Mentions légales du service

Skip to content
Snippets Groups Projects
Commit 3330d6f1 authored by hhakim's avatar hhakim
Browse files

Implement a C++ threaded version of the previous OpenMP impl. of the... 

Implement a C++ threaded version of the previous OpenMP impl. of the parallelization of the Faust-matrix product.
parent d413efc4
No related branches found
No related tags found
No related merge requests found
Hide whitespace changes
Inline Side-by-side
misc/test/src/C++/unit/faust_transform_omp_mul.cpp.in
+ 15
− 8
View file @ 3330d6f1
......@@ -61,7 +61,7 @@ int main(int argc, char* argv[])
}
//RandFaustType t, unsigned int min_num_factors, unsigned int max_num_factors, unsigned int min_dim_size, unsigned int max_dim_size, float density=.1f, bool per_row=true
TransformHelper<FPP,Cpu>* th = TransformHelper<FPP,Cpu>::randFaust(DENSE, 50, 50, 256, 256, .5);
TransformHelper<FPP,Cpu>* th = TransformHelper<FPP,Cpu>::randFaust(DENSE, 15, 15, 256, 256, .5);
th->display();
MatDense<FPP,Cpu>* M = MatDense<FPP,Cpu>::randMat(256,256);
auto start = std::chrono::system_clock::now();
......@@ -71,16 +71,23 @@ int main(int argc, char* argv[])
std::cout << "Time "<< diff.count() << " s\n";
FM.Display();
cout << "FM norm: " << FM.norm() << endl;
th->set_mul_order_opt_mode(4);
// th->set_mul_order_opt_mode(4);
// start = std::chrono::system_clock::now();
// auto FM2 = th->multiply(*M);
// end = std::chrono::system_clock::now();
// diff = end-start;
// std::cout << "Time omp: "<< diff.count() << " s\n";
// FM2.Display();
// cout << "FM2 norm: " << FM.norm() << endl;
th->set_mul_order_opt_mode(5);
start = std::chrono::system_clock::now();
auto FM2 = th->multiply(*M);
auto FM3 = th->multiply(*M);
end = std::chrono::system_clock::now();
diff = end-start;
std::cout << "Time omp: "<< diff.count() << " s\n";
cout << "FM2 norm: " << FM.norm() << endl;
std::cout << "Time C++ threads: "<< diff.count() << " s\n";
FM3.Display();
cout << "FM3 norm: " << FM3.norm() << endl;
FM2.Display();
cout << FM2.norm() << endl;
return 0;
}
src/faust_linear_operator/CPU/faust_Transform.h
+ 7
− 1
View file @ 3330d6f1
......@@ -47,7 +47,9 @@
#include "faust_transform_algebra.h"
#include "faust_RefManager.h"
#include <thread>
#include <condition_variable>
#include <mutex>
//#include "faust_Vect.h"
//#include "faust_MatDense.h"
//#include "faust_MatSparse.h"
......@@ -79,6 +81,9 @@ namespace Faust
Faust::MatDense<FPP,Cpu> operator*(const Transform<FPP,Cpu>& f, const Faust::MatDense<FPP,Cpu>& M);
template<typename FPP>
void multiply_par_run(int nth, int thid, int num_per_th, int data_size, char opThis, std::vector<Faust::MatGeneric<FPP, Cpu>*>& data, std::vector<Faust::MatDense<FPP,Cpu>*>& mats, std::vector<std::thread*>& threads, std::mutex &, std::condition_variable &, int &);
template<typename FPP>
class Transform<FPP,Cpu> : public Faust::LinearOperator<FPP,Cpu>
{
......@@ -260,6 +265,7 @@ namespace Faust
return this->multiply(x,'N');
}
Faust::MatDense<FPP,Cpu> multiply_omp(const Faust::MatDense<FPP,Cpu> A, const char opThis) const;
Faust::MatDense<FPP,Cpu> multiply_par(const Faust::MatDense<FPP,Cpu> A, const char opThis) const;
/*!
