ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-18 20:09:25 +00:00
Code Issues Packages Projects Releases Wiki Activity

Reorganize project folders (#6)

Browse Source
This commit is contained in:
Joseph Macaranas
2025-04-30 13:46:39 -04:00
committed by GitHub
commit 1eb2e57380
3952 changed files with 654944 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

29
example/35_splitK_gemm/CMakeLists.txt Normal file
View File

@@ -0,0 +1,29 @@
add_custom_target(example_splitK_gemm_xdl)
add_example_executable(example_splitK_gemm_xdl_fp32 splitK_gemm_xdl_fp32.cpp)
add_example_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_fp32)
add_example_executable(example_splitK_gemm_xdl_fp16 splitK_gemm_xdl_fp16.cpp)
add_example_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_fp16)
add_example_executable(example_splitK_gemm_xdl_fp16_fp8 splitK_gemm_xdl_fp16_fp8.cpp)
add_example_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_fp16_fp8)
add_example_executable(example_splitK_gemm_xdl_lds_direct_load_fp16 splitK_gemm_xdl_lds_direct_load_fp16.cpp)
add_example_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_lds_direct_load_fp16)
add_example_executable(example_splitK_gemm_xdl_bf16 splitK_gemm_xdl_bf16.cpp)
add_example_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_bf16)
add_example_executable(example_splitK_gemm_xdl_int8 splitK_gemm_xdl_int8.cpp)
add_example_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_int8)
if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_splitK_gemm_xdl_int4 splitK_gemm_xdl_int4.cpp)
add_example_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_int4)
endif()
add_example_executable(example_gemm_xdl_splitk_reduce_multi_d_fp16 gemm_xdl_splitk_reduce_multi_d_fp16.cpp)
add_example_executable(example_gemm_xdl_splitk_reduce_multi_d_bf16 gemm_xdl_splitk_reduce_multi_d_bf16.cpp)
add_example_executable(example_gemm_xdl_splitk_reduce_bf16A_i8B gemm_xdl_splitk_reduce_bf16A_i8B.cpp)
add_example_executable(example_gemm_xdl_splitk_reduce_bfp16 gemm_xdl_splitk_reduce_bf16.cpp)

101
example/35_splitK_gemm/common.hpp Normal file
View File

@@ -0,0 +1,101 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <cstdlib>
#include <iostream>
#include <initializer_list>
#include <numeric>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/utility/data_type.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/fill.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm_multiple_d.hpp"
struct ProblemSizeSplitK final
{
ck::index_t M = 256;
ck::index_t N = 1024;
ck::index_t K = 512;
ck::index_t StrideA = K;
ck::index_t StrideB = N;
ck::index_t StrideC = N;
ck::index_t KBatch = 2;
};
struct ExecutionConfig final
{
bool do_verification = true;
int init_method = 2;
bool time_kernel = true;
};
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using Add = ck::tensor_operation::element_wise::Add;
bool parse_cmd_args(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;
}

58
example/35_splitK_gemm/gemm_xdl_splitk_reduce_bf16.cpp Normal file
View File

@@ -0,0 +1,58 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3r1.hpp"
using ADataType = ck::bhalf_t;
using BDataType = ck::bhalf_t;
using AccDataType = float;
using CShuffleDataType = ck::bhalf_t;
using CDataType = ck::bhalf_t;
using ReduceDataType = ck::bhalf_t;
using D0DataType = ck::bhalf_t;
using DsDataType = ck::Tuple<>;
using ALayout = Row;
using BLayout = Row;
using CLayout = Row;
using D0Layout = CLayout;
using DsLayout = ck::Tuple<>;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNPadding;
// clang-format off
using DeviceGemmV2Instance =
ck::tensor_operation::device::DeviceGemm_Xdl_CShuffleV3R1<
ALayout, BLayout, DsLayout, CLayout,
ADataType, BDataType, DsDataType, CDataType, AccDataType, CShuffleDataType,
AElementOp, BElementOp, CDEElementOp, GemmDefault,
256,
128, 128, 64,
8, 4,
32, 32,
2, 2,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>,
2, 8, 8, 0,
S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>,
1, 8, 4, 0,
1, 1, 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,
PassThrough>;
#include "run_gemm_splitk_reduce_multi_d_example.inc"
int main(int argc, char* argv[]) { return !run_gemm_splitk_example(argc, argv); }

