstring>
|
|
#include <iostream>
|
|
#include <ostream>
|
|
#include <string>
|
|
#include <tuple>
|
|
#include <type_traits>
|
|
|
|
#include "ck_tile/host.hpp"
|
|
#include "mx_flatmm.hpp"
|
|
|
|
template <typename Layout>
|
|
static constexpr inline auto is_row_major(Layout layout_)
|
|
{
|
|
return ck_tile::bool_constant<std::is_same_v<ck_tile::remove_cvref_t<decltype(layout_)>,
|
|
ck_tile::tensor_layout::gemm::RowMajor>>{};
|
|
}
|
|
|
|
template <typename MXFlatmmArchTraits,
|
|
typename ADataType,
|
|
typename BDataType,
|
|
typename DsDatatype,
|
|
typename AccDataType,
|
|
typename CDataType,
|
|
typename ALayout,
|
|
typename BLayout,
|
|
typename DsLayout,
|
|
typename CLayout,
|
|
typename ScaleA,
|
|
typename ScaleB,
|
|
bool UsePersistentKernel = false,
|
|
typename CDEElementWise = ck_tile::element_wise::PassThrough>
|
|
float invoke_mx_flatmm(ck_tile::DeviceMem& a_dev_buf,
|
|
ck_tile::DeviceMem& b_shuffle_dev_buf,
|
|
ck_tile::DeviceMem& c_dev_buf,
|
|
ck_tile::index_t M,
|
|
ck_tile::index_t N,
|
|
ck_tile::index_t K,
|
|
ck_tile::index_t stride_A,
|
|
ck_tile::index_t stride_B,
|
|
ck_tile::index_t stride_C,
|
|
ScaleA scale_a,
|
|
ScaleB scale_b,
|
|
int n_warmup,
|
|
int n_repeat)
|
|
{
|
|
using FlatmmConfig = typename MXFlatmmArchTraits::Config;
|
|
|
|
ck_tile::ScaleFlatmmHostArgs<ScaleA, ScaleB> args = {a_dev_buf.GetDeviceBuffer(),
|
|
b_shuffle_dev_buf.GetDeviceBuffer(),
|
|
{},
|
|
c_dev_buf.GetDeviceBuffer(),
|
|
1,
|
|
M,
|
|
N,
|
|
K,
|
|
stride_A,
|
|
stride_B,
|
|
{},
|
|
stride_C,
|
|
scale_a,
|
|
scale_b};
|
|
|
|
using FlatmmShape = ck_tile::TileGemmShape<
|
|
ck_tile::sequence<FlatmmConfig::M_Tile, FlatmmConfig::N_Tile, FlatmmConfig::K_Tile>,
|
|
ck_tile::sequence<FlatmmConfig::M_Warp, FlatmmConfig::N_Warp, FlatmmConfig::K_Warp>,
|
|
ck_tile::sequence<FlatmmConfig::M_Warp_Tile,
|
|
FlatmmConfig::N_Warp_Tile,
|
|
FlatmmConfig::K_Warp_Tile>>;
|
|
|
|
using TilePartitioner =
|
|
ck_tile::GemmSpatiallyLocalTilePartitioner<FlatmmShape,
|
|
FlatmmConfig::TileParitionerGroupNum,
|
|
FlatmmConfig::TileParitionerM01>;
|
|
|
|
using Traits = ck_tile::TileGemmTraits<FlatmmConfig::kPadM,
|
|
FlatmmConfig::kPadN,
|
|
FlatmmConfig::kPadK,
|
|
ALayout,
|
|
BLayout,
|
|
CLayout,
|
|
FlatmmConfig::NumWaveGroups>;
|
|
using GemmPipelineProblem =
|
|
ck_tile::GemmPipelineProblem<ADataType, BDataType, AccDataType, FlatmmShape, Traits>;
|
|
|
|
using BaseFlatmmPipeline = ck_tile::BaseFlatmmPipelineAGmemBGmemCRegV1<GemmPipelineProblem>;
|
|
|
|
const ck_tile::index_t k_grain = FlatmmConfig::K_Tile;
|
|
const ck_tile::index_t k_split = (K + k_grain - 1) / k_grain * k_grain;
|
|
const ck_tile::index_t num_loop = TilePartitioner::GetLoopNum(k_split);
|
|
const bool has_hot_loop = BaseFlatmmPipeline::BlockHasHotloop(num_loop);
|
|
const ck_tile::TailNumber tail_num = BaseFlatmmPipeline::GetBlockLoopTailNum(num_loop);
|
|
|
|
float ave_time = BaseFlatmmPipeline::template TailHandler<true>(
|
|
[&](auto has_hot_loop_, auto tail_num_) {
