merge develop

library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/avg_pool3d_bwd_ndhwc_instance_common.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_avgpool3d_bwd_ndhwc_ndhwc.hpp"
+#include "ck/utility/data_type.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using I32   = int32_t;
+using F16   = ck::half_t;
+using BF16  = ck::bhalf_t;
+using F32   = float;
+using NDHWC = ck::tensor_layout::convolution::NDHWC;
+
+using device_avgpool_bwd_ndhwc_f16_instances =
+    // clang-format off
+    std::tuple <
+        DeviceAvgPool3dBwd_NDHWC_NDHWC<F16, F16, F32, 256, 256, 1, 1, 1, 1>,
+        DeviceAvgPool3dBwd_NDHWC_NDHWC<F16, F16, F32, 256, 256, 1, 2, 2, 2>,
+        DeviceAvgPool3dBwd_NDHWC_NDHWC<F16, F16, F32, 256, 256, 1, 4, 4, 4>,
+        DeviceAvgPool3dBwd_NDHWC_NDHWC<F16, F16, F32, 256, 256, 1, 8, 8, 8>,
+        DeviceAvgPool3dBwd_NDHWC_NDHWC<F16, F16, F32, 256, 32, 8, 8, 8, 8>
+               // clang-format on
+               >;
+
+using device_avgpool_bwd_ndhwc_bf16_instances =
+    // clang-format off
+    std::tuple <
+        DeviceAvgPool3dBwd_NDHWC_NDHWC<BF16, BF16, F32, 256, 256, 1, 1, 1, 1>,
+        DeviceAvgPool3dBwd_NDHWC_NDHWC<BF16, BF16, F32, 256, 256, 1, 2, 2, 2>,
+        DeviceAvgPool3dBwd_NDHWC_NDHWC<BF16, BF16, F32, 256, 256, 1, 4, 4, 4>,
+        DeviceAvgPool3dBwd_NDHWC_NDHWC<BF16, BF16, F32, 256, 256, 1, 8, 8, 8>,
+        DeviceAvgPool3dBwd_NDHWC_NDHWC<BF16, BF16, F32, 256, 32, 8, 8, 8, 8>
+               // clang-format on
+               >;
+
+using device_avgpool_bwd_ndhwc_f32_instances =
+    // clang-format off
+    std::tuple <
+        DeviceAvgPool3dBwd_NDHWC_NDHWC<F32, F32, F32, 256, 256, 1, 1, 1, 1>,
+        DeviceAvgPool3dBwd_NDHWC_NDHWC<F32, F32, F32, 256, 256, 1, 2, 2, 2>,
+        DeviceAvgPool3dBwd_NDHWC_NDHWC<F32, F32, F32, 256, 256, 1, 4, 4, 4>,
+        DeviceAvgPool3dBwd_NDHWC_NDHWC<F32, F32, F32, 256, 256, 1, 8, 8, 8>,
+        DeviceAvgPool3dBwd_NDHWC_NDHWC<F32, F32, F32, 256, 32, 8, 8, 8, 8>
+               // clang-format on
+               >;
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/device_avg_pool3d_bwd_ndhwc_bf16_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "avg_pool3d_bwd_ndhwc_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_avgpool_bwd_ndhwc_bf16_instances(
+    std::vector<std::unique_ptr<DeviceAvgPoolBwd<3, BF16, BF16, NDHWC, NDHWC>>>& instances)
+{
+    add_device_operation_instances(instances, device_avgpool_bwd_ndhwc_bf16_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/device_avg_pool3d_bwd_ndhwc_f16_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "avg_pool3d_bwd_ndhwc_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_avgpool_bwd_ndhwc_f16_instances(
+    std::vector<std::unique_ptr<DeviceAvgPoolBwd<3, F16, F16, NDHWC, NDHWC>>>& instances)
+{
+    add_device_operation_instances(instances, device_avgpool_bwd_ndhwc_f16_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/device_avg_pool3d_bwd_ndhwc_f32_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "avg_pool3d_bwd_ndhwc_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_avgpool_bwd_ndhwc_f32_instances(
+    std::vector<std::unique_ptr<DeviceAvgPoolBwd<3, F32, F32, NDHWC, NDHWC>>>& instances)
+{
+    add_device_operation_instances(instances, device_avgpool_bwd_ndhwc_f32_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/max_pool_bwd/CMakeLists.txt

+set(DEVICE_MAXPOOL_BWD_INSTANCES)
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    list(APPEND DEVICE_MAXPOOL_BWD_INSTANCES device_max_pool_bwd_f16_instance.cpp)
+endif()
+if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+    list(APPEND DEVICE_MAXPOOL_BWD_INSTANCES device_max_pool_bwd_bf16_instance.cpp)
+endif()
+if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+    list(APPEND DEVICE_MAXPOOL_BWD_INSTANCES device_max_pool_bwd_f32_instance.cpp)
+endif()
+add_instance_library(device_max_pool_bwd_instance ${DEVICE_MAXPOOL_BWD_INSTANCES})

