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


#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);
  }
  

  if (pad == 0 && dilation == 1 && stride == 1 && (S % 16 == 0)){
    #pragma omp parallel for collapse(2) num_threads(100)
    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) {
            __m512 vo = _mm512_setzero_ps();
            for (int c = 0; c < C; ++c) {
              __m512 vi, vf;
              for (int r = 0; r < R; ++r) {
                for (int s = 0; s < S; s+=16) {
                  int h = oh + r;
                  int w = ow + s;
                  vf = _mm512_loadu_ps(&filter[k*C*R*S+c*R*S+r*S+s]);
                  vi = _mm512_loadu_ps(&input[n*C*H*W+c*H*W+h*W+w]);
                  vo = _mm512_fmadd_ps(vf,vi,vo);
                }
              }
            }
            output[n*K*OH*OW+k*OH*OW+oh*OW+ow] = _mm512_reduce_add_ps(vo);
          }
        }
      }
    }
  }
  else {
    #pragma omp parallel for collapse(3) num_threads(100)
    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;
                }
              }
            }
            output[n * K * OH * OW + k * OH * OW + oh * OW + ow] = o;
          }
        }
      }
    }
  }

  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;

  MPI_Comm_rank(MPI_COMM_WORLD, &mpi_rank);
  MPI_Comm_size(MPI_COMM_WORLD, &mpi_world_size);
  
  OH = (H + 2 * pad - dilation * (R - 1) - 1) / stride + 1;
  OW = (W + 2 * pad - dilation * (S - 1) - 1) / stride + 1;
  
  //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);
  }
  
}

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