first commit

example/10_convnd_fwd_multiple_d_multiple_reduce/common.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <algorithm>
+#include <cassert>
+#include <cstdint>
+#include <cstdlib>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+#include <type_traits>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_multiple_r_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/fill.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
+
+using BF16 = ck::bhalf_t;
+using FP16 = ck::half_t;
+using FP32 = float;
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+using I4 = ck::int4_t;
+#endif
+using I8  = std::int8_t;
+using I32 = std::int32_t;
+
+template <typename ALay, typename BLay, typename DELay, typename RLay>
+struct LayoutSetting
+{
+    using ALayout  = ALay;
+    using BLayout  = BLay;
+    using DELayout = DELay;
+    using RLayout  = RLay;
+};
+
+template <ck::index_t NDimSpatial>
+struct LayoutSettingSelector;
+
+namespace ctl = ck::tensor_layout::convolution;
+
+template <>
+struct LayoutSettingSelector<1> final : LayoutSetting<ctl::GNWC, ctl::GKXC, ctl::GNWK, ctl::GNW>
+{
+};
+
+template <>
+struct LayoutSettingSelector<2> final : LayoutSetting<ctl::GNHWC, ctl::GKYXC, ctl::GNHWK, ctl::GNHW>
+{
+};
+
+template <>
+struct LayoutSettingSelector<3> final
+    : LayoutSetting<ctl::GNDHWC, ctl::GKZYXC, ctl::GNDHWK, ctl::GNDHW>
+{
+};
+
+template <ck::index_t NDimSpatial>
+using ALayout = typename LayoutSettingSelector<NDimSpatial>::ALayout;
+
+template <ck::index_t NDimSpatial>
+using BLayout = typename LayoutSettingSelector<NDimSpatial>::BLayout;
+
+template <ck::index_t NDimSpatial>
+using DELayout = typename LayoutSettingSelector<NDimSpatial>::DELayout;
+
+template <ck::index_t NDimSpatial>
+using RLayout = typename LayoutSettingSelector<NDimSpatial>::RLayout;
+
+struct ExecutionConfig final
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+};
+
+inline void print_help_msg()
+{
+    std::cerr << "arg1: verification (0=no, 1=yes)\n"
+              << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+              << "arg3: time kernel (0=no, 1=yes)\n"
+              << ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
+}
+
+inline bool parse_cmd_args(int argc,
+                           char* argv[],
+                           ck::utils::conv::ConvParam& problem_size,
+                           ExecutionConfig& config)
+{
+    constexpr int num_execution_config_args =
+        3; // arguments for do_verification, init_method, time_kernel
+    constexpr int num_conv_param_leading_args = 5; // arguments for num_dim_spatial_, G_, N_, K_, C_
+
+    constexpr int threshold_to_catch_partial_args = 1 + num_execution_config_args;
+    constexpr int threshold_to_catch_all_args =
+        threshold_to_catch_partial_args + num_conv_param_leading_args;
+
+    if(argc == 1)
+    {
+        // use default
+    }
+    // catch only ExecutionConfig arguments
+    else if(argc == threshold_to_catch_partial_args)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+    }
+    // catch both ExecutionConfig & ConvParam arguments
+    else if(threshold_to_catch_all_args < argc && ((argc - threshold_to_catch_all_args) % 3 == 0))
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+
+        const ck::index_t num_dim_spatial = std::stoi(argv[4]);
+        problem_size                      = ck::utils::conv::parse_conv_param(
+            num_dim_spatial, threshold_to_catch_partial_args, argv);
+    }
+    else
+    {
+        print_help_msg();
+        return false;
+    }
+
+    return true;
+}
+
+inline HostTensorDescriptor
+make_r0_host_tensor_descriptor(const ck::utils::conv::ConvParam& problem_size)
+{
+    std::vector<ck::index_t> dimensions{problem_size.G_, problem_size.N_};
+
+    ck::ranges::copy(problem_size.output_spatial_lengths_, std::back_inserter(dimensions));
+
+    return HostTensorDescriptor(dimensions);
+}
+
+template <typename Lengths, typename Strides>
+void unpack_host_tensor_descriptor(const HostTensorDescriptor& descriptor,
+                                   Lengths& lengths,
+                                   Strides& strides)
+{
+    assert(size(descriptor.GetLengths()) == size(lengths));
+    std::copy_n(begin(descriptor.GetLengths()), size(descriptor.GetLengths()), begin(lengths));
+
+    assert(size(descriptor.GetStrides()) == size(strides));
+    std::copy_n(begin(descriptor.GetStrides()), size(descriptor.GetStrides()), begin(strides));
+}

example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_bf16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+using ADataType         = BF16;
+using BDataType         = BF16;
+using AccDataType       = FP32;
+using CShuffleDataType  = FP32;
+using DsDataType        = ck::Tuple<>;
+using EDataType         = BF16;
+using ReduceAccDataType = FP32;
+using R0DataType        = FP32;
+using RsDataType        = ck::Tuple<R0DataType>;
+
+#include "run_convnd_fwd_max_example.inc"
+
+int main(int argc, char* argv[]) { return !run_convnd_fwd_max_example(argc, argv); }

example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+using ADataType         = FP16;
+using BDataType         = FP16;
+using AccDataType       = FP32;
+using CShuffleDataType  = FP32;
+using DsDataType        = ck::Tuple<>;
+using EDataType         = FP16;
+using ReduceAccDataType = FP32;
+using R0DataType        = FP32;
+using RsDataType        = ck::Tuple<R0DataType>;
+
+#include "run_convnd_fwd_max_example.inc"
+
+int main(int argc, char* argv[]) { return !run_convnd_fwd_max_example(argc, argv); }

example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_fp32.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+using ADataType         = FP32;
+using BDataType         = FP32;
+using AccDataType       = FP32;
+using CShuffleDataType  = FP32;
+using DsDataType        = ck::Tuple<>;
+using EDataType         = FP32;
+using ReduceAccDataType = FP32;
+using R0DataType        = FP32;
+using RsDataType        = ck::Tuple<R0DataType>;
+
+#include "run_convnd_fwd_max_example.inc"
+
+int main(int argc, char* argv[]) { return !run_convnd_fwd_max_example(argc, argv); }

example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_int4.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+#error Should compile this file with ck::int4_t support
+#endif
+
+#define BUILD_INT4_EXAMPLE
+
+#include "common.hpp"
+
+using ADataType         = I4;
+using BDataType         = I4;
+using KernelADataType   = I8;
+using KernelBDataType   = I8;
+using AccDataType       = I32;
+using CShuffleDataType  = I32;
+using DsDataType        = ck::Tuple<>;
+using EDataType         = I32;
+using ReduceAccDataType = I32;
+using R0DataType        = I32;
+using RsDataType        = ck::Tuple<R0DataType>;
+
+#include "run_convnd_fwd_max_example.inc"
+
+int main(int argc, char* argv[]) { return !run_convnd_fwd_max_example(argc, argv); }

example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_int8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+using ADataType         = I8;
+using BDataType         = I8;
+using AccDataType       = I32;
+using CShuffleDataType  = I32;
+using DsDataType        = ck::Tuple<>;
+using EDataType         = I32;
+using ReduceAccDataType = I32;
+using R0DataType        = I32;
+using RsDataType        = ck::Tuple<R0DataType>;
+
+#include "run_convnd_fwd_max_example.inc"
+
+int main(int argc, char* argv[]) { return !run_convnd_fwd_max_example(argc, argv); }

