ColinLeeo opened a new issue, #1073:
URL: https://github.com/apache/tsfile/issues/1073
## Problem
`ENABLE_SIMD` only defines `-DENABLE_SIMD`. The library targets the portable
x86-64 baseline (SSE2) and never compiles with `-mavx2`/`-mssse3`, so simde's
AVX2/SSSE3 intrinsics (`vpgatherqq`, `vpshufb`, ...) compile to simde's
portable fallback.
Micro-benchmark of the RLE bit-unpack (same header, only difference is
`-mavx2`):
- width 14: 99 → 644 M values/s (6.5×)
- width 20: 99 → 612 M values/s (6.2×)
So the SIMD paths added in #823 / #898 (PLAIN, GORILLA, TS_2DIFF) and the
RLE packers were not running real AVX2 in shipped builds.
## Proposed changes
- Add `common/simd_dispatch.h` with `cpu_has_avx2()`
(`__builtin_cpu_supports("avx2")`, GCC/Clang on x86 only).
- Add `__attribute__((target("avx2")))` intrinsic twins of the bit-unpack
kernels, dispatched at runtime; keep the simde path as the portable fallback.
Implemented for RLE and TS_2DIFF in #1072. PLAIN / GORILLA kernels can
follow the same pattern.
Related: #1072
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]