[FFmpeg-devel] [PATCH] h264: remove clear_blocks call in threading init.

[Ronald S. Bultje]

Hi,

On Tue, Feb 12, 2013 at 7:39 PM, Michael Niedermayer <michaelni at gmx.at> wrote:
> On Tue, Feb 12, 2013 at 06:30:15PM +0100, Michael Niedermayer wrote:
>> On Tue, Feb 12, 2013 at 09:18:34AM -0800, Ronald S. Bultje wrote:
>> > Hi,
>> >
>> > On Tue, Feb 12, 2013 at 9:05 AM, Michael Niedermayer <michaelni at gmx.at> wrote:
>> > > On Mon, Feb 11, 2013 at 06:08:45PM -0800, Ronald S. Bultje wrote:
>> > >> From: "Ronald S. Bultje" <rsbultje at gmail.com>
>> > >>
>> > >> Init code in that if statement goes down from 26716 cycles to 26047
>> > >> cycles, i.e. the removal of the clear_blocks and smaller memcpy()
>> > >> together save around 670 cycles.
>> > >> ---
>> > >>  libavcodec/h264.c | 7 +++----
>> > >>  1 file changed, 3 insertions(+), 4 deletions(-)
>> > >>
>> > >> diff --git a/libavcodec/h264.c b/libavcodec/h264.c
>> > >> index 2d6c5e6..994e002 100644
>> > >> --- a/libavcodec/h264.c
>> > >> +++ b/libavcodec/h264.c
>> > >> @@ -1253,7 +1253,9 @@ static int decode_update_thread_context(AVCodecContext *dst,
>> > >>
>> > >>          // copy all fields after MpegEnc
>> > >>          memcpy(&h->s + 1, &h1->s + 1,
>> > >> -               sizeof(H264Context) - sizeof(MpegEncContext));
>> > >> +               offsetof(H264Context, intra_gb) - sizeof(MpegEncContext));
>> > >> +        memcpy(&h->cabac, &h1->cabac,
>> > >> +               sizeof(H264Context) - offsetof(H264Context, cabac));
>> > >>          memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
>> > >>          memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
>> > >>
>> > >
>> > > If either of these fields is moved in the structure or a new field
>> > > added between then threading can mysteriously break.
>> > > also the comment above would not match the code anymore
>> > > iam not sure this is a good idea but if it is done then the fields
>> > > should be documented in the struct so a developer touching them would
>> > > know they are special ...
>> >
>> > Isn't that the case already? The non-init code in
>> > update_thread_context() for mpeg and h264 100% depends on ordering of
>> > the fields, any field ordering change will break the threading.
>>
>> yes, the problem exists already with some fields
>
> also what i forgot, iam happy to add these comments/documentation to
> the struct/fields myself if you prefer ...

Yeah that'd be helpful, thanks.

(And yes I sometimes fall off the radar on weekdays - work = busy.)

Ronald