ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-19 12:30:16 +00:00
Code Issues Packages Projects Releases Wiki Activity

Add performance and large tensor tests for grouped conv (#1456)

Browse Source 
* Add performance and large tensor tests for grouped conv

* Resize tests

* Resize tests

* update the python script to parse the grouped_conv results

* Remove int8 tests

* change bwd wei layout

---------

Co-authored-by: illsilin <Illia.Silin@amd.com>

[ROCm/composable_kernel commit: 2581727d2a]
This commit is contained in:
Bartłomiej Kocot
2024-08-16 16:48:30 +02:00
committed by GitHub
parent e77e18da19
commit dffd5eacc0
11 changed files with 305 additions and 90 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
test/grouped_convnd_fwd/CMakeLists.txt
View File

@@ -7,6 +7,12 @@ if(GPU_TARGETS MATCHES "gfx9" OR GPU_TARGETS MATCHES "gfx11")
endif()
endif()
if(GPU_TARGETS MATCHES "gfx9")
add_executable(test_grouped_convnd_fwd_large_cases_xdl test_grouped_convnd_fwd_large_cases_xdl.cpp)
target_compile_options(test_grouped_convnd_fwd_large_cases_xdl PRIVATE -Wno-global-constructors -Wno-undef)
target_link_libraries(test_grouped_convnd_fwd_large_cases_xdl PRIVATE gtest_main getopt::getopt utility device_grouped_conv1d_fwd_instance device_grouped_conv2d_fwd_instance device_grouped_conv3d_fwd_instance)
endif()
add_gtest_executable(test_grouped_convnd_fwd_multi_ab_interface test_grouped_convnd_fwd_multi_ab_interface.cpp)
if(result EQUAL 0)
target_link_libraries(test_grouped_convnd_fwd_multi_ab_interface PRIVATE utility)

