Merge branch 'develop' into bwroblew/dpp8

example/37_batched_gemm_add_add_relu_gemm_add/CMakeLists.txt

-add_example_executable(example_batched_gemm_add_add_relu_gemm_add_xdl_fp16 batched_gemm_add_add_relu_gemm_add_xdl_fp16.cpp)
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_batched_gemm_add_add_relu_gemm_add_xdl_fp16 batched_gemm_add_add_relu_gemm_add_xdl_fp16.cpp)
+endif()

example/38_grouped_conv_bwd_data_multiple_d/CMakeLists.txt

+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
 list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
 set(target 0)
 foreach(gpu IN LISTS GPU_TARGETS)
@@ -10,4 +11,5 @@ foreach(gpu IN LISTS GPU_TARGETS)
    add_dependencies(example_grouped_conv_bwd_data example_grouped_conv_bwd_data_bias_relu_fp16)
    set(target 1)
  endif()
-endforeach()
+endforeach()
\ No newline at end of file
+endif()

example/39_permute/CMakeLists.txt

-add_custom_target(example_permute)
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_custom_target(example_permute)
 
-add_example_executable(example_permute_1xHxW_fp16 permute_1xHxW_fp16.cpp)
-add_example_executable(example_permute_NxHxW_fp16 permute_NxHxW_fp16.cpp)
-add_example_executable(example_permute_HxWx4_fp16 permute_HxWx4_fp16.cpp)
+    add_example_executable(example_permute_1xHxW_fp16 permute_1xHxW_fp16.cpp)
+    add_example_executable(example_permute_NxHxW_fp16 permute_NxHxW_fp16.cpp)
+    add_example_executable(example_permute_HxWx4_fp16 permute_HxWx4_fp16.cpp)
 
-add_dependencies(example_permute example_permute_1xHxW_fp16)
-add_dependencies(example_permute example_permute_NxHxW_fp16)
-add_dependencies(example_permute example_permute_HxWx4_fp16)
+    add_dependencies(example_permute example_permute_1xHxW_fp16)
+    add_dependencies(example_permute example_permute_NxHxW_fp16)
+    add_dependencies(example_permute example_permute_HxWx4_fp16)
+endif()

example/40_conv2d_fwd_quantization/CMakeLists.txt

@@ -10,21 +10,19 @@ foreach(gpu IN LISTS GPU_TARGETS)
    set(target 1)
  endif()
 endforeach()
-# Conv perlayer quantization
-add_example_executable(example_conv2d_fwd_dl_perlayer_quantization_int8 conv2d_fwd_dl_perlayer_quantization_int8.cpp)
 
-# Conv perchannel quantization
-add_example_executable(example_conv2d_fwd_dl_perchannel_quantization_int8 conv2d_fwd_dl_perchannel_quantization_int8.cpp)
-
-# Conv + bias + relu perlayer quantization
-add_example_executable(example_conv2d_fwd_dl_bias_relu_perlayer_quantization_int8 conv2d_fwd_dl_bias_relu_perlayer_quantization_int8.cpp)
-
-# Conv + bias + relu perchannel quantization
-add_example_executable(example_conv2d_fwd_dl_bias_relu_perchannel_quantization_int8 conv2d_fwd_dl_bias_relu_perchannel_quantization_int8.cpp)
-
-# Conv + bias + tanh perlayer quantization
-add_example_executable(example_conv2d_fwd_dl_bias_tanh_perlayer_quantization_int8 conv2d_fwd_dl_bias_tanh_perlayer_quantization_int8.cpp)
-
-# Conv + bias + tanh perchannel quantization
-add_example_executable(example_conv2d_fwd_dl_bias_tanh_perchannel_quantization_int8 conv2d_fwd_dl_bias_tanh_perchannel_quantization_int8.cpp)
+  if(DL_KERNELS)
+  # Conv perlayer quantization
+  add_example_executable(example_conv2d_fwd_dl_perlayer_quantization_int8 conv2d_fwd_dl_perlayer_quantization_int8.cpp)
+  # Conv perchannel quantization
+  add_example_executable(example_conv2d_fwd_dl_perchannel_quantization_int8 conv2d_fwd_dl_perchannel_quantization_int8.cpp)
+  # Conv + bias + relu perlayer quantization
+  add_example_executable(example_conv2d_fwd_dl_bias_relu_perlayer_quantization_int8 conv2d_fwd_dl_bias_relu_perlayer_quantization_int8.cpp)
+  # Conv + bias + relu perchannel quantization
+  add_example_executable(example_conv2d_fwd_dl_bias_relu_perchannel_quantization_int8 conv2d_fwd_dl_bias_relu_perchannel_quantization_int8.cpp)
+  # Conv + bias + tanh perlayer quantization
+  add_example_executable(example_conv2d_fwd_dl_bias_tanh_perlayer_quantization_int8 conv2d_fwd_dl_bias_tanh_perlayer_quantization_int8.cpp)
+  # Conv + bias + tanh perchannel quantization
+  add_example_executable(example_conv2d_fwd_dl_bias_tanh_perchannel_quantization_int8 conv2d_fwd_dl_bias_tanh_perchannel_quantization_int8.cpp)
+  endif()
 endif()
\ No newline at end of file