58
example/35_splitK_gemm/gemm_xdl_splitk_reduce_bf16A_i8B.cpp Normal file
View File

@@ -0,0 +1,58 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3r1.hpp"
using ADataType = ck::bhalf_t;
using BDataType = int8_t;
using AccDataType = float;
using CShuffleDataType = ck::bhalf_t;
using CDataType = ck::bhalf_t;
using ReduceDataType = float;
using D0DataType = ck::bhalf_t;
using DsDataType = ck::Tuple<>;
using ALayout = Row;
using BLayout = Row;
using CLayout = Row;
using D0Layout = Row;
using DsLayout = ck::Tuple<>;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNPadding;
// clang-format off
using DeviceGemmV2Instance =
ck::tensor_operation::device::DeviceGemm_Xdl_CShuffleV3R1<
ALayout, BLayout, DsLayout, CLayout,
ADataType, BDataType, DsDataType, CDataType, AccDataType, CShuffleDataType,
AElementOp, BElementOp, CDEElementOp, GemmDefault,
256,
128, 128, 64,
8, 4,
32, 32,
2, 2,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>,
2, 8, 8, 0,
S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>,
1, 8, 4, 0,
1, 1, S<1, 32, 1, 8>, 8,
ck::BlockGemmPipelineScheduler::Intrawave,ck::BlockGemmPipelineVersion::v3, ReduceDataType>;
// clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
CDataType,
AccDataType,
AElementOp,
BElementOp,
PassThrough>;
#include "run_gemm_splitk_reduce_multi_d_example.inc"
int main(int argc, char* argv[]) { return !run_gemm_splitk_example(argc, argv); }

58
example/35_splitK_gemm/gemm_xdl_splitk_reduce_multi_d_bf16.cpp Normal file
View File

@@ -0,0 +1,58 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3r1.hpp"
using ADataType = ck::bhalf_t;
using BDataType = ck::bhalf_t;
using AccDataType = float;
using CShuffleDataType = ck::bhalf_t;
using CDataType = ck::bhalf_t;
using ReduceDataType = float;
using D0DataType = ck::bhalf_t;
using DsDataType = ck::Tuple<D0DataType>;
using ALayout = Row;
using BLayout = Row;
using CLayout = Row;
using D0Layout = CLayout;
using DsLayout = ck::Tuple<D0Layout>;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = Add;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNPadding;
// clang-format off
using DeviceGemmV2Instance =
ck::tensor_operation::device::DeviceGemm_Xdl_CShuffleV3R1<
ALayout, BLayout, DsLayout, CLayout,
ADataType, BDataType, DsDataType, CDataType, AccDataType, CShuffleDataType,
AElementOp, BElementOp, CDEElementOp, GemmDefault,
256,
128, 128, 64,
8, 4,
32, 32,
2, 2,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>,
2, 8, 8, 0,
S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>,
1, 8, 4, 0,
1, 1, S<1, 32, 1, 8>, 8,
ck::BlockGemmPipelineScheduler::Intrawave,ck::BlockGemmPipelineVersion::v3, ReduceDataType>;
// clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
CDataType,
AccDataType,
AElementOp,
BElementOp,
PassThrough>;
#include "run_gemm_splitk_reduce_multi_d_example.inc"
int main(int argc, char* argv[]) { return !run_gemm_splitk_example(argc, argv); }

58
example/35_splitK_gemm/gemm_xdl_splitk_reduce_multi_d_fp16.cpp Normal file
View File

@@ -0,0 +1,58 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3r1.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 ReduceDataType = float;
using D0DataType = ck::half_t;
using DsDataType = ck::Tuple<D0DataType>;
using ALayout = Row;
using BLayout = Row;
using CLayout = Row;
using D0Layout = CLayout;
using DsLayout = ck::Tuple<D0Layout>;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = Add;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNPadding;
// clang-format off
using DeviceGemmV2Instance =
ck::tensor_operation::device::DeviceGemm_Xdl_CShuffleV3R1<
ALayout, BLayout, DsLayout, CLayout,
ADataType, BDataType, DsDataType, CDataType, AccDataType, CShuffleDataType,
AElementOp, BElementOp, CDEElementOp, GemmDefault,
256,
128, 128, 64,
8, 4,
32, 32,
2, 2,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>,
2, 8, 8, 0,
S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>,
1, 8, 4, 0,
1, 1, S<1, 32, 1, 8>, 8,
ck::BlockGemmPipelineScheduler::Intrawave,ck::BlockGemmPipelineVersion::v2, ReduceDataType>;
// clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
CDataType,
AccDataType,
AElementOp,
BElementOp,
PassThrough>;
#include "run_gemm_splitk_reduce_multi_d_example.inc"
int main(int argc, char* argv[]) { return !run_gemm_splitk_example(argc, argv); }

