composable_kernel/profiler/include/profiler/profile_max_pool2d_bwd_impl.hpp

// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.

#pragma once

#include <iomanip>

#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/gpu/pool3d_fwd.hpp"
#include "ck/library/tensor_operation_instance/gpu/max_pool_bwd.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_pool_fwd.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_maxpool_bwd.hpp"

namespace ck {
namespace profiler {

template <typename InDataType,
          typename OutDataType,
          typename IndexDataType,
          typename DOutDataType,
          typename DInDataType,
          bool PropagateNan>
bool profile_max_pool2d_bwd_impl(int do_verification,
                                 int init_method,
                                 bool do_log,
                                 bool time_kernel,
                                 std::vector<index_t> in_length,
                                 std::vector<index_t> window_spatial_lengths,
                                 std::vector<index_t> window_strides,
                                 std::vector<index_t> window_dilations,
                                 std::vector<index_t> input_left_pads,
                                 std::vector<index_t> input_right_pads)
{
    // AtomicAdd only support f32 for now. ComputeDataType must be float32
    using ComputeDataType = float;

    constexpr index_t InOutRank  = 4;
    constexpr index_t WindowRank = 2;

    using PassThrough = ck::tensor_operation::element_wise::PassThrough;

    if(in_length.size() != InOutRank || window_spatial_lengths.size() != WindowRank ||
       window_strides.size() != WindowRank || window_dilations.size() != WindowRank ||
       input_left_pads.size() != WindowRank || input_right_pads.size() != WindowRank)
    {
        std::cout << "Parameter is incorrect" << std::endl;
        return false;
    }

    std::vector<index_t> out_length(InOutRank);

    int N = in_length[0];
    int C = in_length[1];

    out_length[0] = N;
    out_length[1] = C;

    // Calculate Ho, Wo
    for(unsigned i = 2; i < InOutRank; ++i)
    {
        const int idx         = i - 2;
        auto pad1             = input_left_pads[idx];
        auto pad2             = input_right_pads[idx];
        auto windows_size     = window_spatial_lengths[idx];
        auto windows_stride   = window_strides[idx];
        auto windows_dilation = window_dilations[idx];
        auto eff              = (windows_size - 1) * windows_dilation + 1;
        out_length[i]         = (in_length[i] + pad1 + pad2 - eff) / windows_stride + 1;
    }

    int Hi = in_length[2];
    int Wi = in_length[3];
    int Ho = out_length[2];
    int Wo = out_length[3];

    auto f_host_tensor_descriptor =
        [](std::size_t N_, std::size_t C_, std::size_t H, std::size_t W) {
            using namespace ck::literals;

            return HostTensorDescriptor({N_, C_, H, W}, {C_ * H * W, 1_uz, W * C_, C_});
        };

    Tensor<InDataType> in_n_c_hi_wi(f_host_tensor_descriptor(N, C, Hi, Wi));
    Tensor<OutDataType> out_n_c_ho_wo(f_host_tensor_descriptor(N, C, Ho, Wo));
    Tensor<IndexDataType> out_indices_n_c_ho_wo(f_host_tensor_descriptor(N, C, Ho, Wo));
    Tensor<DOutDataType> dout_n_c_ho_wo(f_host_tensor_descriptor(N, C, Ho, Wo));
    Tensor<DInDataType> din_n_c_hi_wi_host(f_host_tensor_descriptor(N, C, Hi, Wi));

    Tensor<DInDataType> din_n_c_hi_wi_device(f_host_tensor_descriptor(N, C, Hi, Wi));

    switch(init_method)
    {
    case 0: {
        in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_1<InDataType>{});
        dout_n_c_ho_wo.GenerateTensorValue(GeneratorTensor_1<DOutDataType>{});
        break;
    }
    case 1: {
        in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
        dout_n_c_ho_wo.GenerateTensorValue(GeneratorTensor_2<DOutDataType>{-5, 5});
        break;
    }
    default: {
        in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_3<InDataType>{-0.5, 0.5});
        dout_n_c_ho_wo.GenerateTensorValue(GeneratorTensor_3<DOutDataType>{-0.5, 0.5});
    }
    }

    DeviceMem indices_device_buf(sizeof(IndexDataType) *
                                 out_indices_n_c_ho_wo.mDesc.GetElementSpaceSize());
    DeviceMem dout_device_buf(sizeof(DOutDataType) * dout_n_c_ho_wo.mDesc.GetElementSpaceSize());
    DeviceMem din_device_buf(sizeof(DInDataType) *
                             din_n_c_hi_wi_device.mDesc.GetElementSpaceSize());

    // Generate index data from forwarding
    {
        using ReferencePoolingFwdInstance =
            ck::tensor_operation::host::ReferencePoolingFwd<InOutRank,
                                                            WindowRank,
                                                            InDataType,
                                                            OutDataType,
                                                            ComputeDataType,
                                                            IndexDataType,
                                                            ck::ReduceTensorOp::MAX,
                                                            false,
                                                            true>;

