Repository navigation
[C++][Parquet] Optimize GreaterThanDynamicFunction for Neon #55313
Description
Describe the enhancement requested
GreaterThanBitmapImpl runs slow for neon, since below might not optimized in ARM:
template <typename Predicate>
inline uint64_t LevelsToBitmap(const int16_t* levels, int64_t num_levels,
Predicate predicate) {
// Both clang and GCC can vectorize this automatically with SSE4/AVX2.
uint64_t mask = 0;
for (int x = 0; x < num_levels; x++) {
mask |= static_cast<uint64_t>(predicate(levels[x]) ? 1 : 0) << x;
}
return ::arrow::bit_util::ToLittleEndian(mask);
}
Component(s)
C++, Parquet
Activity
JonSnow1807 commentedon Oct 30, 2026
I have a NEON version of GreaterThanBitmap along the lines of the FindMinMax kernels from #51478 (NEON-only, x86 untouched), with a randomized test against the scalar loop and a benchmark. On an M2 Pro the 64-length case drops from 15.6 ns to 4.9 ns and the DefinitionLevelsToBitmap benchmarks improve 1.8x to 3.0x. I am at the open-PR limit for new contributors right now and will open it once one of my other PRs merges or closes, unless you would rather take it yourself.