309
example/35_splitK_gemm/run_gemm_splitk_reduce_multi_d_example.inc Normal file
View File

@@ -0,0 +1,309 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
template <typename DataType>
inline __host__ __device__ constexpr double get_rtol()
{
if constexpr(std::is_same_v<DataType, float>)
{
return 1e-3;
}
else if constexpr(std::is_same_v<DataType, double>)
{
return 1e-6;
}
else if constexpr(std::is_same_v<DataType, ck::half_t>)
{
return 1e-3;
}
else if constexpr(std::is_same_v<DataType, ck::bhalf_t>)
{
return 5e-2;
}
else if constexpr(std::is_same_v<DataType, int32_t>)
{
return 1e-1;
}
else if constexpr(std::is_same_v<DataType, int8_t>)
{
return 1e-1;
}
else if constexpr(std::is_same_v<DataType, ck::f8_t>)
{
return 1e-1; // 240 and 224 are acceptable
}
else if constexpr(std::is_same_v<DataType, ck::bf8_t>)
{
return 1.5e-1; // 57344 and 49152 are acceptable
}
else
{
return 1e-3;
}
}
template <typename DataType>
inline __host__ __device__ constexpr double get_atol()
{
if constexpr(std::is_same_v<DataType, float>)
{
return 1e-3;
}
else if constexpr(std::is_same_v<DataType, double>)
{
return 1e-6;
}
else if constexpr(std::is_same_v<DataType, ck::half_t>)
{
return 1e-3;
}
else if constexpr(std::is_same_v<DataType, ck::bhalf_t>)
{
return 5e-2;
}
else if constexpr(std::is_same_v<DataType, int32_t>)
{
return 1e-1;
}
else if constexpr(std::is_same_v<DataType, int8_t>)
{
return 1e-1;
}
else if constexpr(std::is_same_v<DataType, ck::f8_t>)
{
return 16.1; // 240 and 224 are acceptable
}
else if constexpr(std::is_same_v<DataType, ck::bf8_t>)
{
return 8192.1; // 57344 and 49152 are acceptable
}
else
{
return 1e-3;
}
}
template <typename ProblemType>
bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
{
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 StrideD0 = 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{});
StrideD0 = f_get_default_stride(M, N, StrideD0, D0Layout{});
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{}));
Tensor<D0DataType> d0_m_n(f_host_tensor_descriptor(M, N, StrideD0, D0Layout{}));
switch(config.init_method)
{
case 0:
a_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
b_k_n.GenerateTensorValue(GeneratorTensor_1<BDataType>{1});
d0_m_n.GenerateTensorValue(GeneratorTensor_1<D0DataType>{1});
break;
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{-0.5, 0.5});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
d0_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{-0.5, 0.5});
break;
case 2:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
d0_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-2, 2});
break;
case 3:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b_k_n.GenerateTensorValue(GeneratorTensor_1<BDataType>{1});
d0_m_n.GenerateTensorValue(GeneratorTensor_1<D0DataType>{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});
d0_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{-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;
std::cout << "init method: " << config.init_method << std::endl;
std::cout << "KBatch: " << KBatch << std::endl;
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());
DeviceMem d0_m_n_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpaceSize());
a_m_k_device_buf.ToDevice(a_m_k.mData.data());
b_k_n_device_buf.ToDevice(b_k_n.mData.data());
d0_m_n_device_buf.ToDevice(d0_m_n.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto c_element_op = CDEElementOp{};
// do GEMM
auto gemm = DeviceGemmV2Instance{};
auto invoker = gemm.MakeInvoker();
float ave_time = 0;
auto get_argment = [&]() {
if constexpr(DsDataType::Size() > 0)
{
return gemm.MakeArgument(static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
{d0_m_n_device_buf.GetDeviceBuffer()},
static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
M,
N,
K,
StrideA,
StrideB,
{StrideD0},
StrideC,
KBatch,
a_element_op,
b_element_op,
c_element_op);
}
else
{
return gemm.MakeArgument(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()),
M,
N,
K,
StrideA,
StrideB,
{},
StrideC,
KBatch,
a_element_op,
b_element_op,
c_element_op);
}
};
auto argument = get_argment();
if(!gemm.IsSupportedArgument(argument))
{
std::cerr << gemm.GetTypeString() << " does not support this problem" << std::endl;
return true;
}
DeviceMem gemm_workspace_dev(gemm.GetWorkSpaceSize(&argument));
gemm.SetWorkSpacePointer(&argument, gemm_workspace_dev.GetDeviceBuffer(), StreamConfig{});
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});
c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
if constexpr(DsDataType::Size() > 0)
{
c_m_n_host_result.ForEach(
[&](auto& self, auto idx) { c_element_op(self(idx), self(idx), d0_m_n(idx)); });
}
pass &= ck::utils::check_err(c_m_n_device_result,
c_m_n_host_result,
"Error: Incorrect results!",
get_rtol<CDataType>(),
get_atol<CDataType>());
}
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);
}