library/src/tensor_operation_instance/gpu/max_pool_bwd/device_max_pool_bwd_bf16_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "max_pool_bwd_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_maxpool_bwd_bf16_instances(
+    std::vector<std::unique_ptr<DeviceMaxPoolBwd<BF16, I32, BF16>>>& instances)
+{
+    add_device_operation_instances(instances, device_maxpool_bwd_instances<BF16, I32, BF16>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/max_pool_bwd/device_max_pool_bwd_f16_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "max_pool_bwd_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_maxpool_bwd_f16_instances(
+    std::vector<std::unique_ptr<DeviceMaxPoolBwd<F16, I32, F16>>>& instances)
+{
+    add_device_operation_instances(instances, device_maxpool_bwd_instances<F16, I32, F16>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/max_pool_bwd/device_max_pool_bwd_f32_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "max_pool_bwd_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_maxpool_bwd_f32_instances(
+    std::vector<std::unique_ptr<DeviceMaxPoolBwd<F32, I32, F32>>>& instances)
+{
+    add_device_operation_instances(instances, device_maxpool_bwd_instances<F32, I32, F32>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/max_pool_bwd/max_pool_bwd_instance_common.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_max_pool_bwd_impl.hpp"
+#include "ck/utility/data_type.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using I32  = int32_t;
+using F16  = ck::half_t;
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+template <typename DOutDataType, typename IndexDataType, typename DInDataType>
+using device_maxpool_bwd_instances =
+    // clang-format off
+    std::tuple <
+        DeviceMaxPoolBwdImpl<DOutDataType, IndexDataType, DInDataType, 1>,
+        DeviceMaxPoolBwdImpl<DOutDataType, IndexDataType, DInDataType, 2>,
+        DeviceMaxPoolBwdImpl<DOutDataType, IndexDataType, DInDataType, 4>
+               // clang-format on
+               >;
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/pool3d_fwd/CMakeLists.txt

@@ -3,6 +3,10 @@ if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
     list(APPEND DEVICE_POOL3D_FWD_INSTANCES device_avg_pool3d_fwd_ndhwc_f16_instance.cpp
                                           device_max_pool3d_fwd_ndhwc_f16_instance.cpp)
 endif()
+if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+    list(APPEND DEVICE_POOL3D_FWD_INSTANCES device_avg_pool3d_fwd_ndhwc_bf16_instance.cpp
+                                          device_max_pool3d_fwd_ndhwc_bf16_instance.cpp)
+endif()
 if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
     list(APPEND DEVICE_POOL3D_FWD_INSTANCES device_avg_pool3d_fwd_ndhwc_f32_instance.cpp
                                           device_max_pool3d_fwd_ndhwc_f32_instance.cpp)

library/src/tensor_operation_instance/gpu/pool3d_fwd/device_avg_pool3d_fwd_ndhwc_bf16_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "pool_fwd_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr auto ReduceOpId = ck::ReduceTensorOp::AVG;
+
+void add_device_pool3d_fwd_ndhwc_bf16_instances(
+    std::vector<
+        std::unique_ptr<DevicePoolFwd<5, 3, BF16, BF16, I32, NDHWC, NDHWC, ReduceOpId, false>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances, device_pool3d_fwd_ndhwc_instances<BF16, BF16, I32, F32, ReduceOpId, false>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/pool3d_fwd/device_max_pool3d_fwd_ndhwc_bf16_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "pool_fwd_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr auto ReduceOpId = ck::ReduceTensorOp::MAX;
+
+void add_device_pool3d_fwd_ndhwc_bf16_instances(
+    std::vector<
+        std::unique_ptr<DevicePoolFwd<5, 3, BF16, BF16, I32, NDHWC, NDHWC, ReduceOpId, false>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances, device_pool3d_fwd_ndhwc_instances<BF16, BF16, I32, BF16, ReduceOpId, false>{});
+}
+
+void add_device_pool3d_fwd_ndhwc_index_bf16_instances(
+    std::vector<
+        std::unique_ptr<DevicePoolFwd<5, 3, BF16, BF16, I32, NDHWC, NDHWC, ReduceOpId, true>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances, device_pool3d_fwd_ndhwc_instances<BF16, BF16, I32, BF16, ReduceOpId, true>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/pool3d_fwd/pool_fwd_instance_common.hpp

@@ -17,6 +17,7 @@ namespace instance {
 
 using I32   = int32_t;
 using F16   = ck::half_t;
+using BF16  = ck::bhalf_t;
 using F32   = float;
 using NDHWC = ck::tensor_layout::convolution::NDHWC;
 

