#include "convolution.h"
#include <stdio.h>
#include <mpi.h>
#include <cuda_runtime.h>
#include "util.h"

#define CUDA_CALL(f)                                                         \
{                                                                            \
cudaError_t err = (f);                                                       \
  if (err != cudaSuccess) {                                                  \
    fprintf(stderr, "CUDA error at [%s:%d] %d %s\n", __FILE__, __LINE__,     \
            err, cudaGetErrorString(err));                                   \
    exit(1);                                                                 \
  }                                                                          \
}

//#define TS 32 // 1979.05 GFLOPS
//#define TS 16 // 3198.13 GFLOPS
#define TS 8    // 4617.74 GFLOPS
//#define TS 4  // 4151.70 GFLOPS

#define MAX_NUM_GPU 4

static float *input, *output, *filter;
static float *in_d[MAX_NUM_GPU], *out_d[MAX_NUM_GPU], *filter_d[MAX_NUM_GPU];
static int N, C, H, W;
static int K, R, S;
static int OH, OW;
static int pad;
static int dilation;
static int stride;
static int mpi_rank, mpi_world_size;
static int num_devices = 1;

int size[2];
int Qnum[MAX_NUM_GPU];


__global__ void cuda2Dconv(
    float *_input, float *_output, float *_filter, 
    int _N, int _C, int _H, int _W,
    int _K, int _R, int _S,
    int _pad, int _dilation, int _stride) {


    const int globalRow = blockDim.x*blockIdx.x + threadIdx.x;
    const int globalCol = blockDim.y*blockIdx.y + threadIdx.y;
    int OH, OW;

    OH = (_H + 2*_pad - _dilation*(_R - 1) - 1)/_stride + 1;
    OW = (_W + 2*_pad - _dilation*(_S - 1) - 1)/_stride + 1;

    int n, k, w;
    w = globalCol;
    n = w/(_K*OW);
    w = w - n*(_K*OW);
    k = w/OW;
    w = w - k*OW;
    int col = w;
    int row = globalRow;

    if (globalRow >= OH || globalCol >= _N*_K*OW)
        return;

    int start_row = row * _stride - _pad;
    int start_col = col * _stride - _pad;

    float o = 0.0f;
    for (int c = 0 ; c < _C ; c++) {
        for (int i = 0 ; i < _R ; i++) {
            for (int j = 0 ; j < _S ; j++) {
                int h = start_row + i * _dilation;
                int w = start_col + j * _dilation;
                if (h < 0 || w < 0 || h >= _H || w >= _W)
                        continue;
                float in = _input[n*_C*_W*_H + c*_W*_H + h*_W + w];
                float fil = _filter[k*_C*_R*_S + c*_R*_S + i*_S + j];
                o += in * fil;
            }
        }
    } 

    _output[n*_K*OH*OW + k*OH*OW + row*OW + col] = o;
}

void convolution(
    float *_input, float *_output, float *_filter, 
    int _N, int _C, int _H, int _W,
    int _K, int _R, int _S,
    int _pad, int _dilation, int _stride) {

    int offset = 0;

    input = _input;
    output = _output;
    filter = _filter;

    MPI_Request request;
    MPI_Status status;

    if (mpi_rank == 0 && mpi_world_size == 2 && size[1] != 0) {
        MPI_Isend(&input[size[0]*C*H*W], size[1]*C*H*W,
                  MPI_FLOAT, 1, 0, MPI_COMM_WORLD, &request);
        MPI_Isend(filter, _K*_C*_R*_S, MPI_FLOAT, 1, 0, MPI_COMM_WORLD, &request);

        if (size[mpi_rank] < MAX_NUM_GPU)
            num_devices = size[mpi_rank];
        //printf("[hong] num_devices = %d\n", num_devices);
    } else if (mpi_rank == 1 && size[mpi_rank] != 0) {
        alloc_tensor(&input, size[1], C, H, W);
        alloc_tensor(&output, size[1], K, OH, OW);
        alloc_tensor(&filter, _K, _C, _R, _S);

        MPI_Recv(input, size[1]*C*H*W, MPI_FLOAT, 0, 0, MPI_COMM_WORLD, &status);
        MPI_Recv(filter, _K*_C*_R*_S, MPI_FLOAT, 0, 0, MPI_COMM_WORLD, &status);

        if (size[mpi_rank] < MAX_NUM_GPU)
            num_devices = size[mpi_rank];
        //printf("[hong] num_devices = %d\n", num_devices);
    }

