mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-04-20 06:49:15 +00:00
Wmma support for multiple ABD GEMM (#2803)
* multi_abd wmma support: - Add multiple A and B support to multiple D implementation (gridwise level) - Add multi_abd GEMM (device level) - Add instances (xdl parity) - Add tests (both xdl and wmma) - Add examples - Add ckProfiler support (both xdl and wmma) * Fix bug in device print function * Fix unused template parameter * Fix batched gemm for multiABD gridwise implementation * Fix gemm_universal_reduce with multiABDs gridwise implementation --------- Co-authored-by: Illia Silin <98187287+illsilin@users.noreply.github.com>
This commit is contained in:
Enrico Degregori
@@ -1,3 +1,7 @@
|
||||
add_example_executable(example_gemm_multi_ABD_wmma_fp16 gemm_multi_ABD_wmma_fp16.cpp)
|
||||
add_example_executable(example_gemm_multi_ABD_wmma_bias_fastgelu_bf16_i8 gemm_multi_ABD_wmma_bias_fastgelu_bf16_i8.cpp)
|
||||
add_example_executable(example_gemm_multi_ABD_wmma_multiply_bias_fastgelu_bf16_i8 gemm_multi_ABD_wmma_multiply_bias_fastgelu_bf16_i8.cpp)
|
||||
add_example_executable(example_gemm_multi_ABD_wmma_fastgelu_bf16_i8 gemm_multi_ABD_wmma_fastgelu_bf16_i8.cpp)
|
||||
add_example_executable(example_gemm_multi_ABD_xdl_fp16 gemm_multi_ABD_xdl_fp16.cpp)
|
||||
add_example_executable(example_gemm_multi_ABD_xdl_bias_fastgelu_bf16_i8 gemm_multi_ABD_xdl_bias_fastgelu_bf16_i8.cpp)
|
||||
add_example_executable(example_gemm_multi_ABD_xdl_multiply_bias_fastgelu_bf16_i8 gemm_multi_ABD_xdl_multiply_bias_fastgelu_bf16_i8.cpp)
|
||||
|
||||
@@ -0,0 +1,307 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2025, 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_multiple_abd_wmma_cshuffle_v3.hpp"
|
||||
#include "ck/tensor_operation/gpu/element/element_wise_operation.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/utility/literals.hpp"
|
||||
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
|
||||
#include "ck/library/utility/check_err.hpp"
|
||||
|
||||
#include "ck/utility/blkgemmpipe_scheduler.hpp"
|
||||
|
||||
template <ck::index_t... Is>
|
||||
using S = ck::Sequence<Is...>;
|
||||
|
||||
using BF16 = ck::bhalf_t;
|
||||
using I8 = int8_t;
|
||||
using F32 = float;
|
||||
|
||||
using Row = ck::tensor_layout::gemm::RowMajor;
|
||||
|
||||
using A0DataType = BF16;
|
||||
using AsDataType = ck::Tuple<A0DataType>;
|
||||
using B0DataType = I8;
|
||||
using B1DataType = BF16;
|
||||
using BsDataType = ck::Tuple<B0DataType, B1DataType>;
|
||||
using AccDataType = F32;
|
||||
using CShuffleDataType = BF16;
|
||||
using D0DataType = BF16;
|
||||
using DsDataType = ck::Tuple<D0DataType>;
|
||||
using EDataType = BF16;
|
||||
|
||||
using A0Layout = Row;
|
||||
using AsLayout = ck::Tuple<A0Layout>;
|
||||
using B0Layout = Row;
|
||||
using B1Layout = B0Layout;
|
||||
using BsLayout = ck::Tuple<B0Layout, B1Layout>;
|
||||
using D0Layout = Row;
|
||||
using DsLayout = ck::Tuple<D0Layout>;
|
||||
using ELayout = Row;
|
||||
|
||||
using Multiply = ck::tensor_operation::element_wise::Multiply;
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
using AddFastGelu = ck::tensor_operation::element_wise::AddFastGelu;
|
||||
|
||||
using AElementOp = PassThrough;
|
||||
using BElementOp = Multiply;
|
||||
using CDEElementOp = AddFastGelu;
|
||||
|
||||
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
|
||||
|
||||
using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleABD_Wmma_CShuffleV3<
|
||||
AsLayout,
|
||||
BsLayout,
|
||||
DsLayout,
|
||||
ELayout,
|
||||
AsDataType,
|
||||
BsDataType,
|
||||
AccDataType,
|
||||
CShuffleDataType,
|
||||
DsDataType,
|
||||
EDataType,
|
||||
AElementOp,
|
||||
BElementOp,
|
||||
CDEElementOp,
|
||||
GemmSpec,
|
||||
256,
|
||||
128,
|
||||
128,
|
||||
64,
|
||||
8,
|
||||
8,
|
||||
16,
|
||||
16,
|
||||
4,
|
||||
2,
|
||||
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,
|
||||
1,
|
||||
8,
|
||||
0,
|
||||
1,
|
||||
1,
|
||||
S<1, 32, 1, 8>,
|
||||
S<8, 8, 8>,
|
||||
ck::BlockGemmPipelineScheduler::Intrawave,
|
||||
ck::BlockGemmPipelineVersion::v3>;
|
||||
|
||||
int main(int argc, char* argv[])
|
||||
{
|
||||
bool do_verification = true;
|
||||
int init_method = 2;
|
||||
bool time_kernel = false;
|
||||
|
||||
// GEMM shape
|
||||
ck::index_t M = 4096;
|
||||
ck::index_t N = 768;
|
||||
ck::index_t K = 6144;
|
||||
|
||||
ck::index_t StrideA = K;
|
||||
ck::index_t StrideB = N;
|
||||
ck::index_t StrideD = N;
|
||||
ck::index_t StrideE = N;
|
||||
|
||||
if(argc == 1)
|
||||
{
|
||||
// use default case
|
||||
}
|
||||
else if(argc == 4)
|
||||
{
|
||||
do_verification = std::stoi(argv[1]);
|
||||
init_method = std::stoi(argv[2]);
|
||||