example/41_grouped_conv_conv_fwd/CMakeLists.txt

@@ -3,9 +3,15 @@ list(APPEND gpu_list2 gfx908 gfx90a)
 set(target 0)
 foreach(gpu IN LISTS GPU_TARGETS)
  if(gpu IN_LIST gpu_list1 AND target EQUAL 0)
-   add_example_executable(example_grouped_conv_conv_fwd_xdl_fp32 grouped_conv_conv_fwd_xdl_fp32.cpp)
-   add_example_executable(example_grouped_conv_conv_fwd_xdl_fp16 grouped_conv_conv_fwd_xdl_fp16.cpp)
-   add_example_executable(example_grouped_conv_conv_fwd_xdl_bf16 grouped_conv_conv_fwd_xdl_bf16.cpp)
+   if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+      add_example_executable(example_grouped_conv_conv_fwd_xdl_fp32 grouped_conv_conv_fwd_xdl_fp32.cpp)
+   endif()
+   if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+      add_example_executable(example_grouped_conv_conv_fwd_xdl_fp16 grouped_conv_conv_fwd_xdl_fp16.cpp)
+   endif()
+   if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+      add_example_executable(example_grouped_conv_conv_fwd_xdl_bf16 grouped_conv_conv_fwd_xdl_bf16.cpp)
+   endif()
    if(USE_BITINT_EXTENSION_INT4)
       add_example_executable(example_grouped_conv_conv_fwd_xdl_int4 grouped_conv_conv_fwd_xdl_int4.cpp)
    endif(USE_BITINT_EXTENSION_INT4)
@@ -14,5 +20,7 @@ foreach(gpu IN LISTS GPU_TARGETS)
 endforeach()
 
 if(NOT GPU_TARGETS MATCHES "gfx94" AND NOT GPU_TARGETS MATCHES "gfx1")
-   add_example_executable(example_grouped_conv_conv_fwd_xdl_int8 grouped_conv_conv_fwd_xdl_int8.cpp)
+   if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+      add_example_executable(example_grouped_conv_conv_fwd_xdl_int8 grouped_conv_conv_fwd_xdl_int8.cpp)
+   endif()
 endif()

example/42_groupnorm/CMakeLists.txt

-add_example_executable(example_groupnorm_sigmoid_mul_fp16 groupnorm_sigmoid_mul_fp16.cpp)
-add_example_executable(example_groupnorm_splitk_fp16 groupnorm_splitk_fp16.cpp)
-add_example_executable(example_groupnorm_swish_fp16 groupnorm_swish_fp16.cpp)
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_groupnorm_sigmoid_mul_fp16 groupnorm_sigmoid_mul_fp16.cpp)
+    add_example_executable(example_groupnorm_splitk_fp16 groupnorm_splitk_fp16.cpp)
+    add_example_executable(example_groupnorm_swish_fp16 groupnorm_swish_fp16.cpp)
+endif()

example/43_splitk_gemm_bias_e_permute/CMakeLists.txt

-add_example_executable(example_splitk_gemm_bias_e_permute_xdl_fp16 splitk_gemm_bias_e_permute_xdl_fp16.cpp)
-add_example_executable(example_splitk_gemm_bias_e_permute_xdl_fp32 splitk_gemm_bias_e_permute_xdl_fp32.cpp)
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_splitk_gemm_bias_e_permute_xdl_fp16 splitk_gemm_bias_e_permute_xdl_fp16.cpp)
+endif()
+if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+    add_example_executable(example_splitk_gemm_bias_e_permute_xdl_fp32 splitk_gemm_bias_e_permute_xdl_fp32.cpp)
+endif()

