C++ parallelisation with C++17

dstocco · March 31, 2020, 12:51pm

Dear core experts,
I’m trying to parallelise part of my code using the C++17 feature:

std::for_each(std::execution::par,...)

However, the code breaks as soon as I do:
#include <execution>
with the error:
fatal error: 'execution' file not found

I think that the parallel execution is not implemented in clang, apple clang, or even gcc 7 (?).
Is there an equivalent way to write some parallel code in O2?

Thanks in advance,
cheers,
Diego

eulisse · March 31, 2020, 1:21pm

You’ll need GCC 9 for this and even in that case you will need a backend library (e.g. TBB). I guess openmp is the best we can hope for, for now.

shahoian · March 31, 2020, 1:39pm

In the ITS tracking @iouribelikov use std::async:

github.com

AliceO2Group/AliceO2/blob/7506aa656613d0d191d62385bdfb82bac40afc3e/Detectors/ITSMFT/ITS/reconstruction/src/CookedTracker.cxx#L519


if (!numOfClusters) {
  return {0, 0};
}


std::vector<std::future<std::vector<TrackITSExt>>> futures(mNumOfThreads);
std::vector<std::vector<TrackITSExt>> seedArray(mNumOfThreads);


for (Int_t t = 0, first = 0; t < mNumOfThreads; t++) {
  Int_t rem = t < (numOfClusters % mNumOfThreads) ? 1 : 0;
  Int_t last = first + (numOfClusters / mNumOfThreads) + rem;
  futures[t] = std::async(std::launch::async, &CookedTracker::trackInThread, this, first, last);
  first = last;
}
Int_t nSeeds = 0, ngood = 0;
int nAllTracks = 0, nTracks = 0;
for (Int_t t = 0; t < mNumOfThreads; t++) {
  seedArray[t] = futures[t].get();
  nSeeds += seedArray[t].size();
  for (auto& track : seedArray[t]) {
    if (track.getNumberOfClusters() < kminNumberOfClusters)
      continue;

dstocco · April 1, 2020, 6:30am

Dear @eulisse and @shahoian,
thanks a lot for the suggestions. I’ll give a try to the async command.
Thanks!
Cheers,
Diego