mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-21 13:29:20 +00:00
Pool3d fwd (#697)
* Expand the base class of pool2d, prepare to share base class with pool3d
* Add pool3d device op
* Add pool3d f16 example
* Refactor the base class. implement generic pooling in the future
* clang format
* get original index in max pooling
* Add outputindex to base class
* Fix dimension
* Add pooling instance
* Use indexType instead
* Remove useless header
* Extract IndexDataType to template
* Extract pooling reference code
* clang format
* clang format
* Fix typo
* Add tensor stride
* Add missing header
* Add index stride and output stride
* Refine naming
* Add type to base class
* Rename file
* Use proper size
* Fix typo
* Refine naming
* Modify the argument into vector.
* Add max pool profiler
* Refine naming
* Support f32 pool
* Fix typo
* Add avg pool2d fwd in profiler
* clang format
* Rename AccDatatype to ComputeDatatype
* Fix init
* test pool
* Extract variable
* Add client example
* Check the pooling dim
* clang format
* Connect argv and arg_parser
* Add found check
* Remove useless header
* Refine naming
* Adjust the order of device_pool_fwd
[ROCm/composable_kernel commit: 76ec0089fb]
This commit is contained in:
rocking
264
profiler/include/profiler/profile_pool2d_fwd_impl.hpp
Normal file
264
profiler/include/profiler/profile_pool2d_fwd_impl.hpp
Normal file
@@ -0,0 +1,264 @@
|
||||
// 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/library/tensor_operation_instance/gpu/pool2d_fwd.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/literals.hpp"
|
||||
#include "ck/library/reference_tensor_operation/cpu/reference_pool_fwd.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace profiler {
|
||||
|
||||
template <typename InDataType,
|
||||
typename OutDataType,
|
||||
typename ComputeDataType,
|
||||
typename IndexDataType,
|
||||
ck::ReduceTensorOp ReduceOpId,
|
||||
bool PropagateNan,
|
||||
bool OutputIndex>
|
||||
bool profile_pool2d_fwd_impl(int do_verification,
|
||||
int init_method,
|
||||
bool do_log,
|
||||
bool time_kernel,
|
||||
std::vector<index_t> in_length, // NCHW
|
||||
std::vector<index_t> window_spatial_lengths,
|
||||
std::vector<index_t> window_strides,
|
||||
std::vector<index_t> input_left_pads,
|
||||
std::vector<index_t> input_right_pads)
|
||||
{
|
||||
constexpr index_t InOutRank = 4;
|
||||
constexpr index_t WindowRank = 2;
|
||||
|
||||
if(in_length.size() != InOutRank || window_spatial_lengths.size() != WindowRank ||
|
||||
window_strides.size() != WindowRank || input_left_pads.size() != WindowRank ||
|
||||
input_right_pads.size() != WindowRank)
|
||||
return false;
|
||||
|
||||
std::vector<index_t> out_length(InOutRank);
|
||||
|
||||
int N = in_length[0];
|
||||
int C = in_length[1];
|
||||
|
||||
out_length[0] = N;
|
||||
out_length[1] = C;
|
||||
|
||||
// Calculate Ho, Wo
|
||||
for(int i = 2; i < InOutRank; ++i)
|
||||
{
|
||||
auto pad1 = input_left_pads[i - 2];
|
||||
auto pad2 = input_right_pads[i - 2];
|
||||
auto windows_size = window_spatial_lengths[i - 2];
|
||||
auto windows_stride = window_strides[i - 2];
|
||||
out_length[i] = (in_length[i] + pad1 + pad2 - windows_size) / windows_stride + 1;
|
||||
}
|
||||
|
||||
int Hi = in_length[2];
|
||||
int Wi = in_length[3];
|
||||
int Ho = out_length[2];
|
||||
int Wo = out_length[3];
|
||||
|
||||
auto f_host_tensor_descriptor =
|
||||
[](std::size_t N_, std::size_t C_, std::size_t H, std::size_t W) {
|
||||
using namespace ck::literals;
|
||||
return HostTensorDescriptor({N_, C_, H, W}, {C_ * H * W, 1_uz, W * C_, C_});
|
||||
};
|
||||
|
||||
Tensor<InDataType> in_n_c_hi_wi(f_host_tensor_descriptor(N, C, Hi, Wi));
|
||||