220
example/35_splitK_gemm/run_splitK_gemm_example.inc Normal file
View File

@@ -0,0 +1,220 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
struct ProblemSize final
{
ck::index_t M = 3840;
ck::index_t N = 4096;
ck::index_t K = 4096;
ck::index_t stride_A = K;
ck::index_t stride_B = K;
ck::index_t stride_C = N;
ck::index_t k_batch = 4;
};
struct ExecutionConfig final
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
};
bool run_splitK_gemm(const ProblemSize& problem_size, const ExecutionConfig& config)
{
using namespace ck::literals;
#if defined(BUILD_INT4_EXAMPLE) && defined(CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4)
static_assert(sizeof(ck::int4_t) == sizeof(int8_t));
static_assert(sizeof(ADataType) == sizeof(KernelADataType));
static_assert(sizeof(BDataType) == sizeof(KernelBDataType));
#endif
auto& [M, N, K, StrideA, StrideB, StrideC, KBatch] = problem_size;
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
using namespace ck::literals;
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor({row, col}, {stride, 1_uz});
}
else
{
return HostTensorDescriptor({row, col}, {1_uz, stride});
}
};
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{}));
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_device_result.mDesc << std::endl;
switch(config.init_method)
{
case 0: break;
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
break;
case 2:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_Sequential<ADataType, 0>{});
b_k_n.GenerateTensorValue(GeneratorTensor_Sequential<BDataType, 1>{});
}
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());
#ifdef BUILD_INT4_EXAMPLE
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
a_m_k_device_buf.ToDevice(a_m_k.mData.data());
b_k_n_device_buf.ToDevice(b_k_n.mData.data());
#endif
c_m_n_device_buf.SetZero();
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto c_element_op = CElementOp{};
// do GEMM
auto gemm = DeviceGemmInstance{};
auto invoker = gemm.MakeInvoker();
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()),
#else
static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
#endif
static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
M,
N,
K,
StrideA,
StrideB,
StrideC,
a_element_op,
b_element_op,
c_element_op,
KBatch);
if(!gemm.IsSupportedArgument(argument))
{
std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
return 0;
}
invoker.Run(argument, StreamConfig{nullptr, false});
bool pass = true;
if(config.do_verification)
{
c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
CDataType,
AccDataType,
AElementOp,
BElementOp,
CElementOp>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
auto ref_argument = ref_gemm.MakeArgument(
a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op);
ref_invoker.Run(ref_argument);
if(std::is_same<CDataType, ck::half_t>::value)
{
pass &= ck::utils::check_err(
c_m_n_device_result, c_m_n_host_result, "fp16 incorrect result", 3e-3, 1e-3);
}
else
{
pass &= ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
}
}
if(config.time_kernel)
{
float ave_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel, 1});
std::size_t flop = std::size_t(2) * 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_splitK_gemm_example(int argc, char* argv[])
{
ProblemSize problem_size;
ExecutionConfig config;
if(argc == 1)
{
// use default case
}
else if(argc == 5)
{
config.do_verification = std::stoi(argv[1]);
config.init_method = std::stoi(argv[2]);
config.time_kernel = std::stoi(argv[3]);
problem_size.k_batch = std::stoi(argv[4]);
}
else if(argc == 11)
{
config.do_verification = std::stoi(argv[1]);
config.init_method = std::stoi(argv[2]);
config.time_kernel = std::stoi(argv[3]);
problem_size.k_batch = std::stoi(argv[4]);
problem_size.M = std::stoi(argv[5]);
problem_size.N = std::stoi(argv[6]);
problem_size.K = std::stoi(argv[7]);
problem_size.stride_A = std::stoi(argv[8]);
problem_size.stride_B = std::stoi(argv[9]);
problem_size.stride_C = std::stoi(argv[10]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=no, 1=yes)\n");
printf("arg4: KBatch\n");
printf("arg5 to 11: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n");
exit(0);
}
return run_splitK_gemm(problem_size, config);
}

