Hi,

On Thu, Jul 7, 2016 at 10:53 AM, Ronald S. Bultje <[email protected]>
wrote:

> Hi,
>
> On Thu, Jul 7, 2016 at 9:52 AM, Alexandra Hájková <
> [email protected]> wrote:
>
>> On Thu, Jul 7, 2016 at 1:53 PM, Ronald S. Bultje <[email protected]>
>> wrote:
>> > On Thu, Jul 7, 2016 at 5:25 AM, Alexandra Hájková <
>> > [email protected]> wrote:
>> > > +    s->hevcdsp.add_residual[log2_trafo_size - 2](dst, coeffs,
>> stride);
>> >
>> > Won't this be slower since there's a memory store intermediate?
>> >
>> > (I know it's faster now because you don't have inverse transform simd,
>> but
>> > you should fix that by writing inverse transform simd, not by splitting
>> the
>> > transform and the add.)
>>
>> Separating adding residual from the transform seems to cause certain
>> slow down but  is needed to separate dc from idct which is faster overall,
>> which I consider a good reason to do this.
>
>
> I'm not sure I understand why, could you elaborate on this?
>
> Sure, simd IDCT is needed and I'm working on it.
>
>
> Great!
>

Btw I'm just noticing that all my comments apply to the ffmpeg codebase
also, so perhaps you should just ignore my comments and we can fix that
later on...

Ronald
_______________________________________________
libav-devel mailing list
[email protected]
https://lists.libav.org/mailman/listinfo/libav-devel

Reply via email to