mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-04-19 22:39:03 +00:00
init ck_tile mxfp4 flatmm
This commit is contained in:
mtgu0705
478
example/ck_tile/18_flatmm/mxgemm/mx_flatmm.cpp
Normal file
478
example/ck_tile/18_flatmm/mxgemm/mx_flatmm.cpp
Normal file
@@ -0,0 +1,478 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#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_prec_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 FlatmmConfig,
|
||||
typename ADataType,
|
||||
typename BDataType,
|
||||
typename DsDatatype,
|
||||
typename AccDataType,
|
||||
typename CDataType,
|
||||
typename ALayout,
|
||||
typename BLayout,
|
||||
typename DsLayout,
|
||||
typename ELayout,
|
||||
typename ScaleM,
|
||||
typename ScaleN,
|
||||
bool persistent,
|
||||
typename CDEElementWise>
|
||||
float mx_flatmm_calc(const ck_tile::ScaleFlatmmHostArgs<ScaleM, ScaleN>& args,
|
||||
const ck_tile::stream_config& s)
|
||||
{
|
||||
using CodegenFlatmmShape = 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<CodegenFlatmmShape,
|
||||
FlatmmConfig::TileParitionerGroupNum,
|
||||
FlatmmConfig::TileParitionerM01>;
|
||||
|
||||
using Traits = ck_tile::TileGemmTraits<FlatmmConfig::kPadM,
|
||||
FlatmmConfig::kPadN,
|
||||
FlatmmConfig::kPadK,
|
||||
ALayout,
|
||||
BLayout,
|
||||
ELayout,
|
||||
FlatmmConfig::NumWaveGroups>;
|
||||
|
||||
using CodegenGemmTraits = ck_tile::TileGemmUniversalTraits<FlatmmConfig::kPadM,
|
||||
FlatmmConfig::kPadN,
|
||||
FlatmmConfig::kPadK,
|
||||
FlatmmConfig::DoubleSmemBuffer,
|
||||
ALayout,
|
||||
BLayout,
|
||||
ELayout,
|
||||
FlatmmConfig::TransposeC,
|
||||
FlatmmConfig::UseStructuredSparsity,
|
||||
persistent,
|
||||
FlatmmConfig::NumWaveGroups,
|
||||
true>;
|
||||
|
||||
using ComputeDataType = ADataType;
|
||||
static_assert(sizeof(ComputeDataType) >= sizeof(BDataType),
|
||||
"mixed_prec_flatmm requires ADataType is a wider type than BDataType");
|
||||
|
||||
using GemmPipelineProblem = ck_tile::GemmPipelineProblem<ComputeDataType,
|
||||
ComputeDataType,
|
||||
AccDataType,
|
||||
CodegenFlatmmShape,
|
||||
Traits>;
|
||||
|
||||
using BaseGemmPipeline = ck_tile::BaseFlatmmPipelineAGmemBGmemCRegV1<GemmPipelineProblem>;
|
||||
|
||||
const ck_tile::index_t k_grain = args.k_batch * FlatmmConfig::K_Tile;
|
||||
const ck_tile::index_t K_split = (args.K + k_grain - 1) / k_grain * FlatmmConfig::K_Tile;
|
||||
const ck_tile::index_t num_loop = TilePartitioner::GetLoopNum(K_split);
|
||||
const bool has_hot_loop = BaseGemmPipeline::BlockHasHotloop(num_loop);
|
||||
const ck_tile::TailNumber tail_num = BaseGemmPipeline::GetBlockLoopTailNum(num_loop);
|
||||
float ave_time{0};
|
||||
|
||||
const auto Run = [&](const auto has_hot_loop_,
|
||||
const auto tail_number_,
|
||||
const auto memory_operation_) {
|
||||
constexpr bool has_hot_loop_v = has_hot_loop_.value;
|
||||
constexpr auto tail_number_v = tail_number_.value;
|
||||
constexpr auto scheduler = FlatmmConfig::Scheduler;
|
||||
constexpr auto memory_operation = memory_operation_.value;
|
||||
|
||||
constexpr int BlockedXDLN_PerWarp = 2; // determined by scale shuffle pattern
|
||||
|
||||
using CodegenPipelineProblem = ck_tile::F16xMXF4FlatmmPipelineProblem<ADataType,
|
||||
BDataType,
|
||||
