mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-04-19 22:39:03 +00:00
MaxPool & AvgPool bwd instances, test, ckProfiler, client example (#861)
* Add maxpool instances * Rename index pool to max pool. * Add maxpool bwd bf16 instances * Add avg pool bwd instances * Rename avgpool and maxpool to avg_pool3d and max_pool * Add bf16 pool fwd instances * Add max pool bwd to ckProfiler * Add avg pool3d bwd to ckProfiler * Add avg pool bwd test * Fix bug of reference pool fwd (dilation) * Fix bug of max pool bwd (dilation and initZero) * Support bf16 compute data type * Force compute type be f32. Because atomicAdd only support f32 * Add max pool bwd test * Rename folder * Rename pool * Add max pool bwd client example * Add avg pool bwd client example * Add missing workspace * clang format * Rename macro * remove useless header * remove useless layout
This commit is contained in:
rocking
@@ -100,6 +100,10 @@ int main(int argc, char* argv[])
|
||||
|
||||
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
|
||||
{
|
||||
size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
|
||||
SimpleDeviceMem workspace(workspace_sz);
|
||||
op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
|
||||
|
||||
float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
|
||||
|
||||
std::size_t num_byte = sizeof(XDataType) * M * N + sizeof(GammaDataType) * N +
|
||||
@@ -153,6 +157,10 @@ int main(int argc, char* argv[])
|
||||
|
||||
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
|
||||
{
|
||||
size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
|
||||
SimpleDeviceMem workspace(workspace_sz);
|
||||
op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
|
||||
|
||||
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
|
||||
}
|
||||
|
||||
|
||||
@@ -129,6 +129,10 @@ int main(int argc, char* argv[])
|
||||
|
||||
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
|
||||
{
|
||||
size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
|
||||
SimpleDeviceMem workspace(workspace_sz);
|
||||
op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
|
||||
|
||||
float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
|
||||
|
||||
std::size_t num_byte =
|
||||
@@ -184,6 +188,10 @@ int main(int argc, char* argv[])
|
||||
|
||||
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
|
||||
{
|
||||
size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
|
||||
SimpleDeviceMem workspace(workspace_sz);
|
||||
op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
|
||||
|
||||
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
|
||||
}
|
||||
|
||||
|
||||
11
client_example/19_pool/CMakeLists.txt
Normal file
11
client_example/19_pool/CMakeLists.txt
Normal file
@@ -0,0 +1,11 @@
|
||||
add_executable(client_max_pool2d_fwd max_pool2d_fwd.cpp)
|
||||
target_link_libraries(client_max_pool2d_fwd PRIVATE composable_kernel::device_operations)
|
||||
|
||||
add_executable(client_max_pool2d_bwd max_pool2d_bwd.cpp)
|
||||
target_link_libraries(client_max_pool2d_bwd PRIVATE composable_kernel::device_operations)
|
||||
|
||||
add_executable(client_avg_pool3d_fwd avg_pool3d_fwd.cpp)
|
||||
target_link_libraries(client_avg_pool3d_fwd PRIVATE composable_kernel::device_operations)
|
||||
|
||||
add_executable(client_avg_pool3d_bwd avg_pool3d_bwd.cpp)
|
||||
target_link_libraries(client_avg_pool3d_bwd PRIVATE composable_kernel::device_operations)
|
||||
191
client_example/19_pool/avg_pool3d_bwd.cpp
Normal file
191
client_example/19_pool/avg_pool3d_bwd.cpp
Normal file
@@ -0,0 +1,191 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, 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/element/element_wise_operation.hpp"
|
||||
|
||||
#include "ck/library/tensor_operation_instance/gpu/avg_pool3d_bwd.hpp"
|
||||
|
||||
using DOutDataType = ck::half_t;
|
||||
using DInDataType = ck::half_t;
