mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-04-20 14:59:17 +00:00
Gemm+Bilinear (#316)
* refactor * update example * update example * gemm bilinear * clean * update
This commit is contained in:
Chao Liu
1
example/02_gemm_bilinear/CMakeLists.txt
Normal file
1
example/02_gemm_bilinear/CMakeLists.txt
Normal file
@@ -0,0 +1 @@
|
||||
add_example_executable(example_gemm_bilinear_xdl_fp16 gemm_bilinear_xdl_fp16.cpp)
|
||||
28
example/02_gemm_bilinear/README.md
Normal file
28
example/02_gemm_bilinear/README.md
Normal file
@@ -0,0 +1,28 @@
|
||||
# Instructions for ```example_gemm_bilinear_xdl_fp16```
|
||||
|
||||
## Run ```example_gemm_bilinear_xdl_fp16```
|
||||
```bash
|
||||
#arg1: verification (0=no, 1=yes)
|
||||
#arg2: initialization (0=no init, 1=integer value, 2=decimal value)
|
||||
#arg3: time kernel (0=no, 1=yes)
|
||||
#arg4 to 10: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD, StrideE
|
||||
#arg11 to 12: alpha, beta
|
||||
./bin/example_gemm_bilinear_xdl_fp16 1 1 1 3840 4096 4096 4096 4096 4096 4096 0.5 0.5
|
||||
```
|
||||
Result (MI100 @ 1502Mhz, 184.6TFlops peak FP16)
|
||||
```
|
||||
a_m_k: dim 2, lengths {3840, 4096}, strides {4096, 1}
|
||||
b_k_n: dim 2, lengths {4096, 4096}, strides {1, 4096}
|
||||
c0_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
|
||||
c_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
|
||||
arg.a_grid_desc_k0_m_k1_{512, 3840, 8}
|
||||
arg.b_grid_desc_k0_n_k1_{512, 4096, 8}
|
||||
arg.c0_grid_desc_m_n_{ 3840, 4096}
|
||||
arg.c_grid_desc_m_n_{ 3840, 4096}
|
||||
launch_and_time_kernel: grid_dim {480, 1, 1}, block_dim {256, 1, 1}
|
||||
Warm up
|
||||
Start running 1 times...
|
||||
Perf: 0.936965 ms, 137.517 TFlops, 102.959 GB/s
|
||||
error: 0
|
||||
max_diff: 0, 558.5, 558.5
|
||||
```
|
||||
305
example/02_gemm_bilinear/gemm_bilinear_xdl_fp16.cpp
Normal file
305
example/02_gemm_bilinear/gemm_bilinear_xdl_fp16.cpp
Normal file
@@ -0,0 +1,305 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2022, 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/device_gemm_multiple_d_xdl_cshuffle.hpp"
|
||||
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
|
||||
|
||||
#include "ck/library/host_tensor/device_memory.hpp"
|
||||
#include "ck/library/host_tensor/host_tensor.hpp"
|
||||
#include "ck/library/host_tensor/host_tensor_generator.hpp"
|
||||
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
|
||||
#include "ck/library/utility/check_err.hpp"
|
||||
|
||||
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_;
|
||||
};
|
||||
|
||||
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 CShuffleDataType = F32;
|
||||
using DDataType = F16;
|
||||
using DsDataType = ck::Tuple<DDataType>;
|
||||
using EDataType = F16;
|
||||
|
||||
using ALayout = Row;
|
||||
using BLayout = Col;
|
||||
using DELayout = Row;
|
||||
|
||||
using AElementOp = PassThrough;
|
||||
using BElementOp = PassThrough;
|
||||
using CDEElementOp = AlphaBetaAdd;
|
||||
|
||||
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
|
||||
|
||||
using DeviceOpInstance =
|
||||
ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_CShuffle<ALayout,
|
||||
BLayout,
|
||||
DELayout,
|
||||
ADataType,
|
||||
BDataType,
|
||||
AccDataType,
|
||||
CShuffleDataType,
|
||||
DsDataType,
|
||||
EDataType,
|
||||
AElementOp,
|
||||
