rmuir closed issue #12621: Make `byte[]` vector comparisons faster! (if
possible)
URL: https://github.com/apache/lucene/issues/12621
--
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 comm
rmuir commented on issue #12621:
URL: https://github.com/apache/lucene/issues/12621#issuecomment-1763056704
From my analysis, code being generated is correct. recommend to explore
half-float instead for better performance and space tradeoffs.
--
This is an automated message from the Apach
rmuir commented on issue #12621:
URL: https://github.com/apache/lucene/issues/12621#issuecomment-1753202064
@benwtrent I think a big source of confusion is that while the data might be
`byte`, the related functions return 4-byte `int` and 4-byte `float` so from a
vector api perspective, the
benwtrent commented on issue #12621:
URL: https://github.com/apache/lucene/issues/12621#issuecomment-1753021709
Thank you @rmuir && @ChrisHegarty for digging into this!
The current Panama Vector API makes doing this kind of thing frustrating.
Thank y'all for wrestling with it to make
rmuir commented on issue #12621:
URL: https://github.com/apache/lucene/issues/12621#issuecomment-1751903073
@benwtrent I looked into this more and eeked a bit more out:
https://github.com/apache/lucene/pull/12632
--
This is an automated message from the Apache Git Service.
To respond to t
rmuir commented on issue #12621:
URL: https://github.com/apache/lucene/issues/12621#issuecomment-1747775200
> Actually it is worse: Java 20 introduced conversion between short/float,
but we got neither a native `float16` datatype nor vector support. In short:
completely unuseable.
We
uschindler commented on issue #12621:
URL: https://github.com/apache/lucene/issues/12621#issuecomment-1747206287
See https://github.com/openjdk/jdk/pull/9422 (Java 20)
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use th
uschindler commented on issue #12621:
URL: https://github.com/apache/lucene/issues/12621#issuecomment-1747204954
Actually it is worse: Java 20 introduced conversion between short/float, but
we got neither a native `float16` datatype nor vector support. In short:
completely unuseable. 🤮
--
rmuir commented on issue #12621:
URL: https://github.com/apache/lucene/issues/12621#issuecomment-1747066837
My recommendation: stop messing around with `byte` and start thinking about
the new 16-bit half-float support that is present in Java 21. Unfortunately the
half-float *vectorization*
rmuir commented on issue #12621:
URL: https://github.com/apache/lucene/issues/12621#issuecomment-1747044386
As far as the ARM goes, the fact it has only 128-bit SIMD is the limiting
factor.
For e.g. AVX-256, we use 64-bit vector of 8 byte values -> 128 bit vector of
8 short values ->
rmuir commented on issue #12621:
URL: https://github.com/apache/lucene/issues/12621#issuecomment-1747026111
Also their suggested replacement of 3 instructions for the `VPDPBUSD` is:
> Likewise, for 8-bit values, three instructions are needed - VPMADDUBSW
which is used to multiply two
rmuir commented on issue #12621:
URL: https://github.com/apache/lucene/issues/12621#issuecomment-1747002969
the type conversions are what makes it slow. for float case it is the equiv
of:
```
float x = something;
float y = something;
float z = something;
// no conversions
f
12 matches
Mail list logo