Tensor<OutDataType> out_n_c_ho_wo_host(f_host_tensor_descriptor(N, C, Ho, Wo));
|
||||
Tensor<IndexDataType> out_indices_n_c_ho_wo_host(f_host_tensor_descriptor(N, C, Ho, Wo));
|
||||
|
||||
Tensor<OutDataType> out_n_c_ho_wo_device(f_host_tensor_descriptor(N, C, Ho, Wo));
|
||||
Tensor<IndexDataType> out_indices_n_c_ho_wo_device(f_host_tensor_descriptor(N, C, Ho, Wo));
|
||||
|
||||
switch(init_method)
|
||||
{
|
||||
case 0: in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_1<InDataType>{}); break;
|
||||
case 1: in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5}); break;
|
||||
default: in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_3<InDataType>{-0.5, 0.5});
|
||||
}
|
||||
|
||||
DeviceMem in_device_buf(sizeof(InDataType) * in_n_c_hi_wi.mDesc.GetElementSpaceSize());
|
||||
DeviceMem out_device_buf(sizeof(OutDataType) *
|
||||
out_n_c_ho_wo_device.mDesc.GetElementSpaceSize());
|
||||
DeviceMem out_indices_device_buf(sizeof(IndexDataType) *
|
||||
out_indices_n_c_ho_wo_device.mDesc.GetElementSpaceSize());
|
||||
|
||||
in_device_buf.ToDevice(in_n_c_hi_wi.mData.data());
|
||||
|
||||
// add device normalization instances
|
||||
using DeviceOp = ck::tensor_operation::device::DevicePoolFwd<InOutRank,
|
||||
WindowRank,
|
||||
InDataType,
|
||||
OutDataType,
|
||||
IndexDataType,
|
||||
ReduceOpId,
|
||||
OutputIndex>;
|
||||
|
||||
// get device op instances
|
||||
const auto instance_ptrs =
|
||||
ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
|
||||
DeviceOp>::GetInstances();
|
||||
|
||||
std::cout << "found " << instance_ptrs.size() << " instances" << std::endl;
|
||||
|
||||
std::string best_instance_name;
|
||||
float best_avg_time = std::numeric_limits<float>::max();
|
||||
float best_gb_per_sec = 0;
|
||||
|
||||
if(do_verification)
|
||||
{
|
||||
using ReferenceInstance = ck::tensor_operation::host::ReferencePoolingFwd<InOutRank,
|
||||
WindowRank,
|
||||
InDataType,
|
||||
OutDataType,
|
||||
ComputeDataType,
|
||||
IndexDataType,
|
||||
ReduceOpId,
|
||||
PropagateNan,
|
||||
OutputIndex>;
|
||||
|
||||
ReferenceInstance ref;
|
||||
auto ref_argument = ref.MakeArgument(in_n_c_hi_wi,
|
||||
out_n_c_ho_wo_host,
|
||||
out_indices_n_c_ho_wo_host,
|
||||
window_spatial_lengths,
|
||||
window_strides,
|
||||
input_left_pads,
|
||||
input_right_pads);
|
||||
auto ref_invoker = ref.MakeInvoker();
|
||||
ref_invoker.Run(ref_argument);
|
||||
}
|
||||
|
||||
int num_kernel = 0;
|
||||
|
||||
for(auto& inst_ptr : instance_ptrs)
|
||||
{
|
||||
auto argument_ptr = inst_ptr->MakeArgumentPointer(
|
||||
static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
|
||||
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
|
||||
static_cast<IndexDataType*>(out_indices_device_buf.GetDeviceBuffer()),
|
||||
in_length,
|
||||
window_spatial_lengths,
|
||||
out_length,
|
||||
{C * Hi * Wi, 1, Wi * C, C},
|
||||
{C * Ho * Wo, 1, Wo * C, C},
|
||||
{C * Ho * Wo, 1, Wo * C, C},
|
||||
window_strides,
|
||||
input_left_pads,
|
||||
input_right_pads,
|
||||
{2, 3});
|
||||
|
||||
if(inst_ptr->IsSupportedArgument(argument_ptr.get()))
|
||||
{
|
||||
++num_kernel;
|
||||
}
|
||||
else
|
||||
{
|
||||
if(time_kernel)
|
||||
{
|
||||
std::cout << inst_ptr->GetTypeString() << " skipped due to unsupported argument: ";
|
||||
LogRange(std::cout << "input lengths = ", in_length, ", ") << std::endl;
|
||||
}
|
||||
|
||||
continue;
|
||||
}
|
||||
|
||||
auto invoker_ptr = inst_ptr->MakeInvokerPointer();
|
||||
|
||||
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
|
||||
|
||||
std::size_t num_bytes = in_n_c_hi_wi.mDesc.GetElementSize() * sizeof(InDataType) +
|
||||
out_n_c_ho_wo_host.mDesc.GetElementSize() * sizeof(OutDataType);
|
||||
|
||||