profiler/include/profiler/profile_avg_pool3d_bwd_impl.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/pool3d_fwd.hpp"
+#include "ck/library/tensor_operation_instance/gpu/avg_pool3d_bwd.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"
+
+namespace ck {
+namespace profiler {
+
+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::NDHWC>::value)
+        return {D * C_ * H * W, 1_uz, C_ * H * W, W * C_, C_};
+    else
+        throw std::runtime_error("not supported yet");
+};
+
+template <typename DOutDataType,
+          typename DInDataType,
+          typename ComputeDataType,
+          typename DOutLayout,
+          typename DInLayout>
+bool profile_avg_pool3d_bwd_impl(int do_verification,
+                                 int init_method,
+                                 bool do_log,
+                                 bool time_kernel,
+                                 std::vector<index_t> in_length, // NCDHW
+                                 std::vector<index_t> window_spatial_lengths,
+                                 std::vector<index_t> window_strides,
+                                 std::vector<index_t> window_dilations,
+                                 std::vector<index_t> input_left_pads,
+                                 std::vector<index_t> input_right_pads)
+{
+    constexpr index_t InOutRank  = 5;
+    constexpr index_t WindowRank = 3;
+
+    if(in_length.size() != InOutRank || window_spatial_lengths.size() != WindowRank ||
+       window_strides.size() != WindowRank || window_dilations.size() != WindowRank ||
+       input_left_pads.size() != WindowRank || input_right_pads.size() != WindowRank)
+    {
+        std::cout << "Parameter is incorrect" << std::endl;
+        return false;
+    }
+
+    std::vector<index_t> out_length(InOutRank);
+
+    int N = in_length[0];
+    int C = in_length[1];
+
+    out_length[0] = N;
+    out_length[1] = C;
+
+    // Calculate Do, Ho, Wo
+    for(int i = 2; i < InOutRank; ++i)
+    {
+        auto pad1             = input_left_pads[i - 2];
+        auto pad2             = input_right_pads[i - 2];
+        auto windows_size     = window_spatial_lengths[i - 2];
+        auto windows_stride   = window_strides[i - 2];
+        auto windows_dilation = window_dilations[i - 2];
+        auto eff              = (windows_size - 1) * windows_dilation + 1;
+        out_length[i]         = (in_length[i] + pad1 + pad2 - eff) / windows_stride + 1;
+    }
+
+    int Di = in_length[2];
+    int Hi = in_length[3];
+    int Wi = in_length[4];
+    int Do = out_length[2];
+    int Ho = out_length[3];
+    int Wo = out_length[4];
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t N_, std::size_t C_, std::size_t D, std::size_t H, std::size_t W) {
+            using namespace ck::literals;
+
+            return HostTensorDescriptor({N_, C_, D, H, W},
+                                        {D * C_ * H * W, 1_uz, C_ * H * W, W * C_, C_});
+        };
+
+    Tensor<DOutDataType> dout_n_c_do_ho_wo(f_host_tensor_descriptor(N, C, Do, Ho, Wo));
+    Tensor<DInDataType> din_n_c_di_hi_wi_device(f_host_tensor_descriptor(N, C, Di, Hi, Wi));
+    Tensor<DInDataType> din_n_c_di_hi_wi_host(f_host_tensor_descriptor(N, C, Di, Hi, Wi));
+
+    switch(init_method)
+    {
+    case 0: dout_n_c_do_ho_wo.GenerateTensorValue(GeneratorTensor_1<DOutDataType>{}); break;
+    case 1: dout_n_c_do_ho_wo.GenerateTensorValue(GeneratorTensor_2<DOutDataType>{-5, 5}); break;
+    default: dout_n_c_do_ho_wo.GenerateTensorValue(GeneratorTensor_3<DOutDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem dout_device_buf(sizeof(DOutDataType) * dout_n_c_do_ho_wo.mDesc.GetElementSpaceSize());
+    DeviceMem din_device_buf(sizeof(DInDataType) *
+                             din_n_c_di_hi_wi_device.mDesc.GetElementSpaceSize());
+
+    dout_device_buf.ToDevice(dout_n_c_do_ho_wo.mData.data());
+
+    using DeviceOp = ck::tensor_operation::device::
+        DeviceAvgPoolBwd<3, DOutDataType, DInDataType, DOutLayout, DInLayout>;
+
+    // get device op instances
+    const auto instance_ptrs =
+        ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+            DeviceOp>::GetInstances();
+
+    std::cout << "found " << instance_ptrs.size() << " instances" << std::endl;
+
+    std::string best_instance_name;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    if(do_verification)
+    {
+        using ReferencePoolingBwdInstance =
+            ck::tensor_operation::host::ReferenceAvgPoolBwd<3, DInDataType, DOutDataType>;
+
+        ReferencePoolingBwdInstance ref_pooling_bwd;
+        auto ref_pooling_bwd_argument = ref_pooling_bwd.MakeArgument(din_n_c_di_hi_wi_host,
+                                                                     dout_n_c_do_ho_wo,
+                                                                     window_spatial_lengths,
+                                                                     window_strides,
+                                                                     window_dilations,
+                                                                     input_left_pads,
+                                                                     input_right_pads);
+
+        auto ref_invoker = ref_pooling_bwd.MakeInvoker();
+        ref_invoker.Run(ref_pooling_bwd_argument);
+    }
+
+    int num_kernel = 0;
+
+    for(auto& inst_ptr : instance_ptrs)
+    {
+        auto argument_ptr = inst_ptr->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_spatial_lengths,
+            window_strides,
+            window_dilations,
+            input_left_pads,
+            input_right_pads);
+
+        if(inst_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            ++num_kernel;
+        }
+        else
+        {
+            if(time_kernel)
+            {
+                std::cout << inst_ptr->GetTypeString() << " skipped due to unsupported argument: ";
+                LogRange(std::cout << "doutput lengths = ", out_length, ", ") << std::endl;
+            }
+
+            continue;
+        }
+
+        din_device_buf.SetZero();
+
+        auto invoker_ptr = inst_ptr->MakeInvokerPointer();
+        float avg_time   = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+        std::size_t num_bytes =
+            dout_n_c_do_ho_wo.mDesc.GetElementSize() * sizeof(DOutDataType) +
+            din_n_c_di_hi_wi_device.mDesc.GetElementSize() * sizeof(DInDataType);
+
+        float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+        if(time_kernel)
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << gb_per_sec << " GB/s, "
+                      << inst_ptr->GetTypeString() << std::endl;
+
+        if(avg_time < best_avg_time)
+        {
+            best_instance_name = inst_ptr->GetTypeString();
+            best_avg_time      = avg_time;
+            best_gb_per_sec    = gb_per_sec;
+        }
+
+        if(do_verification)
+        {
+            din_device_buf.FromDevice(din_n_c_di_hi_wi_device.mData.data());
+            bool pass = ck::utils::check_err(din_n_c_di_hi_wi_device.mData,
+                                             din_n_c_di_hi_wi_host.mData,
+                                             "Error: Incorrect results",
+                                             1e-3,
+                                             1e-3);
+
+            if(do_log)
+            {
+                LogRangeAsType<float>(
+                    std::cout << "din_n_c_di_hi_wi_device: ", din_n_c_di_hi_wi_device.mData, ",")
+                    << std::endl;
+
+                LogRangeAsType<float>(
+                    std::cout << "din_n_c_di_hi_wi_host: ", din_n_c_di_hi_wi_host.mData, ",")
+                    << std::endl;
+            }
+
+            if(!pass)
+            {
+                std::cout << inst_ptr->GetTypeString() << " failed verification: ";
+                LogRange(std::cout << "doutput lengths = [", out_length, ", ") << "]." << std::endl;
+                return false;
+            }
+            else
+            {
+                if(time_kernel)
+                    std::cout << "pass" << std::endl;
+            }
+        }
+    }
+
+    if(time_kernel)
+    {
+        LogRange(std::cout << "length = ", out_length, ",") << std::endl;
+        std::cout << "best perf = " << best_avg_time << " ms, " << best_gb_per_sec << " GB/s, "
+                  << best_instance_name << std::endl;
+    }
+
+    if(num_kernel == 0)
+    {
+        std::cout << "Error: No kernel is applicable" << std::endl;
+        return false;
+    }
+
+    return true;
+}
+
+} // namespace profiler
+} // namespace ck

