Skip to content

GitLab

Commit 2f463a94 authored by carlushuang's avatar carlushuang
Browse files

Merge remote-tracking branch 'origin/develop' into stream-k-initial-impl

parents ca8b5c79 ac9e01e2
Hide whitespace changes
Inline Side-by-side
Showing with 1571 additions and 999 deletions
profiler/include/profiler/profile_convnd_bwd_data_impl.hpp deleted 100644 → 0
View file @ ca8b5c79
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_conv_bwd_data.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/conv_util.hpp"
#include "ck/library/host_tensor/device_memory.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
#include "ck/library/host_tensor/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_conv_bwd_data.hpp"
using F16 = ck::half_t;
using F32 = float;
using BF16 = ck::bhalf_t;
using INT8 = int8_t;
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using DeviceConvBwdDataNoOpPtr =
DeviceConvBwdDataPtr<ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>;
void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f16_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_bf16_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f16_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
namespace ck {
namespace profiler {
using DeviceConvBwdDataNoOpPtr = ck::tensor_operation::device::instance::DeviceConvBwdDataNoOpPtr;
template <typename InLayout>
HostTensorDescriptor get_input_host_tensor_descriptor(const std::vector<std::size_t>& dims,
int num_dim_spatial = 2)
{
namespace tl = ck::tensor_layout::convolution;
switch(num_dim_spatial)
{
case 3: {
return ck::utils::conv::get_host_tensor_descriptor(dims, InLayout{});
}
case 2: {
return ck::utils::conv::get_host_tensor_descriptor(dims, InLayout{});
}
case 1: {
return ck::utils::conv::get_host_tensor_descriptor(dims, InLayout{});
}
default: {
throw std::runtime_error("Unsupported number of spatial dimensions provided!");
}
}
}
template <typename WeiLayout>
HostTensorDescriptor get_filters_host_tensor_descriptor(const std::vector<std::size_t>& dims,
int num_dim_spatial = 2)
{
namespace tl = ck::tensor_layout::convolution;
switch(num_dim_spatial)
{
case 3: {
return ck::utils::conv::get_host_tensor_descriptor(dims, WeiLayout{});
}
case 2: {
return ck::utils::conv::get_host_tensor_descriptor(dims, WeiLayout{});
}
case 1: {
return ck::utils::conv::get_host_tensor_descriptor(dims, WeiLayout{});
}
default: {
throw std::runtime_error("Unsupported number of spatial dimensions provided!");
}
}
}
template <typename OutLayout>
HostTensorDescriptor get_output_host_ensor_descriptor(const std::vector<std::size_t>& dims,
int num_dim_spatial = 2)
{
namespace tl = ck::tensor_layout::convolution;
switch(num_dim_spatial)
{
case 3: {
return ck::utils::conv::get_host_tensor_descriptor(dims, OutLayout{});
}
case 2: {
return ck::utils::conv::get_host_tensor_descriptor(dims, OutLayout{});
}
case 1: {
return ck::utils::conv::get_host_tensor_descriptor(dims, OutLayout{});
}
default: {
throw std::runtime_error("Unsupported number of spatial dimensions provided!");
}
}
}
template <typename InDataType, typename WeiDataType, typename OutDataType>
void get_device_conv_bwd_data_op_ptr(
InDataType, WeiDataType, OutDataType, std::vector<DeviceConvBwdDataNoOpPtr>&, int)
{
std::cout << "can not find device conv bwd data" << std::endl;
exit(1);
}
template <>
void get_device_conv_bwd_data_op_ptr(
F32, F32, F32, std::vector<DeviceConvBwdDataNoOpPtr>& conv_ptrs, int num_dim_spatial)
{
switch(num_dim_spatial)
{
case 1:
ck::tensor_operation::device::instance::
add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instances(conv_ptrs);
break;
case 2:
ck::tensor_operation::device::instance::
add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances(conv_ptrs);
break;
case 3:
ck::tensor_operation::device::instance::
add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instances(conv_ptrs);
break;
default: break;
}
}
template <>
void get_device_conv_bwd_data_op_ptr(
F16, F16, F16, std::vector<DeviceConvBwdDataNoOpPtr>& conv_ptrs, int num_dim_spatial)
{
switch(num_dim_spatial)
{
case 1:
ck::tensor_operation::device::instance::
add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f16_instances(conv_ptrs);
break;
case 2:
ck::tensor_operation::device::instance::
add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instances(conv_ptrs);
break;
case 3:
ck::tensor_operation::device::instance::
add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f16_instances(conv_ptrs);
break;
default: break;
}
}
template <>
void get_device_conv_bwd_data_op_ptr(
BF16, BF16, BF16, std::vector<DeviceConvBwdDataNoOpPtr>& conv_ptrs, int num_dim_spatial)
{
switch(num_dim_spatial)
{
case 1:
ck::tensor_operation::device::instance::
add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_bf16_instances(conv_ptrs);
break;
case 2:
ck::tensor_operation::device::instance::
add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instances(conv_ptrs);
break;
case 3:
ck::tensor_operation::device::instance::
add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instances(conv_ptrs);
break;
default: break;
}
}
template <>
void get_device_conv_bwd_data_op_ptr(
INT8, INT8, INT8, std::vector<DeviceConvBwdDataNoOpPtr>& conv_ptrs, int num_dim_spatial)
{
switch(num_dim_spatial)
{
case 1:
ck::tensor_operation::device::instance::
add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instances(conv_ptrs);
break;
case 2:
ck::tensor_operation::device::instance::
add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances(conv_ptrs);
break;
case 3:
ck::tensor_operation::device::instance::
add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instances(conv_ptrs);
break;
default: break;
}
}
template <typename T>
static bool check_out(const Tensor<T>& ref, const Tensor<T>& result)
{
float max_diff = 1e-6;
for(std::size_t i = 0; i < ref.mData.size(); ++i)
{
float diff = std::abs(double(ref.mData[i]) - double(result.mData[i]));
if(max_diff < diff)
{
return false;
}
}
return true;
}
template <typename DataType>
void show_data_nhwc_layout(Tensor<DataType>& nhwc)
{
std::cout << "[";
for(int n = 0; n < ck::type_convert<int>(nhwc.mDesc.GetLengths()[0]); n++)
{
std::cout << "[";
for(int hi = 0; hi < ck::type_convert<int>(nhwc.mDesc.GetLengths()[2]); hi++)
{
std::cout << "[";
for(int wi = 0; wi < ck::type_convert<int>(nhwc.mDesc.GetLengths()[3]); wi++)
{
std::cout << "[";
for(int c = 0; c < ck::type_convert<int>(nhwc.mDesc.GetLengths()[1]); c++)
{
std::cout << static_cast<float>(nhwc(n, c, hi, wi)) << " ";
}
std::cout << "]";
}
std::cout << "]";
}
std::cout << "]";
}
std::cout << "]";
}
template <int NDimSpatial,
typename InDataType,
typename WeiDataType,
typename OutDataType,
typename AccDataType,
typename InLayout,
typename WeiLayout,
typename OutLayout>
bool profile_convnd_bwd_data_impl(int do_verification,
int init_method,
bool do_log,
bool time_kernel,
ck::index_t N,
ck::index_t K,
ck::index_t C,
const std::vector<ck::index_t>& input_spatial_lengths,
const std::vector<ck::index_t>& filter_spatial_lengths,
const std::vector<ck::index_t>& output_spatial_lengths,
const std::vector<ck::index_t>& conv_filter_strides,
const std::vector<ck::index_t>& conv_filter_dilations,
const std::vector<ck::index_t>& input_left_pads,
