[FFmpeg-devel] [PATCH 4/4] swscale/aarch64/rgb2rgb: add neon implementation for rgb24toyv12

[Martin Storsjö]

On Thu, 29 Aug 2024, Ramiro Polla wrote:

> On Wed, Aug 28, 2024 at 11:23 PM Martin Storsjö <martin at martin.st> wrote:
>> On Wed, 28 Aug 2024, Ramiro Polla wrote:
>
>>> +2:
>>> +        // load first line
>>> +        ld3             {v16.8b, v17.8b, v18.8b}, [x0], #24
>>> +        ld3             {v19.8b, v20.8b, v21.8b}, [x0], #24
>>
>> Hmm, can't we do just one single ld3 with .16b registers, instead of two
>> separate ones?
>>
>> If you want to keep the same register layout as now, load into v19-v21,
>> then do "uxtl v16.8h, v19.8b; uxtl2 v19.8h, v19.16b".
>
> Thanks, that made it faster.
>
>>> +        uxtl            v16.8h, v16.8b              // v16 = B11
>>> +        uxtl            v17.8h, v17.8b              // v17 = G11
>>> +        uxtl            v18.8h, v18.8b              // v18 = R11
>>> +        uxtl            v19.8h, v19.8b              // v19 = B12
>>> +        uxtl            v20.8h, v20.8b              // v20 = G12
>>> +        uxtl            v21.8h, v21.8b              // v21 = R12
>>> +
>>> +        // calculate Y values for first line
>>> +        rgbconv16       v24, v16, v17, v18, BY, GY, RY // v24 = Y11
>>> +        rgbconv16       v25, v19, v20, v21, BY, GY, RY // v25 = Y12
>>> +
>>> +        // pairwise add and save rgb values to calculate average
>>> +        addp            v5.8h, v16.8h, v19.8h
>>> +        addp            v6.8h, v17.8h, v20.8h
>>> +        addp            v7.8h, v18.8h, v21.8h
>>> +
>>> +        // load second line
>>> +        ld3             {v16.8b, v17.8b, v18.8b}, [x10], #24
>>> +        ld3             {v19.8b, v20.8b, v21.8b}, [x10], #24
>>
>> It's a shame we can't start this load earlier. But as essentially
>> everything depends on the input as it is, in v16-v21, we'd pretty much
>> need to use different registers here in order to do that.
>>
>> If you wanted to, you could try loading earlier, into different registers
>> (I think v26-v31 are free at this point?), while then doing the uxtl into
>> the same registers as before, which shouldn't require any further changes.
>
> Thanks, that also led to a small improvement.
>
> New patch attached.

The new version LGTM, thanks!

// Martin