65
test/grouped_convnd_fwd/test_grouped_convnd_fwd.cpp
View File

@@ -17,7 +17,7 @@ class TestGroupedConvndFwd : public ::testing::Test
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 = std::tuple_element_t<4, Tuple>;
using IndexType = ck::index_t;
std::vector<ck::utils::conv::ConvParam> conv_params;
@@ -50,31 +50,28 @@ class TestGroupedConvndFwd : public ::testing::Test
using namespace ck::tensor_layout::convolution;
using KernelTypes1d = ::testing::Types<std::tuple<float, GNWC, GKXC, GNWK, ck::index_t>,
std::tuple<ck::half_t, GNWC, GKXC, GNWK, ck::index_t>,
std::tuple<ck::bhalf_t, GNWC, GKXC, GNWK, ck::index_t>,
std::tuple<int8_t, GNWC, GKXC, GNWK, ck::index_t>>;
using KernelTypes1d = ::testing::Types<std::tuple<float, GNWC, GKXC, GNWK>,
std::tuple<ck::half_t, GNWC, GKXC, GNWK>,
std::tuple<ck::bhalf_t, GNWC, GKXC, GNWK>,
std::tuple<int8_t, GNWC, GKXC, GNWK>>;
using KernelTypes2d = ::testing::Types<std::tuple<float, GNHWC, GKYXC, GNHWK, ck::index_t>,
std::tuple<ck::half_t, GNHWC, GKYXC, GNHWK, ck::index_t>,
std::tuple<ck::bhalf_t, GNHWC, GKYXC, GNHWK, ck::index_t>,
std::tuple<int8_t, GNHWC, GKYXC, GNHWK, ck::index_t>,
std::tuple<float, NHWGC, GKYXC, NHWGK, ck::index_t>,
std::tuple<ck::half_t, NHWGC, GKYXC, NHWGK, ck::index_t>,
std::tuple<ck::bhalf_t, NHWGC, GKYXC, NHWGK, ck::index_t>,
std::tuple<int8_t, NHWGC, GKYXC, NHWGK, ck::index_t>>;
using KernelTypes2d = ::testing::Types<std::tuple<float, GNHWC, GKYXC, GNHWK>,
std::tuple<ck::half_t, GNHWC, GKYXC, GNHWK>,
std::tuple<ck::bhalf_t, GNHWC, GKYXC, GNHWK>,
std::tuple<int8_t, GNHWC, GKYXC, GNHWK>,
std::tuple<float, NHWGC, GKYXC, NHWGK>,
std::tuple<ck::half_t, NHWGC, GKYXC, NHWGK>,
std::tuple<ck::bhalf_t, NHWGC, GKYXC, NHWGK>,
std::tuple<int8_t, NHWGC, GKYXC, NHWGK>>;
using KernelTypes3d = ::testing::Types<std::tuple<float, GNDHWC, GKZYXC, GNDHWK, ck::index_t>,
std::tuple<ck::half_t, GNDHWC, GKZYXC, GNDHWK, ck::index_t>,
std::tuple<ck::bhalf_t, GNDHWC, GKZYXC, GNDHWK, ck::index_t>,
std::tuple<int8_t, GNDHWC, GKZYXC, GNDHWK, ck::index_t>,
std::tuple<float, NDHWGC, GKZYXC, NDHWGK, ck::index_t>,
std::tuple<ck::half_t, NDHWGC, GKZYXC, NDHWGK, ck::index_t>,
std::tuple<ck::bhalf_t, NDHWGC, GKZYXC, NDHWGK, ck::index_t>,
std::tuple<int8_t, NDHWGC, GKZYXC, NDHWGK, ck::index_t>>;
using KernelTypes2dLargeCases =
::testing::Types<std::tuple<float, NHWGC, GKYXC, NHWGK, ck::long_index_t>>;
using KernelTypes3d = ::testing::Types<std::tuple<float, GNDHWC, GKZYXC, GNDHWK>,
std::tuple<ck::half_t, GNDHWC, GKZYXC, GNDHWK>,
std::tuple<ck::bhalf_t, GNDHWC, GKZYXC, GNDHWK>,
std::tuple<int8_t, GNDHWC, GKZYXC, GNDHWK>,
std::tuple<float, NDHWGC, GKZYXC, NDHWGK>,
std::tuple<ck::half_t, NDHWGC, GKZYXC, NDHWGK>,
std::tuple<ck::bhalf_t, NDHWGC, GKZYXC, NDHWGK>,
std::tuple<int8_t, NDHWGC, GKZYXC, NDHWGK>>;
template <typename Tuple>
class TestGroupedConvndFwd1d : public TestGroupedConvndFwd<Tuple>
@@ -91,15 +88,9 @@ class TestGroupedConvndFwd3d : public TestGroupedConvndFwd<Tuple>
{
};
template <typename Tuple>
class TestGroupedConvndFwd2dLargeCases : public TestGroupedConvndFwd<Tuple>
{
};
TYPED_TEST_SUITE(TestGroupedConvndFwd1d, KernelTypes1d);
TYPED_TEST_SUITE(TestGroupedConvndFwd2d, KernelTypes2d);
TYPED_TEST_SUITE(TestGroupedConvndFwd3d, KernelTypes3d);
TYPED_TEST_SUITE(TestGroupedConvndFwd2dLargeCases, KernelTypes2dLargeCases);
TYPED_TEST(TestGroupedConvndFwd1d, Test1D)
{
@@ -149,17 +140,3 @@ TYPED_TEST(TestGroupedConvndFwd3d, Test3D)
{3, 96, 1, 1, 1, {3, 3, 3}, {4, 30, 160}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
this->template Run<3>();
}
TYPED_TEST(TestGroupedConvndFwd2dLargeCases, Test2DLargeCases)
{
// Case larger than 2GB
this->conv_params.push_back(
{2, 1, 64, 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, 1, 1, 256, 256, {3, 3}, {4096, 2048}, {1024, 1024}, {3, 3}, {1, 1}, {1, 1}});
this->template Run<2>();
}

127
test/grouped_convnd_fwd/test_grouped_convnd_fwd_large_cases_xdl.cpp Normal file
View File

@@ -0,0 +1,127 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, 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_impl.hpp"
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_impl<NDimSpatial,
InLayout,
WeiLayout,
OutLayout,
DataType,
DataType,
DataType,
DataType,
DataType,
IndexType>(
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<float, NHWGC, GKYXC, NHWGK>,
std::tuple<ck::half_t, NHWGC, GKYXC, NHWGK>,
std::tuple<ck::bhalf_t, NHWGC, GKYXC, NHWGK>>;
using KernelTypes3d = ::testing::Types<std::tuple<float, NDHWGC, GKZYXC, NDHWGK>,
std::tuple<ck::half_t, NDHWGC, GKZYXC, NDHWGK>,
std::tuple<ck::bhalf_t, NDHWGC, GKZYXC, NDHWGK>>;
template <typename Tuple>
class TestGroupedConvndFwd2d : public TestGroupedConvndFwd<Tuple>
{
};
template <typename Tuple>
class TestGroupedConvndFwd3d : public TestGroupedConvndFwd<Tuple>
{
};
TYPED_TEST_SUITE(TestGroupedConvndFwd2d, KernelTypes2d);
TYPED_TEST_SUITE(TestGroupedConvndFwd3d, KernelTypes3d);
TYPED_TEST(TestGroupedConvndFwd2d, 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}});
this->template Run<2>();
}
TYPED_TEST(TestGroupedConvndFwd3d, 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>();
}