const std::vector<ck::index_t>& input_right_pads)
{
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{};
std::vector<std::size_t> input_dims{static_cast<std::size_t>(N), static_cast<std::size_t>(C)};
input_dims.insert(
std::end(input_dims), std::begin(input_spatial_lengths), std::end(input_spatial_lengths));
std::vector<std::size_t> filter_dims{static_cast<std::size_t>(K), static_cast<std::size_t>(C)};
filter_dims.insert(std::end(filter_dims),
std::begin(filter_spatial_lengths),
std::end(filter_spatial_lengths));
std::vector<std::size_t> output_dims{static_cast<std::size_t>(N), static_cast<std::size_t>(K)};
output_dims.insert(std::end(output_dims),
std::begin(output_spatial_lengths),
std::end(output_spatial_lengths));
Tensor<InDataType> input_host_result(
get_input_host_tensor_descriptor<InLayout>(input_dims, NDimSpatial));
Tensor<InDataType> input_device_result(
get_input_host_tensor_descriptor<InLayout>(input_dims, NDimSpatial));
Tensor<WeiDataType> weights(
get_filters_host_tensor_descriptor<WeiLayout>(filter_dims, NDimSpatial));
Tensor<OutDataType> output(
get_output_host_ensor_descriptor<OutLayout>(output_dims, NDimSpatial));
std::cout << "input: " << input_host_result.mDesc << std::endl;
std::cout << "weights: " << weights.mDesc << std::endl;
std::cout << "output: " << output.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
output.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
weights.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
break;
default:
output.GenerateTensorValue(GeneratorTensor_1<OutDataType>{1});
weights.GenerateTensorValue(GeneratorTensor_1<WeiDataType>{1});
}
DeviceMem in_device_buf(sizeof(InDataType) * input_device_result.mDesc.GetElementSpace());
DeviceMem wei_device_buf(sizeof(WeiDataType) * weights.mDesc.GetElementSpace());
DeviceMem out_device_buf(sizeof(OutDataType) * output.mDesc.GetElementSpace());
out_device_buf.ToDevice(output.mData.data());
wei_device_buf.ToDevice(weights.mData.data());
// reset input to zero
in_device_buf.SetZero();
if(do_verification)
{
auto RunReference = [&](auto& ref_conv) {
auto ref_invoker = ref_conv.MakeInvoker();
auto ref_argument = ref_conv.MakeArgument(input_host_result,
weights,
output,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
InElementOp{},
WeiElementOp{},
OutElementOp{});
ref_invoker.Run(ref_argument);
};
auto ref_conv = ck::tensor_operation::host::ReferenceConvBwdData<InDataType,
WeiDataType,
OutDataType,
AccDataType,
InElementOp,
WeiElementOp,
OutElementOp,
NDimSpatial>();
RunReference(ref_conv);
}
// add device Conv instances
std::vector<DeviceConvBwdDataNoOpPtr> conv_ptrs;
get_device_conv_bwd_data_op_ptr(
InDataType{}, WeiDataType{}, OutDataType{}, conv_ptrs, NDimSpatial);
if(conv_ptrs.size() <= 0)
{
throw std::runtime_error("wrong! no device Conv instance found");
}
std::string best_conv_name;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
// profile device Conv instances
bool success = true;
for(auto& conv_ptr : conv_ptrs)
{
auto argument_ptr = conv_ptr->MakeArgumentPointer(
static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
N,
K,
C,
input_spatial_lengths,
filter_spatial_lengths,
output_spatial_lengths,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
in_element_op,
wei_element_op,
out_element_op);
auto invoker_ptr = conv_ptr->MakeInvokerPointer();
if(conv_ptr->IsSupportedArgument(argument_ptr.get()))
{
std::string conv_name = conv_ptr->GetTypeString();
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop =
ck::utils::conv::get_flops(N, C, K, filter_spatial_lengths, output_spatial_lengths);
std::size_t num_btype =
ck::utils::conv::get_btype<InDataType, WeiDataType, OutDataType>(
N, C, K, input_spatial_lengths, filter_spatial_lengths, output_spatial_lengths);
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
<< " GB/s" << std::endl;
if(tflops > best_tflops)
{
best_conv_name = conv_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
in_device_buf.FromDevice(input_device_result.mData.data());
if(!check_out(input_host_result, input_device_result))
{
std::cout << "Fail Info: " << conv_ptr->GetTypeString() << std::endl;
success = false;
}
else
{
std::cout << "Pass Info: " << conv_ptr->GetTypeString() << std::endl;
}
success = ck::utils::check_err(input_host_result, input_device_result);
if(do_log)
{
std::cout << "in : ";
show_data_nhwc_layout(output);
std::cout << std::endl;
std::cout << "wei: ";
show_data_nhwc_layout(weights);
std::cout << std::endl;
std::cout << "out_host : ";
show_data_nhwc_layout(input_host_result);
std::cout << std::endl;
std::cout << "out_device: ";
show_data_nhwc_layout(input_device_result);
std::cout << std::endl;
}
}
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_conv_name << std::endl;
return success;
}
} // namespace profiler
} // namespace ck
profiler/include/profiler/profile_convnd_bwd_weight_impl.hpp deleted 100644 → 0
View file @ ca8b5c79
#pragma once
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_conv_backward_weight.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/conv_util.hpp"
#include "ck/library/host_tensor/device_memory.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
#include "ck/library/host_tensor/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_conv_backward_weight.hpp"
using F16 = ck::half_t;
using F32 = float;
using BF16 = ck::bhalf_t;
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using DeviceConvndBwdWeightNoOpPtr =
DeviceConvBwdWeightPtr<ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>;
void add_device_conv1d_bwd_weight_xdl_nwc_kxc_nwk_f32_instances(
std::vector<DeviceConvndBwdWeightNoOpPtr>&);
void add_device_convnd_bwd_weight_xdl_nhwc_kyxc_nhwk_f32_instances(
std::vector<DeviceConvndBwdWeightNoOpPtr>&);
void add_device_conv3d_bwd_weight_xdl_ndhwc_kzyxc_ndhwk_f32_instances(
std::vector<DeviceConvndBwdWeightNoOpPtr>&);
void add_device_conv1d_bwd_weight_xdl_nwc_kxc_nwk_f16_instances(
std::vector<DeviceConvndBwdWeightNoOpPtr>&);
void add_device_convnd_bwd_weight_xdl_nhwc_kyxc_nhwk_f16_instances(
std::vector<DeviceConvndBwdWeightNoOpPtr>&);
void add_device_conv3d_bwd_weight_xdl_ndhwc_kzyxc_ndhwk_f16_instances(
std::vector<DeviceConvndBwdWeightNoOpPtr>&);
void add_device_conv1d_bwd_weight_xdl_nwc_kxc_nwk_bf16_instances(
std::vector<DeviceConvndBwdWeightNoOpPtr>&);
void add_device_conv2d_bwd_weight_xdl_nhwc_kyxc_nhwk_bf16_instances(
std::vector<DeviceConvndBwdWeightNoOpPtr>&);
void add_device_conv3d_bwd_weight_xdl_ndhwc_kzyxc_ndhwk_bf16_instances(
std::vector<DeviceConvndBwdWeightNoOpPtr>&);
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
namespace ck {
namespace profiler {
using DeviceConvndBwdWeightNoOpPtr =
ck::tensor_operation::device::instance::DeviceConvndBwdWeightNoOpPtr;
template <typename InLayout>
HostTensorDescriptor get_input_host_tensor_descriptor(const std::vector<std::size_t>& dims,
int num_dim_spatial = 2)
{
namespace tl = ck::tensor_layout::convolution;
switch(num_dim_spatial)
{
case 3: {
return ck::utils::conv::get_host_tensor_descriptor(dims, InLayout{});
}
case 2: {
return ck::utils::conv::get_host_tensor_descriptor(dims, InLayout{});
}
case 1: {
return ck::utils::conv::get_host_tensor_descriptor(dims, InLayout{});
}
default: {
throw std::runtime_error("Unsupported number of spatial dimensions provided!");
