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

#define MAX_NUM_THREADS 100
#define NTILESIZE 32
#define KTILESIZE 4

//static int (int x, int y) { return (x < y)? x : y; }

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 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;

  /*---------- seperating for nodes ----------*/
  //MPI_Status status;
  //MPI_Request request = MPI_REQUEST_NULL;

  if (mpi_rank >= mpi_world_size) return;

  if (mpi_rank != 0){
    input = (float *) aligned_alloc(32, sizeof(float) * N*C*H*W);
    filter = (float *) aligned_alloc(32, sizeof(float) * K*C*R*S);
    output = (float *) aligned_alloc(32, sizeof(float) * N*K*OH*OW);
  }

  int ns[mpi_world_size], ne[mpi_world_size];
  for (int i=0; i<mpi_world_size; ++i){
    ns[i] = N / mpi_world_size * i;
    ne[i] = N / mpi_world_size * (i+1);
  }
  ne[mpi_world_size -1] = N;

  if (mpi_rank == 0){
    for (int i=1; i<mpi_world_size; ++i){
      MPI_Send(&input[ ns[i] * C*H*W ], (ne[i] - ns[i]) * C*H*W, MPI_FLOAT, i, 0, MPI_COMM_WORLD);
    }
  } else {
    MPI_Recv (&input[ ns[mpi_rank] * C*H*W], (ne[mpi_rank] - ns[mpi_rank]) * C*H*W, MPI_FLOAT, 0, 0, MPI_COMM_WORLD, nullptr);
  }
  MPI_Bcast(filter, K*C*R*S, MPI_FLOAT, 0, MPI_COMM_WORLD);

  /*---------- calculating ----------*/
  int num_threads = omp_get_num_threads();
  //for (int nn = ns[mpi_rank]; nn < ne[mpi_rank]; nn += NTILESIZE) {
  //for (int kk = 0; kk < K; kk += KTILESIZE) {
          //for (int n = nn; n < std::min(ne[mpi_rank], nn+NTILESIZE); ++n){
          //for (int k = kk; k < std::min(K, kk+KTILESIZE); ++k){
  //printf("mpi_rank: %d, ( %d ~ %d )\n",mpi_rank, ns[mpi_rank], ne[mpi_rank]);
#pragma omp parallel for collapse(3) num_threads (MAX_NUM_THREADS) schedule(dynamic)
  for (int n = ns[mpi_rank]; n < ne[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 = 0.0f;
          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;
        }
      }
    }
  }

  /*---------- merging output ----------*/
  if (mpi_rank == 0){
    for(int i=1; i<mpi_world_size; ++i){
      MPI_Recv (&output[ns[i] * K*OH*OW], (ne[i] - ns[i]) * K*OH*OW, MPI_FLOAT, i, 0, MPI_COMM_WORLD, nullptr);
    }
  } else {
    MPI_Send (&output[ns[mpi_rank] * K*OH*OW], (ne[mpi_rank] - ns[mpi_rank]) * K*OH*OW, MPI_FLOAT, 0, 0, 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) {


  if (mpi_rank >= mpi_world_size) return;

  if (mpi_rank != 0){
    free(input);
    free(filter);
    free(output);
  }
}