#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;

int num_threads = 100;

void convolution1(
    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) {

  int size[2];
  MPI_Request request;
  MPI_Status status;

  input = _input;
  output = _output;
  filter = _filter;

  if (mpi_world_size == 2)
          size[1] = _N / 2;
  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);
  }
#pragma omp parallel for num_threads(num_threads) collapse(3) schedule(dynamic)
    for (int n = 0; n < size[mpi_rank]; ++n) {
//#pragma omp parallel for num_threads(num_threads) collapse(3) schedule(dynamic)
      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 && 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 convolution8(
    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) {

  int size[2];
  MPI_Request request;
  MPI_Status status;

  input = _input;
  output = _output;
  filter = _filter;

  if (mpi_world_size == 2)
          size[1] = _N / 2;
  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);
  }

#pragma omp parallel for num_threads(num_threads) collapse(3) schedule(dynamic)
    for (int n = 0; n < size[mpi_rank]; ++n) {
//#pragma omp parallel for num_threads(num_threads) collapse(3) schedule(dynamic)
      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=s+8) {
                  int h = oh * stride - pad + r * dilation;
                  int w = ow * stride - pad + s * 8 * dilation;
                  int w1 = ow * stride - pad + (s *8 + 1) * dilation;
                  int w2 = ow * stride - pad + (s *8 + 2) * dilation;
                  int w3 = ow * stride - pad + (s *8 + 3) * dilation;
                  int w4 = ow * stride - pad + (s *8 + 4) * dilation;
                  int w5 = ow * stride - pad + (s *8 + 5) * dilation;
                  int w6 = ow * stride - pad + (s *8 + 6) * dilation;
                  int w7 = ow * stride - pad + (s *8 + 7) * 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 i1 = input[n * C * H * W + c * H * W + h * W + w1];
                  float i2 = input[n * C * H * W + c * H * W + h * W + w2];
                  float i3 = input[n * C * H * W + c * H * W + h * W + w3];
                  float i4 = input[n * C * H * W + c * H * W + h * W + w4];
                  float i5 = input[n * C * H * W + c * H * W + h * W + w5];
                  float i6 = input[n * C * H * W + c * H * W + h * W + w6];
                  float i7 = input[n * C * H * W + c * H * W + h * W + w7];
                  float f = filter[k * C * R * S + c * R * S + r * S + s*8];
                  float f1 = filter[k * C * R * S + c * R * S + r * S + s*8+1];
                  float f2 = filter[k * C * R * S + c * R * S + r * S + s*8+2];
                  float f3 = filter[k * C * R * S + c * R * S + r * S + s*8+3];
                  float f4 = filter[k * C * R * S + c * R * S + r * S + s*8+4];
                  float f5 = filter[k * C * R * S + c * R * S + r * S + s*8+5];
                  float f6 = filter[k * C * R * S + c * R * S + r * S + s*8+6];
                  float f7 = filter[k * C * R * S + c * R * S + r * S + s*8+7];
                  o += i * f;
                  o += i1 * f1;
                  o += i2 * f2;
                  o += i3 * f3;
                  o += i4 * f4;
                  o += i5 * f5;
                  o += i6 * f6;
                  o += i7 * f7;
                }
              }
            }
            output[n * K * OH * OW + k * OH * OW + oh * OW + ow] = o;
          }
        }
      }
    }

  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 convolutionsimd(
    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) {

  int size[2];
  MPI_Request request;
  MPI_Status status;

  input = _input;
  output = _output;
  filter = _filter;

