ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-16 10:59:55 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge commit '87dd073887933fc2c75c234871e3885cee970a98' into develop

Browse Source
This commit is contained in:
assistant-librarian[bot]
2025-12-18 00:34:53 +00:00
parent 3c59d702ca
commit 334ae1c494
82 changed files with 7696 additions and 622 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
test/grouped_convnd_bwd_weight/CMakeLists.txt
View File

@@ -5,16 +5,19 @@ if(GPU_TARGETS MATCHES "gfx9|gfx11|gfx12")
add_gtest_executable(test_grouped_convnd_bwd_weight test_grouped_convnd_bwd_weight.cpp)
target_link_libraries(test_grouped_convnd_bwd_weight PRIVATE utility device_grouped_conv1d_bwd_weight_instance device_grouped_conv2d_bwd_weight_instance device_grouped_conv3d_bwd_weight_instance device_grouped_convnd_bwd_weight_instance)
add_gtest_executable(test_grouped_convnd_bwd_weight_bilinear test_grouped_convnd_bwd_weight_bilinear.cpp)
target_link_libraries(test_grouped_convnd_bwd_weight_bilinear PRIVATE utility device_grouped_conv3d_bwd_weight_bilinear_instance)
add_gtest_executable(test_grouped_convnd_bwd_weight_scale test_grouped_convnd_bwd_weight_scale.cpp)
target_link_libraries(test_grouped_convnd_bwd_weight_scale PRIVATE utility device_grouped_conv3d_bwd_weight_scale_instance)
add_executable(test_grouped_convnd_bwd_weight_dataset_xdl test_grouped_convnd_bwd_weight_dataset_xdl.cpp)
target_compile_options(test_grouped_convnd_bwd_weight_dataset_xdl PRIVATE -Wno-global-constructors -Wno-undef)
target_link_libraries(test_grouped_convnd_bwd_weight_dataset_xdl PRIVATE gtest_main getopt::getopt utility device_grouped_conv1d_bwd_weight_instance device_grouped_conv2d_bwd_weight_instance device_grouped_conv3d_bwd_weight_instance device_grouped_convnd_bwd_weight_instance)
elseif(DL_KERNELS)
add_gtest_executable(test_grouped_convnd_bwd_weight test_grouped_convnd_bwd_weight.cpp)
target_link_libraries(test_grouped_convnd_bwd_weight PRIVATE utility device_grouped_conv1d_bwd_weight_instance device_grouped_conv2d_bwd_weight_instance device_grouped_conv3d_bwd_weight_instance)
elseif(GPU_TARGETS MATCHES "gfx11")
add_gtest_executable(test_grouped_convnd_bwd_weight test_grouped_convnd_bwd_weight.cpp)
target_link_libraries(test_grouped_convnd_bwd_weight PRIVATE utility device_grouped_conv3d_bwd_weight_instance)
endif()
add_gtest_executable(test_grouped_convnd_bwd_weight_interface_xdl test_grouped_convnd_bwd_weight_interface_xdl.cpp)
if(result EQUAL 0)
target_link_libraries(test_grouped_convnd_bwd_weight_interface_xdl PRIVATE utility)
@@ -27,7 +30,3 @@ add_gtest_executable(test_grouped_convnd_bwd_weight_interface_wmma test_grouped_
if(result EQUAL 0)
target_link_libraries(test_grouped_convnd_bwd_weight_interface_wmma PRIVATE utility)
endif()
add_gtest_executable(test_grouped_conv_bwd_weight_xdl_bilinear test_grouped_conv_bwd_weight_xdl_bilinear.cpp)
if(result EQUAL 0)
target_link_libraries(test_grouped_conv_bwd_weight_xdl_bilinear PRIVATE utility device_grouped_conv3d_bwd_weight_bilinear_instance)
endif()