if constexpr(OutputIndex)
|
||||
num_bytes += out_indices_n_c_ho_wo_host.mDesc.GetElementSize() * sizeof(IndexDataType);
|
||||
|
||||
float gb_per_sec = num_bytes / 1.E6 / avg_time;
|
||||
|
||||
if(time_kernel)
|
||||
std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << gb_per_sec << " GB/s, "
|
||||
<< inst_ptr->GetTypeString() << std::endl;
|
||||
|
||||
if(avg_time < best_avg_time)
|
||||
{
|
||||
best_instance_name = inst_ptr->GetTypeString();
|
||||
best_avg_time = avg_time;
|
||||
best_gb_per_sec = gb_per_sec;
|
||||
}
|
||||
|
||||
if(do_verification)
|
||||
{
|
||||
out_device_buf.FromDevice(out_n_c_ho_wo_device.mData.data());
|
||||
|
||||
bool pass = ck::utils::check_err(out_n_c_ho_wo_device.mData,
|
||||
out_n_c_ho_wo_host.mData,
|
||||
"Error: Incorrect results",
|
||||
1e-3,
|
||||
1e-3);
|
||||
|
||||
if constexpr(OutputIndex)
|
||||
{
|
||||
out_indices_device_buf.FromDevice(out_indices_n_c_ho_wo_device.mData.data());
|
||||
|
||||
pass = pass && ck::utils::check_err(out_indices_n_c_ho_wo_device,
|
||||
out_indices_n_c_ho_wo_host);
|
||||
}
|
||||
|
||||
if(do_log)
|
||||
{
|
||||
LogRangeAsType<float>(std::cout << "in_n_c_hi_wi : ", in_n_c_hi_wi.mData, ",")
|
||||
<< std::endl;
|
||||
LogRangeAsType<float>(
|
||||
std::cout << "out_n_c_ho_wo_host : ", out_n_c_ho_wo_host.mData, ",")
|
||||
<< std::endl;
|
||||
LogRangeAsType<float>(
|
||||
std::cout << "out_n_c_ho_wo_device : ", out_n_c_ho_wo_device.mData, ",")
|
||||
<< std::endl;
|
||||
|
||||
if constexpr(OutputIndex)
|
||||
LogRangeAsType<float>(std::cout << "out_indices_n_c_ho_wo_device : ",
|
||||
out_indices_n_c_ho_wo_device.mData,
|
||||
",")
|
||||
<< std::endl;
|
||||
}
|
||||
|
||||
if(!pass)
|
||||
{
|
||||
std::cout << inst_ptr->GetTypeString() << " failed verification: ";
|
||||
LogRange(std::cout << "lengths = [", in_length, ", ") << "]." << std::endl;
|
||||
return false;
|
||||
}
|
||||
else
|
||||
{
|
||||
if(time_kernel)
|
||||
std::cout << "pass" << std::endl;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
if(time_kernel)
|
||||
{
|
||||
LogRange(std::cout << "length = ", in_length, ",") << std::endl;
|
||||
std::cout << "best perf = " << best_avg_time << " ms, " << best_gb_per_sec << " GB/s, "
|
||||
<< best_instance_name << std::endl;
|
||||
}
|
||||
|
||||
if(num_kernel == 0)
|
||||
{
|
||||
std::cout << "Error: No kernel is applicable" << std::endl;
|
||||
return false;
|
||||
}
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
} // namespace profiler
|
||||
} // namespace ck
|
||||
271
profiler/include/profiler/profile_pool3d_fwd_impl.hpp
Normal file
271
profiler/include/profiler/profile_pool3d_fwd_impl.hpp
Normal file
@@ -0,0 +1,271 @@
|
||||
// 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/library/tensor_operation_instance/gpu/pool3d_fwd.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/literals.hpp"
|
||||
#include "ck/library/reference_tensor_operation/cpu/reference_pool_fwd.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace profiler {
|
||||
|
||||
template <typename InDataType,
|
||||
typename OutDataType,
|
||||
typename ComputeDataType,
|
||||
typename IndexDataType,
|
||||
ck::ReduceTensorOp ReduceOpId,
|
||||
bool PropagateNan,
|
||||
bool OutputIndex>
|
||||
bool profile_pool3d_fwd_impl(int do_verification,
|
||||
int init_method,
|
||||
bool do_log,
|
||||
bool time_kernel,
|
||||
std::vector<index_t> in_length, // NCDHW
|
||||
std::vector<index_t> window_spatial_lengths,
|
||||
std::vector<index_t> window_strides,
|
||||
std::vector<index_t> input_left_pads,
|
||||
std::vector<index_t> input_right_pads)
|
||||
{
|
||||
constexpr index_t InOutRank = 5;
|
||||
constexpr index_t WindowRank = 3;
|
||||
|
||||
