[FFmpeg-devel] [PATCH v4 1/2][GSoC 2024] libavcodec/x86/vvc: Add AVX2 DMVR SAD functions for VVC

[Stone Chen]

On Mon, May 20, 2024 at 7:23 AM Ronald S. Bultje <rsbultje at gmail.com> wrote:

> Hi,
>
> This is mostly good, the following is tiny nitpicks.
>
> On Sun, May 19, 2024 at 8:46 PM Stone Chen <chen.stonechen at gmail.com>
> wrote:
>
>> +%macro INIT_OFFSET 6 ; src1, src2, dxq, dyq, off1, off2
>>
>
> The macro is only used once, so you could inline it in the calling
> function.
>
>>
>> +    imul            %5, 128
>> +    imul            %6, 128
>>
>
> I believe shl is typically preferred over imul for powers of two.
>
>
>> +    add             %5, 2
>> +    add             %6, 2
>>
>
> And these can be integrated as a constant offset in the lea below (lea %1,
> [%1 + %5 * 2 + 2 * 2], same for %2).
>
>
>> +    add             %5, %3
>> +    sub             %6, %3
>> +
>> +    lea             %1, [%1 + %5 * 2]
>> +    lea             %2, [%2 + %6 * 2]
>
> [..]
>
>> +cglobal vvc_sad, 6, 11, 5, src1, src2, dx, dy, block_w, block_h, off1,
>> off2, row_idx, dx2, dy2
>> +    movsxd           dx2q, dxd
>> +    movsxd           dy2q, dyd
>>
>
> If you change the argument type from int to intptr_t, this is not
> necessary anymore.
>
>
>> +    vvc_sad_16_128:
>> +        .loop_height:
>> +        mov         off1q, src1q
>> +        mov         off2q, src2q
>> +        mov      row_idxd, block_wd
>> +        sar      row_idxd, 4
>>
>
> You could right-shift block_wd by 4 outside the loop (before .loop_height).
>
> Ronald
>

On Mon, May 20, 2024 at 11:53 AM Ronald S. Bultje <rsbultje at gmail.com>
wrote:

> Hi,
>
> one more, I forgot.
>
> On Sun, May 19, 2024 at 8:46 PM Stone Chen <chen.stonechen at gmail.com>
> wrote:
>
>> +pw_1: dw 1
>>
> [..]
>
>> +    vpbroadcastw       m4, [pw_1]
>>
>
> We typically suggest to use vpbroadcastd, not w (and then pw_1: times 2 dw
> 1). agner shows that on e.g. Haswell, the former (d) is 1 uops with 5
> cycles latency, whereas the latter (w) is 3 uops with 7 cycles latency, or
> more generally d is faster then w.
>
> Ronald
>

Hi Ronald,

I've sent a v5 incorporating all the above, thank you for the feedback!

-Stone