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

void my_alloc_tensor(float **t, int D0, int D1, int D2, int D3) {
  *t = (float *) aligned_alloc(32, sizeof(float) * D0 * D1 * D2 * D3);
  if (*t == NULL) {
    printf("Failed to allocate memory for matrix.\n");
    exit(0);
  }
}

int min(int a, int b) { return a < b ? a : b; }

void compute_conv(int is, int ie) {
//	int tNums = omp_get_num_threads(); 
//#pragma omp parallel for num_threads(tNums)
#pragma omp parallel for collapse(3) num_threads(100) schedule(dynamic)
  for (int n = is; n < ie; ++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;
        }
      }
    }
  }
}

void compute_conv_org() {
//	int tNums = omp_get_num_threads(); 
//#pragma omp parallel for num_threads(tNums)
#pragma omp parallel for collapse(3) num_threads(100) schedule(dynamic)
  for (int n = 0; n < N; ++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;
        }
      }
    }
  }
}

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;

  OH = (H + 2 * pad - dilation * (R - 1) - 1) / stride + 1;
  OW = (W + 2 * pad - dilation * (S - 1) - 1) / stride + 1;
	
	if (mpi_world_size > 1){
	  // Calculate is and ie redundantly on every processes
	  int is[mpi_world_size], ie[mpi_world_size];
	  for (int i = 0; i < mpi_world_size; i++) {
	    is[i] = N / mpi_world_size * i;
	    ie[i] = N / mpi_world_size * (i + 1);
	  }
	  ie[mpi_world_size - 1] = N;
		
		if (mpi_rank != 0) {
	    my_alloc_tensor(&input, N, C, H, W);
	    my_alloc_tensor(&output, N, K, OH, OW);
	    my_alloc_tensor(&filter, K, C, R, S);
	  }
	
	  // Scatter A
	  if (mpi_rank == 0) {
	    for (int i = 1; i < mpi_world_size; i++) {
	      MPI_Send(input+is[i]*C*H*W, (ie[i]-is[i])*C*H*W, MPI_FLOAT, i, 0, MPI_COMM_WORLD);
	    }
	  } 
		else {
	    MPI_Recv(input+is[mpi_rank]*C*H*W, (ie[mpi_rank]-is[mpi_rank])*C*H*W, MPI_FLOAT, 0, 0, MPI_COMM_WORLD, nullptr);
	  }
	
	  // Broadcast B
	  MPI_Bcast(filter, K*C*R*S, MPI_FLOAT, 0, MPI_COMM_WORLD);
	  compute_conv(is[mpi_rank], ie[mpi_rank]);
	
	  // Gather C
	  if (mpi_rank == 0) {
	    for (int i = 1; i < mpi_world_size; i++) {
	      MPI_Recv(output+is[i]*K*OH*OW, (ie[i]-is[i])*K*OH*OW, MPI_FLOAT, i, 0, MPI_COMM_WORLD, nullptr);
	    }
	  } 
		else {
	    MPI_Send(output+is[mpi_rank]*K*OH*OW, (ie[mpi_rank]-is[mpi_rank])*K*OH*OW, MPI_FLOAT, 0, 0, MPI_COMM_WORLD);
	  }
	}
	else {
		compute_conv_org();
	}
}

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) {
}