[FFmpeg-devel] Interpolation algorithms for FFmpeg
Tue Aug 3 22:55:10 CEST 2010
On Tue, Aug 3, 2010 at 1:08 PM, Sebastian Vater
<cdgs.basty at googlemail.com> wrote:
> Loren Merritt a ?crit :
>> On Tue, 3 Aug 2010, Sebastian Vater wrote:
>>> If I'm not wrong then most of the speed gain can be achieved by doing a
>>> 100% integer (fixed point) version?
>> float->int16 approximately doubles throughput. Or converting only most
>> of it to int16 (which I have done) is 1.8x so far (over nnedi3-0.9.1).
>> int8 would be faster, but also loses a signficant amount of precision,
>> so I ruled that out.
>> The remaining optimization idea I had involves exploiting isotropy to
>> reduce the number of edge orientations the ANN has to deal with. But
>> that involves retraining, so it'll take longer to test.
> Did you test float->int also? Also target CPU "optimal" int? Is there
> any difference?
What is "int"? Do you mean 32-bit int? 32-bit int will be exactly
twice as slow as 16-bit, given that you'll be able to fit 2x fewer
elements in each register. Possibly worse, given the lack of many
important 32-bit operations in SSE (and the abundance of 16-bit ones).
More information about the ffmpeg-devel