  if (mpi_world_size == 2)
          size[1] = _N / 2;
  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);
  }
#pragma omp parallel for num_threads(num_threads) collapse(3) 
    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) {
          for (int ow = 0; ow < OW/8; ow=ow+8) {
            //float o = 0.f;
            __m256 o = _mm256_set_ps(0.,0.,0.,0.,0.,0.,0.,0.);
            for (int c = 0; c < C; ++c) {
              for (int r = 0; r < R; ++r) {
                for (int s = 0; s < S/8; s=s+8) {
                  int h = oh * stride - pad + r * dilation;
                  int w = ow * stride - pad + s * dilation;
                  /*int w = ow * stride - pad + s * 4 * dilation;
                  int w1 = ow * stride - pad + (s *4 + 1) * dilation;
                  int w2 = ow * stride - pad + (s *4 + 2) * dilation;
                  int w3 = ow * stride - pad + (s *4 + 3) * dilation;*/
                  if (h < 0 || h >= H || w < 0 || w >= W) continue;
                  __m256 i = _mm256_load_ps(&input[n * C * H * W + c * H * W + h * W + w]);
                  /*float i = input[n * C * H * W + c * H * W + h * W + w];
                  float i1 = input[n * C * H * W + c * H * W + h * W + w1];
                  float i2 = input[n * C * H * W + c * H * W + h * W + w2];
                  float i3 = input[n * C * H * W + c * H * W + h * W + w3]; */
                  __m256 f = _mm256_load_ps(&filter[k * C * R * S + c * R * S + r * S + s]);
                  /*float f = filter[k * C * R * S + c * R * S + r * S + s*4];
                  float f1 = filter[k * C * R * S + c * R * S + r * S + s*4+1];
                  float f2 = filter[k * C * R * S + c * R * S + r * S + s*4+2];
                  float f3 = filter[k * C * R * S + c * R * S + r * S + s*4+3]; */
                  o += i * f;
                }
              }
            }
            //output[n * K * OH * OW + k * OH * OW + oh * OW + ow] = o;
            output[n * K * OH * OW + k * OH * OW + oh * OW + ow] = o[0];
            output[n * K * OH * OW + k * OH * OW + oh * OW + ow+1] = o[1];
            output[n * K * OH * OW + k * OH * OW + oh * OW + ow+2] = o[2];
            output[n * K * OH * OW + k * OH * OW + oh * OW + ow+3] = o[3];
            output[n * K * OH * OW + k * OH * OW + oh * OW + ow+4] = o[4];
            output[n * K * OH * OW + k * OH * OW + oh * OW + ow+5] = o[5];
            output[n * K * OH * OW + k * OH * OW + oh * OW + ow+6] = o[6];
            output[n * K * OH * OW + k * OH * OW + oh * OW + ow+7] = o[7];
          }
        }
      }
    }

  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 convolution4(
    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) {

  int size[2];
  MPI_Request request;
  MPI_Status status;

  input = _input;
  output = _output;
  filter = _filter;

  if (mpi_world_size == 2) 
	  size[1] = _N / 2;
  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);
  }

#pragma omp parallel for num_threads(num_threads) collapse(3) schedule(dynamic)
//#pragma omp parallel for num_threads(num_threads) collapse(3) 
    for (int n = 0; n < size[mpi_rank]; ++n) {
//#pragma omp parallel for num_threads(num_threads) collapse(3) schedule(dynamic)
      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/4; ++s) {
                  int h = oh * stride - pad + r * dilation;
                  int w = ow * stride - pad + s * 4 * dilation;
                  int w1 = ow * stride - pad + (s *4 + 1) * dilation;
                  int w2 = ow * stride - pad + (s *4 + 2) * dilation;
                  int w3 = ow * stride - pad + (s *4 + 3) * 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 i1 = input[n * C * H * W + c * H * W + h * W + w1];
                  float i2 = input[n * C * H * W + c * H * W + h * W + w2];
                  float i3 = input[n * C * H * W + c * H * W + h * W + w3]; 
                  float f = filter[k * C * R * S + c * R * S + r * S + s*4];
                  float f1 = filter[k * C * R * S + c * R * S + r * S + s*4+1];
                  float f2 = filter[k * C * R * S + c * R * S + r * S + s*4+2];
                  float f3 = filter[k * C * R * S + c * R * S + r * S + s*4+3]; 
                  o += i * f;
                  o += i1 * f1;
                  o += i2 * f2;
                  o += i3 * f3;
                } 
              } 
            } 
            output[n * K * OH * OW + k * OH * OW + oh * OW + ow] = o;
          } 
        } 
      } 
    } 

  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(
    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) {

    //if(_S % 8 == 0)
    //    convolution8(_input,_output,_filter,_N,_C,_H,_W,_K,_R,_S,_pad,_dilation,_stride);
    
    //convolutionsimd(_input,_output,_filter,_N,_C,_H,_W,_K,_R,_S,_pad,_dilation,_stride);
    if(_S % 4 == 0)
        convolution4(_input,_output,_filter,_N,_C,_H,_W,_K,_R,_S,_pad,_dilation,_stride);
    else
        convolution1(_input,_output,_filter,_N,_C,_H,_W,_K,_R,_S,_pad,_dilation,_stride); 
}

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) {
}