chundoong-lab-ta/SamsungDS22/submissions/final/jj15.kim/tmp-A/convolution.cpp

#include "convolution.h"
#include <mpi.h>
#include <stdio.h>
#include "omp.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;

static int num_threads = 100;

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 req;
	MPI_Status status;
	int size[2];

	size[0] = (mpi_world_size == 2)? _N / mpi_world_size : _N;
	size[1] = _N - size[0];

  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 && mpi_world_size == 2) {
		MPI_Isend(&input[size[0] * C * H * W], size[1] * C * H * W, MPI_FLOAT, 1, 0, MPI_COMM_WORLD, &req);
		MPI_Isend(filter, K * C * R * S, MPI_FLOAT, 1, 0, MPI_COMM_WORLD, &req);
	}
	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);
	}

	//omp_set_num_threads(num_threads);
	if (dilation == 1 && pad == 0 && stride == 1 && !(S % 16)) {
#pragma omp parallel for num_threads(num_threads) collapse(2) schedule(dynamic)
  	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) {
						__m256 vout0, vout1;
						vout0 = _mm256_setzero_ps();
						vout1 = _mm256_setzero_ps();
						for (int c = 0; c < C; ++c) {
							__m256 vin0, vin1, vfi0, vfi1;
							for (int r = 0; r < R; ++r) {
								for (int s = 0; s < S; s = s + 16) {
									int h = oh + r;
									int w = ow + s;
									vfi0 = _mm256_load_ps(&filter[k * C * R * S + c * R * S + r * S + s]);
									vfi1 = _mm256_load_ps(&filter[k * C * R * S + c * R * S + r * S + s + 8]);
									vin0 = _mm256_load_ps(&input[n * C * H * W + c * H * W + h * W + w]);
									vin1 = _mm256_load_ps(&input[n * C * H * W + c * H * W + h * W + w + 8]);
									vout0 = _mm256_fmadd_ps(vfi0, vin0, vout0);
									vout1 = _mm256_fmadd_ps(vfi1, vin1, vout1);
								}
							}
						}
						vout0 = _mm256_add_ps(vout0, vout1);
						float *tmp = (float*)&vout0;
          	output[n * K * OH * OW + k * OH * OW + oh * OW + ow] = 
							tmp[0] + tmp[1] + tmp[2] + tmp[3] + tmp[4] + tmp[5] + tmp[6] + tmp[7];
        	}
				}
      }
		}
  }
  else {
#pragma omp parallel for num_threads(num_threads) collapse(3) schedule(dynamic)
    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 && 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, &req);
	}

}

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) {
}