time_kernel = std::stoi(argv[3]);
|
||||
}
|
||||
else if(argc == 11)
|
||||
{
|
||||
do_verification = std::stoi(argv[1]);
|
||||
init_method = std::stoi(argv[2]);
|
||||
time_kernel = std::stoi(argv[3]);
|
||||
|
||||
M = std::stoi(argv[4]);
|
||||
N = std::stoi(argv[5]);
|
||||
K = std::stoi(argv[6]);
|
||||
|
||||
StrideA = std::stoi(argv[7]);
|
||||
StrideB = std::stoi(argv[8]);
|
||||
StrideD = std::stoi(argv[9]);
|
||||
StrideE = 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 to 10: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD, StrideE\n");
|
||||
exit(0);
|
||||
}
|
||||
|
||||
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<A0DataType> a0_m_k(f_host_tensor_descriptor(M, K, StrideA, A0Layout{}));
|
||||
Tensor<B0DataType> b0_k_n(f_host_tensor_descriptor(K, N, StrideB, B0Layout{}));
|
||||
Tensor<B1DataType> b1_k_n(f_host_tensor_descriptor(K, N, StrideB, B1Layout{}));
|
||||
Tensor<D0DataType> d_m_n(f_host_tensor_descriptor(M, N, StrideD, D0Layout{}));
|
||||
Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
|
||||
Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
|
||||
|
||||
std::cout << "a0_m_k: " << a0_m_k.mDesc << std::endl;
|
||||
std::cout << "b0_k_n: " << b0_k_n.mDesc << std::endl;
|
||||
std::cout << "b1_k_n: " << b1_k_n.mDesc << std::endl;
|
||||
std::cout << "d_m_n: " << d_m_n.mDesc << std::endl;
|
||||
std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
|
||||
|
||||
switch(init_method)
|
||||
{
|
||||
case 0: break;
|
||||
case 1:
|
||||
a0_m_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-5, 5});
|
||||
b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
|
||||
b1_k_n.GenerateTensorValue(GeneratorTensor_2<B1DataType>{0, 5});
|
||||
d_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-5, 5});
|
||||
break;
|
||||
default:
|
||||
a0_m_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
|
||||
b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
|
||||
b1_k_n.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0, 5});
|
||||
d_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{-0.5, 0.5});
|
||||
}
|
||||
|
||||
DeviceMem a0_device_buf(sizeof(A0DataType) * a0_m_k.mDesc.GetElementSpaceSize());
|
||||
DeviceMem b0_device_buf(sizeof(B0DataType) * b0_k_n.mDesc.GetElementSpaceSize());
|
||||
DeviceMem b1_device_buf(sizeof(B1DataType) * b1_k_n.mDesc.GetElementSpaceSize());
|
||||
DeviceMem d_device_buf(sizeof(D0DataType) * d_m_n.mDesc.GetElementSpaceSize());
|
||||
DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
|
||||
|
||||
a0_device_buf.ToDevice(a0_m_k.mData.data());
|
||||
b0_device_buf.ToDevice(b0_k_n.mData.data());
|
||||
b1_device_buf.ToDevice(b1_k_n.mData.data());
|
||||
d_device_buf.ToDevice(d_m_n.mData.data());
|
||||
e_device_buf.ToDevice(e_m_n_device_result.mData.data());
|
||||
|
||||
auto a_element_op = AElementOp{};
|
||||
auto b_element_op = BElementOp{};
|
||||
auto cde_element_op = CDEElementOp{};
|
||||
|
||||
constexpr ck::index_t NumATensor = 1;
|
||||
constexpr ck::index_t NumBTensor = 2;
|
||||
constexpr ck::index_t NumDTensor = 1;
|
||||
|
||||
// do GEMM
|
||||
auto device_op = DeviceOpInstance{};
|
||||
auto invoker = device_op.MakeInvoker();
|
||||
auto argument =
|
||||
device_op.MakeArgument(std::array<const void*, NumATensor>{a0_device_buf.GetDeviceBuffer()},
|
||||
std::array<const void*, NumBTensor>{b0_device_buf.GetDeviceBuffer(),
|
||||
b1_device_buf.GetDeviceBuffer()},
|
||||
std::array<const void*, NumDTensor>{d_device_buf.GetDeviceBuffer()},
|
||||
e_device_buf.GetDeviceBuffer(),
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
std::array<ck::index_t, NumATensor>{StrideA},
|
||||
std::array<ck::index_t, NumBTensor>{StrideB, StrideB},
|
||||
std::array<ck::index_t, NumDTensor>{StrideD},
|
||||
StrideE,
|
||||
1,
|
||||
a_element_op,
|
||||
b_element_op,
|
||||
cde_element_op);
|
||||
|
||||
if(!device_op.IsSupportedArgument(argument))
|
||||
{
|
||||
throw std::runtime_error(
|
||||
"wrong! device_gemm with the specified compilation parameters does "
|
||||
"not support this GEMM problem");
|
||||
}
|
||||
|
||||
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
|
||||
|
||||
std::size_t flop = std::size_t(2) * M * N * K;
|
||||
std::size_t num_btype =
|
||||
sizeof(A0DataType) * M * K + sizeof(B0DataType) * K * N + sizeof(EDataType) * 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"
|
||||
<< std::endl;
|
||||
|
||||
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
|
||||
|
||||
if(do_verification)
|
||||
{
|
||||
Tensor<CShuffleDataType> c_m_n({M, N});
|
||||
|
||||
Tensor<B1DataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, B0Layout{}));
|
||||
|
||||
for(int n = 0; n < N; ++n)
