[FFmpeg-devel] [PATCH 14/17 v2] vp8: Add hwaccel hooks

[Mark Thompson]

On 09/01/17 00:11, Mark Thompson wrote:
> Also adds some extra fields to the main context structure that may
> be needed by a hwaccel decoder.
> 
> (cherry picked from commit 4e528206bc4d968706401206cf54471739250ec7)
> ---
> On 08/01/17 23:37, Michael Niedermayer wrote:
>> On Mon, Jan 09, 2017 at 12:32:30AM +0100, Michael Niedermayer wrote:
>>> On Sun, Jan 08, 2017 at 06:14:45PM +0000, Mark Thompson wrote:
>>>> Also adds some extra fields to the main context structure that may
>>>> be needed by a hwaccel decoder.
>>>>
>>>> (cherry picked from commit 4e528206bc4d968706401206cf54471739250ec7)
>>>> ---
>>>>  libavcodec/vp8.c | 193 +++++++++++++++++++++++++++++++++++++------------------
>>>>  libavcodec/vp8.h |  32 +++++++++
>>>>  2 files changed, 161 insertions(+), 64 deletions(-)
>>>
>>> This breaks
>>>
>>> make fate-webp-rgba-lossy-q80
>>>
>>> TEST    webp-rgba-lossy-q80
>>> --- ./tests/ref/fate/webp-rgba-lossy-q80        2017-01-08 23:01:44.560754747 +0100
>>> +++ tests/data/fate/webp-rgba-lossy-q80 2017-01-09 00:32:50.896869909 +0100
>>> @@ -1,6 +0,0 @@
>>> -#tb 0: 1/25
>>> -#media_type 0: video
>>> -#codec_id 0: rawvideo
>>> -#dimensions 0: 12x8
>>> -#sar 0: 0/1
>>> -0,          0,          0,        1,      240, 0xc56c96c3
>>> Test webp-rgba-lossy-q80 failed. Look at tests/data/fate/webp-rgba-lossy-q80.err for details.
>>> make: *** [fate-webp-rgba-lossy-q80] Error 139
>>
>> ==5840== Invalid write of size 1
>> ==5840==    at 0xC3FCB5: vp8_lossy_decode_alpha (webp.c:1304)
>> ==5840==    by 0xC3FE51: vp8_lossy_decode_frame (webp.c:1346)
>> ==5840==    by 0xC40146: webp_decode_frame (webp.c:1404)
>> ==5840==    by 0xB41EB8: avcodec_decode_video2 (utils.c:2258)
>> ==5840==    by 0xB43E6D: do_decode (utils.c:2789)
>> ==5840==    by 0xB44268: avcodec_send_packet (utils.c:2878)
>> ==5840==    by 0x756D6A: try_decode_frame (utils.c:2983)
>> ==5840==    by 0x759BCD: avformat_find_stream_info (utils.c:3684)
>> ==5840==    by 0x41BA22: open_input_file (ffmpeg_opt.c:1021)
>> ==5840==    by 0x425205: open_files (ffmpeg_opt.c:3222)
>> ==5840==    by 0x425388: ffmpeg_parse_options (ffmpeg_opt.c:3262)
>> ==5840==    by 0x43AE28: main (ffmpeg.c:4553)
>> ==5840==  Address 0x0 is not stack'd, malloc'd or (recently) free'd
> 
> Hmm, thanks for noticing that.
> 
> The webp decoder for an image with an alpha plane is abusing the vp8 decoder by overwriting the pix_fmt and then assuming that it won't be set again via ff_get_format().  The crash comes because the webp decoder has set pix_fmt to YUVA420P and assumed that the vp8 decoder will allocate its frames as that, but the vp8 decoder has found a YUV420P stream (as it always does, since that is the only format VP8 can be) and therefore allocated YUV420P frames.  When the webp decoder then tries to copy into the alpha plane it dies because there is no alpha plane.
> 
> I can hack around it with something like following?  (Whole patch including it surrounding.)
> 
> -    if (!is_vp7 && s->pix_fmt == AV_PIX_FMT_NONE) {
> +    if (!is_vp7 && avctx->pix_fmt == AV_PIX_FMT_YUVA420P) {
> +        // Hack: this is actually the webp decoder, don't call ff_get_format().
> +        s->pix_fmt = AV_PIX_FMT_YUV420P;
> +    } else if (!is_vp7 && s->pix_fmt == AV_PIX_FMT_NONE) {
> 
> Alternatively, we could add something explicit to signal to the vp8 decoder that the intended format is fixed and ff_get_format() shouldn't be called?

(Patches 1-13 applied.)

Ping for any opinion what to do with the webp use of the vp8 decoder.

Thanks,

- Mark