|
|
constexpr auto has_hot_loop_v = has_hot_loop_.value;
|
|
constexpr auto tail_num_v = tail_num_.value;
|
|
return mx_flatmm_calc<MXFlatmmArchTraits,
|
|
ADataType,
|
|
BDataType,
|
|
DsDatatype,
|
|
AccDataType,
|
|
CDataType,
|
|
ALayout,
|
|
BLayout,
|
|
DsLayout,
|
|
CLayout,
|
|
ScaleA,
|
|
ScaleB,
|
|
UsePersistentKernel,
|
|
CDEElementWise,
|
|
false,
|
|
has_hot_loop_v,
|
|
tail_num_v>(
|
|
args, ck_tile::stream_config{nullptr, true, 1, n_warmup, n_repeat, true, true, 50});
|
|
},
|
|
has_hot_loop,
|
|
tail_num);
|
|
|
|
constexpr int APackedSize = ck_tile::numeric_traits<ADataType>::PackedSize;
|
|
constexpr int BPackedSize = ck_tile::numeric_traits<BDataType>::PackedSize;
|
|
|
|
std::size_t flop = std::size_t(2) * M * N * K + std::size_t(2) * M * N * K / 32;
|
|
std::size_t num_byte = sizeof(ADataType) * M * K / APackedSize +
|
|
sizeof(BDataType) * N * K / BPackedSize + sizeof(CDataType) * M * N +
|
|
sizeof(ck_tile::e8m0_t) * M * K / 32 +
|
|
sizeof(ck_tile::e8m0_t) * N * K / 32;
|
|
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
|
|
float gb_per_sec = num_byte / 1.E6 / ave_time;
|
|
|
|
std::cout << "Run " << ck_tile::gemm_prec_str<ADataType, BDataType>() << " Flatmm kernel " //
|
|
<< " M = " << M << " N = " << N << " K = " << K << " StrideA = " << stride_A
|
|
<< " StrideB = " << stride_B << " StrideC = " << stride_C << " : " << ave_time
|
|
<< " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, " << std::endl;
|
|
|
|
return ave_time;
|
|
}
|
|
|
|
auto create_args(int argc, char* argv[])
|
|
{
|
|
ck_tile::ArgParser arg_parser;
|
|
arg_parser.insert("m", "32", "m dimension")
|
|
.insert("n", "512", "n dimension")
|
|
.insert("k", "256", "k dimension")
|
|
.insert("a_layout", "R", "A tensor data layout - Row by default")
|
|
.insert("b_layout", "C", "B tensor data layout - Row by default")
|
|
.insert("c_layout", "R", "C tensor data layout - Row by default")
|
|
.insert("stride_a", "0", "Tensor A stride")
|
|
.insert("stride_b", "0", "Tensor B stride")
|
|
.insert("stride_c", "0", "Tensor C stride")
|
|
.insert("v", "1", "0. No validation, 1. Validation on CPU, 2. Validation on GPU")
|
|
.insert("mx_prec",
|
|
"fp4xfp4",
|
|
"data type for activation and weight, support: fp4xfp4, fp6xfp6, fp8xfp8, fp8xfp4 "
|
|
"and fp4xfp8")
|
|
.insert("warmup", "50", "number of iterations before benchmark the kernel")
|
|
.insert("repeat", "100", "number of iterations to benchmark the kernel")
|
|
.insert("timer", "gpu", "gpu:gpu timer, cpu:cpu timer")
|
|
.insert("init", "0", "0:random, 1:constant(1)")
|
|
.insert("persistent", "0", "0: no persistent, 1: persistent kernel")
|
|
.insert("warp_tile", "0", "0: 16x16x128 on gfx950.");
|
|
bool result = arg_parser.parse(argc, argv);
|
|
return std::make_tuple(result, arg_parser);
|
|
}
|
|
|
|
#include "run_mx_flatmm.inc"
|
|
|
|
int run_mx_flatmm_example(const ck_tile::ArgParser& arg_parser)
|
|
{
|
|
using Row = ck_tile::tensor_layout::gemm::RowMajor;
|
|
