[Libva] [PATCH 2/4] Set the pipeline to use the new VP8 encoding shaders on BSW

Thu Jan 12 22:10:45 UTC 2017

On 12/01/17 07:30, Xiang, Haihao wrote:
> 
> Hi Mark,
> 
> Can you reproduce the issue you mentioned below? If yes, I would like to fix it
> in the new version of the patch series. 

Hi,

Yes, I can reproduce it consistently on both Skylake and Kaby Lake.  Some detailed instructions follow...

Get standard test input: <http://distribution.bbb3d.renderfarming.net/video/mp4/bbb_sunflower_1080p_60fps_normal.mp4>.  (The input file does matter somewhat: I tried some others and found it harder to reproduce.  Maybe the highly variable complexity here helps to show the problem.)

Get current libav from git: <git://git.libav.org/libav.git>.

Apply patches adding framerate configuration and VP8 encode support to libav: <http://ixia.jkqxz.net/~mrt/libva/vp8/0001-vaapi_encode-Pass-framerate-parameters-to-driver.patch>, <http://ixia.jkqxz.net/~mrt/libva/vp8/0002-vaapi_encode-Add-VP8-support.patch>.

Configure libav with --enable-vaapi and build.

Now run a transcode from the H.264 of the input file to VP8.  What happens varies with the bitrate selected (this is now on Skylake GT2, 6300):

./avconv -y -threads 1 -vaapi_device /dev/dri/renderD128 -hwaccel vaapi -hwaccel_output_format vaapi -i bbb_sunflower_1080p_60fps_normal.mp4 -an -c:v vp8_vaapi -r 60 -b:v 5M out.webm

(CBR at 5Mbps)  Everything works and the output looks good: yay!  (This is an immense improvement over the current driver - if you run the same command there, the output is working but terrible quality.)

./avconv -y -threads 1 -vaapi_device /dev/dri/renderD128 -hwaccel vaapi -hwaccel_output_format vaapi -i bbb_sunflower_1080p_60fps_normal.mp4 -an -c:v vp8_vaapi -r 60 -b:v 5M out.webm

(CBR at 2Mbps)  Mostly works, but the output is broken at times and the bitrate target is often missed by a long way.

./avconv -y -threads 1 -vaapi_device /dev/dri/renderD128 -hwaccel vaapi -hwaccel_output_format vaapi -i bbb_sunflower_1080p_60fps_normal.mp4 -an -c:v vp8_vaapi -r 60 -b:v 1M out.webm

(CBR at 1Mbps)  Consistently hangs at around frame 570.

[1321697.079583] drm/i915: Resetting chip after gpu hang
[1321697.081571] [drm] GuC firmware load skipped
[1321699.063831] [drm] RC6 on

/sys/class/drm/card0/error for one case: <http://ixia.jkqxz.net/~mrt/libva/vp8/sys_class_drm_card0_error>.

Backtrace of userspace at the hang point:

#0  0x00007ffff6351cc7 in ioctl () at ../sysdeps/unix/syscall-template.S:84
#1  0x00007ffff5821708 in drmIoctl () from /usr/lib/x86_64-linux-gnu/libdrm.so.2
#2  0x00007ffff48e3e00 in ?? () from /usr/lib/x86_64-linux-gnu/libdrm_intel.so.1
#3  0x00007ffff48e4036 in ?? () from /usr/lib/x86_64-linux-gnu/libdrm_intel.so.1
#4  0x00007ffff4bded57 in intel_batchbuffer_flush (batch=0x555556c03240) at ../../src/intel_batchbuffer.c:147
#5  0x00007ffff4b9c821 in i965_run_kernel_media_object (ctx=0x555556ad15b0, encoder_context=0x555556bfff10, gpe_context=0x555556b6da38, media_function=11, param=0x7fffffffd5f0) at ../../src/i965_encoder_vp8.c:2174
#6  0x00007ffff4baa044 in i965_encoder_vp8_pak_tpu (ctx=0x555556ad15b0, encode_state=0x555556ae5cd0, encoder_context=0x555556bfff10) at ../../src/i965_encoder_vp8.c:6503
#7  0x00007ffff4baa5a7 in i965_encoder_vp8_pak_pipeline (ctx=0x555556ad15b0, profile=VAProfileVP8Version0_3, encode_state=0x555556ae5cd0, encoder_context=0x555556bfff10) at ../../src/i965_encoder_vp8.c:6620
#8  0x00007ffff4b988b3 in intel_encoder_end_picture (ctx=0x555556ad15b0, profile=VAProfileVP8Version0_3, codec_state=0x555556ae5cd0, hw_context=0x555556bfff10) at ../../src/i965_encoder.c:1313
#9  0x00007ffff4b8aa23 in i965_EndPicture (ctx=0x555556ad15b0, context=33554433) at ../../src/i965_drv_video.c:3588
#10 0x00007ffff76744ca in vaEndPicture (dpy=0x555556ad1540, context=33554433) at ../../git/va/va.c:1285
#11 0x0000555555df76d8 in vaapi_encode_issue (avctx=0x555556b5f100, pic=0x555556c64d20) at /home/mrt/video/libav/vp8/libavcodec/vaapi_encode.c:387
#12 0x0000555555df7f29 in vaapi_encode_step (avctx=0x555556b5f100, target=0x555556c64d20) at /home/mrt/video/libav/vp8/libavcodec/vaapi_encode.c:608
#13 0x0000555555df8be9 in ff_vaapi_encode2 (avctx=0x555556b5f100, pkt=0x555556b19380, input_image=0x555556b188c0, got_packet=0x7fffffffdcfc) at /home/mrt/video/libav/vp8/libavcodec/vaapi_encode.c:895
#14 0x00005555557a409b in avcodec_encode_video2 (avctx=0x555556b5f100, avpkt=0x555556b19380, frame=0x555556b188c0, got_packet_ptr=0x7fffffffdcfc) at /home/mrt/video/libav/vp8/libavcodec/encode.c:231
#15 0x00005555557a4280 in do_encode (avctx=0x555556b5f100, frame=0x555556b188c0, got_packet=0x7fffffffdcfc) at /home/mrt/video/libav/vp8/libavcodec/encode.c:278
#16 0x00005555557a444a in avcodec_send_frame (avctx=0x555556b5f100, frame=0x555556b188c0) at /home/mrt/video/libav/vp8/libavcodec/encode.c:327
#17 0x00005555555c3c5c in do_video_out (of=0x555556c61ba0, ost=0x555556c61680, in_picture=0x555556b188c0, frame_size=0x7fffffffe1cc) at /home/mrt/video/libav/vp8/avconv.c:579
#18 0x00005555555c43ce in poll_filter (ost=0x555556c61680) at /home/mrt/video/libav/vp8/avconv.c:718
#19 0x00005555555c4753 in poll_filters () at /home/mrt/video/libav/vp8/avconv.c:792
#20 0x00005555555cab23 in transcode () at /home/mrt/video/libav/vp8/avconv.c:2730
#21 0x00005555555cb063 in main (argc=20, argv=0x7fffffffe438) at /home/mrt/video/libav/vp8/avconv.c:2898

avconv there also supports CQP mode as well as CBR (pass "-qindex N" for N between 0 and 127 instead of the "-b:v M" above): this is always fine for me.

Can you reproduce a hang with that?  Is there any other information I can provide to help?

Thanks,

- Mark

>>> On Tue, Jan 10, 2017 at 4:21 PM, Mark Thompson <sw at jkqxz.net> wrote:
>>>> On 10/01/17 22:02, Sean V Kelley wrote:
>>>>> From: "Xiang, Haihao" <haihao.xiang at intel.com>
>>>>>
>>>>> Currently only one temporal layer is supported
>>>>>
>>>>> Signed-off-by: Xiang, Haihao <haihao.xiang at intel.com>
>>>>> Reviewed-by: Sean V Kelley <seanvk at posteo.de>
>>>>> ---
>>>>>   src/Makefile.am        |    3 +
>>>>>   src/gen8_encoder_vp8.c |  140 +
>>>>>   src/gen8_mfc.c         |    8 +-
>>>>>   src/gen8_vme.c         |    5 +
>>>>>   src/i965_defines.h     |   10 +
>>>>>   src/i965_encoder.c     |    2 +
>>>>>   src/i965_encoder_vp8.c | 6697
>>>> ++++++++++++++++++++++++++++++++++++++++++++++++
>>>>>   src/i965_encoder_vp8.h | 2643 +++++++++++++++++++
>>>>>   8 files changed, 9507 insertions(+), 1 deletion(-)
>>>>
>>>> I had a go with this on Kaby Lake.  In general, big win - looks like it
>>>> can
>>>> be under half the bitrate at comparable quality (though it was pretty
>>>> terrible before...).
>>>>
>>>> However, the rate control seems to do odd things at low bitrate relative
>>>> to
>>>> the frame size?  I can get GPU hangs and wildly varying output bitrate
>>>> with
>>>> it, though it seems ok at high bitrate.
>>> That's a concern.  Please report the If it really is a GPU hang, I need
>>> the error report for the DRM card0 log.
>>>
>>> cat /sys/class/drm/card0/error
>>>
>>> Please rerun and capture the DRM (i915) card0 error log.
>>>  
>>>>  
>>>> I had a look around the rate control and found two minor issues in the RC
>>>> configuration, though I don't think either of them are relevant to my
>>>> problem (see below).  I can try to make a reproducer if this is not
>>>> already
>>>> known?
>>>>
>>> Please do attempt to reproduce.  That's why I've put the patches out here to
>>> test.
>>
>> Thanks for testing the patch, could you detail the steps to reproduce this
>> issue?
>>
>>
>>>
>>> Thanks,
>>>
>>> Sean
>>>  
>>>>  Thanks,
>>>>
>>>> - Mark
>>>>