Function GetCachedTID

tensorflow/core/profiler/internal/gpu/cupti_tracer.cc:85–89

GetCachedTID() caches the thread ID in thread-local storage (which is a userspace construct) to avoid unnecessary system calls. Without this caching, it can take roughly 98ns, while it takes roughly 1ns with this caching.

// userspace construct) to avoid unnecessary system calls. Without this caching,
// it can take roughly 98ns, while it takes roughly 1ns with this caching.
pid_t GetCachedTID() {
  static thread_local pid_t current_thread_id =
      Env::Default()->GetCurrentThreadId();
  return current_thread_id;
}

size_t Bytes2D(const CUDA_MEMCPY2D *p) { return p->Height * p->WidthInBytes; }