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

int num_threads = 40;

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

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 size[2];
	MPI_Request request;
	MPI_Status status;

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

	if (mpi_world_size == 2) 
		size[1] = _N / 2;
	else 
		size[1] = 0;

	size[0] = _N - size[1];

  int OH = (_H + 2 * _pad - dilation * (_R - 1) - 1) / _stride + 1;
  int 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);
	}

	int n, k, oh, ow, c, r, s;
	#pragma omp parallel for num_threads(num_threads) private(n,k,oh,ow,c,r,s) collapse(3) schedule(dynamic)
  for (n = 0; n < size[mpi_rank]; ++n) {
    for (k = 0; k < _K; ++k) {
      for (oh = 0; oh < OH; ++oh) {
        for (ow = 0; ow < OW; ++ow) {
          float o = 0.f;
          for (c = 0; c < _C; ++c) {
            for (r = 0; r < _R; ++r) {
              for (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;
              } // 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) {
}