|
||||
{
|
||||
for(int k = 0; k < K; ++k)
|
||||
{
|
||||
b_element_op(b_k_n(k, n), b0_k_n(k, n), b1_k_n(k, n));
|
||||
}
|
||||
}
|
||||
|
||||
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<A0DataType,
|
||||
B1DataType,
|
||||
CShuffleDataType,
|
||||
AccDataType,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>;
|
||||
auto ref_gemm = ReferenceGemmInstance{};
|
||||
auto ref_invoker = ref_gemm.MakeInvoker();
|
||||
|
||||
auto ref_argument = ref_gemm.MakeArgument(
|
||||
a0_m_k, b_k_n, c_m_n, PassThrough{}, PassThrough{}, PassThrough{});
|
||||
|
||||
ref_invoker.Run(ref_argument);
|
||||
|
||||
for(int m = 0; m < M; ++m)
|
||||
{
|
||||
for(int n = 0; n < N; ++n)
|
||||
{
|
||||
cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d_m_n(m, n));
|
||||
}
|
||||
}
|
||||
|
||||
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
|
||||
|
||||
return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result) ? 0 : 1;
|
||||
}
|
||||
|
||||
return 0;
|
||||
}
|
||||
@@ -0,0 +1,299 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2025, 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_multiple_abd_wmma_cshuffle_v3.hpp"
|
||||
#include "ck/tensor_operation/gpu/element/element_wise_operation.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/utility/literals.hpp"
|
||||
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
|
||||
#include "ck/library/utility/check_err.hpp"
|
||||
|
||||
#include "ck/utility/blkgemmpipe_scheduler.hpp"
|
||||
|
||||
template <ck::index_t... Is>
|
||||
using S = ck::Sequence<Is...>;
|
||||
|
||||
using BF16 = ck::bhalf_t;
|
||||
using I8 = int8_t;
|
||||
using F32 = float;
|
||||
|
||||
using Row = ck::tensor_layout::gemm::RowMajor;
|
||||
|
||||
using A0DataType = BF16;
|
||||
using AsDataType = ck::Tuple<A0DataType>;
|
||||
using B0DataType = I8;
|
||||
using B1DataType = BF16;
|
||||
using BsDataType = ck::Tuple<B0DataType, B1DataType>;
|
||||
using AccDataType = F32;
|
||||
using CShuffleDataType = F32;
|
||||
using DsDataType = ck::Tuple<>;
|
||||
using EDataType = BF16;
|
||||
|
||||
using A0Layout = Row;
|
||||
using AsLayout = ck::Tuple<A0Layout>;
|
||||
using B0Layout = Row;
|
||||
using B1Layout = B0Layout;
|
||||
using BsLayout = ck::Tuple<B0Layout, B1Layout>;
|
||||
using DsLayout = ck::Tuple<>;
|
||||
using ELayout = Row;
|
||||
|
||||
using Multiply = ck::tensor_operation::element_wise::Multiply;
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
using FastGelu = ck::tensor_operation::element_wise::FastGelu;
|
||||
|
||||
using AElementOp = PassThrough;
|
||||
using BElementOp = Multiply;
|
||||
using CDEElementOp = FastGelu;
|
||||
|
||||
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
|
||||
|
||||
using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleABD_Wmma_CShuffleV3<
|
||||
AsLayout,
|
||||
BsLayout,
|
||||
DsLayout,
|
||||
ELayout,
|
||||
AsDataType,
|
||||
BsDataType,
|
||||
AccDataType,
|
||||
CShuffleDataType,
|
||||
DsDataType,
|
||||
EDataType,
|
||||
AElementOp,
|
||||
BElementOp,
|
||||
CDEElementOp,
|
||||
GemmSpec,
|
||||
256,
|
||||
128,
|
||||
128,
|
||||
64,
|
||||
8,
|
||||
8,
|
||||
16,
|
||||
16,
|
||||
4,
|
||||
2,
|
||||
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,
|
||||
1,
|
||||
8,
|
||||
0,
|
||||
1,
|
||||
1,
|
||||
S<1, 32, 1, 8>,
|
||||
S<8, 8, 8>,
|
||||
ck::BlockGemmPipelineScheduler::Intrawave,
|
||||
ck::BlockGemmPipelineVersion::v3>;
|
||||
|
||||
int main(int argc, char* argv[])
|
||||
{
|
||||
bool do_verification = true;
|
||||
int init_method = 2;
|
||||
bool time_kernel = false;
|
||||
|
||||
// GEMM shape
|
||||
ck::index_t M = 4096;
|
||||
ck::index_t N = 768;
|
||||
ck::index_t K = 6144;
|
||||
|
||||
ck::index_t StrideA = K;
|
||||
ck::index_t StrideB = N;
|
||||
ck::index_t StrideE = N;
|
||||
|
||||
if(argc == 1)
|
||||
{
|
||||
// use default case
|
||||
}
|
||||
else if(argc == 4)
|
||||
{
|
||||
do_verification = std::stoi(argv[1]);
|
||||
init_method = std::stoi(argv[2]);
|
||||
time_kernel = std::stoi(argv[3]);
|
||||
}
|
||||
else if(argc == 11)
|
||||
{
|
||||
do_verification = std::stoi(argv[1]);
|
||||
init_method = std::stoi(argv[2]);
|
||||
time_kernel = std::stoi(argv[3]);
|
||||
|
||||
M = std::stoi(argv[4]);
|
||||
N = std::stoi(argv[5]);
|
||||
K = std::stoi(argv[6]);
|
||||
|
||||
StrideA = std::stoi(argv[7]);
|
||||
StrideB = std::stoi(argv[8]);
|
||||
StrideE = std::stoi(argv[9]);
|
||||
}
|
||||
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 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideE\n");
|
||||
exit(0);
|
||||
}
|
||||
|
||||
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<A0DataType> a0_m_k(f_host_tensor_descriptor(M, K, StrideA, A0Layout{}));
