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

#define NUM_THREADS 80
#define T_SIZE_OH 16
#define T_SIZE_OW 8

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;

MPI_Request request;
MPI_Status status;

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 divided_N = N / mpi_world_size;
  int modular_N = N % mpi_world_size;
  int stride_N[4];
  int offset[4];

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

  offset[0] = 0;
  stride_N[0] = divided_N;

  for (int i = 1; i < mpi_world_size; i++) {
  	if (i <= modular_N) stride_N[i] = divided_N + 1;
  	else stride_N[i] = divided_N;
  	offset[1] = divided_N;
  	if (i > 1) offset[i] = offset[i-1] + stride_N[i-1]; 
  }

  if (mpi_rank != 0) {
	alloc_tensor(&input, stride_N[mpi_rank], C, H, W);
	alloc_tensor(&output, stride_N[mpi_rank], K, OH, OW);
	alloc_tensor(&filter, K, C, R, S);
  }

  MPI_Bcast(filter, K*C*R*S, MPI_FLOAT, 0, MPI_COMM_WORLD);

  if (mpi_rank == 0) {
	for (int i = 1; i < mpi_world_size; i++)
  		MPI_Isend(&input[offset[i]*C*H*W], stride_N[i]*C*H*W, MPI_FLOAT, i, 1, MPI_COMM_WORLD, &request);
  }
  else {
		MPI_Recv(input, stride_N[mpi_rank]*C*H*W, MPI_FLOAT, 0, 1, MPI_COMM_WORLD, &status);
  }

//  if (mpi_rank == 0) {
 #pragma omp parallel for num_threads(NUM_THREADS) collapse(3) schedule(dynamic)
    for (int n = 0; n < stride_N[mpi_rank]; ++n) {
      for (int k = 0; k < K; ++k) {
//        for (int t_oh = 0; t_oh < OH; t_oh += T_SIZE_OH) {
// #pragma omp parallel for num_threads(NUM_THREADS) shared(N, K, OH, OW, R, S, C) schedule(dynamic)
        for (int oh = 0; oh < OH; ++oh) {
// #pragma omp parallel for num_threads(NUM_THREADS) shared(N, K, OH, OW, R, S, C) schedule(dynamic)
          for (int ow = 0; ow < OW; ++ow) {
//       for (int t_ow = 0; t_ow < OW; t_ow += T_SIZE_OW) {
//        for (int oh = t_oh; oh < t_oh+T_SIZE_OH; ++oh) {
//          for (int ow = t_ow; ow < t_ow+T_SIZE_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) {
	for (int i = 1; i < mpi_world_size; i++)
  		MPI_Recv(&output[offset[i]*K*OH*OW], stride_N[i]*K*OH*OW, MPI_FLOAT, i, 2, MPI_COMM_WORLD, &status);
  } 
  else {
  	MPI_Send(output, stride_N[mpi_rank]*K*OH*OW, MPI_FLOAT, 0, 2, MPI_COMM_WORLD);
  }
}
//}

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) {
}