ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-14 18:17:44 +00:00
Code Issues Packages Projects Releases Wiki Activity

[GEMM] Gemm universal device operation (#1154)

Browse Source 
* Optimize GEMM on MI200/300:
1. Add new blockwise gemm pipeline
2. Add irregular splitk intances

* clang format + typo fix

* Fix a bug

* initial commit

* Add more instances to irregular splitk

* blkgemm pipeline v1~4 prototype

* Sanity Checked. Known issue:
1. Poor performance of splitk
2. Register spill on blkgemmpipeline v3

* Sanity and Performance fix:
1. fix a bug related to sanity in grouped b2c mapping
2. fix a bug related to sanity and performance in splitk offset

* Sanity and API update:
1. Remove prefetch stage
2. Fix valid check bug
3, Add first gemm_universal instance into ckProfiler

* Add NN instances for gemm universal

* 1. Add NT instances for gemm_universal
2. Fix a bug about Kpadding in gemm_universal

* Fix a bug regarding padding Odd K number

* remove kernel print

* Fix KPadding bug...

* Update safety check

* another try to fix kpadding..

* Sanity checked

* new instances..

* clang format+typo fix

* remove clang format script's change

* Add non-hotloop compile option

* 1. Add fp16xfp8 example
2. pull packed convert f8 from pr1150

* Some miscs.. opt and fix

* Add pipeline description docs

* Split universal gemm instance library to cut profiler compiling time

* uncomment cmakefile

* Fix a bug caused by blockwise_gemm_pipe_v2

* reduce default splitk to 1

* Add 224x256x64 tile size

* update, including:
1. Experiment pipeline 5~7
2. Optimization for pipeline 4
3. Organized instance library

* temp save

* temp save

* Permuted lds layout, sanity and function checked

* clang format

* Move OOB check from RunRead to RunWrite, for better software pipeline.
TODO: agpr spill when NN layout

* clangformat

* A/B splitpipe scheduler for v3

* Fix two bugs

* bug fix

* fix a bug in oob check

* Example for mixed fp16_fp8 gemm

* Clean experimental code blocks

* Add mixed precision gemm into profiler

* tempsave

* optimize m/n major lds layout

* Add RRR GEMM  mixed precision instances

* Optimize f8 matrix transpose

* Add test_gemm_universal

* A/B spilt schedule for blkpip v5

* Take ds_read2 into iglp scheduling scheme

* format

* fixed cmake

* Add llvm-option into CI cmake flag

---------

Co-authored-by: Jing Zhang <jizhan@amd.com>

[ROCm/composable_kernel commit: f83e9701e9]
This commit is contained in:
Haocong WANG
2024-04-14 10:03:18 +08:00
committed by GitHub
parent f77868ae5c
commit ceaecc86ca
107 changed files with 10902 additions and 118 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
example/01_gemm/CMakeLists.txt
View File

@@ -22,6 +22,13 @@ add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp16)
add_example_executable(example_gemm_xdl_fp16_v2 gemm_xdl_fp16_v2.cpp)
add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp16_v2)
add_example_executable(example_gemm_xdl_fp16_v3 gemm_xdl_fp16_v3.cpp)
add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp16_v3)
add_example_executable(example_gemm_xdl_fp8_v3 gemm_xdl_fp8_v3.cpp)
add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp8_v3)
add_example_executable(example_gemm_xdl_fp16_fp8_v3 gemm_xdl_fp16_fp8_v3.cpp)
add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp16_fp8_v3)
add_example_executable(example_gemm_xdl_wavelet_fp16 gemm_xdl_wavelet_fp16.cpp)
add_example_dependencies(example_gemm_xdl example_gemm_xdl_wavelet_fp16)

62
example/01_gemm/common.hpp
View File

@@ -46,6 +46,19 @@ struct ProblemSizeStreamK final
ck::index_t NumSKBlocks = -1;
};
struct ProblemSizeSplitK final
{
ck::index_t M = 3840;
ck::index_t N = 4096;
ck::index_t K = 4096;
ck::index_t StrideA = 4096;
ck::index_t StrideB = 4096;
ck::index_t StrideC = 4096;
ck::index_t KBatch = 1;
};
struct ExecutionConfig final
{
bool do_verification = true;
@@ -158,3 +171,52 @@ bool parse_cmd_args<ProblemSizeStreamK>(int argc,
return true;
}
template <>
bool parse_cmd_args<ProblemSizeSplitK>(int argc,
char* argv[],
ProblemSizeSplitK& problem_size,
ExecutionConfig& config)
{
if(argc == 1)
{
// use default case
}
else if(argc == 4)
{
config.do_verification = std::stoi(argv[1]);
config.init_method = std::stoi(argv[2]);
config.time_kernel = std::stoi(argv[3]);
}
else if(argc >= 10)
{
config.do_verification = std::stoi(argv[1]);
config.init_method = std::stoi(argv[2]);
config.time_kernel = std::stoi(argv[3]);
problem_size.M = std::stoi(argv[4]);
problem_size.N = std::stoi(argv[5]);
problem_size.K = std::stoi(argv[6]);
problem_size.StrideA = std::stoi(argv[7]);
problem_size.StrideB = std::stoi(argv[8]);
problem_size.StrideC = std::stoi(argv[9]);
if(argc >= 11)
{
problem_size.KBatch = std::stoi(argv[10]);
}
}
else
{
std::cerr << "arg1: verification (0=no, 1=yes)" << std::endl
<< "arg2: initialization (0=no init, 1=integer value, 2=decimal value)"
<< std::endl
<< "arg3: time kernel (0=no, 1=yes)" << std::endl
<< "arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC" << std::endl
<< "arg10: KBatch" << std::endl;
return false;
}
return true;
}

