chundoong-lab-ta/SamsungDS22/submissions/final/hj614.yoo/A/convolution.cpp

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

#define VEC_BYTE (32)
#define VEC_SIZE (VEC_BYTE>>2)
#define VEC_NUM (2)
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 sn, en, firstSize, modN;

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;
  int vec = S % (VEC_SIZE * VEC_NUM);
  MPI_Request request[2];
  MPI_Status status[2];

  if (N > 1)
  {
    if (mpi_rank == 0)
    {
      int snTmp = firstSize;
      int enTmp = N;
      MPI_Isend(input+snTmp*C*H*W, (enTmp-snTmp)*C*H*W, MPI_FLOAT, 1, 0, MPI_COMM_WORLD, &request[0]);
      MPI_Isend(filter, K*C*R*S, MPI_FLOAT, 1, 0, MPI_COMM_WORLD, &request[1]);
    }
    else
    {
      if (modN > 0)
      {
        alloc_tensor(&input, modN, C, H, W);
        alloc_tensor(&output, modN, K, OH, OW);
        alloc_tensor(&filter, K, C, R, S);
      }
      MPI_Irecv(input, modN*C*H*W, MPI_FLOAT, 0, 0, MPI_COMM_WORLD, &request[0]);
      MPI_Irecv(filter, K*C*R*S, MPI_FLOAT, 0, 0, MPI_COMM_WORLD, &request[1]);
    }
    MPI_Waitall(2,request, status);
  }

  if (vec == 0 && stride == 1 && pad == 0 && dilation == 1)
  {
    #pragma omp parallel for collapse(4) schedule(dynamic, 4)
    for (int n = 0; n < modN; ++n) {
      for (int k = 0; k < K; ++k) {
        for (int oh = 0; oh < OH; ++oh) {
          for (int ow = 0; ow < OW; ++ow) {
            __m256 vo[VEC_NUM];
            vo[0] = _mm256_setzero_ps();
            vo[1] = _mm256_setzero_ps();
            for (int c = 0; c < C; ++c) {
              __m256 vi[VEC_NUM], vf[VEC_NUM];
              for (int r = 0; r < R; r=r+1) {
                for (int s = 0; s < S; s=s+(VEC_SIZE * VEC_NUM)) {
                  int h = oh + r;
                  int w = ow + s;
                  vf[0] = _mm256_load_ps(&filter[k*C*R*S+c*R*S+r*S+s]);
                  vf[1] = _mm256_load_ps(&filter[k*C*R*S+c*R*S+r*S+s+8]);
                  vi[0] = _mm256_load_ps(&input[n*C*H*W+c*H*W+h*W+w]);
                  vi[1] = _mm256_load_ps(&input[n*C*H*W+c*H*W+h*W+w+8]);
                  vo[0] = _mm256_fmadd_ps(vf[0],vi[0],vo[0]);
                  vo[1] = _mm256_fmadd_ps(vf[1],vi[1],vo[1]);
                }
              }
            }
            vo[0] = _mm256_add_ps(vo[0],vo[1]);
            float *out = (float *)&vo[0];
            out[0] += out[1];
            out[2] += out[3];
            out[4] += out[5];
            out[6] += out[7];
            out[0] += out[2];
            out[4] += out[6];
            out[0] += out[4];
            //alloc_tensor(&out, VEC_SIZE, 1, 1, 1);
            //_mm256_store_ps(out,vo[0]);
            output[n*K*OH*OW+k*OH*OW+oh*OW+ow] = out[0];
            //free(out);
          }
        }
      }
    }
  }
  else
  {
    #pragma omp parallel for collapse(4) schedule(dynamic, 4)
    for (int n = 0; n < modN; ++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 (N > 1)
  {
    if (mpi_rank == 0)
    {
      int snTmp = firstSize;
      int enTmp = N;
      MPI_Irecv(output+snTmp*K*OH*OW, (enTmp-snTmp)*K*OH*OW, MPI_FLOAT, 1, 0, MPI_COMM_WORLD, &request[0]);
    }
    else
    {
      MPI_Isend(output, modN*K*OH*OW, MPI_FLOAT, 0, 0, MPI_COMM_WORLD, &request[0]);
    }
    MPI_Wait(&request[0], &status[0]);
  }
}

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);
  firstSize = (N + 1) / mpi_world_size;
  if (mpi_rank == 0)
  {
    sn = 0;
    en = firstSize;
    modN = firstSize;
  }
  else
  {
    sn = firstSize;
    en = N;
    modN = N - firstSize;
  }
}

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