[FFmpeg-devel] [PATCH 07/10] avcodec/vc1: Arm 64-bit NEON inverse transform fast paths

[Martin Storsjö]

On Thu, 31 Mar 2022, Ben Avison wrote:

> On 30/03/2022 14:49, Martin Storsjö wrote:
>> Looks generally reasonable. Is it possible to factorize out the individual 
>> transforms (so that you'd e.g. invoke the same macro twice in the 8x8 and 
>> 4x4 functions) without too much loss?
>
> There is a close analogy here with the vertical/horizontal deblocking 
> filters, because while there are similarities between the two matrix 
> multiplications within a transform, one of them follows a series of loads and 
> the other follows a matrix transposition.
>
> If you look for example at ff_vc1_inv_trans_8x8_neon, you'll see I was able 
> to do a fair amount of overlap between sections of the function - 
> particularly between the transpose and the second matrix multiplication, but 
> to a lesser extent between the loads and the first matrix multiplication and 
> between the second multiplication and the stores. This sort of overlapping is 
> tricky to maintain when using macros. Also, it means the the order of 
> operations within each matrix multiply ended up quite different.
>
> At first sight, you might think that the multiplies from the 8x8 function 
> (which you might also view as kind of 8-tap filter) would be re-usable for 
> the size-8 multiplies in the 8x4 or 4x8 function. Yes, the instructions are 
> similar, save for using .4h elements rather than .8h elements, but that has 
> significant impacts on scheduling. For example, the Cortex-A72, which is my 
> primary target, can only do NEON bit-shifts in one pipeline at once, 
> irrespective of whether the vectors are 64-bit or 128-bit long, while other 
> instructions don't have such restrictions.
>
> So while in theory you could factor some of this code out more, I suspect any 
> attempt to do so would have a detrimental effect on performance.

Ok, fair enough. Yes, it's always a trade off between code simplicity and 
getting the optimal interleaving. As you've spent the effort on making it 
efficient with respect to that, let's go with that then!

(FWIW, for future endeavours, having the checkasm tests in place while 
developing/tuning the implementation does allow getting good empirical 
data on how much you gain from different alternative scheduling choices. I 
usually don't follow the optimization guides for any specific core, but 
track the benchmark numbers for a couple different cores and try to pick a 
scheduling that is a decent compromise for all of them.)

Also, for future work - if you have checkasm tests in place while working 
on the assembly, I usually amend the test with debug printouts that 
visualize the output of the reference and the tested function, and a map 
showing which elements differ - which makes tracking down issues a whole 
lot easier. I don't think any of the checkasm tests in ffmpeg have such 
printouts though, but within e.g. the dav1d project, the checkasm tool is 
extended with helpers for comparing and printing such debug aids.

// Martin