ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-04-19 22:39:03 +00:00
Code Issues Packages Projects Releases Wiki Activity

Support alpha beta scaling for GEMM (#78)

Browse Source 
* [What] Add 2d version of bias, prepare to implement alpha / beta scaling

* Add alpha / beta functor

* Refine parameter of example

* [What] Use real type instead of template
[Why] Prevent implicit cast

* Rename parameter for general operator

* Remove redundant comment

* Fix compile error

Co-authored-by: rocking <chunylai@amd.com>
Co-authored-by: Chao Liu <chao.liu2@amd.com>
This commit is contained in:
rocking5566
2022-02-11 14:48:41 +08:00
committed by GitHub
parent 904cbe2a8f
commit 6f928a0876
6 changed files with 907 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

59
example/8_gemm_xdl_alpha_beta/README.md Normal file
View File

@@ -0,0 +1,59 @@
# Instructions for ```gemm_xdl_alpha_beta``` Example
## Docker script
```bash
docker run \
-it \
--rm \
--privileged \
--group-add sudo \
-w /root/workspace \
-v ${PATH_TO_LOCAL_WORKSPACE}:/root/workspace \
rocm/tensorflow:rocm4.3.1-tf2.6-dev \
/bin/bash
```
## Build ```gemm_xdl_alpha_beta```
```bash
mkdir build && cd build
```
```bash
# Need to specify target ID, example below is gfx908
cmake \
-D BUILD_DEV=OFF \
-D CMAKE_BUILD_TYPE=Release \
-D CMAKE_CXX_FLAGS="-DCK_AMD_GPU_GFX908 --amdgpu-target=gfx908 -O3 " \
-D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc \
-D CMAKE_PREFIX_PATH=/opt/rocm \
..
```
```bash
make -j gemm_xdl_alpha_beta
```
## Run ```gemm_xdl_alpha_beta```
```bash
#arg1: verification (0=no, 1=yes)
#arg2: initialization (0=no init, 1=integer value, 2=decimal value)
#arg3: run kernel # of times (>1)
./example/gemm_xdl_alpha_beta 1 1 1 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
```

272
example/8_gemm_xdl_alpha_beta/gemm_xdl_alpha_beta.cpp Normal file
View File

@@ -0,0 +1,272 @@
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <stdlib.h>
#include <half.hpp>
#include "config.hpp"
#include "print.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "host_gemm.hpp"
#include "device_tensor.hpp"
#include "device_base.hpp"
#include "device_gemm_xdl_c_shuffle_bias_2d.hpp"
#include "element_wise_operation.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using ADataType = ck::half_t;
using BDataType = ck::half_t;
using CDataType = ck::half_t;
using AccDataType = float;
using ALayout = ck::tensor_layout::gemm::RowMajor;
using BLayout = ck::tensor_layout::gemm::ColumnMajor;
using CLayout = ck::tensor_layout::gemm::RowMajor;
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::AlphaBetaAdd;
// clang-format off
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdl_C_Shuffle_Bias_2d<
ADataType, // ADataType
BDataType, // BDataType
CDataType, // CDataType
AccDataType, // AccDataType
ALayout, // ALayout
BLayout, // BLayout
CLayout, // CLayout
AElementOp, // AElementwiseOperation
BElementOp, // BElementwiseOperation
CElementOp, // CElementwiseOperation
256, // BlockSize
256, // MPerBlock
128, // NPerBlock
4, // K0PerBlock
8, // K1
32, // MPerXDL
32, // NPerXDL
4, // MXdlPerWave
2, // NXdlPerWave
S<4, 64, 1>, // ABlockTransferThreadClusterLengths_K0_M_K1
S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
8, // ABlockTransferSrcScalarPerVector
8, // ABlockTransferDstScalarPerVector_K1
true, // ABlockLdsAddExtraM
S<4, 64, 1>, // BBlockTransferThreadClusterLengths_K0_N_K1
S<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
8, // BBlockTransferSrcScalarPerVector
8, // BBlockTransferDstScalarPerVector_K1
true, // BBlockLdsAddExtraN
1, // CShuffleMXdlPerWavePerShuffle
1, // CShuffleNXdlPerWavePerShuffle
S<1, 1, 32, 1, 1, 8>, // CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
8>; // CBlockTransferScalarPerVector_NWaveNPerXdl
// clang-format on
template <typename AType,
typename BType,
typename CType,
typename C0Type,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
static void host_verify(const Tensor<AType>& a_m_k,
const Tensor<BType>& b_k_n,
const Tensor<C0Type>& c0_k_n,
Tensor<CType>& c_m_n,
const AElementwiseOperation& a_element_op,
const BElementwiseOperation& b_element_op,
const CElementwiseOperation& c_element_op)
{
auto f_mk_kn_mn = [&](auto m, auto n) {
const int K = a_m_k.mDesc.GetLengths()[1];
AccDataType v = 0;
AccDataType a = 0;
AccDataType b = 0;
for(int k = 0; k < K; ++k)
{
a_element_op(a, a_m_k(m, k));
b_element_op(b, b_k_n(k, n));
v += a * b;
}
CType y = static_cast<CType>(v);
c_element_op(c_m_n(m, n), y, c0_k_n(m, n));
};
make_ParallelTensorFunctor(f_mk_kn_mn,
c_m_n.mDesc.GetLengths()[0],
c_m_n.mDesc.GetLengths()[1])(std::thread::hardware_concurrency());
}
int main(int argc, char* argv[])
{
bool do_verification = 0;
int init_method = 0;
int nrepeat = 5;
// 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 StrideC = 4096;
float alpha = 1.0f;
float beta = 1.0f;
if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
}
else if(argc == 6)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
alpha = std::stof(argv[4]);
beta = std::stof(argv[5]);
}
else if(argc == 12)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = 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]);
StrideC = std::stoi(argv[9]);
alpha = std::stof(argv[10]);
beta = std::stof(argv[11]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: run kernel # of times (>1)\n");
printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC, 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<BDataType> c0_m_n(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<BDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<BDataType> 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 << "c0_m_n: " << c0_m_n.mDesc << std::endl;
std::cout << "c_m_n: " << c_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});
c0_m_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-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});
c0_m_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-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 c0_m_n_device_buf(sizeof(CDataType) * c0_m_n.mDesc.GetElementSpace());
DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_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());
c0_m_n_device_buf.ToDevice(c0_m_n.mData.data());
c_m_n_device_buf.ToDevice(c_m_n_device_result.mData.data());
// do GEMM
auto gemm = DeviceGemmInstance{};
auto invoker = gemm.MakeInvoker();
auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c0_m_n_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
M,
N,
K,
StrideA,
StrideB,
StrideC,
AElementOp{},
BElementOp{},
CElementOp{alpha, beta});
if(!gemm.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, nrepeat);
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"
<< std::endl;
c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
if(do_verification)
{
host_verify(a_m_k,
b_k_n,
c0_m_n,
c_m_n_host_result,
AElementOp{},
BElementOp{},
CElementOp{alpha, beta});
check_error(c_m_n_host_result, c_m_n_device_result);
}
}

