saymrwulf/onnxruntime
1
0
Fork 0
mirror of https://github.com/saymrwulf/onnxruntime.git synced 2026-07-30 20:18:08 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Use alignas instead of naive padding to avoid false cache sharing (#12514)

Browse source 
PerThread and ChildThreadStat alignas
This commit is contained in:
Dmitri Smirnov 2022-08-10 11:23:20 -07:00 committed by GitHub
parent 25032f1756
commit c10704a501
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
1 changed files with 23 additions and 6 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

29
include/onnxruntime/core/platform/EigenNonBlockingThreadPool.h
View file

@ -268,13 +268,20 @@ class ThreadPoolProfiler {
bool enabled_ = false;
MainThreadStat& GetMainThreadStat(); //return thread local stat
int num_threads_;
struct ChildThreadStat {
#ifdef _MSC_VER
#pragma warning(push)
// C4324: structure was padded due to alignment specifier
#pragma warning(disable : 4324)
#endif // _MSC_VER
struct ORT_ALIGN_TO_AVOID_FALSE_SHARING ChildThreadStat {
std::thread::id thread_id_;
uint64_t num_run_ = 0;
onnxruntime::TimePoint last_logged_point_ = Clock::now();
int32_t core_ = -1; //core that the child thread is running on
PaddingToAvoidFalseSharing padding_; //to prevent false sharing
};
#ifdef _MSC_VER
#pragma warning(pop)
#endif // _MSC_VER
std::vector<ChildThreadStat> child_thread_stats_;
std::string thread_pool_name_;
};
@ -308,9 +315,9 @@ class ExtendedThreadPoolInterface : public Eigen::ThreadPoolInterface {
// Special case alternative to RunInParallelSection for use without
// an existing parallel section. Ideally we would use a single
// iplemenation and a stack-allocated ThreadPoolParallelSection.
// implementation and a stack-allocated ThreadPoolParallelSection.
//
// However, on the BM_ThreadPoolParallelFor microbenchmark I saw
// However, on the BM_ThreadPoolParallelFor micro-benchmark I saw
// ~20% overhead on the resulting single-loop parallel sections.
// There are some additional costs (~5%) for additional invocations
// through lambda functions on loop entry. Most significantly, on
@ -1311,7 +1318,13 @@ class ThreadPoolTempl : public onnxruntime::concurrency::ExtendedThreadPoolInter
// threads in the pool, and their lifetime is managed along with the
// pool.
struct PerThread {
#ifdef _MSC_VER
#pragma warning(push)
// C4324: structure was padded due to alignment specifier
#pragma warning(disable : 4324)
#endif // _MSC_VER
struct ORT_ALIGN_TO_AVOID_FALSE_SHARING PerThread {
constexpr PerThread() : pool(nullptr) {
}
ThreadPoolTempl* pool; // Parent pool, or null for normal threads.
@ -1327,9 +1340,13 @@ class ThreadPoolTempl : public onnxruntime::concurrency::ExtendedThreadPoolInter
// of times that the work-stealing code paths are used for
// rebalancing.
InlinedVector<int> preferred_workers;
PaddingToAvoidFalseSharing padding_2;
};
#ifdef _MSC_VER
#pragma warning(pop)
#endif // _MSC_VER
struct WorkerData {
constexpr WorkerData() : thread(), queue() {
}