|
||||
Tensor<B0DataType> b0_k_n(f_host_tensor_descriptor(K, N, StrideB, B0Layout{}));
|
||||
Tensor<B1DataType> b1_k_n(f_host_tensor_descriptor(K, N, StrideB, B1Layout{}));
|
||||
Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
|
||||
Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
|
||||
|
||||
std::cout << "a0_m_k: " << a0_m_k.mDesc << std::endl;
|
||||
std::cout << "b0_k_n: " << b0_k_n.mDesc << std::endl;
|
||||
std::cout << "b1_k_n: " << b1_k_n.mDesc << std::endl;
|
||||
std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
|
||||
|
||||
switch(init_method)
|
||||
{
|
||||
case 0: break;
|
||||
case 1:
|
||||
a0_m_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-5, 5});
|
||||
b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
|
||||
b1_k_n.GenerateTensorValue(GeneratorTensor_2<B1DataType>{0, 5});
|
||||
break;
|
||||
default:
|
||||
a0_m_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
|
||||
b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
|
||||
b1_k_n.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0, 5});
|
||||
}
|
||||
|
||||
DeviceMem a0_device_buf(sizeof(A0DataType) * a0_m_k.mDesc.GetElementSpaceSize());
|
||||
DeviceMem b0_device_buf(sizeof(B0DataType) * b0_k_n.mDesc.GetElementSpaceSize());
|
||||
DeviceMem b1_device_buf(sizeof(B1DataType) * b1_k_n.mDesc.GetElementSpaceSize());
|
||||
DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
|
||||
|
||||
a0_device_buf.ToDevice(a0_m_k.mData.data());
|
||||
b0_device_buf.ToDevice(b0_k_n.mData.data());
|
||||
b1_device_buf.ToDevice(b1_k_n.mData.data());
|
||||
e_device_buf.ToDevice(e_m_n_device_result.mData.data());
|
||||
|
||||
auto a_element_op = AElementOp{};
|
||||
auto b_element_op = BElementOp{};
|
||||
auto cde_element_op = CDEElementOp{};
|
||||
|
||||
constexpr ck::index_t NumATensor = 1;
|
||||
constexpr ck::index_t NumBTensor = 2;
|
||||
constexpr ck::index_t NumDTensor = 0;
|
||||
|
||||
// do GEMM
|
||||
auto device_op = DeviceOpInstance{};
|
||||
auto invoker = device_op.MakeInvoker();
|
||||
auto argument =
|
||||
device_op.MakeArgument(std::array<const void*, NumATensor>{a0_device_buf.GetDeviceBuffer()},
|
||||
std::array<const void*, NumBTensor>{b0_device_buf.GetDeviceBuffer(),
|
||||
b1_device_buf.GetDeviceBuffer()},
|
||||
std::array<const void*, NumDTensor>{},
|
||||
e_device_buf.GetDeviceBuffer(),
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
std::array<ck::index_t, NumATensor>{StrideA},
|
||||
std::array<ck::index_t, NumBTensor>{StrideB, StrideB},
|
||||
std::array<ck::index_t, NumDTensor>{},
|
||||
StrideE,
|
||||
1,
|
||||
a_element_op,
|
||||
b_element_op,
|
||||
cde_element_op);
|
||||
|
||||
if(!device_op.IsSupportedArgument(argument))
|
||||
{
|
||||
throw std::runtime_error(
|
||||
"wrong! device_gemm with the specified compilation parameters does "
|
||||
"not support this GEMM problem");
|
||||
}
|
||||
|
||||
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
|
||||
|
||||
std::size_t flop = std::size_t(2) * M * N * K;
|
||||
std::size_t num_btype =
|
||||
sizeof(A0DataType) * M * K + sizeof(B0DataType) * K * N + sizeof(EDataType) * 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"
|
||||
<< std::endl;
|
||||
|
||||
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
|
||||
|
||||
if(do_verification)
|
||||
{
|
||||
Tensor<CShuffleDataType> c_m_n({M, N});
|
||||
|
||||
Tensor<A0DataType> a_m_k({M, K});
|
||||
|
||||
Tensor<B1DataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, B0Layout{}));
|
||||
|
||||
for(int n = 0; n < N; ++n)
|
||||
{
|
||||
for(int k = 0; k < K; ++k)
|
||||
{
|
||||
b_element_op(b_k_n(k, n), b0_k_n(k, n), b1_k_n(k, n));
|
||||
}
|
||||
}
|
||||
|
||||
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<A0DataType,
|
||||
B1DataType,
|
||||
CShuffleDataType,
|
||||
AccDataType,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>;
|
||||
auto ref_gemm = ReferenceGemmInstance{};
|
||||
auto ref_invoker = ref_gemm.MakeInvoker();
|
||||
|
||||
auto ref_argument = ref_gemm.MakeArgument(
|
||||
a0_m_k, b_k_n, c_m_n, PassThrough{}, PassThrough{}, PassThrough{});
|
||||
|
||||
ref_invoker.Run(ref_argument);
|
||||
|
||||
for(int m = 0; m < M; ++m)
|
||||
{
|
||||
for(int n = 0; n < N; ++n)
|
||||
{
|
||||
cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n));
|
||||
}
|
||||
}
|
||||
|
||||
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
|
||||
|
||||
return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result) ? 0 : 1;
|
||||
}
|
||||
|
||||
return 0;
|
||||
}
|
||||
362
example/60_gemm_multi_ABD/gemm_multi_ABD_wmma_fp16.cpp
Normal file
362
example/60_gemm_multi_ABD/gemm_multi_ABD_wmma_fp16.cpp
