#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "tensor_layout.hpp"
#include "device_tensor.hpp"
#include "device_convnd_fwd_avx2_nhwc_kyxc_nhwk.hpp"
#include "element_wise_operation_cpu.hpp"
#include "reference_conv_fwd.hpp"
#include "element_wise_operation_cpu.hpp"
#include "dynamic_buffer_cpu.hpp"
#include <omp.h>

#define AVX2_DATA_ALIGNMENT 32
using F32 = float;
using F16 = ck::half_t;

namespace ck {
namespace tensor_operation {
namespace cpu {
namespace device {
namespace device_conv2d_fwd_avx2_instance {

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

void add_device_conv2d_fwd_avx2_nhwc_kyxc_nhwk(
    std::vector<DeviceConvFwdPtr<PassThrough, PassThrough, PassThrough>>& instances);

} // namespace device_conv2d_fwd_avx2_instance
} // namespace device
} // namespace cpu
} // namespace tensor_operation
} // namespace ck

using InElementOp  = ck::tensor_operation::cpu::element_wise::PassThrough;
using WeiElementOp = ck::tensor_operation::cpu::element_wise::PassThrough;
using OutElementOp = ck::tensor_operation::cpu::element_wise::PassThrough;

template <typename T>
static bool
check_out(const Tensor<T>& ref, const Tensor<T>& result, double nrms, int per_pixel_check = 0)
{
    int error_count = 0;
    float max_diff  = 1e-5;

    double square_difference = .0;
    double mag1              = .0;
    double mag2              = .0;

    for(int i = 0; i < ref.mData.size(); ++i)
    {
        double ri = (double)ref.mData[i];
        double pi = (double)result.mData[i];
        double d  = ri - pi;

        if(per_pixel_check)
        {
            if(max_diff < std::abs(d))
            {
                error_count++;
                printf("idx:%3d, ref:%f, res:%f (diff:%f)\n",
                       i,
                       double(ref.mData[i]),
                       double(result.mData[i]),
                       d);
            }
        }

        square_difference += d * d;
        if(std::abs(mag1) < std::abs(ri))
            mag1 = ri;
        if(std::abs(mag2) < std::abs(pi))
            mag2 = pi;
    }

    double mag = std::max({std::fabs(mag1), std::fabs(mag2), std::numeric_limits<double>::min()});
    double computed_nrms = std::sqrt(square_difference) / (std::sqrt(ref.mData.size()) * mag);

    if(computed_nrms >= nrms)
        printf("nrms:%lf, mag1:%lf, mag2:%lf, expected_nrms is %1f\n",
               computed_nrms,
               mag1,
               mag2,
               nrms);

    return computed_nrms < nrms && error_count == 0;
}

float calculate_gflops() {}

int main(int argc, char* argv[])
{
    int data_type   = 0;
    int init_method = 0;

    // Conv shape
    ck::index_t N               = 2;
    ck::index_t K               = 256;
    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 conv_stride_h   = 1;
    ck::index_t conv_stride_w   = 1;
    ck::index_t conv_dilation_h = 1;
    ck::index_t conv_dilation_w = 1;
    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)
    {
        data_type   = 0;
        init_method = 1;
    }
    else if(argc == 3)
    {
        data_type   = std::stoi(argv[1]);
        init_method = std::stoi(argv[2]);
    }
    else if(argc == 18)
    {
        data_type   = std::stoi(argv[1]);
        init_method = std::stoi(argv[2]);

        N               = std::stoi(argv[3]);
        K               = 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]);
        conv_stride_h   = std::stoi(argv[10]);
        conv_stride_w   = std::stoi(argv[11]);
        conv_dilation_h = std::stoi(argv[12]);
        conv_dilation_w = std::stoi(argv[13]);
        in_left_pad_h   = std::stoi(argv[14]);
        in_left_pad_w   = std::stoi(argv[15]);
        in_right_pad_h  = std::stoi(argv[16]);
        in_right_pad_w  = std::stoi(argv[17]);
    }
    else
    {
        printf("arg1: data type (0=fp32, 1=fp16)\n");
        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
        printf("arg3 to 17: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
               "RightPx\n");
        exit(1);
    }

    auto Run = [&](auto input_type, auto wei_type, auto out_type) {
        using InDataType  = decltype(input_type);
        using WeiDataType = decltype(wei_type);
        using OutDataType = decltype(out_type);

        using ReferenceConvFwdInstance = ck::tensor_operation::host::ReferenceConvFwd<InDataType,
                                                                                      WeiDataType,
                                                                                      OutDataType,
                                                                                      InElementOp,
                                                                                      WeiElementOp,
                                                                                      OutElementOp>;

