Merge branch 'develop' into lwpck-739

profiler/src/profile_grouped_conv_bwd_data.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_grouped_conv_bwd_data_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+namespace {
+
+enum struct ConvLayout
+{
+    GNHWC_GKYXC_GNHWK, // 0
+    NHWGC_GKYXC_NHWGK, // 1
+};
+
+enum struct ConvDataType
+{
+    F32_F32_F32,    // 0
+    F16_F16_F16,    // 1
+    BF16_BF16_BF16, // 2
+};
+
+#define OP_NAME "grouped_conv_bwd_data"
+#define OP_DESC "Grouped Convolution Backward Data"
+
+static void print_helper_msg()
+{
+    std::cout
+        // clang-format off
+        << "arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"
+        << "arg2: data type (0: Output fp32, Weight fp32, Input fp32\n"
+        << "                 1: Output fp16, Weight fp16, Input fp16\n"
+        << "                 2: Output bf16, Weight bf16, Input bf16\n"
+        << "arg3: tensor layout (0: Output[G, N, Hi, Wi, C], Weight[G, K, Y, X, C], Input[G, N, Ho, Wo, K]\n"
+        << "                     1: Output[N, Hi, Wi, G, C], Weight[G, K, Y, X, C], Input[N, Ho, Wo, G, K])\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"
+        << ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
+    // clang-format on
+}
+
+} // namespace
+
+int profile_grouped_conv_bwd_data(int argc, char* argv[])
+{
+    // 8 for control, 1 for num_dim_spatial
+    if(argc < 9)
+    {
+        print_helper_msg();
+        return 1;
+    }
+
+    const auto data_type       = static_cast<ConvDataType>(std::stoi(argv[2]));
+    const auto layout          = static_cast<ConvLayout>(std::stoi(argv[3]));
+    const bool do_verification = std::stoi(argv[4]);
+    const int init_method      = std::stoi(argv[5]);
+    const bool do_log          = std::stoi(argv[6]);
+    const bool time_kernel     = std::stoi(argv[7]);
+    const int num_dim_spatial  = std::stoi(argv[8]);
+
+    // 8 for control, 1 for num_dim_spatial, 4 for G/N/K/C, and 6 * num_dim_spatial
+    if(argc != 8 + 1 + 4 + 6 * num_dim_spatial)
+    {
+        print_helper_msg();
+        return 1;
+    }
+
+    const auto params = ck::utils::conv::parse_conv_param(num_dim_spatial, 9, argv);
+
+    using F32  = float;
+    using F16  = ck::half_t;
+    using BF16 = ck::bhalf_t;
+
+    using GNHWC = ck::tensor_layout::convolution::GNHWC;
+    using NHWGC = ck::tensor_layout::convolution::NHWGC;
+
+    using GKYXC = ck::tensor_layout::convolution::GKYXC;
+
+    using GNHWK = ck::tensor_layout::convolution::GNHWK;
+    using NHWGK = ck::tensor_layout::convolution::NHWGK;
+
+    constexpr auto I2 = ck::Number<2>{};
+
+    auto profile = [&](auto num_dim_spatial_tmp,
+                       auto out_layout,
+                       auto wei_layout,
+                       auto in_layout,
+                       auto wei_type,
+                       auto out_type,
+                       auto in_type) {
+        constexpr ck::index_t NDimSpatial = num_dim_spatial_tmp.value;
+
+        using OutLayout = decltype(out_layout);
+        using WeiLayout = decltype(wei_layout);
+        using InLayout  = decltype(in_layout);
+
+        using OutDataType = decltype(out_type);
+        using WeiDataType = decltype(wei_type);
+        using InDataType  = decltype(in_type);
+
+        bool pass = ck::profiler::profile_grouped_conv_bwd_data_impl<NDimSpatial,
+                                                                     OutLayout,
+                                                                     WeiLayout,
+                                                                     InLayout,
+                                                                     OutDataType,
+                                                                     WeiDataType,
+                                                                     InDataType>(
+            do_verification, init_method, do_log, time_kernel, params);
+
+        return pass ? 0 : 1;
+    };
+
+    // GNHWC_GKYXC_GNHWK
+    if(num_dim_spatial == 2 && layout == ConvLayout::GNHWC_GKYXC_GNHWK)
+    {
+        if(data_type == ConvDataType::F32_F32_F32)
+        {
+            return profile(I2, GNHWK{}, GKYXC{}, GNHWC{}, F32{}, F32{}, F32{});
+        }
+        else if(data_type == ConvDataType::F16_F16_F16)
+        {
+            return profile(I2, GNHWK{}, GKYXC{}, GNHWC{}, F16{}, F16{}, F16{});
+        }
+        else if(data_type == ConvDataType::BF16_BF16_BF16)
+        {
+            return profile(I2, GNHWK{}, GKYXC{}, GNHWC{}, BF16{}, BF16{}, BF16{});
+        }
+    }
+    // NHWGC_GKYXC_NHWGK
+    else if(num_dim_spatial == 2 && layout == ConvLayout::NHWGC_GKYXC_NHWGK)
+    {
+        if(data_type == ConvDataType::F32_F32_F32)
+        {
+            return profile(I2, NHWGK{}, GKYXC{}, NHWGC{}, F32{}, F32{}, F32{});
+        }
+        else if(data_type == ConvDataType::F16_F16_F16)
+        {
+            return profile(I2, NHWGK{}, GKYXC{}, NHWGC{}, F16{}, F16{}, F16{});
+        }
+        else if(data_type == ConvDataType::BF16_BF16_BF16)
+        {
+            return profile(I2, NHWGK{}, GKYXC{}, NHWGC{}, BF16{}, BF16{}, BF16{});
+        }
+    }
+
+    std::cout << "this data_type & layout is not implemented" << std::endl;
+
+    return 1;
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_grouped_conv_bwd_data);