AccDataType,
|
||||
CodegenFlatmmShape,
|
||||
CodegenGemmTraits,
|
||||
scheduler,
|
||||
has_hot_loop_v,
|
||||
tail_number_v>;
|
||||
|
||||
using CodegenFlatmmPipeline =
|
||||
ck_tile::F16xMXF4FlatmmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem>;
|
||||
|
||||
using GemmEpilogue = ck_tile::CShuffleEpilogue<
|
||||
ck_tile::CShuffleEpilogueProblem<ComputeDataType,
|
||||
ComputeDataType,
|
||||
DsDatatype,
|
||||
AccDataType,
|
||||
CDataType,
|
||||
DsLayout,
|
||||
ELayout,
|
||||
CDEElementWise,
|
||||
CodegenPipelineProblem::kBlockSize,
|
||||
TilePartitioner::MPerBlock,
|
||||
TilePartitioner::NPerBlock,
|
||||
FlatmmConfig::M_Warp,
|
||||
FlatmmConfig::N_Warp,
|
||||
FlatmmConfig::M_Warp_Tile,
|
||||
FlatmmConfig::N_Warp_Tile,
|
||||
FlatmmConfig::K_Warp_Tile,
|
||||
CodegenPipelineProblem::TransposeC,
|
||||
memory_operation,
|
||||
FlatmmConfig::NumWaveGroups,
|
||||
false, // FixedVectorSize
|
||||
1, // VectorSizeC
|
||||
FlatmmConfig::TiledMMAPermuteN,
|
||||
BlockedXDLN_PerWarp>>;
|
||||
|
||||
using Kernel =
|
||||
ck_tile::F16xMXF4FlatmmKernel<TilePartitioner, CodegenFlatmmPipeline, GemmEpilogue>;
|
||||
|
||||
auto kargs = Kernel::MakeKernelArgs(args);
|
||||
|
||||
const dim3 grids = Kernel::GridSize(kargs);
|
||||
constexpr dim3 blocks = Kernel::BlockSize();
|
||||
|
||||
if(!Kernel::IsSupportedArgument(kargs))
|
||||
{
|
||||
throw std::runtime_error("Wrong! Arguments not supported! Skipping gemm!\n");
|
||||
}
|
||||
|
||||
if(s.log_level_ > 0)
|
||||
{
|
||||
std::cout << "Launching kernel with args:" << CodegenFlatmmShape::GetName() << "\n"
|
||||
<< "Shape: " << CodegenFlatmmShape::GetName() << "\n"
|
||||
<< "problem: " << CodegenPipelineProblem::GetName() << "\n"
|
||||
<< "pipeline: " << CodegenFlatmmPipeline::GetName() << "\n"
|
||||
<< "grid: {" << grids.x << ", " << grids.y << ", " << grids.z << "}"
|
||||
<< ", blocks: {" << blocks.x << ", " << blocks.y << ", " << blocks.z << "}"
|
||||
<< std::endl;
|
||||
}
|
||||
|
||||
if(s.flush_cache_)
|
||||
{
|
||||
std::cout << "Flushing cache..." << std::endl;
|
||||
constexpr ck_tile::index_t APackedSize = ck_tile::numeric_traits<ADataType>::PackedSize;
|
||||
constexpr ck_tile::index_t BPackedSize = ck_tile::numeric_traits<BDataType>::PackedSize;
|
||||
|
||||
ck_tile::HostTensor<ADataType> a_m(ck_tile::host_tensor_descriptor(
|
||||
args.M, args.K, args.stride_A, is_row_major(ALayout{})));
|
||||
ck_tile::HostTensor<BDataType> b_n(ck_tile::host_tensor_descriptor(
|
||||
args.K, args.N, args.stride_B, is_row_major(BLayout{})));
|
||||
|
||||
auto size_a_buffer = a_m.get_element_space_size_in_bytes() / APackedSize;
|
||||
auto size_b_buffer = b_n.get_element_space_size_in_bytes() / BPackedSize;
|
||||
|
||||
ck_tile::RotatingMemWrapper<ADataType, BDataType> rotating_mem(
|
||||
kargs.a_ptr, kargs.b_ptr, s.rotating_count_, size_a_buffer, size_b_buffer);
|
||||
rotating_mem.Print();
|
||||
|
||||
auto run_flush_cache = [&]() {
|
||||
// flush icache
|
||||
ck_tile::flush_icache();
|
||||
// rotating mem
|
||||
rotating_mem.Next();
|
||||
// clear c mem
|
||||
if(args.k_batch > 1)
|
||||
hipGetErrorString(hipMemsetAsync(
|
||||
args.e_ptr, 0, args.M * args.N * sizeof(CDataType), s.stream_id_));
|
||||
};
|
||||