Normal file
@@ -0,0 +1,362 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2025, 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_multiple_abd_wmma_cshuffle_v3.hpp"
|
||||
#include "ck/tensor_operation/gpu/element/element_wise_operation.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/utility/literals.hpp"
|
||||
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
|
||||
#include "ck/library/utility/check_err.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 PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
|
||||
using ADataType = F16;
|
||||
using BDataType = F16;
|
||||
using AccDataType = F32;
|
||||
using CShuffleDataType = F32;
|
||||
using DDataType = F16;
|
||||
using EDataType = F16;
|
||||
|
||||
using ALayout = Row;
|
||||
using BLayout = Row;
|
||||
using DLayout = Row;
|
||||
using ELayout = Row;
|
||||
|
||||
struct AddScale
|
||||
{
|
||||
static constexpr auto I0 = ck::Number<0>{};
|
||||
static constexpr auto I1 = ck::Number<1>{};
|
||||
static constexpr auto I2 = ck::Number<2>{};
|
||||
static constexpr auto I3 = ck::Number<3>{};
|
||||
|
||||
__host__ __device__ constexpr void
|
||||
operator()(ck::half4_t& a, const ck::half4_t& a0, const ck::half4_t& a1) const
|
||||
{
|
||||
const auto a0_v_t = ck::vector_type<ck::half_t, 4>{a0};
|
||||
const auto a1_v_t = ck::vector_type<ck::half_t, 4>{a1};
|
||||
|
||||
auto r_v_t = ck::vector_type<ck::half_t, 4>{};
|
||||
|
||||
r_v_t.AsType<ck::half_t>()(I0) =
|
||||
scale * (a0_v_t.AsType<ck::half_t>()[I0] + a1_v_t.AsType<ck::half_t>()[I0]);
|
||||
r_v_t.AsType<ck::half_t>()(I1) =
|
||||
scale * (a0_v_t.AsType<ck::half_t>()[I1] + a1_v_t.AsType<ck::half_t>()[I1]);
|
||||
r_v_t.AsType<ck::half_t>()(I2) =
|
||||
scale * (a0_v_t.AsType<ck::half_t>()[I2] + a1_v_t.AsType<ck::half_t>()[I2]);
|
||||
r_v_t.AsType<ck::half_t>()(I3) =
|
||||
scale * (a0_v_t.AsType<ck::half_t>()[I3] + a1_v_t.AsType<ck::half_t>()[I3]);
|
||||
|
||||
a = r_v_t.AsType<ck::half4_t>()[I0];
|
||||
}
|
||||
|
||||
__host__ __device__ constexpr void
|
||||
operator()(ck::half_t& a, const ck::half_t& a0, const ck::half_t& a1) const
|
||||
{
|
||||
a = scale * (a0 + a1);
|
||||
}
|
||||
|
||||
// this attribute controls the copy_function applying element_wise_op with
|
||||
// pack4_data
|
||||
constexpr const static bool is_pack4_invocable = true;
|
||||
|
||||
float scale = 1.0;
|
||||
};
|
||||
|
||||
struct AlphaBetaAdd
|
||||
{
|
||||
AlphaBetaAdd(float alpha, float beta) : alpha_(alpha), beta_(beta){};
|
||||
|
||||
template <typename E, typename C, typename D>
|
||||
__host__ __device__ constexpr void operator()(E& e, const C& c, const D& d) const;
|
||||
|
||||
template <>
|
||||
__host__ __device__ constexpr void operator()<ck::half_t, float, ck::half_t>(
|
||||
ck::half_t& e, const float& c, const ck::half_t& d) const
|
||||
{
|
||||
e = ck::type_convert<ck::half_t>(alpha_ * c + beta_ * ck::type_convert<float>(d));
|
||||
};
|
||||
|
||||
float alpha_;
|
||||
float beta_;
|
||||
};
|
||||
|
||||
using AElementOp = AddScale;
|
||||
using BElementOp = PassThrough;
|
||||
using CDEElementOp = AlphaBetaAdd;
|
||||
|
||||
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
|
||||
|
||||
using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleABD_Wmma_CShuffleV3<
|
||||
ck::Tuple<ALayout, ALayout>,
|
||||
ck::Tuple<BLayout>,
|
||||
ck::Tuple<DLayout>,
|
||||
ELayout,
|
||||
ck::Tuple<ADataType, ADataType>,
|
||||
ck::Tuple<BDataType>,
|
||||
AccDataType,
|
||||
CShuffleDataType,
|
||||
ck::Tuple<DDataType>,
|
||||
EDataType,
|
||||
AElementOp,
|
||||
BElementOp,
|
||||
CDEElementOp,
|
||||
GemmSpec,
|
||||
256,
|
||||
256,
|
||||
128,
|
||||
32,
|
||||
8,
|
||||
8,
|
||||
16,
|
||||
16,
|
||||
4,
|
||||
4,
|
||||
S<4, 64, 1>,
|
||||
S<1, 0, 2>,
|
||||
S<1, 0, 2>,
|
||||
2,
|
||||
8,
|
||||
8,
|
||||
0,
|
||||
S<4, 64, 1>,
|
||||
S<1, 0, 2>,
|
||||
S<1, 0, 2>,
|
||||
1,
|
||||
1,
|
||||
8,
|
||||
0,
|
||||
1,
|
||||
1,
|
||||
S<1, 64, 1, 4>,
|
||||
S<8, 8, 8>>;
|
||||
|
||||
int main(int argc, char* argv[])
|
||||
{
|
||||
bool do_verification = true;
|
||||
int init_method = 1;
|
||||
bool time_kernel = false;
|
||||
|
||||
// GEMM shape
|
||||
ck::index_t M = 3840;
|
||||
ck::index_t N = 4096;
|
||||
ck::index_t K = 4096;
|
||||
|
||||
ck::index_t StrideA = K;
|
||||
ck::index_t StrideB = N;
|
||||
ck::index_t StrideD = N;
|
||||
ck::index_t StrideE = N;
|
||||
|
||||
float alpha = 1.0f;
|
||||
float beta = 1.0f;
|
||||
|
||||
if(argc == 1)
|
||||
{
|
||||
// use default case
|
||||
}
|
||||
else if(argc == 4)
|
||||
{
|
||||
do_verification = std::stoi(argv[1]);
|
||||
init_method = std::stoi(argv[2]);
|
||||
