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
251 lines
11 KiB
C++
251 lines
11 KiB
C++
// SPDX-License-Identifier: MIT
|
|
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
|
|
|
|
#pragma once
|
|
|
|
#include <iomanip>
|
|
#include <iostream>
|
|
#include <typeinfo>
|
|
|
|
#include "ck/ck.hpp"
|
|
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
|
|
#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp"
|
|
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
|
|
|
|
#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.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/utility/convolution_parameter.hpp"
|
|
#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
|
|
#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
|
|
|
|
namespace ck {
|
|
namespace profiler {
|
|
|
|
template <ck::index_t NDimSpatial,
|
|
typename InLayout,
|
|
typename WeiLayout,
|
|
typename OutLayout,
|
|
typename InDataType,
|
|
typename WeiDataType,
|
|
typename OutDataType>
|
|
bool profile_grouped_conv_fwd_impl(int do_verification,
|
|
int init_method,
|
|
bool do_log,
|
|
bool time_kernel,
|
|
const ck::utils::conv::ConvParam& conv_param)
|
|
{
|
|
using InElementOp = ck::tensor_operation::element_wise::PassThrough;
|
|
using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
|
|
using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
|
|
|
|
const auto in_element_op = InElementOp{};
|
|
const auto wei_element_op = WeiElementOp{};
|
|
const auto out_element_op = OutElementOp{};
|
|
|
|
const auto in_g_n_c_wis_desc =
|
|
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_param);
|
|
|
|
const auto wei_g_k_c_xs_desc =
|
|
ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(conv_param);
|
|
|
|
const auto out_g_n_k_wos_desc =
|
|
ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(conv_param);
|
|
|
|
std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_lengths{};
|
|
std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_strides{};
|
|
std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_lengths{};
|
|
std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_strides{};
|
|
std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_lengths{};
|
|
std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_strides{};
|
|
std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
|
|
std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
|
|
std::array<ck::index_t, NDimSpatial> input_left_pads{};
|
|
std::array<ck::index_t, NDimSpatial> input_right_pads{};
|
|
|
|
auto copy = [](auto& x, auto& y) { std::copy(x.begin(), x.end(), y.begin()); };
|
|
|
|
copy(in_g_n_c_wis_desc.GetLengths(), a_g_n_c_wis_lengths);
|
|
copy(in_g_n_c_wis_desc.GetStrides(), a_g_n_c_wis_strides);
|
|
copy(wei_g_k_c_xs_desc.GetLengths(), b_g_k_c_xs_lengths);
|
|
copy(wei_g_k_c_xs_desc.GetStrides(), b_g_k_c_xs_strides);
|
|
copy(out_g_n_k_wos_desc.GetLengths(), e_g_n_k_wos_lengths);
|
|
copy(out_g_n_k_wos_desc.GetStrides(), e_g_n_k_wos_strides);
|
|
copy(conv_param.conv_filter_strides_, conv_filter_strides);
|
|
copy(conv_param.conv_filter_dilations_, conv_filter_dilations);
|
|
copy(conv_param.input_left_pads_, input_left_pads);
|
|
copy(conv_param.input_right_pads_, input_right_pads);
|
|
|
|
Tensor<InDataType> input(in_g_n_c_wis_desc);
|
|
Tensor<WeiDataType> weight(wei_g_k_c_xs_desc);
|
|
Tensor<OutDataType> host_output(out_g_n_k_wos_desc);
|
|
Tensor<OutDataType> device_output(out_g_n_k_wos_desc);
|
|
|
|
std::cout << "input: " << input.mDesc << std::endl;
|
|
std::cout << "weight: " << weight.mDesc << std::endl;
|
|
std::cout << "output: " << host_output.mDesc << std::endl;
|
|
|
|
switch(init_method)
|
|
{
|
|
case 0: break;
|
|
case 1:
|
|
input.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
|
|
weight.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
|
|
break;
|
|
default:
|
|
input.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
|
|
weight.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