if(in_length.size() != InOutRank || window_spatial_lengths.size() != WindowRank ||
|
||||
window_strides.size() != WindowRank || input_left_pads.size() != WindowRank ||
|
||||
input_right_pads.size() != WindowRank)
|
||||
return false;
|
||||
|
||||
std::vector<index_t> out_length(InOutRank);
|
||||
|
||||
int N = in_length[0];
|
||||
int C = in_length[1];
|
||||
|
||||
out_length[0] = N;
|
||||
out_length[1] = C;
|
||||
|
||||
// Calculate Do, Ho, Wo
|
||||
for(int i = 2; i < InOutRank; ++i)
|
||||
{
|
||||
auto pad1 = input_left_pads[i - 2];
|
||||
auto pad2 = input_right_pads[i - 2];
|
||||
auto windows_size = window_spatial_lengths[i - 2];
|
||||
auto windows_stride = window_strides[i - 2];
|
||||
out_length[i] = (in_length[i] + pad1 + pad2 - windows_size) / windows_stride + 1;
|
||||
}
|
||||
|
||||
int Di = in_length[2];
|
||||
int Hi = in_length[3];
|
||||
int Wi = in_length[4];
|
||||
int Do = out_length[2];
|
||||
int Ho = out_length[3];
|
||||
int Wo = out_length[4];
|
||||
|
||||
auto f_host_tensor_descriptor =
|
||||
[](std::size_t N_, std::size_t C_, std::size_t D, std::size_t H, std::size_t W) {
|
||||
using namespace ck::literals;
|
||||
|
||||
return HostTensorDescriptor({N_, C_, D, H, W},
|
||||
{D * C_ * H * W, 1_uz, C_ * H * W, W * C_, C_});
|
||||
};
|
||||
|
||||
Tensor<InDataType> in_n_c_di_hi_wi(f_host_tensor_descriptor(N, C, Di, Hi, Wi));
|
||||
Tensor<OutDataType> out_n_c_do_ho_wo_host(f_host_tensor_descriptor(N, C, Do, Ho, Wo));
|
||||
Tensor<IndexDataType> out_indices_n_c_do_ho_wo_host(f_host_tensor_descriptor(N, C, Do, Ho, Wo));
|
||||
|
||||
Tensor<OutDataType> out_n_c_do_ho_wo_device(f_host_tensor_descriptor(N, C, Do, Ho, Wo));
|
||||
Tensor<IndexDataType> out_indices_n_c_do_ho_wo_device(
|
||||
f_host_tensor_descriptor(N, C, Do, Ho, Wo));
|
||||
|
||||
switch(init_method)
|
||||
{
|
||||
case 0: in_n_c_di_hi_wi.GenerateTensorValue(GeneratorTensor_1<InDataType>{}); break;
|
||||
case 1: in_n_c_di_hi_wi.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5}); break;
|
||||
default: in_n_c_di_hi_wi.GenerateTensorValue(GeneratorTensor_3<InDataType>{-0.5, 0.5});
|
||||
}
|
||||
|
||||
DeviceMem in_device_buf(sizeof(InDataType) * in_n_c_di_hi_wi.mDesc.GetElementSpaceSize());
|
||||
DeviceMem out_device_buf(sizeof(OutDataType) *
|
||||
out_n_c_do_ho_wo_device.mDesc.GetElementSpaceSize());
|
||||
DeviceMem out_indices_device_buf(sizeof(IndexDataType) *
|
||||
out_indices_n_c_do_ho_wo_device.mDesc.GetElementSpaceSize());
|
||||
|
||||
in_device_buf.ToDevice(in_n_c_di_hi_wi.mData.data());
|
||||
|
||||
// add device normalization instances
|
||||
using DeviceOp = ck::tensor_operation::device::DevicePoolFwd<InOutRank,
|
||||
WindowRank,
|
||||
InDataType,
|
||||
OutDataType,
|
||||
IndexDataType,
|
||||
ReduceOpId,
|
||||
OutputIndex>;
|
||||
|
||||
// get device op instances
|
||||
const auto instance_ptrs =
|
||||
ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
|
||||
DeviceOp>::GetInstances();
|
||||
|
||||
std::cout << "found " << instance_ptrs.size() << " instances" << std::endl;
|
||||
|
||||
std::string best_instance_name;
|
||||
float best_avg_time = std::numeric_limits<float>::max();
|
||||
float best_gb_per_sec = 0;
|
||||
|
||||
if(do_verification)
|
||||
{
|
||||
using ReferenceInstance = ck::tensor_operation::host::ReferencePoolingFwd<InOutRank,
|
||||
WindowRank,
|
||||
InDataType,
|
||||
OutDataType,
|
||||
ComputeDataType,
|
||||
IndexDataType,
|
||||
ReduceOpId,
|
||||
PropagateNan,
|
||||
OutputIndex>;
|
||||
|
||||
ReferenceInstance ref;
|
||||
auto ref_argument = ref.MakeArgument(in_n_c_di_hi_wi,
|
||||
out_n_c_do_ho_wo_host,
|
||||
out_indices_n_c_do_ho_wo_host,
|
||||
window_spatial_lengths,
|
||||
window_strides,