......
src/faust_linear_operator/CPU/faust_Transform.hpp
+ 156
− 0
View file @ 3330d6f1
......@@ -1093,6 +1093,162 @@ Faust::MatDense<FPP,Cpu> Faust::Transform<FPP,Cpu>::multiply(const Faust::MatDen
}
template<typename FPP>
Faust::MatDense<FPP,Cpu> Faust::Transform<FPP,Cpu>::multiply_par(const Faust::MatDense<FPP,Cpu> A, const char opThis) const
{
int nth = 2; //TODO: https://en.cppreference.com/w/cpp/thread/thread/hardware_concurrency
int barrier_count = 0;
int rem_size;
Faust::MatDense<FPP, Cpu> *M;
std::vector<std::thread*> threads;
std::vector<Faust::MatDense<FPP,Cpu>*> mats(nth);
std::vector<Faust::MatGeneric<FPP, Cpu>*> data(this->data.size()+1);
std::mutex barrier_mutex;
std::condition_variable barrier_cv;
int data_size = this->data.size()+1; // counting A in addition to Transform factors
int num_per_th = data_size;
num_per_th /= nth;
num_per_th = num_per_th<2?2:num_per_th;
// recompute effective needed nth
nth = data_size / num_per_th;
// compute rem_size: the number of factors to treat afterward
rem_size = data_size - nth*num_per_th; // facts of ids data_size-rem_size to data_size-1
int i = 0;
for(auto ptr: this->data)
{
cout << "fac i=" << i << ": " << ptr << endl;
data[i++] = ptr;
}
data[i] = const_cast<Faust::MatDense<FPP,Cpu>*>(&A);
cout << "A ptr: " << &A << endl;
cout << A.norm() << endl;
std::thread *th;
for(i=0; i < nth; i++)
{
// th = new std::thread(&Faust::Transform<FPP,Cpu>::multiply_par_run, nth, i, num_per_th, data_size, opThis, data, mats, threads);
th = new std::thread(Faust::multiply_par_run<FPP>, nth, i, num_per_th, data_size, opThis, std::ref(data), std::ref(mats), std::ref(threads), std::ref(barrier_mutex), std::ref(barrier_cv), std::ref(barrier_count));
threads.push_back(th);
}
for(auto t: threads)
if(t->get_id() != std::this_thread::get_id() )
t->join();
M = dynamic_cast<Faust::MatDense<FPP,Cpu>*>(mats[0]);
cout << "M:" << M << endl;
MatDense<FPP,Cpu> M_;
if(! M)
{
M_ = Faust::MatDense<FPP,Cpu>(*dynamic_cast<Faust::MatSparse<FPP,Cpu>*>(mats[0]));
M = &M_;
}
//TODO: delete other thread mats
//TODO: return a ptr instead of a copy
//TODO: delete threads
return *M;
}
template<typename FPP>
void Faust::multiply_par_run(int nth, int thid, int num_per_th, int data_size, char opThis, std::vector<Faust::MatGeneric<FPP, Cpu>*>& data, std::vector<Faust::MatDense<FPP,Cpu>*>& mats, std::vector<std::thread*>& threads, std::mutex & barrier_mutex, std::condition_variable & barrier_cv, int & barrier_count)
{
Faust::MatSparse<FPP, Cpu> * sM;
Faust::MatDense<FPP, Cpu> *M;
Faust::MatDense<FPP,Cpu>* tmp; // (data[end_id-1]);
// std::cout << "omp th: " << nth << " num_per_th: " << num_per_th << std::endl;
std::cout << "num_per_th: " << num_per_th << std::endl;
while(num_per_th > 1)
{
if(thid < nth)
{
// {
// // build barrier
// std::lock_guard<std::mutex> lk4(barrier_mutex);
// barrier_count++;
// if(barrier_count == nth) barrier_cv.notify_all();
// }
int first_id, end_id, id;
first_id = num_per_th*thid;
