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

extern void cuda_mem_init(int _N, int _C, int _H, int _W, int _K, int _R, int _S, int _pad, int _dilation, int _stride);
extern void cuda_conv_init(float *input, float *filter, float *output);
extern void cuda_conv_final(float *output);
extern void cuda_conv();

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;


#define MAX_NODE 2

static int size[MAX_NODE], offset[MAX_NODE];

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) {
  MPI_Request request;
  MPI_Status status;

  if(mpi_rank == 0) {
    input = _input;
    output = _output;
    filter = _filter;
   
    for (int node=1; node < mpi_world_size; node++) {
      MPI_Isend(&input[offset[node]*C*H*W], size[node]*C*H*W, MPI_FLOAT, node, 0, MPI_COMM_WORLD, &request);
      MPI_Isend(filter, K*C*R*S, MPI_FLOAT, node, 0, MPI_COMM_WORLD, &request);
    }
  }
  else {
    MPI_Recv(input, size[mpi_rank]*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);
  }
 
  cuda_conv_init(input, filter, output); 

  cuda_conv(); 
  
  cuda_conv_final(output);

  if(mpi_rank == 0) {
    for (int node=1; node < mpi_world_size; node++) {
      MPI_Recv(&output[offset[node]*K*OH*OW], size[node]*K*OH*OW, MPI_FLOAT, node, 0, MPI_COMM_WORLD, &status);
    }
  }
  else {
    MPI_Isend(output, size[mpi_rank]*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;

  OH = (H + 2 * pad - dilation * (R - 1) - 1) / stride + 1;
  OW = (W + 2 * pad - dilation * (S - 1) - 1) / stride + 1;
 
  MPI_Comm_rank(MPI_COMM_WORLD, &mpi_rank);
  MPI_Comm_size(MPI_COMM_WORLD, &mpi_world_size);
 
  //Calc partitioned size
  for (int i=0; i < mpi_world_size; i++) {
    int st = i * (N/mpi_world_size);
    int ed = i == mpi_world_size-1 ? N : (i+1)*(N/mpi_world_size);

    size[i] = ed - st;
    offset[i] = st;
  }
  
  if(mpi_rank != 0) {
    alloc_tensor(&input, size[mpi_rank], C, H, W);
    alloc_tensor(&output, size[mpi_rank], K, OH, OW);
    alloc_tensor(&filter, K, C, R, S);
  }
 
  cuda_mem_init(size[mpi_rank], C, H, W, K, R, S, pad, dilation, stride);
}

void convolution_final(
    int _N, int _C, int _H, int _W,
    int _K, int _R, int _S,
    int _pad, int _dilation, int _stride) {

}