192 lines
9 KiB
C++
192 lines
9 KiB
C++
#include "dnn.hpp"
|
|
#include <omp.h>
|
|
|
|
void run_omp(MNIST& D, unsigned num_threads) {
|
|
|
|
// Create a task flow graph
|
|
const auto iter_num = D.images.rows()/D.batch_size;
|
|
const auto num_storage = num_threads;
|
|
|
|
// number of concurrent shuffle tasks
|
|
const auto num_par_shf = std::min(num_storage, D.epoch);
|
|
std::vector<Eigen::MatrixXf> mats(num_par_shf, D.images);
|
|
std::vector<Eigen::VectorXi> vecs(num_par_shf, D.labels);
|
|
|
|
const int num_layers = D.acts.size();
|
|
|
|
// Propagation per epoch
|
|
const auto prop_per_e = num_layers*iter_num;
|
|
|
|
auto dep_s = new int [D.epoch];
|
|
auto dep_f = new int [D.epoch * iter_num];
|
|
auto dep_b = new int [D.epoch * prop_per_e];
|
|
auto dep_u = new int [D.epoch * prop_per_e];
|
|
|
|
omp_set_num_threads(num_threads);
|
|
|
|
#pragma omp parallel
|
|
{
|
|
#pragma omp single
|
|
{
|
|
for(size_t e=0; e<D.epoch; e++) {
|
|
// Shuffle Tasks
|
|
if(e < num_par_shf) {
|
|
#pragma omp task depend (out: dep_s[e]) firstprivate(e, num_par_shf) shared(D, mats, vecs)
|
|
{
|
|
if(e != 0) {
|
|
D.shuffle(mats[e%num_par_shf], vecs[e%num_par_shf], D.images.rows());
|
|
}
|
|
}
|
|
}
|
|
else {
|
|
#pragma omp task depend (in: dep_s[e-num_par_shf], dep_b[(1+e-num_par_shf)*prop_per_e-num_layers]) depend (out: dep_s[e]) firstprivate(e, num_par_shf) shared(D, mats, vecs)
|
|
{
|
|
D.shuffle(mats[e%num_par_shf], vecs[e%num_par_shf], D.images.rows());
|
|
}
|
|
}
|
|
|
|
// DNN operations
|
|
for(size_t i=0; i<iter_num; i++) {
|
|
// Forward tasks
|
|
if(e == 0) {
|
|
if(i == 0) {
|
|
// The first task!!
|
|
#pragma omp task depend (in: dep_s[e]) depend (out: dep_f[i]) firstprivate(i, e, num_par_shf) shared(D, mats, vecs)
|
|
{
|
|
forward_task(D, i, e%num_par_shf, mats, vecs);
|
|
}
|
|
}
|
|
else {
|
|
// use openmp array sections syntax [lower_bound: length]
|
|
//#pragma omp task depend (in: dep_u[(i-1)*num_layers: num_layers]) depend (out: dep_f[i]) firstprivate(i, e, num_par_shf) shared(D, mats, vecs)
|
|
|
|
switch(num_layers) {
|
|
case 2:
|
|
#pragma omp task depend (in: dep_u[(i-1)*num_layers], dep_u[(i-1)*num_layers+1]) depend (out: dep_f[i]) firstprivate(i, e, num_par_shf) shared(D, mats, vecs)
|
|
{
|
|
forward_task(D, i, e%num_par_shf, mats, vecs);
|
|
}
|
|
break;
|
|
case 3:
|
|
#pragma omp task depend (in: dep_u[(i-1)*num_layers], dep_u[(i-1)*num_layers+1], dep_u[(i-1)*num_layers+2]) depend (out: dep_f[i]) firstprivate(i, e, num_par_shf) shared(D, mats, vecs)
|
|
{
|
|
forward_task(D, i, e%num_par_shf, mats, vecs);
|
|
}
|
|
break;
|
|
case 4:
|
|
#pragma omp task depend (in: dep_u[(i-1)*num_layers], dep_u[(i-1)*num_layers+1], dep_u[(i-1)*num_layers+2], dep_u[(i-1)*num_layers+3]) depend (out: dep_f[i]) firstprivate(i, e, num_par_shf) shared(D, mats, vecs)
|
|
{
|
|
forward_task(D, i, e%num_par_shf, mats, vecs);
|
|
}
|
|
break;
|
|
case 5:
|
|
#pragma omp task depend (in: dep_u[(i-1)*num_layers], dep_u[(i-1)*num_layers+1], dep_u[(i-1)*num_layers+2], dep_u[(i-1)*num_layers+3], dep_u[(i-1)*num_layers+4]) depend (out: dep_f[i]) firstprivate(i, e, num_par_shf) shared(D, mats, vecs)
|
|
{
|
|
forward_task(D, i, e%num_par_shf, mats, vecs);
|
|
}
|
|
break;
|
|
default: assert(false); break;
|
|
}
|
|
}
|
|
}
|
|
else {
|
|
if(i == 0) {
|
|
switch(num_layers) {
|
|
case 2:
|
|
#pragma omp task depend (in: dep_s[e], dep_u[e*prop_per_e - num_layers], dep_u[e*prop_per_e - num_layers+1]) depend (out: dep_f[e*iter_num+i]) firstprivate(i, e, num_par_shf) shared(D, mats, vecs)
|
|
{
|
|
forward_task(D, i, e%num_par_shf, mats, vecs);
|
|
}
|
|
break;
|
|
case 3:
|
|
#pragma omp task depend (in: dep_s[e], dep_u[e*prop_per_e - num_layers], dep_u[e*prop_per_e - num_layers+1], dep_u[e*prop_per_e - num_layers+2]) depend (out: dep_f[e*iter_num+i]) firstprivate(i, e, num_par_shf) shared(D, mats, vecs)