}
}
}
template <typename WeiLayout>
HostTensorDescriptor get_filters_host_tensor_descriptor(const std::vector<std::size_t>& dims,
int num_dim_spatial = 2)
{
namespace tl = ck::tensor_layout::convolution;
switch(num_dim_spatial)
{
case 3: {
return ck::utils::conv::get_host_tensor_descriptor(dims, WeiLayout{});
}
case 2: {
return ck::utils::conv::get_host_tensor_descriptor(dims, WeiLayout{});
}
case 1: {
return ck::utils::conv::get_host_tensor_descriptor(dims, WeiLayout{});
}
default: {
throw std::runtime_error("Unsupported number of spatial dimensions provided!");
}
}
}
template <typename OutLayout>
HostTensorDescriptor get_output_host_ensor_descriptor(const std::vector<std::size_t>& dims,
int num_dim_spatial = 2)
{
namespace tl = ck::tensor_layout::convolution;
switch(num_dim_spatial)
{
case 3: {
return ck::utils::conv::get_host_tensor_descriptor(dims, OutLayout{});
}
case 2: {
return ck::utils::conv::get_host_tensor_descriptor(dims, OutLayout{});
}
case 1: {
return ck::utils::conv::get_host_tensor_descriptor(dims, OutLayout{});
}
default: {
throw std::runtime_error("Unsupported number of spatial dimensions provided!");
}
}
}
template <typename InDataType, typename WeiDataType, typename OutDataType>
void get_device_conv_bwd_weight_op_ptr(
InDataType, WeiDataType, OutDataType, std::vector<DeviceConvndBwdWeightNoOpPtr>&, int)
{
std::cout << "can not find device conv bwd weight" << std::endl;
exit(1);
}
template <>
void get_device_conv_bwd_weight_op_ptr(
F32, F32, F32, std::vector<DeviceConvndBwdWeightNoOpPtr>& conv_ptrs, int num_dim_spatial)
{
switch(num_dim_spatial)
{
case 1:
ck::tensor_operation::device::instance::
add_device_conv1d_bwd_weight_xdl_nwc_kxc_nwk_f32_instances(conv_ptrs);
break;
case 2:
ck::tensor_operation::device::instance::
add_device_convnd_bwd_weight_xdl_nhwc_kyxc_nhwk_f32_instances(conv_ptrs);
break;
case 3:
ck::tensor_operation::device::instance::
add_device_conv3d_bwd_weight_xdl_ndhwc_kzyxc_ndhwk_f32_instances(conv_ptrs);
break;
default: break;
}
}
template <>
void get_device_conv_bwd_weight_op_ptr(
F16, F16, F16, std::vector<DeviceConvndBwdWeightNoOpPtr>& conv_ptrs, int num_dim_spatial)
{
switch(num_dim_spatial)
{
case 1:
ck::tensor_operation::device::instance::
add_device_conv1d_bwd_weight_xdl_nwc_kxc_nwk_f16_instances(conv_ptrs);
break;
case 2:
ck::tensor_operation::device::instance::
add_device_convnd_bwd_weight_xdl_nhwc_kyxc_nhwk_f16_instances(conv_ptrs);
break;
case 3:
ck::tensor_operation::device::instance::
add_device_conv3d_bwd_weight_xdl_ndhwc_kzyxc_ndhwk_f16_instances(conv_ptrs);
break;
default: break;
}
}
template <>
void get_device_conv_bwd_weight_op_ptr(
BF16, BF16, BF16, std::vector<DeviceConvndBwdWeightNoOpPtr>& conv_ptrs, int num_dim_spatial)
{
switch(num_dim_spatial)
{
case 1:
ck::tensor_operation::device::instance::
add_device_conv1d_bwd_weight_xdl_nwc_kxc_nwk_bf16_instances(conv_ptrs);
break;
case 2:
ck::tensor_operation::device::instance::
add_device_conv2d_bwd_weight_xdl_nhwc_kyxc_nhwk_bf16_instances(conv_ptrs);
break;
case 3:
ck::tensor_operation::device::instance::
add_device_conv3d_bwd_weight_xdl_ndhwc_kzyxc_ndhwk_bf16_instances(conv_ptrs);
break;
default: break;
}
}
template <typename DataType>
void show_data_nhwc_layout(Tensor<DataType>& nhwc)
{
std::cout << "[";
for(int n = 0; n < ck::type_convert<int>(nhwc.mDesc.GetLengths()[0]); n++)
{
std::cout << "[";
for(int hi = 0; hi < ck::type_convert<int>(nhwc.mDesc.GetLengths()[2]); hi++)
{
std::cout << "[";
for(int wi = 0; wi < ck::type_convert<int>(nhwc.mDesc.GetLengths()[3]); wi++)
{
std::cout << "[";
for(int c = 0; c < ck::type_convert<int>(nhwc.mDesc.GetLengths()[1]); c++)
{
std::cout << static_cast<float>(nhwc(n, c, hi, wi)) << " ";
}
std::cout << "]";
}
std::cout << "]";
}
std::cout << "]";
}
std::cout << "]";
}
template <int NDimSpatial,
typename InDataType,
typename WeiDataType,
typename OutDataType,
typename InLayout,
typename WeiLayout,
typename OutLayout>
bool profile_convnd_bwd_weight_impl(int do_verification,
int init_method,
bool do_log,
bool time_kernel,
ck::index_t N,
ck::index_t K,
ck::index_t C,
std::vector<ck::index_t> input_spatial_lengths,
std::vector<ck::index_t> filter_spatial_lengths,
std::vector<ck::index_t> output_spatial_lengths,
std::vector<ck::index_t> conv_filter_strides,
std::vector<ck::index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads,
ck::index_t split_k)
{
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{};
std::vector<std::size_t> input_dims{static_cast<std::size_t>(N), static_cast<std::size_t>(C)};
input_dims.insert(
std::end(input_dims), std::begin(input_spatial_lengths), std::end(input_spatial_lengths));
std::vector<std::size_t> filter_dims{static_cast<std::size_t>(K), static_cast<std::size_t>(C)};
filter_dims.insert(std::end(filter_dims),
std::begin(filter_spatial_lengths),
std::end(filter_spatial_lengths));
std::vector<std::size_t> output_dims{static_cast<std::size_t>(N), static_cast<std::size_t>(K)};
output_dims.insert(std::end(output_dims),
std::begin(output_spatial_lengths),
std::end(output_spatial_lengths));
Tensor<InDataType> input(get_input_host_tensor_descriptor<InLayout>(input_dims, NDimSpatial));
Tensor<WeiDataType> weights_host_result(
get_filters_host_tensor_descriptor<WeiLayout>(filter_dims, NDimSpatial));
Tensor<WeiDataType> weights_device_result(
get_filters_host_tensor_descriptor<WeiLayout>(filter_dims, NDimSpatial));
Tensor<OutDataType> output(
get_output_host_ensor_descriptor<OutLayout>(output_dims, NDimSpatial));
std::cout << "input: " << input.mDesc << std::endl;
std::cout << "weights: " << weights_host_result.mDesc << std::endl;
std::cout << "output: " << output.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
input.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-2, 2});
output.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-2, 2});
break;
default:
input.GenerateTensorValue(GeneratorTensor_1<OutDataType>{1});
output.GenerateTensorValue(GeneratorTensor_1<WeiDataType>{1});
}
DeviceMem in_device_buf(sizeof(InDataType) * input.mDesc.GetElementSpace());
DeviceMem wei_device_buf(sizeof(WeiDataType) * weights_device_result.mDesc.GetElementSpace());
DeviceMem out_device_buf(sizeof(OutDataType) * output.mDesc.GetElementSpace());
in_device_buf.ToDevice(input.mData.data());
out_device_buf.ToDevice(output.mData.data());
// reset input to zero
wei_device_buf.SetZero();
if(do_verification)
{
auto RunReference = [&](auto& ref_conv) {
auto ref_invoker = ref_conv.MakeInvoker();
auto ref_argument = ref_conv.MakeArgument(input,
weights_host_result,
output,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
InElementOp{},
WeiElementOp{},
OutElementOp{});
ref_invoker.Run(ref_argument);
};
auto ref_conv = ck::tensor_operation::host::ReferenceConvBwdWeight<InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
NDimSpatial>();
RunReference(ref_conv);
}
// add device Conv instances
std::vector<DeviceConvndBwdWeightNoOpPtr> conv_ptrs;
get_device_conv_bwd_weight_op_ptr(
InDataType{}, WeiDataType{}, OutDataType{}, conv_ptrs, NDimSpatial);
if(conv_ptrs.size() <= 0)
{
throw std::runtime_error("wrong! no device Conv instance found");
}
std::string best_conv_name;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
// profile device Conv instances