example/10_convnd_fwd_multiple_d_multiple_reduce/run_convnd_fwd_max_example.inc

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+using QsElementOp  = ck::Tuple<PassThrough>;
+using RsElementOp  = ck::Tuple<PassThrough>;
+
+// ReduceOp
+using RsThreadReduceOp = ck::Tuple<ck::reduce::Max>;
+
+using RsGlobalReduceOp =
+    ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicMax>;
+
+static constexpr auto ConvSpec =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+template <ck::index_t NDimSpatial>
+using DeviceInstance =
+    ck::tensor_operation::device::DeviceGroupedConvFwdMultipleDMultipleR_Xdl_CShuffle
+//######| NDimSpatial|              ALayout|              BLayout|              DELayout|              RLayout|           AData|           BData|     AccData|         CShuffle|     DsData|     EData|     ReduceAccData|     RsData|           A|           B|          CDE|          Qs|          Rs|           Thread|           Global|           Conv|          GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|    CDRThreadTransfer|                  CDE|    RThreadTransfer|
+//######|            |                     |                     |                      |                     |            Type|            Type|        Type|         DataType|       Type|      Type|              Type|       Type| Elementwise| Elementwise|  Elementwise| Elementwise| Elementwise|           Reduce|           Reduce|            Fwd|Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|       ClusterLengths| ReduceThreadTransfer| DstScalarPerVector|
+//######|            |                     |                     |                      |                     |                |                |            |                 |           |          |                  |           |   Operation|   Operation|    Operation|   Operation|   Operation|        Operation|        Operation| Specialization|              |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _MPerBlock_NPerBlock|      ScalarPerVector|         _MPerBlock|
+//######|            |                     |                     |                      |                     |                |                |            |                 |           |          |                  |           |            |            |             |            |            |                 |                 |               |              |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                     |           _NPerBlock|                   |
+#ifdef BUILD_INT4_EXAMPLE
+        < NDimSpatial, ALayout<NDimSpatial>, BLayout<NDimSpatial>, DELayout<NDimSpatial>, RLayout<NDimSpatial>, KernelADataType, KernelBDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, ReduceAccDataType, RsDataType,  AElementOp,  BElementOp, CDEElementOp, QsElementOp, RsElementOp, RsThreadReduceOp, RsGlobalReduceOp,       ConvSpec,   GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<64, 4>,                    4,                  1>;
+#else
+        < NDimSpatial, ALayout<NDimSpatial>, BLayout<NDimSpatial>, DELayout<NDimSpatial>, RLayout<NDimSpatial>,       ADataType,       BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, ReduceAccDataType, RsDataType,  AElementOp,  BElementOp, CDEElementOp, QsElementOp, RsElementOp, RsThreadReduceOp, RsGlobalReduceOp,       ConvSpec,   GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<64, 4>,                    4,                  1>;
+#endif
+
+template <ck::index_t NDimSpatial>
+using HostInstance = ck::tensor_operation::host::ReferenceConvFwd
+        <NDimSpatial, ADataType, BDataType, EDataType, AElementOp, BElementOp, PassThrough>;
+// clang-format on
+
+template <ck::index_t NDimSpatial>
+bool run_convnd_fwd_max(const ck::utils::conv::ConvParam& problem_size,
+                        const ExecutionConfig& config)
+{
+    static_assert(1 <= NDimSpatial && NDimSpatial <= 3, "Unsupported NDimSpatial");
+
+#if defined(BUILD_INT4_EXAMPLE) && defined(CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4)
+    static_assert(sizeof(ck::int4_t) == sizeof(int8_t));
+#endif
+
+    const auto conv_input_g_n_c_wis_desc =
+        ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<ALayout<NDimSpatial>>(
+            problem_size);
+
+    const auto conv_weight_g_k_c_xs_desc =
+        ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<BLayout<NDimSpatial>>(
+            problem_size);
+
+    const auto conv_output_g_n_k_wos_desc =
+        ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<DELayout<NDimSpatial>>(
+            problem_size);
+
+    const auto r0_desc = make_r0_host_tensor_descriptor(problem_size);
+
+    Tensor<ADataType> conv_input(conv_input_g_n_c_wis_desc);
+    Tensor<BDataType> conv_weight(conv_weight_g_k_c_xs_desc);
+    Tensor<EDataType> conv_output_device(conv_output_g_n_k_wos_desc);
+    Tensor<R0DataType> r0_device(r0_desc);
+
+    switch(config.init_method)
+    {
+    case 0: break;
+    case 1:
+        ck::utils::FillUniformDistributionIntegerValue<ADataType>{-8, 7}(conv_input);
+        ck::utils::FillUniformDistributionIntegerValue<BDataType>{-8, 7}(conv_weight);
+        break;
+    default:
+        ck::utils::FillUniformDistribution<ADataType>{-5, 5}(conv_input);
+        ck::utils::FillUniformDistribution<BDataType>{-5, 5}(conv_weight);
+    }
+
+    DeviceMem conv_input_device_buf(sizeof(ADataType) * conv_input.mDesc.GetElementSpaceSize());
+    DeviceMem conv_weight_device_buf(sizeof(BDataType) * conv_weight.mDesc.GetElementSpaceSize());
+    DeviceMem conv_output_device_buf(sizeof(EDataType) *
+                                     conv_output_device.mDesc.GetElementSpaceSize());
+    DeviceMem r0_device_buf(sizeof(R0DataType) * r0_device.mDesc.GetElementSpaceSize());
+
+#ifdef BUILD_INT4_EXAMPLE
+    const Tensor<KernelADataType> conv_input_converted(conv_input);
+    const Tensor<KernelBDataType> conv_weight_converted(conv_weight);
+
+    conv_input_device_buf.ToDevice(conv_input_converted.mData.data());
+    conv_weight_device_buf.ToDevice(conv_weight_converted.mData.data());
+#else
+    conv_input_device_buf.ToDevice(conv_input.mData.data());
+    conv_weight_device_buf.ToDevice(conv_weight.mData.data());
+#endif
+
+    std::array<ck::index_t, NDimSpatial + 3> conv_input_g_n_c_wis_lengths{},
+        conv_input_g_n_c_wis_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> conv_weight_g_k_c_xs_lengths{},
+        conv_weight_g_k_c_xs_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> conv_output_g_n_k_wos_lengths{},
+        conv_output_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial + 2> r0_lengths{}, r0_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_strides{}, conv_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input_left_pads{}, input_right_pads{};
+
+    unpack_host_tensor_descriptor(
+        conv_input_g_n_c_wis_desc, conv_input_g_n_c_wis_lengths, conv_input_g_n_c_wis_strides);
+    unpack_host_tensor_descriptor(
+        conv_weight_g_k_c_xs_desc, conv_weight_g_k_c_xs_lengths, conv_weight_g_k_c_xs_strides);
+    unpack_host_tensor_descriptor(
+        conv_output_g_n_k_wos_desc, conv_output_g_n_k_wos_lengths, conv_output_g_n_k_wos_strides);
+    unpack_host_tensor_descriptor(r0_desc, r0_lengths, r0_strides);
+
+    ck::ranges::copy(problem_size.conv_filter_strides_, begin(conv_filter_strides));
+    ck::ranges::copy(problem_size.conv_filter_dilations_, begin(conv_filter_dilations));
+    ck::ranges::copy(problem_size.input_left_pads_, begin(input_left_pads));
+    ck::ranges::copy(problem_size.input_right_pads_, begin(input_right_pads));
+
+    // run Conv + Reduction on device
+    auto conv     = DeviceInstance<NDimSpatial>{};
+    auto invoker  = conv.MakeInvoker();
+    auto argument = conv.MakeArgument(conv_input_device_buf.GetDeviceBuffer(),
+                                      conv_weight_device_buf.GetDeviceBuffer(),
+                                      std::array<const void*, 0>{},
+                                      conv_output_device_buf.GetDeviceBuffer(),
+                                      {r0_device_buf.GetDeviceBuffer()},
+                                      conv_input_g_n_c_wis_lengths,
+                                      conv_input_g_n_c_wis_strides,
+                                      conv_weight_g_k_c_xs_lengths,
+                                      conv_weight_g_k_c_xs_strides,
+                                      std::array<std::array<ck::index_t, NDimSpatial + 3>, 0>{{}},
+                                      std::array<std::array<ck::index_t, NDimSpatial + 3>, 0>{{}},
+                                      conv_output_g_n_k_wos_lengths,
+                                      conv_output_g_n_k_wos_strides,
+                                      r0_lengths,
+                                      r0_strides,
+                                      conv_filter_strides,
+                                      conv_filter_dilations,
+                                      input_left_pads,
+                                      input_right_pads,
+                                      AElementOp{},
+                                      BElementOp{},
+                                      CDEElementOp{},
+                                      QsElementOp{},
+                                      RsElementOp{});
+
+    if(!conv.IsSupportedArgument(argument))
+    {
+        std::cerr << "wrong! device_conv with the specified compilation parameters does "
+                     "not support this Conv problem"
+                  << std::endl;
+        return false;
+    }
+
+    const float avg_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
+
+    const std::size_t flop      = problem_size.GetFlops();
+    const std::size_t num_btype = problem_size.GetByte<ADataType, BDataType, EDataType>();
+
+    const float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+    const float gb_per_sec = num_btype / 1.E6 / avg_time;
+    std::cout << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << conv.GetTypeString() << std::endl;
+
+    if(config.do_verification)
+    {
+        Tensor<EDataType> conv_output_host(conv_output_g_n_k_wos_desc);
+
+        // run Conv + Reduction on host
+        auto ref_conv     = HostInstance<NDimSpatial>{};
+        auto ref_invoker  = ref_conv.MakeInvoker();
+        auto ref_argument = ref_conv.MakeArgument(conv_input,
+                                                  conv_weight,
+                                                  conv_output_host,
+                                                  problem_size.conv_filter_strides_,
+                                                  problem_size.conv_filter_dilations_,
+                                                  problem_size.input_left_pads_,
+                                                  problem_size.input_right_pads_,
+                                                  AElementOp{},
+                                                  BElementOp{},
+                                                  PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        Tensor<R0DataType> r0_host(r0_device.mDesc);
+
+        auto reduce0_op = RsThreadReduceOp{}[ck::Number<0>{}];
+
+        auto& output_dims = conv_output_g_n_k_wos_desc.GetLengths();
+
+        if constexpr(NDimSpatial == 1)
+        {
+            for(std::size_t g = 0; g < output_dims[0]; ++g)
+            {
+                for(std::size_t n = 0; n < output_dims[1]; ++n)
+                {
+                    for(std::size_t w = 0; w < output_dims[3]; ++w)
+                    {
+                        auto reduce0_acc = reduce0_op.GetIdentityValue<ReduceAccDataType>();
+                        for(std::size_t k = 0; k < output_dims[2]; ++k)
+                        {
+
+                            auto e_val =
+                                ck::type_convert<ReduceAccDataType>(conv_output_host(g, n, k, w));
+                            reduce0_op(reduce0_acc, e_val);
+                        }
+                        r0_host(g, n, w) = ck::type_convert<R0DataType>(reduce0_acc);
+                    }
+                }
+            }
+        }
+        else if constexpr(NDimSpatial == 2)
+        {
+            for(std::size_t g = 0; g < output_dims[0]; ++g)
+            {
+                for(std::size_t n = 0; n < output_dims[1]; ++n)
+                {
+                    for(std::size_t h = 0; h < output_dims[3]; ++h)
+                    {
+                        for(std::size_t w = 0; w < output_dims[4]; ++w)
+                        {
+                            auto reduce0_acc = reduce0_op.GetIdentityValue<ReduceAccDataType>();
+                            for(std::size_t k = 0; k < output_dims[2]; ++k)
+                            {
+
+                                auto e_val = ck::type_convert<ReduceAccDataType>(
+                                    conv_output_host(g, n, k, h, w));
+                                reduce0_op(reduce0_acc, e_val);
+                            }
+                            r0_host(g, n, h, w) = ck::type_convert<R0DataType>(reduce0_acc);
+                        }
+                    }
+                }
+            }
+        }
+        else if constexpr(NDimSpatial == 3)
+        {
+            for(std::size_t g = 0; g < output_dims[0]; ++g)
+            {
+                for(std::size_t n = 0; n < output_dims[1]; ++n)
+                {
+                    for(std::size_t d = 0; d < output_dims[3]; ++d)
+                    {
+                        for(std::size_t h = 0; h < output_dims[4]; ++h)
+                        {
+                            for(std::size_t w = 0; w < output_dims[5]; ++w)
+                            {
+                                auto reduce0_acc = reduce0_op.GetIdentityValue<ReduceAccDataType>();
+                                for(std::size_t k = 0; k < output_dims[2]; ++k)
+                                {
+
+                                    auto e_val = ck::type_convert<ReduceAccDataType>(
+                                        conv_output_host(g, n, k, d, h, w));
+                                    reduce0_op(reduce0_acc, e_val);
+                                }
+                                r0_host(g, n, d, h, w) = ck::type_convert<R0DataType>(reduce0_acc);
+                            }
+                        }
+                    }
+                }
+            }
+        }
+
+        conv_output_device_buf.FromDevice(conv_output_device.mData.data());
+        r0_device_buf.FromDevice(r0_device.mData.data());
+
+        return ck::utils::check_err(conv_output_device,
+                                    conv_output_host,
+                                    "Error: incorrect results! (Matrix E)",
+                                    1e-5f,
+                                    1e-4f) &&
+               ck::utils::check_err(
+                   r0_device, r0_host, "Error: incorrect results! (Matrix R0)", 1e-5f, 1e-4f);
+    }
+
+    return true;
+}
+
+bool run_convnd_fwd_max_example(int argc, char* argv[])
+{
+    ck::utils::conv::ConvParam problem_size{
+        2, 1, 128, 256, 192, {3, 3}, {71, 71}, {2, 2}, {1, 1}, {1, 1}, {1, 1}};
+    ExecutionConfig config;
+
+    if(!parse_cmd_args(argc, argv, problem_size, config))
+    {
+        return false;
+    }
+
+    switch(problem_size.num_dim_spatial_)
+    {
+    case 1: return run_convnd_fwd_max<1>(problem_size, config);
+    case 2: return run_convnd_fwd_max<2>(problem_size, config);
+    case 3: return run_convnd_fwd_max<3>(problem_size, config);
+    }
+
+    return false;
+}

example/12_reduce/CMakeLists.txt

+add_example_executable(example_reduce_blockwise reduce_blockwise.cpp)
+add_example_executable(example_reduce_multiblock_atomic_add reduce_multiblock_atomic_add.cpp)
+add_example_executable(example_reduce_blockwise_two_call reduce_blockwise_two_call.cpp)

example/12_reduce/README.md

+# Instructions for ```example_reduce_blockwise```
+
+## Run ```example_reduce_blockwise```
+```bash
+# -D <xxx> : input 3d/4d/5d tensor lengths
+# -R <xxx> : reduce dimension ids
+# -v <x> :   verification (0=no, 1=yes)
+#arg1: data type (0: fp16, 1: fp32, 3: int8, 5: bp16, 6: fp64, 7: int4)
+#arg2: initialization (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
+#arg3: time kernel (0=no, 1=yes)
+./bin/example_reduce_blockwise -D 16,64,32,960 -v 1 0 2 1
+```
+
+Result
+```
+./bin/example_reduce_blockwise -D 16,64,32,960 -v 1 0 2 1
+launch_and_time_kernel: grid_dim {240, 1, 1}, block_dim {256, 1, 1} 
+Warm up 1 time
+Start running 10 times...
+Perf: 0.238063 ms, 264.285 GB/s, DeviceReduceBlockWise<256,M_C4_S1,K_C64_S1,InSrcVectorDim_0_InSrcVectorSize_1_OutDstVectorSize_1>
+```
+
+## Run ```example_reduce_multiblock_atomic_add```
+```bash
+# -D <xxx> : input 3d/4d/5d tensor lengths
+# -R <xxx> : reduce dimension ids
+# -v <x> :   verification (0=no, 1=yes)
+#arg1: data type (0: fp32, 1: fp64)
+#arg2: initialization (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
+#arg3: time kernel (0=no, 1=yes)
+./bin/example_reduce_multiblock_atomic_add -D 16,64,32,960 -v 1 0 2 0
+```
+
+Result
+```
+./bin/example_reduce_multiblock_atomic_add -D 16,64,32,960 -v 1 0 2 0
+Perf: 0 ms, inf GB/s, DeviceReduceMultiBlock<256,M_C4_S1,K_C64_S1,InSrcVectorDim_0_InSrcVectorSize_1_OutDstVectorSize_1>
+echo $?
+0
+```
+
+# Instructions for ```example_reduce_blockwise_two_call```
+
+## Run ```example_reduce_blockwise_two_call```
+```bash
+#arg1:  verification (0=no, 1=yes(
+#arg2:  initialization (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
+#arg3:  time kernel (0=no, 1=yes)
+./bin/example_reduce_blockwise_two_call 1 2 1
+```
+
+Result
+```
+./bin/example_reduce_blockwise_two_call 1 2 1
+launch_and_time_kernel: grid_dim {204800, 1, 1}, block_dim {256, 1, 1}
+Warm up 1 time
+Start running 10 times...
+launch_and_time_kernel: grid_dim {6400, 1, 1}, block_dim {256, 1, 1}
+Warm up 1 time
+Start running 10 times...
+Perf: 2.1791 ms, 771.42 GB/s, DeviceReduceBlockWise<256,M_C32_S1,K_C8_S1,InSrcVectorDim_1_InSrcVectorSize_1_OutDstVectorSize_1> => DeviceReduceBlockWise<256,M_C256_S1,K_C1_S1,InSrcVectorDim_1_InSrcVectorSize_1_OutDstVectorSize_1>
+```

