Skip to content

GitLab

"fmoe/vscode:/vscode.git/clone" did not exist on "33fc3aca2108838ad84e2bf8af2ccc436ef57d5d"
Unverified Commit 261f1759 authored by Bartłomiej Kocot's avatar Bartłomiej Kocot Committed by GitHub
Browse files

Support large batch tensors in grouped conv bwd data (#1711) 

* Support large batch tensors in grouped conv bwd data

* Fix multiD

* fixes

* fixes

* fixes
parent 58e7f37f
Expand all Show whitespace changes
Inline Side-by-side
Showing with 1067 additions and 843 deletions
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_wmma_cshuffle.hpp
View file @ 261f1759
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once #pragma once
...@@ -106,8 +106,7 @@ struct DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle ...@@ -106,8 +106,7 @@ struct DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle
static constexpr auto I3 = Number<3>{}; static constexpr auto I3 = Number<3>{};
static constexpr index_t KPerBlock = K0PerBlock * K1; static constexpr index_t KPerBlock = K0PerBlock * K1;
static constexpr auto transform_conv_to_gemm = using ConvToGemmBwdDataTransform = TransformConvBwdDataToGemm_v1<NDimSpatial,
TransformConvBwdDataToGemm_v1<NDimSpatial,
ConvBackwardDataSpecialization, ConvBackwardDataSpecialization,
K1, K1,
K1, K1,
...@@ -115,80 +114,27 @@ struct DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle ...@@ -115,80 +114,27 @@ struct DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle
NPerBlock, NPerBlock,
KPerBlock, KPerBlock,
true /* DoPadGemmM */, true /* DoPadGemmM */,
true /* DoPadGemmN */>{}; true /* DoPadGemmN */,
ALayout,
BLayout,
ELayout>;
static auto GetDummyABDsEGridDescriptor() static auto
GetDummyABDsEGridDescriptor(const ConvToGemmBwdDataTransform& conv_to_gemm_transform)
{ {
const std::array<index_t, NDimSpatial + 3> dummy_tensor_lengths = {1}; const auto a_grid_desc_ak0_m_ak1 = conv_to_gemm_transform.MakeADescriptor_AK0_M_AK1();
const std::array<index_t, NDimSpatial + 3> dummy_tensor_strides = {1}; const auto b_grid_desc_bk0_n_bk1 = conv_to_gemm_transform.MakeBDescriptor_BK0_N_BK1();
const std::array<index_t, NDimSpatial> dummy_spatial_lengths = {1}; const auto ds_grid_desc_m_n =
generate_tuple([&](auto) { return conv_to_gemm_transform.MakeCDescriptor_M_N(); },
const auto a_grid_desc_ak0_m_ak1 =
transform_conv_to_gemm.template MakeADescriptor_AK0_M_AK1<ALayout>(
dummy_tensor_lengths,
dummy_tensor_strides,
dummy_tensor_lengths,
dummy_tensor_strides,
dummy_tensor_lengths,
dummy_tensor_strides,
dummy_spatial_lengths,
dummy_spatial_lengths,
dummy_spatial_lengths,
dummy_spatial_lengths,
dummy_spatial_lengths);
const auto b_grid_desc_bk0_n_bk1 =
transform_conv_to_gemm.template MakeBDescriptor_BK0_N_BK1<BLayout>(
dummy_tensor_lengths,
dummy_tensor_strides,
dummy_tensor_lengths,
dummy_tensor_strides,
dummy_tensor_lengths,
dummy_tensor_strides,
dummy_spatial_lengths,
dummy_spatial_lengths,
dummy_spatial_lengths,
dummy_spatial_lengths,
dummy_spatial_lengths);
const auto ds_grid_desc_m_n = generate_tuple(
[&](auto i) {
using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
return transform_conv_to_gemm.template MakeCDescriptor_M_N<DLayout>(
dummy_tensor_lengths,
dummy_tensor_strides,
dummy_tensor_lengths,
dummy_tensor_strides,
dummy_tensor_lengths,
dummy_tensor_strides,
dummy_spatial_lengths,
dummy_spatial_lengths,
dummy_spatial_lengths,
dummy_spatial_lengths,
dummy_spatial_lengths);
},
Number<NumDTensor>{}); Number<NumDTensor>{});
const auto e_grid_desc_m_n = conv_to_gemm_transform.MakeCDescriptor_M_N();
