mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-04-19 14:29:05 +00:00
Add multiD Gemm client APIs (#534)
* start add example * fix config * fix showinfo bug * add an elementop * change to padding * add xdl example * change elementwiseop * add instance * add instance to profiler * change file name * fix deive not support issue * add client example * fix client gemm_add_multiply name * change AddMultiply elementwiseop * fix elementwiseop * fix client example * fix addmultiply op * fix comments and fun name Co-authored-by: letaoqin <letaoqin@amd.com>
This commit is contained in:
ltqin
3
client_example/15_gemm_add_multiply/CMakeLists.txt
Normal file
3
client_example/15_gemm_add_multiply/CMakeLists.txt
Normal file
@@ -0,0 +1,3 @@
|
||||
|
||||
add_executable(client_gemm_add_multiply gemm_add_multiply.cpp)
|
||||
target_link_libraries(client_gemm_add_multiply PRIVATE composable_kernel::device_operations)
|
||||
241
client_example/15_gemm_add_multiply/gemm_add_multiply.cpp
Normal file
241
client_example/15_gemm_add_multiply/gemm_add_multiply.cpp
Normal file
@@ -0,0 +1,241 @@
|
||||
// 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/gemm_add_multiply.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 AddMultiply = ck::tensor_operation::element_wise::AddMultiply;
|
||||
|
||||
using AElementOp = PassThrough;
|
||||
using BElementOp = PassThrough;
|
||||
using CDEElementOp = AddMultiply;
|
||||
|
||||
using ADataType = F16;
|
||||
using BDataType = F16;
|
||||
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{}));
|
||||
|
||||
using DeviceOp =
|
||||
ck::tensor_operation::device::DeviceGemmMultipleD<ALayout,
|
||||
BLayout,
|
||||
ck::Tuple<D0Layout, D1Layout>,
|
||||
ELayout,
|
||||
ADataType,
|
||||
BDataType,
|
||||
ck::Tuple<D0DataType, D1DataType>,
|
||||
EDataType,
|
||||
AElementOp,
|
||||
BElementOp,
|
||||
CDEElementOp>;
|
||||
|
||||
// get device op instances
|
||||
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
|
||||
DeviceOp>::GetInstances();
|
||||
|
||||
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});
|
||||
}
|
||||
|
||||
std::cout << "Done" << std::endl;
|
||||
}
|
||||
|
||||
return 0;
|
||||
}
|
||||
Reference in New Issue
Block a user