[FFmpeg-devel] [PATCH 5/6] avcodec/qtrle: return last frame even if unchanged
James Almer
jamrial at gmail.com
Mon Aug 26 17:41:37 EEST 2019
On 8/26/2019 4:32 AM, Michael Niedermayer wrote:
> On Sun, Aug 25, 2019 at 08:04:03PM -0300, James Almer wrote:
>> On 8/25/2019 6:46 PM, Michael Niedermayer wrote:
>>> On Sun, Aug 25, 2019 at 01:22:22PM -0300, James Almer wrote:
>>>> On 8/24/2019 3:18 PM, Michael Niedermayer wrote:
>>>>> Fixes: Ticket7880
>>>>>
>>>>> Signed-off-by: Michael Niedermayer <michael at niedermayer.cc>
>>>>> ---
>>>>> libavcodec/qtrle.c | 30 ++++++++++++++++++++++++++----
>>>>> tests/ref/fate/qtrle-8bit | 1 +
>>>>> 2 files changed, 27 insertions(+), 4 deletions(-)
>>>>>
>>>>> diff --git a/libavcodec/qtrle.c b/libavcodec/qtrle.c
>>>>> index 2c29547e5a..c22a1a582d 100644
>>>>> --- a/libavcodec/qtrle.c
>>>>> +++ b/libavcodec/qtrle.c
>>>>> @@ -45,6 +45,7 @@ typedef struct QtrleContext {
>>>>>
>>>>> GetByteContext g;
>>>>> uint32_t pal[256];
>>>>> + AVPacket flush_pkt;
>>>>> } QtrleContext;
>>>>>
>>>>> #define CHECK_PIXEL_PTR(n) \
>>>>> @@ -454,11 +455,27 @@ static int qtrle_decode_frame(AVCodecContext *avctx,
>>>>> int has_palette = 0;
>>>>> int ret, size;
>>>>>
>>>>> + if (!avpkt->data) {
>>>>> + if (avctx->internal->need_flush) {
>>>>> + avctx->internal->need_flush = 0;
>>>>> + ret = ff_setup_buffered_frame_for_return(avctx, data, s->frame, &s->flush_pkt);
>>>>> + if (ret < 0)
>>>>> + return ret;
>>>>> + *got_frame = 1;
>>>>> + }
>>>>> + return 0;
>>>>> + }
>>>>> + s->flush_pkt = *avpkt;
>>>>> + s->frame->pkt_dts = s->flush_pkt.dts;
>>>>> +
>>>>> bytestream2_init(&s->g, avpkt->data, avpkt->size);
>>>>>
>>>>> /* check if this frame is even supposed to change */
>>>>> - if (avpkt->size < 8)
>>>>> + if (avpkt->size < 8) {
>>>>> + avctx->internal->need_flush = 1;
>>>>> return avpkt->size;
>>>>> + }
>>>>> + avctx->internal->need_flush = 0;
>>>>>
>>>>> /* start after the chunk size */
>>>>> size = bytestream2_get_be32(&s->g) & 0x3FFFFFFF;
>>>>> @@ -471,14 +488,18 @@ static int qtrle_decode_frame(AVCodecContext *avctx,
>>>>>
>>>>> /* if a header is present, fetch additional decoding parameters */
>>>>> if (header & 0x0008) {
>>>>> - if (avpkt->size < 14)
>>>>> + if (avpkt->size < 14) {
>>>>> + avctx->internal->need_flush = 1;
>>>>> return avpkt->size;
>>>>> + }
>>>>> start_line = bytestream2_get_be16(&s->g);
>>>>> bytestream2_skip(&s->g, 2);
>>>>> height = bytestream2_get_be16(&s->g);
>>>>> bytestream2_skip(&s->g, 2);
>>>>> - if (height > s->avctx->height - start_line)
>>>>> + if (height > s->avctx->height - start_line) {
>>>>> + avctx->internal->need_flush = 1;
>>>>> return avpkt->size;
>>>>> + }
>>>>> } else {
>>>>> start_line = 0;
>>>>> height = s->avctx->height;
>>>>> @@ -572,5 +593,6 @@ AVCodec ff_qtrle_decoder = {
>>>>> .init = qtrle_decode_init,
>>>>> .close = qtrle_decode_end,
>>>>> .decode = qtrle_decode_frame,
>>>>> - .capabilities = AV_CODEC_CAP_DR1,
>>>>> + .caps_internal = FF_CODEC_CAP_SETS_PKT_DTS | FF_CODEC_CAP_SETS_PKT_POS,
>>>>> + .capabilities = AV_CODEC_CAP_DR1 | AV_CODEC_CAP_DELAY,
>>>>> };
>>>>> diff --git a/tests/ref/fate/qtrle-8bit b/tests/ref/fate/qtrle-8bit
>>>>> index 27bb8aad71..39a03b7b6c 100644
>>>>> --- a/tests/ref/fate/qtrle-8bit
>>>>> +++ b/tests/ref/fate/qtrle-8bit
>>>>> @@ -61,3 +61,4 @@
>>>>> 0, 160, 160, 1, 921600, 0xcfd6ad2b
>>>>> 0, 163, 163, 1, 921600, 0x3b372379
>>>>> 0, 165, 165, 1, 921600, 0x36f245f5
>>>>> +0, 166, 166, 1, 921600, 0x36f245f5
>>>>
>>>> Following what i said in the nuv patch, do you still experience timeouts