bool success = true;
for(auto& conv_ptr : conv_ptrs)
{
// using atomic, so need to reset input, setzero is done in invoker
// if(split_k > 1)
//{
// wei_device_buf.SetZero();
//}
auto argument_ptr = conv_ptr->MakeArgumentPointer(
static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
N,
K,
C,
input_spatial_lengths,
filter_spatial_lengths,
output_spatial_lengths,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
in_element_op,
wei_element_op,
out_element_op,
split_k);
if(!conv_ptr->IsSupportedArgument(argument_ptr.get()))
{
std::cout << "wrong! device_conv with the specified compilation parameters does "
"not support this Conv problem"
<< std::endl;
continue;
}
auto invoker_ptr = conv_ptr->MakeInvokerPointer();
std::string conv_name = conv_ptr->GetTypeString();
float ave_time = 0;
if(std::is_same<InDataType, ck::bhalf_t>::value && split_k > 1)
{
// alloc work space
size_t bwd_weight_workspace_size = conv_ptr->GetWorkSpaceSize(argument_ptr.get());
if(bwd_weight_workspace_size <= 0)
{
printf("wrong work space size\n");
exit(1);
}
DeviceMem wei_work_space_device_buf(bwd_weight_workspace_size);
wei_work_space_device_buf.SetZero();
conv_ptr->SetWorkSpacePointer(argument_ptr.get(),
wei_work_space_device_buf.GetDeviceBuffer());
ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
}
else
{
ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
}
std::size_t flop =
ck::utils::conv::get_flops(N, C, K, filter_spatial_lengths, output_spatial_lengths);
std::size_t num_btype = ck::utils::conv::get_btype<InDataType, WeiDataType, OutDataType>(
N, C, K, input_spatial_lengths, filter_spatial_lengths, output_spatial_lengths);
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
<< " GB/s" << std::endl;
if(tflops > best_tflops)
{
best_conv_name = conv_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
wei_device_buf.FromDevice(weights_device_result.mData.data());
success = ck::utils::check_err(weights_host_result, weights_device_result);
if(success == false)
{
std::cout << "Fail Info: " << conv_ptr->GetTypeString() << std::endl;
}
else
{
std::cout << "Pass Info: " << conv_ptr->GetTypeString() << std::endl;
}
if(do_log)
{
std::cout << "in : ";
show_data_nhwc_layout(output);
std::cout << std::endl;
std::cout << "wei: ";
show_data_nhwc_layout(weights_host_result);
std::cout << std::endl;
std::cout << "out : ";
show_data_nhwc_layout(input);
std::cout << std::endl;
std::cout << "wei_device: ";
show_data_nhwc_layout(weights_device_result);
std::cout << std::endl;
}
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_conv_name << std::endl;
return success;
}
} // namespace profiler
} // namespace ck
profiler/include/profiler/profile_pool2d_fwd_impl.hpp 0 → 100644
View file @ 2f463a94
// 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
profiler/include/profiler/profile_pool3d_fwd_impl.hpp 0 → 100644
View file @ 2f463a94
// 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
profiler/src/CMakeLists.txt
View file @ 2f463a94
...@@ -26,11 +26,15 @@ set(PROFILER_SOURCES ...@@ -26,11 +26,15 @@ set(PROFILER_SOURCES
profile_reduce.cpp profile_reduce.cpp
profile_groupnorm.cpp profile_groupnorm.cpp
profile_layernorm.cpp profile_layernorm.cpp
profile_avg_pool2d_fwd.cpp
profile_max_pool3d_fwd.cpp
profile_softmax.cpp profile_softmax.cpp
profile_batchnorm_fwd.cpp profile_batchnorm_fwd.cpp
profile_batchnorm_bwd.cpp profile_batchnorm_bwd.cpp
profile_batchnorm_infer.cpp profile_batchnorm_infer.cpp
profile_grouped_gemm_fastgelu.cpp profile_grouped_gemm_fastgelu.cpp
profile_contraction_bilinear.cpp
profile_contraction_scale.cpp
) )
set(PROFILER_EXECUTABLE ckProfiler) set(PROFILER_EXECUTABLE ckProfiler)
...@@ -72,4 +76,8 @@ target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_softmax_instance) ...@@ -72,4 +76,8 @@ target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_softmax_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_reduce_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_reduce_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batchnorm_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batchnorm_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fastgelu_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_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) rocm_install(TARGETS ${PROFILER_EXECUTABLE} COMPONENT profiler)
profiler/src/profile_avg_pool2d_fwd.cpp 0 → 100644
View file @ 2f463a94
// 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);
profiler/src/profile_contraction_bilinear.cpp 0 → 100644
View file @ 2f463a94
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <vector>
#include "profiler/profile_contraction_impl.hpp"
#include "profiler/profile_contraction_utils.hpp"
#include "profiler_operation_registry.hpp"
#define OP_NAME "contraction_bilinear"
#define OP_DESC "CONTRACTION+Bilinear"
static void print_helper_msg()
{
std::cout << "arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"
<< "arg2: data type (0: fp32; 1: f64)\n"
<< "arg3: matrix layout (0: A[m0, m1, k0, k1] * B[k0, k1, n0, n1] + "
"D[m0, m1, n0, n1] = E[m0, m1, n0, n1];\n"
<< " 1: A[m0, m1, k0, k1] * B[n0, n1, k0, k1] + "
"D[m0, m1, n0, n1] = E[m0, m1, n0, n1];\n"
<< " 2: A[k0, k1, m0, m1] * B[k0, k1, n0, n1] + "
"D[m0, m1, n0, n1] = E[m0, m1, n0, n1];\n"
<< " 3: A[k0, k1, m0, m1] * B[n0, n1, k0, k1] + "
"D[m0, m1, n0, n1] = E[m0, m1, n0, n1])\n"
<< "arg4: verification (0: no; 1: yes)\n"
<< "arg5: initialization (0: no init; 1: integer value; 2: decimal "
<< "value)\n"
<< "arg6: print tensor value (0: no; 1: yes)\n"
<< "arg7: time kernel (0: no, 1: yes)\n"
<< "arg8 and arg9: alpha and beta\n"
<< "arg10 to 15: M0, M1, N0, N1, K0, K1\n"
<< "arg16 to 31: Strides for A, B, D and E (skip for default)\n"
<< std::endl;
}
int profile_contraction_bilinear(int argc, char* argv[])
{
const bool default_strides = argc == 16;
if(argc != 32 && argc != 16)
{
print_helper_msg();
exit(1);
}
const auto data_type = static_cast<ContractionDataType>(std::stoi(argv[2]));
const auto layout = static_cast<ContractionMatrixLayout>(std::stoi(argv[3]));
const bool do_verification = std::stoi(argv[4]);
const ck::index_t init_method = std::stoi(argv[5]);
const bool do_log = std::stoi(argv[6]);
const bool time_kernel = std::stoi(argv[7]);
const float alpha = std::stof(argv[8]);
const float beta = std::stof(argv[9]);
std::vector<ck::index_t> M;
std::vector<ck::index_t> N;
std::vector<ck::index_t> K;
const ck::index_t dims_arg_num = 10;
collect_index_params(argv, M, dims_arg_num, 2);
collect_index_params(argv, N, dims_arg_num + 2, 2);
collect_index_params(argv, K, dims_arg_num + 4, 2);
std::vector<ck::index_t> StridesA;
std::vector<ck::index_t> StridesB;
std::vector<ck::index_t> StridesE;
std::vector<ck::index_t> StridesD;
if(!default_strides)
{
collect_index_params(argv, StridesA, dims_arg_num + 6, 4);
collect_index_params(argv, StridesB, dims_arg_num + 10, 4);
collect_index_params(argv, StridesE, dims_arg_num + 14, 4);
collect_index_params(argv, StridesD, dims_arg_num + 18, 4);
}
using F32 = float;
using F64 = double;
auto profile = [&](auto a_layout, auto b_layout, auto cde_layout, auto type) {
using ALayout = decltype(a_layout);
using BLayout = decltype(b_layout);
using CDELayout = decltype(cde_layout);
using DataType = decltype(type);
if(default_strides)
{
