News for hackers. Stuff that matters.
Re:Linux founder tells Intel to stop inventing magic instruc... by gameswithgo on Monday July 13, 2020 @03:07PM (#23824369)
Those look like benchmarks comparing just autovectorized compilation of some tools and measuring overall runtime. I am referring to cases of hand tuned use SIMD. The bSIMD project used to have a suite of benchmarks up and some where very very impressive for AVX-512. The page appears to be gone or moved now though.

To be clear I am not suggest "avx-512 will cause programs to run twice as fast or more", just that specific functions that used to use AVX-2, may be more than twice as fast if you move them to AVX-512.

Re:Linux founder tells Intel to stop inventing magic instruc... by rumanator on Monday July 13, 2020 @04:55PM (#23825468)
> Those look like benchmarks comparing just autovectorized compilation of some tools and measuring overall runtime. I am referring to cases of hand tuned use SIMD.

I presented data that showed that AVX-512 in reality only brings margins performance improvements.

If you have any data that shows otherwise then feel free to present it.

Still, hand-tuned assembly is hardly a relevant example, let alone selling point, as it creates hypothetical expectations that are never ever matched in reality.

Re:Linux founder tells Intel to stop inventing magic instruc... by vardump on Monday July 13, 2020 @05:51PM (#23825978)
"Still, hand-tuned assembly is hardly a relevant example, let alone selling point, as it creates hypothetical expectations that are never ever matched in reality."

There's a ton of hand tuned assembly running on your system in numerous libraries. Media and data compression? Hand tuned assembly. Matrix math? You guessed it. C-standard library functions? Yes, really.

Re:Linux founder tells Intel to stop inventing magic instruc... by greyhair on Tuesday July 14, 2020 @12:02PM (#23833519)
Maybe for media, crypto, things like that, but we all have the source for the C-standard libaries, and there is not all that much assembly in there given the size of the code base.

Certainly you have required architecture instructions for things like instruction scheduling (isync/dsync), stack manipulation (setjump/longjump and others) or floating point, but though it looks like a lot of assembly source, it is really just a little bit for each of the many supported architectures in the tree. And most of it isn't performance related, it is just plain required because architectures vary on things like instruction scheduling and stack/cache manipulation.

Re:Linux founder tells Intel to stop inventing magic instruc... by vardump on Monday July 13, 2020 @03:51PM (#23824884)
Some SIMD optimization cases can occasionally exceed scalar versions by more than lane width multiply. Often because of branch elision, but also less dependency chains stalls due to the larger register file. There are also some special instructions that can be very useful in SIMD case (packing, unpacking, permutation, etc.).