        const ck::index_t YEff = (Y - 1) * conv_dilation_h + 1;
        const ck::index_t XEff = (X - 1) * conv_dilation_w + 1;

        const ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - YEff) / conv_stride_h + 1;
        const ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - XEff) / conv_stride_w + 1;

        const std::vector<ck::index_t> input_spatial_lengths{{Hi, Wi}};
        const std::vector<ck::index_t> filter_spatial_lengths{{Y, X}};
        const std::vector<ck::index_t> output_spatial_lengths{{Ho, Wo}};
        const std::vector<ck::index_t> conv_filter_strides{{conv_stride_h, conv_stride_w}};
        const std::vector<ck::index_t> conv_filter_dilations{{conv_dilation_h, conv_dilation_w}};
        const std::vector<ck::index_t> input_left_pads{{in_left_pad_h, in_left_pad_w}};
        const std::vector<ck::index_t> input_right_pads{{in_right_pad_h, in_right_pad_w}};

        auto f_host_tensor_descriptor = [](std::size_t N_,
                                           std::size_t C_,
                                           std::size_t H_,
                                           std::size_t W_) {
            return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H_, W_}),
                                        std::vector<std::size_t>({C_ * H_ * W_, 1, W_ * C_, C_}));
        };

        Tensor<InDataType> in_n_c_hi_wi(f_host_tensor_descriptor(N, C, Hi, Wi));
        Tensor<WeiDataType> wei_k_c_y_x(f_host_tensor_descriptor(K, C, Y, X));
        Tensor<OutDataType> out_n_k_ho_wo_host_result(f_host_tensor_descriptor(N, K, Ho, Wo));
        Tensor<OutDataType> out_n_k_ho_wo_device_result(f_host_tensor_descriptor(N, K, Ho, Wo));

        std::cout << "in (N, C, Hi, Wi): " << in_n_c_hi_wi.mDesc << std::endl;
        std::cout << "wei(K, C,  Y,  X): " << wei_k_c_y_x.mDesc << std::endl;
        std::cout << "out(N, K, Ho, Wo): " << out_n_k_ho_wo_host_result.mDesc << std::endl;
        std::cout << "LPad(H, W):" << in_left_pad_h << "," << in_left_pad_w
                  << ", RPad(H, W):" << in_right_pad_h << "," << in_right_pad_w
                  << ", Stride(H, W):" << conv_stride_h << ", " << conv_stride_w
                  << ", Dilation(H, W):" << conv_dilation_h << ", " << conv_dilation_w
                  << ", Threads:" << omp_get_max_threads() << std::endl;

        int per_pixel_check = 0;
        switch(init_method)
        {
        case 0:
            in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_1<InDataType>{});
            wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_1<WeiDataType>{});
            per_pixel_check = 1;
            break;
        case 1:

            in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
            // in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_1<InDataType>{});
            wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
            // wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_1<WeiDataType>{});
            per_pixel_check = 1;
            break;

        case 2:
            in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
            wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
            break;

        case 3:

#define PACK_32(v24, v16, v8, v0) \
    (((v24 & 0xff) << 24) | ((v16 & 0xff) << 16) | ((v8 & 0xff) << 8) | ((v0 & 0xff) << 0))

            for(auto i_n = 0; i_n < N; i_n++)
            {
                for(auto i_c = 0; i_c < C; i_c++)
                {
                    for(auto i_hi = 0; i_hi < Hi; i_hi++)
                    {
                        for(auto i_wi = 0; i_wi < Wi; i_wi++)
                        {
                            uint32_t v                         = PACK_32(i_n, i_c, i_hi, i_wi);
                            in_n_c_hi_wi(i_n, i_c, i_hi, i_wi) = *reinterpret_cast<float*>(&v);
                        }
                    }
                }
            }

            for(auto i_k = 0; i_k < K; i_k++)
            {
                for(auto i_c = 0; i_c < C; i_c++)
                {
                    for(auto i_y = 0; i_y < Y; i_y++)
                    {
                        for(auto i_x = 0; i_x < X; i_x++)
                        {
                            uint32_t v                      = PACK_32(i_k, i_c, i_y, i_x);
                            wei_k_c_y_x(i_k, i_c, i_y, i_x) = *reinterpret_cast<float*>(&v);
                        }
                    }
                }
            }
            break;
        default:
            in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_3<InDataType>{0, 1});
            wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-1, 1});
        }