59
example/35_splitK_gemm/splitK_gemm_xdl_bf16.cpp Normal file
View File

@@ -0,0 +1,59 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/library/utility/literals.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using BF16 = ck::bhalf_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = BF16;
using BDataType = BF16;
using AccDataType = F32;
using CDataType = F32;
using ComputeType = BF16;
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;
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdlSplitKCShuffle
// clang-format off
//######| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| KPer| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Compute|
//######| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Spacialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Type|
//######| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| |
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout, AElementOp, BElementOp, CElementOp, GemmDefault, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 4, ComputeType>;
// clang-format on
#include "run_splitK_gemm_example.inc"
int main(int argc, char* argv[]) { return !run_splitK_gemm_example(argc, argv); }

58
example/35_splitK_gemm/splitK_gemm_xdl_fp16.cpp Normal file
View File

@@ -0,0 +1,58 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/library/utility/literals.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = F16;
using BDataType = F16;
using AccDataType = F32;
using CDataType = F16;
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::KPadding;
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdlSplitKCShuffle
// clang-format off
//######| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| KPer| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//######| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Spacialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//######| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout, AElementOp, BElementOp, CElementOp, GemmDefault, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8>;
// clang-format on
#include "run_splitK_gemm_example.inc"
int main(int argc, char* argv[]) { return !run_splitK_gemm_example(argc, argv); }

60
example/35_splitK_gemm/splitK_gemm_xdl_fp16_fp8.cpp Normal file
View File

@@ -0,0 +1,60 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024 Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/library/utility/literals.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F8 = ck::f8_t;
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = F16;
using BDataType = F8;
using AccDataType = F32;
using CDataType = F16;
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;
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdlSplitKCShuffle
// clang-format off
//######| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| KPer| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//######| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Spacialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//######| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout, AElementOp, BElementOp, CElementOp, GemmDefault, 128, 16, 64, 8, 16, 16, 16, 1, 2, S<1, 8, 8, 2>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 8, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 16, 16, true, 1, 1, S<1, 16, 1, 8>, 4, F16, ck::PipelineVersion::v1, ck::LoopScheduler::Default, ADataType, BDataType>;
// clang-format on
#include "run_splitK_gemm_example.inc"
int main(int argc, char* argv[]) { return !run_splitK_gemm_example(argc, argv); }

58
example/35_splitK_gemm/splitK_gemm_xdl_fp32.cpp Normal file
View File

@@ -0,0 +1,58 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/library/utility/literals.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = F32;
using BDataType = F32;
using AccDataType = F32;
using CDataType = F32;
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;
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdlSplitKCShuffle
// clang-format off
//######| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| KPer| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//######| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Spacialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//######| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout, AElementOp, BElementOp, CElementOp, GemmDefault, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 4, 4, true, 1, 1, S<1, 32, 1, 8>, 4>;
// clang-format on
#include "run_splitK_gemm_example.inc"
int main(int argc, char* argv[]) { return !run_splitK_gemm_example(argc, argv); }