profiler/include/profiler/profile_max_pool3d_bwd_impl.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/pool3d_fwd.hpp"
+#include "ck/library/tensor_operation_instance/gpu/max_pool_bwd.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_pool_fwd.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_maxpool_bwd.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename InDataType,
+          typename OutDataType,
+          typename IndexDataType,
+          typename DOutDataType,
+          typename DInDataType,
+          bool PropagateNan>
+bool profile_max_pool3d_bwd_impl(int do_verification,
+                                 int init_method,
+                                 bool do_log,
+                                 bool time_kernel,
+                                 std::vector<index_t> in_length, // NCDHW
+                                 std::vector<index_t> window_spatial_lengths,
+                                 std::vector<index_t> window_strides,
+                                 std::vector<index_t> window_dilations,
+                                 std::vector<index_t> input_left_pads,
+                                 std::vector<index_t> input_right_pads)
+{
+    // AtomicAdd only support f32 for now. ComputeDataType must be float32
+    using ComputeDataType = float;
+
+    constexpr index_t InOutRank  = 5;
+    constexpr index_t WindowRank = 3;
+
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+    if(in_length.size() != InOutRank || window_spatial_lengths.size() != WindowRank ||
+       window_strides.size() != WindowRank || window_dilations.size() != WindowRank ||
+       input_left_pads.size() != WindowRank || input_right_pads.size() != WindowRank)
+    {
+        std::cout << "Parameter is incorrect" << std::endl;
+        return false;
+    }
+
+    std::vector<index_t> out_length(InOutRank);
+
+    int N = in_length[0];
+    int C = in_length[1];
+
+    out_length[0] = N;
+    out_length[1] = C;
+
+    // Calculate Do, Ho, Wo
+    for(int i = 2; i < InOutRank; ++i)
+    {
+        auto pad1             = input_left_pads[i - 2];
+        auto pad2             = input_right_pads[i - 2];
+        auto windows_size     = window_spatial_lengths[i - 2];
+        auto windows_stride   = window_strides[i - 2];
+        auto windows_dilation = window_dilations[i - 2];
+        auto eff              = (windows_size - 1) * windows_dilation + 1;
+        out_length[i]         = (in_length[i] + pad1 + pad2 - eff) / windows_stride + 1;
+    }
+
+    int Di = in_length[2];
+    int Hi = in_length[3];
+    int Wi = in_length[4];
+    int Do = out_length[2];
+    int Ho = out_length[3];
+    int Wo = out_length[4];
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t N_, std::size_t C_, std::size_t D, std::size_t H, std::size_t W) {
+            using namespace ck::literals;
+
+            return HostTensorDescriptor({N_, C_, D, H, W},
+                                        {D * C_ * H * W, 1_uz, C_ * H * W, W * C_, C_});
+        };
+
+    Tensor<InDataType> in_n_c_di_hi_wi(f_host_tensor_descriptor(N, C, Di, Hi, Wi));
+    Tensor<OutDataType> out_n_c_do_ho_wo(f_host_tensor_descriptor(N, C, Do, Ho, Wo));
+    Tensor<IndexDataType> out_indices_n_c_do_ho_wo(f_host_tensor_descriptor(N, C, Do, Ho, Wo));
+    Tensor<DOutDataType> dout_n_c_do_ho_wo(f_host_tensor_descriptor(N, C, Do, Ho, Wo));
+    Tensor<DInDataType> din_n_c_di_hi_wi_host(f_host_tensor_descriptor(N, C, Di, Hi, Wi));
+
+    Tensor<DInDataType> din_n_c_di_hi_wi_device(f_host_tensor_descriptor(N, C, Di, Hi, Wi));
+
+    switch(init_method)
+    {
+    case 0:
+        in_n_c_di_hi_wi.GenerateTensorValue(GeneratorTensor_1<InDataType>{});
+        dout_n_c_do_ho_wo.GenerateTensorValue(GeneratorTensor_1<DOutDataType>{});
+        break;
+    case 1:
+        in_n_c_di_hi_wi.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
+        dout_n_c_do_ho_wo.GenerateTensorValue(GeneratorTensor_2<DOutDataType>{-5, 5});
+        break;
+    default:
+        in_n_c_di_hi_wi.GenerateTensorValue(GeneratorTensor_3<InDataType>{-0.5, 0.5});
+        dout_n_c_do_ho_wo.GenerateTensorValue(GeneratorTensor_3<DOutDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem indices_device_buf(sizeof(IndexDataType) *
+                                 out_indices_n_c_do_ho_wo.mDesc.GetElementSpaceSize());
+    DeviceMem dout_device_buf(sizeof(DOutDataType) * dout_n_c_do_ho_wo.mDesc.GetElementSpaceSize());
+    DeviceMem din_device_buf(sizeof(DInDataType) *
+                             din_n_c_di_hi_wi_device.mDesc.GetElementSpaceSize());
+
+    // Generate index data from forwarding
+    {
+        using ReferencePoolingFwdInstance =
+            ck::tensor_operation::host::ReferencePoolingFwd<InOutRank,
+                                                            WindowRank,
+                                                            InDataType,
+                                                            OutDataType,
+                                                            ComputeDataType,
+                                                            IndexDataType,
+                                                            ck::ReduceTensorOp::MAX,
+                                                            false,
+                                                            true>;
+
+        ReferencePoolingFwdInstance ref_pooling_fwd;
+        auto ref_pooling_fwd_argument = ref_pooling_fwd.MakeArgument(in_n_c_di_hi_wi,
+                                                                     out_n_c_do_ho_wo,
+                                                                     out_indices_n_c_do_ho_wo,
+                                                                     window_spatial_lengths,
+                                                                     window_strides,