const auto e_grid_desc_m_n =
transform_conv_to_gemm.template MakeCDescriptor_M_N<ELayout>(dummy_tensor_lengths,
dummy_tensor_strides,
dummy_tensor_lengths,
dummy_tensor_strides,
dummy_tensor_lengths,
dummy_tensor_strides,
dummy_spatial_lengths,
dummy_spatial_lengths,
dummy_spatial_lengths,
dummy_spatial_lengths,
dummy_spatial_lengths);
return make_tuple( return make_tuple(
a_grid_desc_ak0_m_ak1, b_grid_desc_bk0_n_bk1, ds_grid_desc_m_n, e_grid_desc_m_n); a_grid_desc_ak0_m_ak1, b_grid_desc_bk0_n_bk1, ds_grid_desc_m_n, e_grid_desc_m_n);
} }
// desc // desc
using ABDsEGridDesc = decltype(GetDummyABDsEGridDescriptor()); constexpr static ConvToGemmBwdDataTransform dummy_conv_to_gemm_transform;
using ABDsEGridDesc = decltype(GetDummyABDsEGridDescriptor(dummy_conv_to_gemm_transform));
using AGridDesc_AK0_M_AK1 = remove_cvref_t<tuple_element_t<0, ABDsEGridDesc>>; using AGridDesc_AK0_M_AK1 = remove_cvref_t<tuple_element_t<0, ABDsEGridDesc>>;
using BGridDesc_BK0_N_BK1 = remove_cvref_t<tuple_element_t<1, ABDsEGridDesc>>; using BGridDesc_BK0_N_BK1 = remove_cvref_t<tuple_element_t<1, ABDsEGridDesc>>;
...@@ -270,7 +216,7 @@ struct DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle ...@@ -270,7 +216,7 @@ struct DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle
const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths, const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides, const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>&
ds_g_n_c_wis_lengths, /*ds_g_n_c_wis_lengths*/,
const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>&
ds_g_n_c_wis_strides, ds_g_n_c_wis_strides,
const std::array<index_t, NDimSpatial + 3>& e_g_n_c_wis_lengths, const std::array<index_t, NDimSpatial + 3>& e_g_n_c_wis_lengths,
...@@ -291,15 +237,8 @@ struct DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle ...@@ -291,15 +237,8 @@ struct DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle
b_element_op_{b_element_op}, b_element_op_{b_element_op},
cde_element_op_{cde_element_op}, cde_element_op_{cde_element_op},
a_g_n_k_wos_lengths_{a_g_n_k_wos_lengths}, a_g_n_k_wos_lengths_{a_g_n_k_wos_lengths},
a_g_n_k_wos_strides_{a_g_n_k_wos_strides},
b_g_k_c_xs_lengths_{b_g_k_c_xs_lengths}, b_g_k_c_xs_lengths_{b_g_k_c_xs_lengths},
b_g_k_c_xs_strides_{b_g_k_c_xs_strides},
ds_g_n_c_wis_lengths_{ds_g_n_c_wis_lengths},
ds_g_n_c_wis_strides_{ds_g_n_c_wis_strides},
e_g_n_c_wis_lengths_{e_g_n_c_wis_lengths},
e_g_n_c_wis_strides_{e_g_n_c_wis_strides},
conv_filter_strides_{conv_filter_strides}, conv_filter_strides_{conv_filter_strides},
conv_filter_dilations_{conv_filter_dilations},
input_left_pads_{input_left_pads}, input_left_pads_{input_left_pads},
input_right_pads_{input_right_pads} input_right_pads_{input_right_pads}
{ {
...@@ -382,9 +321,7 @@ struct DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle ...@@ -382,9 +321,7 @@ struct DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle
tildes = {i_ztilde, i_ytilde, i_xtilde}; tildes = {i_ztilde, i_ytilde, i_xtilde};
} }
const auto a_grid_desc_ak0_m_ak1 = ConvToGemmBwdDataTransform conv_to_gemm_transform_{a_g_n_k_wos_lengths,
transform_conv_to_gemm.template MakeADescriptor_AK0_M_AK1<ALayout>(
a_g_n_k_wos_lengths,
a_g_n_k_wos_strides, a_g_n_k_wos_strides,
b_g_k_c_xs_lengths, b_g_k_c_xs_lengths,
b_g_k_c_xs_strides, b_g_k_c_xs_strides,
...@@ -394,56 +331,37 @@ struct DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle ...@@ -394,56 +331,37 @@ struct DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle
conv_filter_dilations, conv_filter_dilations,
input_left_pads, input_left_pads,
input_right_pads, input_right_pads,
tildes); tildes};
const auto a_grid_desc_ak0_m_ak1 =
conv_to_gemm_transform_.MakeADescriptor_AK0_M_AK1();
const auto b_grid_desc_bk0_n_bk1 = const auto b_grid_desc_bk0_n_bk1 =
transform_conv_to_gemm.template MakeBDescriptor_BK0_N_BK1<BLayout>( conv_to_gemm_transform_.MakeBDescriptor_BK0_N_BK1();
a_g_n_k_wos_lengths,
a_g_n_k_wos_strides,
b_g_k_c_xs_lengths,
b_g_k_c_xs_strides,
e_g_n_c_wis_lengths,
e_g_n_c_wis_strides,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
tildes);
DsGridDesc_M_N ds_grid_desc_m_n; DsGridDesc_M_N ds_grid_desc_m_n;
// populate Ds desc // populate Ds desc
static_for<0, NumDTensor, 1>{}([&](auto i) { static_for<0, NumDTensor, 1>{}([&](auto i) {
using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>; using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
static_assert(is_same_v<DLayout, ELayout>);
ds_grid_desc_m_n(i) = ConvToGemmBwdDataTransform conv_to_gemm_transform_d{
transform_conv_to_gemm.template MakeCDescriptor_M_N<DLayout>(
a_g_n_k_wos_lengths, a_g_n_k_wos_lengths,
a_g_n_k_wos_strides, a_g_n_k_wos_strides,
b_g_k_c_xs_lengths, b_g_k_c_xs_lengths,
b_g_k_c_xs_strides, b_g_k_c_xs_strides,
ds_g_n_c_wis_lengths[i], e_g_n_c_wis_lengths,
ds_g_n_c_wis_strides[i], ds_g_n_c_wis_strides[i],
conv_filter_strides, conv_filter_strides,
conv_filter_dilations, conv_filter_dilations,
input_left_pads, input_left_pads,
input_right_pads, input_right_pads,
tildes); tildes};
ds_grid_desc_m_n(i) = conv_to_gemm_transform_d.MakeCDescriptor_M_N();
}); });
const auto e_grid_desc_m_n = const auto e_grid_desc_m_n = conv_to_gemm_transform_.MakeCDescriptor_M_N();
transform_conv_to_gemm.template MakeCDescriptor_M_N<ELayout>(
a_g_n_k_wos_lengths,
a_g_n_k_wos_strides,
b_g_k_c_xs_lengths,
b_g_k_c_xs_strides,
e_g_n_c_wis_lengths,
e_g_n_c_wis_strides,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
tildes);
// for check validity // for check validity
ds_grid_desc_m_n_container_.push_back(ds_grid_desc_m_n); ds_grid_desc_m_n_container_.push_back(ds_grid_desc_m_n);
...@@ -522,17 +440,9 @@ struct DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle ...@@ -522,17 +440,9 @@ struct DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle
BElementwiseOp b_element_op_; BElementwiseOp b_element_op_;
CDEElementwiseOp cde_element_op_; CDEElementwiseOp cde_element_op_;
// for checking IsSupportedArgument()
std::array<index_t, NDimSpatial + 3> a_g_n_k_wos_lengths_; std::array<index_t, NDimSpatial + 3> a_g_n_k_wos_lengths_;
std::array<index_t, NDimSpatial + 3> a_g_n_k_wos_strides_;
std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_lengths_; std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_lengths_;
std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_strides_;
std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor> ds_g_n_c_wis_lengths_;
std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor> ds_g_n_c_wis_strides_;
std::array<index_t, NDimSpatial + 3> e_g_n_c_wis_lengths_;
std::array<index_t, NDimSpatial + 3> e_g_n_c_wis_strides_;
std::array<index_t, NDimSpatial> conv_filter_strides_; std::array<index_t, NDimSpatial> conv_filter_strides_;
std::array<index_t, NDimSpatial> conv_filter_dilations_;
std::array<index_t, NDimSpatial> input_left_pads_; std::array<index_t, NDimSpatial> input_left_pads_;
std::array<index_t, NDimSpatial> input_right_pads_; std::array<index_t, NDimSpatial> input_right_pads_;
}; };
......
