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

static float *input, *output, *filter;
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;

int num_threads = 120;
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) {
  input = _input;
  output = _output;
  filter = _filter;

  int size[2];
  MPI_Request request;
  MPI_Status status;
  if ( mpi_world_size ==2) size[1] = _N/2; //TBC
  else size[1] = 0;
  size[0] = N - size[1];

  
  OH = (H + 2 * pad - dilation * (R - 1) - 1) / stride + 1;
  OW = (W + 2 * pad - dilation * (S - 1) - 1) / stride + 1;
  
  if (mpi_rank == 0 && mpi_world_size ==2) 
  {
    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 if(mpi_world_size ==2)
  {
    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);
  }

  float i[8];
  float f[8];
//#pragma omp parallel for num_threads(num_threads) collapse(3) schedule(dynamic) 
//#pragma omp parallel for num_threads(num_threads) collapse(3) schedule(dynamic) 
#pragma omp parallel for num_threads(num_threads) collapse(3) schedule(dynamic) 
  for (int oh = 0; oh < OH; ++oh) {
    for (int ow = 0; ow < OW; ++ow) {
      for (int n = 0; n < size[mpi_rank]; ++n) {
  	for (int k = 0; k < K; ++k) {
          float o = 0.f;
          for (int c = 0; c < C; c=c+1) {
            for (int r = 0; r < R; ++r) {
              int h = oh * stride - pad + r * dilation;
              if (h < 0 || h >= H) continue;
              for (int s = 0; s < S; ++s) {
                int w = ow * stride - pad + s * dilation;
                if (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];
                  //i[0] = input[n * C * H * W +  c * H * W + h * W + w];
                  //f[0] = filter[k * C * R * S + c * R * S + r * S + s];
                  //i[1] = input[n * C * H * W + (c+1) * H * W + h * W + w];
                  //f[1] = filter[k * C * R * S +(c+1) * R * S + r * S + s];
                  //i[2] = input[n * C * H * W + (c+2)  * H * W + h * W + w];
                  //f[2] = filter[k * C * R * S +(c+2) * R * S + r * S + s];
                  //i[3] = input[n * C * H * W + (c+3) * H * W + h * W + w];
                  //f[3] = filter[k * C * R * S + (c+3) * R * S + r * S + s];
                  //i[4] = input[n * C * H * W + (c+4) * H * W + h * W + w];
                  //f[4] = filter[k * C * R * S + (c+4) * R * S + r * S + s];
                  //i[5] = input[n * C * H * W + (c+5) * H * W + h * W + w];
                  //f[5] = filter[k * C * R * S + (c+5) * R * S + r * S + s];
                  //i[6] = input[n * C * H * W + (c+6) * H * W + h * W + w];
                  //f[6] = filter[k * C * R * S + (c+6) * R * S + r * S + s];
                  //i[7] = input[n * C * H * W + (c+7) * H * W + h * W + w];
                  //f[7] = filter[k * C * R * S + (c+7) * R * S + r * S + s];
                  //o += i[0] * f[0] + i[1] * f[1] + i[2] * f[2] + i[3] * f[3] 
		  //        + i[4] * f[4] + i[5] * f[5] + i[6] * f[6] + i[7] * f[7] ;
		  o += i*f;
              }//s
            }//r
          }//c
          output[n * K * OH * OW + k * OH * OW + oh * OW + ow] = o;
        }//ow
      }//oh
    }//k
  }//n

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

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