ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-17 03:19:48 +00:00
Code Issues Packages Projects Releases Wiki Activity

Enable multiple D for grouped conv fwd large tensors (#2572)

Browse Source 
[ROCm/composable_kernel commit: 5b244105d9]
This commit is contained in:
Bartłomiej Kocot
2025-07-28 22:39:07 +02:00
committed by GitHub
parent d3a1842669
commit f25da17c36
7 changed files with 377 additions and 148 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
test/grouped_convnd_fwd_activation/CMakeLists.txt
View File

@@ -7,4 +7,8 @@ if(GPU_TARGETS MATCHES "gfx9")
add_gtest_executable(test_grouped_convnd_fwd_clamp test_grouped_convnd_fwd_clamp.cpp)
target_link_libraries(test_grouped_convnd_fwd_clamp PRIVATE utility device_grouped_conv2d_fwd_clamp_instance device_grouped_conv3d_fwd_clamp_instance)
add_executable(test_grouped_convnd_fwd_bias_clamp_large_cases test_grouped_convnd_fwd_bias_clamp_large_cases.cpp)
target_compile_options(test_grouped_convnd_fwd_bias_clamp_large_cases PRIVATE -Wno-global-constructors -Wno-undef)
target_link_libraries(test_grouped_convnd_fwd_bias_clamp_large_cases PRIVATE gtest_main getopt::getopt utility device_grouped_conv2d_fwd_bias_clamp_instance device_grouped_conv3d_fwd_bias_clamp_instance)
endif()

135
test/grouped_convnd_fwd_activation/test_grouped_convnd_fwd_bias_clamp_large_cases.cpp Normal file
View File

@@ -0,0 +1,135 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <iostream>
#include <initializer_list>
#include <vector>
#include <gtest/gtest.h>
#include "profiler/profile_grouped_conv_fwd_bias_clamp_impl.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
using AddClamp = ck::tensor_operation::element_wise::AddClamp;
template <typename Tuple>
class TestGroupedConvndFwd : public ::testing::Test
{
protected:
using DataType = std::tuple_element_t<0, Tuple>;
using InLayout = std::tuple_element_t<1, Tuple>;
using WeiLayout = std::tuple_element_t<2, Tuple>;
using OutLayout = std::tuple_element_t<3, Tuple>;
using IndexType = ck::long_index_t;
std::vector<ck::utils::conv::ConvParam> conv_params;
template <ck::index_t NDimSpatial>
void Run()
{
EXPECT_FALSE(conv_params.empty());
bool pass = true;
for(auto& param : conv_params)
{
pass = pass && ck::profiler::profile_grouped_conv_fwd_bias_clamp_impl<NDimSpatial,
InLayout,
WeiLayout,
OutLayout,
DataType,
DataType,
DataType,
DataType,
DataType,
IndexType,
false /*BiasGK*/>(
true, // do_verification
1, // init_method: integer value
false, // do_log
false, // time_kernel
param);
}
EXPECT_TRUE(pass);
}
};
using namespace ck::tensor_layout::convolution;
using KernelTypes2d = ::testing::Types<std::tuple<ck::bhalf_t, NHWGC, GKYXC, NHWGK>,
std::tuple<float, NHWGC, GKYXC, NHWGK>,
std::tuple<ck::half_t, NHWGC, GKYXC, NHWGK>>;
using KernelTypes3d = ::testing::Types<std::tuple<ck::bhalf_t, NDHWGC, GKZYXC, NDHWGK>,
std::tuple<float, NDHWGC, GKZYXC, NDHWGK>,
std::tuple<ck::half_t, NDHWGC, GKZYXC, NDHWGK>>;
template <typename Tuple>
class TestGroupedConvndFwdBiasClamp2d : public TestGroupedConvndFwd<Tuple>
{
};
template <typename Tuple>
class TestGroupedConvndFwdBiasClamp3d : public TestGroupedConvndFwd<Tuple>
{
};
TYPED_TEST_SUITE(TestGroupedConvndFwdBiasClamp2d, KernelTypes2d);
TYPED_TEST_SUITE(TestGroupedConvndFwdBiasClamp3d, KernelTypes3d);
TYPED_TEST(TestGroupedConvndFwdBiasClamp2d, Test2D)
{
// Case larger than 2GB
this->conv_params.push_back(
{2, 1, 128, 4, 192, {2, 2}, {224, 224}, {224, 224}, {1, 1}, {0, 0}, {0, 0}});
// With supported NumGroupsToMerge > 1
this->conv_params.push_back(
{2, 32, 64, 1, 1, {2, 2}, {672, 672}, {672, 672}, {1, 1}, {0, 0}, {0, 0}});
// When image is larger than 2GB
this->conv_params.push_back(
{2, 2, 2, 128, 128, {3, 3}, {4096, 2048}, {300, 300}, {3, 3}, {1, 1}, {1, 1}});
// Split N and G > 1
this->conv_params.push_back(
{2, 4, 112, 8, 8, {3, 3}, {469, 724}, {2, 2}, {2, 2}, {1, 1}, {1, 1}});
this->template Run<2>();
}
TYPED_TEST(TestGroupedConvndFwdBiasClamp3d, Test3D)
{
// Case larger than 2GB
this->conv_params.push_back({3,
1,
128,
4,
192,
{2, 2, 2},
{2, 224, 224},
{1, 224, 224},
{1, 1, 1},
{0, 0, 0},
{0, 0, 0}});
// With supported NumGroupsToMerge > 1
this->conv_params.push_back({3,
32,
64,
1,
1,
{2, 2, 2},
{360, 2, 672},
{360, 2, 672},
{1, 1, 1},
{0, 0, 0},
{0, 0, 0}});
// When image is larger than 2GB
this->conv_params.push_back({3,
1,
2,
128,
128,
{3, 1, 3},
{900, 2, 2048},
{300, 1, 300},
{3, 2, 3},
{1, 1, 1},
{1, 1, 1}});
this->template Run<3>();
}