92
example/35_splitK_gemm/splitK_gemm_xdl_int4.cpp Normal file
View File

@@ -0,0 +1,92 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/library/utility/literals.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = ck::int4_t;
using BDataType = ck::int4_t;
using AccDataType = int32_t;
using CDataType = int32_t;
using KernelADataType = int8_t;
using KernelBDataType = int8_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;
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdlSplitKCShuffle
// clang-format off
<KernelADataType, //ADataType
KernelBDataType, //BDataType
CDataType, //EDataType
AccDataType, //AccDataType
ALayout, //ALayout
BLayout, //BLayout
CLayout, //ELayout
AElementOp, //AElementwiseOperation
BElementOp, //BElementwiseOperation
CElementOp, //CElementwiseOperation
GemmDefault, //GEMMSpecialization
256, // BlockSize
256, // MPerBlock
128, // NPerBlock
4, // KPerBlock
16, // K1
32, // MPerXdl
32, // NPerXdl
4, // MXdlPerWave
2, // NXdlPerWave
S<1, 4, 64, 1>, // ABlockTransfer ThreadCluster Lengths_K0_M_K1
S<0, 2, 1, 3>, // ABlockTransfer ThreadCluster ArrangeOrder
S<0, 2, 1, 3>, // ABlockTransfer SrcAccessOrder
3, // ABlockTransfer SrcVectorDim
16, // ABlockTransfer SrcScalarPerVector
16, // ABlockTransfer DstScalarPerVector_K1
true, // ABlockLdsExtraM
S<1, 4, 64, 1>, // BBlockTransfer ThreadCluster Lengths_K0_N_K1
S<0, 1, 3, 2>, // BBlockTransfer ThreadCluster ArrangeOrder
S<0, 1, 3, 2>, // BBlockTransfer SrcAccessOrder
3, // BBlockTransfer SrcVectorDim
16, // BBlockTransfer SrcScalarPerVector
16, // BBlockTransfer DstScalarPerVector_K1
true, // BBlockLdsExtraN
1, // CShuffleMXdlPerWavePerShuffle
1, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 8>, // CBlockTransferClusterLengths _MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
4>; // CBlockTransferScalarPerVector_NWaveNPerXdl
// clang-format on
#define BUILD_INT4_EXAMPLE
#include "run_splitK_gemm_example.inc"
int main(int argc, char* argv[]) { return !run_splitK_gemm_example(argc, argv); }

56
example/35_splitK_gemm/splitK_gemm_xdl_int8.cpp Normal file
View File

@@ -0,0 +1,56 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/library/utility/literals.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = int8_t;
using BDataType = int8_t;
using AccDataType = int32_t;
using CDataType = int32_t;
using ComputeType = int8_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;
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdlSplitKCShuffle
// clang-format off
//######| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| KPer| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Compute|
//######| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Spacialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Type|
//######| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| |
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout, AElementOp, BElementOp, CElementOp, GemmDefault, 256, 256, 128, 4, 16, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 16, 16, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 16, 16, true, 1, 1, S<1, 32, 1, 8>, 4, ComputeType>;
// clang-format on
#include "run_splitK_gemm_example.inc"
int main(int argc, char* argv[]) { return !run_splitK_gemm_example(argc, argv); }

82
example/35_splitK_gemm/splitK_gemm_xdl_lds_direct_load_fp16.cpp Normal file
View File

@@ -0,0 +1,82 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#define DIRECT_LOAD 1
#if DIRECT_LOAD
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle_lds_direct_load.hpp"
#else
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"
#endif
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/library/utility/literals.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = F16;
using BDataType = F16;
using AccDataType = F32;
using CDataType = F16;
using ALayout = Row;
using BLayout = Col;
using CLayout = Row;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CElementOp = PassThrough;
#if DIRECT_LOAD
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdlSplitKCShuffle_LdsDirectLoad
// clang-format off
//######| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| NumGemmK| Block| MPer| NPer| KPer| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//######| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| AddExtraM| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//######| | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | Wave| Wave| Lengths_KBatch_K0_M_K1| | | PerVector| | Lengths_KBatch_K0_N_K1| | | PerVector| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout, AElementOp, BElementOp, CElementOp, GemmDefault, 2, 128, 32, 16, 4, 16, 16, 16, 1, 1, S<1, 2, 8, 8>, S<0, 2, 1, 3>, 3, 2, true, S<1, 2, 8, 8>, S<0, 2, 1, 3>, 3, 2, true, 1, 1, S<1, 32, 1, 4>, 4>;
// clang-format on
#else
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdlSplitKCShuffle
// clang-format off
//######| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| KPer| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//######| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Spacialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//######| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout, AElementOp, BElementOp, CElementOp, GemmDefault, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8>;
// clang-format on
#endif
#include "run_splitK_gemm_example.inc"
int main(int argc, char* argv[]) { return !run_splitK_gemm_example(argc, argv); }