Please provide the following info (tick the boxes after creating this topic):
Software Version
DRIVE OS
Target Operating System
Linux
Hi，
I used trtexec to run deep learning model with DLA and have a question.
I found that there are still 14%-16% GPU usage even if the DLA model all ran on DLA.
Can you tell me the reason？
Dear @xinran.wang2,
Could you share the complete trtexec command and output log? How did you verify the usage?
We use command "./trtexec --onnx=model.onnx --useDLACore=0 --best --allowGPUFallback
The test result prove that “all layers run on DLA”.
And we use tegrastats to monitor the GPU usage. There are still have GPU usage.
Dear @xinran.wang2,
Does it run without allowGPUFallback flag? Do you see No Layer is fallback to GPU and all are mapped to DLA in output log? Could you share steps to reproduce the issue?
" Do you see No Layer is fallback to GPU and all are mapped to DLA in output log? "
Yes the log said that no Layer is fallback to GPU and all are mapped to DLA. As shown in figure.
Dear @xinran.wang2,
Could you check using nsys to confirm if GPU is used. It could be possible that TRT is inserting some reformat layers before DLA work. Also, signalling from GPU to DLA and back happens inside TRT when cuDLA hybrid mode is used. It could be interpreted as GPU usage by tegrastats.
Tegrastats is not recommend to check usage when GPU+DLA is invloved. We recommend to use nsys to get more clarify on the usage.
Can I use nsys on Orin?
Dear @xinran.wang2,
Yes. Please use nsys and check if you see any trace/usage on GPU
Sorry, I find that there are no nsys on driveOS 6.0. Could you tell me how to use it?
Thanks!
Dear @xinran.wang2,
Please use the tool from host and check remote profiling on target. Please check User Guide :: Nsight Systems Documentation
This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.
