HeterogeneousCore/CUDAUtilities/src/StreamCache.cc - master - githubmirror/cmssw

HeterogeneousCore/CUDAUtilities/src/StreamCache.cc

44 строки2 KB

Andrea Bocci

Rename GPU check macros to uppercase

24 июл 2026, 18:36

24 июл 2026, 18:364ca8016

#include "FWCore/Utilities/interface/thread_safety_macros.h"
#include "HeterogeneousCore/CUDAUtilities/interface/StreamCache.h"
#include "HeterogeneousCore/CUDAUtilities/interface/cudaCheck.h"
#include "HeterogeneousCore/CUDAUtilities/interface/currentDevice.h"
#include "HeterogeneousCore/CUDAUtilities/interface/deviceCount.h"
#include "HeterogeneousCore/CUDAUtilities/interface/ScopedSetDevice.h"

namespace cms::cuda {
  void StreamCache::Deleter::operator()(cudaStream_t stream) const {
    if (device_ != -1) {
      ScopedSetDevice deviceGuard{device_};
      CUDA_CHECK(cudaStreamDestroy(stream));
    }
  }

  // StreamCache should be constructed by the first call to
  // getStreamCache() only if we have CUDA devices present
  StreamCache::StreamCache() : cache_(deviceCount()) {}

  SharedStreamPtr StreamCache::get() {
    const auto dev = currentDevice();
    return cache_[dev].makeOrGet([dev]() {
      cudaStream_t stream;
      CUDA_CHECK(cudaStreamCreateWithFlags(&stream, cudaStreamNonBlocking));
      return std::unique_ptr<BareStream, Deleter>(stream, Deleter{dev});
    });
  }

  void StreamCache::clear() {
    // Reset the contents of the caches, but leave an
    // edm::ReusableObjectHolder alive for each device. This is needed
    // mostly for the unit tests, where the function-static
    // StreamCache lives through multiple tests (and go through
    // multiple shutdowns of the framework).
    cache_.clear();
    cache_.resize(deviceCount());
  }

  StreamCache& getStreamCache() {
    // the public interface is thread safe
    CMS_THREAD_SAFE static StreamCache cache;
    return cache;
  }
}  // namespace cms::cuda