Hi,

On Sun, Feb 19, 2012 at 7:42 AM, Ronald S. Bultje <[email protected]> wrote:
> Hi,
>
> On Sun, Feb 19, 2012 at 7:11 AM, Kostya Shishkov
> <[email protected]> wrote:
>> On Sun, Feb 19, 2012 at 07:05:04AM -0800, Ronald S. Bultje wrote:
>>> Hi,
>>>
>>> On Sun, Feb 19, 2012 at 7:02 AM, Kostya Shishkov
>>> <[email protected]> wrote:
>>> > On Sun, Feb 19, 2012 at 06:41:55AM -0800, Ronald S. Bultje wrote:
>>> >> Hi,
>>> >>
>>> >> On Sun, Feb 19, 2012 at 2:02 AM, Luca Barbato <[email protected]> wrote:
>>> >> > On 19/02/12 09:08, Kostya Shishkov wrote:
>>> >> >> This also makes encoder take DCT permutation into account as well.
>>> >> >> ---
>>> >> >>  libavcodec/proresdec.c |   10 +++++-----
>>> >> >>  libavcodec/proresdsp.c |   11 +++++++++++
>>> >> >>  libavcodec/proresdsp.h |    1 +
>>> >> >>  libavcodec/proresenc.c |   14 +++++++-------
>>> >> >>  4 files changed, 24 insertions(+), 12 deletions(-)
>>> >> >
>>> >> > Ok.
>>> >>
>>> >> Wait, what happened to make dspcontext smaller?
>>> >
>>> > nothing, look at the file list again
>>>
>>> Oh I see. So why is this in dsp then? We can't x86-SIMD this since it
>>> uses an indexed table (out[]), where the index is non-linear (scan[]).
>>
>> It's not scantable, it's permutation table. So if there's no permutation you
>> can simply multiply input matrix by quant without reordering.
>
> Right, but that only happens for C, and the annoying thing is that for
> C, there's no way to speed up that multiplication. See the catch-22
> here?

Let me be a little constructive. I'm all for putting this
multiplication in a more generic place, e.g. prores.c, so
encoder/decoder can share it. But I don't think it belongs in a dsp
file, since it cannot be DSPized.

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

Reply via email to