|
||||
|
||||
using DOutLayout = ck::tensor_layout::convolution::NDHWC;
|
||||
using DInLayout = ck::tensor_layout::convolution::NDHWC;
|
||||
|
||||
struct SimpleDeviceMem
|
||||
{
|
||||
SimpleDeviceMem() = delete;
|
||||
|
||||
SimpleDeviceMem(std::size_t mem_size) : p_mem_{}, mMemSize_(mem_size)
|
||||
{
|
||||
(void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
|
||||
}
|
||||
|
||||
void* GetDeviceBuffer() { return p_mem_; }
|
||||
|
||||
void SetZero() const { (void)hipMemset(p_mem_, 0, mMemSize_); }
|
||||
|
||||
~SimpleDeviceMem() { (void)hipFree(p_mem_); }
|
||||
|
||||
void* p_mem_;
|
||||
std::size_t mMemSize_;
|
||||
};
|
||||
|
||||
int main(int argc, char* argv[])
|
||||
{
|
||||
ck::index_t N = 2;
|
||||
ck::index_t C = 32;
|
||||
ck::index_t Z = 2;
|
||||
ck::index_t Y = 2;
|
||||
ck::index_t X = 2;
|
||||
ck::index_t Di = 30;
|
||||
ck::index_t Hi = 30;
|
||||
ck::index_t Wi = 30;
|
||||
ck::index_t window_stride_d = 2;
|
||||
ck::index_t window_stride_h = 2;
|
||||
ck::index_t window_stride_w = 2;
|
||||
ck::index_t window_dilation_d = 1;
|
||||
ck::index_t window_dilation_h = 1;
|
||||
ck::index_t window_dilation_w = 1;
|
||||
ck::index_t in_left_pad_d = 1;
|
||||
ck::index_t in_left_pad_h = 1;
|
||||
ck::index_t in_left_pad_w = 1;
|
||||
ck::index_t in_right_pad_d = 1;
|
||||
ck::index_t in_right_pad_h = 1;
|
||||
ck::index_t in_right_pad_w = 1;
|
||||
|
||||
const ck::index_t Zs = (Z - 1) * window_dilation_d + 1;
|
||||
const ck::index_t Ys = (Y - 1) * window_dilation_h + 1;
|
||||
const ck::index_t Xs = (X - 1) * window_dilation_w + 1;
|
||||
ck::index_t Do = (Di + in_left_pad_d + in_right_pad_d - Zs) / window_stride_d + 1;
|
||||
ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - Ys) / window_stride_h + 1;
|
||||
ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - Xs) / window_stride_w + 1;
|
||||
|
||||
// Pool API only support the order of NCDHW
|
||||
std::vector<ck::index_t> in_length = {N, C, Di, Hi, Wi};
|
||||
std::vector<ck::index_t> out_length = {N, C, Do, Ho, Wo};
|
||||
std::vector<ck::index_t> window_spatial_lengths = {Z, Y, X};
|
||||
std::vector<ck::index_t> window_strides = {window_stride_d, window_stride_h, window_stride_w};
|
||||
std::vector<ck::index_t> window_dilations{
|
||||
window_dilation_d, window_dilation_h, window_dilation_w};
|
||||
std::vector<ck::index_t> input_left_pads = {in_left_pad_d, in_left_pad_h, in_left_pad_w};
|
||||
std::vector<ck::index_t> input_right_pads = {in_right_pad_d, in_right_pad_h, in_right_pad_w};
|
||||
|
||||
std::size_t in_tensor_size = N * C * Di * Hi * Wi;
|
||||
std::size_t out_tensor_size = N * C * Do * Ho * Wo;
|
||||
|
||||
// tensor layout = NDHWC
|
||||
std::vector<ck::index_t> in_tensor_stride = {Di * C * Hi * Wi, 1, C * Hi * Wi, Wi * C, C};
|
||||
std::vector<ck::index_t> out_tensor_stride = {Do * C * Ho * Wo, 1, C * Ho * Wo, Wo * C, C};
|
||||
|
||||
SimpleDeviceMem dout_device_buf(sizeof(DOutDataType) * out_tensor_size);
|
||||
SimpleDeviceMem din_device_buf(sizeof(DInDataType) * in_tensor_size);
|
||||
|
||||
using DeviceOp = ck::tensor_operation::device::
|
||||
DeviceAvgPoolBwd<3, DOutDataType, DInDataType, DOutLayout, DInLayout>;
|
||||
|
||||
// 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;
|
||||
bool found = false;
|
||||
int best_op_id = -1;
|
||||
float best_ave_time = std::numeric_limits<float>::max();
|
||||
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(
|
||||
static_cast<DOutDataType*>(dout_device_buf.GetDeviceBuffer()),
|
||||
static_cast<DInDataType*>(din_device_buf.GetDeviceBuffer()),
|
||||
out_length,
|
||||
in_length,
|
||||
out_tensor_stride,
|
||||
in_tensor_stride,
|
||||
window_spatial_lengths,
|
||||
window_strides,
|
||||
window_dilations,
|
||||
input_left_pads,
|
||||
input_right_pads);
|
||||
|
||||
