[FFmpeg-devel] [PATCH 5/5] avcodec/vc1: store zero MVs for all blocks in a MB

[James Almer]

On 5/19/2018 6:40 PM, Michael Niedermayer wrote:
> On Fri, May 18, 2018 at 05:06:36PM +0200, Jerome Borsboom wrote:
>> Direct prediction for interlace frame B pictures references the mv in the
>> second block in an MB in the backward reference frame for the twomv case.
>> When the backward reference frame is an I frame, this value may be unset.
>>
>> Signed-off-by: Jerome Borsboom <jerome.borsboom at carpalis.nl>
>> ---
>>  libavcodec/vc1_block.c | 6 ++++--
>>  1 file changed, 4 insertions(+), 2 deletions(-)
>>
>> diff --git a/libavcodec/vc1_block.c b/libavcodec/vc1_block.c
>> index 74935ec9e9..9c170a1e3c 100644
>> --- a/libavcodec/vc1_block.c
>> +++ b/libavcodec/vc1_block.c
>> @@ -2678,8 +2678,10 @@ static void vc1_decode_i_blocks_adv(VC1Context *v)
>>              s->bdsp.clear_blocks(block[0]);
>>              mb_pos = s->mb_x + s->mb_y * s->mb_stride;
>>              s->current_picture.mb_type[mb_pos + v->mb_off]                         = MB_TYPE_INTRA;
>> -            s->current_picture.motion_val[1][s->block_index[0] + v->blocks_off][0] = 0;
>> -            s->current_picture.motion_val[1][s->block_index[0] + v->blocks_off][1] = 0;
>> +            for (int i = 0; i < 4; i++) {
>> +                s->current_picture.motion_val[1][s->block_index[i] + v->blocks_off][0] = 0;
>> +                s->current_picture.motion_val[1][s->block_index[i] + v->blocks_off][1] = 0;
>> +            }
> 
> see AV_ZERO*

Those may use mmx instructions, which may require calling emms_c() after
the loop.
memset() may be a better choice here. For a small amount of bytes the
compiler should be able to optimize it into an inlined mov/xor on its own.

Or we could get rid of the mmx implementations from both AV_ZERO* and
AV_COPY* and forget about it.