+                                                                     window_dilations,
+                                                                     input_left_pads,
+                                                                     input_right_pads);
+        auto ref_pooling_fwd_invoker  = ref_pooling_fwd.MakeInvoker();
+        ref_pooling_fwd_invoker.Run(ref_pooling_fwd_argument);
+    }
+
+    indices_device_buf.ToDevice(out_indices_n_c_do_ho_wo.mData.data());
+    dout_device_buf.ToDevice(dout_n_c_do_ho_wo.mData.data());
+
+    using DeviceOp =
+        ck::tensor_operation::device::DeviceMaxPoolBwd<DOutDataType, IndexDataType, DInDataType>;
+
+    // get device op instances
+    const auto instance_ptrs =
+        ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+            DeviceOp>::GetInstances();
+
+    std::cout << "found " << instance_ptrs.size() << " instances" << std::endl;
+
+    std::string best_instance_name;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    if(do_verification)
+    {
+        using ReferencePoolingBwdInstance =
+            ck::tensor_operation::host::ReferenceMaxPoolBwd<DOutDataType,
+                                                            IndexDataType,
+                                                            ComputeDataType,
+                                                            DInDataType,
+                                                            PassThrough>;
+
+        ReferencePoolingBwdInstance ref_pooling_bwd;
+        auto ref_pooling_bwd_argument = ref_pooling_bwd.MakeArgument(
+            dout_n_c_do_ho_wo, out_indices_n_c_do_ho_wo, din_n_c_di_hi_wi_host, PassThrough{});
+        auto ref_invoker = ref_pooling_bwd.MakeInvoker();
+        ref_invoker.Run(ref_pooling_bwd_argument);
+    }
+
+    int num_kernel = 0;
+
+    for(auto& inst_ptr : instance_ptrs)
+    {
+        auto argument_ptr = inst_ptr->MakeArgumentPointer(
+            static_cast<DOutDataType*>(dout_device_buf.GetDeviceBuffer()),
+            static_cast<IndexDataType*>(indices_device_buf.GetDeviceBuffer()),
+            static_cast<DInDataType*>(din_device_buf.GetDeviceBuffer()),
+            dout_n_c_do_ho_wo.mDesc.GetElementSpaceSize(),
+            din_n_c_di_hi_wi_device.mDesc.GetElementSpaceSize(),
+            window_spatial_lengths,
+            window_strides,
+            window_dilations);
+
+        if(inst_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            ++num_kernel;
+        }
+        else
+        {
+            if(time_kernel)
+            {
+                std::cout << inst_ptr->GetTypeString() << " skipped due to unsupported argument: ";
+                LogRange(std::cout << "doutput lengths = ", out_length, ", ") << std::endl;
+            }
+
+            continue;
+        }
+
+        size_t workspace_sz = inst_ptr->GetWorkSpaceSize(argument_ptr.get());
+        DeviceMem workspace_device_buf(workspace_sz);
+        inst_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_device_buf.GetDeviceBuffer());
+
+        auto invoker_ptr = inst_ptr->MakeInvokerPointer();
+        float avg_time   = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+        std::size_t num_bytes =
+            dout_n_c_do_ho_wo.mDesc.GetElementSize() * sizeof(DOutDataType) +
+            out_indices_n_c_do_ho_wo.mDesc.GetElementSize() * sizeof(IndexDataType) +
+            din_n_c_di_hi_wi_device.mDesc.GetElementSize() * sizeof(DInDataType);
+
+        float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+        if(time_kernel)
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << gb_per_sec << " GB/s, "
+                      << inst_ptr->GetTypeString() << std::endl;
+
+        if(avg_time < best_avg_time)
+        {
+            best_instance_name = inst_ptr->GetTypeString();
+            best_avg_time      = avg_time;
+            best_gb_per_sec    = gb_per_sec;
+        }
+
+        if(do_verification)
+        {
+            din_device_buf.FromDevice(din_n_c_di_hi_wi_device.mData.data());
+
+            bool pass = ck::utils::check_err(din_n_c_di_hi_wi_device.mData,
+                                             din_n_c_di_hi_wi_host.mData,
+                                             "Error: Incorrect results",
+                                             1e-3,
+                                             1e-3);
+
+            if(do_log)
+            {
+                LogRangeAsType<float>(
+                    std::cout << "out_indices_n_c_do_ho_wo: ", out_indices_n_c_do_ho_wo.mData, ",")
+                    << std::endl;
+
+                LogRangeAsType<float>(
+                    std::cout << "din_n_c_di_hi_wi_device: ", din_n_c_di_hi_wi_device.mData, ",")
+                    << std::endl;
+
+                LogRangeAsType<float>(
+                    std::cout << "din_n_c_di_hi_wi_host: ", din_n_c_di_hi_wi_host.mData, ",")
+                    << std::endl;
+            }
+
+            if(!pass)
+            {
+                std::cout << inst_ptr->GetTypeString() << " failed verification: ";
+                LogRange(std::cout << "doutput lengths = [", out_length, ", ") << "]." << std::endl;
+                return false;
+            }
+            else
+            {
+                if(time_kernel)
+                    std::cout << "pass" << std::endl;
+            }
+        }
+    }
+
+    if(time_kernel)
+    {
+        LogRange(std::cout << "length = ", out_length, ",") << std::endl;
+        std::cout << "best perf = " << best_avg_time << " ms, " << best_gb_per_sec << " GB/s, "
+                  << best_instance_name << std::endl;
+    }
+
+    if(num_kernel == 0)
+    {
+        std::cout << "Error: No kernel is applicable" << std::endl;
+        return false;
+    }
+
+    return true;
+}
+
+} // namespace profiler
+} // namespace ck