assign_default_strides(a_layout, StridesA, {M[0], M[1], K[0], K[1]});
assign_default_strides(b_layout, StridesB, {K[0], K[1], N[0], N[1]});
assign_default_strides(cde_layout, StridesE, {M[0], M[1], N[0], N[1]});
assign_default_strides(cde_layout, StridesD, {M[0], M[1], N[0], N[1]});
}
bool pass = ck::profiler::profile_contraction_impl<ALayout,
BLayout,
CDELayout,
DataType,
ck::Tuple<DataType>,
Bilinear>(do_verification,
init_method,
do_log,
time_kernel,
Bilinear{alpha, beta},
M,
N,
K,
StridesA,
StridesB,
StridesE,
StridesD);
return pass;
};
if(data_type == ContractionDataType::F32_F32_F32_F32 &&
layout == ContractionMatrixLayout::MK_KN_MN_MN)
{
return profile(Row{}, Row{}, Row{}, F32{});
}
else if(data_type == ContractionDataType::F32_F32_F32_F32 &&
layout == ContractionMatrixLayout::MK_NK_MN_MN)
{
return profile(Row{}, Col{}, Row{}, F32{});
}
else if(data_type == ContractionDataType::F32_F32_F32_F32 &&
layout == ContractionMatrixLayout::KM_KN_MN_MN)
{
return profile(Col{}, Row{}, Row{}, F32{});
}
else if(data_type == ContractionDataType::F32_F32_F32_F32 &&
layout == ContractionMatrixLayout::KM_NK_MN_MN)
{
return profile(Col{}, Col{}, Row{}, F32{});
}
else if(data_type == ContractionDataType::F64_F64_F64_F64 &&
layout == ContractionMatrixLayout::MK_KN_MN_MN)
{
return profile(Row{}, Row{}, Row{}, F64{});
}
else if(data_type == ContractionDataType::F64_F64_F64_F64 &&
layout == ContractionMatrixLayout::MK_NK_MN_MN)
{
return profile(Row{}, Col{}, Row{}, F64{});
}
else if(data_type == ContractionDataType::F64_F64_F64_F64 &&
layout == ContractionMatrixLayout::KM_KN_MN_MN)
{
return profile(Col{}, Row{}, Row{}, F64{});
}
else if(data_type == ContractionDataType::F64_F64_F64_F64 &&
layout == ContractionMatrixLayout::KM_NK_MN_MN)
{
return profile(Col{}, Col{}, Row{}, F64{});
}
else
{
std::cout << "this data_type & layout is not implemented" << std::endl;
return 1;
}
}
REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_contraction_bilinear);
profiler/src/profile_contraction_scale.cpp 0 → 100644
View file @ 2f463a94
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <vector>
#include "profiler/profile_contraction_impl.hpp"
#include "profiler/profile_contraction_utils.hpp"
#include "profiler_operation_registry.hpp"
#define OP_NAME "contraction_scale"
#define OP_DESC "CONTRACTION+Scale"
static void print_helper_msg()
{
std::cout << "arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"
<< "arg2: data type (0: fp32; 1: f64)\n"
<< "arg3: matrix layout (0: A[m0, m1, k0, k1] * B[k0, k1, n0, n1] + "
"D[m0, m1, n0, n1] = E[m0, m1, n0, n1];\n"
<< " 1: A[m0, m1, k0, k1] * B[n0, n1, k0, k1] + "
"D[m0, m1, n0, n1] = E[m0, m1, n0, n1];\n"
<< " 2: A[k0, k1, m0, m1] * B[k0, k1, n0, n1] + "
"D[m0, m1, n0, n1] = E[m0, m1, n0, n1];\n"
<< " 3: A[k0, k1, m0, m1] * B[n0, n1, k0, k1] + "
"D[m0, m1, n0, n1] = E[m0, m1, n0, n1])\n"
<< "arg4: verification (0: no; 1: yes)\n"
<< "arg5: initialization (0: no init; 1: integer value; 2: decimal "
<< "value)\n"
<< "arg6: print tensor value (0: no; 1: yes)\n"
<< "arg7: time kernel (0: no, 1: yes)\n"
<< "arg8: alpha\n"
<< "arg9 to 14: M0, M1, N0, N1, K0, K1\n"
<< "arg15 to 30: Strides for A, B, D and E (skip for default)\n"
<< std::endl;
}
int profile_contraction_scale(int argc, char* argv[])
{
const bool default_strides = argc == 15;
if(argc != 31 && argc != 15)
{
print_helper_msg();
exit(1);
}
const auto data_type = static_cast<ContractionDataType>(std::stoi(argv[2]));
const auto layout = static_cast<ContractionMatrixLayout>(std::stoi(argv[3]));
const bool do_verification = std::stoi(argv[4]);
const ck::index_t init_method = std::stoi(argv[5]);
const bool do_log = std::stoi(argv[6]);
const bool time_kernel = std::stoi(argv[7]);
const float alpha = std::stof(argv[8]);
std::vector<ck::index_t> M;
std::vector<ck::index_t> N;
std::vector<ck::index_t> K;
const ck::index_t dims_arg_num = 9;
collect_index_params(argv, M, dims_arg_num, 2);
collect_index_params(argv, N, dims_arg_num + 2, 2);
collect_index_params(argv, K, dims_arg_num + 4, 2);
std::vector<ck::index_t> StridesA;
std::vector<ck::index_t> StridesB;
std::vector<ck::index_t> StridesE;
std::vector<ck::index_t> StridesD;
if(!default_strides)
{
collect_index_params(argv, StridesA, dims_arg_num + 6, 4);
collect_index_params(argv, StridesB, dims_arg_num + 10, 4);
collect_index_params(argv, StridesE, dims_arg_num + 14, 4);
collect_index_params(argv, StridesD, dims_arg_num + 18, 4);
}
using F32 = float;
using F64 = double;
auto profile = [&](auto a_layout, auto b_layout, auto cde_layout, auto type) {
using ALayout = decltype(a_layout);
using BLayout = decltype(b_layout);
using CDELayout = decltype(cde_layout);
using DataType = decltype(type);
if(default_strides)
{
assign_default_strides(a_layout, StridesA, {M[0], M[1], K[0], K[1]});
assign_default_strides(b_layout, StridesB, {K[0], K[1], N[0], N[1]});
assign_default_strides(cde_layout, StridesE, {M[0], M[1], N[0], N[1]});
assign_default_strides(cde_layout, StridesD, {M[0], M[1], N[0], N[1]});
}
bool pass = ck::profiler::
profile_contraction_impl<ALayout, BLayout, CDELayout, DataType, ck::Tuple<>, Scale>(
do_verification,
init_method,
do_log,
time_kernel,
Scale{alpha},
M,
N,
K,
StridesA,
StridesB,
StridesE,
StridesD);
return pass;
};
if(data_type == ContractionDataType::F32_F32_F32_F32 &&
layout == ContractionMatrixLayout::MK_KN_MN_MN)
{
return profile(Row{}, Row{}, Row{}, F32{});
}
else if(data_type == ContractionDataType::F32_F32_F32_F32 &&
layout == ContractionMatrixLayout::MK_NK_MN_MN)
{
return profile(Row{}, Col{}, Row{}, F32{});
}
else if(data_type == ContractionDataType::F32_F32_F32_F32 &&
layout == ContractionMatrixLayout::KM_KN_MN_MN)
{
return profile(Col{}, Row{}, Row{}, F32{});
}
else if(data_type == ContractionDataType::F32_F32_F32_F32 &&
layout == ContractionMatrixLayout::KM_NK_MN_MN)
{
return profile(Col{}, Col{}, Row{}, F32{});
}
else if(data_type == ContractionDataType::F64_F64_F64_F64 &&
layout == ContractionMatrixLayout::MK_KN_MN_MN)
{
return profile(Row{}, Row{}, Row{}, F64{});
}
else if(data_type == ContractionDataType::F64_F64_F64_F64 &&
layout == ContractionMatrixLayout::MK_NK_MN_MN)
{
return profile(Row{}, Col{}, Row{}, F64{});
}
else if(data_type == ContractionDataType::F64_F64_F64_F64 &&
layout == ContractionMatrixLayout::KM_KN_MN_MN)
{
return profile(Col{}, Row{}, Row{}, F64{});
}
else if(data_type == ContractionDataType::F64_F64_F64_F64 &&
layout == ContractionMatrixLayout::KM_NK_MN_MN)
{
return profile(Col{}, Col{}, Row{}, F64{});
}
else
{
std::cout << "this data_type & layout is not implemented" << std::endl;
return 1;
}
}
REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_contraction_scale);
profiler/src/profile_groupnorm.cpp
View file @ 2f463a94
...@@ -64,7 +64,7 @@ int profile_groupnorm(int argc, char* argv[]) ...@@ -64,7 +64,7 @@ int profile_groupnorm(int argc, char* argv[])
ck::DataTypeEnum data_type = ck::DataTypeEnum::Half; ck::DataTypeEnum data_type = ck::DataTypeEnum::Half;
bool do_verification = false; bool do_verification = false;
int init_method = 0; int init_method = 0;
bool do_log = 0; bool do_log = false;
bool time_kernel = 1; bool time_kernel = 1;
std::vector<index_t> length = {64, 16, 16, 32, 40}; std::vector<index_t> length = {64, 16, 16, 32, 40};
......
