Hi,
I got such crash log when I executed image classification using DSP in sample android app eventhough it worked well.
com.qualcomm.qti.snpe.imageclassifiers: vendor/qcom/proprietary/adsprpc/src/fastrpc_apps_user.c:836::Error 4b: hlist[domain].domainsupport
com.qualcomm.qti.snpe.imageclassifiers: vendor/qcom/proprietary/adsprpc/src/fastrpc_apps_user.c:845::Error 4b: adsp current process handle failed. domain 0
Could you let me know the reason of crash log and it was really fine?
Thanks for trying out SNPE.
Could you let us know which device are you running this on?
Hi, jesliger
I run it on Galaxy S8(SM-G9500). Hope it is helpful information.
We do see errors in the log even when things are working fine. We'll double check on an S8 as well, when we get a chance. If it's working, you don't need to worry about the errors. It might be related to our DSP discovery mechanism trying things.. getting errors.. then trying something else (we do this to discover which target/DSP we are running on.)
Thank you for answer.
I experienced DSP performance was fluctuated according to CPU frequency.
Thus, I wonder there are some possibility of DSP latency due to discovery mechanism error.
We only do discovery during initialization/loading time, we don't do it once SNPE is up and running and executing networks.
Inference time fluctuations can occur due to CPU performance when using the DSP runtime as the CPU has to quantize the input tensor from float to 8bit fixed point before sending it to the DSP. We've seen this take anywhere from 8ms to 25ms for a 299x299x3 input size (for inception V3)
There are features in newer releases that help mitigate this quantization overhead. Which release are you using?
Thank you, jesliger. Your answer is very helpful to understand phenomenon.
I used SNPE v1.8.0 with quantized resnet-50.
Additionally, I discovered DSP performance increasing when CPU frequency was increase for file write simultaneously.
I could not found a way to increase CPU frequency with DSP performance in SNPE v1.8.0 eventhough I set HIGH_PERFORMANCE as performance profile.