profiler/src/CMakeLists.txt

@@ -19,6 +19,8 @@ set(PROFILER_SOURCES
     profile_groupnorm.cpp
     profile_layernorm.cpp
     profile_max_pool3d_fwd.cpp
+    profile_avg_pool3d_bwd.cpp
+    profile_max_pool3d_bwd.cpp
     profile_softmax.cpp
     profile_batchnorm_fwd.cpp
     profile_batchnorm_bwd.cpp
@@ -76,6 +78,8 @@ target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batchnorm_instance)
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_bilinear_instance)
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_scale_instance)
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool3d_fwd_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_avg_pool3d_bwd_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_max_pool_bwd_instance)
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_data_instance)
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_data_instance)
 if(DL_KERNELS)

profiler/src/profile_avg_pool3d_bwd.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <vector>
+#include <unordered_map>
+
+#include "profiler/data_type_enum.hpp"
+#include "profiler/profile_avg_pool3d_bwd_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+using ck::index_t;
+
+struct maxPoolbwdArgParser
+{
+    std::unordered_map<std::string, std::vector<int>> long_opts = {{"length", {}},
+                                                                   {"wsize", {}},
+                                                                   {"wstride", {}},
+                                                                   {"wdilation", {}},
+                                                                   {"pad1", {}},
+                                                                   {"pad2", {}}};
+
+    bool parse_opt(int argc, char* argv[], const std::string& key, int i)
+    {
+        if(std::string("--") + key == argv[i])
+        {
+            int pos = i;
+            while(++i < argc && argv[i][0] != '-') {}
+            int end = i;
+            for(int j = pos + 1; j < end; j++)
+            {
+                long_opts[key].push_back(std::stoi(argv[j]));
+            }
+            return true;
+        }
+        return false;
+    }
+
+    void operator()(int argc, char* argv[])
+    {
+        for(auto& kv : long_opts)
+        {
+            for(int i = 1; i < argc; i++)
+            {
+                if(parse_opt(argc, argv, kv.first, i))
+                    break;
+            }
+        }
+    }
+};
+
+void print_help_avg_pool3d_bwd()
+{
+    std::cout << "arg1: data type (0: fp16; 1: fp32; 5: bf16)\n"
+              << "arg2: verification (0: no; 1: yes)\n"
+              << "arg3: initialization (0: no init; 1: integer value; 2: decimal value)\n"
+              << "arg4: print tensor value (0: no; 1: yes)\n"
+              << "arg5: time kernel (0=no, 1=yes)\n"
+              << "--length: input tensor length for NCDHW(e.g, --length 2 32 30 30 30) \n"
+              << "--wsize: window size for ZYX (e.g, --wsize 2 2 2) \n"
+              << "--wstride: window stride for DHW (e.g, --wstride 2 2 2) \n"
+              << "--wdilation: window dilation for DHW (e.g, --wdilation 1 1 1) \n"
+              << "--pad1: left side of padding in DHW (e.g, --pad1 1 1 1) \n"
+              << "--pad2: right side of padding in DHW (e.g, --pad2 1 1 1) \n"
+              << "eg: ckProfiler avg_pool3d_bwd 0 1 2 0 1 --length 2 32 30 30 30 --wsize 2 2 2 "
+                 "--wstride 2 2 2 --wdilation 1 1 1 --pad1 1 1 1 --pad2 1 1 1"
+              << std::endl;
+}
+
+int profile_avg_pool3d_bwd(int argc, char* argv[])
+{
+    ck::DataTypeEnum data_type = ck::DataTypeEnum::Half;
+    bool do_verification       = true;
+    int init_method            = 0;
+    bool do_log                = false;
+    bool time_kernel           = true;
+
+    std::vector<index_t> in_length = {2, 32, 30, 30, 30};
+    std::vector<index_t> wsize     = {2, 2, 2};
+    std::vector<index_t> wstride   = {2, 2, 2};
+    std::vector<index_t> wdilation = {1, 1, 1};
+    std::vector<index_t> pad1      = {1, 1, 1};
+    std::vector<index_t> pad2      = {1, 1, 1};
+
+    if(argc != 2 && argc != 33)
+    {
+        print_help_avg_pool3d_bwd();
+        return 0;
+    }
+    else if(argc == 33)
+    {
+        data_type       = static_cast<ck::DataTypeEnum>(std::stoi(argv[2]));
+        do_verification = std::stoi(argv[3]);
+        init_method     = std::stoi(argv[4]);
+        do_log          = std::stoi(argv[5]);
+        time_kernel     = std::stoi(argv[6]);
+
+        // parse the long options
+        maxPoolbwdArgParser arg_parser;
+        arg_parser(argc, argv);
+        in_length = arg_parser.long_opts["length"];
+        wsize     = arg_parser.long_opts["wsize"];
+        wstride   = arg_parser.long_opts["wstride"];
+        wdilation = arg_parser.long_opts["wdilation"];
+        pad1      = arg_parser.long_opts["pad1"];
+        pad2      = arg_parser.long_opts["pad2"];
+    }
+
+#ifdef CK_ENABLE_FP16
+    using F16 = ck::half_t;
+#endif
+#ifdef CK_ENABLE_BF16
+    using BF16 = ck::bhalf_t;
+#endif
+#ifdef CK_ENABLE_FP32
+    using F32 = float;
+#endif
+    using NDHWC = ck::tensor_layout::convolution::NDHWC;
+
+    if(false)
+        ;
+#ifdef CK_ENABLE_FP16
+    else if(data_type == ck::DataTypeEnum::Half)
+    {
+        ck::profiler::profile_avg_pool3d_bwd_impl<F16, F16, F16, NDHWC, NDHWC>(do_verification,
+                                                                               init_method,
+                                                                               do_log,
+                                                                               time_kernel,
+                                                                               in_length,
+                                                                               wsize,
+                                                                               wstride,
+                                                                               wdilation,
+                                                                               pad1,
+                                                                               pad2);
+    }
+#endif
+#ifdef CK_ENABLE_BF16
+    else if(data_type == ck::DataTypeEnum::BFloat16)
+    {
+        ck::profiler::profile_avg_pool3d_bwd_impl<BF16, BF16, BF16, NDHWC, NDHWC>(do_verification,
+                                                                                  init_method,
+                                                                                  do_log,
+                                                                                  time_kernel,
+                                                                                  in_length,
+                                                                                  wsize,
+                                                                                  wstride,
+                                                                                  wdilation,
+                                                                                  pad1,
+                                                                                  pad2);
+    }
+#endif
+#ifdef CK_ENABLE_FP32
+    else if(data_type == ck::DataTypeEnum::Float)
+    {
+        ck::profiler::profile_avg_pool3d_bwd_impl<F32, F32, F32, NDHWC, NDHWC>(do_verification,
+                                                                               init_method,
+                                                                               do_log,
+                                                                               time_kernel,
+                                                                               in_length,
+                                                                               wsize,
+                                                                               wstride,
+                                                                               wdilation,
+                                                                               pad1,
+                                                                               pad2);
+    }
+#endif
+    else
+    {
+        throw std::runtime_error("not implemented yet");
+    }
+
+    return 0;
+}
+
+REGISTER_PROFILER_OPERATION("avg_pool3d_bwd", "max_pool bwd", profile_avg_pool3d_bwd);