|
||||
input_left_pads,
|
||||
input_right_pads);
|
||||
auto ref_invoker = ref.MakeInvoker();
|
||||
ref_invoker.Run(ref_argument);
|
||||
}
|
||||
|
||||
int num_kernel = 0;
|
||||
|
||||
for(auto& inst_ptr : instance_ptrs)
|
||||
{
|
||||
auto argument_ptr = inst_ptr->MakeArgumentPointer(
|
||||
static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
|
||||
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
|
||||
static_cast<IndexDataType*>(out_indices_device_buf.GetDeviceBuffer()),
|
||||
in_length,
|
||||
window_spatial_lengths,
|
||||
out_length,
|
||||
{Di * C * Hi * Wi, 1, C * Hi * Wi, Wi * C, C},
|
||||
{Do * C * Ho * Wo, 1, C * Ho * Wo, Wo * C, C},
|
||||
{Do * C * Ho * Wo, 1, C * Ho * Wo, Wo * C, C},
|
||||
window_strides,
|
||||
input_left_pads,
|
||||
input_right_pads,
|
||||
{2, 3, 4});
|
||||
|
||||
if(inst_ptr->IsSupportedArgument(argument_ptr.get()))
|
||||
{
|
||||
++num_kernel;
|
||||
}
|
||||
else
|
||||
{
|
||||
if(time_kernel)
|
||||
{
|
||||
std::cout << inst_ptr->GetTypeString() << " skipped due to unsupported argument: ";
|
||||
LogRange(std::cout << "input lengths = ", in_length, ", ") << std::endl;
|
||||
}
|
||||
|
||||
continue;
|
||||
}
|
||||
|
||||
auto invoker_ptr = inst_ptr->MakeInvokerPointer();
|
||||
|
||||
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
|
||||
|
||||
std::size_t num_bytes = in_n_c_di_hi_wi.mDesc.GetElementSize() * sizeof(InDataType) +
|
||||
out_n_c_do_ho_wo_host.mDesc.GetElementSize() * sizeof(OutDataType);
|
||||
|
||||
if constexpr(OutputIndex)
|
||||
num_bytes +=
|
||||
out_indices_n_c_do_ho_wo_host.mDesc.GetElementSize() * sizeof(IndexDataType);
|
||||
|
||||
float gb_per_sec = num_bytes / 1.E6 / avg_time;
|
||||
|
||||
if(time_kernel)
|
||||
std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << gb_per_sec << " GB/s, "
|
||||
<< inst_ptr->GetTypeString() << std::endl;
|
||||
|
||||
if(avg_time < best_avg_time)
|
||||
{
|
||||
best_instance_name = inst_ptr->GetTypeString();
|
||||
best_avg_time = avg_time;
|
||||
best_gb_per_sec = gb_per_sec;
|
||||
}
|
||||
|
||||
if(do_verification)
|
||||
{
|
||||
out_device_buf.FromDevice(out_n_c_do_ho_wo_device.mData.data());
|
||||
|
||||
bool pass = ck::utils::check_err(out_n_c_do_ho_wo_device.mData,
|
||||
out_n_c_do_ho_wo_host.mData,
|
||||
"Error: Incorrect results",
|
||||
1e-3,
|
||||
1e-3);
|
||||
|
||||
if constexpr(OutputIndex)
|
||||
{
|
||||
out_indices_device_buf.FromDevice(out_indices_n_c_do_ho_wo_device.mData.data());
|
||||
|
||||
pass = pass && ck::utils::check_err(out_indices_n_c_do_ho_wo_device,
|
||||
out_indices_n_c_do_ho_wo_host);
|
||||
}
|
||||
|
||||
if(do_log)
|
||||
{
|
||||
LogRangeAsType<float>(
|
||||
std::cout << "in_n_c_di_hi_wi : ", in_n_c_di_hi_wi.mData, ",")
|
||||
<< std::endl;
|
||||
LogRangeAsType<float>(
|
||||
std::cout << "out_n_c_do_ho_wo_host : ", out_n_c_do_ho_wo_host.mData, ",")
|
||||
<< std::endl;
|
||||
LogRangeAsType<float>(
|
||||
std::cout << "out_n_c_do_ho_wo_device : ", out_n_c_do_ho_wo_device.mData, ",")
|
||||
<< std::endl;
|
||||
|
||||
if constexpr(OutputIndex)
|
||||
LogRangeAsType<float>(std::cout << "out_indices_n_c_do_ho_wo_device : ",
|
||||
out_indices_n_c_do_ho_wo_device.mData,
|
||||
",")
|
||||
<< std::endl;
|
||||
}
|
||||
|
||||
if(!pass)
|
||||
{
|
||||
std::cout << inst_ptr->GetTypeString() << " failed verification: ";
|
||||
LogRange(std::cout << "lengths = [", in_length, ", ") << "]." << std::endl;
|
||||
return false;
|
||||
}
|
||||
else
|
||||
{
|
||||
if(time_kernel)
|
||||
std::cout << "pass" << std::endl;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
if(time_kernel)
|
||||
{
|
||||
LogRange(std::cout << "length = ", in_length, ",") << std::endl;