38
test/grouped_convnd_bwd_weight/test_grouped_convnd_bwd_weight.cpp
View File

@@ -46,44 +46,6 @@ class TestGroupedConvndBwdWeight : public ::testing::Test
return true;
}
}
if(ck::is_gfx11_supported() || ck::is_gfx12_supported())
{
// on gfx11 only support for 3d is implemented
if constexpr(NDimSpatial{} != 3)
{
return true;
}
// on gfx11 only support for i8 and fp16 is implemented
if constexpr(!((std::is_same_v<InDataType, int8_t> &&
std::is_same_v<WeiDataType, int8_t> &&
std::is_same_v<OutDataType, int8_t>) ||
(std::is_same_v<InDataType, ck::half_t> &&
std::is_same_v<WeiDataType, ck::half_t> &&
std::is_same_v<OutDataType, ck::half_t>)))
{
return true;
}
// WMMA kernel is only supported for split_k=1
if(split_k != 1)
{
return true;
}
// Skip due to the lack of kernels for NGCDHW
if constexpr(std::is_same_v<InLayout, NGCW> || std::is_same_v<InLayout, NGCHW> ||
std::is_same_v<InLayout, NGCDHW>)
{
return true;
}
}
else
{
// support for i8 is only implemented on gfx11
if constexpr(std::is_same_v<InDataType, int8_t> &&
std::is_same_v<WeiDataType, int8_t> && std::is_same_v<OutDataType, int8_t>)
{
return true;
}
}
return false;
}

30
test/grouped_convnd_bwd_weight/test_grouped_conv_bwd_weight_xdl_bilinear.cpp → test/grouped_convnd_bwd_weight/test_grouped_convnd_bwd_weight_bilinear.cpp
View File

@@ -212,7 +212,34 @@ class TestGroupedConvndBwdWeight : public ::testing::Test
}
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr});
wei_device_buf.FromDevice(wei_device.mData.data());
passed &= ck::utils::check_err(wei_device, wei_host, "Error: incorrect results!");
using AccDataType = float;
float max_accumulated_value =
*std::max_element(wei_host.mData.begin(), wei_host.mData.end());
const ck::index_t num_accums = out.GetElementSize() / conv_param.K_;
const ck::index_t num_accums_split_k = split_k;
double rtol =
ck::utils::get_relative_threshold<InDataType, WeiDataType, AccDataType>(
num_accums / num_accums_split_k);
double atol =
ck::utils::get_absolute_threshold<InDataType, WeiDataType, AccDataType>(
max_accumulated_value / num_accums_split_k,
num_accums / num_accums_split_k);
// Calculate error due to split_k accumulation
auto rtol_split_k =
ck::utils::get_relative_threshold<WeiDataType, WeiDataType, WeiDataType>(
num_accums_split_k);
auto atol_split_k =
ck::utils::get_absolute_threshold<WeiDataType, WeiDataType, WeiDataType>(
max_accumulated_value, num_accums_split_k);
// Use higher threshold
rtol = std::max(rtol, rtol_split_k);
atol = std::max(atol, atol_split_k);
passed &= ck::utils::check_err(
wei_device, wei_host, "Error: incorrect results!", rtol, atol);
std::size_t flop =
conv_param.GetFlops() +
@@ -236,6 +263,7 @@ class TestGroupedConvndBwdWeight : public ::testing::Test
std::cout << "grouped_conv_bwd_weight_instance (" << instance_index << "/" << num_kernel
<< "): Passed" << std::endl;
}
printf("\033[36mvalids: %d\033[0m\n", num_kernel);
return passed;
}

294
test/grouped_convnd_bwd_weight/test_grouped_convnd_bwd_weight_scale.cpp Normal file
View File