example/12_reduce/reduce_blockwise.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <initializer_list>
+#include <cstdlib>
+#include <getopt.h>
+
+#include "ck/utility/reduction_enums.hpp"
+#include "reduce_blockwise_impl.hpp"
+#include "reduce_example_common.hpp"
+
+using namespace ck;
+using namespace ck::tensor_operation::device;
+
+static struct option long_options[] = {{"inLengths", required_argument, nullptr, 'D'},
+                                       {"verify", required_argument, nullptr, 'v'},
+                                       {"help", no_argument, nullptr, '?'},
+                                       {nullptr, 0, nullptr, 0}};
+
+class SimpleAppArgs
+{
+    private:
+    int option_index = 0;
+
+    public:
+    std::vector<size_t> inLengths = {16, 64, 32, 960};
+    std::vector<int> reduceDims   = {0, 1, 2};
+    std::vector<float> scales     = {1.0f, 0.0f};
+
+    bool do_verification = true;
+    int data_type        = 1;
+    int init_method      = 2;
+    bool time_kernel     = true;
+
+    public:
+    void show_usage(const char* cmd)
+    {
+        std::cout << "Usage of " << cmd << std::endl;
+        std::cout << "--inLengths or -D, comma separated list of input tensor dimension lengths"
+                  << std::endl;
+        std::cout << "--reduceDims or -R, comma separated list of to-reduce dimensions"
+                  << std::endl;
+        std::cout << "--verify or -v, 1/0 to indicate whether to verify the reduction result by "
+                     "comparing with the host-based reduction"
+                  << std::endl;
+        std::cout << "Arg1: data type (0: fp16, 1: fp32, 3: int8, 5: bp16, 6: fp64, 7: int4)"
+                  << std::endl;
+        std::cout << "Arg2 -- init method (0=no init, 1=single integer value, 2=scope integer "
+                     "value, 3=decimal value)"
+                  << std::endl;
+        std::cout << "Arg3 -- time kernel (0=no, 1=yes)" << std::endl;
+    };
+
+    int processArgs(int argc, char* argv[])
+    {
+        using ck::host_common::getTypeValuesFromString;
+
+        int ch;
+
+        while(1)
+        {
+            ch = getopt_long(argc, argv, "D:R:v:l:", long_options, &option_index);
+            if(ch == -1)
+                break;
+            switch(ch)
+            {
+            case 'D':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                inLengths = getTypeValuesFromString<size_t>(optarg);
+                break;
+            case 'R':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                reduceDims = getTypeValuesFromString<int>(optarg);
+                break;
+            case 'v':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                do_verification = static_cast<bool>(std::atoi(optarg));
+                break;
+            case '?':
+                if(std::string(long_options[option_index].name) == "help")
+                {
+                    show_usage(argv[0]);
+                    return (-1);
+                };
+                break;
+            default: show_usage(argv[0]); return (-1);
+            };
+        };
+
+        if(optind + 3 > argc)
+        {
+            throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
+        };
+
+        data_type   = std::atoi(argv[optind++]);
+        init_method = std::atoi(argv[optind++]);
+        time_kernel = static_cast<bool>(std::atoi(argv[optind]));
+
+        if(scales.empty())
+        {
+            scales.push_back(1.0f);
+            scales.push_back(0.0f);
+        };
+
+        return (0);
+    };
+};
+
+template <typename InOutDataType,
+          typename AccDataType,
+          ReduceTensorOp ReduceOpId,
+          index_t PropagateNan,
+          index_t OutputIndex>
+bool reduce_blockwise_test(bool do_verification,
+                           int init_method,
+                           bool time_kernel,
+                           const std::vector<size_t>& inLengths,
+                           const std::vector<int>& reduceDims,
+                           float alpha,
+                           float beta)
+{
+    bool matched = false;
+    int result   = 0;
+
+    const auto tuple_object = reduce_shape_instances{};
+
+    static_for<0, std::tuple_size<reduce_shape_instances>::value, 1>{}([&](auto i) {
+        if(matched)
+            return;
+
+        using ShapeType = remove_cvref_t<decltype(std::get<i>(tuple_object))>;
+
+        if(ShapeType::Rank_ != inLengths.size() || ShapeType::NumReduceDim_ != reduceDims.size())
+            return;
+
+        std::array<int, ShapeType::NumReduceDim_> arrReduceDims;
+
+        ck::ranges::copy(reduceDims, arrReduceDims.begin());
+
+        result = reduce_blockwise_impl<InOutDataType,
+                                       AccDataType,
+                                       ReduceOpId,
+                                       ShapeType::Rank_,
+                                       ShapeType::NumReduceDim_,
+                                       PropagateNan,
+                                       OutputIndex>(
+            do_verification, init_method, time_kernel, inLengths, arrReduceDims, alpha, beta);
+
+        matched = true;
+    });
+
+    return (result == 0) ? true : false;
+};
+
+constexpr ReduceTensorOp ReduceOpId = ReduceTensorOp::AVG;
+constexpr bool PropagateNan         = true;
+constexpr bool OutputIndex          = false;
+
+int main(int argc, char* argv[])
+{
+    bool pass = true;
+
+    if(argc > 1)
+    {
+        SimpleAppArgs arg;
+
+        if(arg.processArgs(argc, argv) < 0)
+            return (-1);
+
+        if(arg.data_type == 0)
+        {
+            pass = reduce_blockwise_test<ck::half_t, float, ReduceOpId, PropagateNan, OutputIndex>(
+                arg.do_verification,
+                arg.init_method,
+                arg.time_kernel,
+                arg.inLengths,
+                arg.reduceDims,
+                arg.scales[0],
+                arg.scales[1]);
+        }
+        else if(arg.data_type == 1)
+        {
+            pass = reduce_blockwise_test<float, float, ReduceOpId, PropagateNan, OutputIndex>(
+                arg.do_verification,
+                arg.init_method,
+                arg.time_kernel,
+                arg.inLengths,
+                arg.reduceDims,
+                arg.scales[0],
+                arg.scales[1]);
+        }
+        else if(arg.data_type == 3)
+        {
+            pass = reduce_blockwise_test<int8_t, float, ReduceOpId, PropagateNan, OutputIndex>(
+                arg.do_verification,
+                arg.init_method,
+                arg.time_kernel,
+                arg.inLengths,
+                arg.reduceDims,
+                arg.scales[0],
+                arg.scales[1]);
+        }
+        else if(arg.data_type == 5)
+        {
+            pass = reduce_blockwise_test<ck::bhalf_t, float, ReduceOpId, PropagateNan, OutputIndex>(
+                arg.do_verification,
+                arg.init_method,
+                arg.time_kernel,
+                arg.inLengths,
+                arg.reduceDims,
+                arg.scales[0],
+                arg.scales[1]);
+        }
+        else if(arg.data_type == 6)
+        {
+            pass = reduce_blockwise_test<double, double, ReduceOpId, PropagateNan, OutputIndex>(
+                arg.do_verification,
+                arg.init_method,
+                arg.time_kernel,
+                arg.inLengths,
+                arg.reduceDims,
+                arg.scales[0],
+                arg.scales[1]);
+        }
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+        else if(arg.data_type == 7)
+        {
+            pass = reduce_blockwise_test<int4_t, int32_t, ReduceTensorOp::AVG, false, false>(
+                arg.do_verification,
+                arg.init_method,
+                arg.time_kernel,
+                arg.inLengths,
+                arg.reduceDims,
+                arg.scales[0],
+                arg.scales[1]);
+
+            pass = pass && reduce_blockwise_test<int4_t, int8_t, ReduceTensorOp::MAX, false, false>(
+                               arg.do_verification,
+                               arg.init_method,
+                               arg.time_kernel,
+                               arg.inLengths,
+                               arg.reduceDims,
+                               arg.scales[0],
+                               arg.scales[1]);
+        }
+#endif
+    }
+    else
+    {
+        // for testing half_t
+        pass =
+            pass && reduce_blockwise_test<ck::half_t, float, ReduceOpId, PropagateNan, OutputIndex>(
+                        true, 2, true, {16, 64, 32, 960}, {0, 1, 2}, 1.0f, 0.0f);
+
+        // for testing float
+        pass = pass && reduce_blockwise_test<float, float, ReduceOpId, PropagateNan, OutputIndex>(
+                           true, 2, true, {16, 64, 32, 960}, {0, 1, 2}, 1.0f, 0.0f);
+
+        // for testing double
+        pass = pass && reduce_blockwise_test<float, float, ReduceOpId, PropagateNan, OutputIndex>(
+                           true, 2, true, {16, 64, 32, 960}, {0, 1, 2}, 1.0f, 0.0f);
+
+        // for testing bhalf_t
+        pass = pass &&
+               reduce_blockwise_test<ck::bhalf_t, float, ReduceOpId, PropagateNan, OutputIndex>(
+                   true, 2, true, {16, 64, 32, 960}, {0, 1, 2}, 1.0f, 0.0f);
+
+        // for testing int8_t
+        pass =
+            pass && reduce_blockwise_test<int8_t, int32_t, ReduceOpId, PropagateNan, OutputIndex>(
+                        true, 2, true, {16, 64, 32, 960}, {0, 1, 2}, 1.0f, 0.0f);
+
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+        // for testing int4_t using AVG operation
+        pass = pass && reduce_blockwise_test<int4_t, int32_t, ReduceTensorOp::AVG, false, false>(
+                           true, 2, true, {16, 64, 32, 960}, {0, 1, 2}, 1.0f, 0.0f);
+
+        // for testing int4_t using MAX operation
+        pass = pass && reduce_blockwise_test<int4_t, int8_t, ReduceTensorOp::MAX, false, false>(
+                           true, 2, true, {16, 64, 32, 960}, {0, 1, 2}, 1.0f, 0.0f);
+#endif
+        // for testing 3D input
+        pass = pass && reduce_blockwise_test<float, float, ReduceOpId, PropagateNan, OutputIndex>(
+                           true, 2, true, {16, 64, 960}, {0, 1}, 1.0f, 0.0f);
+
+        // for testing 5D input
+        pass = pass && reduce_blockwise_test<float, float, ReduceOpId, PropagateNan, OutputIndex>(
+                           true, 2, true, {16, 64, 32, 2, 960}, {0, 1, 2, 3}, 1.0f, 0.0f);
+    };
+
+    return (pass ? 0 : 1);
+};

