03-25-2019 11:51 PM
I've set up host (ubuntu 16.04) and target tools (dnndk v2.08) for my ZCU 104 board. Resnet50, inception and face detection examples work like a charm. Video analysis example runs OK with sample video but starts to freeze after a while (about 30 seconds) with my own video file. Monitor is flickering as well. I changed the fps, bitrate, resolution and codec of my video to make it as similar as possible to sample video. Some fps improvement reveals but then it starts to freeze again and process is killed. I applied the patch provided in https://forums.xilinx.com/t5/Deephi-DNNDK/DNNDK-2-08-Beta-in-ZCU104-has-ERRORS/td-p/927686 but this do not change anything.
03-26-2019 04:52 PM
@nuriI encountered a very similar problem on my ZCU102. I cannot say with 100% certainty that it is the same problem. What (if anything) is reported in the console messages? Are you getting a DPU timeout message, perhaps similar to the message in this thread?
03-26-2019 05:30 PM
@nuriThe most probable cause of the monitor flickering is that the DPU and display interfaces are both contending for PS DDR bandwidth. Often the flickering issue can be cured by reducing the display resolution to some lower resolution sufficient to support display of the output video.
03-27-2019 12:18 AM
@qhall,decreasing screen resolution solves the flickering problem. But freezing problem still happens. Actually whole system is getting slow rather than executable. Even moving the mouse is impossible in that case. I do not know what exactly happens, evantually process is killed
03-28-2019 10:54 AM
@qhall , something weird happened today. Face detection project started to freeze as video analysis did. I saw cpu usage is around %90 when i “top” the processes. I apply power issue patch at every reboot. It seems like my board has some subtle hardware bug. I am kinda disappointed.
03-30-2019 10:00 AM
03-31-2019 01:59 AM - edited 03-31-2019 02:00 AM
There might be a memory leak in video analysis source codes. Here is the memory usage. It eats the memory fastly.
And DPU profiling
I compiled DPU and object detection example in https://github.com/Xilinx/Edge-AI-Platform-Tutorials/tree/master/docs/ML-SSD-PASCAL from scratch. It works perfect. No memory leak and dpu utilization interrupt