diff options
| author | Jacques Lucke <jacques@blender.org> | 2021-12-02 14:54:35 +0300 |
|---|---|---|
| committer | Jacques Lucke <jacques@blender.org> | 2021-12-02 14:56:47 +0300 |
| commit | e1309030603980c6b2f33486adf6ae5c2e4eb965 (patch) | |
| tree | eebf8666d6ab20a75b3bae4299eac12bf0b161fa | |
| parent | 0f89d05848b8a279ac6507404dc535244cd73d43 (diff) | |
BLI: avoid invoking tbb for small workloads
We often call `parallel_for` in places with very variable
sized workloads. When many elements are processed,
using multi-threading is great, but when processing
few elements (possibly many times) using `parallel_for`
can result in significant overhead.
I measured that this improves performance by >20% in
the refactored realize instances code I'm working on
separately. The change might also help with debugging
sometimes, because the stack trace is smaller and contains
fewer irrevelant symbols.
| -rw-r--r-- | source/blender/blenlib/BLI_task.hh | 15 |
1 files changed, 10 insertions, 5 deletions
diff --git a/source/blender/blenlib/BLI_task.hh b/source/blender/blenlib/BLI_task.hh index da7309837c8..84d5cd39bb4 100644 --- a/source/blender/blenlib/BLI_task.hh +++ b/source/blender/blenlib/BLI_task.hh @@ -67,14 +67,19 @@ void parallel_for(IndexRange range, int64_t grain_size, const Function &function return; } #ifdef WITH_TBB - tbb::parallel_for(tbb::blocked_range<int64_t>(range.first(), range.one_after_last(), grain_size), - [&](const tbb::blocked_range<int64_t> &subrange) { - function(IndexRange(subrange.begin(), subrange.size())); - }); + /* Invoking tbb for small workloads has a large overhead. */ + if (range.size() >= grain_size) { + tbb::parallel_for( + tbb::blocked_range<int64_t>(range.first(), range.one_after_last(), grain_size), + [&](const tbb::blocked_range<int64_t> &subrange) { + function(IndexRange(subrange.begin(), subrange.size())); + }); + return; + } #else UNUSED_VARS(grain_size); - function(range); #endif + function(range); } template<typename Value, typename Function, typename Reduction> |