|
||||
std::cout << "best perf = " << best_avg_time << " ms, " << best_gb_per_sec << " GB/s, "
|
||||
<< best_instance_name << std::endl;
|
||||
}
|
||||
|
||||
if(num_kernel == 0)
|
||||
{
|
||||
std::cout << "Error: No kernel is applicable" << std::endl;
|
||||
return false;
|
||||
}
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
} // namespace profiler
|
||||
} // namespace ck
|
||||
@@ -25,6 +25,8 @@ set(PROFILER_SOURCES
|
||||
profile_reduce.cpp
|
||||
profile_groupnorm.cpp
|
||||
profile_layernorm.cpp
|
||||
profile_avg_pool2d_fwd.cpp
|
||||
profile_max_pool3d_fwd.cpp
|
||||
profile_softmax.cpp
|
||||
profile_batchnorm_fwd.cpp
|
||||
profile_batchnorm_bwd.cpp
|
||||
@@ -74,4 +76,6 @@ target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batchnorm_instance)
|
||||
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fastgelu_instance)
|
||||
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_bilinear_instance)
|
||||
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_scale_instance)
|
||||
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool_fwd_instance)
|
||||
|
||||
rocm_install(TARGETS ${PROFILER_EXECUTABLE} COMPONENT profiler)
|
||||
|
||||
141
profiler/src/profile_avg_pool2d_fwd.cpp
Normal file
141
profiler/src/profile_avg_pool2d_fwd.cpp
Normal file
@@ -0,0 +1,141 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#include <iostream>
|
||||
#include <vector>
|
||||
#include <unordered_map>
|
||||
|
||||
#include "profiler/data_type_enum.hpp"
|
||||
#include "profiler/profile_pool2d_fwd_impl.hpp"
|
||||
#include "profiler_operation_registry.hpp"
|
||||
|
||||
using ck::index_t;
|
||||
|
||||
struct avgPoolFwdArgParser
|
||||
{
|
||||
std::unordered_map<std::string, std::vector<int>> long_opts = {
|
||||
{"length", {}}, {"wsize", {}}, {"wstride", {}}, {"pad1", {}}, {"pad2", {}}};
|
||||
|
||||
bool parse_opt(int argc, char* argv[], const std::string& key, int i)
|
||||
{
|
||||
if(std::string("--") + key == argv[i])
|
||||
{
|
||||
int pos = i;
|
||||
while(++i < argc && argv[i][0] != '-') {}
|
||||
int end = i;
|
||||
for(int j = pos + 1; j < end; j++)
|
||||
{
|
||||
long_opts[key].push_back(std::stoi(argv[j]));
|
||||
}
|
||||
return true;
|
||||
}
|
||||
return false;
|
||||
}
|
||||
|
||||
void operator()(int argc, char* argv[])
|
||||
{
|
||||
for(auto& kv : long_opts)
|
||||
{
|
||||
for(int i = 1; i < argc; i++)
|
||||
{
|
||||
if(parse_opt(argc, argv, kv.first, i))
|
||||
break;
|
||||
}
|
||||
}
|
||||
}
|
||||
};
|
||||
|
||||
void print_help_avg_pool2d_fwd()
|
||||
{
|
||||
std::cout << "arg1: data type (0: fp16; 1: fp32)\n"
|
||||
<< "arg2: verification (0: no; 1: yes)\n"
|
||||
<< "arg3: initialization (0: no init; 1: integer value; 2: decimal value)\n"
|
||||
<< "arg4: print tensor value (0: no; 1: yes)\n"
|
||||
<< "arg5: time kernel (0=no, 1=yes)\n"
|
||||
<< "--length: input tensor length for NDHW(e.g, --length 2 32 30 30) \n"
|
||||
<< "--wsize: window size for YX (e.g, --wsize 2 2) \n"
|
||||
<< "--wstride: window stride for HW (e.g, --wstride 2 2) \n"
|
||||
<< "--pad1: left side of padding in HW (e.g, --pad1 1 1) \n"
|
||||
<< "--pad2: right side of padding in HW (e.g, --pad2 1 1) \n"
|
||||
<< "eg: ckProfiler avg_pool2d_fwd 0 1 2 0 1 0 --length 2 32 30 30 --wsize 2 2 "
|
||||
"--wstride 2 2 --pad1 1 1 --pad2 1 1"
|
||||
<< std::endl;
|
||||
}
|
||||
|
||||
int profile_avg_pool2d_fwd(int argc, char* argv[])
|
||||
{
|
||||
ck::DataTypeEnum data_type = ck::DataTypeEnum::Half;
|
||||
bool do_verification = true;
|
||||
int init_method = 0;
|
||||
bool do_log = false;
|
||||
bool time_kernel = true;
|
||||
|
||||
std::vector<index_t> in_length = {2, 32, 30, 30};
|
||||
std::vector<index_t> wsize = {2, 2};