53
example/01_gemm/gemm_xdl_fp16_fp8_v3.cpp Normal file
View File

@@ -0,0 +1,53 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3.hpp"
using ADataType = ck::f8_t;
using BDataType = ck::half_t;
using AccDataType = float;
using CShuffleDataType = ck::half_t;
using CDataType = ck::half_t;
using ALayout = Row;
using BLayout = Col;
using CLayout = Row;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CElementOp = PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// clang-format off
using DeviceGemmV2Instance =
ck::tensor_operation::device::DeviceGemm_Xdl_CShuffleV3<
ALayout, BLayout, CLayout,
ADataType, BDataType, CDataType, AccDataType, CShuffleDataType,
AElementOp, BElementOp, CElementOp, GemmDefault,
64,
16, 16,
64, 16, 8,
16, 16,
1, 1,
S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>,
2, 16, 16, 0,
S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>,
2, 8, 8, 0,
1, 1, S<1, 16, 1, 4>, 4,
ck::BlockGemmPipelineScheduler::Intrawave,ck::BlockGemmPipelineVersion::v1>;
// clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
CDataType,
AccDataType,
PassThrough,
PassThrough,
PassThrough>;
#include "run_gemm_example_v2.inc"
int main(int argc, char* argv[]) { return !run_gemm_splitk_example(argc, argv); }

48
example/01_gemm/gemm_xdl_fp16_v3.cpp Normal file
View File

@@ -0,0 +1,48 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3.hpp"
using ADataType = ck::half_t;
using BDataType = ck::half_t;
using AccDataType = float;
using CShuffleDataType = ck::half_t;
using CDataType = ck::half_t;
using ALayout = Row;
using BLayout = Row;
using CLayout = Row;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CElementOp = PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNPadding;
// clang-format off
using DeviceGemmV2Instance =
ck::tensor_operation::device::DeviceGemm_Xdl_CShuffleV3<
ALayout, BLayout, CLayout,
ADataType, BDataType, CDataType, AccDataType, CShuffleDataType,
PassThrough, PassThrough, PassThrough, GemmDefault,
256,
224, 256,
64, 8, 2,
16, 16,
7, 8,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>,
2, 8, 8, 0,
S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>,
1, 8, 2, 0,
1, 2, S<1, 32, 1, 8>, 8,
ck::BlockGemmPipelineScheduler::Intrawave,ck::BlockGemmPipelineVersion::v3>;
// clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host::
ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
#include "run_gemm_example_v2.inc"
int main(int argc, char* argv[]) { return !run_gemm_splitk_example(argc, argv); }

48
example/01_gemm/gemm_xdl_fp8_v3.cpp Normal file
View File

@@ -0,0 +1,48 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3.hpp"
using ADataType = ck::f8_t;
using BDataType = ck::f8_t;
using AccDataType = float;
using CShuffleDataType = ck::half_t;
using CDataType = ck::half_t;
using ALayout = Row;
using BLayout = Col;
using CLayout = Row;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CElementOp = PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// clang-format off
using DeviceGemmV2Instance =
ck::tensor_operation::device::DeviceGemm_Xdl_CShuffleV3<
ALayout, BLayout, CLayout,
ADataType, BDataType, CDataType, AccDataType, CShuffleDataType,
PassThrough, PassThrough, PassThrough, GemmDefault,
256,
128, 256,
128, 16, 16,
16, 16,
4, 8,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>,
2, 16, 16, 1,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>,
2, 16, 16, 1,
1, 2, S<1, 32, 1, 8>, 8,
ck::BlockGemmPipelineScheduler::Intrawave,ck::BlockGemmPipelineVersion::v3, ck::f8_t>;
// clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host::
ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
#include "run_gemm_example_v2.inc"
int main(int argc, char* argv[]) { return !run_gemm_splitk_example(argc, argv); }

