SNPE cannot work with DSP on Snapdragon 845
Join Date: 8 Jul 21
Posts: 8
Posted: Tue, 2021-10-19 20:08
Hi
I'm trying to run neural network with different runtimes on Snapdragon 845.
The snpe-net-run worked well on CPU and GPU.
But there are some errors when I used snpe-net-run on DSP.
There are the error info:
'''
error_code=910; error_message=DSP runtime system error. Failed to execute network. Error code: -6, layer id: 0, op type: 0, extra: 0, backendErrorCode: 0.; error_component=DSP Runtime; line_no=612; thread_id=547949793296
'''
But snpe-platform-validator --runtime dsp worked will:
'''
PF_VALIDATOR: DEBUG: Calling PlatformValidator->setRuntime
PF_VALIDATOR: DEBUG: Calling PlatformValidator->IsRuntimeAvailable
PF_VALIDATOR: DEBUG: CPU side validation passed.
PF_VALIDATOR: DEBUG: starting calculator test
PF_VALIDATOR: DEBUG: Loading DSP stub: libcalculator.so
PF_VALIDATOR: DEBUG: Successfully loaded DSP library - 'libcalculator.so'. Setting up pointers.
PF_VALIDATOR: DEBUG: Success in executing the sum function
Runtime DSP Prerequisites: Present.
'''
How can I do to solve this problem?
Thanks for you help.
I found a DSP bug, showing the similar error message. In my case, setting all zero values to the input buffer caused the error, and DSP runtime works normally by setting a random image.
I hope this helps.
Thank you
My model has some error.
When I change another model, DSP works good.
Hi,
This error can cause if your model consist of the layer which can not be supported by SNPE for execute of DSP, and in this if you have loaded your model only for DSP accelerator.
Please use All available Hardware Accelerator (['AIP', 'DSP', 'GPU', CPU']) in sequence, which allows to run model on multiple accelerator and if layer not supported it will fall back to another Accelerator and executes your network properly.
Hope It's answered your question.
Regards,
Sahil Bandar
In my situation doing platform checking showed err like bellow:
*******************************************************