3
example/CMakeLists.txt
View File

@@ -19,6 +19,7 @@ set(CONV2D_FWD_XDL_SOURCE 4_conv2d_fwd_xdl/conv2d_fwd_xdl.cpp)
set(CONV2D_FWD_XDL_BIAS_RELU_SOURCE 5_conv2d_fwd_xdl_bias_relu/conv2d_fwd_xdl_bias_relu.cpp)
set(CONV2D_FWD_XDL_BIAS_RELU_ADD_SOURCE 6_conv2d_fwd_xdl_bias_relu_add/conv2d_fwd_xdl_bias_relu_add.cpp)
set(CONV2D_FWD_XDL_BIAS_RELU_ATOMIC_ADD_SOURCE 7_conv2d_fwd_xdl_bias_relu_atomic_add/conv2d_fwd_xdl_bias_relu_atomic_add.cpp)
set(GEMM_XDL_ALPHA_BETA_SOURCE 8_gemm_xdl_alpha_beta/gemm_xdl_alpha_beta.cpp)
add_executable(gemm_xdl ${GEMM_XDL_SOURCE})
add_executable(gemm_xdl_bias_relu ${GEMM_XDL_BIAS_RELU_SOURCE})
@@ -27,6 +28,7 @@ add_executable(conv2d_fwd_xdl ${CONV2D_FWD_XDL_SOURCE})
add_executable(conv2d_fwd_xdl_bias_relu ${CONV2D_FWD_XDL_BIAS_RELU_SOURCE})
add_executable(conv2d_fwd_xdl_bias_relu_add ${CONV2D_FWD_XDL_BIAS_RELU_ADD_SOURCE})
add_executable(conv2d_fwd_xdl_bias_relu_atomic_add ${CONV2D_FWD_XDL_BIAS_RELU_ATOMIC_ADD_SOURCE})
add_executable(gemm_xdl_alpha_beta ${GEMM_XDL_ALPHA_BETA_SOURCE})
target_link_libraries(gemm_xdl PRIVATE host_tensor)
target_link_libraries(gemm_xdl_bias_relu PRIVATE host_tensor)
@@ -35,3 +37,4 @@ target_link_libraries(conv2d_fwd_xdl PRIVATE host_tensor)
target_link_libraries(conv2d_fwd_xdl_bias_relu PRIVATE host_tensor)
target_link_libraries(conv2d_fwd_xdl_bias_relu_add PRIVATE host_tensor)
target_link_libraries(conv2d_fwd_xdl_bias_relu_atomic_add PRIVATE host_tensor)
target_link_libraries(gemm_xdl_alpha_beta PRIVATE host_tensor)