        ReferencePoolingFwdInstance ref_pooling_fwd;
        auto ref_pooling_fwd_argument = ref_pooling_fwd.MakeArgument(in_n_c_hi_wi,
                                                                     out_n_c_ho_wo,
                                                                     out_indices_n_c_ho_wo,
                                                                     window_spatial_lengths,
                                                                     window_strides,
                                                                     window_dilations,
                                                                     input_left_pads,
                                                                     input_right_pads);
        auto ref_pooling_fwd_invoker  = ref_pooling_fwd.MakeInvoker();
        ref_pooling_fwd_invoker.Run(ref_pooling_fwd_argument);
    }

    indices_device_buf.ToDevice(out_indices_n_c_ho_wo.mData.data());
    dout_device_buf.ToDevice(dout_n_c_ho_wo.mData.data());

    using DeviceOp =
        ck::tensor_operation::device::DeviceMaxPoolBwd<DOutDataType, IndexDataType, DInDataType>;

    // get device op instances
    const auto instance_ptrs =
        ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
            DeviceOp>::GetInstances();

    std::cout << "found " << instance_ptrs.size() << " instances" << std::endl;

    std::string best_instance_name;
    float best_avg_time   = std::numeric_limits<float>::max();
    float best_gb_per_sec = 0;

    if(do_verification)
    {
        using ReferencePoolingBwdInstance =
            ck::tensor_operation::host::ReferenceMaxPoolBwd<DOutDataType,
                                                            IndexDataType,
                                                            ComputeDataType,
                                                            DInDataType,
                                                            PassThrough>;

        ReferencePoolingBwdInstance ref_pooling_bwd;
        auto ref_pooling_bwd_argument = ref_pooling_bwd.MakeArgument(
            dout_n_c_ho_wo, out_indices_n_c_ho_wo, din_n_c_hi_wi_host, PassThrough{});
        auto ref_invoker = ref_pooling_bwd.MakeInvoker();
        ref_invoker.Run(ref_pooling_bwd_argument);
    }

    int num_kernel = 0;

    bool pass           = true;
    bool instance_found = false;

    for(auto& inst_ptr : instance_ptrs)
    {
        auto argument_ptr = inst_ptr->MakeArgumentPointer(
            static_cast<DOutDataType*>(dout_device_buf.GetDeviceBuffer()),
            static_cast<IndexDataType*>(indices_device_buf.GetDeviceBuffer()),
            static_cast<DInDataType*>(din_device_buf.GetDeviceBuffer()),
            dout_n_c_ho_wo.mDesc.GetElementSpaceSize(),
            din_n_c_hi_wi_device.mDesc.GetElementSpaceSize(),
            window_spatial_lengths,
            window_strides,
            window_dilations);

        if(inst_ptr->IsSupportedArgument(argument_ptr.get()))
        {
            ++num_kernel;
            instance_found = true;
        }
        else
        {
            if(time_kernel)
            {
                std::cout << inst_ptr->GetTypeString() << " skipped due to unsupported argument: ";
                LogRange(std::cout << "doutput lengths = ", out_length, ", ") << std::endl;
            }

            continue;
        }

        size_t workspace_sz = inst_ptr->GetWorkSpaceSize(argument_ptr.get());
        DeviceMem workspace_device_buf(workspace_sz);
        inst_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_device_buf.GetDeviceBuffer());

        auto invoker_ptr = inst_ptr->MakeInvokerPointer();
        float avg_time   = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});

        std::size_t num_bytes =
            dout_n_c_ho_wo.mDesc.GetElementSize() * sizeof(DOutDataType) +
            out_indices_n_c_ho_wo.mDesc.GetElementSize() * sizeof(IndexDataType) +
            din_n_c_hi_wi_device.mDesc.GetElementSize() * sizeof(DInDataType);

        float gb_per_sec = num_bytes / 1.E6 / avg_time;

        if(time_kernel)
            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << gb_per_sec << " GB/s, "
                      << inst_ptr->GetTypeString() << std::endl;

        if(avg_time < best_avg_time)
        {
            best_instance_name = inst_ptr->GetTypeString();
            best_avg_time      = avg_time;
            best_gb_per_sec    = gb_per_sec;
        }

        if(do_verification)
        {
            din_device_buf.FromDevice(din_n_c_hi_wi_device.mData.data());

            bool local_pass = ck::utils::check_err(din_n_c_hi_wi_device.mData,
                                                   din_n_c_hi_wi_host.mData,
                                                   "Error: Incorrect results",
                                                   1e-3,
                                                   1e-3);

            if(do_log)
            {
                LogRangeAsType<float>(
                    std::cout << "out_indices_n_c_ho_wo: ", out_indices_n_c_ho_wo.mData, ",")
                    << std::endl;

                LogRangeAsType<float>(
                    std::cout << "din_n_c_hi_wi_device: ", din_n_c_hi_wi_device.mData, ",")
                    << std::endl;

                LogRangeAsType<float>(
                    std::cout << "din_n_c_hi_wi_host: ", din_n_c_hi_wi_host.mData, ",")
                    << std::endl;
            }

            if(!local_pass)
            {
                std::cout << inst_ptr->GetTypeString() << " failed verification: ";
                LogRange(std::cout << "doutput lengths = [", out_length, ", ") << "]." << std::endl;
                pass &= local_pass;
            }
            else
            {
                if(time_kernel)
                {
                    std::cout << "pass" << std::endl;
                }
            }
        }
    }

    if(time_kernel)
    {
        LogRange(std::cout << "length = ", out_length, ",") << std::endl;
        std::cout << "best perf = " << best_avg_time << " ms, " << best_gb_per_sec << " GB/s, "
                  << best_instance_name << std::endl;
    }

    if(num_kernel == 0)
    {
        std::cout << "Error: No kernel is applicable" << std::endl;
        return false;
    }

    return pass && instance_found;
}

} // namespace profiler
} // namespace ck