test/CMakeLists.txt

@@ -59,6 +59,7 @@ add_subdirectory(batchnorm)
 add_subdirectory(contraction)
 add_subdirectory(pool_fwd)
 add_subdirectory(batched_gemm_multi_d)
+add_subdirectory(grouped_convnd_bwd_data)
 if(GPU_TARGETS MATCHES "gfx1100")
     add_subdirectory(wmma_op)
 endif()

test/grouped_convnd_bwd_data/CMakeLists.txt

+if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940")
+    add_gtest_executable(test_grouped_convnd_bwd_data test_grouped_convnd_bwd_data.cpp)
+    target_link_libraries(test_grouped_convnd_bwd_data PRIVATE utility device_grouped_conv2d_bwd_data_instance)
+    add_gtest_executable(test_grouped_convnd_bwd_data_interface test_grouped_convnd_bwd_data_interface.cpp)
+    target_link_libraries(test_grouped_convnd_bwd_data_interface PRIVATE utility device_grouped_conv2d_bwd_data_instance)
+endif()
\ No newline at end of file

test/grouped_convnd_bwd_data/test_grouped_convnd_bwd_data.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, 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 TestGroupedConvndBwdData : 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 GNHWC = ck::tensor_layout::convolution::GNHWC;
+using NHWGC = ck::tensor_layout::convolution::NHWGC;
+
+using GKYXC = ck::tensor_layout::convolution::GKYXC;
+
+using GNHWK = ck::tensor_layout::convolution::GNHWK;
+using NHWGK = ck::tensor_layout::convolution::NHWGK;
+
+using KernelTypes = ::testing::Types<std::tuple<float, GNHWK, GKYXC, GNHWC>,
+                                     std::tuple<ck::half_t, GNHWK, GKYXC, GNHWC>,
+                                     std::tuple<ck::bhalf_t, GNHWK, GKYXC, GNHWC>,
+                                     std::tuple<float, NHWGK, GKYXC, NHWGC>,
+                                     std::tuple<ck::half_t, NHWGK, GKYXC, NHWGC>,
+                                     std::tuple<ck::bhalf_t, NHWGK, GKYXC, NHWGC>>;
+TYPED_TEST_SUITE(TestGroupedConvndBwdData, KernelTypes);
+
+TYPED_TEST(TestGroupedConvndBwdData, 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->template Run<2>();
+}