profiler/src/profile_max_pool3d_fwd.cpp 0 → 100644
View file @ 2f463a94
// 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);
test/CMakeLists.txt
View file @ 2f463a94
...@@ -56,6 +56,8 @@ add_subdirectory(normalization) ...@@ -56,6 +56,8 @@ add_subdirectory(normalization)
add_subdirectory(data_type) add_subdirectory(data_type)
add_subdirectory(elementwise_normalization) add_subdirectory(elementwise_normalization)
add_subdirectory(batchnorm) add_subdirectory(batchnorm)
add_subdirectory(contraction)
add_subdirectory(pool_fwd)
if(GPU_TARGETS MATCHES "gfx1100") if(GPU_TARGETS MATCHES "gfx1100")
add_subdirectory(wmma_op) add_subdirectory(wmma_op)
endif() endif()
test/batched_gemm/CMakeLists.txt
View file @ 2f463a94
add_test_executable(test_batched_gemm_fp16 batched_gemm_fp16.cpp) if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940")
target_link_libraries(test_batched_gemm_fp16 PRIVATE utility) add_test_executable(test_batched_gemm_fp16 batched_gemm_fp16.cpp)
target_link_libraries(test_batched_gemm_fp16 PRIVATE device_batched_gemm_instance) target_link_libraries(test_batched_gemm_fp16 PRIVATE utility)
target_link_libraries(test_batched_gemm_fp16 PRIVATE device_batched_gemm_instance)
add_test_executable(test_batched_gemm_fp32 batched_gemm_fp32.cpp) add_test_executable(test_batched_gemm_fp32 batched_gemm_fp32.cpp)
target_link_libraries(test_batched_gemm_fp32 PRIVATE utility) target_link_libraries(test_batched_gemm_fp32 PRIVATE utility)
target_link_libraries(test_batched_gemm_fp32 PRIVATE device_batched_gemm_instance) target_link_libraries(test_batched_gemm_fp32 PRIVATE device_batched_gemm_instance)
add_test_executable(test_batched_gemm_bf16 batched_gemm_bf16.cpp) add_test_executable(test_batched_gemm_bf16 batched_gemm_bf16.cpp)
target_link_libraries(test_batched_gemm_bf16 PRIVATE utility) target_link_libraries(test_batched_gemm_bf16 PRIVATE utility)
target_link_libraries(test_batched_gemm_bf16 PRIVATE device_batched_gemm_instance) target_link_libraries(test_batched_gemm_bf16 PRIVATE device_batched_gemm_instance)
add_test_executable(test_batched_gemm_int8 batched_gemm_int8.cpp) add_test_executable(test_batched_gemm_int8 batched_gemm_int8.cpp)
target_link_libraries(test_batched_gemm_int8 PRIVATE utility) target_link_libraries(test_batched_gemm_int8 PRIVATE utility)
target_link_libraries(test_batched_gemm_int8 PRIVATE device_batched_gemm_instance) target_link_libraries(test_batched_gemm_int8 PRIVATE device_batched_gemm_instance)
endif()
\ No newline at end of file
test/batched_gemm_gemm/CMakeLists.txt
View file @ 2f463a94
add_custom_target(test_batched_gemm_gemm) if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940")
add_custom_target(test_batched_gemm_gemm)
add_gtest_executable(test_batched_gemm_gemm_fp16 test_batched_gemm_gemm_fp16.cpp) add_gtest_executable(test_batched_gemm_gemm_fp16 test_batched_gemm_gemm_fp16.cpp)
target_link_libraries(test_batched_gemm_gemm_fp16 PRIVATE utility device_batched_gemm_gemm_instance) target_link_libraries(test_batched_gemm_gemm_fp16 PRIVATE utility device_batched_gemm_gemm_instance)
add_dependencies(test_batched_gemm_gemm test_batched_gemm_gemm_fp16) add_dependencies(test_batched_gemm_gemm test_batched_gemm_gemm_fp16)
\ No newline at end of file endif()
\ No newline at end of file
test/batched_gemm_reduce/CMakeLists.txt
View file @ 2f463a94
add_test_executable(test_batched_gemm_reduce_fp16 batched_gemm_reduce_fp16.cpp) if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940")
target_link_libraries(test_batched_gemm_reduce_fp16 PRIVATE utility) add_test_executable(test_batched_gemm_reduce_fp16 batched_gemm_reduce_fp16.cpp)
target_link_libraries(test_batched_gemm_reduce_fp16 PRIVATE device_batched_gemm_reduce_instance) target_link_libraries(test_batched_gemm_reduce_fp16 PRIVATE utility)
target_link_libraries(test_batched_gemm_reduce_fp16 PRIVATE device_batched_gemm_reduce_instance)
endif()
test/batched_gemm_softmax_gemm/CMakeLists.txt
View file @ 2f463a94
add_custom_target(test_batched_gemm_softmax_gemm) if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940")
add_custom_target(test_batched_gemm_softmax_gemm)
add_gtest_executable(test_batched_gemm_softmax_gemm_fp16 test_batched_gemm_softmax_gemm_fp16.cpp) add_gtest_executable(test_batched_gemm_softmax_gemm_fp16 test_batched_gemm_softmax_gemm_fp16.cpp)
target_link_libraries(test_batched_gemm_softmax_gemm_fp16 PRIVATE utility device_batched_gemm_softmax_gemm_instance) target_link_libraries(test_batched_gemm_softmax_gemm_fp16 PRIVATE utility device_batched_gemm_softmax_gemm_instance)
add_dependencies(test_batched_gemm_softmax_gemm test_batched_gemm_softmax_gemm_fp16) add_dependencies(test_batched_gemm_softmax_gemm test_batched_gemm_softmax_gemm_fp16)
\ No newline at end of file endif()
\ No newline at end of file
test/batched_gemm_softmax_gemm_permute/CMakeLists.txt
View file @ 2f463a94
add_custom_target(test_batched_gemm_softmax_gemm_permute) if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940")
add_custom_target(test_batched_gemm_softmax_gemm_permute)
add_gtest_executable(test_batched_gemm_softmax_gemm_permute_fp16 test_batched_gemm_softmax_gemm_permute_fp16.cpp) add_gtest_executable(test_batched_gemm_softmax_gemm_permute_fp16 test_batched_gemm_softmax_gemm_permute_fp16.cpp)
add_gtest_executable(test_batched_gemm_softmax_gemm_permute_bf16 test_batched_gemm_softmax_gemm_permute_bf16.cpp) add_gtest_executable(test_batched_gemm_softmax_gemm_permute_bf16 test_batched_gemm_softmax_gemm_permute_bf16.cpp)
target_link_libraries(test_batched_gemm_softmax_gemm_permute_fp16 PRIVATE utility device_batched_gemm_softmax_gemm_permute_instance) target_link_libraries(test_batched_gemm_softmax_gemm_permute_fp16 PRIVATE utility device_batched_gemm_softmax_gemm_permute_instance)
target_link_libraries(test_batched_gemm_softmax_gemm_permute_bf16 PRIVATE utility device_batched_gemm_softmax_gemm_permute_instance) target_link_libraries(test_batched_gemm_softmax_gemm_permute_bf16 PRIVATE utility device_batched_gemm_softmax_gemm_permute_instance)
add_dependencies(test_batched_gemm_softmax_gemm_permute test_batched_gemm_softmax_gemm_permute_fp16) add_dependencies(test_batched_gemm_softmax_gemm_permute test_batched_gemm_softmax_gemm_permute_fp16)
add_dependencies(test_batched_gemm_softmax_gemm_permute test_batched_gemm_softmax_gemm_permute_bf16) add_dependencies(test_batched_gemm_softmax_gemm_permute test_batched_gemm_softmax_gemm_permute_bf16)