@@ -0,0 +1,294 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
#include <algorithm>
#include <iomanip>
#include <iostream>
#include <iterator>
#include <typeinfo>
#include <gtest/gtest.h>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_conv_bwd_weight_multiple_d.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight_scale.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/convolution_parameter.hpp"
#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_conv_bwd_weight.hpp"
template <typename Tuple>
class TestGroupedConvndBwdWeight : public ::testing::Test
{
protected:
using InDataType = std::tuple_element_t<0, Tuple>;
using WeiDataType = std::tuple_element_t<1, Tuple>;
using OutDataType = std::tuple_element_t<2, Tuple>;
using InLayout = ck::tensor_layout::convolution::NDHWGC;
using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
using OutLayout = ck::tensor_layout::convolution::NDHWGK;
using InElementOp = ck::tensor_operation::element_wise::PassThrough;
using WeiElementOp = ck::tensor_operation::element_wise::Scale;
using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
static constexpr ck::index_t NDimSpatial = std::tuple_element_t<3, Tuple>{};
static constexpr float alpha = 2.f;
std::vector<ck::utils::conv::ConvParam> conv_params;
std::vector<ck::index_t> split_ks{1, 2};
void RunReference(ck::utils::conv::ConvParam& conv_param,
ck::Tensor<InDataType>& in,
ck::Tensor<WeiDataType>& wei_host,
ck::Tensor<OutDataType>& out)
{
auto ref_conv =
ck::tensor_operation::host::ReferenceConvBwdWeight<NDimSpatial,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
0, /*Num A Elementwise Tensors*/
0, /*Num B Elementwise Tensors*/
0> /*Num D Elementwise Tensors*/
{};
auto ref_invoker = ref_conv.MakeInvoker();
auto ref_argument = ref_conv.MakeArgument(in,
wei_host,
out,
conv_param.conv_filter_strides_,
conv_param.conv_filter_dilations_,
conv_param.input_left_pads_,
conv_param.input_right_pads_,
InElementOp{},
WeiElementOp{alpha},
OutElementOp{},
{},
{},
{});
ref_invoker.Run(ref_argument);
}
bool PerformConvWeightScale(ck::utils::conv::ConvParam& conv_param, const ck::index_t split_k)
{
bool passed = true;
const auto in_g_n_c_wis_desc =
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
conv_param);
const auto wei_g_k_c_xs_desc =
ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
conv_param);
const auto out_g_n_k_wos_desc =
ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
conv_param);
ck::Tensor<InDataType> in(in_g_n_c_wis_desc);
ck::Tensor<OutDataType> out(out_g_n_k_wos_desc);
ck::Tensor<WeiDataType> wei_host(wei_g_k_c_xs_desc);
ck::Tensor<WeiDataType> wei_device(wei_g_k_c_xs_desc);
std::cout << "in: " << in.mDesc << std::endl;
std::cout << "wei: " << wei_host.mDesc << std::endl;
std::cout << "out: " << out.mDesc << std::endl;
in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
out.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
ck::DeviceMem in_device_buf(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());
ck::DeviceMem out_device_buf(sizeof(OutDataType) * out.mDesc.GetElementSpaceSize());
ck::DeviceMem wei_device_buf(sizeof(WeiDataType) * wei_device.mDesc.GetElementSpaceSize());
in_device_buf.ToDevice(in.mData.data());
wei_device_buf.ToDevice(wei_device.mData.data());
out_device_buf.ToDevice(out.mData.data());
std::array<ck::index_t, NDimSpatial + 3> b_g_n_c_wis_lengths{};
std::array<ck::index_t, NDimSpatial + 3> b_g_n_c_wis_strides{};
std::array<ck::index_t, NDimSpatial + 3> e_g_k_c_xs_lengths{};
std::array<ck::index_t, NDimSpatial + 3> e_g_k_c_xs_strides{};
std::array<ck::index_t, NDimSpatial + 3> a_g_n_k_wos_lengths{};
std::array<ck::index_t, NDimSpatial + 3> a_g_n_k_wos_strides{};
std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
std::array<ck::index_t, NDimSpatial> input_left_pads{};
std::array<ck::index_t, NDimSpatial> input_right_pads{};
auto copy = [](const auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
copy(in_g_n_c_wis_desc.GetLengths(), b_g_n_c_wis_lengths);
copy(in_g_n_c_wis_desc.GetStrides(), b_g_n_c_wis_strides);
copy(wei_g_k_c_xs_desc.GetLengths(), e_g_k_c_xs_lengths);
copy(wei_g_k_c_xs_desc.GetStrides(), e_g_k_c_xs_strides);
copy(out_g_n_k_wos_desc.GetLengths(), a_g_n_k_wos_lengths);
copy(out_g_n_k_wos_desc.GetStrides(), a_g_n_k_wos_strides);
copy(conv_param.conv_filter_strides_, conv_filter_strides);
copy(conv_param.conv_filter_dilations_, conv_filter_dilations);
copy(conv_param.input_left_pads_, input_left_pads);
copy(conv_param.input_right_pads_, input_right_pads);
