[FFmpeg-devel] [PATCH 1/2] lavc/flacdsp: R-V V flac_wasted32

[Rémi Denis-Courmont]

Le sunnuntaina 12. toukokuuta 2024, 21.37.28 EEST James Almer a écrit :
> Not sure if you're taking it into account, but the minimum blocksize is
> 16

Granted, this only fills a single 8-vector vector group (v8-v15), so only a 
quarter of the register bank (v0-v31), which is unusually low. But that 
already adds up to 32 ints per iteration with 128-bit vectors respectively. 
IIUC, the x86 implementation is only half as much.

> and the buffer is always allocated for max_blocksize plus padding,

RVV really wants element-size alignment, so 32/64-bit here. Beyond that, it 
really does not care. (I think Arm SVE works the same way?)

> so you should be able to do more samples per loop than this.

In my experience, this particular hardware would likely exhibit marginally 
better performance with only *half* as many sample per iterations. I just 
don't want to overfit to this relatively early and low-end hardware design. In 
fact, Yuechi already has newer better hardware.

> Same for wasted33.

The wasted33 kernel actually already uses 3 eighth (v8-v11, v16-v23) of the
bank, for 16 ints per iteration. I doubt that unrolling explicitly would help.

The performance (~1.5x) is pretty disappointing to be sure. The root cause is 
RVV's notoriously lacks of widening left-shifts. FWIW, Zvbb only adds the 
unsigned variant, which is not what we need here. Plus there is no 
commercially available hardware with Zvbb yet. So in the end, we have an extra 
size conversion. And then 64-bit shift and 64-bit element stores which are 
half as fast as 32-bit ones on weighed basis.

Maybe widening signed multiplication would be faster though, I will try that.

-- 
Rémi Denis-Courmont
http://www.remlab.net/