On Tue, 21 Jun 2022, Martin Storsjö wrote:

> On Mon, 13 Jun 2022, Swinney, Jonathan wrote:
>
>> - added a test for yuv2plane1 (currently disabled for x86_64)
>
> What's the reason for having it disabled for x86 - is it another case where 
> the current implementations there aren't bitexact? Could we avoid that by 
> setting the bitexact flag for the new yuv2yuv1 test?
>
>> - fixed test for yuv2planeX for aarch64 which was previously not working at 
>> all
>
> Could we make the test fuzzy and allow minor differences from the reference, 
> when the bitexact flag isn't set, and separately test with the bitexact flag 
> and require exact matches?
>
>> @@ -95,7 +210,7 @@ static void check_yuv2yuvX(void)
>>     ff_sws_init_scale(ctx);
>>     for(isi = 0; isi < INPUT_SIZES; ++isi){
>>         dstW = input_sizes[isi];
>> -        for(osi = 0; osi < 64; osi += 16){
>> +        for(osi = 0; osi < 1; osi += 16){
>
> This looks like a stray leftover change?

I had a look at this, trying to fix things up. This now passes tests on 
x86_32, x86_64 and aarch64. See the attached patch, which goes on top of 
yours.

It's not intended as a final version of how things should be necessarily, 
but as a more concrete pointer about how it could be done - it needs at 
least reindenting after adding the outer for loop.

I also had to skip the filter sizes 1 and 3 in check_yuv2yuvX, because 
ff_yuv2planeX_8_sse2 couldn't handle those. I presume that means that in 
practice, those aren't ever used?

// Martin