end_id = num_per_th*(thid+1);
id = 'N' == opThis? end_id-1: first_id;
cout << "thid=" << thid << " end orig adr.:" << data[id] << endl;
if(sM = dynamic_cast<Faust::MatSparse<FPP,Cpu>*>(data[id]))
tmp = new Faust::MatDense<FPP,Cpu>(*sM);
else
tmp = new Faust::MatDense<FPP,Cpu>(*(Faust::MatDense<FPP,Cpu>*)(data[id]));
mats[thid] = tmp;
cout << "thid=" << thid << "mats[thid]=" << mats[thid] << "tmp=" << tmp << endl;
cout << "tmp.norm()" << tmp->norm() << endl;
if(opThis == 'N')
for(int i=end_id-2;i>=first_id; i--)
{
cout << "mul:" << data[i] << " thid: " << thid << endl;
if(sM = dynamic_cast<Faust::MatSparse<FPP,Cpu>*>(data[i]))
sM->multiply(*mats[thid], opThis);
else
dynamic_cast<Faust::MatDense<FPP,Cpu>*>(data[i])->multiply(*mats[thid], opThis);
}
else
for(int i=first_id+1;i < end_id; i++)
if(sM = dynamic_cast<Faust::MatSparse<FPP,Cpu>*>(data[i]))
sM->multiply(*mats[thid],opThis);
else
dynamic_cast<Faust::MatDense<FPP,Cpu>*>(data[i])->multiply(*mats[thid], opThis);
data[thid] = mats[thid];
// mats[thid]->Display();
// cout << mats[thid]->norm() << endl;
cout << "thid=" << thid << "mats[thid]=" << mats[thid] << "data[thid]=" << data[thid] << endl;
}
// //wait until barrier is ready
// {
// std::lock_guard<std::mutex> lk3(barrier_mutex);
// if(barrier_count < nth)
// barrier_cv.wait(lk3, [&barrier_count, &nth]{return barrier_count == nth;});
// }
//
//
// //barrier
// {
// std::lock_guard<std::mutex> lk(barrier_mutex);
// if(thid < nth)
// {
// // decrement barrier
// barrier_count--;
// std::cout << "thid: " << thid << " update barrier count: " << barrier_count << std::endl;
// if(barrier_count <= 0) barrier_cv.notify_all();
// }
// else return;
// }
////
////
// {
////
// std::unique_lock<std::mutex> lk2(barrier_mutex);
// if(nth > 1)
// {
// num_per_th = 2;
// }
// else
// num_per_th = 1;
//
// nth >>= 1;
// if(barrier_count > 0)
// {
// cout << "barrier reached thid:" << thid << endl;
// cout << "thread " << thid << " is waiting" << endl;
// barrier_cv.wait(lk2, [&barrier_count]{return barrier_count <= 0;});
// }
//
// cout << "barrier passed by " << thid << endl;
// }
// for(auto t: threads)
// if(t->get_id() != std::this_thread::get_id() )
// t->join();
cout <<"num_per_th: " << num_per_th << " thid:" << thid << endl;
}
}
template<typename FPP>
Faust::MatDense<FPP,Cpu> Faust::Transform<FPP,Cpu>::multiply_omp(const Faust::MatDense<FPP,Cpu> A, const char opThis) const
......
src/faust_linear_operator/CPU/faust_TransformHelper.hpp
+ 4
− 1
View file @ 3330d6f1
......@@ -203,7 +203,10 @@ namespace Faust {
this->transform->data.erase(this->transform->data.end()-1);
break;
case 4:
M = this->transform->multiply_omp(A, isTransposed2char());
M = this->transform->multiply_omp(A, isTransposed2char());
break;
case 5:
M = this->transform->multiply_par(A, isTransposed2char());
break;
default:
M = this->transform->multiply(A, isTransposed2char());
......
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Please register or to comment