time_kernel = std::stoi(argv[3]);
|
||||
}
|
||||
else if(argc == 6)
|
||||
{
|
||||
do_verification = std::stoi(argv[1]);
|
||||
init_method = std::stoi(argv[2]);
|
||||
time_kernel = std::stoi(argv[3]);
|
||||
|
||||
alpha = std::stof(argv[4]);
|
||||
beta = std::stof(argv[5]);
|
||||
}
|
||||
else if(argc == 13)
|
||||
{
|
||||
do_verification = std::stoi(argv[1]);
|
||||
init_method = std::stoi(argv[2]);
|
||||
time_kernel = std::stoi(argv[3]);
|
||||
|
||||
M = std::stoi(argv[4]);
|
||||
N = std::stoi(argv[5]);
|
||||
K = std::stoi(argv[6]);
|
||||
|
||||
StrideA = std::stoi(argv[7]);
|
||||
StrideB = std::stoi(argv[8]);
|
||||
StrideD = std::stoi(argv[9]);
|
||||
StrideE = std::stoi(argv[10]);
|
||||
|
||||
alpha = std::stof(argv[11]);
|
||||
beta = std::stof(argv[12]);
|
||||
}
|
||||
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 to 12: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD, StrideE, alpha, "
|
||||
"beta\n");
|
||||
exit(0);
|
||||
}
|
||||
|
||||
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> a0_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
|
||||
Tensor<ADataType> a1_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
|
||||
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
|
||||
Tensor<DDataType> d_m_n(f_host_tensor_descriptor(M, N, StrideD, DLayout{}));
|
||||
Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
|
||||
Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
|
||||
|
||||
std::cout << "a0_m_k: " << a0_m_k.mDesc << std::endl;
|
||||
std::cout << "a1_m_k: " << a1_m_k.mDesc << std::endl;
|
||||
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
|
||||
std::cout << "d_m_n: " << d_m_n.mDesc << std::endl;
|
||||
std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
|
||||
|
||||
switch(init_method)
|
||||
{
|
||||
case 0: break;
|
||||
case 1:
|
||||
a0_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
|
||||
a1_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
|
||||
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
|
||||
d_m_n.GenerateTensorValue(GeneratorTensor_2<DDataType>{-5, 5});
|
||||
break;
|
||||
default:
|
||||
a0_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
|
||||
a1_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
|
||||
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
|
||||
d_m_n.GenerateTensorValue(GeneratorTensor_3<DDataType>{-0.5, 0.5});
|
||||
}
|
||||
|
||||
DeviceMem a0_device_buf(sizeof(ADataType) * a0_m_k.mDesc.GetElementSpaceSize());
|
||||
DeviceMem a1_device_buf(sizeof(ADataType) * a1_m_k.mDesc.GetElementSpaceSize());
|
||||
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
|
||||
DeviceMem d_device_buf(sizeof(DDataType) * d_m_n.mDesc.GetElementSpaceSize());
|
||||
DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
|
||||
|
||||
a0_device_buf.ToDevice(a0_m_k.mData.data());
|
||||
a1_device_buf.ToDevice(a1_m_k.mData.data());
|
||||
b_device_buf.ToDevice(b_k_n.mData.data());
|
||||
d_device_buf.ToDevice(d_m_n.mData.data());
|
||||
e_device_buf.ToDevice(e_m_n_device_result.mData.data());
|
||||
|
||||
auto a_element_op = AElementOp{0.2};
|
||||
auto b_element_op = BElementOp{};
|
||||
auto cde_element_op = CDEElementOp{alpha, beta};
|
||||
|
||||
// do GEMM
|
||||
auto device_op = DeviceOpInstance{};
|
||||
auto invoker = device_op.MakeInvoker();
|
||||
auto argument =
|
||||
device_op.MakeArgument(std::array<const void*, 2>{a0_device_buf.GetDeviceBuffer(),
|
||||
a1_device_buf.GetDeviceBuffer()},
|
||||
std::array<const void*, 1>{b_device_buf.GetDeviceBuffer()},
|
||||
std::array<const void*, 1>{d_device_buf.GetDeviceBuffer()},
|
||||
e_device_buf.GetDeviceBuffer(),
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
std::array<ck::index_t, 2>{StrideA, StrideA},
|
||||
std::array<ck::index_t, 1>{StrideB},
|
||||
std::array<ck::index_t, 1>{StrideD},
|
||||
StrideE,
|
||||
1,
|
||||
a_element_op,
|
||||
b_element_op,
|
||||
cde_element_op);
|
||||
|
||||
if(!device_op.IsSupportedArgument(argument))
|
||||
{
|
||||
throw std::runtime_error(
|
||||
"wrong! device_gemm with the specified compilation parameters does "
|
||||
"not support this GEMM problem");
|
||||
}
|
||||
|
||||
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
|
||||
|
||||
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(EDataType) * 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"
|
||||
<< std::endl;
|
||||
|
||||
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
|
||||
|
||||
if(do_verification)
|
||||
{
|
||||
Tensor<CShuffleDataType> c_m_n({M, N});
|
||||
|
||||
Tensor<ADataType> a_m_k({M, K});
|
||||
|
||||
for(int m = 0; m < M; ++m)
|
||||
{
|
||||
for(int k = 0; k < K; ++k)
|
||||
{
|
||||