ave_time = ck_tile::launch_kernel_preprocess(
|
||||
s,
|
||||
run_flush_cache,
|
||||
ck_tile::make_kernel<blocks.x, FlatmmConfig::kBlockPerCu>(
|
||||
Kernel{}, grids, blocks, 0, kargs));
|
||||
}
|
||||
else
|
||||
{
|
||||
// ave_time =
|
||||
// ck_tile::launch_kernel(s,
|
||||
// ck_tile::make_kernel<blocks.x, FlatmmConfig::kBlockPerCu>(
|
||||
// Kernel{}, grids, blocks, 0, kargs));
|
||||
}
|
||||
return ave_time;
|
||||
};
|
||||
|
||||
const auto RunSplitk = [&](const auto has_hot_loop_, const auto tail_number_) {
|
||||
if(args.k_batch == 1)
|
||||
{
|
||||
Run(has_hot_loop_,
|
||||
tail_number_,
|
||||
ck_tile::integral_constant<ck_tile::memory_operation_enum,
|
||||
ck_tile::memory_operation_enum::set>{});
|
||||
}
|
||||
else
|
||||
{
|
||||
Run(has_hot_loop_,
|
||||
tail_number_,
|
||||
ck_tile::integral_constant<ck_tile::memory_operation_enum,
|
||||
ck_tile::memory_operation_enum::atomic_add>{});
|
||||
}
|
||||
};
|
||||
BaseGemmPipeline::TailHandler(RunSplitk, has_hot_loop, tail_num);
|
||||
return ave_time;
|
||||
}
|
||||
|
||||
template <typename FlatmmConfig,
|
||||
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,
|
||||
ck_tile::index_t kbatch,
|
||||
ScaleA scale_a,
|
||||
ScaleB scale_b,
|
||||
int n_warmup,
|
||||
int n_repeat)
|
||||
{
|
||||
ck_tile::ScaleFlatmmHostArgs<ScaleA, ScaleB> args = {a_dev_buf.GetDeviceBuffer(),
|
||||
b_shuffle_dev_buf.GetDeviceBuffer(),
|
||||
{},
|
||||
c_dev_buf.GetDeviceBuffer(),
|
||||
kbatch,
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
stride_A,
|
||||
stride_B,
|
||||
{},
|
||||
stride_C,
|
||||
{},
|
||||
scale_a,
|
||||
scale_b};
|
||||
|
||||
float ave_time = mx_flatmm_calc<FlatmmConfig,
|
||||
ADataType,
|
||||
BDataType,
|
||||
DsDatatype,
|
||||
AccDataType,
|
||||
CDataType,
|
||||
ALayout,
|
||||
BLayout,
|
||||
DsLayout,
|
||||
CLayout,
|
||||
ScaleA,
|
||||
ScaleB,
|
||||
UsePersistentKernel,
|
||||
CDEElementWise>(
|
||||
args, ck_tile::stream_config{nullptr, true, 1, n_warmup, n_repeat, true, true, 50});
|
||||
|
||||
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;
|
||||
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
|
||||
float gb_per_sec = num_byte / 1.E6 / ave_time;
|
||||
|
||||
std::cout << "Run A16W4_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", "128", "n dimension")
|
||||
.insert("k", "512", "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: fp6xfp6, fp8xfp8")
|
||||
.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("split_k", "1", "splitK value")
|
||||
.insert("init", "0", "0:random, 1:constant(1)")
|
||||
.insert("persistent", "0", "0: no persistent, 1: persistent kernel")
|
||||
.insert("warp_tile",
|
||||
"0",
|
||||
"0: 16x16, 1: 32x32, 2: 16x16x128 (950 only), 3: 32x32x64 (950 only)");
|
||||
bool result = arg_parser.parse(argc, argv);
|
||||
return std::make_tuple(result, arg_parser);
|
||||
}
|
||||
|
||||
template <class FlatmmConfig, class IterSrc, class IterDst>
|
||||
void preShuffleWeight(const IterSrc src, IterDst dst, int N, int K)
|
||||
{
|
||||
int KPack = 16;
|
||||
int NLane = FlatmmConfig::N_Warp_Tile;
|
||||
int KLane = 64 / NLane;
|
||||
int K_pk = K / 2;
|
||||
int K0 = K_pk / (KLane * KPack);
|
||||
// K -> K0 KLane KPack
|
||||
// N -> N0 NLane
|
||||