example/44_elementwise_permute/CMakeLists.txt

-add_example_executable(example_elementwise_permute_4D_fp16 elementwise_permute_4D_fp16.cpp)
-add_example_executable(example_elementwise_permute_4D_fp16_2d elementwise_permute_4D_fp16_2d.cpp)
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_elementwise_permute_4D_fp16 elementwise_permute_4D_fp16.cpp)
+    add_example_executable(example_elementwise_permute_4D_fp16_2d elementwise_permute_4D_fp16_2d.cpp)
+endif()

example/46_gemm_add_multiply/CMakeLists.txt

-add_example_executable(example_gemm_add_multiply_dl_fp16 gemm_add_multiply_dl_fp16.cpp)
-add_example_executable(example_gemm_add_multiply_xdl_fp16 gemm_add_multiply_xdl_fp16.cpp)
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    if(DL_KERNELS)
+        add_example_executable(example_gemm_add_multiply_dl_fp16 gemm_add_multiply_dl_fp16.cpp)
+    endif()
+    add_example_executable(example_gemm_add_multiply_xdl_fp16 gemm_add_multiply_xdl_fp16.cpp)
+endif()

example/48_pool3d_fwd/CMakeLists.txt

-add_example_executable(example_pool3d_fwd_fp16 pool3d_fwd_fp16.cpp)
-
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_pool3d_fwd_fp16 pool3d_fwd_fp16.cpp)
+endif()

example/49_maxpool2d_bwd/CMakeLists.txt

-add_example_executable(example_maxpool2d_bwd_bf16 maxpool2d_bwd_bf16.cpp)
-add_example_executable(example_maxpool2d_bwd_fp16 maxpool2d_bwd_fp16.cpp)
-add_example_executable(example_maxpool2d_bwd_fp32 maxpool2d_bwd_fp32.cpp)
+if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_maxpool2d_bwd_bf16 maxpool2d_bwd_bf16.cpp)
+endif()
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_maxpool2d_bwd_fp16 maxpool2d_bwd_fp16.cpp)
+endif()
+if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+    add_example_executable(example_maxpool2d_bwd_fp32 maxpool2d_bwd_fp32.cpp)
+endif()

example/50_put_element/CMakeLists.txt

-add_example_executable(example_put_element_fp16 put_element_fp16.cpp)
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_put_element_fp16 put_element_fp16.cpp)
+endif()

example/51_avgpool3d_bwd/CMakeLists.txt

+add_example_executable(example_avgpool3d_bwd_bf16 avgpool3d_bwd_bf16.cpp)
+add_example_executable(example_avgpool3d_bwd_fp16 avgpool3d_bwd_fp16.cpp)
+add_example_executable(example_avgpool3d_bwd_fp32 avgpool3d_bwd_fp32.cpp)

example/51_avgpool3d_bwd/avgpool3d_bwd_bf16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_avgpool3d_bwd_ndhwc_ndhwc.hpp"
+
+#include "avgpool3d_bwd_common.hpp"
+
+using DOutDataType    = ck::bhalf_t;
+using DInDataType     = ck::bhalf_t;
+using ComputeDataType = float;
+
+#if 1
+using DOutLayout = ck::tensor_layout::convolution::NDHWC;
+using DInLayout  = ck::tensor_layout::convolution::NDHWC;
+#else
+using DOutLayout = ck::tensor_layout::convolution::NCDHW;
+using DInLayout  = ck::tensor_layout::convolution::NCDHW;
+#endif
+
+using DevicePoolBwdInstance =
+    ck::tensor_operation::device::DeviceAvgPool3dBwd_NDHWC_NDHWC<DOutDataType,
+                                                                 DInDataType,
+                                                                 ComputeDataType,
+                                                                 64, // BlockSize
+                                                                 64, // ReduceMThreadClusterSize
+                                                                 1,  // ReduceKThreadClusterSize
+                                                                 1,  // ReduceMThreadSliceSize
+                                                                 1,  // ReduceKThreadSliceSize
+                                                                 1>; // InSrcOutDstVectorSize
+
+int main()
+{
+    std::vector<ck::index_t> window_lengths    = {5, 5, 5};
+    std::vector<ck::index_t> window_strides    = {2, 2, 2};
+    std::vector<ck::index_t> window_dilations  = {2, 2, 2};
+    std::vector<ck::index_t> dinput_left_pads  = {0, 0, 0};
+    std::vector<ck::index_t> dinput_right_pads = {0, 0, 0};
+
+    ck::index_t N  = 1;
+    ck::index_t C  = 16;
+    ck::index_t Di = 40;
+    ck::index_t Hi = 40;
+    ck::index_t Wi = 40;
+
+    pool3d_bwd_test<DevicePoolBwdInstance, DOutDataType, DInDataType, DOutLayout, DInLayout>(
+        true,
+        false,
+        N,
+        C,
+        Di,
+        Hi,
+        Wi,
+        window_lengths,
+        window_strides,
+        window_dilations,
+        dinput_left_pads,
+        dinput_right_pads);
+}

