ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-04-19 14:29:05 +00:00
Code Issues Packages Projects Releases Wiki Activity

External Interface (#304)

Browse Source 
* add client example

* clean

* clean

* reorg

* clean up profiler

* reorg

* clea

* fix profiler

* function for getinstances

* update client example

* update client example

* update client example

* update

* update example

* update Jenkins file

* update cmake

* update Jenkins
This commit is contained in:
Chao Liu
2022-06-26 19:39:02 -05:00
committed by GitHub
parent b653c5eb2e
commit aebd211c36
73 changed files with 2199 additions and 1961 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
client_example/02_gemm_add_add_fastgelu/CMakeLists.txt Normal file
View File

@@ -0,0 +1,2 @@
add_executable(client_gemm_add_add_fastgelu gemm_add_add_fastgelu.cpp)
target_link_libraries(client_gemm_add_add_fastgelu PRIVATE composable_kernel::device_operations)

237
client_example/02_gemm_add_add_fastgelu/gemm_add_add_fastgelu.cpp Normal file
View File

@@ -0,0 +1,237 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iomanip>
#include <vector>
#include <iostream>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/device_gemm_add_add_fastgelu_instance.hpp"
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 AddAddFastGelu = ck::tensor_operation::element_wise::AddAddFastGelu;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = AddAddFastGelu;
using ADataType = F16;
using BDataType = F16;
using AccDataType = F32;
using D0DataType = F16;
using D1DataType = F16;
using EDataType = F16;
using ALayout = Row;
using BLayout = Col;
using D0Layout = Row;
using D1Layout = Row;
using ELayout = Row;
struct SimpleDeviceMem
{
SimpleDeviceMem() = delete;
SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
{
(void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
}
void* GetDeviceBuffer() { return p_mem_; }
~SimpleDeviceMem() { (void)hipFree(p_mem_); }
void* p_mem_;
};
int main(int argc, char* argv[])
{
// 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 StrideD0 = 0;
ck::index_t StrideD1 = 4096;
ck::index_t StrideE = 4096;
if(argc == 1)
{
// use default case
}
else if(argc == 9)
{
M = std::stoi(argv[1]);
N = std::stoi(argv[2]);
K = std::stoi(argv[3]);
StrideA = std::stoi(argv[4]);
StrideB = std::stoi(argv[5]);
StrideD0 = std::stoi(argv[6]);
StrideD1 = std::stoi(argv[7]);
StrideE = std::stoi(argv[8]);
}
else
{
printf("arg1 to 8: M, N, K, StrideA, StrideB, StrideD0, StrideD1, StrideE\n");
exit(0);
}
auto f_matrix_space_size =
[](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
using Layout = decltype(layout);
if(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
{
return (nRow - 1) * stride + nCol;
}
else
{
return (nCol - 1) * stride + nRow;
}
};
SimpleDeviceMem a_device_buf(sizeof(ADataType) * f_matrix_space_size(M, K, StrideA, ALayout{}));
SimpleDeviceMem b_device_buf(sizeof(BDataType) * f_matrix_space_size(K, N, StrideB, BLayout{}));
SimpleDeviceMem d0_m_n_device_buf(sizeof(D0DataType) *
f_matrix_space_size(M, N, StrideD0, D0Layout{}));
SimpleDeviceMem d1_m_n_device_buf(sizeof(D1DataType) *
f_matrix_space_size(M, N, StrideD1, D1Layout{}));
SimpleDeviceMem e_device_buf(sizeof(EDataType) * f_matrix_space_size(M, N, StrideE, ELayout{}));
// add device op instances
const auto op_ptrs = ck::tensor_operation::device::device_gemm_instance::
get_device_gemm_add_add_fastgelu_instances<ADataType,
BDataType,
AccDataType,
D0DataType,
D1DataType,
EDataType,
ALayout,
BLayout,
D0Layout,
D1Layout,
ELayout>();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto cde_element_op = CDEElementOp{};
std::string best_op_name;
bool found = false;
int best_op_id = -1;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
// profile device operation instances
std::cout << "Run all instances and do timing" << std::endl;
for(int i = 0; i < op_ptrs.size(); ++i)
{
auto& op_ptr = op_ptrs[i];
auto argument_ptr = op_ptr->MakeArgumentPointer(
a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
std::array<const void*, 2>{d0_m_n_device_buf.GetDeviceBuffer(),
d1_m_n_device_buf.GetDeviceBuffer()},
e_device_buf.GetDeviceBuffer(),
M,
N,
K,
StrideA,
StrideB,
std::array<ck::index_t, 2>{StrideD0, StrideD1},
StrideE,
a_element_op,
b_element_op,
cde_element_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
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: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << op_name << std::endl;
if(tflops > best_tflops)
{
found = true;
best_op_id = i;
best_op_name = op_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
}
else
{
std::cout << op_name << " does not support this problem" << std::endl;
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
// run the best intance
{
auto& op_ptr = op_ptrs[best_op_id];
std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
<< std::endl;
auto argument_ptr = op_ptr->MakeArgumentPointer(
a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
std::array<const void*, 2>{d0_m_n_device_buf.GetDeviceBuffer(),
d1_m_n_device_buf.GetDeviceBuffer()},
e_device_buf.GetDeviceBuffer(),
M,
N,
K,
StrideA,
StrideB,
std::array<ck::index_t, 2>{StrideD0, StrideD1},
StrideE,
a_element_op,
b_element_op,
cde_element_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
}
}
return 0;
}

9
client_example/CMakeLists.txt Normal file
View File

@@ -0,0 +1,9 @@
cmake_minimum_required(VERSION 3.15)
project(ck_app)
add_compile_options(-std=c++17)
find_package(composable_kernel 1.0.0 COMPONENTS device_operations)
find_package(hip REQUIRED PATHS /opt/rocm)
message(STATUS "Build with HIP ${hip_VERSION}")
add_subdirectory(02_gemm_add_add_fastgelu)

32
client_example/README.md Normal file
View File

@@ -0,0 +1,32 @@
##
Client application links to CK library, and therefore CK library needs to be installed before building client applications.
## Docker script
```bash
docker run \
-it \
--privileged \
--group-add sudo \
-w /root/workspace \
-v ${PATH_TO_LOCAL_WORKSPACE}:/root/workspace \
rocm/tensorflow:rocm5.1-tf2.6-dev \
/bin/bash
```
## Build
```bash
mkdir -p client_example/build
cd client_example/build
```
```bash
cmake \
-D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc \
-D CMAKE_PREFIX_PATH=/opt/rocm \
..
```
### Build client example
```bash
make -j
```