test/grouped_convnd_bwd_data/test_grouped_convnd_bwd_data_interface.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iostream>
+#include <initializer_list>
+#include <tuple>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_backward_data_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp"
+
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+
+#include <gtest/gtest.h>
+
+using DataType    = ck::half_t;
+using AccDataType = float;
+using Pass        = ck::tensor_operation::element_wise::PassThrough;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+using ConvBackwardDataSpecialization =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization;
+
+static constexpr auto ConvBwdDataDefault   = ConvBackwardDataSpecialization::Default;
+static constexpr auto Filter1x1Stride1Pad0 = ConvBackwardDataSpecialization::Filter1x1Stride1Pad0;
+
+template <typename Tuple, ConvBackwardDataSpecialization ConvSpec>
+class TestGroupedConvndBwdData : public ::testing::Test
+{
+    protected:
+    static constexpr ck::index_t NDimSpatial = 2;
+
+    using OutLayout = std::tuple_element_t<0, Tuple>;
+    using WeiLayout = std::tuple_element_t<1, Tuple>;
+    using InLayout  = std::tuple_element_t<2, Tuple>;
+
+    // clang-format off
+    using GroupedConvBwdDataDeviceInstance = ck::tensor_operation::device::DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1
+    // ######| NDimSpatial|   ALayout|   BLayout|   DsLayout|  ELayout|       AData|       BData|     AccData|         CShuffle|       DsData|      EData| AElementwise| BElementwise| CDEElementwise| ConvolutionBackward| DoPad| DoPad|      NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer|    MXdl|    NXdl|    ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|    BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffleMXdl| CShuffleNXdl|   CDEBlockTransfer| CDEBlockTransfer|
+    // ######|            |          |          |           |         |        Type|        Type|        Type|         DataType|         Type|       Type|    Operation|    Operation|      Operation|  DataSpecialization| GemmM| GemmN| PrefetchStage|  Size| Block| Block| Block|    |    |  XDL|  XDL| PerWave| PerWave|     ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraM|     ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraN|      PerWave|      PerWave|  _MBlock_MPerBlock|  ScalarPerVector|
+    // ######|            |          |          |           |         |            |            |            |                 |             |           |             |             |               |                    |      |      |              |      |      |      |      |    |    |     |     |        |        | Lengths_AK0_M_AK1|   ArrangeOrder|               |               |      PerVector|  PerVector_AK1|          | Lengths_BK0_N_BK1|   ArrangeOrder|               |               |      PerVector|  PerVector_BK1|          |   PerShuffle|   PerShuffle|  _NBlock_NPerBlock|       _NPerBlock|
+    // ######|            |          |          |           |         |            |            |            |                 |             |           |             |             |               |                    |      |      |              |      |      |      |      |    |    |     |     |        |        |                  |               |               |               |               |               |          |                  |               |               |               |               |               |          |             |             |                   |                 |
+    < NDimSpatial, OutLayout, WeiLayout, ck::Tuple<>, InLayout, DataType, DataType, AccDataType, DataType,   ck::Tuple<>, DataType, Pass, Pass,    Pass,  ConvSpec,  true,  true,             1,   256,   128,   256,    32,   8,   2,   32,   32,       2,       4,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,            1,            1,     S<1, 32, 1, 8>,                8>;
+    // clang-format on
+
+    ck::utils::conv::ConvParam conv_param;
+
+    template <ck::index_t NDimSpatial>
+    bool Run()
+    {
+
+        const auto out_g_n_k_wos_desc =
+            ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
