168 lines
4.5 KiB
C++
168 lines
4.5 KiB
C++
#include "convolution.h"
|
|
#include <mpi.h>
|
|
#include <stdio.h>
|
|
#include "util.h"
|
|
|
|
#define MAX_NODE 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;
|
|
|
|
static int min(int x, int y){
|
|
return x < y ? x: y;
|
|
}
|
|
#if 0
|
|
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;
|
|
|
|
OH = (H + 2 * pad - dilation * (R - 1) - 1) / stride + 1;
|
|
OW = (W + 2 * pad - dilation * (S - 1) - 1) / stride + 1;
|
|
|
|
if (mpi_rank == 0) {
|
|
|
|
for (int n = 0; n < N; ++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;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
}
|
|
}
|
|
#else
|
|
|
|
int num_threads = 80;
|
|
|
|
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){
|
|
int size[MAX_NODE];
|
|
|
|
int kblock = 4;
|
|
int rblock = 16;
|
|
int sblock = 16;
|
|
int cblock = 24;
|
|
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){
|
|
for(int kk=0; kk< _K; kk+=kblock){
|
|
for(int oh=0; oh < OH; ++oh){
|
|
for(int ow=0; ow< OW; ++ow){
|
|
for(int k= kk; k < min(_K, kk + kblock); k++){
|
|
float o = 0;
|
|
for(int cc=0; cc< _C;cc+=cblock){
|
|
for(int rr = 0; rr< _R; rr+=rblock){
|
|
for(int ss=0; ss< _S; ss+=sblock){
|
|
for(int c = cc; c < min(_C, cc + cblock); c++){
|
|
for(int r= rr; r < min(_R, rr + rblock); r++){
|
|
for(int s=ss; s< min(_S, ss + sblock); 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);
|
|
}
|
|
}
|
|
|
|
|
|
#endif
|
|
|
|
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) {
|
|
}
|