auto invoker_ptr = op_ptr->MakeInvokerPointer();
|
||||
|
||||
std::string op_name = op_ptr->GetTypeString();
|
||||
|
||||
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
|
||||
{
|
||||
din_device_buf.SetZero();
|
||||
|
||||
float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
|
||||
|
||||
std::size_t num_bytes =
|
||||
in_tensor_size * sizeof(DInDataType) + out_tensor_size * sizeof(DOutDataType);
|
||||
|
||||
float gb_per_sec = num_bytes / 1.E6 / ave_time;
|
||||
|
||||
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec << " GB/s, "
|
||||
<< op_name << std::endl;
|
||||
|
||||
if(ave_time < best_ave_time)
|
||||
{
|
||||
found = true;
|
||||
best_op_id = i;
|
||||
best_op_name = op_name;
|
||||
best_ave_time = ave_time;
|
||||
best_gb_per_sec = gb_per_sec;
|
||||
}
|
||||
}
|
||||
else
|
||||
{
|
||||
std::cout << op_name << " does not support this problem" << std::endl;
|
||||
}
|
||||
}
|
||||
|
||||
// run the best intance
|
||||
if(found)
|
||||
{
|
||||
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
|
||||
<< best_op_name << std::endl;
|
||||
|
||||
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(
|
||||
static_cast<DOutDataType*>(dout_device_buf.GetDeviceBuffer()),
|
||||
static_cast<DInDataType*>(din_device_buf.GetDeviceBuffer()),
|
||||
out_length,
|
||||
in_length,
|
||||
out_tensor_stride,
|
||||
in_tensor_stride,
|
||||
window_spatial_lengths,
|
||||
window_strides,
|
||||
window_dilations,
|
||||
input_left_pads,
|
||||
input_right_pads);
|
||||
|
||||
auto invoker_ptr = op_ptr->MakeInvokerPointer();
|
||||
|
||||
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
|
||||
{
|
||||
din_device_buf.SetZero();
|
||||
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
|
||||
}
|
||||
|
||||
std::cout << "Done" << std::endl;
|
||||
}
|
||||
|
||||
return 0;
|
||||
}
|
||||
280
client_example/19_pool/max_pool2d_bwd.cpp
Normal file
280
client_example/19_pool/max_pool2d_bwd.cpp
Normal file
@@ -0,0 +1,280 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, 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_pool_fwd.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/device_max_pool_bwd.hpp"
|
||||
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
|
||||
|
||||
#include "ck/library/tensor_operation_instance/gpu/pool3d_fwd.hpp"
|
||||
#include "ck/library/tensor_operation_instance/gpu/max_pool_bwd.hpp"
|
||||
|
||||
using InDataType = ck::half_t;
|
||||
using OutDataType = ck::half_t;
|
||||
using DOutDataType = ck::half_t;
|
||||
using DInDataType = ck::half_t;
|
||||
using IndexDataType = int32_t;
|
||||
|
||||
// We use pool3d to implement pool2d in this example
|
||||
using InLayout = ck::tensor_layout::convolution::NDHWC;
|
||||
using OutLayout = ck::tensor_layout::convolution::NDHWC;
|
||||
|
||||
constexpr ck::index_t InOutRank = 5;
|
||||
constexpr ck::index_t WindowRank = 3;
|
||||
|
||||
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_;
|
||||
};
|
||||
|
||||
void TransformPool2dparamToPool3d(std::vector<ck::index_t>& input_lengths,
|
||||
std::vector<ck::index_t>& window_lengths,
|
||||
std::vector<ck::index_t>& output_lengths,
|
||||
std::vector<ck::index_t>& input_stride,
|
||||
std::vector<ck::index_t>& output_stride,
|
||||
std::vector<ck::index_t>& indices_stride,
|
||||
std::vector<ck::index_t>& window_strides,
|
||||
std::vector<ck::index_t>& window_dilations,
|
||||
std::vector<ck::index_t>& input_left_pads,
|
||||
std::vector<ck::index_t>& input_right_pads,
|
||||
std::vector<ck::index_t>& pooling_dims)
|
||||
{
|
||||
// NCHW to NCDHW
|
||||
input_lengths.insert(input_lengths.begin() + 2, 1);
|
||||
output_lengths.insert(output_lengths.begin() + 2, 1);
|
||||
input_stride.insert(input_stride.begin() + 2, 0);
|
||||
output_stride.insert(output_stride.begin() + 2, 0);
|
||||