test/grouped_convnd_bwd_data/CMakeLists.txt
View file @ 261f1759
add_gtest_executable(test_grouped_convnd_bwd_data test_grouped_convnd_bwd_data_xdl_wmma.cpp) add_gtest_executable(test_grouped_convnd_bwd_data_xdl test_grouped_convnd_bwd_data_xdl.cpp)
if(result EQUAL 0) if(result EQUAL 0)
target_link_libraries(test_grouped_convnd_bwd_data PRIVATE utility device_grouped_conv2d_bwd_data_instance device_grouped_conv3d_bwd_data_instance) target_link_libraries(test_grouped_convnd_bwd_data_xdl PRIVATE utility device_grouped_conv2d_bwd_data_instance device_grouped_conv3d_bwd_data_instance)
endif()
add_gtest_executable(test_grouped_convnd_bwd_data_wmma test_grouped_convnd_bwd_data_wmma.cpp)
if(result EQUAL 0)
target_link_libraries(test_grouped_convnd_bwd_data_wmma PRIVATE utility device_grouped_conv2d_bwd_data_instance device_grouped_conv3d_bwd_data_instance)
endif() endif()
add_gtest_executable(test_grouped_convnd_bwd_data_interface_xdl test_grouped_convnd_bwd_data_interface_xdl.cpp) add_gtest_executable(test_grouped_convnd_bwd_data_interface_xdl test_grouped_convnd_bwd_data_interface_xdl.cpp)
if(result EQUAL 0) if(result EQUAL 0)
......
test/grouped_convnd_bwd_data/test_grouped_convnd_bwd_data_wmma.cpp 0 → 100644
View file @ 261f1759
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <iostream>
#include <initializer_list>
#include <tuple>
#include <vector>
#include <gtest/gtest.h>
#include "profiler/profile_grouped_conv_bwd_data_impl.hpp"
template <typename Tuple>
class TestGroupedConvndBwdDataWmma : public ::testing::Test
{
protected:
using DataType = std::tuple_element_t<0, Tuple>;
using OutLayout = std::tuple_element_t<1, Tuple>;
using WeiLayout = std::tuple_element_t<2, Tuple>;
using InLayout = std::tuple_element_t<3, Tuple>;
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_bwd_data_impl<NDimSpatial,
OutLayout,
WeiLayout,
InLayout,
DataType,
DataType,
DataType>(
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::half_t, GNHWK, GKYXC, GNHWC>,
std::tuple<int8_t, GNHWK, GKYXC, GNHWC>,
std::tuple<ck::half_t, NHWGK, GKYXC, NHWGC>,
std::tuple<int8_t, NHWGK, GKYXC, NHWGC>>;
using KernelTypes3d = ::testing::Types<std::tuple<ck::half_t, GNDHWK, GKZYXC, GNDHWC>,
std::tuple<int8_t, GNDHWK, GKZYXC, GNDHWC>,
std::tuple<ck::half_t, NDHWGK, GKZYXC, NDHWGC>,
std::tuple<int8_t, NDHWGK, GKZYXC, NDHWGC>>;
template <typename Tuple>
class TestGroupedConvndBwdDataWmma2d : public TestGroupedConvndBwdDataWmma<Tuple>
{
};
template <typename Tuple>
class TestGroupedConvndBwdDataWmma3d : public TestGroupedConvndBwdDataWmma<Tuple>
{
};
TYPED_TEST_SUITE(TestGroupedConvndBwdDataWmma2d, KernelTypes2d);
TYPED_TEST_SUITE(TestGroupedConvndBwdDataWmma3d, KernelTypes3d);
TYPED_TEST(TestGroupedConvndBwdDataWmma2d, Test2D)
{
this->conv_params.clear();
this->conv_params.push_back(
{2, 2, 4, 192, 192, {3, 3}, {28, 28}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
this->conv_params.push_back(