example/12_reduce/reduce_blockwise_impl.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_reduce_multiblock.hpp"
+
+#include "ck/library/utility/algorithm.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/host_common_util.hpp"
+#include "ck/library/utility/host_reduction.hpp"
+
+#include "reduce_example_common.hpp"
+
+template <typename InOutDataType,
+          typename AccDataType,
+          ck::ReduceTensorOp ReduceOpId,
+          ck::index_t Rank,
+          ck::index_t NumReduceDim,
+          bool PropagateNan,
+          bool OutputIndex>
+int reduce_blockwise_impl(bool do_verification,
+                          int init_method,
+                          bool time_kernel,
+                          const std::vector<size_t>& inLengths,
+                          const std::array<int, NumReduceDim>& reduceDims,
+                          float alpha,
+                          float beta)
+
+{
+    using namespace ck;
+    using namespace ck::tensor_operation::device;
+
+    constexpr index_t NumOutDim = (Rank - NumReduceDim == 0) ? 1 : Rank - NumReduceDim;
+
+    constexpr bool op_support_indices =
+        (ReduceOpId == ReduceTensorOp::MIN || ReduceOpId == ReduceTensorOp::MAX ||
+         ReduceOpId == ReduceTensorOp::AMAX);
+
+    constexpr bool invalid_reduce_1 = OutputIndex && !op_support_indices;
+
+    // 1) If InOutDataType is half_t, must use half_t as AccDataType for indexable reduction
+    // operations 2) If InOutDataType is half_t, must use float as AccDataType for non-indexable
+    // reduction operations
+    constexpr bool invalid_reduce_2 =
+        std::is_same<InOutDataType, half_t>::value &&
+        ((!op_support_indices && !std::is_same<AccDataType, float>::value) ||
+         (op_support_indices && !std::is_same<AccDataType, half_t>::value));
+
+    // 1) If InOutDataType is float, must use float as AccDataType for indexable reduction
+    // operations
+    constexpr bool invalid_reduce_3 =
+        std::is_same<InOutDataType, float>::value &&
+        (op_support_indices && !std::is_same<AccDataType, float>::value);
+
+    // 1) If InOutDataType is int8_t or int4_t, must use int8_t as AccDataType for indexable
+    // reduction operations 2) If InOutDataType is int8_t or int4_t, must use int32_t as AccDataType
+    // for non-indexable reduction operations
+    constexpr bool invalid_reduce_4 =
+        std::is_same<InOutDataType, int8_t>::value &&
+        ((!op_support_indices && !std::is_same<AccDataType, int32_t>::value) ||
+         (op_support_indices && !std::is_same<AccDataType, int8_t>::value));
+
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+    constexpr bool invalid_reduce_4_2 =
+        std::is_same<InOutDataType, int4_t>::value &&
+        ((!op_support_indices && !std::is_same<AccDataType, int32_t>::value) ||
+         (op_support_indices && !std::is_same<AccDataType, int8_t>::value));
+#endif
+
+    // 1) If InOutDataType is int8_t or int4_t, the supported operation must be either indexable
+    // operations or ADD/AVG
+    constexpr bool invalid_reduce_5 = std::is_same<InOutDataType, int8_t>::value &&
+                                      (!op_support_indices && ReduceOpId != ReduceTensorOp::ADD &&
+                                       ReduceOpId != ReduceTensorOp::AVG);
+
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+    constexpr bool invalid_reduce_5_2 = std::is_same<InOutDataType, int4_t>::value &&
+                                        (!op_support_indices && ReduceOpId != ReduceTensorOp::ADD &&
+                                         ReduceOpId != ReduceTensorOp::AVG);
+#endif
+
+    // 1) If InOutDataType is bhalf_t, must use float as AccDataType for all reduction operations
+    constexpr bool invalid_reduce_6 =
+        std::is_same<InOutDataType, bhalf_t>::value && !std::is_same<AccDataType, float>::value;
+
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+    constexpr bool invalid_reduce =
+        (invalid_reduce_1 || invalid_reduce_2 || invalid_reduce_3 || invalid_reduce_4 ||
+         invalid_reduce_5 || invalid_reduce_6 || invalid_reduce_4_2 || invalid_reduce_5_2);
+#else
+    constexpr bool invalid_reduce = (invalid_reduce_1 || invalid_reduce_2 || invalid_reduce_3 ||
+                                     invalid_reduce_4 || invalid_reduce_5 || invalid_reduce_6);
+#endif
+
+    if constexpr(invalid_reduce)
+    {
+        std::cerr << "The reduction setting is invalid, exiting!" << std::endl;
+        return (-1);
+    };
+
+    using ReduceOperation = typename reduce_binary_operator<ReduceOpId>::opType;
+    using InElementwiseOperation =
+        typename reduce_unary_operator<ReduceOpId, true, true>::InElementwiseOperation;
+    using AccElementwiseOperation =
+        typename reduce_unary_operator<ReduceOpId, true, true>::AccElementwiseOperation;
+
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+    using InOutDataTypeInDevice = typename std::
+        conditional<std::is_same<InOutDataType, int4_t>::value, int8_t, InOutDataType>::type;
+#else
+    using InOutDataTypeInDevice   = InOutDataType;
+#endif
+
+    using DeviceReduceInstance =
+        ck::tensor_operation::device::DeviceReduceMultiBlock<InOutDataTypeInDevice,
+                                                             AccDataType,
+                                                             InOutDataTypeInDevice,
+                                                             Rank,
+                                                             NumReduceDim,
+                                                             ReduceOperation,
+                                                             InElementwiseOperation,
+                                                             AccElementwiseOperation,
+                                                             InMemoryDataOperationEnum::Set,
+                                                             PropagateNan,
+                                                             OutputIndex,
+                                                             false, // HaveIndexInputIfOutputIndex
+                                                             256,   // BlockSize
+                                                             4,     // MThreadClusterSize
+                                                             64,    // KThreadClusterSize
+                                                             1,     // MThreadSliceSize
+                                                             1,     // KThreadSliceSize
+                                                             0,     // InSrcVectorDim
+                                                             1,     // InSrceVectorSize
+                                                             1>;    // OutDstVectorSize
+
+    Tensor<InOutDataType> in(inLengths);
+
+    std::vector<size_t> outLengths;
+
+    auto invariantDims = get_invariant_dims<Rank, NumReduceDim>(reduceDims);
+
+    if(invariantDims.empty())
+        outLengths.push_back(1);
+    else
+        for(auto dim : invariantDims)
+            outLengths.push_back(inLengths[dim]);
+
+    Tensor<InOutDataType> out_ref(outLengths);
+    Tensor<InOutDataType> out(outLengths);
+    Tensor<int> out_indices_ref(outLengths);
+    Tensor<int> out_indices(outLengths);
+
+    auto inStrides  = in.mDesc.GetStrides();
+    auto outStrides = out.mDesc.GetStrides();
+
+    size_t invariant_total_length = out.mDesc.GetElementSize();
+    size_t reduce_total_length    = in.mDesc.GetElementSize() / invariant_total_length;
+
+    std::size_t num_thread = 1;
+
+    if(do_verification)
+    {
+        switch(init_method)
+        {
+        case 0: break;
+        case 1:
+            in.GenerateTensorValue(GeneratorTensor_1<InOutDataType>{1}, num_thread);
+            if(beta != 0.0f)
+                out_ref.GenerateTensorValue(GeneratorTensor_1<InOutDataType>{1}, num_thread);
+            break;
+        case 2:
+            in.GenerateTensorValue(GeneratorTensor_2<InOutDataType>{-5, 5}, num_thread);
+            if(beta != 0.0f)
+                out_ref.GenerateTensorValue(GeneratorTensor_2<InOutDataType>{-5, 5}, num_thread);
+            break;
+        default:
+            in.GenerateTensorValue(GeneratorTensor_3<InOutDataType>{-5.0, 5.0}, num_thread);
+            if(beta != 0.0f)
+                out_ref.GenerateTensorValue(GeneratorTensor_3<InOutDataType>{-5.0, 5.0},
+                                            num_thread);
+        }
+
+        if(beta != 0.0f)
+            for(size_t i = 0; i < out_ref.mDesc.GetElementSpaceSize(); i++)
+                out.mData[i] = out_ref.mData[i];
+    };
+
+    // these buffers are usually provided by the user application
+    DeviceMem in_dev(sizeof(InOutDataTypeInDevice) * in.mDesc.GetElementSpaceSize());
+    DeviceMem out_dev(sizeof(InOutDataTypeInDevice) * out.mDesc.GetElementSpaceSize());
+
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+    if(std::is_same<InOutDataType, int4_t>::value)
+    {
+        std::vector<InOutDataTypeInDevice> tmp_buf(in.mData.size());
+
+        std::copy_n(in.mData.data(), in.mData.size(), tmp_buf.data());
+        in_dev.ToDevice(tmp_buf.data());
+    }
+    else
+#endif
+        in_dev.ToDevice(in.mData.data());
+
+    if(beta != 0.0f)
+    {
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+        if(std::is_same<InOutDataType, int4_t>::value)
+        {
+            std::vector<InOutDataTypeInDevice> tmp_buf(in.mData.size());
+
+            std::copy_n(out.mData.data(), out.mData.size(), tmp_buf.data());
+            out_dev.ToDevice(tmp_buf.data());
+        }
+        else
+#endif
+            out_dev.ToDevice(out.mData.data());
+    };
+
+    size_t indicesSizeInBytes = OutputIndex ? out.mDesc.GetElementSize() * sizeof(int32_t) : 0;
+
+    DeviceMem out_index_dev(indicesSizeInBytes);
+
+    InElementwiseOperation in_elementwise_op;
+    AccElementwiseOperation acc_elementwise_op;
+
+    std::tie(in_elementwise_op, acc_elementwise_op) =
+        reduce_unary_operator<ReduceOpId, true, true>::GetElementwiseOperator(
+            static_cast<int32_t>(reduce_total_length));
+
+    if(do_verification)
+    {
+        ReductionHost<InOutDataType,
+                      AccDataType,
+                      InOutDataType,
+                      ReduceOperation,
+                      InElementwiseOperation,
+                      AccElementwiseOperation,
+                      Rank,
+                      NumReduceDim,
+                      PropagateNan,
+                      OutputIndex>
+            hostReduce(in.mDesc, out_ref.mDesc, invariantDims, reduceDims);
+
+        hostReduce.Run(alpha,
+                       in.mData.data(),
+                       beta,
+                       out_ref.mData.data(),
+                       out_indices_ref.mData.data(),
+                       in_elementwise_op,
+                       acc_elementwise_op);
+    };
+
+    std::array<index_t, Rank> arrInLengths;
+    std::array<index_t, Rank> arrInStrides;
+    std::array<index_t, NumOutDim> arrOutLengths;
+    std::array<index_t, NumOutDim> arrOutStrides;
+
+    ck::ranges::copy(inLengths, arrInLengths.begin());
+    ck::ranges::copy(inStrides, arrInStrides.begin());
+    ck::ranges::copy(outLengths, arrOutLengths.begin());
+    ck::ranges::copy(outStrides, arrOutStrides.begin());
+
+    auto reduce = DeviceReduceInstance{};
+
+    auto argument_ptr = reduce.MakeArgumentPointer(arrInLengths,
+                                                   arrInStrides,
+                                                   arrOutLengths,
+                                                   arrOutStrides,
+                                                   reduceDims,
+                                                   alpha,
+                                                   beta,
+                                                   in_dev.GetDeviceBuffer(),
+                                                   nullptr,
+                                                   out_dev.GetDeviceBuffer(),
+                                                   out_index_dev.GetDeviceBuffer(),
+                                                   in_elementwise_op,
+                                                   acc_elementwise_op);
+
+    if(!reduce.IsSupportedArgument(argument_ptr.get()))
+    {
+        std::cerr
+            << "The runtime parameters seems not supported by the DeviceReduce instance, exiting!"
+            << std::endl;
+
+        return (-2);
+    };
+
+    std::string reduce_name = reduce.GetTypeString();
+
+    auto invoker_ptr = reduce.MakeInvokerPointer();
+
+    float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+    std::size_t num_bytes = invariant_total_length * reduce_total_length * sizeof(InOutDataType) +
+                            invariant_total_length * sizeof(InOutDataType);
+
+    float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+    std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s, " << reduce_name
+              << std::endl;
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+        if(std::is_same<InOutDataType, int4_t>::value)
+        {
+            std::vector<InOutDataTypeInDevice> tmp_buf(out.mData.size());
+
+            out_dev.FromDevice(tmp_buf.data());
+
+            std::copy_n(tmp_buf.data(), out.mData.size(), out.mData.data());
+        }
+        else
+#endif
+            out_dev.FromDevice(out.mData.data());
+
+        pass = pass && ck::utils::check_err(out, out_ref);
+
+        if(OutputIndex)
+        {
+            out_index_dev.FromDevice(out_indices.mData.data());
+            pass = pass && ck::utils::check_err(out_indices, out_indices_ref);
+        };
+    };
+
+    return (pass ? 0 : 1);
+}