|
||||
std::vector<index_t> wstride = {2, 2};
|
||||
std::vector<index_t> pad1 = {1, 1};
|
||||
std::vector<index_t> pad2 = {1, 1};
|
||||
|
||||
if(argc != 2 && argc != 25)
|
||||
{
|
||||
print_help_avg_pool2d_fwd();
|
||||
return 0;
|
||||
}
|
||||
else if(argc == 25)
|
||||
{
|
||||
data_type = static_cast<ck::DataTypeEnum>(std::stoi(argv[2]));
|
||||
do_verification = std::stoi(argv[3]);
|
||||
init_method = std::stoi(argv[4]);
|
||||
do_log = std::stoi(argv[5]);
|
||||
time_kernel = std::stoi(argv[6]);
|
||||
|
||||
// parse the long options
|
||||
avgPoolFwdArgParser arg_parser;
|
||||
arg_parser(argc, argv);
|
||||
in_length = arg_parser.long_opts["length"];
|
||||
wsize = arg_parser.long_opts["wsize"];
|
||||
wstride = arg_parser.long_opts["wstride"];
|
||||
pad1 = arg_parser.long_opts["pad1"];
|
||||
pad2 = arg_parser.long_opts["pad2"];
|
||||
}
|
||||
|
||||
using F16 = ck::half_t;
|
||||
using F32 = float;
|
||||
using I32 = int32_t;
|
||||
constexpr auto ReduceOpId = ck::ReduceTensorOp::AVG;
|
||||
|
||||
if(data_type == ck::DataTypeEnum::Half)
|
||||
{
|
||||
ck::profiler::profile_pool2d_fwd_impl<F16, F16, F32, I32, ReduceOpId, false, false>(
|
||||
do_verification,
|
||||
init_method,
|
||||
do_log,
|
||||
time_kernel,
|
||||
in_length,
|
||||
wsize,
|
||||
wstride,
|
||||
pad1,
|
||||
pad2);
|
||||
}
|
||||
else if(data_type == ck::DataTypeEnum::Float)
|
||||
{
|
||||
ck::profiler::profile_pool2d_fwd_impl<F32, F32, F32, I32, ReduceOpId, false, false>(
|
||||
do_verification,
|
||||
init_method,
|
||||
do_log,
|
||||
time_kernel,
|
||||
in_length,
|
||||
wsize,
|
||||
wstride,
|
||||
pad1,
|
||||
pad2);
|
||||
}
|
||||
else
|
||||
{
|
||||
throw std::runtime_error("not implemented yet");
|
||||
}
|
||||
|
||||
return 0;
|
||||
}
|
||||
|
||||
REGISTER_PROFILER_OPERATION("avg_pool2d_fwd", "avg_pool2d fwd", profile_avg_pool2d_fwd);
|
||||
@@ -64,7 +64,7 @@ int profile_groupnorm(int argc, char* argv[])
|
||||
ck::DataTypeEnum data_type = ck::DataTypeEnum::Half;
|
||||
bool do_verification = false;
|
||||
int init_method = 0;
|
||||
bool do_log = 0;
|
||||
bool do_log = false;
|
||||
bool time_kernel = 1;
|
||||
std::vector<index_t> length = {64, 16, 16, 32, 40};
|
||||
|
||||
|
||||
168
profiler/src/profile_max_pool3d_fwd.cpp
Normal file
168
profiler/src/profile_max_pool3d_fwd.cpp
Normal file
@@ -0,0 +1,168 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#include <iostream>
|
||||
#include <vector>
|
||||
#include <unordered_map>
|
||||
|
||||
#include "profiler/data_type_enum.hpp"
|
||||
#include "profiler/profile_pool3d_fwd_impl.hpp"
|
||||
#include "profiler_operation_registry.hpp"
|
||||
|
||||
using ck::index_t;
|
||||
|
||||
struct maxPoolFwdArgParser
|
||||
{
|
||||
std::unordered_map<std::string, std::vector<int>> long_opts = {
|
||||
{"length", {}}, {"wsize", {}}, {"wstride", {}}, {"pad1", {}}, {"pad2", {}}};
|
||||
|
||||
bool parse_opt(int argc, char* argv[], const std::string& key, int i)
|
||||
{
|
||||
if(std::string("--") + key == argv[i])
|
||||
{
|
||||
int pos = i;
|
||||
while(++i < argc && argv[i][0] != '-') {}
|
||||
int end = i;
|
||||
for(int j = pos + 1; j < end; j++)
|
||||
{
|
||||
long_opts[key].push_back(std::stoi(argv[j]));
|
||||
}
|
||||
return true;
|
||||
}
|
||||
return false;
|
||||
}
|
||||
|
||||
void operator()(int argc, char* argv[])
|
||||
{
|
||||
for(auto& kv : long_opts)
|
||||
{
|
||||
for(int i = 1; i < argc; i++)
|
||||
{
|
||||
if(parse_opt(argc, argv, kv.first, i))
|
||||
break;
|
||||
}
|
||||
}
|
||||
}
|
||||
};
|
||||
|
||||
void print_help_max_pool3d_fwd()
|
||||
{
|
||||
std::cout << "arg1: data type (0: fp16; 1: fp32)\n"
|
||||
<< "arg2: verification (0: no; 1: yes)\n"