add_gtest_executable(test_batched_gemm_bias_softmax_gemm_permute_fp16 test_batched_gemm_bias_softmax_gemm_permute_fp16.cpp) add_gtest_executable(test_batched_gemm_bias_softmax_gemm_permute_fp16 test_batched_gemm_bias_softmax_gemm_permute_fp16.cpp)
add_gtest_executable(test_batched_gemm_bias_softmax_gemm_permute_bf16 test_batched_gemm_bias_softmax_gemm_permute_bf16.cpp) add_gtest_executable(test_batched_gemm_bias_softmax_gemm_permute_bf16 test_batched_gemm_bias_softmax_gemm_permute_bf16.cpp)
target_link_libraries(test_batched_gemm_bias_softmax_gemm_permute_fp16 PRIVATE utility device_batched_gemm_softmax_gemm_permute_instance) target_link_libraries(test_batched_gemm_bias_softmax_gemm_permute_fp16 PRIVATE utility device_batched_gemm_softmax_gemm_permute_instance)
target_link_libraries(test_batched_gemm_bias_softmax_gemm_permute_bf16 PRIVATE utility device_batched_gemm_softmax_gemm_permute_instance) target_link_libraries(test_batched_gemm_bias_softmax_gemm_permute_bf16 PRIVATE utility device_batched_gemm_softmax_gemm_permute_instance)
add_dependencies(test_batched_gemm_softmax_gemm_permute test_batched_gemm_bias_softmax_gemm_permute_fp16) add_dependencies(test_batched_gemm_softmax_gemm_permute test_batched_gemm_bias_softmax_gemm_permute_fp16)
add_dependencies(test_batched_gemm_softmax_gemm_permute test_batched_gemm_bias_softmax_gemm_permute_bf16) add_dependencies(test_batched_gemm_softmax_gemm_permute test_batched_gemm_bias_softmax_gemm_permute_bf16)
\ No newline at end of file endif()
\ No newline at end of file
test/contraction/CMakeLists.txt 0 → 100644
View file @ 2f463a94
add_gtest_executable(test_contraction test_contraction.cpp)
target_link_libraries(test_contraction PRIVATE utility device_contraction_bilinear_instance device_contraction_scale_instance)
if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940")
add_gtest_executable(test_contraction_interface test_contraction_interface.cpp)
target_link_libraries(test_contraction_interface PRIVATE utility device_contraction_bilinear_instance device_contraction_scale_instance)
endif()
test/contraction/test_contraction.cpp 0 → 100644
View file @ 2f463a94
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <iostream>
#include <memory>
#include <initializer_list>
#include <vector>
#include <tuple>
#include <gtest/gtest.h>
#include "profiler/profile_contraction_impl.hpp"
using F32 = float;
using F64 = double;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using Bilinear = ck::tensor_operation::element_wise::Bilinear;
using Scale = ck::tensor_operation::element_wise::Scale;
struct MemoryParams
{
std::vector<ck::index_t> M;
std::vector<ck::index_t> N;
std::vector<ck::index_t> K;
std::vector<ck::index_t> StridesA;
std::vector<ck::index_t> StridesB;
std::vector<ck::index_t> StridesC;
std::vector<ck::index_t> StridesD;
};
template <typename Tuple>
class TestContraction : public ::testing::Test
{
protected:
using ALayout = std::tuple_element_t<0, Tuple>;
using BLayout = std::tuple_element_t<1, Tuple>;
using CDLayout = std::tuple_element_t<2, Tuple>;
using DataType = std::tuple_element_t<3, Tuple>;
using DTupleDataType = std::tuple_element_t<4, Tuple>;
using CDElementOp = std::tuple_element_t<5, Tuple>;
std::vector<MemoryParams> list_of_memory_params = {{{32, 32},
{32, 32},
{32, 32},
{32768, 1024, 32, 1},
{32768, 1024, 32, 1},
{32768, 1024, 32, 1},
{32768, 1024, 32, 1}},
{{16, 16},
{32, 32},
{16, 16},
{4096, 256, 16, 1},
{16, 1, 8192, 256},
{16384, 1024, 32, 1},
{16384, 1024, 32, 1}}};
std::vector<ck::index_t> init_methods = {0, 1, 2};
std::unique_ptr<CDElementOp> p_cd_element_op;
void Run()
{
for(auto& memory_params : list_of_memory_params)
{
for(const ck::index_t init_method : init_methods)
{
bool pass =
ck::profiler::profile_contraction_impl<ALayout,
BLayout,
CDLayout,
DataType,
DTupleDataType,
CDElementOp>(true /*do_verification*/,
init_method,
false /*do_logs*/,
false /*time_kernel*/,
*p_cd_element_op,
memory_params.M,
memory_params.N,
memory_params.K,
memory_params.StridesA,
memory_params.StridesB,
memory_params.StridesC,
memory_params.StridesD);
EXPECT_TRUE(pass);
}
}
}
};
template <typename Tuple>
class TestContractionScale : public TestContraction<Tuple>
{
};
template <typename Tuple>
class TestContractionBilinear : public TestContraction<Tuple>
{
};
using BilinearKernelTypes =
::testing::Types<std::tuple<Row, Row, Row, F32, ck::Tuple<F32>, Bilinear>,
std::tuple<Row, Col, Row, F32, ck::Tuple<F32>, Bilinear>,
std::tuple<Col, Row, Row, F32, ck::Tuple<F32>, Bilinear>,
std::tuple<Col, Col, Row, F32, ck::Tuple<F32>, Bilinear>,
std::tuple<Row, Row, Row, F64, ck::Tuple<F32>, Bilinear>,
std::tuple<Row, Col, Row, F64, ck::Tuple<F32>, Bilinear>,
std::tuple<Col, Row, Row, F64, ck::Tuple<F32>, Bilinear>,
std::tuple<Col, Col, Row, F64, ck::Tuple<F32>, Bilinear>>;
using ScaleKernelTypes = ::testing::Types<std::tuple<Row, Row, Row, F32, ck::Tuple<>, Scale>,
std::tuple<Row, Col, Row, F32, ck::Tuple<>, Scale>,
std::tuple<Col, Row, Row, F32, ck::Tuple<>, Scale>,
std::tuple<Col, Col, Row, F32, ck::Tuple<>, Scale>,
std::tuple<Row, Row, Row, F64, ck::Tuple<>, Scale>,
std::tuple<Row, Col, Row, F64, ck::Tuple<>, Scale>,
std::tuple<Col, Row, Row, F64, ck::Tuple<>, Scale>,
std::tuple<Col, Col, Row, F64, ck::Tuple<>, Scale>>;
TYPED_TEST_SUITE(TestContractionBilinear, BilinearKernelTypes);
TYPED_TEST_SUITE(TestContractionScale, ScaleKernelTypes);
TYPED_TEST(TestContractionBilinear, bilinear)
{
this->p_cd_element_op = std::make_unique<Bilinear>(1.f, 1.f);
this->Run();
this->p_cd_element_op = std::make_unique<Bilinear>(-0.5f, 0.5f);
this->Run();
}
TYPED_TEST(TestContractionScale, scale)
{
this->p_cd_element_op = std::make_unique<Scale>(1.f);
this->Run();
this->p_cd_element_op = std::make_unique<Scale>(0.5f);
this->Run();
}
test/contraction/test_contraction_interface.cpp 0 → 100644
View file @ 2f463a94
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#include <stdexcept>
#include <vector>
#include "gtest/gtest.h"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/device_contraction_multiple_d.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp"
#include "ck/library/tensor_operation_instance/gpu/contraction_bilinear.hpp"
#include "ck/library/utility/device_memory.hpp"
using Pass = ck::tensor_operation::element_wise::PassThrough;
using Bilinear = ck::tensor_operation::element_wise::Bilinear;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F32 = float;