using Col = ck_tile::tensor_layout::gemm::ColumnMajor;
|
|
|
|
std::string mx_prec = arg_parser.get_str("mx_prec");
|
|
std::string a_layout = arg_parser.get_str("a_layout");
|
|
std::string b_layout = arg_parser.get_str("b_layout");
|
|
int persistent_opt = arg_parser.get_int("persistent");
|
|
|
|
std::cout << "Using default warptile of 16x16x128." << std::endl;
|
|
|
|
if(a_layout == "R" && b_layout == "C")
|
|
{
|
|
if(mx_prec == "fp4" || mx_prec == "fp4xfp4")
|
|
{
|
|
if(persistent_opt == 0)
|
|
return run_mx_flatmm_with_layouts<ck_tile::pk_fp4_t,
|
|
ck_tile::pk_fp4_t,
|
|
ck_tile::fp16_t,
|
|
MXFlatmm_GFX950_FP4FP4_Traits,
|
|
false>(arg_parser, Row{}, Col{}, Row{});
|
|
else
|
|
throw std::runtime_error("Only non-persistent kernels are supported currently!");
|
|
}
|
|
else if(mx_prec == "fp6" || mx_prec == "fp6xfp6")
|
|
{
|
|
if(persistent_opt == 0)
|
|
return run_mx_flatmm_with_layouts<ck_tile::pk_fp6x16_t,
|
|
ck_tile::pk_fp6x16_t,
|
|
ck_tile::fp16_t,
|
|
MXFlatmm_GFX950_FP6FP6_Traits,
|
|
false>(arg_parser, Row{}, Col{}, Row{});
|
|
else
|
|
throw std::runtime_error("Only support non-persistent kernel now!");
|
|
}
|
|
else if(mx_prec == "fp8" || mx_prec == "fp8xfp8")
|
|
{
|
|
if(persistent_opt == 0)
|
|
return run_mx_flatmm_with_layouts<ck_tile::fp8_t,
|
|
ck_tile::fp8_t,
|
|
ck_tile::fp16_t,
|
|
MXFlatmm_GFX950_FP8FP8_Traits,
|
|
false>(arg_parser, Row{}, Col{}, Row{});
|
|
else
|
|
throw std::runtime_error("Only support non-persistent kernel now!");
|
|
}
|
|
else if(mx_prec == "fp8xfp4")
|
|
{
|
|
if(persistent_opt == 0)
|
|
return run_mx_flatmm_with_layouts<ck_tile::fp8_t,
|
|
ck_tile::pk_fp4_t,
|
|
ck_tile::fp16_t,
|
|
MXFlatmm_GFX950_FP8FP4_Traits,
|
|
false>(arg_parser, Row{}, Col{}, Row{});
|
|
else
|
|
throw std::runtime_error("Only support non-persistent kernel now!");
|
|
}
|
|
else if(mx_prec == "fp4xfp8")
|
|
{
|
|
if(persistent_opt == 0)
|
|
return run_mx_flatmm_with_layouts<ck_tile::pk_fp4_t,
|
|
ck_tile::fp8_t,
|
|
ck_tile::fp16_t,
|
|
MXFlatmm_GFX950_FP4FP8_Traits,
|
|
false>(arg_parser, Row{}, Col{}, Row{});
|
|
else
|
|
throw std::runtime_error("Only support non-persistent kernel now!");
|
|
}
|
|
else
|
|
{
|
|
throw std::runtime_error("Unsupported data_type!");
|
|
}
|
|
}
|
|
else
|
|
{
|
|
throw std::runtime_error("Unsupported data layout configuration for A,B and C tensors!");
|
|
}
|
|
}
|
|
|
|
int main(int argc, char* argv[])
|
|
{
|
|
auto [result, arg_parser] = create_args(argc, argv);
|
|
if(!result)
|
|
return EXIT_FAILURE;
|
|
try
|
|
{
|
|
int warp_tile = arg_parser.get_int("warp_tile");
|
|
if(warp_tile == 0)
|
|
{
|
|
return run_mx_flatmm_example(arg_parser);
|
|
}
|
|
else if(warp_tile == 1)
|
|
{
|
|
throw std::runtime_error("Only support MFMA_16x16x128 now!");
|
|
}
|
|
else
|
|
{
|
|
throw std::runtime_error("Unsupported warp_tile!");
|
|
}
|
|
}
|
|
catch(const std::runtime_error& e)
|
|
{
|
|
std::cerr << "Runtime error: " << e.what() << '\n';
|
|
return EXIT_FAILURE;
|
|
}
|
|
}
|