Cycles: use intrinsics for fast_rint #107371

Xavier Hallade · 2023-04-26T15:25:53+02:00

Xavier Hallade commented

2023-04-26 15:25:53 +02:00

MSVC can't optimize it out and even keeps an external call to CRT
function rintf: https://godbolt.org/z/Ex9vjf8vj

It does translate to a real speedup on windows on some scenes, here are the ratios I had on my 13900K:
classroom 101.53%
junkshop 100.71%
monster 100.76%
attic 107.98%
bistro 113.00%

MSVC can't optimize it out and even keeps an external call to CRT function rintf: https://godbolt.org/z/Ex9vjf8vj It does translate to a real speedup on windows on some scenes, here are the ratios I had on my 13900K: classroom 101.53% junkshop 100.71% monster 100.76% attic 107.98% bistro 113.00%

Xavier Hallade added 1 commit 2023-04-26 15:26:02 +02:00

buildbot/vexp-code-patch-coordinator Build done. Details

587f2e9583 Cycles: use intrinsics for fast_rint

MSVC can't optimize it out and even keeps an external call to CRT
function rintf: https://godbolt.org/z/Ex9vjf8vj

Xavier Hallade added the

Module

Render & Cycles

label 2023-04-26 15:26:28 +02:00

Xavier Hallade added this to the 3.6 LTS milestone 2023-04-26 15:26:33 +02:00

Xavier Hallade added this to the Render & Cycles project 2023-04-26 15:26:41 +02:00

Xavier Hallade requested review from Brecht Van Lommel 2023-04-26 15:26:55 +02:00

Xavier Hallade requested review from Sergey Sharybin 2023-04-26 15:26:56 +02:00

Brecht Van Lommel approved these changes 2023-04-26 19:28:53 +02:00

Brecht Van Lommel commented

2023-04-26 19:29:02 +02:00

@blender-bot build

Sergey Sharybin referenced this pull request

2023-04-27 11:18:31 +02:00

Deps: Bump version of the sse2neon library #107396

Sergey Sharybin commented

2023-04-27 11:19:37 +02:00

That is quite some speedup! Just to confirm, does it keep the same precision?

As for the failure in the arm64 i've created #107396 which should solve it.

That is quite some speedup! Just to confirm, does it keep the same precision? As for the failure in the arm64 i've created #107396 which should solve it.

Xavier Hallade commented

2023-04-27 12:44:41 +02:00

Yes precision is the same - you can check the assembly on the godbolt.org link.
The only difference in gcc's generated assembly is the rounding control value, which when you use rintf() is set to _MM_FROUND_CUR_DIRECTION and then uses what's in MXCSR register which is nearest by default, while in my proposed version I've put (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC) explicitly.

Yes precision is the same - you can check the assembly on the godbolt.org link. The only difference in gcc's generated assembly is the rounding control value, which when you use rintf() is set to _MM_FROUND_CUR_DIRECTION and then uses what's in MXCSR register which is nearest by default, while in my proposed version I've put (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC) explicitly.

Sergey Sharybin commented

2023-04-27 14:17:54 +02:00

Thanks for the explanation. The code seems fine, just need to avoid breaking arm64 platform (as in, need to land #107396 first).