example/51_avgpool3d_bwd/avgpool3d_bwd_common.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include "ck/ck.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/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_avgpool_bwd.hpp"
+
+template <typename TensorLayout>
+std::vector<ck::index_t> f_tensor_strides_ncdhw(ck::index_t N_,
+                                                ck::index_t C_,
+                                                ck::index_t D,
+                                                ck::index_t H,
+                                                ck::index_t W,
+                                                TensorLayout layout)
+{
+    using namespace ck::literals;
+    (void)N_;
+    if constexpr(ck::is_same<decltype(layout), ck::tensor_layout::convolution::NCDHW>::value)
+        return {C_ * D * H * W, D * H * W, H * W, W, 1_uz};
+    else if constexpr(ck::is_same<decltype(layout), ck::tensor_layout::convolution::NDHWC>::value)
+        return {D * C_ * H * W, 1_uz, C_ * H * W, W * C_, C_};
+};
+
+template <typename TensorLayout>
+HostTensorDescriptor f_host_tensor_descriptor(std::size_t N_,
+                                              std::size_t C_,
+                                              std::size_t D,
+                                              std::size_t H,
+                                              std::size_t W,
+                                              TensorLayout layout)
+{
+    using namespace ck::literals;
+
+    if constexpr(ck::is_same<decltype(layout), ck::tensor_layout::convolution::NCDHW>::value)
+    {
+        return HostTensorDescriptor({N_, C_, D, H, W}, {C_ * D * H * W, D * H * W, H * W, W, 1_uz});
+    }
+    else if constexpr(ck::is_same<decltype(layout), ck::tensor_layout::convolution::NDHWC>::value)
+    {
+        return HostTensorDescriptor({N_, C_, D, H, W},
+                                    {D * C_ * H * W, 1_uz, C_ * H * W, W * C_, C_});
+    }
+};
+
+template <typename DevicePoolBwdInstance,
+          typename DOutDataType,
+          typename DInDataType,
+          typename DOutLayout,
+          typename DInLayout>
+bool pool3d_bwd_test(bool do_verification,
+                     bool time_kernel,
+                     ck::index_t N,
+                     ck::index_t C,
+                     ck::index_t Di,
+                     ck::index_t Hi,
+                     ck::index_t Wi,
+                     std::vector<ck::index_t> window_lengths,
+                     std::vector<ck::index_t> window_strides,
+                     std::vector<ck::index_t> window_dilations,
+                     std::vector<ck::index_t> dinput_left_pads,
+                     std::vector<ck::index_t> dinput_right_pads)
+{
+    auto OutSpatialLength = [&](auto InSpatialLength, int index) {
+        ck::index_t left_pad   = dinput_left_pads[index];
+        ck::index_t right_pad  = dinput_right_pads[index];
+        ck::index_t window_len = window_lengths[index];
+        ck::index_t stride     = window_strides[index];
+        ck::index_t dilation   = window_dilations[index];
+        ck::index_t eff        = (window_len - 1) * dilation + 1;
+        return (InSpatialLength + left_pad + right_pad - eff) / stride + 1;
+    };
+
+    ck::index_t Do = OutSpatialLength(Di, 0);
+    ck::index_t Ho = OutSpatialLength(Hi, 1);
+    ck::index_t Wo = OutSpatialLength(Wi, 2);
+
+    Tensor<DOutDataType> dout(f_host_tensor_descriptor(N, C, Do, Ho, Wo, DOutLayout{}));
+    Tensor<DInDataType> din_dev(f_host_tensor_descriptor(N, C, Di, Hi, Wi, DInLayout{}));
+    Tensor<DInDataType> din_host(f_host_tensor_descriptor(N, C, Di, Hi, Wi, DInLayout{}));
+
+    std::cout << "dout: " << dout.mDesc << std::endl;
+    std::cout << "din_host: " << din_host.mDesc << std::endl;
+
+    dout.GenerateTensorValue(GeneratorTensor_3<DOutDataType>{0.0, 1.0});
+
+    DeviceMem dout_device_buf(sizeof(DOutDataType) * dout.mDesc.GetElementSpaceSize());
+    DeviceMem din_device_buf(sizeof(DInDataType) * din_dev.mDesc.GetElementSpaceSize());
+
+    dout_device_buf.ToDevice(dout.mData.data());
+    din_device_buf.SetZero();
+
+    auto pool        = DevicePoolBwdInstance{};
+    auto invoker_ptr = pool.MakeInvokerPointer();
+    auto argument_ptr =
+        pool.MakeArgumentPointer(static_cast<DOutDataType*>(dout_device_buf.GetDeviceBuffer()),
+                                 static_cast<DInDataType*>(din_device_buf.GetDeviceBuffer()),
+                                 {N, C, Do, Ho, Wo},
+                                 {N, C, Di, Hi, Wi},
+                                 f_tensor_strides_ncdhw(N, C, Do, Ho, Wo, DOutLayout{}),
+                                 f_tensor_strides_ncdhw(N, C, Di, Hi, Wi, DInLayout{}),
+                                 window_lengths,
+                                 window_strides,
+                                 window_dilations,
+                                 dinput_left_pads,
+                                 dinput_right_pads);
+
+    if(!pool.IsSupportedArgument(argument_ptr.get()))
+    {
+        throw std::runtime_error("wrong! device_op with the specified compilation parameters does "
+                                 "not support this problem");
+    }
+
+    float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+    std::cout << "Perf: " << ave_time << std::endl;
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        auto ref_pool =
+            ck::tensor_operation::host::ReferenceAvgPoolBwd<3, DInDataType, DOutDataType>();
+
+        auto ref_invoker = ref_pool.MakeInvoker();
+
+        auto ref_argument = ref_pool.MakeArgument(din_host,
+                                                  dout,
+                                                  window_lengths,
+                                                  window_strides,
+                                                  window_dilations,
+                                                  dinput_left_pads,
+                                                  dinput_right_pads);
+
+        ref_invoker.Run(ref_argument);
+
+        din_device_buf.FromDevice(din_dev.mData.data());
+        pass = ck::utils::check_err(din_dev, din_host);
+    }
+
+    return pass;
+}