// N, K -> N0 K0 KLane NLane KPack
|
||||
int tempk;
|
||||
for(int n = 0; n < N; ++n)
|
||||
{
|
||||
for(int k = 0; k < K_pk; ++k)
|
||||
{
|
||||
int n0 = n / NLane;
|
||||
int n1 = n % NLane;
|
||||
|
||||
int k0 = k / (KLane * KPack);
|
||||
tempk = k % (KLane * KPack);
|
||||
int k1 = tempk / KPack;
|
||||
int k2 = tempk % KPack;
|
||||
|
||||
int outputIndex = n0 * KPack * NLane * KLane * K0 + k0 * KPack * NLane * KLane +
|
||||
k1 * KPack * NLane + n1 * KPack + k2;
|
||||
|
||||
dst[outputIndex] = src[n * K_pk + k];
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
template <class FlatmmConfig, class T>
|
||||
auto preShuffleScale(const ck_tile::HostTensor<T>& scale)
|
||||
{
|
||||
assert(scale.get_lengths().size() == 2);
|
||||
int n_ = scale.get_lengths()[1];
|
||||
int k_ = scale.get_lengths()[0];
|
||||
|
||||
constexpr int K_Pack = 2; // fixed for mxfp4
|
||||
constexpr int N_Pack = 2; // fixed for mxfp4
|
||||
constexpr int GranularityK = 32; // fixed for mxfp4
|
||||
|
||||
constexpr int K_Lane = 64 / FlatmmConfig::N_Warp_Tile; // 4
|
||||
|
||||
static_assert(FlatmmConfig::N_Warp_Tile == 16, "only support XDL_N == 16");
|
||||
static_assert(FlatmmConfig::N_Repeat % N_Pack == 0);
|
||||
static_assert(FlatmmConfig::K_Tile % (K_Pack * K_Lane * GranularityK) == 0);
|
||||
|
||||
ck_tile::HostTensor<T> shfl_scale({
|
||||
k_ / K_Pack / K_Lane,
|
||||
K_Pack,
|
||||
K_Lane,
|
||||
n_ / FlatmmConfig::N_Warp_Tile / N_Pack,
|
||||
N_Pack,
|
||||
FlatmmConfig::N_Warp_Tile,
|
||||
});
|
||||
std::copy(scale.begin(), scale.end(), shfl_scale.begin());
|
||||
return ck_tile::reference_permute(shfl_scale, {3, 0, 2, 5, 1, 4});
|
||||
}
|
||||
|
||||
#include "run_mx_prec_flatmm.inc"
|
||||
|
||||
template <typename FlatmmConfig>
|
||||
int run_mx_flatmm_example(int argc, char* argv[])
|
||||
{
|
||||
auto [result, arg_parser] = create_args(argc, argv);
|
||||
if(!result)
|
||||
return -1;
|
||||
|
||||
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");
|
||||
|
||||
if(a_layout == "R" && b_layout == "C")
|
||||
{
|
||||
if(mx_prec == "fp4xfp4")
|
||||
{
|
||||
if(persistent_opt == 0)
|
||||
{
|
||||
run_mx_flatmm_with_layouts<ck_tile::pk_fp4_t,
|
||||
ck_tile::pk_fp4_t,
|
||||
ck_tile::fp16_t,
|
||||
FlatmmConfig,
|
||||
false>(argc, argv, Row{}, Col{}, Row{});
|
||||
}
|
||||
else
|
||||
{
|
||||
run_mx_flatmm_with_layouts<ck_tile::pk_fp4_t,
|
||||
ck_tile::pk_fp4_t,
|
||||
ck_tile::fp16_t,
|
||||
FlatmmConfig,
|
||||
true>(argc, argv, Row{}, Col{}, Row{});
|
||||
}
|
||||
}
|
||||
else if(mx_prec == "fp6xfp6")
|
||||
{
|
||||
throw std::runtime_error("Only support fp4xfp4 now!");
|
||||
}
|
||||
else if(mx_prec == "fp8xfp8")
|
||||
{
|
||||
throw std::runtime_error("Only support fp4xfp4 now!");
|
||||
}
|
||||
else
|
||||
{
|
||||
throw std::runtime_error("Unsupported data_type!");
|
||||
}
|
||||
}
|
||||
else
|
||||
{
|
||||
throw std::runtime_error("Unsupported data layout configuration for A,B and C tensors!");
|
||||
}
|
||||
return -1;
|
||||
}
|
||||
|
||||
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<MXfp4_FlatmmConfig16>(argc, argv);
|
||||
}
|
||||
else if(warp_tile == 1)
|
||||
{
|
||||
thow 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;
|
||||
}
|
||||
}
|
