I am trying to profile the system resource consumption for models running on SNPE. For profiling DSP usage for any usecase, I am not able to find any system level entries or utility. Below are the system configuration of my setup.
Chipset: QCS605
SDK: Linux Embedded (r45)
SNPE version: 1.39
I have obtained the CPU and GPU usage with system stats with top command and /sys/kernel/gpu/gpu_busy entry. I have also benchmarked the performances with SNPE utilities. Similarly, I would like to profile the DSP usage of any model running in SNPE (DSP runtime).
Can I get a possible solution for measuring DSP usage at any instance ? If it is not possible at system level, is there any application or utility to achieve the same ?
Thanks,
Varshini Rajendran
Hi varshini.rajendran,
Yes, you can use the Benchmarking tool from NPE to get the model analysis and inference time with DSP as runtime.
Kindly use the instructions on the usage of the tool from the below link,
https://developer.qualcomm.com/software/qualcomm-neural-processing-sdk/learning-resources/vision-based-ai-use-cases/performance-analysis-using-benchmarking-tools
You can also use the same tool for CPU & GPU analysis on your end device.
Thanks for the response !
I have used the benchmarking tool and obtained inferences per second. But I'm concerned more of measuring the usage percentage of DSP at any instance. This data can be used to plan any concurrency of activities in DSP runtime.
I would appreciate any help on this.
Thanks,
Varshini Rajendran.