mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-04 13:41:24 +00:00
500fa99512
* convnd_fwd fp16 example * update example * update example * update instance * updating refernce conv * update reference conv * update conv fwd profiler * update conv 1d and 3d instance * update include path * clean * update profiler for conv bwd data and weight * update conv bwd weight * clean * update conv example * update profiler for conv bwd weight * update ckprofiler for conv bwd data * fix reference conv bwd data bug; update conv bwd data test * update examples * fix initialization issue * update test for conv fwd * clean * clean * remove test case too sensitive to error threshhold * fix test * clean * fix build * adding conv multiple d * adding conv multiple D * add matrix padder * add gemm padding to convnd * adding group conv * update gemm multi-d * refactor * refactor * refactor * clean * clean * refactor * refactor * reorg * add ds * add bias * clean * add G * adding group * adding group * adding group * update Tensor * clean * update example * update DeviceGemmMultipleD_Xdl_CShuffle * update conv bwd-data and bwd-weight * upate contraction example * update gemm and batch gemm with e permute * fix example build * instance for grouped conv1d * update example * adding group conv instance * update gemm bilinear instance * update gemm+add+add+fastgelu instance * update profiler * update profiler * update test * update test and client example * clean * add grouped conv into profiler * update profiler * clean * add test grouped conv, update all conv test to gtest * update test
244 lines
9.3 KiB
C++
244 lines
9.3 KiB
C++
// SPDX-License-Identifier: MIT
|
|
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
|
|
|
|
#pragma once
|
|
|
|
#include <iomanip>
|
|
|
|
#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_add_fastgelu.hpp"
|
|
|
|
#include "ck/library/utility/check_err.hpp"
|
|
#include "ck/library/utility/device_memory.hpp"
|
|
#include "ck/library/utility/host_tensor.hpp"
|
|
#include "ck/library/utility/host_tensor_generator.hpp"
|
|
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
|
|
|
|
namespace ck {
|
|
namespace profiler {
|
|
|
|
template <typename ADataType,
|
|
typename BDataType,
|
|
typename AccDataType,
|
|
typename D0DataType,
|
|
typename D1DataType,
|
|
typename EDataType,
|
|
typename ALayout,
|
|
typename BLayout,
|
|
typename D0Layout,
|
|
typename D1Layout,
|
|
typename ELayout>
|
|
bool profile_gemm_add_add_fastgelu_impl(int do_verification,
|
|
int init_method,
|
|
bool /*do_log*/,
|
|
bool time_kernel,
|
|
int M,
|
|
int N,
|
|
int K,
|
|
int StrideA,
|
|
int StrideB,
|
|
int StrideD0,
|
|
int StrideD1,
|
|
int StrideE)
|
|
{
|
|
auto f_host_tensor_descriptor =
|
|
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
|
|
if(is_same<decltype(layout), 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<D0DataType> d0_m_n(f_host_tensor_descriptor(M, N, StrideD0, D0Layout{}));
|
|
Tensor<D1DataType> d1_m_n(f_host_tensor_descriptor(M, N, StrideD1, D1Layout{}));
|
|
Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
|
|
Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
|
|
|
|
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
|
|
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
|
|
std::cout << "d0_m_n: " << d0_m_n.mDesc << std::endl;
|
|
std::cout << "d1_m_n: " << d1_m_n.mDesc << std::endl;
|
|
std::cout << "e_m_n: " << e_m_n_device_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});
|
|
d0_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-5, 5});
|
|
d1_m_n.GenerateTensorValue(GeneratorTensor_2<D1DataType>{-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});
|
|
d0_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{0.0, 1.0});
|
|
d1_m_n.GenerateTensorValue(GeneratorTensor_3<D1DataType>{0.0, 1.0});
|
|
}
|
|
|
|
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;
|
|
|
|
const auto a_element_op = AElementOp{};
|
|
const auto b_element_op = BElementOp{};
|
|
const auto cde_element_op = CDEElementOp{};
|
|
|
|
using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<
|
|
ALayout,
|
|
BLayout,
|
|
ck::Tuple<D0Layout, D1Layout>,
|
|
ELayout,
|
|
ADataType,
|
|
BDataType,
|
|
ck::Tuple<D0DataType, D1DataType>,
|
|
EDataType,
|
|
ck::tensor_operation::element_wise::PassThrough,
|
|
ck::tensor_operation::element_wise::PassThrough,
|
|
ck::tensor_operation::element_wise::AddAddFastGelu>;
|
|
|
|
// 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;
|
|
|
|
// run reference
|
|
if(do_verification)
|
|
{
|
|
Tensor<AccDataType> 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,
|
|
AccDataType,
|
|
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), d0_m_n(m, n), d1_m_n(m, n));
|
|
}
|
|
}
|
|
}
|
|
|
|
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
|
|
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
|
|
DeviceMem d0_m_n_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpaceSize());
|
|
DeviceMem d1_m_n_device_buf(sizeof(D1DataType) * d1_m_n.mDesc.GetElementSpaceSize());
|
|
DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
|
|
|
|
a_device_buf.ToDevice(a_m_k.mData.data());
|
|
b_device_buf.ToDevice(b_k_n.mData.data());
|
|
d0_m_n_device_buf.ToDevice(d0_m_n.mData.data());
|
|
d1_m_n_device_buf.ToDevice(d1_m_n.mData.data());
|
|
|
|
std::string best_op_name;
|
|
float best_ave_time = 0;
|
|
float best_tflops = 0;
|
|
float best_gb_per_sec = 0;
|
|
|
|
bool pass = true;
|
|
|
|
// profile device operation instances
|
|
for(auto& op_ptr : op_ptrs)
|
|
{
|
|
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()))
|
|
{
|
|
// re-init E to zero before profiling a kernel
|
|
e_device_buf.SetZero();
|
|
|
|
float ave_time =
|
|
invoker_ptr->Run(argument_ptr.get(), 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: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
|
|
<< gb_per_sec << " GB/s, " << op_name << std::endl;
|
|
|
|
if(tflops > best_tflops)
|
|
{
|
|
best_op_name = op_name;
|
|
best_tflops = tflops;
|
|
best_ave_time = ave_time;
|
|
best_gb_per_sec = gb_per_sec;
|
|
}
|
|
|
|
if(do_verification)
|
|
{
|
|
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
|
|
|
|
pass = pass &&
|
|
ck::utils::check_err(e_m_n_device_result.mData, e_m_n_host_result.mData);
|
|
}
|
|
}
|
|
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;
|
|
|
|
return pass;
|
|
}
|
|
|
|
} // namespace profiler
|
|
} // namespace ck
|