>>>> with the current codebase, or even if you revert commit
>>>> a9dacdeea6168787a142209bd19fdd74aefc9dd6? Creating a reference to an
>>>> existing frame buffer shouldn't be expensive anymore for the fuzzer
>>>> after my ref counting changes to target_dec_fuzzer.c
>>>>
>>>> This is a very ugly solution to a problem that was already solved when
>>>> reference counting was introduced. Returning duplicate frames to achieve
>>>> cfr in decoders where it's expected shouldn't affect performance.
>>>
>>> Maybe i should ask this backward to make it clearer what this is trying
>>> to fix.
>>>
>>> Consider a patch that would return every frame twice with the same ref
>>> of course.
>>> Would you consider this to have 0 performance impact ?
>>
>> No, but it would be negligible.
>
> ok, lets see some actual numbers
>
> This comparission is about fixing ticket 7880 (just saying so we dont loose
> track of what we compare here)
>
> this qtrle patchset which fixes the last frame
> time ./ffmpeg -i clock.mov -vf scale=1920x1080 -qscale 2 -y test-new.avi
> real 0m0.233s
> user 0m0.404s
> sys 0m0.036s
> -rw-r----- 1 michael michael 769212 Aug 26 08:38 test-new.avi
> time ./ffmpeg -i test-new.avi -f null -
> real 0m0.095s
> user 0m0.131s
> sys 0m0.038s
>
> The alternative of reverting the code that discards duplicate frames
> (this i believe is what the people opposing this patchset here want)
> git revert a9dacdeea6168787a142209bd19fdd74aefc9dd6
> time ./ffmpeg -i clock.mov -vf scale=1920x1080 -qscale 2 -y test.avi
> real 0m1.208s
> user 0m2.866s
> sys 0m0.168s
> -rw-r----- 1 michael michael 3274430 Aug 26 08:44 test.avi
> time ./ffmpeg -i test.avi -f null -
> real 0m0.185s
> user 0m0.685s
> sys 0m0.076s
>
> https://samples.ffmpeg.org/ffmpeg-bugs/trac/ticket7880/clock.mov
>
> As we can see the solution preferred by the opposition to this patchset
> will make the code 6 times slower and result in 4 times higher bitrate for
> the same quality.
> I would say this is not "negligible"
> you know this is not 0.6%, its not 6% its not 60% it is 600%
>
> I do not understand why we have a discussion here about this.
> Do we want 600% slower code ?
> Do our users want 600% slower code ?
>
> I do not want 600% slower code
I was talking about the decoder outputting the frames would have
negligible performance effect. It's after all generating a new reference
to an existing buffer. I never said passing 100 frames instead of 20 to
libavfilter and expect it to process them all would be just as fast,
because that's obviously not going to happen.
By flagging frames as "disposable" however, libavfilter (or any library
user) could drop them on sight, achieving the speed up you mentioned. Or
by flagging them as discard at the user's request, the generic
libavcodec would never propagate them (aka, generate the current forced
vfr behavior post a9dacdeea6).
I'm fine with either solution, but i want other people's opinion. One
requires a new frame flag and no extra work in generic lavc code (what
to do with them is up to the user), whereas the other will require a new
lavc option or avctx flag to let the user request dropping these frames.
libavfilter is uncharted land for me, btw, so i have no idea how to have
it look for the aforementioned disposable flag and properly drop them
without disturbing the filtering process. But perhaps that should be the
job of whatever calls libavfilter instead.
More information about the ffmpeg-devel
mailing list