a_element_op(a_m_k(m, k), a0_m_k(m, k), a1_m_k(m, k));
|
||||
}
|
||||
}
|
||||
|
||||
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
|
||||
BDataType,
|
||||
CShuffleDataType,
|
||||
AccDataType,
|
||||
PassThrough,
|
||||
BElementOp,
|
||||
PassThrough>;
|
||||
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, PassThrough{}, b_element_op, PassThrough{});
|
||||
|
||||
ref_invoker.Run(ref_argument);
|
||||
|
||||
for(int m = 0; m < M; ++m)
|
||||
{
|
||||
for(int n = 0; n < N; ++n)
|
||||
{
|
||||
cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d_m_n(m, n));
|
||||
}
|
||||
}
|
||||
|
||||
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
|
||||
|
||||
return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result) ? 0 : 1;
|
||||
}
|
||||
|
||||
return 0;
|
||||
}
|
||||
@@ -0,0 +1,296 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2025, 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_multiple_abd_wmma_cshuffle_v3.hpp"
|
||||
#include "ck/tensor_operation/gpu/element/element_wise_operation.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/utility/literals.hpp"
|
||||
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
|
||||
#include "ck/library/utility/check_err.hpp"
|
||||
|
||||
#include "ck/utility/blkgemmpipe_scheduler.hpp"
|
||||
|
||||
template <ck::index_t... Is>
|
||||
using S = ck::Sequence<Is...>;
|
||||
|
||||
using BF16 = ck::bhalf_t;
|
||||
using I8 = int8_t;
|
||||
using F32 = float;
|
||||
|
||||
using Row = ck::tensor_layout::gemm::RowMajor;
|
||||
|
||||
using A0DataType = BF16;
|
||||
using AsDataType = ck::Tuple<A0DataType>;
|
||||
using B0DataType = I8;
|
||||
using BsDataType = ck::Tuple<B0DataType>;
|
||||
using AccDataType = F32;
|
||||
using CShuffleDataType = F32;
|
||||
using D0DataType = BF16;
|
||||
using D1DataType = BF16;
|
||||
using DsDataType = ck::Tuple<D0DataType, D1DataType>;
|
||||
using EDataType = BF16;
|
||||
|
||||
using A0Layout = Row;
|
||||
using AsLayout = ck::Tuple<A0Layout>;
|
||||
using B0Layout = Row;
|
||||
using BsLayout = ck::Tuple<B0Layout>;
|
||||
using D0Layout = Row;
|
||||
using D1Layout = D0Layout;
|
||||
using DsLayout = ck::Tuple<D0Layout, D1Layout>;
|
||||
using ELayout = Row;
|
||||
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
using MultiplyAddFastGelu = ck::tensor_operation::element_wise::MultiplyAddFastGelu;
|
||||
|
||||
using AElementOp = PassThrough;
|
||||
using BElementOp = PassThrough;
|
||||
using CDEElementOp = MultiplyAddFastGelu;
|
||||
|
||||
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
|
||||
|
||||
using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleABD_Wmma_CShuffleV3<
|
||||
AsLayout,
|
||||
BsLayout,
|
||||
DsLayout,
|
||||
ELayout,
|
||||
AsDataType,
|
||||
BsDataType,
|
||||
AccDataType,
|
||||
CShuffleDataType,
|
||||
DsDataType,
|
||||
EDataType,
|
||||
AElementOp,
|
||||
BElementOp,
|
||||
CDEElementOp,
|
||||
GemmSpec,
|
||||
256,
|
||||
128,
|
||||
128,
|
||||
64,
|
||||
8,
|
||||
8,
|
||||
16,
|
||||
16,
|
||||
4,
|
||||
2,
|
||||
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,
|
||||
1,
|
||||
8,
|
||||
0,
|
||||
1,
|
||||
1,
|
||||
S<1, 32, 1, 8>,
|
||||
S<8, 8, 8>,
|
||||
ck::BlockGemmPipelineScheduler::Intrawave,
|
||||
ck::BlockGemmPipelineVersion::v3>;
|
||||
|
||||
int main(int argc, char* argv[])
|
||||
{
|
||||
bool do_verification = true;
|
||||
int init_method = 1;
|
||||
bool time_kernel = false;
|
||||
|
||||
// GEMM shape
|
||||
ck::index_t M = 4096;
|
||||
ck::index_t N = 768;
|
||||
ck::index_t K = 6144;
|
||||
|
||||
ck::index_t StrideA = K;
|
||||
ck::index_t StrideB = N;
|
||||
ck::index_t StrideD = N;
|
||||
ck::index_t StrideE = N;
|
||||
|
||||
if(argc == 1)
|
||||
{
|
||||
// use default case
|
||||
}
|
||||
else if(argc == 4)
|
||||
{
|
||||
do_verification = std::stoi(argv[1]);
|
||||
init_method = std::stoi(argv[2]);
|
||||
time_kernel = std::stoi(argv[3]);
|
||||
}
|
||||
else if(argc == 11)
|
||||
{
|
||||
do_verification = std::stoi(argv[1]);
|
||||
init_method = std::stoi(argv[2]);
|
||||
time_kernel = std::stoi(argv[3]);
|
||||
|
||||
M = std::stoi(argv[4]);
|
||||
N = std::stoi(argv[5]);
|
||||
K = std::stoi(argv[6]);
|
||||
|
||||
StrideA = std::stoi(argv[7]);
|
||||
StrideB = std::stoi(argv[8]);
|
||||
StrideD = std::stoi(argv[9]);
|
||||
StrideE = 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 to 10: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD, StrideE\n");
|
||||
exit(0);
|
||||
}
|
||||
|
||||
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<A0DataType> a0_m_k(f_host_tensor_descriptor(M, K, StrideA, A0Layout{}));
|
||||
Tensor<B0DataType> b0_k_n(f_host_tensor_descriptor(K, N, StrideB, B0Layout{}));