|
||||
<< "arg3: initialization (0: no init; 1: integer value; 2: decimal value)\n"
|
||||
<< "arg4: print tensor value (0: no; 1: yes)\n"
|
||||
<< "arg5: time kernel (0=no, 1=yes)\n"
|
||||
<< "arg6: return index (0=no, 1=yes)\n"
|
||||
<< "--length: input tensor length for NCDHW(e.g, --length 2 32 30 30 30) \n"
|
||||
<< "--wsize: window size for ZYX (e.g, --wsize 2 2 2) \n"
|
||||
<< "--wstride: window stride for DHW (e.g, --wstride 2 2 2) \n"
|
||||
<< "--pad1: left side of padding in DHW (e.g, --pad1 1 1 1) \n"
|
||||
<< "--pad2: right side of padding in DHW (e.g, --pad2 1 1 1) \n"
|
||||
<< "eg: ckProfiler max_pool3d_fwd 0 1 2 0 1 0 --length 2 32 30 30 30 --wsize 2 2 2 "
|
||||
"--wstride 2 2 2 --pad1 1 1 1 --pad2 1 1 1"
|
||||
<< std::endl;
|
||||
}
|
||||
|
||||
int profile_max_pool3d_fwd(int argc, char* argv[])
|
||||
{
|
||||
ck::DataTypeEnum data_type = ck::DataTypeEnum::Half;
|
||||
bool do_verification = true;
|
||||
int init_method = 0;
|
||||
bool do_log = false;
|
||||
bool time_kernel = true;
|
||||
bool return_index = false;
|
||||
|
||||
std::vector<index_t> in_length = {2, 32, 30, 30, 30};
|
||||
std::vector<index_t> wsize = {2, 2, 2};
|
||||
std::vector<index_t> wstride = {2, 2, 2};
|
||||
std::vector<index_t> pad1 = {1, 1, 1};
|
||||
std::vector<index_t> pad2 = {1, 1, 1};
|
||||
|
||||
if(argc != 2 && argc != 30)
|
||||
{
|
||||
print_help_max_pool3d_fwd();
|
||||
return 0;
|
||||
}
|
||||
else if(argc == 30)
|
||||
{
|
||||
data_type = static_cast<ck::DataTypeEnum>(std::stoi(argv[2]));
|
||||
do_verification = std::stoi(argv[3]);
|
||||
init_method = std::stoi(argv[4]);
|
||||
do_log = std::stoi(argv[5]);
|
||||
time_kernel = std::stoi(argv[6]);
|
||||
return_index = std::stoi(argv[7]);
|
||||
|
||||
// parse the long options
|
||||
maxPoolFwdArgParser arg_parser;
|
||||
arg_parser(argc, argv);
|
||||
in_length = arg_parser.long_opts["length"];
|
||||
wsize = arg_parser.long_opts["wsize"];
|
||||
wstride = arg_parser.long_opts["wstride"];
|
||||
pad1 = arg_parser.long_opts["pad1"];
|
||||
pad2 = arg_parser.long_opts["pad2"];
|
||||
}
|
||||
|
||||
using F16 = ck::half_t;
|
||||
using F32 = float;
|
||||
using I32 = int32_t;
|
||||
constexpr auto ReduceOpId = ck::ReduceTensorOp::MAX;
|
||||
|
||||
if(data_type == ck::DataTypeEnum::Half)
|
||||
{
|
||||
if(return_index)
|
||||
ck::profiler::profile_pool3d_fwd_impl<F16, F16, F16, I32, ReduceOpId, false, true>(
|
||||
do_verification,
|
||||
init_method,
|
||||
do_log,
|
||||
time_kernel,
|
||||
in_length,
|
||||
wsize,
|
||||
wstride,
|
||||
pad1,
|
||||
pad2);
|
||||
else
|
||||
ck::profiler::profile_pool3d_fwd_impl<F16, F16, F16, I32, ReduceOpId, false, false>(
|
||||
do_verification,
|
||||
init_method,
|
||||
do_log,
|
||||
time_kernel,
|
||||
in_length,
|
||||
wsize,
|
||||
wstride,
|
||||
pad1,
|
||||
pad2);
|
||||
}
|
||||
else if(data_type == ck::DataTypeEnum::Float)
|
||||
{
|
||||
if(return_index)
|
||||
ck::profiler::profile_pool3d_fwd_impl<F32, F32, F32, I32, ReduceOpId, false, true>(
|
||||
do_verification,
|
||||
init_method,
|
||||
do_log,
|
||||
time_kernel,
|
||||
in_length,
|
||||
wsize,
|
||||
wstride,
|
||||
pad1,
|
||||
pad2);
|
||||
else
|
||||
ck::profiler::profile_pool3d_fwd_impl<F32, F32, F32, I32, ReduceOpId, false, false>(
|
||||
do_verification,
|
||||
init_method,
|
||||
do_log,
|
||||
time_kernel,
|
||||
in_length,
|
||||
wsize,
|
||||
wstride,
|
||||
pad1,
|
||||
pad2);
|
||||
}
|
||||
else
|
||||
{
|
||||
throw std::runtime_error("not implemented yet");
|
||||
}
|
||||
|
||||
return 0;
|
||||
}
|
||||
|
||||
REGISTER_PROFILER_OPERATION("max_pool3d_fwd", "max_pool3d fwd", profile_max_pool3d_fwd);
|
||||
Reference in New Issue
Block a user