211
example/01_gemm/run_gemm_example_v2.inc Normal file
View File

@@ -0,0 +1,211 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
template <typename ProblemType>
bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
{
#if defined(BUILD_INT4_EXAMPLE) && defined(CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4)
static_assert(sizeof(ck::int4_t) == sizeof(int8_t));
#endif
using namespace ck::literals;
auto M = problem_size.M;
auto N = problem_size.N;
auto K = problem_size.K;
auto StrideA = problem_size.StrideA;
auto StrideB = problem_size.StrideB;
auto StrideC = problem_size.StrideC;
auto KBatch = problem_size.KBatch;
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
if constexpr(std::is_same_v<decltype(layout), ck::tensor_layout::gemm::RowMajor>)
{
return HostTensorDescriptor({row, col}, {stride, 1_uz});
}
else
{
return HostTensorDescriptor({row, col}, {1_uz, stride});
}
};
auto f_get_default_stride =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
if(stride == 0)
{
// give a chance if stride is zero, return a default packed stride
if constexpr(std::is_same_v<decltype(layout), ck::tensor_layout::gemm::RowMajor>)
{
return col;
}
else
{
return row;
}
}
else
return stride;
};
StrideA = f_get_default_stride(M, K, StrideA, ALayout{});
StrideB = f_get_default_stride(K, N, StrideB, BLayout{});
StrideC = f_get_default_stride(M, N, StrideC, CLayout{});
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
switch(config.init_method)
{
case 0:
a_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
b_k_n.GenerateTensorValue(GeneratorTensor_1<BDataType>{1});
break;
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
break;
case 2:
a_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
break;
case 3:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b_k_n.GenerateTensorValue(GeneratorTensor_1<BDataType>{1});
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
}
#if 0
printf("B matrix:\n");
for (int in = 0; in < N; in++)
{
for (int ik = 0; ik < K; ik++)
{
printf("%02x ", *(reinterpret_cast<uint8_t*>(&b_k_n(ik,in))));
if(ik%8==7) printf("|");
}
printf("\n");
}
#endif
Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
#ifdef BUILD_INT4_EXAMPLE
DeviceMem a_m_k_device_buf(sizeof(KernelADataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_k_n_device_buf(sizeof(KernelBDataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem c_m_n_device_buf(sizeof(KernelCDataType) *
c_m_n_device_result.mDesc.GetElementSpaceSize());
const Tensor<KernelADataType> a_m_k_converted(a_m_k);
const Tensor<KernelBDataType> b_k_n_converted(b_k_n);
a_m_k_device_buf.ToDevice(a_m_k_converted.mData.data());
b_k_n_device_buf.ToDevice(b_k_n_converted.mData.data());
#else
DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
a_m_k_device_buf.ToDevice(a_m_k.mData.data());
b_k_n_device_buf.ToDevice(b_k_n.mData.data());
#endif
DeviceMem workspace;
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto c_element_op = CElementOp{};
// do GEMM
auto gemm = DeviceGemmV2Instance{};
auto invoker = gemm.MakeInvoker();
float ave_time = 0;
auto argument = gemm.MakeArgument(
#ifdef BUILD_INT4_EXAMPLE
static_cast<KernelADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
static_cast<KernelBDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
static_cast<KernelCDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
#else
static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
#endif
M,
N,
K,
StrideA,
StrideB,
StrideC,
KBatch,
a_element_op,
b_element_op,
c_element_op);
if(!gemm.IsSupportedArgument(argument))
{
std::cerr << gemm.GetTypeString() << " does not support this problem" << std::endl;
return true;
}
bool pass = true;
if(config.do_verification)
{
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(
a_m_k, b_k_n, c_m_n_host_result, PassThrough{}, PassThrough{}, PassThrough{});
ref_invoker.Run(ref_argument);
ave_time = invoker.Run(argument, StreamConfig{nullptr, false, 1});
#ifdef BUILD_INT4_EXAMPLE
Tensor<CDataType> c_m_n_device_result_converted(c_m_n_host_result.mDesc);
c_m_n_device_buf.FromDevice(c_m_n_device_result_converted.mData.data());
c_m_n_device_result = c_m_n_device_result_converted.CopyAsType<CDataType>();
return ck::utils::check_err(c_m_n_device_result_converted, c_m_n_host_result);
#else
c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
pass &= ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
#endif
}
if(config.time_kernel)
{
ave_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
std::size_t flop = 2_uz * M * N * K;
std::size_t num_btype =
sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
<< " GB/s, " << gemm.GetTypeString() << std::endl;
}
return pass;
}
bool run_gemm_splitk_example(int argc, char* argv[])
{
ProblemSizeSplitK problem_size;
ExecutionConfig config;
return !parse_cmd_args(argc, argv, problem_size, config) || run_gemm(problem_size, config);
}