|
||||
Tensor<D0DataType> d0_m_n(f_host_tensor_descriptor(M, N, StrideD, D0Layout{}));
|
||||
Tensor<D1DataType> d1_m_n(f_host_tensor_descriptor(M, N, StrideD, D1Layout{}));
|
||||
Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
|
||||
Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
|
||||
|
||||
std::cout << "a0_m_k: " << a0_m_k.mDesc << std::endl;
|
||||
std::cout << "b0_k_n: " << b0_k_n.mDesc << std::endl;
|
||||
std::cout << "d0_m_n: " << d0_m_n.mDesc << std::endl;
|
||||
std::cout << "d1_m_n: " << d1_m_n.mDesc << std::endl;
|
||||
std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
|
||||
|
||||
switch(init_method)
|
||||
{
|
||||
case 0: break;
|
||||
case 1:
|
||||
a0_m_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-5, 5});
|
||||
b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
|
||||
d0_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-5, 5});
|
||||
d1_m_n.GenerateTensorValue(GeneratorTensor_2<D1DataType>{-5, 5});
|
||||
break;
|
||||
default:
|
||||
a0_m_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
|
||||
b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
|
||||
d0_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{-0.5, 0.5});
|
||||
d1_m_n.GenerateTensorValue(GeneratorTensor_3<D1DataType>{-0.5, 0.5});
|
||||
}
|
||||
|
||||
DeviceMem a0_device_buf(sizeof(A0DataType) * a0_m_k.mDesc.GetElementSpaceSize());
|
||||
DeviceMem b0_device_buf(sizeof(B0DataType) * b0_k_n.mDesc.GetElementSpaceSize());
|
||||
DeviceMem d0_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpaceSize());
|
||||
DeviceMem d1_device_buf(sizeof(D1DataType) * d1_m_n.mDesc.GetElementSpaceSize());
|
||||
DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
|
||||
|
||||
a0_device_buf.ToDevice(a0_m_k.mData.data());
|
||||
b0_device_buf.ToDevice(b0_k_n.mData.data());
|
||||
d0_device_buf.ToDevice(d0_m_n.mData.data());
|
||||
d1_device_buf.ToDevice(d1_m_n.mData.data());
|
||||
e_device_buf.ToDevice(e_m_n_device_result.mData.data());
|
||||
|
||||
auto a_element_op = AElementOp{};
|
||||
auto b_element_op = BElementOp{};
|
||||
auto cde_element_op = CDEElementOp{};
|
||||
|
||||
constexpr ck::index_t NumATensor = 1;
|
||||
constexpr ck::index_t NumBTensor = 1;
|
||||
constexpr ck::index_t NumDTensor = 2;
|
||||
|
||||
// do GEMM
|
||||
auto device_op = DeviceOpInstance{};
|
||||
auto invoker = device_op.MakeInvoker();
|
||||
auto argument =
|
||||
device_op.MakeArgument(std::array<const void*, NumATensor>{a0_device_buf.GetDeviceBuffer()},
|
||||
std::array<const void*, NumBTensor>{b0_device_buf.GetDeviceBuffer()},
|
||||
std::array<const void*, NumDTensor>{d0_device_buf.GetDeviceBuffer(),
|
||||
d1_device_buf.GetDeviceBuffer()},
|
||||
e_device_buf.GetDeviceBuffer(),
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
std::array<ck::index_t, NumATensor>{StrideA},
|
||||
std::array<ck::index_t, NumBTensor>{StrideB},
|
||||
std::array<ck::index_t, NumDTensor>{StrideD, StrideD},
|
||||
StrideE,
|
||||
1,
|
||||
a_element_op,
|
||||
b_element_op,
|
||||
cde_element_op);
|
||||
|
||||
if(!device_op.IsSupportedArgument(argument))
|
||||
{
|
||||
throw std::runtime_error(
|
||||
"wrong! device_gemm with the specified compilation parameters does "
|
||||
"not support this GEMM problem");
|
||||
}
|
||||
|
||||
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
|
||||
|
||||
std::size_t flop = std::size_t(2) * M * N * K;
|
||||
std::size_t num_btype =
|
||||
sizeof(A0DataType) * M * K + sizeof(B0DataType) * K * N + sizeof(EDataType) * 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"
|
||||
<< std::endl;
|
||||
|
||||
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
|
||||
|
||||
if(do_verification)
|
||||
{
|
||||
Tensor<CShuffleDataType> c_m_n({M, N});
|
||||
|
||||
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<A0DataType,
|
||||
B0DataType,
|
||||
CShuffleDataType,
|
||||
AccDataType,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>;
|
||||
auto ref_gemm = ReferenceGemmInstance{};
|
||||
auto ref_invoker = ref_gemm.MakeInvoker();
|
||||
|
||||
auto ref_argument = ref_gemm.MakeArgument(
|
||||
a0_m_k, b0_k_n, c_m_n, PassThrough{}, PassThrough{}, PassThrough{});
|
||||
|
||||
ref_invoker.Run(ref_argument);
|
||||
|
||||
for(int m = 0; m < M; ++m)
|
||||
{
|
||||
for(int n = 0; n < N; ++n)
|
||||
{
|
||||
cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d0_m_n(m, n), d1_m_n(m, n));
|
||||
}
|
||||
}
|
||||
|
||||
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
|
||||
|
||||
return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result) ? 0 : 1;
|
||||
}
|
||||
|
||||
return 0;
|
||||
}
|
||||
Reference in New Issue
Block a user