|
|
{
|
|
forward_task(D, i, e%num_par_shf, mats, vecs);
|
|
}
|
|
break;
|
|
case 4:
|
|
#pragma omp task depend (in: dep_s[e], dep_u[e*prop_per_e - num_layers], dep_u[e*prop_per_e - num_layers+1], dep_u[e*prop_per_e - num_layers+2], dep_u[e*prop_per_e - num_layers+3]) depend (out: dep_f[e*iter_num+i]) firstprivate(i, e, num_par_shf) shared(D, mats, vecs)
|
|
{
|
|
forward_task(D, i, e%num_par_shf, mats, vecs);
|
|
}
|
|
break;
|
|
case 5:
|
|
#pragma omp task depend (in: dep_s[e], dep_u[e*prop_per_e - num_layers], dep_u[e*prop_per_e - num_layers+1], dep_u[e*prop_per_e - num_layers+2], dep_u[e*prop_per_e - num_layers+3], dep_u[e*prop_per_e - num_layers+4]) depend (out: dep_f[e*iter_num+i]) firstprivate(i, e, num_par_shf) shared(D, mats, vecs)
|
|
{
|
|
forward_task(D, i, e%num_par_shf, mats, vecs);
|
|
}
|
|
break;
|
|
default: assert(false); break;
|
|
}
|
|
//#pragma omp task depend (in: dep_s[e], dep_u[e*prop_per_e - num_layers: num_layers]) depend (out: dep_f[e*iter_num+i]) firstprivate(i, e, num_par_shf) shared(D, mats, vecs)
|
|
|
|
}
|
|
else {
|
|
switch(num_layers) {
|
|
case 2:
|
|
#pragma omp task depend (in: dep_u[e*prop_per_e+(i-1)*num_layers], dep_u[e*prop_per_e+(i-1)*num_layers+1]) depend (out: dep_f[e*iter_num+i]) firstprivate(i ,e, num_par_shf) shared(D, mats, vecs)
|
|
{
|
|
forward_task(D, i, e%num_par_shf, mats, vecs);
|
|
}
|
|
break;
|
|
case 3:
|
|
#pragma omp task depend (in: dep_u[e*prop_per_e+(i-1)*num_layers], dep_u[e*prop_per_e+(i-1)*num_layers+1], dep_u[e*prop_per_e+(i-1)*num_layers+2]) depend (out: dep_f[e*iter_num+i]) firstprivate(i ,e, num_par_shf) shared(D, mats, vecs)
|
|
{
|
|
forward_task(D, i, e%num_par_shf, mats, vecs);
|
|
}
|
|
break;
|
|
case 4:
|
|
#pragma omp task depend (in: dep_u[e*prop_per_e+(i-1)*num_layers], dep_u[e*prop_per_e+(i-1)*num_layers+1], dep_u[e*prop_per_e+(i-1)*num_layers+2], dep_u[e*prop_per_e+(i-1)*num_layers+3]) depend (out: dep_f[e*iter_num+i]) firstprivate(i ,e, num_par_shf) shared(D, mats, vecs)
|
|
{
|
|
forward_task(D, i, e%num_par_shf, mats, vecs);
|
|
}
|
|
break;
|
|
case 5:
|
|
#pragma omp task depend (in: dep_u[e*prop_per_e+(i-1)*num_layers], dep_u[e*prop_per_e+(i-1)*num_layers+1], dep_u[e*prop_per_e+(i-1)*num_layers+2], dep_u[e*prop_per_e+(i-1)*num_layers+3], dep_u[e*prop_per_e+(i-1)*num_layers+4]) depend (out: dep_f[e*iter_num+i]) firstprivate(i ,e, num_par_shf) shared(D, mats, vecs)
|
|
{
|
|
forward_task(D, i, e%num_par_shf, mats, vecs);
|
|
}
|
|
break;
|
|
default: assert(false); break;
|
|
}
|
|
//#pragma omp task depend (in: dep_u[e*prop_per_e+(i-1)*num_layers: num_layers]) depend (out: dep_f[e*iter_num+i]) firstprivate(i ,e, num_par_shf) shared(D, mats, vecs)
|
|
}
|
|
}
|
|
|
|
// Backward tasks
|
|
for(int j=num_layers-1; j>=0; j--) {
|
|
if(j == num_layers-1) {
|
|
#pragma omp task depend (in: dep_f[e*iter_num + i]) depend (out: dep_b[e*prop_per_e + i*num_layers + j]) firstprivate(j, e, num_par_shf) shared(D, mats)
|
|
{
|
|
backward_task(D, j, e%num_par_shf, mats);
|
|
}
|
|
}
|
|
else {
|
|
#pragma omp task depend (in: dep_b[e*prop_per_e + i*num_layers + j + 1]) depend (out: dep_b[e*prop_per_e + i*num_layers + j]) firstprivate(j, e, num_par_shf) shared(D, mats)
|
|
{
|
|
backward_task(D, j, e%num_par_shf, mats);
|
|
}
|
|
}
|
|
}
|
|
|
|
// Update tasks
|
|
for(int j=num_layers-1; j>=0; j--) {
|
|
#pragma omp task depend (in: dep_b[e*prop_per_e + i*num_layers + j]) depend (out: dep_u[e*prop_per_e + i*num_layers + j]) firstprivate(j) shared(D)
|
|
{
|
|
D.update(j);
|
|
}
|
|
}
|
|
|
|
} // End of one iteration
|
|
} // End of one epoch
|
|
|
|
#pragma omp taskwait
|
|
} // End of omp single
|
|
} // End of omp parallel
|
|
|
|
delete [] dep_s;
|
|
delete [] dep_f;
|
|
delete [] dep_b;
|
|
delete [] dep_u;
|
|
}
|
|
|