example/12_reduce/reduce_blockwise_two_call.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <sstream>
+#include <initializer_list>
+#include <cstdlib>
+#include <getopt.h>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_reduce_multiblock.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/host_common_util.hpp"
+#include "ck/library/utility/host_reduction.hpp"
+
+using namespace ck;
+using namespace ck::tensor_operation::device;
+
+using InOutDataType = ck::half_t;
+using InOutDataType = ck::half_t;
+using AccDataType   = float;
+
+constexpr ReduceTensorOp ReduceOpId = ReduceTensorOp::NORM2;
+constexpr bool PropagateNan         = true;
+constexpr bool OutputIndex          = false;
+
+using ReduceOperation = typename reduce_binary_operator<ReduceOpId>::opType;
+using InElementwiseOperation =
+    typename reduce_unary_operator<ReduceOpId, true, true>::InElementwiseOperation;
+using AccElementwiseOperation =
+    typename reduce_unary_operator<ReduceOpId, true, true>::AccElementwiseOperation;
+
+using PassThroughOp = tensor_operation::element_wise::PassThrough;
+
+using DeviceReduceInstance_1 = DeviceReduceMultiBlock<InOutDataType,
+                                                      AccDataType,
+                                                      InOutDataType,
+                                                      5, // Rank
+                                                      1, // NumReduceDim
+                                                      ReduceOperation,
+                                                      InElementwiseOperation,
+                                                      PassThroughOp,
+                                                      InMemoryDataOperationEnum::Set,
+                                                      PropagateNan,
+                                                      OutputIndex,
+                                                      false, // HaveIndexInputIfOutputIndex
+                                                      256,
+                                                      32,
+                                                      8,
+                                                      1,
+                                                      1,
+                                                      1, // vector dim
+                                                      1,
+                                                      1>;
+
+using DeviceReduceInstance_2 = DeviceReduceMultiBlock<InOutDataType,
+                                                      AccDataType,
+                                                      InOutDataType,
+                                                      4, // Rank
+                                                      1, // NumReduceDim
+                                                      ReduceOperation,
+                                                      PassThroughOp,
+                                                      AccElementwiseOperation,
+                                                      InMemoryDataOperationEnum::Set,
+                                                      PropagateNan,
+                                                      OutputIndex,
+                                                      false, // HaveIndexInputIfOutputIndex
+                                                      256,
+                                                      128,
+                                                      2,
+                                                      1,
+                                                      1,
+                                                      1, // vector dim
+                                                      1,
+                                                      1>;
+
+static bool do_verify;
+static int init_method;
+static float alpha;
+static float beta;
+static bool time_kernel;
+
+int main(int argc, char* argv[])
+{
+    // used by the device reduction
+    const std::array<int, 1> reduceDims_1 = {4};
+    // const std::array<int, 4> invariantDims_1 = {0, 1, 2, 3};
+
+    const std::array<int, 1> reduceDims_2 = {3};
+    // const std::array<int, 3> invariantDims_2 = {0, 1, 2};
+
+    // used by the host reduction
+    const std::array<int, 2> reduceDims    = {3, 4};
+    const std::array<int, 3> invariantDims = {0, 1, 2};
+
+    const std::vector<size_t> inLengths_1 = {64, 320, 80, 4, 128};
+
+    // input lengths of the second reduction, which is also the output lengths of the first
+    // reduction
+    const std::vector<size_t> inLengths_2 = {64, 320, 80, 4};
+
+    const std::vector<size_t> outLengths = {64, 320, 80};
+
+    if(argc == 1)
+    {
+        do_verify   = true;
+        init_method = 2;
+        time_kernel = true;
+    }
+    else if(argc == 4)
+    {
+        do_verify   = static_cast<bool>(argv[1]);
+        init_method = atoi(argv[2]);
+        time_kernel = static_cast<bool>(atoi(argv[3]));
+    }
+    else
+    {
+        std::ostringstream ostr;
+
+        ostr << "Wrong parameter! " << std::endl
+             << "Usage: " << argv[0] << "[verify 0/1] init_method time_kernel" << std::endl;
+
+        throw std::runtime_error(ostr.str());
+    };
+
+    alpha = 1.0f;
+    beta  = 0.0f;
+
+    Tensor<InOutDataType> in_1(inLengths_1);
+
+    Tensor<InOutDataType> out_ref(outLengths);
+    Tensor<InOutDataType> in_2(inLengths_2); // also the output tensor of the first reduction
+    Tensor<InOutDataType> out(outLengths);
+
+    auto inStrides_1 = in_1.mDesc.GetStrides();
+    auto inStrides_2 = in_2.mDesc.GetStrides();
+    auto outStrides  = out.mDesc.GetStrides();
+
+    size_t invariant_total_length = out.mDesc.GetElementSize();
+    size_t reduce_total_length    = in_1.mDesc.GetElementSize() / invariant_total_length;
+
+    std::size_t num_thread = 1;
+
+    if(do_verify)
+    {
+        switch(init_method)
+        {
+        case 0: break;
+        case 1:
+            in_1.GenerateTensorValue(GeneratorTensor_1<InOutDataType>{1}, num_thread);
+            if(beta != 0.0f)
+                out_ref.GenerateTensorValue(GeneratorTensor_1<InOutDataType>{1}, num_thread);
+            break;
+        case 2:
+            in_1.GenerateTensorValue(GeneratorTensor_2<InOutDataType>{-5, 5}, num_thread);
+            if(beta != 0.0f)
+                out_ref.GenerateTensorValue(GeneratorTensor_2<InOutDataType>{-5, 5}, num_thread);
+            break;
+        default:
+            in_1.GenerateTensorValue(GeneratorTensor_3<InOutDataType>{-5.0, 5.0}, num_thread);
+            if(beta != 0.0f)
+                out_ref.GenerateTensorValue(GeneratorTensor_3<InOutDataType>{-5.0, 5.0},
+                                            num_thread);
+        }
+
+        if(beta != 0.0f)
+            for(size_t i = 0; i < out_ref.mDesc.GetElementSpaceSize(); i++)
+                out.mData[i] = out_ref.mData[i];
+    };
+
+    DeviceMem in_1_dev(sizeof(InOutDataType) * in_1.mDesc.GetElementSpaceSize());
+    DeviceMem in_2_dev(sizeof(InOutDataType) * in_2.mDesc.GetElementSpaceSize());
+    DeviceMem out_dev(sizeof(InOutDataType) * out.mDesc.GetElementSpaceSize());
+
+    in_1_dev.ToDevice(in_1.mData.data());
+
+    if(beta != 0.0f)
+        out_dev.ToDevice(out.mData.data());
+
+    InElementwiseOperation in_elementwise_op;
+    AccElementwiseOperation acc_elementwise_op;
+
+    std::tie(in_elementwise_op, acc_elementwise_op) =
+        reduce_unary_operator<ReduceOpId, true, true>::GetElementwiseOperator(
+            static_cast<int32_t>(reduce_total_length));
+
+    if(do_verify)
+    {
+        ReductionHost<InOutDataType,
+                      AccDataType,
+                      InOutDataType,
+                      ReduceOperation,
+                      InElementwiseOperation,
+                      AccElementwiseOperation,
+                      5, // Rank
+                      2, // NumReduceDim
+                      PropagateNan,
+                      OutputIndex>
+            hostReduce(in_1.mDesc, out_ref.mDesc, invariantDims, reduceDims);
+
+        hostReduce.Run(alpha,
+                       in_1.mData.data(),
+                       beta,
+                       out_ref.mData.data(),
+                       nullptr,
+                       in_elementwise_op,
+                       acc_elementwise_op);
+    };
+
+    std::array<index_t, 5> arrInLengths_1;
+    std::array<index_t, 5> arrInStrides_1;
+    std::array<index_t, 4> arrInLengths_2;
+    std::array<index_t, 4> arrInStrides_2;
+    std::array<index_t, 3> arrOutLengths;
+    std::array<index_t, 3> arrOutStrides;
+
+    ck::ranges::copy(inLengths_1, arrInLengths_1.begin());
+    ck::ranges::copy(inStrides_1, arrInStrides_1.begin());
+    ck::ranges::copy(inLengths_2, arrInLengths_2.begin());
+    ck::ranges::copy(inStrides_2, arrInStrides_2.begin());
+    ck::ranges::copy(outLengths, arrOutLengths.begin());
+    ck::ranges::copy(outStrides, arrOutStrides.begin());
+
+    auto reduce_1 = DeviceReduceInstance_1{};
+
+    auto argument_ptr_1 = reduce_1.MakeArgumentPointer(arrInLengths_1,
+                                                       arrInStrides_1,
+                                                       arrInLengths_2,
+                                                       arrInStrides_2,
+                                                       reduceDims_1,
+                                                       1.0f,
+                                                       0.0f,
+                                                       in_1_dev.GetDeviceBuffer(),
+                                                       nullptr,
+                                                       in_2_dev.GetDeviceBuffer(),
+                                                       nullptr,
+                                                       in_elementwise_op,
+                                                       PassThroughOp{});
+
+    if(!reduce_1.IsSupportedArgument(argument_ptr_1.get()))
+    {
+        std::cout
+            << "The runtime parameters seems not supported by the DeviceReduce instance, exiting!"
+            << std::endl;
+    };
+
+    auto invoker_ptr_1 = reduce_1.MakeInvokerPointer();
+
+    auto reduce_2 = DeviceReduceInstance_2{};
+
+    auto argument_ptr_2 = reduce_2.MakeArgumentPointer(arrInLengths_2,
+                                                       arrInStrides_2,
+                                                       arrOutLengths,
+                                                       arrOutStrides,
+                                                       reduceDims_2,
+                                                       alpha,
+                                                       beta,
+                                                       in_2_dev.GetDeviceBuffer(),
+                                                       nullptr,
+                                                       out_dev.GetDeviceBuffer(),
+                                                       nullptr,
+                                                       PassThroughOp{},
+                                                       acc_elementwise_op);
+
+    if(!reduce_2.IsSupportedArgument(argument_ptr_2.get()))
+    {
+        std::cout
+            << "The runtime parameters seems not supported by the DeviceReduce instance, exiting!"
+            << std::endl;
+    };
+
+    auto invoker_ptr_2 = reduce_2.MakeInvokerPointer();
+
+    float avg_time_1 = invoker_ptr_1->Run(argument_ptr_1.get(), StreamConfig{nullptr, time_kernel});
+    float avg_time_2 = invoker_ptr_2->Run(argument_ptr_2.get(), StreamConfig{nullptr, time_kernel});
+
+    std::size_t num_bytes = invariant_total_length * reduce_total_length * sizeof(InOutDataType) +
+                            invariant_total_length * sizeof(InOutDataType);
+
+    float gb_per_sec = num_bytes / 1.E6 / (avg_time_1 + avg_time_2);
+
+    std::cout << "Perf: " << avg_time_1 + avg_time_2 << " ms, " << gb_per_sec << " GB/s, "
+              << reduce_1.GetTypeString() << " => " << reduce_2.GetTypeString() << std::endl;
+
+    bool pass = true;
+
+    if(do_verify)
+    {
+        out_dev.FromDevice(out.mData.data());
+        pass = pass && ck::utils::check_err(out, out_ref);
+    };
+
+    return (pass ? 0 : 1);
+}