indices_stride.insert(indices_stride.begin() + 2, 0);
|
||||
|
||||
// YX to ZYX
|
||||
window_lengths.insert(window_lengths.begin(), 1);
|
||||
window_strides.insert(window_strides.begin(), 0);
|
||||
window_dilations.insert(window_dilations.begin(), 0);
|
||||
input_left_pads.insert(input_left_pads.begin(), 0);
|
||||
input_right_pads.insert(input_right_pads.begin(), 0);
|
||||
|
||||
pooling_dims = {2, 3, 4};
|
||||
}
|
||||
|
||||
int main(int argc, char* argv[])
|
||||
{
|
||||
ck::index_t N = 2;
|
||||
ck::index_t C = 32;
|
||||
ck::index_t Y = 2;
|
||||
ck::index_t X = 2;
|
||||
ck::index_t Hi = 30;
|
||||
ck::index_t Wi = 30;
|
||||
ck::index_t window_stride_h = 2;
|
||||
ck::index_t window_stride_w = 2;
|
||||
ck::index_t window_dilation_h = 1;
|
||||
ck::index_t window_dilation_w = 1;
|
||||
ck::index_t in_left_pad_h = 1;
|
||||
ck::index_t in_left_pad_w = 1;
|
||||
ck::index_t in_right_pad_h = 1;
|
||||
ck::index_t in_right_pad_w = 1;
|
||||
|
||||
const ck::index_t Ys = (Y - 1) * window_dilation_h + 1;
|
||||
const ck::index_t Xs = (X - 1) * window_dilation_w + 1;
|
||||
ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - Ys) / window_stride_h + 1;
|
||||
ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - Xs) / window_stride_w + 1;
|
||||
|
||||
// Pool API only support the order of NCHW
|
||||
std::vector<ck::index_t> in_length = {N, C, Hi, Wi};
|
||||
std::vector<ck::index_t> out_length = {N, C, Ho, Wo};
|
||||
std::vector<ck::index_t> window_spatial_lengths = {Y, X};
|
||||
std::vector<ck::index_t> window_strides = {window_stride_h, window_stride_w};
|
||||
std::vector<ck::index_t> window_dilations = {window_dilation_h, window_dilation_w};
|
||||
std::vector<ck::index_t> input_left_pads = {in_left_pad_h, in_left_pad_w};
|
||||
std::vector<ck::index_t> input_right_pads = {in_right_pad_h, in_right_pad_w};
|
||||
std::vector<ck::index_t> pooling_dims = {2, 3};
|
||||
|
||||
std::size_t in_tensor_size = N * C * Hi * Wi;
|
||||
std::size_t out_tensor_size = N * C * Ho * Wo;
|
||||
|
||||
// tensor layout = NHWC
|
||||
std::vector<ck::index_t> in_tensor_stride = {C * Hi * Wi, 1, Wi * C, C};
|
||||
std::vector<ck::index_t> out_tensor_stride = {C * Ho * Wo, 1, Wo * C, C};
|
||||
|
||||
TransformPool2dparamToPool3d(in_length,
|
||||
window_spatial_lengths,
|
||||
out_length,
|
||||
in_tensor_stride,
|
||||
out_tensor_stride,
|
||||
out_tensor_stride,
|
||||
window_strides,
|
||||
window_dilations,
|
||||
input_left_pads,
|
||||
input_right_pads,
|
||||
pooling_dims);
|
||||
|
||||
SimpleDeviceMem in_device_buf(sizeof(InDataType) * in_tensor_size);
|
||||
SimpleDeviceMem out_device_buf(sizeof(OutDataType) * out_tensor_size);
|
||||
SimpleDeviceMem indices_device_buf(sizeof(IndexDataType) * out_tensor_size);
|
||||
SimpleDeviceMem dout_device_buf(sizeof(DOutDataType) * out_tensor_size);
|
||||
SimpleDeviceMem din_device_buf(sizeof(DInDataType) * in_tensor_size);
|
||||
|
||||
// Generate index data from max pool forward
|
||||
{
|
||||
using MaxPoolFwdDeviceOp =
|
||||
ck::tensor_operation::device::DevicePoolFwd<InOutRank,
|
||||
WindowRank,
|
||||
InDataType,
|
||||
OutDataType,
|
||||
IndexDataType,
|
||||
InLayout,
|
||||
OutLayout,
|
||||
ck::ReduceTensorOp::MAX,
|
||||
true>;
|
||||
|
||||
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
|
||||
MaxPoolFwdDeviceOp>::GetInstances();
|
||||
|
||||
auto& op_ptr = op_ptrs[0];
|
||||
|
||||
auto argument_ptr = op_ptr->MakeArgumentPointer(
|
||||
static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
|
||||
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
|
||||
static_cast<IndexDataType*>(indices_device_buf.GetDeviceBuffer()),
|
||||
in_length,
|
||||
window_spatial_lengths,
|
||||
out_length,
|
||||
in_tensor_stride,
|
||||
out_tensor_stride,
|
||||
out_tensor_stride,
|
||||
window_strides,
|
||||
window_dilations,
