packed_simd requires nightly

[alamb]

Note: migrated from original JIRA: https://issues.apache.org/jira/browse/ARROW-6718

See [https://github.com/rust-lang/rfcs/pull/2366] for more info on stabilization of this crate.

 
{code:java}
error[E0554]: #![feature] may not be used on the stable release channel
--> /home/andy/.cargo/registry/src/github.com-1ecc6299db9ec823/packed_simd-0.3.3/src/lib.rs:202:1
|
202 | / #![feature(
203 | | repr_simd,
204 | | const_fn,
205 | | platform_intrinsics,
... |
215 | | custom_inner_attributes
216 | | )]
| |__^
{code}

[alamb]

Comment from Neville Dipale(nevi_me) @ 2019-12-23T04:13:36.550+0000:

If everyone's happy with the performance of arithmetic and compute kernels after the improvements to the non-SIMD functions (ARROW-7460), I'd like to propose removing packed_simd [~paddyhoran]

Comment from Paddy Horan(paddyhoran) @ 2019-12-24T13:47:15.418+0000:

I would love to get back on stable.  I added a feature to disable explicit SIMD to try and make progress toward this goal.  Although the main thing we need is specialization on stable.

 

If we can get the same level of performance then I'm all for removing packed_simd.  At the time we adopted it, the author was trying to get it adopted into std.  Since then he has stopped driving this forward until other features land.

 

I'll take a look in the next few days to compare performance, etc. 

Comment from Kyle Strand(batmanaod) @ 2020-09-30T22:58:09.616+0000:

There is now a portable-SIMD Rust language working group: https://blog.rust-lang.org/inside-rust/2020/09/29/Portable-SIMD-PG.html

Their goal is to make "a portable SIMD API available to stable Rust users." We should follow their progress and adopt their recommendations.

[velvia]

@alamb any updates on performance of stable vs nightly? I'd be surprised if the non SIMD performance was quite as good, but hey all power to it.....

[alamb]

I'd be surprised if the non SIMD performance was quite as good, but hey all power to it.....

I don't think 'non SIMD' is quite accurate -- it is more like "Clang produced (auto vectorized) SIMD" where the compiler generates the SIMD instructions

@jhorstmann did some work that showed several kernels were actually faster using autovectorized code: #1221

I believe @jorgecarleitao did some work in arrow2 to migrate it away from using packed_simd (and thus no longer require nightly rust): jorgecarleitao/arrow2#747

[jhorstmann]

See also #1492 for a small discussion about packed_simd vs portable_simd

[jorgecarleitao]

@alamb, note that std::simd requires nightly, as it is gated under #![feature(portable_simd)], rust-lang/rust#86656

[alamb]

@alamb, note that std::simd requires nightly, as it is gated under #![feature(portable_simd)], rust-lang/rust#86656

Ah -- my bad. Thanks you for the clarification

[velvia]

Hmmm interesting. BTW packed_simd as well as the portable SIMD work in std::simd both require nightly still. However, good idea to just use stable for now. :)

…

On Apr 1, 2022, at 10:22 AM, Andrew Lamb ***@***.***> wrote: I'd be surprised if the non SIMD performance was quite as good, but hey all power to it..... I don't think 'non SIMD' is quite accurate -- it is more like "Clang produced (auto vectorized) SIMD" where the compiler generates the SIMD instructions @jhorstmann <https://github.com/jhorstmann> did some work that showed several kernels were actually faster using autovectorized code: #1221 <#1221> I believe @jorgecarleitao <https://github.com/jorgecarleitao> did some work in arrow2 to migrate it to use packed_simd (and thus no longer require nightly rust): jorgecarleitao/arrow2#747 <jorgecarleitao/arrow2#747> — Reply to this email directly, view it on GitHub <#54 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAIDPWYRIVVKY4CP6EFHJS3VC4WDTANCNFSM43SV3AKQ>. You are receiving this because you commented.