{2, 2, 128, 128, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
this->conv_params.push_back(
{2, 2, 128, 128, 256, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}});
this->conv_params.push_back(
{2, 2, 128, 128, 256, {1, 1}, {3, 3}, {1, 1}, {1, 1}, {0, 0}, {0, 0}});
this->conv_params.push_back({2, 1, 1, 1, 32, {8, 8}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
this->conv_params.push_back({2, 1, 1, 64, 3, {8, 8}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
this->conv_params.push_back({2, 1, 1, 1, 1, {8, 8}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
this->template Run<2>();
}
TYPED_TEST(TestGroupedConvndBwdDataWmma3d, Test3D)
{
this->conv_params.clear();
this->conv_params.push_back(
{3, 2, 16, 128, 256, {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, 256, {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, 256, {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->template Run<3>();
}
test/grouped_convnd_bwd_data/test_grouped_convnd_bwd_data_xdl_wmma.cpp → test/grouped_convnd_bwd_data/test_grouped_convnd_bwd_data_xdl.cpp
View file @ 261f1759
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib> #include <cstdlib>
#include <iostream> #include <iostream>
...@@ -12,7 +12,7 @@ ...@@ -12,7 +12,7 @@
#include "profiler/profile_grouped_conv_bwd_data_impl.hpp" #include "profiler/profile_grouped_conv_bwd_data_impl.hpp"
template <typename Tuple> template <typename Tuple>
class TestGroupedConvndBwdData : public ::testing::Test class TestGroupedConvndBwdDataXdl : public ::testing::Test
{ {
protected: protected:
using DataType = std::tuple_element_t<0, Tuple>; using DataType = std::tuple_element_t<0, Tuple>;
...@@ -51,35 +51,31 @@ using namespace ck::tensor_layout::convolution; ...@@ -51,35 +51,31 @@ using namespace ck::tensor_layout::convolution;
using KernelTypes2d = ::testing::Types<std::tuple<float, GNHWK, GKYXC, GNHWC>, using KernelTypes2d = ::testing::Types<std::tuple<float, GNHWK, GKYXC, GNHWC>,
std::tuple<ck::half_t, GNHWK, GKYXC, GNHWC>, std::tuple<ck::half_t, GNHWK, GKYXC, GNHWC>,
std::tuple<ck::bhalf_t, GNHWK, GKYXC, GNHWC>, std::tuple<ck::bhalf_t, GNHWK, GKYXC, GNHWC>,
std::tuple<int8_t, GNHWK, GKYXC, GNHWC>,
std::tuple<float, NHWGK, GKYXC, NHWGC>, std::tuple<float, NHWGK, GKYXC, NHWGC>,
std::tuple<ck::half_t, NHWGK, GKYXC, NHWGC>, std::tuple<ck::half_t, NHWGK, GKYXC, NHWGC>,
std::tuple<ck::bhalf_t, NHWGK, GKYXC, NHWGC>, std::tuple<ck::bhalf_t, NHWGK, GKYXC, NHWGC>>;
std::tuple<int8_t, NHWGK, GKYXC, NHWGC>>;
using KernelTypes3d = ::testing::Types<std::tuple<float, GNDHWK, GKZYXC, GNDHWC>, using KernelTypes3d = ::testing::Types<std::tuple<float, GNDHWK, GKZYXC, GNDHWC>,
std::tuple<ck::half_t, GNDHWK, GKZYXC, GNDHWC>, std::tuple<ck::half_t, GNDHWK, GKZYXC, GNDHWC>,
std::tuple<ck::bhalf_t, GNDHWK, GKZYXC, GNDHWC>, std::tuple<ck::bhalf_t, GNDHWK, GKZYXC, GNDHWC>,
std::tuple<int8_t, GNDHWK, GKZYXC, GNDHWC>,
