#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;
static int num_threads;


#if 0
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_rank == 0) {

    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;
          }
        }
      }
    }

  }
}
#else

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;
  
  // Divide input to nodes (# of nodes: 1 ~ 2)
  int N_s[mpi_world_size], N_e[mpi_world_size];
  int K_s[mpi_world_size], K_e[mpi_world_size];
  if (N != 1) {
    for (int i = 0; i < mpi_world_size; i++) {
      N_s[i] = N / mpi_world_size * i;
      N_e[i] = N / mpi_world_size * (i + 1);
      K_s[i] = 0;
      K_e[i] = K;
    }
    N_e[mpi_world_size - 1] = N;
  }
  else if (K != 1) { // N = 1
    for (int i = 0; i < mpi_world_size; i++) {
      N_s[i] = 0;
      N_e[i] = N;
      K_s[i] = K / mpi_world_size * i;
      K_e[i] = K / mpi_world_size * (i + 1);
    }
    K_e[mpi_world_size - 1] = K;
  }
  else { // N = 1 & K = 1
    for (int i = 0; i < mpi_world_size; i++) {
      N_s[i] = 0;
      N_e[i] = N;
      K_s[i] = 0;
      K_e[i] = K;
    }
    if (mpi_world_size == 2) {  //Use single node
      N_e[1] = 0;
      K_e[1] = 0;
    }
  }

  // Scatter
  if (mpi_rank == 0) {
    for (int i = 1; i < mpi_world_size; i++) {
      MPI_Send(&input[N_s[i] * C * H * W], ((N_e[i] - N_s[i]) * C * H * W), MPI_FLOAT, i, 0, MPI_COMM_WORLD);
      MPI_Send(&filter[K_s[i] * C * R * S], ((K_e[i] - K_s[i]) * C * R * S), MPI_FLOAT, i, 0, MPI_COMM_WORLD);
    }
  }
  else {
    alloc_tensor(&input, N, C, H, W);
    alloc_tensor(&output, N, K, OH, OW);
    alloc_tensor(&filter, K, C, R, S);
    
    MPI_Recv(&input[N_s[mpi_rank] * C * H * W], ((N_e[mpi_rank] - N_s[mpi_rank]) * C * H * W), MPI_FLOAT, 0, 0, MPI_COMM_WORLD, nullptr);
    MPI_Recv(&filter[K_s[mpi_rank] * C * R * S], ((K_e[mpi_rank] - K_s[mpi_rank]) * C * R * S), MPI_FLOAT, 0, 0, MPI_COMM_WORLD, nullptr);
  }
  
  // Compute convolution
  #pragma omp parallel for num_threads(num_threads) collapse(3) schedule(dynamic)
  for (int n = N_s[mpi_rank]; n < N_e[mpi_rank]; ++n) {
    for (int k = K_s[mpi_rank]; k < K_e[mpi_rank]; ++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;
        }
      }
    }
  }

  // Gather output
  if (mpi_rank == 0) {
    for (int i = 1; i < mpi_world_size; i++) {
      if (N != 1) {
        MPI_Recv(&output[N_s[i] * K_e[i] * OH * OW], ((N_e[i] - N_s[i]) * (K_e[i] - K_s[i]) * OH * OW), MPI_FLOAT, i, 1, MPI_COMM_WORLD, nullptr);
      }
      else if (K != 1) {
        MPI_Recv(&output[N_e[i] * K_s[i] * OH * OW], ((N_e[i] - N_s[i]) * (K_e[i] - K_s[i]) * OH * OW), MPI_FLOAT, i, 1, MPI_COMM_WORLD, nullptr);
      }
      else {
        MPI_Recv(&output[N_e[i] * K_e[i] * OH * OW], ((N_e[i] - N_s[i]) * (K_e[i] - K_s[i]) * OH * OW), MPI_FLOAT, i, 1, MPI_COMM_WORLD, nullptr);
      }
    }
  } else {
    if (N != 1) {
      MPI_Send(&output[N_s[mpi_rank] * K_e[mpi_rank] * OH * OW], ((N_e[mpi_rank] - N_s[mpi_rank]) * (K_e[mpi_rank] - K_s[mpi_rank]) * OH * OW), MPI_FLOAT, 0, 1, MPI_COMM_WORLD);
    }
    else if (K != 1) {
      MPI_Send(&output[N_e[mpi_rank] * K_s[mpi_rank] * OH * OW], ((N_e[mpi_rank] - N_s[mpi_rank]) * (K_e[mpi_rank] - K_s[mpi_rank]) * OH * OW), MPI_FLOAT, 0, 1, MPI_COMM_WORLD);
    }
    else {
      MPI_Send(&output[N_e[mpi_rank] * K_e[mpi_rank] * OH * OW], ((N_e[mpi_rank] - N_s[mpi_rank]) * (K_e[mpi_rank] - K_s[mpi_rank]) * OH * OW), MPI_FLOAT, 0, 1, MPI_COMM_WORLD);
    }
  }
  
}


#endif

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;

  num_threads = 100;

  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) {
}