#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 = 160;
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);
  }
  int CHW = C*H*W;
  int HW = H*W;

if (C%8==0 && R==16 && S==16 && dilation==1 && pad==0 && stride==1)
{
//#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 n = 0; n < size[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=c+16) {
          for (int c = 0; c < C; c=c+8) {
            for (int r = 0; r < 16; r++) {
                //int h = oh + r;
                //if (h < 0 || h >= H) continue;
              for (int s = 0; s < 16; s++) {
                //int w = ow + s ;
                //if (w < 0 || w >= W) continue;
                float i0 = input[n * CHW + c * HW + (oh+r) * W + ow+s];
                float f0 = filter[k * C * 16 * 16 + c * 16 * 16 + r * 16 + s];
                float i1 = input[n * CHW + (c+1) * HW + (oh+r) * W + ow+s];
                float f1 = filter[k * C * 16 * 16 + (c+1) * 16 * 16 + r * 16 + s];
                float i2 = input[n * CHW + (c+2) * HW + (oh+r) * W + ow+s];
                float f2 = filter[k * C * 16 * 16 + (c+2) * 16 * 16 + r * 16 + s];
                float i3 = input[n * CHW + (c+3) * HW + (oh+r) * W + ow+s];
                float f3 = filter[k * C * 16 * 16 + (c+3) * 16 * 16 + r * 16 + s];
                float i4 = input[n * CHW + (c+4) * HW + (oh+r) * W + ow+s];
                float f4 = filter[k * C * 16 * 16 + (c+4) * 16 * 16 + r * 16 + s];
                float i5 = input[n * CHW + (c+5) * HW + (oh+r) * W + ow+s];
                float f5 = filter[k * C * 16 * 16 + (c+5) * 16 * 16 + r * 16 + s];
                float i6 = input[n * CHW + (c+6) * HW + (oh+r) * W + ow+s];
                float f6 = filter[k * C * 16 * 16 + (c+6) * 16 * 16 + r * 16 + s];
                float i7 = input[n * CHW + (c+7) * HW + (oh+r) * W + ow+s];
                float f7 = filter[k * C * 16 * 16 + (c+7) * 16 * 16 + r * 16 + s];

                //float i8 = input[n * C * H * W + (c+8) * H * W + h * W + w];
                //float f8 = filter[k * C * R * S + (c+8) * R * S + r * S + s];
                //float i9 = input[n * C * H * W + (c+9) * H * W + h * W + w];
                //float f9 = filter[k * C * R * S + (c+9) * R * S + r * S + s];
                //float i10 = input[n * C * H * W + (c+10) * H * W + h * W + w];
                //float f10 = filter[k * C * R * S + (c+10) * R * S + r * S + s];
                //float i11 = input[n * C * H * W + (c+11) * H * W + h * W + w];
                //float f11 = filter[k * C * R * S + (c+11) * R * S + r * S + s];
                //float i12 = input[n * C * H * W + (c+12) * H * W + h * W + w];
                //float f12 = filter[k * C * R * S + (c+12) * R * S + r * S + s];
                //float i13 = input[n * C * H * W + (c+13) * H * W + h * W + w];
                //float f13 = filter[k * C * R * S + (c+13) * R * S + r * S + s];
                //float i14 = input[n * C * H * W + (c+14) * H * W + h * W + w];
                //float f14 = filter[k * C * R * S + (c+14) * R * S + r * S + s];
                //float i15 = input[n * C * H * W + (c+15) * H * W + h * W + w];
                //float f15 = filter[k * C * R * S + (c+15) * R * S + r * S + s];

                //o += i0*f0 + i1*f1 + i2*f2 + i3*f3 + i4*f4 + i5*f5 + i6*f6 + i7*f7 
		//    +i8*f8 + i9*f9 + i10*f10 + i11*f11 + i12*f12 + i13*f13 + i14*f14 + i15*f15 ;
                o += i0*f0 + i1*f1 + i2*f2 + i3*f3 + i4*f4 + i5*f5 + i6*f6 + i7*f7;
              }//s
            }//r
          }//c
          output[n * K * OH * OW + k * OH * OW + oh * OW + ow] = o;
        }//ow
      }//oh
    }//k
  }//n
}
else
{
#pragma omp parallel for num_threads(num_threads) collapse(3) schedule(dynamic) 
  for (int n = 0; n < size[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;
              }//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) {
}