    //test: num_devices = 1;

    offset = 0;
    for (int i = 0 ; i < num_devices ; i++) {
        CUDA_CALL( cudaMemcpy(in_d[i], input + offset, Qnum[i]*C*H*W*sizeof(float),
                   cudaMemcpyHostToDevice) );
        CUDA_CALL( cudaMemcpy(filter_d[i], filter, K*C*R*S*sizeof(float),
                   cudaMemcpyHostToDevice) );
        offset += Qnum[i]*C*H*W;
    }

    for (int i = 0; i < num_devices; i++) {
        CUDA_CALL( cudaSetDevice(i) );
        CUDA_CALL( cudaDeviceSynchronize() );
    }

    for (int i = 0; i < num_devices; i++) {
        //dim3 gridDim((OH+TS-1)/TS, (Qnum[0]*K*OW + TS - 1)/TS, 1);
        //dim3 blockDim(TS, TS, 1);
        dim3 gridDim((OH+TS-1)/TS, (Qnum[i]*K*OW + TS - 1)/TS, 1);
        dim3 blockDim(TS, TS, 1);

        CUDA_CALL( cudaSetDevice(i) );
        cuda2Dconv<<<gridDim, blockDim>>>(in_d[i], out_d[i], filter_d[i], Qnum[i],
                                          _C, _H, _W, _K, _R, _S, _pad, _dilation, _stride);
    }

    // DO NOT REMOVE; NEEDED FOR TIME MEASURE
    for (int i = 0; i < num_devices; i++) {
        CUDA_CALL( cudaSetDevice(i) );
        CUDA_CALL( cudaDeviceSynchronize() );
    }

    offset = 0;
    for (int i = 0; i < num_devices; i++) {
        CUDA_CALL( cudaSetDevice(i) );
        CUDA_CALL( cudaMemcpy(output + offset, out_d[i],
                   Qnum[i]*K*OH*OW * sizeof(float), cudaMemcpyDeviceToHost) );
        offset += Qnum[i]*K*OH*OW;
    }

    for (int i = 0; i < num_devices; i++) {
        CUDA_CALL( cudaSetDevice(i) );
        CUDA_CALL( cudaDeviceSynchronize() );
    }

    if (mpi_rank == 0 && mpi_world_size == 2 && size[1] != 0)
            MPI_Recv(&output[size[0]*K*OH*OW], size[1]*K*OH*OW,
                     MPI_FLOAT, 1, 0, MPI_COMM_WORLD, &status);
    else if (mpi_rank == 1 && size[1] != 0)
            MPI_Isend(output, size[1]*K*OH*OW, MPI_FLOAT, 0, 0, MPI_COMM_WORLD, &request);
}

void convolution_init(
    int _N, int _C, int _H, int _W,
    int _K, int _R, int _S,
    int _pad, int _dilation, int _stride) {
    N = _N; C = _C; H = _H; W = _W; K = _K; R = _R; S = _S;
    pad = _pad;
    dilation = _dilation;
    stride = _stride;

    MPI_Comm_size(MPI_COMM_WORLD, &mpi_world_size);
    MPI_Comm_rank(MPI_COMM_WORLD, &mpi_rank);

    OH = (H + 2*pad - dilation*(R - 1) - 1)/stride + 1;
    OW = (W + 2*pad - dilation*(S - 1) - 1)/stride + 1;

    //if (mpi_world_size == 2 && _N > 4)
    if (mpi_world_size == 2)
        size[1] = _N / 2;
    else
        size[1] = 0;
    size[0] = N - size[1];

    if (size[mpi_rank] < MAX_NUM_GPU) {
        num_devices = size[mpi_rank];

        for (int i = 0 ; i < size[mpi_rank] ; i++)
            Qnum[i] = 1;
    } else {
        num_devices = MAX_NUM_GPU;
        int remain = size[mpi_rank] % MAX_NUM_GPU;
        int quot = size[mpi_rank] / MAX_NUM_GPU;
        for (int i = 0 ; i < MAX_NUM_GPU ; i++) {
            Qnum[i] = quot;
            if (i < remain)
                Qnum[i]++;
        }
    }

    for (int i = 0 ; i < num_devices ; i++) {
        CUDA_CALL( cudaSetDevice(i) );
        //printf("[hong] cudaMallock - i = %d\n", i);

        CUDA_CALL( cudaMalloc(&in_d[i], Qnum[i]*C*H*W*sizeof(float)) );
        CUDA_CALL( cudaMalloc(&out_d[i], Qnum[i]*K*OH*OW*sizeof(float)) );
        CUDA_CALL( cudaMalloc(&filter_d[i], K*C*R*S*sizeof(float)) );
    }
}

void convolution_final(
                int _N, int _C, int _H, int _W,
                int _K, int _R, int _S,
                int _pad, int _dilation, int _stride) {
}