BElementOp,
|
||||
CDEElementOp,
|
||||
GemmDefault,
|
||||
1,
|
||||
256,
|
||||
256,
|
||||
128,
|
||||
32,
|
||||
8,
|
||||
8,
|
||||
32,
|
||||
32,
|
||||
4,
|
||||
2,
|
||||
S<4, 64, 1>,
|
||||
S<1, 0, 2>,
|
||||
S<1, 0, 2>,
|
||||
2,
|
||||
8,
|
||||
8,
|
||||
1,
|
||||
S<4, 64, 1>,
|
||||
S<1, 0, 2>,
|
||||
S<1, 0, 2>,
|
||||
2,
|
||||
8,
|
||||
8,
|
||||
1,
|
||||
1,
|
||||
1,
|
||||
S<1, 32, 1, 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 = 4096;
|
||||
ck::index_t StrideB = 4096;
|
||||
ck::index_t StrideD = 4096;
|
||||
ck::index_t StrideE = 4096;
|
||||
|
||||
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 9: 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) {
|
||||
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
|
||||
std::vector<std::size_t>({stride, 1}));
|
||||
}
|
||||
else
|
||||
{
|
||||
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
|
||||
std::vector<std::size_t>({1, 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<DDataType> d_m_n(f_host_tensor_descriptor(M, N, StrideD, DELayout{}));
|
||||
Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, DELayout{}));
|
||||
Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, DELayout{}));
|
||||
|
||||
std::cout << "a_m_k: " << a_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:
|
||||
a_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:
|
||||
a_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 a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
|
||||
DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
|
||||
DeviceMem d_m_n_device_buf(sizeof(DDataType) * d_m_n.mDesc.GetElementSpace());
|
||||
DeviceMem e_m_n_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpace());
|
||||
|
||||
a_m_k_device_buf.ToDevice(a_m_k.mData.data());
|
||||
b_k_n_device_buf.ToDevice(b_k_n.mData.data());
|
||||
d_m_n_device_buf.ToDevice(d_m_n.mData.data());
|
||||
e_m_n_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{alpha, beta};
|
||||
|
||||
// do GEMM
|
||||
auto device_op = DeviceOpInstance{};
|
||||
auto invoker = device_op.MakeInvoker();
|
||||
auto argument =
|
||||
device_op.MakeArgument(a_m_k_device_buf.GetDeviceBuffer(),
|
||||
b_k_n_device_buf.GetDeviceBuffer(),
|
||||
std::array<const void*, 1>{d_m_n_device_buf.GetDeviceBuffer()},
|
||||
e_m_n_device_buf.GetDeviceBuffer(),
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
StrideA,
|
||||
StrideB,
|
||||
std::array<ck::index_t, 1>{StrideD},
|
||||
StrideE,
|
||||
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_m_n_device_buf.FromDevice(e_m_n_device_result.mData.data());
|
||||
|
||||
if(do_verification)
|
||||
{
|
||||
Tensor<CShuffleDataType> c_m_n(HostTensorDescriptor(
|
||||
std::vector<std::size_t>{static_cast<std::size_t>(M), static_cast<std::size_t>(N)}));
|
||||
|
||||
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
|
||||
BDataType,
|
||||
CShuffleDataType,
|
||||
AccDataType,
|
||||
AElementOp,
|
||||
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, a_element_op, 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_m_n_device_buf.FromDevice(e_m_n_device_result.mData.data());
|
||||
|
||||
return ck::utils::check_err(e_m_n_device_result.mData, e_m_n_host_result.mData) ? 0 : 1;
|
||||
}
|
||||
|
||||
return 0;
|
||||
}
|
||||
Reference in New Issue
Block a user