        DeviceAlignedMemCPU in_device_buf(sizeof(InDataType) * in_n_c_hi_wi.mDesc.GetElementSpace(),
                                          AVX2_DATA_ALIGNMENT);
        DeviceAlignedMemCPU wei_device_buf(
            sizeof(WeiDataType) * wei_k_c_y_x.mDesc.GetElementSpace(), AVX2_DATA_ALIGNMENT);
        DeviceAlignedMemCPU out_device_buf(sizeof(OutDataType) *
                                               out_n_k_ho_wo_host_result.mDesc.GetElementSpace(),
                                           AVX2_DATA_ALIGNMENT);

        in_device_buf.ToDevice(in_n_c_hi_wi.mData.data());
        wei_device_buf.ToDevice(wei_k_c_y_x.mData.data());

        // get host result
        {
            auto ref_conv    = ReferenceConvFwdInstance{};
            auto ref_invoker = ref_conv.MakeInvoker();

            auto ref_argument = ref_conv.MakeArgument(in_n_c_hi_wi,
                                                      wei_k_c_y_x,
                                                      out_n_k_ho_wo_host_result,
                                                      conv_filter_strides,
                                                      conv_filter_dilations,
                                                      input_left_pads,
                                                      input_right_pads,
                                                      InElementOp{},
                                                      WeiElementOp{},
                                                      OutElementOp{});
            ref_invoker.Run(ref_argument);
        }

        using PassThrough          = ck::tensor_operation::cpu::element_wise::PassThrough;
        using DeviceConvFwdNoOpPtr = ck::tensor_operation::cpu::device::
            DeviceConvFwdPtr<PassThrough, PassThrough, PassThrough>;

        // add device Conv instances
        std::vector<DeviceConvFwdNoOpPtr> conv_ptrs;

        if constexpr(ck::is_same_v<ck::remove_cv_t<InDataType>, float> &&
                     ck::is_same_v<ck::remove_cv_t<WeiDataType>, float> &&
                     ck::is_same_v<ck::remove_cv_t<OutDataType>, float>)
        {
            ck::tensor_operation::cpu::device::device_conv2d_fwd_avx2_instance::
                add_device_conv2d_fwd_avx2_nhwc_kyxc_nhwk(conv_ptrs);
        }

        if(conv_ptrs.size() <= 0)
        {
            throw std::runtime_error("wrong! no device Conv instance found");
        }

        // profile device Conv instances
        bool success                    = true;
        double fastest_kernel_time      = std::numeric_limits<double>::max();
        std::string fastest_kernel_name = "";
        double fastest_kernel_gflops    = 0;
        for(auto& conv_ptr : conv_ptrs)
        {
            auto argument_ptr = conv_ptr->MakeArgumentPointer(
                static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
                static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
                static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
                N,
                K,
                C,
                input_spatial_lengths,
                filter_spatial_lengths,
                output_spatial_lengths,
                conv_filter_strides,
                conv_filter_dilations,
                input_left_pads,
                input_right_pads,
                InElementOp{},
                WeiElementOp{},
                OutElementOp{});

            if(conv_ptr->IsSupportedArgument(argument_ptr.get()))
            {
                auto invoker_ptr = conv_ptr->MakeInvokerPointer();
                double time      = invoker_ptr->Run(argument_ptr.get(), 10);

                double total_flop = static_cast<double>(2) * N * C * Ho * Wo * K * Y * X;

                double gflops = (total_flop * 1e-6) / time;

                out_device_buf.FromDevice(out_n_k_ho_wo_device_result.mData.data());

                if(!check_out(out_n_k_ho_wo_host_result,
                              out_n_k_ho_wo_device_result,
                              1e-6,
                              per_pixel_check))
                {
                    std::cout << "Fail Info: " << conv_ptr->GetTypeString() << std::endl;
                    success = false;
                }
                else
                {
                    std::cout << "Pass Info: " << conv_ptr->GetTypeString() << ", Time:" << time
                              << "ms, Gflops:" << gflops << std::endl;

                    if(time < fastest_kernel_time)
                    {
                        fastest_kernel_time   = time;
                        fastest_kernel_name   = conv_ptr->GetTypeString();
                        fastest_kernel_gflops = gflops;
                    }
                }
            }
            else
            {
                std::cout << "Not support Info: " << conv_ptr->GetTypeString() << std::endl;
            }
        }

        if(fastest_kernel_time != std::numeric_limits<double>::max())
        {
            std::cout << "  fastest:" << fastest_kernel_name << ", time:" << fastest_kernel_time
                      << "ms, Gflops:" << fastest_kernel_gflops << std::endl;
        }
        return 0;
        // if(success)
        // {
        //     std::cout << "test conv2d fwd cpu : Pass" << std::endl;
        //     return 0;
        // }
        // else
        // {
        //     std::cout << "test conv2d fwd cpu: Fail " << std::endl;
        //     return -1;
        // }
    };

    if(data_type == 0)
    {
        return Run(F32(), F32(), F32());
    }
    else
    {
        return 1;
    }
}