example/12_reduce/reduce_example_common.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+
+template <int Rank, int NumReduceDim>
+static inline std::array<int, Rank - NumReduceDim>
+get_invariant_dims(const std::array<int, NumReduceDim>& reduceDims)
+{
+    int reduceFlag = 0;
+
+    // flag the bits for the reduceDims
+    for(int i = 0; i < NumReduceDim; i++)
+    {
+        reduceFlag |= 1 << reduceDims[i];
+    };
+
+    std::array<int, Rank - NumReduceDim> invariantDims;
+
+    // collect invariant dimensions
+    int dim = 0;
+    for(int i = 0; i < Rank; i++)
+        if((reduceFlag & (1 << i)) == 0)
+        {
+            invariantDims[dim] = i;
+            dim++;
+        };
+
+    return invariantDims;
+};
+
+template <ck::index_t Rank, ck::index_t NumReduceDim>
+struct ReduceShape
+{
+    static constexpr ck::index_t Rank_         = Rank;
+    static constexpr ck::index_t NumReduceDim_ = NumReduceDim;
+};
+
+using reduce_shape_instances = std::tuple<ReduceShape<3, 1>,
+                                          ReduceShape<3, 2>,
+                                          ReduceShape<4, 1>,
+                                          ReduceShape<4, 2>,
+                                          ReduceShape<4, 3>,
+                                          ReduceShape<5, 1>,
+                                          ReduceShape<5, 2>,
+                                          ReduceShape<5, 3>,
+                                          ReduceShape<5, 4>>;

example/12_reduce/reduce_multiblock_atomic_add.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <initializer_list>
+#include <cstdlib>
+#include <getopt.h>
+
+#include "ck/utility/reduction_enums.hpp"
+#include "reduce_multiblock_atomic_add_impl.hpp"
+#include "reduce_example_common.hpp"
+
+using namespace ck;
+using namespace ck::tensor_operation::device;
+
+static struct option long_options[] = {{"inLengths", required_argument, nullptr, 'D'},
+                                       {"verify", required_argument, nullptr, 'v'},
+                                       {"help", no_argument, nullptr, '?'},
+                                       {nullptr, 0, nullptr, 0}};
+
+class SimpleAppArgs
+{
+    private:
+    int option_index = 0;
+
+    public:
+    std::vector<size_t> inLengths = {16, 64, 32, 960};
+    std::vector<int> reduceDims   = {0, 1, 2};
+    std::vector<float> scales     = {1.0f, 0.0f};
+
+    bool do_verification = true;
+    int data_type        = 1;
+    int init_method      = 2;
+    bool time_kernel     = true;
+
+    public:
+    void show_usage(const char* cmd)
+    {
+        std::cout << "Usage of " << cmd << std::endl;
+        std::cout << "--inLengths or -D, comma separated list of input tensor dimension lengths"
+                  << std::endl;
+        std::cout << "--reduceDims or -R, comma separated list of to-reduce dimensions"
+                  << std::endl;
+        std::cout << "--verify or -v, 1/0 to indicate whether to verify the reduction result by "
+                     "comparing with the host-based reduction"
+                  << std::endl;
+        std::cout << "Arg1: data type (0: fp32, 1: fp64)" << std::endl;
+        std::cout << "Arg2 -- init method (0=no init, 1=single integer value, 2=scope integer "
+                     "value, 3=decimal value)"
+                  << std::endl;
+        std::cout << "Arg3 -- time kernel (0=no, 1=yes)" << std::endl;
+    };
+
+    int processArgs(int argc, char* argv[])
+    {
+        using ck::host_common::getTypeValuesFromString;
+
+        int ch;
+
+        while(1)
+        {
+            ch = getopt_long(argc, argv, "D:R:v:l:", long_options, &option_index);
+            if(ch == -1)
+                break;
+            switch(ch)
+            {
+            case 'D':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                inLengths = getTypeValuesFromString<size_t>(optarg);
+                break;
+            case 'R':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                reduceDims = getTypeValuesFromString<int>(optarg);
+                break;
+            case 'v':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                do_verification = static_cast<bool>(std::atoi(optarg));
+                break;
+            case '?':
+                if(std::string(long_options[option_index].name) == "help")
+                {
+                    show_usage(argv[0]);
+                    return (-1);
+                };
+                break;
+            default: show_usage(argv[0]); return (-1);
+            };
+        };
+
+        if(optind + 3 > argc)
+        {
+            throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
+        };
+
+        data_type   = std::atoi(argv[optind++]);
+        init_method = std::atoi(argv[optind++]);
+        time_kernel = static_cast<bool>(std::atoi(argv[optind]));
+
+        if(scales.empty())
+        {
+            scales.push_back(1.0f);
+            scales.push_back(0.0f);
+        };
+
+        return (0);
+    };
+};
+
+template <typename InOutDataType,
+          typename AccDataType,
+          ReduceTensorOp ReduceOpId,
+          index_t PropagateNan>
+bool reduce_multiblock_atomic_add_test(bool do_verification,
+                                       int init_method,
+                                       bool time_kernel,
+                                       const std::vector<size_t>& inLengths,
+                                       const std::vector<int>& reduceDims,
+                                       float alpha,
+                                       float beta)
+{
+    bool matched = false;
+    int result   = 0;
+
+    const auto tuple_object = reduce_shape_instances{};
+
+    static_for<0, std::tuple_size<reduce_shape_instances>::value, 1>{}([&](auto i) {
+        if(matched)
+            return;
+
+        using ShapeType = remove_cvref_t<decltype(std::get<i>(tuple_object))>;
+
+        if(ShapeType::Rank_ != inLengths.size() || ShapeType::NumReduceDim_ != reduceDims.size())
+            return;
+
+        std::array<int, ShapeType::NumReduceDim_> a_reduceDims;
+
+        ck::ranges::copy(reduceDims, a_reduceDims.begin());
+
+        result = reduce_multiblock_atomic_add_impl<InOutDataType,
+                                                   AccDataType,
+                                                   ReduceOpId,
+                                                   ShapeType::Rank_,
+                                                   ShapeType::NumReduceDim_,
+                                                   PropagateNan>(
+            do_verification, init_method, time_kernel, inLengths, a_reduceDims, alpha, beta);
+
+        matched = true;
+    });
+
+    return (result == 0) ? true : false;
+};
+
+constexpr ReduceTensorOp ReduceOpId = ReduceTensorOp::AVG;
+constexpr bool PropagateNan         = true;
+
+int main(int argc, char* argv[])
+{
+    bool pass = true;
+
+    if(argc > 1)
+    {
+        SimpleAppArgs arg;
+
+        if(arg.processArgs(argc, argv) < 0)
+            return (-1);
+
+        if(arg.data_type == 0)
+        {
+            pass = reduce_multiblock_atomic_add_test<float, float, ReduceOpId, PropagateNan>(
+                arg.do_verification,
+                arg.init_method,
+                arg.time_kernel,
+                arg.inLengths,
+                arg.reduceDims,
+                arg.scales[0],
+                arg.scales[1]);
+        }
+        else if(arg.data_type == 1)
+        {
+            pass = reduce_multiblock_atomic_add_test<double, double, ReduceOpId, PropagateNan>(
+                arg.do_verification,
+                arg.init_method,
+                arg.time_kernel,
+                arg.inLengths,
+                arg.reduceDims,
+                arg.scales[0],
+                arg.scales[1]);
+        }
+    }
+    else
+    {
+        // for testing float
+        pass = pass && reduce_multiblock_atomic_add_test<float, float, ReduceOpId, PropagateNan>(
+                           true, 2, false, {16, 64, 32, 960}, {0, 1, 2}, 1.0f, 0.0f);
+
+        // for testing double
+        pass = pass && reduce_multiblock_atomic_add_test<double, double, ReduceOpId, PropagateNan>(
+                           true, 2, false, {16, 64, 32, 960}, {0, 1, 2}, 1.0f, 0.0f);
+
+        // for testing 3D input
+        pass = pass && reduce_multiblock_atomic_add_test<float, float, ReduceOpId, PropagateNan>(
+                           true, 2, false, {16, 64, 960}, {0, 1}, 1.0f, 0.0f);
+
+        // for testing 5D input
+        pass = pass && reduce_multiblock_atomic_add_test<float, float, ReduceOpId, PropagateNan>(
+                           true, 2, false, {16, 64, 32, 2, 960}, {0, 1, 2, 3}, 1.0f, 0.0f);
+    };
+
+    return (pass ? 0 : 1);
+};

