mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-04-20 06:49:15 +00:00
ckProfiler and device-level XDL GEMM operator (#48)
* add DeviceGemmXdl * update script * fix naming issue * fix comment * output HostTensorDescriptor * rename * padded GEMM for fwd v4r4r4 nhwc * refactor * refactor * refactor * adding ckProfiler * adding ckProfiler * refactor * fix tuning parameter bug * add more gemm instances * add more fp16 GEMM instances * fix profiler driver * fix bug in tuning parameter * add fp32 gemm instances * small fix * refactor * rename * refactor gemm profiler; adding DeviceConv and conv profiler * refactor * fix * add conv profiler * refactor * adding more GEMM and Conv instance * Create README.md Add build instruction for ckProfiler * Create README.md Add Readme for gemm_xdl example * Update README.md Remove build instruction from top most folder * Update README.md * clean up
This commit is contained in:
Chao Liu
56
example/1_gemm_xdl/README.md
Normal file
56
example/1_gemm_xdl/README.md
Normal file
@@ -0,0 +1,56 @@
|
||||
# Instructions for ```gemm_xdl``` 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```
|
||||
```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
|
||||
```
|
||||
|
||||
## Run ```gemm_xdl```
|
||||
```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.sh 0 1 5
|
||||
```
|
||||
|
||||
Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
|
||||
```
|
||||
a_m_k: dim 2, lengths {3840, 4096}, strides {4096, 1}
|
||||
b_k_n: dim 2, lengths {4096, 4096}, strides {1, 4096}
|
||||
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.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 5 times...
|
||||
Perf: 1.19685 ms, 107.657 TFlops, 78.8501 GB/s
|
||||
```
|
||||
202
example/1_gemm_xdl/gemm_xdl.cpp
Normal file
202
example/1_gemm_xdl/gemm_xdl.cpp
Normal file
@@ -0,0 +1,202 @@
|
||||
#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 "gemm_common.hpp"
|
||||
#include "host_gemm.hpp"
|
||||
#include "device_tensor.hpp"
|
||||
#include "device_base.hpp"
|
||||
#include "device_gemm_xdl.hpp"
|
||||
|
||||
template <typename ADataType,
|
||||
typename BDataType,
|
||||
typename CDataType,
|
||||
typename ALayout,
|
||||
typename BLayout,
|
||||
typename CLayout>
|
||||
struct DeviceGemmInstance;
|
||||
|
||||
template <>
|
||||
struct DeviceGemmInstance<ck::half_t,
|
||||
ck::half_t,
|
||||
ck::half_t,
|
||||
ck::tensor_layout::gemm::RowMajor,
|
||||
ck::tensor_layout::gemm::ColumnMajor,
|
||||
ck::tensor_layout::gemm::RowMajor>
|
||||
{
|
||||
using F16 = ck::half_t;
|
||||
using F32 = float;
|
||||
|
||||
using Row = ck::tensor_layout::gemm::RowMajor;
|
||||
using Col = ck::tensor_layout::gemm::ColumnMajor;
|
||||
|
||||
template <ck::index_t... Is>
|
||||
using S = ck::Sequence<Is...>;
|
||||
|
||||
// Compilation parameters for NT problem
|
||||
// clang-format off
|
||||
using type =
|
||||
//########################################| AData| BData| CData| AccData| ALayout| BLayout| CLayout| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| ABlockLds| BBlockLds|
|
||||
//########################################| Type| Type| Type| Type| | | | Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadSlice| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ThreadSlice| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| SrcDstVectorDim| DstScalar| AddExtraM| AddExtraN|
|
||||
//########################################| | | | | | | | | | | | | | | Wave| Wave| Lengths_K0_N_K1| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| Lengths_K0_N_K1| Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerVector| | |
|
||||
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
ck::tensor_operation::device::DeviceGemmXdl< F16, F16, F16, F32, Row, Col, Row, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 8>, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, S<1, 2, 8>, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 7, 1, true, true>;
|
||||
// clang-format on
|
||||
};
|
||||
|
||||
template <>
|
||||
struct DeviceGemmInstance<float,
|
||||
float,
|
||||
float,
|
||||
ck::tensor_layout::gemm::RowMajor,
|
||||
ck::tensor_layout::gemm::ColumnMajor,
|
||||
ck::tensor_layout::gemm::RowMajor>
|
||||
{
|
||||
using F16 = ck::half_t;
|
||||
using F32 = float;
|
||||
|
||||
