cpu_engine: OpenMP regions are ~2x slower on Apple Silicon because libomp's default KMP_BLOCKTIME is 0

[Nor-s]

Summary

On arm64 macOS, LLVM libomp defaults KMP_BLOCKTIME to 0.

Worker threads therefore go to sleep right after every #pragma omp parallel region,
and the next region has to wake them again through a condition variable.

The SW engine opens many small parallel regions per frame (one per image, or one per shape).
On this platform the wake/sleep cost exceeds the work being parallelized.

Setting only KMP_BLOCKTIME=1 (1 ms) doubles the CPU benchmark FPS.

We should either:

• (A) set KMP_BLOCKTIME in the benchmark runs so the numbers reflect ThorVG and not the libomp idle policy, or
• (B) call kmp_set_blocktime() inside the SW engine when OpenMP is enabled. ( LLVM/Intel libomp)

Environment

• Apple M5 Pro (6P + 12E cores), macOS 26 (Darwin 25)
• ThorVG CPU engine, -Dextra=openmp (the meson default), libomp 22.1.x (Homebrew libomp / llvm)
• thorvg.benchmark, --backend=cpu --scene=default, default --threads=4, 2560x1440

Results

multiimagebench (5000 instances from 25 assets, 1000 frames) x2.03

./build/multiimagebench_thorvg_sdl
KMP_BLOCKTIME=1 ./build/multiimagebench_thorvg_sdl

Precedent

These projects set blocktime or wait policy programmatically:

project	what	value	code
Tencent/ncnn	set_kmp_blocktime() -> kmp_set_blocktime() around each Extractor::extract(), restored after	20 ms (Option::openmp_blocktime)	cpu.cpp#L3284-L3291, net.cpp#L3216-L3217
Android NNAPI CPU executor	RAII ScopedOpenmpSettings: kmp_set_blocktime(20), restored on destruction	20 ms	CpuExecutor.cpp#L1874-L1901 (AOSP mirror)
TensorFlow (oneDNN/MKL build)	sets blocktime once unless the user set KMP_BLOCKTIME	1 ms	threadpool_device.cc#L90-L98
ggml / llama.cpp	setenv("KMP_BLOCKTIME", "200", 0) when built with OpenMP	200 ms	ggml-cpu.c#L3909-L3924
primecount	app sets OMP_WAIT_POLICY=ACTIVE + KMP_BLOCKTIME=30ms unless the user set them	30 ms	main.cpp#L85-L110
PyTorch / IPEX CPU launchers	KMP_BLOCKTIME=1	1 ms	run_cpu.py#L434-L438, launcher_base.py#L314-L317
DynEarthSol	#if __APPLE__ && _OPENMP: works around blocktime=0 on Apple Silicon	active	dynearthsol.cxx#L602-L613
LAMMPS (INTEL)	kmp_set_blocktime(0)	0	fix_intel.cpp#L153-L156
oneTBB python module	KMP_BLOCKTIME=0 unless set, to avoid spinning next to TBB's pool	0	__init__.py#L323-L324
FalkorDB	KMP_BLOCKTIME=0 unless set (Redis module sharing the host process)	0	module_init.rs#L232-L240

Libraries with many short regions per call (ncnn, NNAPI, TensorFlow, ggml) pick a small non-zero value and respect user overrides. Projects that pick 0 do so because they share cores with other pools, or for power reasons. ThorVG's case, thousands of small regions per frame, is the former.