zhanchangbao-sanechips opened a new pull request, #2706:
URL: https://github.com/apache/orc/pull/2706
# Add RVV acceleration for RLEv2 bit unpacking
The scalar RLEv2 decoder is a bottleneck when we're reading wide integer
columns. Since RVV hardware is becoming much more common, this patch adds a
vectorized path to unpack bits in parallel.
## Changes
- Add `c++/src/BpackingRvv.{hh,cc}`: Covers bit widths 3..32, 40, 48 and 56
using gather-based and strided-load vector paths; falls back to the scalar
implementation for widths 1, 2, 64 and for bit-level misalignment / buffer
boundaries.
- `c++/src/Dispatch.hh`: add `DispatchLevel::RVV` and the corresponding
`isSupported` check.
- `c++/src/RleDecoderV2.cc`: register the RVV implementation in the dispatch
table when `ORC_HAVE_RUNTIME_RVV` is defined.
- `c++/src/CMakeLists.txt` and `c++/test/CMakeLists.txt`: compile the new
source under RVV builds and reuse `TestRleVectorDecoder.cc` for vector decoder
test coverage.
## Test
1. Build with `-DBUILD_ENABLE_RVV=ON` on a RISC-V machine(SG2044) with Gcc
support RVV(I used gcc 14.2.1).
2. use ORC_USER_SIMD_LEVEL=RVV ./orc-test and ORC_USER_SIMD_LEVEL=NONE
./orc-test，all unittest pass.
3. I write a small standalone benchmark(bench_rvv.cc) to compare scalar and
RVV decoding speeds. Switch paths via `ORC_USER_SIMD_LEVEL=NONE|RVV`. but i
dont left `bench_rvv.cc` at the repo.
## performance
| bits | NONE_Mvals/s | RVV_Mvals/s | speedup |
|------|-------------|-------------|---------|
| 1 | 80.8 | 78.5 | 0.97 |
| 2 | 106.2 | 102.4 | 0.96 |
| 3 | 114.1 | 160.8 | 1.41 |
| 4 | 217.9 | 235.4 | 1.08 |
| 5 | 96.2 | 240.1 | 2.50 |
| 6 | 79.9 | 245.3 | 3.07 |
| 7 | 75.4 | 246.7 | 3.27 |
| 8 | 237.3 | 276.0 | 1.16 |
| 9 | 70.8 | 175.0 | 2.47 |
| 10 | 69.5 | 124.3 | 1.79 |
| 12 | 60.7 | 110.7 | 1.82 |
| 14 | 52.8 | 102.1 | 1.93 |
| 16 | 184.6 | 238.1 | 1.29 |
| 17 | 55.4 | 173.9 | 3.14 |
| 20 | 48.2 | 173.9 | 3.61 |
| 24 | 156.3 | 176.9 | 1.13 |
| 26 | 41.3 | 135.0 | 3.27 |
| 28 | 35.2 | 135.6 | 3.85 |
| 30 | 32.3 | 135.9 | 4.21 |
| 32 | 155.2 | 162.0 | 1.04 |
| 40 | 139.7 | 146.6 | 1.05 |
| 48 | 126.8 | 134.7 | 1.06 |
| 56 | 108.3 | 100.1 | 0.92 |
| 64 | 111.0 | 111.3 | 1.00 |
## Screenshot
## unittest
<img width="1340" height="495" alt="功能测试"
src="https://github.com/user-attachments/assets/384ec71d-9150-4885-9657-7a5402c03096";
/>
## performance
<img width="708" height="1260" alt="性能测试"
src="https://github.com/user-attachments/assets/d4c9956d-357d-4d14-82ac-3313120568fe";
/>
## bench_rvv.cc
<img width="1136" height="1861" alt="bench_rvv cc"
src="https://github.com/user-attachments/assets/0affb9c6-6762-474b-9222-5a31b5b30ef6";
/>
--
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]
