[FFmpeg-devel] Questions about Video Memory in qsv decoders

[Ivan Uskov]

Hello Michael,

Sunday, July 24, 2016, 12:16:59 AM, you wrote:

MN> On Sat, Jul 23, 2016 at 11:19:59PM +0300, Ivan Uskov wrote:
>> Hello 张玉晓,
>> 
>> Friday, July 22, 2016, 4:10:36 AM, you wrote:
>> 
>> 张> I have a question when learning ffmpeg qsv decoder and Intel media sdk.
>> 
>> 张> The intel media sdk suggest to use Video Memory while doing Hardware
>> 张> decoding, use System Memory while doing Software decoding.
>> 
>> 张> FFmpeg only used System Memory to decode with qsv (Intel msdk) in file
>> 张> libavcodec/qsvdec.c. Is there any problem when use Video Memory? 
>> 
>> As qsv modules maintainer I can confirm, yes, there are problems.
>> 
>> First (objective) problem that in common case there are 3 qsv components
>> should be linked by a pipeline through GPU memory: 
>> *qsv video decoder
>> *qsv vpp filter
>> *qsv video encoder
>> But current ffmpeg architecture have restrictions to share one common qsv
>> session between libavcodec and libavfilter.
>> 
>> Second  (subjective) problem that nobody in ffmpeg community is interested to
>> advance  QSV  codecs. Most of  patches at this year provided by me or by my
>> colleagues (including tries to solve issue with GPU memory) were rejected or
>> ignored silently. Moreover, if you will try to clone a latest ffmpeg you found qsv
>> decoder broken, it crashes at any try of use. The patch to fix it was posted
>> by me at April 25 but was ignored.

MN> which patches should be applied to fix the issues ?
MN> can you repost them ?
Currently I believe that broken merge from libav should be not improved/fixed
by some way but just reverted because it provides nothing helpful, only
complexity. I will provide new patch soon.

MN> [...]



-- 
Best regards,
 Ivan                            mailto:ivan.uskov at nablet.com