std::tuple<float, NDHWGK, GKZYXC, NDHWGC>, std::tuple<float, NDHWGK, GKZYXC, NDHWGC>,
std::tuple<ck::half_t, NDHWGK, GKZYXC, NDHWGC>, std::tuple<ck::half_t, NDHWGK, GKZYXC, NDHWGC>,
std::tuple<ck::bhalf_t, NDHWGK, GKZYXC, NDHWGC>, std::tuple<ck::bhalf_t, NDHWGK, GKZYXC, NDHWGC>>;
std::tuple<int8_t, NDHWGK, GKZYXC, NDHWGC>>;
template <typename Tuple> template <typename Tuple>
class TestGroupedConvndBwdData2d : public TestGroupedConvndBwdData<Tuple> class TestGroupedConvndBwdDataXdl2d : public TestGroupedConvndBwdDataXdl<Tuple>
{ {
}; };
template <typename Tuple> template <typename Tuple>
class TestGroupedConvndBwdData3d : public TestGroupedConvndBwdData<Tuple> class TestGroupedConvndBwdDataXdl3d : public TestGroupedConvndBwdDataXdl<Tuple>
{ {
}; };
TYPED_TEST_SUITE(TestGroupedConvndBwdData2d, KernelTypes2d); TYPED_TEST_SUITE(TestGroupedConvndBwdDataXdl2d, KernelTypes2d);
TYPED_TEST_SUITE(TestGroupedConvndBwdData3d, KernelTypes3d); TYPED_TEST_SUITE(TestGroupedConvndBwdDataXdl3d, KernelTypes3d);
TYPED_TEST(TestGroupedConvndBwdData2d, Test2D) TYPED_TEST(TestGroupedConvndBwdDataXdl2d, Test2D)
{ {
this->conv_params.clear(); this->conv_params.clear();
...@@ -94,10 +90,13 @@ TYPED_TEST(TestGroupedConvndBwdData2d, Test2D) ...@@ -94,10 +90,13 @@ TYPED_TEST(TestGroupedConvndBwdData2d, Test2D)
this->conv_params.push_back({2, 1, 1, 1, 32, {8, 8}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}}); this->conv_params.push_back({2, 1, 1, 1, 32, {8, 8}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
this->conv_params.push_back({2, 1, 1, 64, 3, {8, 8}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}}); this->conv_params.push_back({2, 1, 1, 64, 3, {8, 8}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
this->conv_params.push_back({2, 1, 1, 1, 1, {8, 8}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}}); this->conv_params.push_back({2, 1, 1, 1, 1, {8, 8}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
// SplitN case
this->conv_params.push_back(
{2, 1, 128, 4, 192, {2, 2}, {224, 224}, {224, 224}, {1, 1}, {0, 0}, {0, 0}});
this->template Run<2>(); this->template Run<2>();
} }
TYPED_TEST(TestGroupedConvndBwdData3d, Test3D) TYPED_TEST(TestGroupedConvndBwdDataXdl3d, Test3D)
{ {
this->conv_params.clear(); this->conv_params.clear();
this->conv_params.push_back( this->conv_params.push_back(
...@@ -112,5 +111,17 @@ TYPED_TEST(TestGroupedConvndBwdData3d, Test3D) ...@@ -112,5 +111,17 @@ TYPED_TEST(TestGroupedConvndBwdData3d, Test3D)
{3, 1, 1, 64, 3, {3, 3, 3}, {32, 32, 32}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}}); {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( 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}}); {3, 1, 1, 1, 1, {3, 3, 3}, {32, 32, 32}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
// SplitN case
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}});
this->template Run<3>(); this->template Run<3>();
} }
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