Hi, I want to run my model with dsp but it only work on cpu.
Did i miss anything need to do?
I tried on both quantized and unquantized dlc, they gave same results.
Btw, I got same results when running snpe_bench.py
Env setting
Devices: nokia8.1 ta-1131
Soc: Snapdragon 845
snpe: 1.55.0.2958
Error message
1|PNXN:/data/local/tmp $ snpe-net-run --container split_head.dlc --input_list inputlist.txt --use_dsp
error_code=910; error_message=DSP runtime system error. error_code=910; error_message=DSP runtime system error. Failed to load network. Error code: 1, Backend Error Code: 0.; error_component=DSP Runtime; line_no=554; thread_id=521562877184; error_component=DSP Runtime; line_no=262; thread_id=534281082104
CPU result
1|PNXN:/data/local/tmp $ snpe-net-run --container split_head.dlc --input_list inputlist.txt
-------------------------------------------------------------------------------
Model String: N/A
SNPE v1.55.0.2958
-------------------------------------------------------------------------------
Processing DNN input(s):
test.raw
Applying padding
Applying padding
Applying padding
PNXN:/data/local/tmp $
Dear customer,
Did you quantize this model successfully on your side or any errors happened?
You can check the DSP logs through qxdm or mini-dm tool to capture the runing logs.
BR.
Wei
Hi,
I found this issue which caused by wrong .raw format.
I only feed 1/4 size compared to normal input.
Thanks for your help.
Hi kuanhungliu,
Please do make sure of that you have tried setting up the library path of the DSP.
Please follow the steps given in SNPE Docs in order to setup the DSP lib paths,
https://developer.qualcomm.com/sites/default/files/docs/snpe/dsp_runtime...
After this it should work well on DSP environment.
Solution to avoid Applying padding in input, please rescale your image in 0.0 to 1.0 value range and convert the dtype to float32.
After conversion store the file as a raw input file and test the inference.
Hope this might resolve your issue.
Thanks & Regards,
Sahil Bandar