profiler/src/profile_max_pool3d_bwd.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <vector>
+#include <unordered_map>
+
+#include "profiler/data_type_enum.hpp"
+#include "profiler/profile_max_pool3d_bwd_impl.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "profiler_operation_registry.hpp"
+
+using ck::index_t;
+
+struct maxPoolbwdArgParser
+{
+    std::unordered_map<std::string, std::vector<int>> long_opts = {{"length", {}},
+                                                                   {"wsize", {}},
+                                                                   {"wstride", {}},
+                                                                   {"wdilation", {}},
+                                                                   {"pad1", {}},
+                                                                   {"pad2", {}}};
+
+    bool parse_opt(int argc, char* argv[], const std::string& key, int i)
+    {
+        if(std::string("--") + key == argv[i])
+        {
+            int pos = i;
+            while(++i < argc && argv[i][0] != '-') {}
+            int end = i;
+            for(int j = pos + 1; j < end; j++)
+            {
+                long_opts[key].push_back(std::stoi(argv[j]));
+            }
+            return true;
+        }
+        return false;
+    }
+
+    void operator()(int argc, char* argv[])
+    {
+        for(auto& kv : long_opts)
+        {
+            for(int i = 1; i < argc; i++)
+            {
+                if(parse_opt(argc, argv, kv.first, i))
+                    break;
+            }
+        }
+    }
+};
+
+void print_help_max_pool3d_bwd()
+{
+    std::cout << "arg1: data type (0: fp16; 1: fp32; 5: bf16)\n"
+              << "arg2: verification (0: no; 1: yes)\n"
+              << "arg3: initialization (0: no init; 1: integer value; 2: decimal value)\n"
+              << "arg4: print tensor value (0: no; 1: yes)\n"
+              << "arg5: time kernel (0=no, 1=yes)\n"
+              << "--length: input tensor length for NCDHW(e.g, --length 2 32 30 30 30) \n"
+              << "--wsize: window size for ZYX (e.g, --wsize 2 2 2) \n"
+              << "--wstride: window stride for DHW (e.g, --wstride 2 2 2) \n"
+              << "--wdilation: window dilation for DHW (e.g, --wdilation 1 1 1) \n"
+              << "--pad1: left side of padding in DHW (e.g, --pad1 1 1 1) \n"
+              << "--pad2: right side of padding in DHW (e.g, --pad2 1 1 1) \n"
+              << "eg: ckProfiler max_pool3d_bwd 0 1 2 0 1 --length 2 32 30 30 30 --wsize 2 2 2 "
+                 "--wstride 2 2 2 --wdilation 1 1 1 --pad1 1 1 1 --pad2 1 1 1"
+              << std::endl;
+}
+
+int profile_max_pool3d_bwd(int argc, char* argv[])
+{
+    ck::DataTypeEnum data_type = ck::DataTypeEnum::Half;
+    bool do_verification       = true;
+    int init_method            = 0;
+    bool do_log                = false;
+    bool time_kernel           = true;
+
+    std::vector<index_t> in_length = {2, 32, 30, 30, 30};
+    std::vector<index_t> wsize     = {2, 2, 2};
+    std::vector<index_t> wstride   = {2, 2, 2};
+    std::vector<index_t> wdilation = {1, 1, 1};
+    std::vector<index_t> pad1      = {1, 1, 1};
+    std::vector<index_t> pad2      = {1, 1, 1};
+
+    if(argc != 2 && argc != 33)
+    {
+        print_help_max_pool3d_bwd();
+        return 0;
+    }
+    else if(argc == 33)
+    {
+        data_type       = static_cast<ck::DataTypeEnum>(std::stoi(argv[2]));
+        do_verification = std::stoi(argv[3]);
+        init_method     = std::stoi(argv[4]);
+        do_log          = std::stoi(argv[5]);
+        time_kernel     = std::stoi(argv[6]);
+
+        // parse the long options
+        maxPoolbwdArgParser arg_parser;
+        arg_parser(argc, argv);
+        in_length = arg_parser.long_opts["length"];
+        wsize     = arg_parser.long_opts["wsize"];
+        wstride   = arg_parser.long_opts["wstride"];
+        wdilation = arg_parser.long_opts["wdilation"];
+        pad1      = arg_parser.long_opts["pad1"];
+        pad2      = arg_parser.long_opts["pad2"];
+    }
+
+#ifdef CK_ENABLE_FP16
+    using F16 = ck::half_t;
+#endif
+#ifdef CK_ENABLE_BF16
+    using BF16 = ck::bhalf_t;
+#endif
+#ifdef CK_ENABLE_FP32
+    using F32 = float;
+#endif
+    using I32 = int32_t;
+
+    if(false)
+        ;
+#ifdef CK_ENABLE_FP16
+    else if(data_type == ck::DataTypeEnum::Half)
+    {
+        ck::profiler::profile_max_pool3d_bwd_impl<F16, F16, I32, F16, F16, false>(do_verification,
+                                                                                  init_method,
+                                                                                  do_log,
+                                                                                  time_kernel,
+                                                                                  in_length,
+                                                                                  wsize,
+                                                                                  wstride,
+                                                                                  wdilation,
+                                                                                  pad1,
+                                                                                  pad2);
+    }
+#endif
+#ifdef CK_ENABLE_BF16
+    else if(data_type == ck::DataTypeEnum::BFloat16)
+    {
+        ck::profiler::profile_max_pool3d_bwd_impl<BF16, BF16, I32, BF16, BF16, false>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            in_length,
+            wsize,
+            wstride,
+            wdilation,
+            pad1,
+            pad2);
+    }
+#endif
+#ifdef CK_ENABLE_FP32
+    else if(data_type == ck::DataTypeEnum::Float)
+    {
+        ck::profiler::profile_max_pool3d_bwd_impl<F32, F32, I32, F32, F32, false>(do_verification,
+                                                                                  init_method,
+                                                                                  do_log,
+                                                                                  time_kernel,
+                                                                                  in_length,
+                                                                                  wsize,
+                                                                                  wstride,
+                                                                                  wdilation,
+                                                                                  pad1,
+                                                                                  pad2);
+    }
+#endif
+    else
+    {
+        throw std::runtime_error("not implemented yet");
+    }
+
+    return 0;
+}
+
+REGISTER_PROFILER_OPERATION("max_pool3d_bwd", "max_pool3d bwd", profile_max_pool3d_bwd);