using F64 = double;
template <ck::index_t ABlockTransferSrcVectorDim,
ck::index_t BBlockTransferSrcVectorDim,
ck::index_t CDEBlockTransferScalarPerVector>
class ContractionInstanceWrapper
{
public:
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
static constexpr ck::index_t NumDim = 2;
// clang-format off
using ContractionDeviceInstance = ck::tensor_operation::device::
//#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//#####################################| | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//#####################################| | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceContractionMultipleD_Xdl_CShuffle< NumDim, NumDim, NumDim, F32, F32, F32, F32, ck::Tuple<F32>, F32, Pass, Pass, Bilinear, GemmSpec, 1, 256, 256, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, ABlockTransferSrcVectorDim, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, BBlockTransferSrcVectorDim, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, CDEBlockTransferScalarPerVector>;
// clang-format on
bool isSupported(std::vector<ck::index_t>& ADims,
std::vector<ck::index_t>& BDims,
std::vector<ck::index_t>& DDims,
std::vector<ck::index_t>& EDims,
std::vector<ck::index_t>& AStrides,
std::vector<ck::index_t>& BStrides,
std::vector<ck::index_t>& DStrides,
std::vector<ck::index_t>& EStrides) const
{
auto contraction = ContractionDeviceInstance{};
auto argument = contraction.MakeArgument(nullptr,
nullptr,
std::array<const void*, 1>{nullptr},
nullptr,
ADims,
AStrides,
BDims,
BStrides,
std::array<std::vector<ck::index_t>, 1>{DDims},
std::array<std::vector<ck::index_t>, 1>{DStrides},
EDims,
EStrides,
Pass{},
Pass{},
Bilinear{1.f, 1.f});
return contraction.IsSupportedArgument(argument);
}
};
template <typename DataTypeA,
typename DataTypeB,
typename DataTypeC,
typename DataTypeD,
ck::index_t NumDim>
class ContractionDeviceOpWrapper
{
protected:
using DeviceOp = ck::tensor_operation::device::DeviceContractionMultipleD<NumDim,
NumDim,
NumDim,
DataTypeA,
DataTypeB,
ck::Tuple<DataTypeC>,
DataTypeD,
Pass,
Pass,
Bilinear>;
public:
bool IsSupportedInstance(std::vector<ck::index_t>& Dims,
std::vector<ck::index_t>& Strides) const
{
bool supported = false;
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
for(auto& op_ptr : op_ptrs)
{
auto argument_ptr =
op_ptr->MakeArgumentPointer(nullptr,
nullptr,
std::array<const void*, 1>{nullptr},
nullptr,
Dims,
Strides,
Dims,
Strides,
std::array<std::vector<ck::index_t>, 1>{Dims},
std::array<std::vector<ck::index_t>, 1>{Strides},
Dims,
Strides,
Pass{},
Pass{},
Bilinear{1.f, 1.f});
supported = supported || op_ptr->IsSupportedArgument(argument_ptr.get());
}
return supported;
}
};
TEST(TestContractionInterface, IncorrectNumDims)
{
std::vector<std::vector<ck::index_t>> Dims = {{4, 4}, {4, 4, 4, 4}, {4, 4, 4, 4, 4, 4}};
std::vector<std::vector<ck::index_t>> Strides = {{1, 1}, {1, 1, 1, 1}, {1, 1, 1, 1, 1, 1}};
ContractionDeviceOpWrapper<F32, F32, F32, F32, 1> wrapper_1d;
ContractionDeviceOpWrapper<F32, F32, F32, F32, 2> wrapper_2d;
ContractionDeviceOpWrapper<F32, F32, F32, F32, 3> wrapper_3d;
EXPECT_FALSE(wrapper_1d.IsSupportedInstance(Dims[0], Strides[0]));
EXPECT_TRUE(wrapper_2d.IsSupportedInstance(Dims[1], Strides[1]));
EXPECT_FALSE(wrapper_3d.IsSupportedInstance(Dims[2], Strides[2]));
}
TEST(TestContractionInterface, IncorrectDataTypes)
{
std::vector<ck::index_t> Dims = {4, 4, 4, 4};
std::vector<ck::index_t> Strides = {64, 16, 4, 1};
ContractionDeviceOpWrapper<F32, F32, F64, F64, 2> wrapper_1;
ContractionDeviceOpWrapper<F64, F64, F32, F32, 2> wrapper_2;
EXPECT_FALSE(wrapper_1.IsSupportedInstance(Dims, Strides));
EXPECT_FALSE(wrapper_2.IsSupportedInstance(Dims, Strides));
}
TEST(TestContractionSupportedArgs, ABMemoryAccess)
{
std::vector<ck::index_t> Dims = {4, 4, 4, 4};
std::vector<ck::index_t> Strides = {64, 16, 4, 1};
std::vector<ck::index_t> StridesM1 = {4, 1, 64, 16};
std::vector<ck::index_t> StridesK1 = {64, 16, 4, 1};
std::vector<ck::index_t> InvalidStrides = {4, 4, 4, 4};
// Memory access to A
ContractionInstanceWrapper<1, 2, 4> wrapperA1;
ContractionInstanceWrapper<2, 2, 4> wrapperA2;
EXPECT_FALSE(
wrapperA1.isSupported(Dims, Dims, Dims, Dims, InvalidStrides, Strides, Strides, Strides));
EXPECT_FALSE(
wrapperA2.isSupported(Dims, Dims, Dims, Dims, InvalidStrides, Strides, Strides, Strides));
EXPECT_TRUE(
wrapperA1.isSupported(Dims, Dims, Dims, Dims, StridesM1, Strides, Strides, Strides));
EXPECT_TRUE(
wrapperA2.isSupported(Dims, Dims, Dims, Dims, StridesK1, Strides, Strides, Strides));
// Memory access to B
ContractionInstanceWrapper<2, 1, 4> wrapperB1;
ContractionInstanceWrapper<2, 2, 4> wrapperB2;
EXPECT_FALSE(
wrapperB1.isSupported(Dims, Dims, Dims, Dims, Strides, InvalidStrides, Strides, Strides));
EXPECT_FALSE(
wrapperB2.isSupported(Dims, Dims, Dims, Dims, Strides, InvalidStrides, Strides, Strides));
EXPECT_TRUE(
wrapperB1.isSupported(Dims, Dims, Dims, Dims, Strides, StridesM1, Strides, Strides));
EXPECT_TRUE(
wrapperB2.isSupported(Dims, Dims, Dims, Dims, Strides, StridesK1, Strides, Strides));
}
TEST(TestContractionSupportedArgs, DEMemoryAccess)
{
std::vector<ck::index_t> Dims = {4, 4, 4, 4};
std::vector<ck::index_t> Strides = {64, 16, 4, 1};
std::vector<ck::index_t> InvalidStrides = {64, 16, 1, 4};
ContractionInstanceWrapper<2, 2, 4> wrapper;
// Memory access to D
EXPECT_FALSE(
wrapper.isSupported(Dims, Dims, Dims, Dims, Strides, Strides, InvalidStrides, Strides));
EXPECT_TRUE(wrapper.isSupported(Dims, Dims, Dims, Dims, Strides, Strides, Strides, Strides));
// Memory access to E
EXPECT_FALSE(
wrapper.isSupported(Dims, Dims, Dims, Dims, Strides, Strides, Strides, InvalidStrides));
EXPECT_TRUE(wrapper.isSupported(Dims, Dims, Dims, Dims, Strides, Strides, Strides, Strides));
}
test/convnd_bwd_data/CMakeLists.txt
View file @ 2f463a94
add_gtest_executable(test_convnd_bwd_data convnd_bwd_data.cpp) if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940")
target_link_libraries(test_convnd_bwd_data PRIVATE utility device_conv1d_bwd_data_instance device_conv2d_bwd_data_instance device_conv3d_bwd_data_instance) add_gtest_executable(test_convnd_bwd_data convnd_bwd_data.cpp)
target_link_libraries(test_convnd_bwd_data PRIVATE utility device_conv1d_bwd_data_instance device_conv2d_bwd_data_instance device_conv3d_bwd_data_instance)
endif()
\ No newline at end of file
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment