Rewrite pool2d fwd (#1462)

* added pool2d fwd

* add tests

* add reviewers changes

* Revert "Merge remote-tracking branch 'origin/develop' into jakpiase/pool2d_fwd_new"

This reverts commit 6b2ba7ff8960b0a6ddbe30d8dac53eeb55a8597e, reversing
changes made to 22c82bea0caf3e0f29399100c1bb67b8003fc042.

* Revert "add reviewers changes"

This reverts commit 22c82bea0caf3e0f29399100c1bb67b8003fc042.

* added reviewers comments

* revert some old files

* add reviewers requests

---------
Co-authored-by: Adam Osewski <19374865+aosewski@users.noreply.github.com>

library/include/ck/library/tensor_operation_instance/gpu/pool2d_fwd.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_pool_fwd.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr auto InOutRank  = 4;
+static constexpr auto WindowRank = 2;
+
+static constexpr auto MaxOp = ck::ReduceTensorOp::MAX;
+static constexpr auto AvgOp = ck::ReduceTensorOp::AVG;
+#ifdef CK_ENABLE_FP16
+// FP16
+void add_device_pool2d_fwd_nhwc_f16_instances(
+    std::vector<std::unique_ptr<
+        DevicePoolFwd<InOutRank, WindowRank, F16, F16, I32, NHWC, NHWC, MaxOp, false>>>&);
+
+void add_device_pool2d_fwd_nhwc_f16_instances(
+    std::vector<std::unique_ptr<
+        DevicePoolFwd<InOutRank, WindowRank, F16, F16, I32, NHWC, NHWC, AvgOp, false>>>&);
+
+// FP16 - return index
+void add_device_pool2d_fwd_nhwc_index_f16_instances(
+    std::vector<std::unique_ptr<
+        DevicePoolFwd<InOutRank, WindowRank, F16, F16, I32, NHWC, NHWC, MaxOp, true>>>&);
+#endif
+#ifdef CK_ENABLE_BF16
+// BF16
+void add_device_pool2d_fwd_nhwc_bf16_instances(
+    std::vector<std::unique_ptr<
+        DevicePoolFwd<InOutRank, WindowRank, BF16, BF16, I32, NHWC, NHWC, MaxOp, false>>>&);
+
+void add_device_pool2d_fwd_nhwc_bf16_instances(
+    std::vector<std::unique_ptr<
+        DevicePoolFwd<InOutRank, WindowRank, BF16, BF16, I32, NHWC, NHWC, AvgOp, false>>>&);
+
+// BF16 - return index
+void add_device_pool2d_fwd_nhwc_index_bf16_instances(
+    std::vector<std::unique_ptr<
+        DevicePoolFwd<InOutRank, WindowRank, BF16, BF16, I32, NHWC, NHWC, MaxOp, true>>>&);
+#endif
+#ifdef CK_ENABLE_FP32
+// FP32
+void add_device_pool2d_fwd_nhwc_f32_instances(
+    std::vector<std::unique_ptr<
+        DevicePoolFwd<InOutRank, WindowRank, F32, F32, I32, NHWC, NHWC, MaxOp, false>>>&);
+
+void add_device_pool2d_fwd_nhwc_f32_instances(
+    std::vector<std::unique_ptr<
+        DevicePoolFwd<InOutRank, WindowRank, F32, F32, I32, NHWC, NHWC, AvgOp, false>>>&);
+
+// FP32 - return index
+void add_device_pool2d_fwd_nhwc_index_f32_instances(
+    std::vector<std::unique_ptr<
+        DevicePoolFwd<InOutRank, WindowRank, F32, F32, I32, NHWC, NHWC, MaxOp, true>>>&);
+#endif
+template <typename InDataType,
+          typename OutDataType,
+          typename IndexDataType,
+          typename InLayout,
+          typename OutLayout,
+          ck::ReduceTensorOp ReduceOpId,
+          bool OutputIndex>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DevicePoolFwd<InOutRank,
+                                                                                  WindowRank,
+                                                                                  InDataType,
+                                                                                  OutDataType,
+                                                                                  IndexDataType,
+                                                                                  InLayout,
+                                                                                  OutLayout,
+                                                                                  ReduceOpId,
+                                                                                  OutputIndex>>
+{
+    using DeviceOp = DevicePoolFwd<InOutRank,
+                                   WindowRank,
+                                   InDataType,
+                                   OutDataType,
+                                   IndexDataType,
+                                   InLayout,
+                                   OutLayout,
+                                   ReduceOpId,
+                                   OutputIndex>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+        if constexpr(is_same_v<InLayout, NHWC> && is_same_v<OutLayout, NHWC>)
+        {
+#ifdef CK_ENABLE_FP16
+            if constexpr(is_same_v<InDataType, F16> && is_same_v<OutDataType, F16> &&
+                         is_same_v<IndexDataType, I32>)
+            {
+                if constexpr(OutputIndex && ReduceOpId == MaxOp)
+                {
+                    add_device_pool2d_fwd_nhwc_index_f16_instances(op_ptrs);
+                }
+                else
+                {
+                    add_device_pool2d_fwd_nhwc_f16_instances(op_ptrs);
+                }
+            }
+#endif
+#ifdef CK_ENABLE_BF16
+            else if constexpr(is_same_v<InDataType, BF16> && is_same_v<OutDataType, BF16> &&
+                              is_same_v<IndexDataType, I32>)
+            {
+                if constexpr(OutputIndex && ReduceOpId == MaxOp)
+                {
+                    add_device_pool2d_fwd_nhwc_index_bf16_instances(op_ptrs);
+                }
+                else
+                {
+                    add_device_pool2d_fwd_nhwc_bf16_instances(op_ptrs);
+                }
+            }
+#endif
+#ifdef CK_ENABLE_FP32
+            else if constexpr(is_same_v<InDataType, F32> && is_same_v<OutDataType, F32> &&
+                              is_same_v<IndexDataType, I32>)
+            {
+                if constexpr(OutputIndex && ReduceOpId == MaxOp)
+                {
+                    add_device_pool2d_fwd_nhwc_index_f32_instances(op_ptrs);
+                }
+                else
+                {
+                    add_device_pool2d_fwd_nhwc_f32_instances(op_ptrs);
+                }
+            }
+#endif
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/pool2d_fwd/CMakeLists.txt

+set(DEVICE_POOL2D_FWD_INSTANCES)
+list(APPEND DEVICE_POOL2D_FWD_INSTANCES device_avg_pool2d_fwd_nhwc_f16_instance.cpp
+                                        device_max_pool2d_fwd_nhwc_f16_instance.cpp
+                                        device_avg_pool2d_fwd_nhwc_f32_instance.cpp
+                                        device_max_pool2d_fwd_nhwc_f32_instance.cpp
+                                        device_avg_pool2d_fwd_nhwc_bf16_instance.cpp
+                                        device_max_pool2d_fwd_nhwc_bf16_instance.cpp)
+add_instance_library(device_pool2d_fwd_instance ${DEVICE_POOL2D_FWD_INSTANCES})

library/src/tensor_operation_instance/gpu/pool2d_fwd/device_avg_pool2d_fwd_nhwc_bf16_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "pool2d_fwd_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr auto ReduceOpId = ck::ReduceTensorOp::AVG;
+
+void add_device_pool2d_fwd_nhwc_bf16_instances(
+    std::vector<
+        std::unique_ptr<DevicePoolFwd<4, 2, BF16, BF16, I32, NHWC, NHWC, ReduceOpId, false>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances, device_pool2d_fwd_nhwc_instances<BF16, BF16, I32, F32, ReduceOpId, false>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/pool2d_fwd/device_avg_pool2d_fwd_nhwc_f16_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "pool2d_fwd_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr auto ReduceOpId = ck::ReduceTensorOp::AVG;
+
+void add_device_pool2d_fwd_nhwc_f16_instances(
+    std::vector<std::unique_ptr<DevicePoolFwd<4, 2, F16, F16, I32, NHWC, NHWC, ReduceOpId, false>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances, device_pool2d_fwd_nhwc_instances<F16, F16, I32, F32, ReduceOpId, false>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/pool2d_fwd/device_avg_pool2d_fwd_nhwc_f32_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "pool2d_fwd_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr auto ReduceOpId = ck::ReduceTensorOp::AVG;
+
+void add_device_pool2d_fwd_nhwc_f32_instances(
+    std::vector<std::unique_ptr<DevicePoolFwd<4, 2, F32, F32, I32, NHWC, NHWC, ReduceOpId, false>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances, device_pool2d_fwd_nhwc_instances<F32, F32, I32, F32, ReduceOpId, false>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/pool2d_fwd/device_max_pool2d_fwd_nhwc_bf16_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "pool2d_fwd_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr auto ReduceOpId = ck::ReduceTensorOp::MAX;
+
+void add_device_pool2d_fwd_nhwc_bf16_instances(
+    std::vector<
+        std::unique_ptr<DevicePoolFwd<4, 2, BF16, BF16, I32, NHWC, NHWC, ReduceOpId, false>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances, device_pool2d_fwd_nhwc_instances<BF16, BF16, I32, F32, ReduceOpId, false>{});
+}
+
+void add_device_pool2d_fwd_nhwc_index_bf16_instances(
+    std::vector<
+        std::unique_ptr<DevicePoolFwd<4, 2, BF16, BF16, I32, NHWC, NHWC, ReduceOpId, true>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances, device_pool2d_fwd_nhwc_instances<BF16, BF16, I32, F32, ReduceOpId, true>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/pool2d_fwd/device_max_pool2d_fwd_nhwc_f16_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "pool2d_fwd_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr auto ReduceOpId = ck::ReduceTensorOp::MAX;
+
+void add_device_pool2d_fwd_nhwc_f16_instances(
+    std::vector<std::unique_ptr<DevicePoolFwd<4, 2, F16, F16, I32, NHWC, NHWC, ReduceOpId, false>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances, device_pool2d_fwd_nhwc_instances<F16, F16, I32, F32, ReduceOpId, false>{});
+}
+
+void add_device_pool2d_fwd_nhwc_index_f16_instances(
+    std::vector<std::unique_ptr<DevicePoolFwd<4, 2, F16, F16, I32, NHWC, NHWC, ReduceOpId, true>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances, device_pool2d_fwd_nhwc_instances<F16, F16, I32, F32, ReduceOpId, true>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/pool2d_fwd/device_max_pool2d_fwd_nhwc_f32_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "pool2d_fwd_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr auto ReduceOpId = ck::ReduceTensorOp::MAX;
+
+void add_device_pool2d_fwd_nhwc_f32_instances(
+    std::vector<std::unique_ptr<DevicePoolFwd<4, 2, F32, F32, I32, NHWC, NHWC, ReduceOpId, false>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances, device_pool2d_fwd_nhwc_instances<F32, F32, I32, F32, ReduceOpId, false>{});
+}
+
+void add_device_pool2d_fwd_nhwc_index_f32_instances(
+    std::vector<std::unique_ptr<DevicePoolFwd<4, 2, F32, F32, I32, NHWC, NHWC, ReduceOpId, true>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances, device_pool2d_fwd_nhwc_instances<F32, F32, I32, F32, ReduceOpId, true>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/pool2d_fwd/pool2d_fwd_instance_common.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_pool2d_fwd_nhwc_nhwc.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 NHWC = ck::tensor_layout::convolution::NHWC;
+
+template <typename InDataType,
+          typename OutDataType,
+          typename IndexDataType,
+          typename ComputeDataType,
+          ReduceTensorOp ReduceOpId,
+          bool OutputIndex>
+using device_pool2d_fwd_nhwc_instances =
+    // clang-format off
+    std::tuple <
+        DevicePool2dFwd_NHWC_NHWC<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 1, 1, 1>,
+        DevicePool2dFwd_NHWC_NHWC<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 2, 1, 2>,
+        DevicePool2dFwd_NHWC_NHWC<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 4, 1, 4>
+               // clang-format on
+               >;
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

profiler/include/profiler/profile_pool2d_fwd_impl.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/pool2d_fwd.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"
+
+namespace ck {
+namespace profiler {
+
+template <typename InDataType,
+          typename OutDataType,
+          typename ComputeDataType,
+          typename IndexDataType,
+          typename InLayout,
+          typename OutLayout,
+          ck::ReduceTensorOp ReduceOpId,
+          bool PropagateNan,
+          bool OutputIndex>
+bool profile_pool2d_fwd_impl(int do_verification,
+                             int init_method,
+                             bool do_log,
+                             bool time_kernel,
+                             std::vector<index_t> in_length, // NCHW
+                             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  = 4;
+    constexpr index_t WindowRank = 2;
+
+    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)
+        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 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 Hi = in_length[2];
+    int Wi = in_length[3];
+    int Ho = out_length[2];
+    int Wo = out_length[3];
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t N_, std::size_t C_, std::size_t H, std::size_t W) {
+            using namespace ck::literals;
+
+            return HostTensorDescriptor({N_, C_, H, W}, {C_ * H * W, 1_uz, W * C_, C_});
+        };
+
+    Tensor<InDataType> in_n_c_hi_wi(f_host_tensor_descriptor(N, C, Hi, Wi));
+    Tensor<OutDataType> out_n_c_ho_wo_host(f_host_tensor_descriptor(N, C, Ho, Wo));
+    Tensor<IndexDataType> out_indices_n_c_ho_wo_host(f_host_tensor_descriptor(N, C, Ho, Wo));
+
+    Tensor<OutDataType> out_n_c_ho_wo_device(f_host_tensor_descriptor(N, C, Ho, Wo));
+    Tensor<IndexDataType> out_indices_n_c_ho_wo_device(f_host_tensor_descriptor(N, C, Ho, Wo));
+
+    switch(init_method)
+    {
+    case 0: in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_1<InDataType>{}); break;
+    case 1: in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5}); break;
+    default: in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_3<InDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem in_device_buf(sizeof(InDataType) * in_n_c_hi_wi.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) *
+                             out_n_c_ho_wo_device.mDesc.GetElementSpaceSize());
+    DeviceMem out_indices_device_buf(sizeof(IndexDataType) *
+                                     out_indices_n_c_ho_wo_device.mDesc.GetElementSpaceSize());
+
+    in_device_buf.ToDevice(in_n_c_hi_wi.mData.data());
+
+    // add device normalization instances
+    using DeviceOp = ck::tensor_operation::device::DevicePoolFwd<InOutRank,
+                                                                 WindowRank,
+                                                                 InDataType,
+                                                                 OutDataType,
+                                                                 IndexDataType,
+                                                                 InLayout,
+                                                                 OutLayout,
+                                                                 ReduceOpId,
+                                                                 OutputIndex>;
+
+    // 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 ReferenceInstance = ck::tensor_operation::host::ReferencePoolingFwd<InOutRank,
+                                                                                  WindowRank,
+                                                                                  InDataType,
+                                                                                  OutDataType,
+                                                                                  ComputeDataType,
+                                                                                  IndexDataType,
+                                                                                  ReduceOpId,
+                                                                                  PropagateNan,
+                                                                                  OutputIndex>;
+
+        ReferenceInstance ref;
+        auto ref_argument = ref.MakeArgument(in_n_c_hi_wi,
+                                             out_n_c_ho_wo_host,
+                                             out_indices_n_c_ho_wo_host,
+                                             window_spatial_lengths,
+                                             window_strides,
+                                             window_dilations,
+                                             input_left_pads,
+                                             input_right_pads);
+        auto ref_invoker  = ref.MakeInvoker();
+        ref_invoker.Run(ref_argument);
+    }
+
+    int num_kernel = 0;
+
+    for(auto& inst_ptr : instance_ptrs)
+    {
+        auto argument_ptr = inst_ptr->MakeArgumentPointer(
+            static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
+            static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
+            static_cast<IndexDataType*>(out_indices_device_buf.GetDeviceBuffer()),
+            in_length,
+            window_spatial_lengths,
+            out_length,
+            {C * Hi * Wi, 1, Wi * C, C},
+            {C * Ho * Wo, 1, Wo * C, C},
+            {C * Ho * Wo, 1, Wo * C, C},
+            window_strides,
+            window_dilations,
+            input_left_pads,
+            input_right_pads,
+            {2, 3});
+
+        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 << "input lengths = ", in_length, ", ") << std::endl;
+            }
+
+            continue;
+        }
+
+        auto invoker_ptr = inst_ptr->MakeInvokerPointer();
+
+        float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+        std::size_t num_bytes = in_n_c_hi_wi.mDesc.GetElementSize() * sizeof(InDataType) +
+                                out_n_c_ho_wo_host.mDesc.GetElementSize() * sizeof(OutDataType);
+
+        if constexpr(OutputIndex)
+            num_bytes += out_indices_n_c_ho_wo_host.mDesc.GetElementSize() * sizeof(IndexDataType);
+
+        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)
+        {
+            out_device_buf.FromDevice(out_n_c_ho_wo_device.mData.data());
+
+            bool pass = ck::utils::check_err(out_n_c_ho_wo_device.mData,
+                                             out_n_c_ho_wo_host.mData,
+                                             "Error: Incorrect results",
+                                             1e-3,
+                                             1e-3);
+
+            if constexpr(OutputIndex)
+            {
+                out_indices_device_buf.FromDevice(out_indices_n_c_ho_wo_device.mData.data());
+
+                pass = pass && ck::utils::check_err(out_indices_n_c_ho_wo_device,
+                                                    out_indices_n_c_ho_wo_host);
+            }
+
+            if(do_log)
+            {
+                LogRangeAsType<float>(std::cout << "in_n_c_hi_wi  : ", in_n_c_hi_wi.mData, ",")
+                    << std::endl;
+                LogRangeAsType<float>(
+                    std::cout << "out_n_c_ho_wo_host  : ", out_n_c_ho_wo_host.mData, ",")
+                    << std::endl;
+                LogRangeAsType<float>(
+                    std::cout << "out_n_c_ho_wo_device  : ", out_n_c_ho_wo_device.mData, ",")
+                    << std::endl;
+
+                if constexpr(OutputIndex)
+                    LogRangeAsType<float>(std::cout << "out_indices_n_c_ho_wo_device  : ",
+                                          out_indices_n_c_ho_wo_device.mData,
+                                          ",")
+                        << std::endl;
+            }
+
+            if(!pass)
+            {
+                std::cout << inst_ptr->GetTypeString() << " failed verification: ";
+                LogRange(std::cout << "lengths = [", in_length, ", ") << "]." << std::endl;
+                return false;
+            }
+            else
+            {
+                if(time_kernel)
+                    std::cout << "pass" << std::endl;
+            }
+        }
+    }
+
+    if(time_kernel)
+    {
+        LogRange(std::cout << "length = ", in_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

@@ -9,6 +9,7 @@ set(PROFILER_SOURCES
     profile_layernorm_bwd_gamma_beta.cpp
     profile_groupnorm_bwd_gamma_beta.cpp
     profile_layernorm_fwd.cpp
+    profile_max_pool2d_fwd.cpp
     profile_max_pool3d_fwd.cpp
     profile_avg_pool3d_bwd.cpp
     profile_max_pool3d_bwd.cpp
@@ -98,6 +99,7 @@ target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_bwd_ga
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_softmax_instance)
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_reduce_instance)
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batchnorm_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool2d_fwd_instance)
 target_link_libraries(${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)

profiler/src/profile_max_pool2d_fwd.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <vector>
+#include <unordered_map>
+
+#include "profiler/data_type_enum.hpp"
+#include "profiler/profile_pool2d_fwd_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+using ck::index_t;
+
+struct maxPoolFwdArgParser
+{
+    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_pool2d_fwd()
+{
+    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"
+              << "arg6: return index (0=no, 1=yes)\n"
+              << "--length: input tensor length for NCHW(e.g, --length 2 32 30 30) \n"
+              << "--wsize: window size for YX (e.g, --wsize 2 2) \n"
+              << "--wstride: window stride for HW (e.g, --wstride 2 2) \n"
+              << "--wdilation: window dilation for HW (e.g, --wdilation 1 1) \n"
+              << "--pad1: left side of padding in HW (e.g, --pad1 1 1) \n"
+              << "--pad2: right side of padding in HW (e.g, --pad2 1 1) \n"
+              << "eg: ckProfiler max_pool2d_fwd 0 1 2 0 1 0 --length 2 32 30 30 --wsize 2 2"
+                 "--wstride 2 2 --wdilation 1 1 --pad1 1 1 --pad2 1 1"
+              << std::endl;
+}
+
+int profile_max_pool2d_fwd(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;
+    bool return_index          = false;
+
+    std::vector<index_t> in_length = {2, 32, 30, 30};
+    std::vector<index_t> wsize     = {2, 2};
+    std::vector<index_t> wstride   = {2, 2};
+    std::vector<index_t> wdilation = {1, 1};
+    std::vector<index_t> pad1      = {1, 1};
+    std::vector<index_t> pad2      = {1, 1};
+
+    if(argc != 2 && argc != 28)
+    {
+        print_help_max_pool2d_fwd();
+        return 0;
+    }
+    else if(argc == 28)
+    {
+        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]);
+        return_index    = std::stoi(argv[7]);
+
+        // parse the long options
+        maxPoolFwdArgParser 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"];
+    }
+
+    using F16  = ck::half_t;
+    using BF16 = ck::bhalf_t;
+    using F32  = float;
+    using I32  = int32_t;
+    using NHWC = ck::tensor_layout::convolution::NHWC;
+
+    constexpr auto ReduceOpId = ck::ReduceTensorOp::MAX;
+
+    if(data_type == ck::DataTypeEnum::Half)
+    {
+        if(return_index)
+        {
+            ck::profiler::
+                profile_pool2d_fwd_impl<F16, F16, F16, I32, NHWC, NHWC, ReduceOpId, false, true>(
+                    do_verification,
+                    init_method,
+                    do_log,
+                    time_kernel,
+                    in_length,
+                    wsize,
+                    wstride,
+                    wdilation,
+                    pad1,
+                    pad2);
+        }
+        else
+        {
+            ck::profiler::
+                profile_pool2d_fwd_impl<F16, F16, F16, I32, NHWC, NHWC, ReduceOpId, false, false>(
+                    do_verification,
+                    init_method,
+                    do_log,
+                    time_kernel,
+                    in_length,
+                    wsize,
+                    wstride,
+                    wdilation,
+                    pad1,
+                    pad2);
+        }
+    }
+    else if(data_type == ck::DataTypeEnum::BFloat16)
+    {
+        if(return_index)
+        {
+            ck::profiler::
+                profile_pool2d_fwd_impl<BF16, BF16, BF16, I32, NHWC, NHWC, ReduceOpId, false, true>(
+                    do_verification,
+                    init_method,
+                    do_log,
+                    time_kernel,
+                    in_length,
+                    wsize,
+                    wstride,
+                    wdilation,
+                    pad1,
+                    pad2);
+        }
+        else
+        {
+            ck::profiler::profile_pool2d_fwd_impl<BF16,
+                                                  BF16,
+                                                  BF16,
+                                                  I32,
+                                                  NHWC,
+                                                  NHWC,
+                                                  ReduceOpId,
+                                                  false,
+                                                  false>(do_verification,
+                                                         init_method,
+                                                         do_log,
+                                                         time_kernel,
+                                                         in_length,
+                                                         wsize,
+                                                         wstride,
+                                                         wdilation,
+                                                         pad1,
+                                                         pad2);
+        }
+    }
+    else if(data_type == ck::DataTypeEnum::Float)
+    {
+        if(return_index)
+        {
+            ck::profiler::
+                profile_pool2d_fwd_impl<F32, F32, F32, I32, NHWC, NHWC, ReduceOpId, false, true>(
+                    do_verification,
+                    init_method,
+                    do_log,
+                    time_kernel,
+                    in_length,
+                    wsize,
+                    wstride,
+                    wdilation,
+                    pad1,
+                    pad2);
+        }
+        else
+        {
+            ck::profiler::
+                profile_pool2d_fwd_impl<F32, F32, F32, I32, NHWC, NHWC, ReduceOpId, false, false>(
+                    do_verification,
+                    init_method,
+                    do_log,
+                    time_kernel,
+                    in_length,
+                    wsize,
+                    wstride,
+                    wdilation,
+                    pad1,
+                    pad2);
+        }
+    }
+    else
+    {
+        throw std::runtime_error("not implemented yet");
+    }
+
+    return 0;
+}
+
+REGISTER_PROFILER_OPERATION("max_pool2d_fwd", "max_pool2d fwd", profile_max_pool2d_fwd);

test/pool/CMakeLists.txt

@@ -4,13 +4,19 @@ add_gtest_executable(test_avg_pool3d_bwd test_avg_pool3d_bwd.cpp)
 add_gtest_executable(test_max_pool3d_bwd test_max_pool3d_bwd.cpp)
 add_gtest_executable(test_avg_pool3d_fwd test_avg_pool3d_fwd.cpp)
 add_gtest_executable(test_max_pool3d_fwd test_max_pool3d_fwd.cpp)
+add_gtest_executable(test_avg_pool2d_fwd test_avg_pool2d_fwd.cpp)
+add_gtest_executable(test_max_pool2d_fwd test_max_pool2d_fwd.cpp)
 
 target_link_libraries(test_avg_pool3d_bwd PRIVATE utility device_avg_pool3d_bwd_instance)
 target_link_libraries(test_max_pool3d_bwd PRIVATE utility device_max_pool_bwd_instance)
 target_link_libraries(test_avg_pool3d_fwd PRIVATE utility device_pool3d_fwd_instance)
 target_link_libraries(test_max_pool3d_fwd PRIVATE utility device_pool3d_fwd_instance)
+target_link_libraries(test_avg_pool2d_fwd PRIVATE utility device_pool2d_fwd_instance)
+target_link_libraries(test_max_pool2d_fwd PRIVATE utility device_pool2d_fwd_instance)
 
 add_dependencies(test_pool test_avg_pool3d_bwd)
 add_dependencies(test_pool test_max_pool3d_bwd)
 add_dependencies(test_pool test_avg_pool3d_fwd)
 add_dependencies(test_pool test_max_pool3d_fwd)
+add_dependencies(test_pool test_avg_pool2d_fwd)
+add_dependencies(test_pool test_max_pool2d_fwd)

test/pool/test_avg_pool2d_fwd.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gtest/gtest.h"
+#include "profiler/profile_pool2d_fwd_impl.hpp"
+#include "test_pool_fwd_common.hpp"
+
+template <typename Tuple>
+class TestAvgPool2dFwd : public ::testing::Test
+{
+    protected:
+    using InDataType      = std::tuple_element_t<0, Tuple>;
+    using OutDataType     = std::tuple_element_t<1, Tuple>;
+    using ComputeDataType = std::tuple_element_t<2, Tuple>;
+    using IndexDataType   = std::tuple_element_t<3, Tuple>;
+
+    std::vector<PoolingParam> params;
+
+    void Run()
+    {
+        for(auto param : params)
+        {
+            // avg pool
+            bool success =
+                ck::profiler::profile_pool2d_fwd_impl<InDataType,
+                                                      OutDataType,
+                                                      ComputeDataType,
+                                                      IndexDataType,
+                                                      ck::tensor_layout::convolution::NHWC,
+                                                      ck::tensor_layout::convolution::NHWC,
+                                                      ck::ReduceTensorOp::AVG,
+                                                      false,
+                                                      false>(true,
+                                                             2,
+                                                             false,
+                                                             false,
+                                                             param.length_,
+                                                             param.window_spatial_lengths_,
+                                                             param.window_strides_,
+                                                             param.window_dilations_,
+                                                             param.input_left_pads_,
+                                                             param.input_right_pads_);
+            EXPECT_TRUE(success);
+        }
+    }
+};
+
+using KernelTypes = std::conditional_t<
+    CK_ENABLE_FP16 && CK_ENABLE_BF16,
+    ::testing::Types<std::tuple<F16, F16, F32, I32>,
+                     std::tuple<F16, F16, F32, I32>,
+                     std::tuple<BF16, BF16, F32, I32>,
+                     std::tuple<BF16, BF16, F32, I32>,
+                     std::tuple<F32, F32, F32, I32>,
+                     std::tuple<F32, F32, F32, I32>>,
+    ::testing::Types<std::tuple<F32, F32, F32, I32>, std::tuple<F32, F32, F32, I32>>>;
+
+TYPED_TEST_SUITE(TestAvgPool2dFwd, KernelTypes);
+TYPED_TEST(TestAvgPool2dFwd, Test_Pool)
+{
+    // length, window_length, window_stride, window_dilation, left_pad, right_pad
+    this->params = {{{1, 1, 1, 1}, {1, 1}, {1, 1}, {1, 1}, {0, 0}, {0, 0}},
+                    {{2, 16, 64, 64}, {64, 64}, {1, 1}, {1, 1}, {0, 0}, {0, 0}},
+                    {{2, 16, 64, 64}, {4, 4}, {4, 4}, {2, 2}, {0, 0}, {0, 0}},
+                    {{2, 32, 30, 30}, {2, 2}, {2, 2}, {1, 1}, {1, 1}, {1, 1}}};
+
+    this->Run();
+}

test/pool/test_max_pool2d_fwd.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gtest/gtest.h"
+#include "profiler/profile_pool2d_fwd_impl.hpp"
+#include "test_pool_fwd_common.hpp"
+
+template <typename Tuple>
+class TestMaxPool2dFwd : public ::testing::Test
+{
+    protected:
+    using InDataType                  = std::tuple_element_t<0, Tuple>;
+    using OutDataType                 = std::tuple_element_t<1, Tuple>;
+    using ComputeDataType             = std::tuple_element_t<2, Tuple>;
+    using IndexDataType               = std::tuple_element_t<3, Tuple>;
+    static constexpr bool ReturnIndex = std::tuple_element_t<4, Tuple>::value;
+
+    std::vector<PoolingParam> params;
+
+    void Run()
+    {
+        for(auto param : params)
+        {
+            // max pool
+            bool success =
+                ck::profiler::profile_pool2d_fwd_impl<InDataType,
+                                                      OutDataType,
+                                                      ComputeDataType,
+                                                      IndexDataType,
+                                                      ck::tensor_layout::convolution::NHWC,
+                                                      ck::tensor_layout::convolution::NHWC,
+                                                      ck::ReduceTensorOp::MAX,
+                                                      false,
+                                                      ReturnIndex>(true,
+                                                                   2,
+                                                                   false,
+                                                                   false,
+                                                                   param.length_,
+                                                                   param.window_spatial_lengths_,
+                                                                   param.window_strides_,
+                                                                   param.window_dilations_,
+                                                                   param.input_left_pads_,
+                                                                   param.input_right_pads_);
+            EXPECT_TRUE(success);
+        }
+    }
+};
+
+using true_t  = std::integral_constant<bool, true>;
+using false_t = std::integral_constant<bool, false>;
+
+using KernelTypes = std::conditional_t<CK_ENABLE_FP16 && CK_ENABLE_BF16,
+                                       ::testing::Types<std::tuple<F16, F16, F32, I32, true_t>,
+                                                        std::tuple<F16, F16, F32, I32, false_t>,
+                                                        std::tuple<BF16, BF16, F32, I32, true_t>,
+                                                        std::tuple<BF16, BF16, F32, I32, false_t>,
+                                                        std::tuple<F32, F32, F32, I32, true_t>,
+                                                        std::tuple<F32, F32, F32, I32, false_t>>,
+                                       ::testing::Types<std::tuple<F32, F32, F32, I32, true_t>,
+                                                        std::tuple<F32, F32, F32, I32, false_t>>>;
+
+TYPED_TEST_SUITE(TestMaxPool2dFwd, KernelTypes);
+TYPED_TEST(TestMaxPool2dFwd, Test_Pool)
+{
+    // length, window_length, window_stride, window_dilation, left_pad, right_pad
+    this->params = {{{1, 1, 1, 1}, {1, 1}, {1, 1}, {1, 1}, {0, 0}, {0, 0}},
+                    {{2, 16, 64, 64}, {64, 64}, {1, 1}, {1, 1}, {0, 0}, {0, 0}},
+                    {{2, 16, 64, 64}, {4, 4}, {4, 4}, {2, 2}, {0, 0}, {0, 0}},
+                    {{2, 32, 30, 30}, {2, 2}, {2, 2}, {1, 1}, {1, 1}, {1, 1}}};
+
+    this->Run();
+}