jhorstmann opened a new issue, #10841:
URL: https://github.com/apache/arrow-rs/issues/10841
There are some open issues or prs that want to make use of cpu specific
instructions, and dispatch either at compile time or runtime:
- BMI `pext` for bitmap filtering #10136
- AArch64 SVE gather for dictionary decoding #10066
- AVX512 `compresstore` for filtering #1949
- BMI or AVX512-vbmi for bitpack decoding #8102
Previously we avoided target specific code and relied on compiler flags and
auto-vectorization to generate good code. We also [recommended compiling for a
specific
`target-cpu`](https://github.com/apache/arrow-rs/blob/main/arrow/README.md#performance-tips).
If we want to switch to runtime dispatch for some cases, I think we should
implement some common framework for feature detection and dispatch. And decide
on the level at which dispatch happens, for example dispatching for a single
`pext` instruction would not be beneficial, the cost would need to be amortized
across a bigger batch of data.
--
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]