|
||||
input_left_pads,
|
||||
input_right_pads,
|
||||
pooling_dims);
|
||||
|
||||
auto invoker_ptr = op_ptr->MakeInvokerPointer();
|
||||
|
||||
std::string op_name = op_ptr->GetTypeString();
|
||||
|
||||
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
|
||||
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
|
||||
}
|
||||
|
||||
// Run MaxPool bwd
|
||||
using MaxPoolBwdDeviceOp =
|
||||
ck::tensor_operation::device::DeviceMaxPoolBwd<DOutDataType, IndexDataType, DInDataType>;
|
||||
|
||||
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
|
||||
MaxPoolBwdDeviceOp>::GetInstances();
|
||||
|
||||
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
|
||||
|
||||
std::string best_op_name;
|
||||
bool found = false;
|
||||
int best_op_id = -1;
|
||||
float best_ave_time = std::numeric_limits<float>::max();
|
||||
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(
|
||||
static_cast<InDataType*>(dout_device_buf.GetDeviceBuffer()),
|
||||
static_cast<IndexDataType*>(indices_device_buf.GetDeviceBuffer()),
|
||||
static_cast<DInDataType*>(din_device_buf.GetDeviceBuffer()),
|
||||
out_tensor_size,
|
||||
in_tensor_size,
|
||||
window_spatial_lengths,
|
||||
window_strides,
|
||||
window_dilations);
|
||||
|
||||
auto invoker_ptr = op_ptr->MakeInvokerPointer();
|
||||
|
||||
std::string op_name = op_ptr->GetTypeString();
|
||||
|
||||
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
|
||||
{
|
||||
size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
|
||||
|
||||
SimpleDeviceMem workspace(workspace_sz);
|
||||
|
||||
op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
|
||||
|
||||
float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
|
||||
|
||||
std::size_t num_bytes = in_tensor_size * sizeof(DInDataType) +
|
||||
out_tensor_size * sizeof(IndexDataType) +
|
||||
out_tensor_size * sizeof(DOutDataType);
|
||||
|
||||
float gb_per_sec = num_bytes / 1.E6 / ave_time;
|
||||
|
||||
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec << "GB / s,"
|
||||
<< op_name << std::endl;
|
||||
|
||||
if(ave_time < best_ave_time)
|
||||
{
|
||||
found = true;
|
||||
best_op_id = i;
|
||||
best_op_name = op_name;
|
||||
best_ave_time = ave_time;
|
||||
best_gb_per_sec = gb_per_sec;
|
||||
}
|
||||
}
|
||||
else
|
||||
{
|
||||
std::cout << op_name << " does not support this problem" << std::endl;
|
||||
}
|
||||
}
|
||||
|
||||
// run the best intance
|
||||
if(found)
|
||||
{
|
||||
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
|
||||
<< best_op_name << std::endl;
|
||||
|
||||
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(
|
||||
static_cast<InDataType*>(dout_device_buf.GetDeviceBuffer()),
|
||||
static_cast<IndexDataType*>(indices_device_buf.GetDeviceBuffer()),
|
||||
static_cast<DInDataType*>(din_device_buf.GetDeviceBuffer()),
|
||||
out_tensor_size,
|
||||
in_tensor_size,
|
||||
window_spatial_lengths,
|
||||
window_strides,
|
||||
window_dilations);
|
||||
|
||||
auto invoker_ptr = op_ptr->MakeInvokerPointer();
|
||||
|
||||
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
|
||||
{
|
||||
size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
|
||||
|
||||
SimpleDeviceMem workspace(workspace_sz);
|
||||
|
||||
op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
|
||||
|
||||
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
|
||||
}
|
||||
|
||||
std::cout << "Done" << std::endl;
|
||||
}
|
||||
|
||||
return 0;
|
||||
}
|
||||
@@ -1,5 +0,0 @@
|
||||
add_executable(client_max_pool2d_fwd max_pool2d_fwd.cpp)
|
||||
target_link_libraries(client_max_pool2d_fwd PRIVATE composable_kernel::device_operations)
|
||||
|
||||
add_executable(client_avg_pool3d_fwd avg_pool3d_fwd.cpp)
|
||||
target_link_libraries(client_avg_pool3d_fwd PRIVATE composable_kernel::device_operations)
|
||||
Reference in New Issue
Block a user