|
|
}
|
|
|
|
DeviceMem in_device_buf(sizeof(InDataType) * input.mDesc.GetElementSpaceSize());
|
|
DeviceMem wei_device_buf(sizeof(WeiDataType) * weight.mDesc.GetElementSpaceSize());
|
|
DeviceMem out_device_buf(sizeof(OutDataType) * device_output.mDesc.GetElementSpaceSize());
|
|
|
|
in_device_buf.ToDevice(input.mData.data());
|
|
wei_device_buf.ToDevice(weight.mData.data());
|
|
|
|
// run reference op
|
|
if(do_verification)
|
|
{
|
|
auto ref_conv = ck::tensor_operation::host::ReferenceConvFwd<NDimSpatial,
|
|
InDataType,
|
|
WeiDataType,
|
|
OutDataType,
|
|
InElementOp,
|
|
WeiElementOp,
|
|
OutElementOp>{};
|
|
|
|
auto ref_invoker = ref_conv.MakeInvoker();
|
|
auto ref_argument = ref_conv.MakeArgument(input,
|
|
weight,
|
|
host_output,
|
|
conv_param.conv_filter_strides_,
|
|
conv_param.conv_filter_dilations_,
|
|
conv_param.input_left_pads_,
|
|
conv_param.input_right_pads_,
|
|
in_element_op,
|
|
wei_element_op,
|
|
out_element_op);
|
|
|
|
// init host output to zero
|
|
host_output.SetZero();
|
|
|
|
ref_invoker.Run(ref_argument);
|
|
}
|
|
|
|
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NDimSpatial,
|
|
InLayout,
|
|
WeiLayout,
|
|
ck::Tuple<>,
|
|
OutLayout,
|
|
InDataType,
|
|
WeiDataType,
|
|
ck::Tuple<>,
|
|
OutDataType,
|
|
InElementOp,
|
|
WeiElementOp,
|
|
OutElementOp>;
|
|
|
|
// 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;
|
|
|
|
std::string best_op_name;
|
|
float best_avg_time = 0;
|
|
float best_tflops = 0;
|
|
float best_gb_per_sec = 0;
|
|
|
|
// profile device op instances
|
|
bool pass = true;
|
|
|
|
for(auto& op_ptr : op_ptrs)
|
|
{
|
|
auto argument_ptr =
|
|
op_ptr->MakeArgumentPointer(in_device_buf.GetDeviceBuffer(),
|
|
wei_device_buf.GetDeviceBuffer(),
|
|
std::array<const void*, 0>{},
|
|
out_device_buf.GetDeviceBuffer(),
|
|
a_g_n_c_wis_lengths,
|
|
a_g_n_c_wis_strides,
|
|
b_g_k_c_xs_lengths,
|
|
b_g_k_c_xs_strides,
|
|
std::array<std::array<ck::index_t, NDimSpatial + 3>, 0>{{}},
|
|
std::array<std::array<ck::index_t, NDimSpatial + 3>, 0>{{}},
|
|
e_g_n_k_wos_lengths,
|
|
e_g_n_k_wos_strides,
|
|
conv_filter_strides,
|
|
conv_filter_dilations,
|
|
input_left_pads,
|
|
input_right_pads,
|
|
in_element_op,
|
|
wei_element_op,
|
|
out_element_op);
|
|
|
|
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
|
|
{
|
|
// re-init output to zero before profiling next kernel
|
|
out_device_buf.SetZero();
|
|
|
|
std::string op_name = op_ptr->GetTypeString();
|
|
|
|
auto invoker_ptr = op_ptr->MakeInvokerPointer();
|
|
|
|
float avg_time =
|
|
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
|
|
|
|
std::size_t flop = conv_param.GetFlops();
|
|
std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>();
|
|
|
|
float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
|
|
|
|
float gb_per_sec = num_btype / 1.E6 / avg_time;
|
|
|
|
std::cout << "Perf: " << std::setw(10) << avg_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_avg_time = avg_time;
|
|
best_gb_per_sec = gb_per_sec;
|
|
}
|
|
|
|
if(do_verification)
|
|
{
|
|
out_device_buf.FromDevice(device_output.mData.data());
|
|
|
|
pass = pass & ck::utils::check_err(device_output.mData, host_output.mData);
|
|
|
|
if(do_log)
|
|
{
|
|
LogRangeAsType<float>(std::cout << "input : ", input.mData, ",") << std::endl;
|
|
LogRangeAsType<float>(std::cout << "weight: ", weight.mData, ",") << std::endl;
|
|
LogRangeAsType<float>(std::cout << "host_output : ", host_output.mData, ",")
|
|
<< std::endl;
|
|
LogRangeAsType<float>(std::cout << "device_output: ", device_output.mData, ",")
|
|
<< std::endl;
|
|
}
|
|
}
|
|
}
|
|
else
|
|
{
|
|
std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
|
|
}
|
|
}
|
|
|
|
std::cout << "Best configuration parameters:"
|
|
<< "\nname: " << best_op_name << "\navg_time: " << best_avg_time
|
|
<< "\ntflops: " << best_tflops << "\nGB/s: " << best_gb_per_sec << std::endl;
|
|
|
|
return pass;
|
|
}
|
|
|
|
} // namespace profiler
|
|
} // namespace ck
|