+                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 in_g_n_c_wis_desc =
+            ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
+                conv_param);
+
+        std::array<ck::index_t, NDimSpatial + 3> out_lengths{};
+        std::array<ck::index_t, NDimSpatial + 3> out_strides{};
+        std::array<ck::index_t, NDimSpatial + 3> wei_lengths{};
+        std::array<ck::index_t, NDimSpatial + 3> wei_strides{};
+        std::array<ck::index_t, NDimSpatial + 3> in_lengths{};
+        std::array<ck::index_t, NDimSpatial + 3> in_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(out_g_n_k_wos_desc.GetLengths(), out_lengths);
+        copy(out_g_n_k_wos_desc.GetStrides(), out_strides);
+        copy(wei_g_k_c_xs_desc.GetLengths(), wei_lengths);
+        copy(wei_g_k_c_xs_desc.GetStrides(), wei_strides);
+        copy(in_g_n_c_wis_desc.GetLengths(), in_lengths);
+        copy(in_g_n_c_wis_desc.GetStrides(), in_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);
+
+        auto conv = GroupedConvBwdDataDeviceInstance{};
+
+        auto argument = conv.MakeArgument(nullptr,
+                                          nullptr,
+                                          std::array<const void*, 0>{},
+                                          nullptr,
+                                          out_lengths,
+                                          out_strides,
+                                          wei_lengths,
+                                          wei_strides,
+                                          {},
+                                          {},
+                                          in_lengths,
+                                          in_strides,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads,
+                                          Pass{},
+                                          Pass{},
+                                          Pass{});
+        return conv.IsSupportedArgument(argument);
+    }
+};
+
+using GNHWC = ck::tensor_layout::convolution::GNHWC;
+using NHWGC = ck::tensor_layout::convolution::NHWGC;
+
+using GKYXC = ck::tensor_layout::convolution::GKYXC;
+
+using GNHWK = ck::tensor_layout::convolution::GNHWK;
+using NHWGK = ck::tensor_layout::convolution::NHWGK;
+
+using KernelTypes =
+    ::testing::Types<std::tuple<GNHWK, GKYXC, GNHWC>, std::tuple<NHWGK, GKYXC, NHWGC>>;
+
+template <typename Tuple>
+class TestGroupedConvndBwdDataDefault : public TestGroupedConvndBwdData<Tuple, ConvBwdDataDefault>
+{
+};
+
+template <typename Tuple>
+class TestGroupedConvndBwdDataFilter1x1
+    : public TestGroupedConvndBwdData<Tuple, Filter1x1Stride1Pad0>
+{
+};
+
+TYPED_TEST_SUITE(TestGroupedConvndBwdDataDefault, KernelTypes);
+TYPED_TEST_SUITE(TestGroupedConvndBwdDataFilter1x1, KernelTypes);
+
+TYPED_TEST(TestGroupedConvndBwdDataFilter1x1, SpecializationCheck)
+{
+    // Check filter 3,3 instead of 1,1
+    this->conv_param  = {2, 2, 4, 192, 192, {3, 3}, {28, 28}, {1, 1}, {1, 1}, {0, 0}, {0, 0}};
+    bool is_supported = this->template Run<2>();
+    EXPECT_FALSE(is_supported);
+
+    // Check strides 2,2 instead of 1,1
+    this->conv_param = {2, 2, 4, 192, 192, {1, 1}, {28, 28}, {2, 2}, {1, 1}, {0, 0}, {0, 0}};
+    is_supported     = this->template Run<2>();
+    EXPECT_FALSE(is_supported);
+
+    // Check with pad
+    this->conv_param = {2, 2, 4, 192, 192, {1, 1}, {28, 28}, {1, 1}, {1, 1}, {1, 1}, {1, 1}};
+    is_supported     = this->template Run<2>();
+    EXPECT_FALSE(is_supported);
+
+    // Supported version
+    this->conv_param = {2, 2, 4, 192, 192, {1, 1}, {28, 28}, {1, 1}, {1, 1}, {0, 0}, {0, 0}};
+    is_supported     = this->template Run<2>();
+    EXPECT_TRUE(is_supported);
+}
+
+TYPED_TEST(TestGroupedConvndBwdDataDefault, VectorLoadCheck)
+{
+    // vector load for A
+    this->conv_param  = {2, 2, 128, 129, 256, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}};
+    bool is_supported = this->template Run<2>();
+    EXPECT_FALSE(is_supported);
+    // vector load for B, E, Ds
+    this->conv_param = {2, 2, 128, 128, 257, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}};
+    is_supported     = this->template Run<2>();
+    EXPECT_FALSE(is_supported);
+}