RunReference(conv_param, in, wei_host, out);
using DeviceOp =
ck::tensor_operation::device::DeviceGroupedConvBwdWeightMultipleD<NDimSpatial,
InLayout,
WeiLayout,
OutLayout,
ck::Tuple<>,
InDataType,
WeiDataType,
OutDataType,
ck::Tuple<>,
InElementOp,
WeiElementOp,
OutElementOp>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
int num_kernel = 0;
for(std::size_t i = 0; i < op_ptrs.size(); ++i)
{
auto& op_ptr = op_ptrs[i];
auto argument_ptr = op_ptr->MakeArgumentPointer(
static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
std::array<const void*, 0>{},
b_g_n_c_wis_lengths,
b_g_n_c_wis_strides,
e_g_k_c_xs_lengths,
e_g_k_c_xs_strides,
a_g_n_k_wos_lengths,
a_g_n_k_wos_strides,
std::array<std::array<ck::index_t, NDimSpatial + 3>, 0>{},
std::array<std::array<ck::index_t, NDimSpatial + 3>, 0>{},
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
InElementOp{},
WeiElementOp{alpha},
OutElementOp{},
split_k);
ck::DeviceMem workspace_buf(op_ptr->GetWorkSpaceSize(argument_ptr.get()));
op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_buf.GetDeviceBuffer());
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
num_kernel++;
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr});
wei_device_buf.FromDevice(wei_device.mData.data());
using AccDataType = float;
float max_accumulated_value =
*std::max_element(wei_host.mData.begin(), wei_host.mData.end());
const ck::index_t num_accums = out.GetElementSize() / conv_param.K_;
const ck::index_t num_accums_split_k = split_k;
double rtol =
ck::utils::get_relative_threshold<InDataType, WeiDataType, AccDataType>(
num_accums / num_accums_split_k);
double atol =
ck::utils::get_absolute_threshold<InDataType, WeiDataType, AccDataType>(
max_accumulated_value / num_accums_split_k,
num_accums / num_accums_split_k);
// Calculate error due to split_k accumulation
auto rtol_split_k =
ck::utils::get_relative_threshold<WeiDataType, WeiDataType, WeiDataType>(
num_accums_split_k);
auto atol_split_k =
ck::utils::get_absolute_threshold<WeiDataType, WeiDataType, WeiDataType>(
max_accumulated_value, num_accums_split_k);
// Use higher threshold
rtol = std::max(rtol, rtol_split_k);
atol = std::max(atol, atol_split_k);
passed &= ck::utils::check_err(
wei_device, wei_host, "Error: incorrect results!", rtol, atol);
std::size_t flop =
conv_param.GetFlops() +
3 * conv_param.GetOutputByte<WeiDataType>() / sizeof(WeiDataType);
std::size_t num_bytes = conv_param.GetByte<InDataType, WeiDataType, OutDataType>() +
conv_param.GetOutputByte<WeiDataType>();
float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
float gb_per_sec = num_bytes / 1.E6 / avg_time;
std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops
<< " TFlops, " << gb_per_sec << " GB/s, " << op_name << std::endl;
}
else
{
std::cerr << op_name << " does not support this problem" << std::endl;
}
}
printf("\033[36mvalids: %d\033[0m\n", num_kernel);
return passed;
}
void Run()
{
EXPECT_FALSE(conv_params.empty());
bool pass = true;
for(auto split_k : split_ks)
{
for(auto& param : conv_params)
{
pass = pass && PerformConvWeightScale(param, split_k);
}
}
EXPECT_TRUE(pass);
}
};
template <typename Tuple>
class TestGroupedConvndBwdWeight3d : public TestGroupedConvndBwdWeight<Tuple>
{
};
using KernelTypes3d =
::testing::Types<std::tuple<float, float, float, ck::Number<3>>,
std::tuple<ck::half_t, ck::half_t, ck::half_t, ck::Number<3>>,
std::tuple<ck::bhalf_t, float, ck::bhalf_t, ck::Number<3>>>;
TYPED_TEST_SUITE(TestGroupedConvndBwdWeight3d, KernelTypes3d);
TYPED_TEST(TestGroupedConvndBwdWeight3d, Test3D)
{
this->conv_params.clear();
this->conv_params.push_back(
{3, 2, 16, 128, 128, {1, 1, 1}, {7, 7, 7}, {2, 2, 2}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
this->conv_params.push_back(
{3, 2, 2, 128, 128, {3, 3, 3}, {14, 14, 3}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
this->conv_params.push_back(
{3, 2, 32, 128, 128, {1, 1, 1}, {3, 3, 3}, {1, 1, 1}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
this->conv_params.push_back(
{3, 1, 1, 1, 32, {3, 3, 3}, {32, 32, 32}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
this->conv_params.push_back(
{3, 1, 1, 64, 3, {3, 3, 3}, {32, 32, 32}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
this->conv_params.push_back(
{3, 1, 1, 1, 1, {3, 3, 3}, {32, 32, 32}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
this->conv_params.push_back(
{3, 1, 1, 4, 4, {3, 3, 3}, {14, 28, 28}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
this->Run();
}