||||
40
example/ck_tile/18_flatmm/mxgemm/mx_flatmm.hpp
Normal file
40
example/ck_tile/18_flatmm/mxgemm/mx_flatmm.hpp
Normal file
@@ -0,0 +1,40 @@
|
||||
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "ck_tile/core.hpp"
|
||||
|
||||
// GEMM config with 16x16 warp tile
|
||||
struct MXfp4_FlatmmConfig16
|
||||
{
|
||||
static constexpr ck_tile::index_t M_Tile = 32;
|
||||
static constexpr ck_tile::index_t N_Tile = 128;
|
||||
static constexpr ck_tile::index_t K_Tile = 256;
|
||||
|
||||
static constexpr ck_tile::index_t M_Warp = 1;
|
||||
static constexpr ck_tile::index_t N_Warp = 4;
|
||||
static constexpr ck_tile::index_t K_Warp = 1;
|
||||
|
||||
static constexpr ck_tile::index_t M_Warp_Tile = 16;
|
||||
static constexpr ck_tile::index_t N_Warp_Tile = 16;
|
||||
static constexpr ck_tile::index_t K_Warp_Tile = 128;
|
||||
|
||||
static constexpr bool kPadM = false;
|
||||
static constexpr bool kPadN = false;
|
||||
static constexpr bool kPadK = false;
|
||||
|
||||
static constexpr bool TransposeC = false;
|
||||
static constexpr bool UseStructuredSparsity = false;
|
||||
|
||||
static constexpr int kBlockPerCu = 1;
|
||||
static constexpr int TileParitionerGroupNum = 8;
|
||||
static constexpr int TileParitionerM01 = 4;
|
||||
static constexpr auto Scheduler = ck_tile::GemmPipelineScheduler::Default;
|
||||
static constexpr ck_tile::index_t NumWaveGroups = 1;
|
||||
static constexpr bool DoubleSmemBuffer = false;
|
||||
|
||||
static constexpr int N_Repeat = N_Tile / N_Warp_Tile / N_Warp;
|
||||
static constexpr bool TiledMMAPermuteN = false;
|
||||
};
|
||||
301
example/ck_tile/18_flatmm/mxgemm/run_mx_flatmm.inc
Normal file
301
example/ck_tile/18_flatmm/mxgemm/run_mx_flatmm.inc
Normal file
@@ -0,0 +1,301 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
template <typename PrecActType,
|
||||
typename PrecWeightType,
|
||||
typename CDataType,
|
||||
typename FlatmmConfig,
|
||||
bool UsePersistentKernel = false,
|
||||
typename ALayout,
|
||||
typename BLayout,
|
||||
typename CLayout>
|
||||
int run_mx_flatmm_with_layouts(int argc,
|
||||
char* argv[],
|
||||
const ALayout a_layout = ALayout{},
|
||||
const BLayout b_layout = BLayout{},
|
||||
[[maybe_unused]] const CLayout c_layout = CLayout{})
|
||||
{
|
||||
auto [result, arg_parser] = create_args(argc, argv);
|
||||
if(!result)
|
||||
return -1;
|
||||
|
||||
using ADataType = PrecActType;
|
||||
using BDataType = PrecWeightType;
|
||||
using CDataType = CDataType;
|
||||
using AccDataType = float;
|
||||
|
||||
using ScaleType = ck_tile::e8m0_t;
|
||||
|
||||
constexpr int ScaleGranularityM = 1;
|
||||
constexpr int ScaleGranularityN = 1;
|
||||
constexpr int ScaleGranularityK = 32;
|
||||
|
||||
ck_tile::index_t M = arg_parser.get_int("m");
|
||||
ck_tile::index_t N = arg_parser.get_int("n");
|
||||
ck_tile::index_t K = arg_parser.get_int("k");
|
||||
|
||||
ck_tile::index_t stride_A = arg_parser.get_int("stride_a");
|
||||
ck_tile::index_t stride_B = arg_parser.get_int("stride_b");
|
||||
ck_tile::index_t stride_C = arg_parser.get_int("stride_c");
|
||||
|
||||
ck_tile::index_t kbatch = arg_parser.get_int("split_k");
|
||||
ck_tile::index_t init_method = arg_parser.get_int("init");
|
||||
ck_tile::index_t n_warmup = arg_parser.get_int("warmup");
|
||||
ck_tile::index_t n_repeat = arg_parser.get_int("repeat");
|
||||
|
||||
stride_A = ck_tile::get_default_stride(M, K, stride_A, is_row_major(a_layout));
|
||||
stride_B = ck_tile::get_default_stride(K, N, stride_B, is_row_major(b_layout));
|
||||
stride_C = ck_tile::get_default_stride(M, N, stride_C, is_row_major(CLayout{}));
|
||||
|
||||
auto scale_stride_A = ck_tile::get_default_stride(
|
||||
M / ScaleGranularityM, K / ScaleGranularityK, 0, is_row_major(a_layout));
|
||||
auto scale_stride_B = ck_tile::get_default_stride(
|
||||
K / ScaleGranularityK, N / ScaleGranularityN, 0, is_row_major(b_layout));
|
||||
|
||||
if(K % DequantGranularityK != 0)
|
||||
{
|
||||
thow std::runtime_error("wrong! K must be multiple of ScaleGranularityK.");
|
||||
}
|
||||
if(K % ck_tile::packed_size_v<ADataType> != 0 || K % ck_tile::packed_size_v<BDataType> != 0)
|
||||
{
|
||||
throw std::runtime_error("wrong! K must be multiple of packed size.");
|
||||
}
|
||||
|
||||
ck_tile ::HostTensor<ADataType> a_host(
|
||||
ck_tile::host_tensor_descriptor(M, K, stride_A, is_row_major(a_layout)));
|
||||
ck_tile::HostTensor<BDataType> b_origin_host(
|
||||
ck_tile::host_tensor_descriptor(K, N, stride_B, is_row_major(b_layout)));
|
||||
ck_tile::HostTensor<CDataType> c_rslt_host(
|
||||
ck_tile::host_tensor_descriptor(M, N, stride_C, is_row_major(CLayout{})));
|
||||
|
||||
ck_tile::HostTensor<ScaleType> scale_a(ck_tile::host_tensor_descriptor(
|
||||
M / ScaleGranularityM, K / ScaleGranularityK, scale_stride_A, is_row_major(a_layout)));
|
||||
ck_tile::HostTensor<ScaleType> scale_b(ck_tile::host_tensor_descriptor(
|
||||
K / ScaleGranularityK, N / ScaleGranularityN, scale_stride_B, is_row_major(b_layout)));
|
||||
|
||||
if(init_method == 0)
|
||||
{
|
||||
ck_tile::FillUniformDistribution<ADataType>{0.0f, 1.0f}(a_host);
|
||||
ck_tile::FillUniformDistribution<BDataType>{-.5f, .5f}(b_origin_host);
|
||||
ck_tile::FillUniformDistribution<ScaleType>{-2.f, 2.f}(scale_a);
|
||||
ck_tile::FillUniformDistribution<ScaleType>{-2.f, 2.f}(scale_b);
|
||||
}
|
||||
else if(init_method == 1)
|
||||
{
|
||||
ck_tile::FillUniformDistribution<ADataType>{1.f, 1.f}(a_host);
|
||||
ck_tile::FillUniformDistribution<BDataType>{1.f, 1.f}(b_origin_host);
|
||||
ck_tile::FillUniformDistribution<ScaleType>{1.f, 1.f}(scale_a);
|
||||
ck_tile::FillUniformDistribution<ScaleType>{1.f, 1.f}(scale_b);
|
||||
}
|
||||
else if(init_method == 2)
|
||||
{
|
||||
ck_tile::FillUniformDistribution<ADataType>{0.0f, 1.0f}(a_host);
|
||||
ck_tile::FillUniformDistribution<BDataType>{1.f, 1.f}(b_origin_host);
|
||||
ck_tile::FillUniformDistribution<ScaleType>{-2.f, 2.f}(scale_a);
|
||||
ck_tile::FillUniformDistribution<ScaleType>{-2.f, 2.f}(scale_b);
|
||||
}
|
||||
|
||||
#if 0
|
||||
#if 1
|
||||
printf("printf a_host: \n");
|
||||
for(int m = 0; m < M; m++)
|
||||
{
|
||||
for(int k = 0; k < K; k++)
|
||||
{
|
||||
printf("%.2f ", ck_tile::type_convert<float>(a_host(m, k)));
|
||||
}
|
||||
printf("\n");
|
||||
}
|
||||
printf("\n");
|
||||
|
||||
printf("printf b_origin_host: \n");
|
||||
#if 1
|
||||
for(int n = 0; n < N; n++)
|
||||
{
|
||||
for(int k = 0; k < K; k++)
|
||||
{
|
||||
auto b_f4x2 = b_origin_host(k, n);
|
||||
if(k % 2 == 0)
|
||||
printf("%.2f ", ck_tile::type_convert<float>(b_f4x2.unpack(ck_tile::number<1>{})));
|
||||
else
|
||||
printf("%.2f ", ck_tile::type_convert<float>(b_f4x2.unpack(ck_tile::number<0>{})));
|
||||
}
|
||||
printf("\n");
|
||||
}
|
||||
#else
|
||||
for(int n = 0; n < N; n++)
|
||||
{
|
||||
for(int k = 0; k < K; k++)
|
||||
{
|
||||
printf("0x%02x ", (b_origin_host(k, n).get()));
|
||||
}
|
||||
printf("\n");
|
||||
}
|
||||
#endif
|
||||
printf("\n");
|
||||
|
||||
printf("printf scale_b: \n");
|
||||
for(int n = 0; n < N / DequantGranularityN; n++)
|
||||
{
|
||||
for(int k = 0; k < K / DequantGranularityK; k++)
|
||||
{
|
||||
printf("%.2f ", ck_tile::type_convert<float>(scale_b(k, n)));
|
||||
}
|
||||
printf("\n");
|
||||
}
|
||||
#else
|
||||
printf("printf a_host: \n");
|
||||
for(int m = 0; m < M; m++)
|
||||
{
|
||||
for(int k = 0; k < K;)
|
||||
{
|
||||
printf("0x%08x ", *(reinterpret_cast<uint32_t*>(&a_host(m, k))));
|
||||
k += 2;
|
||||
}
|
||||
printf("\n");
|
||||
}
|
||||
printf("\n");
|
||||
|
||||
printf("printf b_origin_host: \n");
|
||||
for(int n = 0; n < N; n++)
|
||||
{
|
||||
for(int k = 0; k < K;)
|
||||
{
|
||||
printf("0x%08x ", *(reinterpret_cast<uint32_t*>(&b_origin_host(k, n).data)));
|
||||
k += 8;
|
||||
}
|
||||
printf("\n");
|
||||
}
|
||||
|
||||
printf("\n");
|
||||
|
||||
printf("printf scale_b: \n");
|
||||
for(int n = 0; n < N / DequantGranularityN; n++)
|
||||
{
|
||||
for(int k = 0; k < K / DequantGranularityK;)
|
||||
{
|
||||
printf("0x%08x ", *(reinterpret_cast<uint32_t*>(&scale_b(k, n).data)));
|
||||
k += 4;
|
||||
}
|
||||
printf("\n");
|
||||
}
|
||||
#endif
|
||||
#endif
|
||||
|
||||
ck_tile::HostTensor<BDataType> b_shuffled_host(
|
||||
ck_tile::host_tensor_descriptor(K, N, stride_B, is_row_major(b_layout)));
|
||||
preShuffleWeight<FlatmmConfig>(b_origin_host.begin(), b_shuffled_host.begin(), N, K);
|
||||
|
||||
ck_tile::HostTensor<ScaleType> scale_a_shuffled = preShuffleScale<FlatmmConfig>(scale_a);
|
||||
ck_tile::HostTensor<ScaleType> scale_b_shuffled = preShuffleScale<FlatmmConfig>(scale_b);
|
||||
|
||||
ck_tile::DeviceMem a_dev_buf(a_host.get_element_space_size_in_bytes());
|
||||
ck_tile::DeviceMem b_shuffled_dev_buf(b_shuffled_host.get_element_space_size_in_bytes());
|
||||
ck_tile::DeviceMem c_dev_buf(c_rslt_host.get_element_space_size_in_bytes());
|
||||
|
||||
ck_tile::DeviceMem scale_a_dev_buf(scale_a_shuffled.get_element_space_size_in_bytes());
|
||||
ck_tile::DeviceMem scale_b_dev_buf(scale_b_shuffled.get_element_space_size_in_bytes());
|
||||
|
||||
a_dev_buf.ToDevice(a_host.data());
|
||||
b_shuffle_dev_buf.ToDevice(b_shuffled_host.data());
|
||||
c_rslt_host.SetZero();
|
||||
scale_a_dev_buf.ToDevice(scale_a_shuffled.data());
|
||||
scale_b_dev_buf.ToDevice(scale_b_shuffled.data());
|
||||
|
||||
auto scale_a_dev_ptr = ck_tile::FlatmmScalePointer<DequantGranularityM, DequantGranularityK>{
|
||||
static_cast<float*>(scale_a_dev_buf.GetDeviceBuffer()), M / DequantGranularityM};
|
||||
auto scale_b_dev_ptr = ck_tile::FlatmmScalePointer<DequantGranularityN, DequantGranularityK>{
|
||||
static_cast<float*>(scale_b_dev_buf.GetDeviceBuffer()), N / DequantGranularityN};
|
||||
|
||||
invoke_mx_flatmm<FlatmmConfig,
|
||||
ADataType,
|
||||
BDataType,
|
||||
ck_tile::tuple<>,
|
||||
AccDataType,
|
||||
CDataType,
|
||||
ALayout,
|
||||
BLayout,
|
||||
ck_tile::tuple<>,
|
||||
CLayout,
|
||||
decltype(scale_a_dev_ptr),
|
||||
decltype(scale_b_dev_ptr),
|
||||
UsePersistentKernel>(a_dev_buf,
|
||||
b_shuffle_dev_buf,
|
||||
c_dev_buf,
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
stride_A,
|
||||
stride_B,
|
||||
stride_C,
|
||||
kbatch,
|
||||
scale_a_dev_ptr,
|
||||
scale_b_dev_ptr,
|
||||
n_warmup,
|
||||
n_repeat);
|
||||
|
||||
c_dev_buf.FromDevice(c_rslt_host.data());
|
||||
|
||||
bool pass = true;
|
||||
if(arg_parser.get_int("v") == 1)
|
||||
{
|
||||
ck_tile::DeviceMem b_origin_dev_buf(b_origin_host.get_element_space_size_in_bytes());
|
||||
b_origin_dev_buf.ToDevice(b_origin_host.data());
|
||||
|
||||
ck_tile::HostTensor<CDataType> c_gpu_ref_host(
|
||||
ck_tile::host_tensor_descriptor(M, N, stride_C, is_row_major(CLayout{})));
|
||||
ck_tile::DeviceMem c_gpu_ref_dev_buf(c_gpu_ref_host.get_element_space_size_in_bytes());
|
||||
|
||||
ck_tile::HostTensor<AccDataType> scale_A(
|
||||
ck_tile::HostTensorDescriptor({1, K / DequantGranularityK}, {1, 1}));
|
||||
|
||||
// scaleA = 1 has no effect on the result
|
||||
ck_tile::FillUniformDistribution<AccDataType>{1.f, 1.f}(scale_A);
|
||||
ck_tile::DeviceMem scale_A_dev_buf(scale_A.get_element_space_size_in_bytes());
|
||||
scale_A_dev_buf.ToDevice(scale_A.data());
|
||||
|
||||
// convert scale_b from e8m0 to float
|
||||
ck_tile::HostTensor<AccDataType> scale_b_float(ck_tile::HostTensorDescriptor(
|
||||
{K / DequantGranularityK, N / DequantGranularityN}, {N / DequantGranularityN, 1}));
|
||||
std::copy(scale_b.begin(), scale_b.end(), scale_b_float.begin());
|
||||
ck_tile::DeviceMem scale_b_float_dev_buf(scale_b_float.get_element_space_size_in_bytes());
|
||||
scale_b_float_dev_buf.ToDevice(scale_b_float.data());
|
||||
|
||||
c_gpu_ref_dev_buf.SetZero();
|
||||
ck_tile::reference_blockwise_gemm_gpu<ADataType,
|
||||
BDataType,
|
||||
AccDataType,
|
||||
CDataType,
|
||||
ALayout,
|
||||
BLayout,
|
||||
CLayout>(
|
||||
static_cast<ADataType*>(a_dev_buf.GetDeviceBuffer()),
|
||||
static_cast<BDataType*>(b_origin_dev_buf.GetDeviceBuffer()),
|
||||
static_cast<CDataType*>(c_gpu_ref_dev_buf.GetDeviceBuffer()),
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
stride_A,
|
||||
stride_B,
|
||||
stride_C,
|
||||
M,
|
||||
DequantGranularityN,
|
||||
DequantGranularityK,
|
||||
static_cast<float*>(scale_A_dev_buf.GetDeviceBuffer()),
|
||||
static_cast<float*>(scale_b_float_dev_buf.GetDeviceBuffer()));
|
||||
|
||||
c_gpu_ref_dev_buf.FromDevice(c_gpu_ref_host.data());
|
||||
|
||||
const float rtol = std::is_same_v<ADataType, ck_tile::half_t> ? 1e-3 : 1e-2;
|
||||
const float atol = std::is_same_v<ADataType, ck_tile::half_t> ? 1e-3 : 1e-2;
|
||||
|
||||
pass = ck_tile::check_err(
|
||||
c_rslt_host, c_gpu_ref_host, "Error: Incorrect results!", rtol, atol);
|
||||
|
||||
std::cout << "Relative error threshold: " << rtol << " Absolute error threshold: " << atol
|
||||
<< std::endl;
|
||||
std::cout << "The GPU veification result is: " << (pass ? "correct" : "fail") << std::endl;
|
||||
}
|
||||
|
||||
return pass;
|
||||
}
|
||||
Reference in New Issue
Block a user