profiler/src/profile_max_pool3d_fwd.cpp

@@ -51,7 +51,7 @@ struct maxPoolFwdArgParser
 
 void print_help_max_pool3d_fwd()
 {
-    std::cout << "arg1: data type (0: fp16; 1: fp32)\n"
+    std::cout << "arg1: data type (0: fp16; 1: fp32; 5: bf16)\n"
               << "arg2: verification (0: no; 1: yes)\n"
               << "arg3: initialization (0: no init; 1: integer value; 2: decimal value)\n"
               << "arg4: print tensor value (0: no; 1: yes)\n"
@@ -109,8 +109,15 @@ int profile_max_pool3d_fwd(int argc, char* argv[])
         pad2      = arg_parser.long_opts["pad2"];
     }
 
-    using F16   = ck::half_t;
-    using F32   = float;
+#ifdef CK_ENABLE_FP16
+    using F16 = ck::half_t;
+#endif
+#ifdef CK_ENABLE_BF16
+    using BF16 = ck::bhalf_t;
+#endif
+#ifdef CK_ENABLE_FP32
+    using F32 = float;
+#endif
     using I32   = int32_t;
     using NDHWC = ck::tensor_layout::convolution::NDHWC;
 
@@ -120,7 +127,10 @@ int profile_max_pool3d_fwd(int argc, char* argv[])
     constexpr auto ReduceOpId = ck::ReduceTensorOp::AVG;
 #endif
 
-    if(data_type == ck::DataTypeEnum::Half)
+    if(false)
+        ;
+#ifdef CK_ENABLE_FP16
+    else if(data_type == ck::DataTypeEnum::Half)
     {
         if(return_index)
             ck::profiler::
@@ -149,6 +159,51 @@ int profile_max_pool3d_fwd(int argc, char* argv[])
                     pad1,
                     pad2);
     }
+#endif
+#ifdef CK_ENABLE_BF16
+    else if(data_type == ck::DataTypeEnum::BFloat16)
+    {
+        if(return_index)
+            ck::profiler::profile_pool3d_fwd_impl<BF16,
+                                                  BF16,
+                                                  BF16,
+                                                  I32,
+                                                  NDHWC,
+                                                  NDHWC,
+                                                  ReduceOpId,
+                                                  false,
+                                                  true>(do_verification,
+                                                        init_method,
+                                                        do_log,
+                                                        time_kernel,
+                                                        in_length,
+                                                        wsize,
+                                                        wstride,
+                                                        wdilation,
+                                                        pad1,
+                                                        pad2);
+        else
+            ck::profiler::profile_pool3d_fwd_impl<BF16,
+                                                  BF16,
+                                                  BF16,
+                                                  I32,
+                                                  NDHWC,
+                                                  NDHWC,
+                                                  ReduceOpId,
+                                                  false,
+                                                  false>(do_verification,
+                                                         init_method,
+                                                         do_log,
+                                                         time_kernel,
+                                                         in_length,
+                                                         wsize,
+                                                         wstride,
+                                                         wdilation,
+                                                         pad1,
+                                                         pad2);
+    }
+#endif
+#ifdef CK_ENABLE_FP32
     else if(data_type == ck::DataTypeEnum::Float)
     {
         if(return_index)
@@ -178,6 +233,7 @@ int profile_max_pool3d_fwd(int argc, char* argv[])
                     pad1,
                     pad2);
     }
+#endif
     else
     {
         throw std::runtime_error("not implemented yet");

test/CMakeLists.txt

@@ -57,7 +57,7 @@ add_subdirectory(data_type)
 add_subdirectory(elementwise_normalization)
 add_subdirectory(batchnorm)
 add_subdirectory(contraction)
-add_subdirectory(pool_fwd)
+add_subdirectory(pool)
 add_subdirectory(batched_gemm_multi_d)
 add_subdirectory(grouped_convnd_bwd_data)
 if(GPU_TARGETS MATCHES "gfx11")