example/51_avgpool3d_bwd/avgpool3d_bwd_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_avgpool3d_bwd_ndhwc_ndhwc.hpp"
+
+#include "avgpool3d_bwd_common.hpp"
+
+using DOutDataType    = ck::half_t;
+using DInDataType     = ck::half_t;
+using ComputeDataType = float;
+
+#if 1
+using DOutLayout = ck::tensor_layout::convolution::NDHWC;
+using DInLayout  = ck::tensor_layout::convolution::NDHWC;
+#else
+using DOutLayout = ck::tensor_layout::convolution::NCDHW;
+using DInLayout  = ck::tensor_layout::convolution::NCDHW;
+#endif
+
+using DevicePoolBwdInstance =
+    ck::tensor_operation::device::DeviceAvgPool3dBwd_NDHWC_NDHWC<DOutDataType,
+                                                                 DInDataType,
+                                                                 ComputeDataType,
+                                                                 64, // BlockSize
+                                                                 64, // ReduceMThreadClusterSize
+                                                                 1,  // ReduceKThreadClusterSize
+                                                                 1,  // ReduceMThreadSliceSize
+                                                                 1,  // ReduceKThreadSliceSize
+                                                                 1>; // InSrcOutDstVectorSize
+
+int main()
+{
+    std::vector<ck::index_t> window_lengths    = {5, 5, 5};
+    std::vector<ck::index_t> window_strides    = {2, 2, 2};
+    std::vector<ck::index_t> window_dilations  = {2, 2, 2};
+    std::vector<ck::index_t> dinput_left_pads  = {0, 0, 0};
+    std::vector<ck::index_t> dinput_right_pads = {0, 0, 0};
+
+    ck::index_t N  = 1;
+    ck::index_t C  = 16;
+    ck::index_t Di = 40;
+    ck::index_t Hi = 40;
+    ck::index_t Wi = 40;
+
+    pool3d_bwd_test<DevicePoolBwdInstance, DOutDataType, DInDataType, DOutLayout, DInLayout>(
+        true,
+        false,
+        N,
+        C,
+        Di,
+        Hi,
+        Wi,
+        window_lengths,
+        window_strides,
+        window_dilations,
+        dinput_left_pads,
+        dinput_right_pads);
+}