example/12_reduce/reduce_multiblock_atomic_add_impl.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_reduce_multiblock.hpp"
+
+#include "ck/library/utility/algorithm.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/host_common_util.hpp"
+#include "ck/library/utility/host_reduction.hpp"
+
+#include "reduce_example_common.hpp"
+
+template <typename InOutDataType,
+          typename AccDataType,
+          ck::ReduceTensorOp ReduceOpId,
+          ck::index_t Rank,
+          ck::index_t NumReduceDim,
+          bool PropagateNan>
+int reduce_multiblock_atomic_add_impl(bool do_verification,
+                                      int init_method,
+                                      bool time_kernel,
+                                      const std::vector<size_t>& inLengths,
+                                      const std::array<int, NumReduceDim>& reduceDims,
+                                      float alpha,
+                                      float beta)
+
+{
+    using namespace ck;
+    using namespace ck::tensor_operation::device;
+
+    constexpr index_t NumOutDim = (Rank - NumReduceDim == 0) ? 1 : Rank - NumReduceDim;
+
+    constexpr bool op_support_atomic_add =
+        (ReduceOpId == ReduceTensorOp::ADD || ReduceOpId == ReduceTensorOp::AVG);
+
+    constexpr bool invalid_reduce_1 = !op_support_atomic_add;
+    constexpr bool invalid_reduce_2 =
+        !(std::is_same<InOutDataType, float>::value || std::is_same<InOutDataType, double>::value);
+
+    constexpr bool invalid_reduce = (invalid_reduce_1 || invalid_reduce_2);
+
+    if(invalid_reduce)
+    {
+        std::cerr << "The reduction setting is invalid, exiting!" << std::endl;
+        return (-1);
+    };
+
+    using ReduceOperation = typename reduce_binary_operator<ReduceOpId>::opType;
+    using InElementwiseOperation =
+        typename reduce_unary_operator<ReduceOpId, true, true>::InElementwiseOperation;
+    using AccElementwiseOperation =
+        typename reduce_unary_operator<ReduceOpId, true, true>::AccElementwiseOperation;
+
+    using DeviceReduceInstance =
+        ck::tensor_operation::device::DeviceReduceMultiBlock<InOutDataType,
+                                                             AccDataType,
+                                                             InOutDataType,
+                                                             Rank,
+                                                             NumReduceDim,
+                                                             ReduceOperation,
+                                                             InElementwiseOperation,
+                                                             AccElementwiseOperation,
+                                                             InMemoryDataOperationEnum::AtomicAdd,
+                                                             PropagateNan,
+                                                             false,
+                                                             false, // HaveIndexInputIfOutputIndex
+                                                             256,
+                                                             4,
+                                                             64,
+                                                             1,
+                                                             1,
+                                                             0,
+                                                             1,
+                                                             1>;
+
+    Tensor<InOutDataType> in(inLengths);
+
+    std::vector<size_t> outLengths;
+
+    auto invariantDims = get_invariant_dims<Rank, NumReduceDim>(reduceDims);
+
+    if(invariantDims.empty())
+        outLengths.push_back(1);
+    else
+        for(auto dim : invariantDims)
+            outLengths.push_back(inLengths[dim]);
+
+    Tensor<InOutDataType> out_ref(outLengths);
+    Tensor<InOutDataType> out(outLengths);
+
+    auto inStrides  = in.mDesc.GetStrides();
+    auto outStrides = out.mDesc.GetStrides();
+
+    size_t invariant_total_length = out.mDesc.GetElementSize();
+    size_t reduce_total_length    = in.mDesc.GetElementSize() / invariant_total_length;
+
+    std::size_t num_thread = 1;
+
+    if(do_verification)
+    {
+        switch(init_method)
+        {
+        case 0: break;
+        case 1:
+            in.GenerateTensorValue(GeneratorTensor_1<InOutDataType>{1}, num_thread);
+            if(beta != 0.0f)
+                out_ref.GenerateTensorValue(GeneratorTensor_1<InOutDataType>{1}, num_thread);
+            break;
+        case 2:
+            in.GenerateTensorValue(GeneratorTensor_2<InOutDataType>{-5, 5}, num_thread);
+            if(beta != 0.0f)
+                out_ref.GenerateTensorValue(GeneratorTensor_2<InOutDataType>{-5, 5}, num_thread);
+            break;
+        default:
+            in.GenerateTensorValue(GeneratorTensor_3<InOutDataType>{-5.0, 5.0}, num_thread);
+            if(beta != 0.0f)
+                out_ref.GenerateTensorValue(GeneratorTensor_3<InOutDataType>{-5.0, 5.0},
+                                            num_thread);
+        }
+
+        if(beta != 0.0f)
+            for(size_t i = 0; i < out_ref.mDesc.GetElementSpaceSize(); i++)
+                out.mData[i] = out_ref.mData[i];
+    };
+
+    // these buffers are usually provided by the user application
+    DeviceMem in_dev(sizeof(InOutDataType) * in.mDesc.GetElementSpaceSize());
+    DeviceMem out_dev(sizeof(InOutDataType) * out.mDesc.GetElementSpaceSize());
+
+    in_dev.ToDevice(in.mData.data());
+
+    if(beta != 0.0f)
+        out_dev.ToDevice(out.mData.data());
+
+    InElementwiseOperation in_elementwise_op;
+    AccElementwiseOperation acc_elementwise_op;
+
+    std::tie(in_elementwise_op, acc_elementwise_op) =
+        reduce_unary_operator<ReduceOpId, true, true>::GetElementwiseOperator(
+            static_cast<int32_t>(reduce_total_length));
+
+    if(do_verification)
+    {
+        ReductionHost<InOutDataType,
+                      AccDataType,
+                      InOutDataType,
+                      ReduceOperation,
+                      InElementwiseOperation,
+                      AccElementwiseOperation,
+                      Rank,
+                      NumReduceDim,
+                      PropagateNan,
+                      false>
+            hostReduce(in.mDesc, out_ref.mDesc, invariantDims, reduceDims);
+
+        hostReduce.Run(alpha,
+                       in.mData.data(),
+                       beta,
+                       out_ref.mData.data(),
+                       nullptr,
+                       in_elementwise_op,
+                       acc_elementwise_op);
+    };
+
+    std::array<index_t, Rank> arrInLengths;
+    std::array<index_t, Rank> arrInStrides;
+    std::array<index_t, NumOutDim> arrOutLengths;
+    std::array<index_t, NumOutDim> arrOutStrides;
+
+    ck::ranges::copy(inLengths, arrInLengths.begin());
+    ck::ranges::copy(inStrides, arrInStrides.begin());
+    ck::ranges::copy(outLengths, arrOutLengths.begin());
+    ck::ranges::copy(outStrides, arrOutStrides.begin());
+
+    auto reduce = DeviceReduceInstance{};
+
+    auto argument_ptr = reduce.MakeArgumentPointer(arrInLengths,
+                                                   arrInStrides,
+                                                   arrOutLengths,
+                                                   arrOutStrides,
+                                                   reduceDims,
+                                                   alpha,
+                                                   beta,
+                                                   in_dev.GetDeviceBuffer(),
+                                                   nullptr,
+                                                   out_dev.GetDeviceBuffer(),
+                                                   nullptr,
+                                                   in_elementwise_op,
+                                                   acc_elementwise_op);
+
+    if(!reduce.IsSupportedArgument(argument_ptr.get()))
+    {
+        std::cerr
+            << "The runtime parameters seems not supported by the DeviceReduce instance, exiting!"
+            << std::endl;
+
+        return (-2);
+    };
+
+    std::string reduce_name = reduce.GetTypeString();
+
+    auto invoker_ptr = reduce.MakeInvokerPointer();
+
+    float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+    std::size_t num_bytes = invariant_total_length * reduce_total_length * sizeof(InOutDataType) +
+                            invariant_total_length * sizeof(InOutDataType);
+
+    float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+    std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s, " << reduce_name
+              << std::endl;
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        out_dev.FromDevice(out.mData.data());
+        pass = pass && ck::utils::check_err(out, out_ref);
+    };
+
+    return (pass ? 0 : 1);
+}

example/13_pool2d_fwd/CMakeLists.txt

+add_example_executable(example_pool2d_fwd_fp16 pool2d_fwd_fp16.cpp)
+add_example_executable(example_pool2d_fwd_fp32 pool2d_fwd_fp32.cpp)
+

example/13_pool2d_fwd/README.md

+# Instructions for ```example_pool2d_fwd``` Examples
+
+## Run ```example_pool2d_fwd_fp16```
+```bash
+#arg1: verification (0=no, 1=yes)
+#arg2: initialization (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
+#arg3: time kernel (0=no, 1=yes)
+#arg4 to 15: N, C, Y, X, Hi, Wi, Sy, Sx, LeftPy, LeftPx, RightPy, RightPx
+./bin/example_pool2d_fwd_fp16 1 1 1
+```
+
+Result 
+```
+in_n_c_hi_wi: dim 4, lengths {128, 192, 71, 71}, strides {967872, 1, 13632, 192}
+out_n_c_ho_wo: dim 4, lengths {128, 192, 36, 36}, strides {248832, 1, 6912, 192}
+launch_and_time_kernel: grid_dim {124416, 1, 1}, block_dim {64, 1, 1} 
+Warm up 1 time
+Start running 10 times...
+Perf: 0.397436 ms, 1.44252 TFlops, 783.713 GB/s
+```
+
+## Run ```example_pool2d_fwd_fp32```
+```bash
+#arg1: verification (0=no, 1=yes)
+#arg2: initialization (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
+#arg3: time kernel (0=no, 1=yes)
+#arg4 to 15: N, C, Y, X, Hi, Wi, Sy, Sx, LeftPy, LeftPx, RightPy, RightPx
+./bin/example_pool2d_fwd_fp32 1 1 1
+```
+
+
+Result 
+```
+./bin/example_pool2d_fwd_fp32 1 1 1
+in_n_c_hi_wi: dim 4, lengths {128, 192, 71, 71}, strides {967872, 1, 13632, 192}
+out_n_c_ho_wo: dim 4, lengths {128, 192, 36, 36}, strides {248832, 1, 6912, 192}
+launch_and_time_kernel: grid_dim {124416, 1, 1}, block_dim {64, 1, 1} 
+Warm up 1 time
+Start running 10 times...
+Perf: 1.01823 ms, 0.563045 TFlops, 611.8 GB/s
+```

