Disclosure: This report was drafted with AI assistance (ChatGPT). The reproduction commands and reported behavior were tested and verified on actual hardware.
Environment
-
Jetson AGX Thor Developer Kit
-
Jetson Linux R38.2.0
-
GStreamer 1.24.2
-
NVIDIA
nvvideo4linux2plugin 1.14.0
Issue
With num-B-Frames=2, nvv4l2h264enc produces actual B-coded H.264 pictures but assigns identical, monotonically increasing PTS and DTS to every output buffer, despite picture reordering.
This results in incorrect MP4 presentation timestamps and breaks frame-accurate seeking.
Reproduction
gst-launch-1.0 -e -q \
videotestsrc num-buffers=18 pattern=ball ! \
'video/x-raw,format=I420,width=320,height=180,framerate=25/1' ! \
nvvidconv ! 'video/x-raw(memory:NVMM),format=NV12' ! \
nvv4l2h264enc profile=2 num-B-Frames=2 \
num-Ref-Frames=2 iframeinterval=8 idrinterval=8 ! \
h264parse ! qtmux ! filesink location=/tmp/test.mp4
ffprobe -v error -select_streams v:0 \
-show_entries packet=pts_time,dts_time \
-of csv=p=0 /tmp/test.mp4
ffprobe -v error -select_streams v:0 \
-show_entries frame=pict_type \
-of csv=p=0 /tmp/test.mp4
Actual result: Every packet has PTS=DTS (0.000/0.000, 0.040/0.040, etc.), although the stream contains 8 B-coded pictures out of 18.
Expected result: Presentation timestamps reflect the correct picture display order, while DTS reflects decoding order.
Additional findings
-
Incorrect timestamps are observable directly after
nvv4l2h264enc, beforeh264parse. -
Main/High profiles avoid the default Baseline+B-frame EOS hang but do not correct the timestamps.
-
nvh264encon the same Thor device produces correctly reordered timestamps and passes independent MP4 decoding and seeking tests.
Question: Is this a known V4L2 encoder issue on Thor R38.2.0? Is a driver or plugin fix available?