example/51_avgpool3d_bwd/avgpool3d_bwd_fp32.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_avgpool3d_bwd_ndhwc_ndhwc.hpp"
+
+#include "avgpool3d_bwd_common.hpp"
+
+using DOutDataType    = float;
+using DInDataType     = float;
+using ComputeDataType = float;
+
+#if 1
+using DOutLayout = ck::tensor_layout::convolution::NDHWC;
+using DInLayout  = ck::tensor_layout::convolution::NDHWC;
+#else
+using DOutLayout = ck::tensor_layout::convolution::NCDHW;
+using DInLayout  = ck::tensor_layout::convolution::NCDHW;
+#endif
+
+using DevicePoolBwdInstance =
+    ck::tensor_operation::device::DeviceAvgPool3dBwd_NDHWC_NDHWC<DOutDataType,
+                                                                 DInDataType,
+                                                                 ComputeDataType,
+                                                                 64, // BlockSize
+                                                                 64, // ReduceMThreadClusterSize
+                                                                 1,  // ReduceKThreadClusterSize
+                                                                 1,  // ReduceMThreadSliceSize
+                                                                 1,  // ReduceKThreadSliceSize
+                                                                 1>; // InSrcOutDstVectorSize
+
+int main()
+{
+    std::vector<ck::index_t> window_lengths    = {5, 5, 5};
+    std::vector<ck::index_t> window_strides    = {2, 2, 2};
+    std::vector<ck::index_t> window_dilations  = {2, 2, 2};
+    std::vector<ck::index_t> dinput_left_pads  = {0, 0, 0};
+    std::vector<ck::index_t> dinput_right_pads = {0, 0, 0};
+
+    ck::index_t N  = 1;
+    ck::index_t C  = 16;
+    ck::index_t Di = 40;
+    ck::index_t Hi = 40;
+    ck::index_t Wi = 40;
+
+    pool3d_bwd_test<DevicePoolBwdInstance, DOutDataType, DInDataType, DOutLayout, DInLayout>(
+        true,
+        false,
+        N,
+        C,
+        Di,
+        Hi,
+        Wi,
+        window_lengths,
+        window_strides,
+        window_dilations,
+        dinput_left_pads,
+        dinput_right_pads);
+}

include/ck/ck.hpp

@@ -201,7 +201,7 @@
 #define CK_WORKAROUND_SWDEV_388832 1
 
 // workaround: Grouped Conv2d_bwd_data fails for already implemented instance
-#define CK_WORKAROUND_SWDEV_3318619 0
+#define CK_WORKAROUND_GITHUB_ISSUE_824 1
 
 // flag to enable (1) or disable (0) the debugging output in some kernels
 #define DEBUG_LOG 0

include/ck/tensor_operation/gpu/device/device_avgpool_bwd.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <index_t NDimSpatial,
+          typename DOutDataType,
+          typename DInDataType,
+          typename DOutLayout,
+          typename DInLayout>
+struct DeviceAvgPoolBwd : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_dout,
+                        void* p_din,
+                        std::vector<ck::index_t> dout_n_k_wos_lengths,
+                        std::vector<ck::index_t> dout_n_k_wos_strides,
+                        std::vector<ck::index_t> din_n_k_wos_length,
+                        std::vector<ck::index_t> din_n_k_wos_strides,
+                        std::vector<ck::index_t> window_k_c_xs_lengths,
+                        std::vector<ck::index_t> window_strides,
+                        std::vector<ck::index_t> window_dilations,
+                        std::vector<ck::index_t> input_left_pads,
+                        std::vector<ck::index_t> input_right_pads) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

include/ck/tensor_operation/gpu/device/device_grouped_conv_bwd_weight.hpp

@@ -27,15 +27,12 @@ struct DeviceGroupedConvBwdWeight : public BaseOperator
     MakeArgumentPointer(const void* p_in,
                         void* p_wei,
                         const void* p_out,
-                        const ck::index_t G,
-                        const ck::index_t N,
-                        const ck::index_t K,
-                        const ck::index_t C,
-                        const std::array<ck::index_t, NDimSpatial>& input_spatial_lengths,
-                        const std::array<ck::index_t, NDimSpatial>& filter_spatial_lengths,
-                        const std::array<ck::index_t, NDimSpatial>& output_spatial_lengths,
-                        const std::array<ck::index_t, NDimSpatial + 3>& input_strides,
-                        const std::array<ck::index_t, NDimSpatial + 3>& output_strides,
+                        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths, // input
+                        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths, // weight
+                        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths, // output
+                        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
                         const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
                         const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
                         const std::array<ck::index_t, NDimSpatial>& input_left_pads,