using Row = ck::tensor_layout::gemm::RowMajor;
|
||||
using Col = ck::tensor_layout::gemm::ColumnMajor;
|
||||
|
||||
template <ck::index_t... Is>
|
||||
using S = ck::Sequence<Is...>;
|
||||
|
||||
// Compilation parameters for NT problem
|
||||
// clang-format off
|
||||
using type =
|
||||
//########################################| AData| BData| CData| AccData| ALayout| BLayout| CLayout| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| ABlockLds| BBlockLds|
|
||||
//########################################| Type| Type| Type| Type| | | | Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadSlice| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ThreadSlice| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| SrcDstVectorDim| DstScalar| AddExtraM| AddExtraN|
|
||||
//########################################| | | | | | | | | | | | | | | Wave| Wave| Lengths_K0_N_K1| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| Lengths_K0_N_K1| Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerVector| | |
|
||||
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
ck::tensor_operation::device::DeviceGemmXdl< F32, F32, F32, F32, Row, Col, Row, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 4>, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, S<1, 2, 4>, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 7, 1, true, true>;
|
||||
// clang-format on
|
||||
};
|
||||
|
||||
int main(int argc, char* argv[])
|
||||
{
|
||||
if(argc != 4)
|
||||
{
|
||||
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");
|
||||
exit(0);
|
||||
}
|
||||
|
||||
const bool do_verification = std::stoi(argv[1]);
|
||||
const int init_method = std::stoi(argv[2]);
|
||||
const int nrepeat = std::stoi(argv[3]);
|
||||
|
||||
// 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;
|
||||
|
||||
// matrix data type
|
||||
using ADataType = ck::half_t;
|
||||
using BDataType = ck::half_t;
|
||||
using CDataType = ck::half_t;
|
||||
|
||||
// matrix layout
|
||||
using ALayout = ck::tensor_layout::gemm::RowMajor;
|
||||
using BLayout = ck::tensor_layout::gemm::ColumnMajor;
|
||||
using CLayout = ck::tensor_layout::gemm::RowMajor;
|
||||
|
||||
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> 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 << "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{-5, 5});
|
||||
b_k_n.GenerateTensorValue(GeneratorTensor_2{-5, 5});
|
||||
break;
|
||||
default:
|
||||
a_m_k.GenerateTensorValue(GeneratorTensor_3<float>{0.0, 1.0});
|
||||
b_k_n.GenerateTensorValue(GeneratorTensor_3<float>{-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 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());
|
||||
c_m_n_device_buf.ToDevice(c_m_n_device_result.mData.data());
|
||||
|
||||
// do GEMM
|
||||
auto gemm =
|
||||
typename DeviceGemmInstance<ADataType, BDataType, CDataType, ALayout, BLayout, CLayout>::
|
||||
type{};
|
||||
|
||||
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*>(c_m_n_device_buf.GetDeviceBuffer()),
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
StrideA,
|
||||
StrideB,
|
||||
StrideC);
|
||||
|
||||
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 * M + 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_gemm_mk_kn_mn(a_m_k, b_k_n, c_m_n_host_result);
|
||||
|
||||
check_error(c_m_n_host_result, c_m_n_device_result);
|
||||
}
|
||||
}
|
||||
18
example/CMakeLists.txt
Normal file
18
example/CMakeLists.txt
Normal file
@@ -0,0 +1,18 @@
|
||||
include_directories(BEFORE
|
||||
include
|
||||
${PROJECT_SOURCE_DIR}/host/host_tensor/include
|
||||
${PROJECT_SOURCE_DIR}/host/device/include
|
||||
${PROJECT_SOURCE_DIR}/device_operation/include
|
||||
${PROJECT_SOURCE_DIR}/composable_kernel/include
|
||||
${PROJECT_SOURCE_DIR}/composable_kernel/include/utility
|
||||
${PROJECT_SOURCE_DIR}/composable_kernel/include/tensor_description
|
||||
${PROJECT_SOURCE_DIR}/composable_kernel/include/tensor_operation
|
||||
${PROJECT_SOURCE_DIR}/composable_kernel/include/problem_transform
|
||||
${PROJECT_SOURCE_DIR}/external/rocm/include
|
||||
)
|
||||
|
||||
set(GEMM_XDL_SOURCE 1_gemm_xdl/gemm_xdl.cpp)
|
||||
|
||||
add_executable(gemm_xdl ${GEMM_XDL_SOURCE})
|
||||
|
||||
target_link_libraries(gemm_xdl PRIVATE host_tensor)
|
||||
Reference in New Issue
Block a user