#include "convolution.h"
#include "util.h"
#include <mpi.h>
#include <stdio.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 = 100;

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 N_start[mpi_world_size];
  int N_end[mpi_world_size];
	MPI_Request request;
	MPI_Status status;

	for(int i=0; i<mpi_world_size; i++){
		N_start[i] = _N/mpi_world_size*i;
		N_end[i] = _N/mpi_world_size*(i+1);
	}
  N_end[mpi_world_size-1] = _N;

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

  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[N_start[1]*C*H*W],(N_end[1]-N_start[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, (N_end[1]-N_start[1]),C,H,W);
    //alloc_tensor(&input, N,C,H,W);
		alloc_tensor(&output, (N_end[1]-N_start[1]),K,OH,OW);
		//alloc_tensor(&output, N,K,OH,OW);
		alloc_tensor(&filter, K,C,R,S);
		//MPI_Recv(&input[N_start[1]*C*H*W],(N_end[1]-N_start[1])*C*H*W,MPI_FLOAT,0,0,MPI_COMM_WORLD,&status);
		MPI_Recv(input,(N_end[1]-N_start[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);
	}

  
  #pragma omp parallel for num_threads(num_threads) collapse(3) schedule(dynamic)
    //for (int n = N_start[mpi_rank]; n < N_end[mpi_rank]; ++n) {
    for (int n = 0; n < (N_end[mpi_rank]-N_start[mpi_rank]); ++n) {
      for (int k = 0; k < K; ++k) {
        for (int oh = 0; oh < OH; ++oh) {
          for (int ow = 0; ow < OW; ++ow) {
            float o = 0.f;
            for (int c = 0; c < C; ++c) {
              for (int r = 0; r < R; ++r) {
                for (int s = 0; s < S; ++s) {
                  int h = oh * stride - pad + r * dilation;
                  int w = ow * stride - pad + 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 + oh * OW + ow] = o;
          }
        }
      }
    }

		if(mpi_rank == 0 && mpi_world_size == 2){
		  MPI_Recv(&output[N_start[1]*K*OH*OW],(N_end[1]-N_start[1])*K*OH*OW, MPI_FLOAT,1,0,MPI_COMM_WORLD, &status);
		}
		else if(mpi_world_size == 2){
		  //MPI_Isend(&output[N_start[1]*K*OH*OW],(N_end[1]-N_start[1])*K*OH*OW, MPI_FLOAT,0,0,MPI_COMM_WORLD, &request);
		  MPI_Isend(output,(N_end[1]-N_start[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) {
}