• Hardware Platform (Jetson / GPU): dGPU • DeepStream Version: 6.0 GA • TensorRT Version: 8.0.0.1 • NVIDIA GPU Driver Version (valid for GPU only): 470.82 • Issue Type( questions, new requirements, bugs): Feature regression • How to reproduce the issue ?: Change file-loop to 1 in the deepstream-app sample config source4_1080p_dec_infer-resnet_tracker_sgie_tiled_display_int8.txtand run it with deepstream-app -c source4_1080p_dec_infer-resnet_tracker_sgie_tiled_display_int8.txt, application exits at EOS instead of looping.
The configuration file is apparently being parsed correctly with restart_stream_buf_prob (sources/apps/apps-common/src/deepstream_source_bin.c) being added to nvv4l2decoder but the EOS event never reaches the probe to trigger seek_decode, with the only events getting triggered there being caps and stream-start .
Is it related to this issue in the release notes? Though setting qos=1 didn’t seem to change anything.
GStreamer sink elements sending upstream QOS events lead to memory leaks in the
DeepStream pipeline. Applications should make sure to set the property “qos” to
FALSE explicitly. For reference applications, this can be done by setting qos=0 in all
sink groups of the configuration file
Is there any package in specific that could affect that behavior? In the same environment I can make loop work using the NGC image and the main difference seems to be quite a few package versions mismatching.
I have been installing the packages as needed, the version differences were mostly minor (e.g. 21.0.3-0ubuntu0.3~20.04.2 vs 21.0.3-0ubuntu0.3~20.04.3 for libgles2-mesa and such).
I was trying something similar to this dockerfile:
so that display works properly, but the in the resulting container setting file-loop to 1 does not work as expected.
For now I’ve decided to just use nvcr.io/nvidia/deepstream:6.0-triton as it seems to be the only Ubuntu 20.04 one and everything works as expected, only downside is the image size.