example/13_pool2d_fwd/pool2d_fwd_common.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/utility/reduction_functions_accumulate.hpp"
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_pool2d_fwd_nhwc_nhwc.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.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"
+
+template <typename InDataType,
+          typename OutDataType,
+          typename AccDataType,
+          typename IndexDataType,
+          ck::ReduceTensorOp ReduceOpId,
+          bool PropagateNan,
+          bool OutputIndex>
+static void pool_host_verify(const Tensor<InDataType>& in,
+                             Tensor<OutDataType>& out,
+                             Tensor<IndexDataType>& out_indices,
+                             const std::array<ck::index_t, 2>& window_spatial_lengths,
+                             const std::array<ck::index_t, 2>& window_strides,
+                             const std::array<ck::index_t, 2>& in_left_pads,
+                             const std::array<ck::index_t, 2>& /*in_right_pads*/)
+{
+    const int32_t reduceLength = window_spatial_lengths[0] * window_spatial_lengths[1];
+
+    using ReduceOperation = typename ck::reduce_binary_operator<ReduceOpId>::opType;
+
+    auto elementwise_ops =
+        ck::reduce_unary_operator<ReduceOpId, true, true>::GetElementwiseOperator(reduceLength);
+
+    auto in_elementwise_op  = std::get<0>(elementwise_ops);
+    auto acc_elementwise_op = std::get<1>(elementwise_ops);
+
+    if constexpr(!OutputIndex)
+    {
+        using Accumulation =
+            ck::detail::AccumulateWithNanCheck<PropagateNan, ReduceOperation, AccDataType>;
+
+        auto f_nchw = [&](auto n, auto c, auto ho, auto wo) {
+            auto accuVal = ReduceOperation::template GetIdentityValue<AccDataType>();
+
+            for(ck::index_t y = 0; y < window_spatial_lengths[0]; ++y)
+            {
+                ck::index_t hi = ho * window_strides[0] + y - in_left_pads[0];
+                for(ck::index_t x = 0; x < window_spatial_lengths[1]; ++x)
+                {
+                    ck::index_t wi = wo * window_strides[1] + x - in_left_pads[1];
+                    if(hi >= 0 && hi < static_cast<ck::index_t>(in.mDesc.GetLengths()[2]) &&
+                       wi >= 0 && wi < static_cast<ck::index_t>(in.mDesc.GetLengths()[3]))
+                    {
+                        AccDataType currVal = static_cast<AccDataType>(in(n, c, hi, wi));
+
+                        in_elementwise_op(currVal, currVal);
+
+                        Accumulation::Calculate(accuVal, currVal);
+                    }
+                }
+            }
+
+            acc_elementwise_op(accuVal, accuVal);
+
+            out(n, c, ho, wo) = accuVal;
+        };
+
+        make_ParallelTensorFunctor(f_nchw,
+                                   out.mDesc.GetLengths()[0],
+                                   out.mDesc.GetLengths()[1],
+                                   out.mDesc.GetLengths()[2],
+                                   out.mDesc.GetLengths()[3])(std::thread::hardware_concurrency());
+    }
+    else
+    {
+        using Accumulation = ck::detail::AccumulateWithIndexAndNanCheck<PropagateNan,
+                                                                        ReduceOperation,
+                                                                        AccDataType,
+                                                                        IndexDataType>;
+        auto f_nchw        = [&](auto n, auto c, auto ho, auto wo) {
+            auto accuVal            = ReduceOperation::template GetIdentityValue<AccDataType>();
+            IndexDataType accuIndex = 0;
+
+            for(ck::index_t y = 0; y < window_spatial_lengths[0]; ++y)
+            {
+                ck::index_t hi = ho * window_strides[0] + y - in_left_pads[0];
+                for(ck::index_t x = 0; x < window_spatial_lengths[1]; ++x)
+                {
+                    ck::index_t wi = wo * window_strides[1] + x - in_left_pads[1];
+                    if(hi >= 0 && hi < in.mDesc.GetLengths()[2] && wi >= 0 &&
+                       wi < in.mDesc.GetLengths()[3])
+                    {
+                        AccDataType currVal     = static_cast<AccDataType>(in(n, c, hi, wi));
+                        IndexDataType currIndex = y * window_spatial_lengths[1] + x;
+
+                        in_elementwise_op(currVal, currVal);
+
+                        Accumulation::Calculate(accuVal, currVal, accuIndex, currIndex);
+                    }
+                }
+            }
+
+            acc_elementwise_op(accuVal, accuVal);
+
+            out(n, c, ho, wo)         = accuVal;
+            out_indices(n, c, ho, wo) = accuIndex;
+        };
+
+        make_ParallelTensorFunctor(f_nchw,
+                                   out.mDesc.GetLengths()[0],
+                                   out.mDesc.GetLengths()[1],
+                                   out.mDesc.GetLengths()[2],
+                                   out.mDesc.GetLengths()[3])(std::thread::hardware_concurrency());
+    };
+}
+
+template <typename InDataType,
+          typename OutDataType,
+          typename AccDataType,
+          typename IndexDataType,
+          typename InLayout,
+          typename OutLayout,
+          ck::ReduceTensorOp ReduceOpId,
+          bool PropagateNan,
+          bool OutputIndex>
+bool pool_test(bool do_verification,
+               int init_method,
+               bool time_kernel,
+               ck::index_t N,
+               ck::index_t C,
+               ck::index_t Y,
+               ck::index_t X,
+               ck::index_t Hi,
+               ck::index_t Wi,
+               ck::index_t window_stride_h,
+               ck::index_t window_stride_w,
+               ck::index_t in_left_pad_h,
+               ck::index_t in_left_pad_w,
+               ck::index_t in_right_pad_h,
+               ck::index_t in_right_pad_w)
+{
+    using DevicePoolFwdInstance =
+        ck::tensor_operation::device::DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C<
+            InDataType,  // InDataType
+            OutDataType, // OutDataType
+            AccDataType, // AccDataType
+            ReduceOpId,
+            OutputIndex,
+            64, // BlockSize
+            64, // ReduceMThreadClusterSize
+            1,  // ReduceKThreadClusterSize
+            4,  // ReduceMThreadSliceSize
+            1,  // ReduceKThreadSliceSize
+            4>; // InSrcOutDstVectorSize
+
+    const ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - Y) / window_stride_h + 1;
+    const ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - X) / window_stride_w + 1;
+
+    const std::array<ck::index_t, 2> window_spatial_lengths{{Y, X}};
+    const std::array<ck::index_t, 2> window_strides{{window_stride_h, window_stride_w}};
+    const std::array<ck::index_t, 2> input_left_pads{{in_left_pad_h, in_left_pad_w}};
+    const std::array<ck::index_t, 2> input_right_pads{{in_right_pad_h, in_right_pad_w}};
+
+    // tensor layout
+    auto f_host_tensor_descriptor =
+        [](std::size_t N_, std::size_t C_, std::size_t H, std::size_t W, auto layout) {
+            using namespace ck::literals;
+
+            if constexpr(ck::is_same<decltype(layout), ck::tensor_layout::convolution::NCHW>::value)
+            {
+                return HostTensorDescriptor({N_, C_, H, W}, {C_ * H * W, H * W, W, 1_uz});
+            }
+            else if constexpr(ck::is_same<decltype(layout),
+                                          ck::tensor_layout::convolution::NHWC>::value)
+            {
+                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, InLayout{}));
+    Tensor<OutDataType> out_n_c_ho_wo_host(f_host_tensor_descriptor(N, C, Ho, Wo, OutLayout{}));
+    Tensor<IndexDataType> out_indices_n_c_ho_wo_host(
+        f_host_tensor_descriptor(N, C, Ho, Wo, OutLayout{}));
+    Tensor<OutDataType> out_n_c_ho_wo_device(f_host_tensor_descriptor(N, C, Ho, Wo, OutLayout{}));
+    Tensor<IndexDataType> out_indices_n_c_ho_wo_device(
+        f_host_tensor_descriptor(N, C, Ho, Wo, OutLayout{}));
+
+    std::cout << "in_n_c_hi_wi: " << in_n_c_hi_wi.mDesc << std::endl;
+    std::cout << "out_n_c_ho_wo: " << out_n_c_ho_wo_host.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1: in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_1<InDataType>{1}); break;
+    case 2: in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5}); break;
+    default: in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_3<InDataType>{-5.0, 5.0});
+    }
+
+    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());
+
+    auto pool         = DevicePoolFwdInstance{};
+    auto invoker_ptr  = pool.MakeInvokerPointer();
+    auto argument_ptr = pool.MakeArgumentPointer(
+        static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
+        static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
+        static_cast<IndexDataType*>(out_indices_device_buf.GetDeviceBuffer()),
+        N,
+        C,
+        std::array<ck::index_t, 2>{{Hi, Wi}},
+        std::array<ck::index_t, 2>{{Y, X}},
+        std::array<ck::index_t, 2>{{Ho, Wo}},
+        window_strides,
+        input_left_pads,
+        input_right_pads);
+
+    if(!pool.IsSupportedArgument(argument_ptr.get()))
+    {
+        throw std::runtime_error("wrong! device_op with the specified compilation parameters does "
+                                 "not support this problem");
+    }
+
+    float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop = std::size_t(2) * N * C * Ho * Wo * Y * X;
+
+    std::size_t num_btype =
+        sizeof(InDataType) * (N * C * Hi * Wi) + sizeof(OutDataType) * (N * C * Ho * Wo);
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
+              << std::endl;
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        pool_host_verify<InDataType,
+                         OutDataType,
+                         AccDataType,
+                         IndexDataType,
+                         ReduceOpId,
+                         PropagateNan,
+                         OutputIndex>(in_n_c_hi_wi,
+                                      out_n_c_ho_wo_host,
+                                      out_indices_n_c_ho_wo_host,
+                                      window_spatial_lengths,
+                                      window_strides,
+                                      input_left_pads,
+                                      input_right_pads);
+
+        out_device_buf.FromDevice(out_n_c_ho_wo_device.mData.data());
+
+        pass = pass && ck::utils::check_err(out_n_c_ho_wo_device, out_n_c_ho_wo_host);
+
+        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);
+        };
+    }
+
+    return (pass);
+};

example/13_pool2d_fwd/pool2d_fwd_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "pool2d_fwd_common.hpp"
+
+using InDataType  = ck::half_t;
+using OutDataType = ck::half_t;
+using AccDataType = float;
+
+using IndexDataType = int32_t;
+
+using InLayout  = ck::tensor_layout::convolution::NHWC;
+using OutLayout = ck::tensor_layout::convolution::NHWC;
+
+#if 1
+static constexpr auto ReduceOpId = ck::ReduceTensorOp::MAX;
+#else
+static constexpr auto ReduceOpId = ck::ReduceTensorOp::AVG;
+#endif
+
+static constexpr bool OutputIndex  = false;
+static constexpr bool PropagateNan = false;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification;
+    int init_method;
+    bool time_kernel;
+
+    // Pool shape
+    ck::index_t N               = 128;
+    ck::index_t C               = 192;
+    ck::index_t Y               = 3;
+    ck::index_t X               = 3;
+    ck::index_t Hi              = 71;
+    ck::index_t Wi              = 71;
+    ck::index_t window_stride_h = 2;
+    ck::index_t window_stride_w = 2;
+    ck::index_t in_left_pad_h   = 1;
+    ck::index_t in_left_pad_w   = 1;
+    ck::index_t in_right_pad_h  = 1;
+    ck::index_t in_right_pad_w  = 1;
+
+    if(argc == 1)
+    {
+        do_verification = true;
+        init_method     = 1;
+        time_kernel     = true;
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = static_cast<bool>(std::stoi(argv[3]));
+    }
+    else if(argc == 16)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = static_cast<bool>(std::stoi(argv[3]));
+
+        N               = std::stoi(argv[4]);
+        C               = std::stoi(argv[5]);
+        Y               = std::stoi(argv[6]);
+        X               = std::stoi(argv[7]);
+        Hi              = std::stoi(argv[8]);
+        Wi              = std::stoi(argv[9]);
+        window_stride_h = std::stoi(argv[10]);
+        window_stride_w = std::stoi(argv[11]);
+        in_left_pad_h   = std::stoi(argv[12]);
+        in_left_pad_w   = std::stoi(argv[13]);
+        in_right_pad_h  = std::stoi(argv[14]);
+        in_right_pad_w  = std::stoi(argv[15]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 15: N, C, Y, X, Hi, Wi, Sy, Sx, LeftPy, LeftPx, RightPy, "
+               "RightPx\n");
+        exit(0);
+    }
+
+    bool pass = pool_test<InDataType,
+                          OutDataType,
+                          AccDataType,
+                          IndexDataType,
+                          InLayout,
+                          OutLayout,
+                          ReduceOpId,
+                          PropagateNan,
+                          OutputIndex>(do_verification,
+                                       init_method,
+                                       time_kernel,
+                                       N,
+                                       C,
+                                       Y,
+                                       X,
+                                       Hi,
+                                       Wi,
+                                       window_stride_h,
+                                       window_stride_w,
+                                       in_left_pad_h,
+                                       in_left_pad_w,
+                                       in_right_pad_h,
+                                       in_right_pad_w);
+
+    return (pass ? 0 : 1);
+}

example/13_pool2d_fwd/pool2d_fwd_fp32.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+#include "pool2d_fwd_common.hpp"
+
+using InDataType  = float;
+using OutDataType = float;
+using AccDataType = float;
+
+using IndexDataType = int32_t;
+
+using InLayout  = ck::tensor_layout::convolution::NHWC;
+using OutLayout = ck::tensor_layout::convolution::NHWC;
+
+#if 1
+static constexpr auto ReduceOpId = ck::ReduceTensorOp::MAX;
+#else
+static constexpr auto ReduceOpId = ck::ReduceTensorOp::AVG;
+#endif
+
+static constexpr bool OutputIndex  = false;
+static constexpr bool PropagateNan = false;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification;
+    int init_method;
+    bool time_kernel;
+
+    // Pool shape
+    ck::index_t N               = 128;
+    ck::index_t C               = 192;
+    ck::index_t Y               = 3;
+    ck::index_t X               = 3;
+    ck::index_t Hi              = 71;
+    ck::index_t Wi              = 71;
+    ck::index_t window_stride_h = 2;
+    ck::index_t window_stride_w = 2;
+    ck::index_t in_left_pad_h   = 1;
+    ck::index_t in_left_pad_w   = 1;
+    ck::index_t in_right_pad_h  = 1;
+    ck::index_t in_right_pad_w  = 1;
+
+    if(argc == 1)
+    {
+        do_verification = true;
+        init_method     = 1;
+        time_kernel     = true;
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = static_cast<bool>(std::stoi(argv[3]));
+    }
+    else if(argc == 16)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = static_cast<bool>(std::stoi(argv[3]));
+
+        N               = std::stoi(argv[4]);
+        C               = std::stoi(argv[5]);
+        Y               = std::stoi(argv[6]);
+        X               = std::stoi(argv[7]);
+        Hi              = std::stoi(argv[8]);
+        Wi              = std::stoi(argv[9]);
+        window_stride_h = std::stoi(argv[10]);
+        window_stride_w = std::stoi(argv[11]);
+        in_left_pad_h   = std::stoi(argv[12]);
+        in_left_pad_w   = std::stoi(argv[13]);
+        in_right_pad_h  = std::stoi(argv[14]);
+        in_right_pad_w  = std::stoi(argv[15]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 15: N, C, Y, X, Hi, Wi, Sy, Sx, LeftPy, LeftPx, RightPy, "
+               "RightPx\n");
+        exit(0);
+    }
+
+    bool pass = pool_test<InDataType,
+                          OutDataType,
+                          AccDataType,
+                          IndexDataType,
+                          InLayout,
+                          OutLayout,
+                          ReduceOpId,
+                          PropagateNan,
+                          OutputIndex>(do_verification,
+                                       init_method,
+                                       time_kernel,
+                                       N,
+                                       C,
+                                       Y,
+                                       X,
+                                       Hi,
+                                       Wi,
+                                       window_stride_h,
+                                       window_stride_w,
+                                       in_left_pad_h,
+                                       in_left_pad_w,
+                                       in_right_pad_h,
+                                       in_right_pad_w);
+
+    return (pass ? 0 : 1);
+}