[FFmpeg-devel] [PATCH 3/3] avcodec/hevc: Add wasm simd128 idct

[Zhao Zhili]

> Le 13 novembre 2024 16:16:36 GMT+02:00, Zhao Zhili <quinkblack at foxmail.com <https://ffmpeg.org/mailman/listinfo/ffmpeg-devel>> a écrit :
> >From: Zhao Zhili <zhilizhao at tencent.com <https://ffmpeg.org/mailman/listinfo/ffmpeg-devel>>
> >
> >We don't use intrinsics normally for performance reasons. However,
> >WASM isn't machine instruction.
> 
> I don't think that's relevant to the matter.
> 
> What may be relevant is that WASM is an infinite register ISA, so that the compiler's register allocator will never run out and spill.
> 
> But then again, it will have to be recompiled to a finite register machine ISA, with probably 8, 16 or 32 vector registers, and if we break whatever the machine's limit is, performance will suck.
> 
> So I actually expect that assembler will be about as relevant for WASM as it is for the machine ISA with the fewest vector registers... on that architecture.
> 
> But in any case, I cannot agree with that unsubstantiated statement. Unless proven otherwise, assembler is relevant here too.
OK, I can remove this misleading information from commit message. I still
have no interest to write WASM in handwritten assembly. Does anyone
has a plan on WASM SIMD support? If the answer is yes, I can stop at here.
The work is only a proof of concept.