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

#include "util.h"
#include <cstdio>
#include <cuda_runtime.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 MAX_NUM_GPU 4
#define TS 8

__global__ void conv(
        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 OH, int OW) {

    const int globalRow = blockDim.x * blockIdx.x + threadIdx.x;
    const int globalCol = blockDim.y * blockIdx.y + threadIdx.y;
    int n, k;

    n = globalCol/(_K*OW);
    k = (globalCol-n*(_K*OW))/OW;
    int col = (globalCol-n*(_K*OW))-k*OW;
    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 r = 0; r<_R; r++) {
            for (int s = 0; s < _S; ++s) {
                int h = start_row + r * _dilation;
                int w = start_col + s * _dilation;
                if (h < 0 || h >= _H || w < 0 || w >= _W) continue;
                float i = _input[n * _C * _H * _W + c * _H * _W + h * _W + w];
                float f = _filter[k * _C * _R * _S + c * _R * _S + r * _S + s];
                o += i * f;
            }
        }
    }

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

static float *input, *output, *filter;
static float *input_d[MAX_NUM_GPU];
static float *filter_d[MAX_NUM_GPU];
static float *output_d[MAX_NUM_GPU];
static int N, C, H, W;
static int K, R, S;
static int pad;
static int dilation;
static int stride;
static int mpi_rank, mpi_world_size;
int num_devices = 0;
static int size[2];
static int Mbegin[MAX_NUM_GPU], NN[MAX_NUM_GPU];
static int OH, OW;

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) {
  MPI_Request request;
  MPI_Status status;

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

  if (size[1] != 0) {
      if (mpi_rank == 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);
      } else {
          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] != 0) {
      for (int i = 0; i < num_devices; i++) {
          CUDA_CALL( cudaMemcpy(input_d[i], input + Mbegin[i]*C*H*W,
                      NN[i]*C*H*W * sizeof(float),
                      cudaMemcpyHostToDevice) );
          CUDA_CALL( cudaMemcpy(filter_d[i], filter, K*C*R*S * sizeof(float), cudaMemcpyHostToDevice) );
      }

      // Launch kernel on every GPU
      for (int i = 0; i < num_devices; i++) {
          dim3 gridDim((OH+TS-1)/TS, (NN[i]*K*OW + TS - 1)/TS, 1);
          dim3 blockDim(TS, TS, 1);

          CUDA_CALL( cudaSetDevice(i) );
          conv<<<gridDim, blockDim>>>(input_d[i], output_d[i], filter_d[i], NN[i], _C, _H, _W, _K, _R, _S, _pad, _dilation, _stride, OH, OW);
      }

      // Download C matrix from GPUs
      for (int i = 0; i < num_devices; i++) {
          CUDA_CALL( cudaMemcpy(output + Mbegin[i] * K*OH*OW, output_d[i],
                      NN[i] * K*OH*OW * sizeof(float),
                      cudaMemcpyDeviceToHost) );
      }
  }

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

      MPI_Wait(&request, &status);
  }
}

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

    if (mpi_world_size == 2 && _N > 4) 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++)
        {
            NN[i] = 1;
            Mbegin[i] = i;
        }
    } else {
        num_devices = MAX_NUM_GPU;
        int q = size[mpi_rank] / MAX_NUM_GPU;
        int r = size[mpi_rank] % MAX_NUM_GPU;
        int sum = 0;
        for (int i = 0 ; i < MAX_NUM_GPU ; i++) {
            NN[i] = q;
            Mbegin[i] = sum;
            if (i == MAX_NUM_GPU - 1) {
                NN[i] += r;
            }
            sum += NN[i]; 
        }
    }

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

    // possigle to MPI
    CUDA_CALL( cudaGetDeviceCount(&num_devices) );

    for (int i = 0; i < num_devices; i++) {
        cudaDeviceProp prop;
        CUDA_CALL( cudaGetDeviceProperties(&prop, i) );
    }
    if (num_devices <= 0) {
        exit(1);
    }

    // Allocate device memory for each GPU
    for (int i = 0; i < num_devices; i++) {
        CUDA_CALL( cudaSetDevice(i) );
        CUDA_CALL( cudaMalloc(&input_d[i], NN[i]*C*H*W * sizeof(float)) );
        CUDA_CALL( cudaMalloc(&filter_d[i], K*C*R*S * sizeof(float)) );
        CUDA_CALL( cudaMalloc(&output_d[i], NN[i]*K*OH*OW * 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) {
}