Hi Support Team,
Right now trying to run the ARMNN CaffeTest on iMX8 Board. I tried to run Inception-BNbatchsize1.
caffemodel with GPU Acceleration but the performance is not as expected. We are following the
i.MX_Machine_Learning_User's_Guide for running the ARMNN Tests.
We are running the Arm NN Caffe binary as below :
With CPU :
-----------
root@imx8qxpmek:/usr/bin/ArmnnTests# CaffeInception_BN-Armnn --compute=CpuAcc --data-dir=data --model-dir=models
ArmNN v20190801
= Prediction values for test #0
Top(1) prediction is 3694 with value: 0.940841
Top(2) prediction is 3197 with value: 0.000752777
Top(3) prediction is 2411 with value: 0.000105063
Top(4) prediction is 1972 with value: 4.33676e-05
Top(5) prediction is 1081 with value: 4.27122e-05
Total time for 1 test cases: 0.831 seconds
Average time per test case: 831.004 ms
Overall accuracy: 1.000
With GPU Backend:
----------------------------
root@imx8qxpmek:/usr/bin/ArmnnTests# CaffeInception_BN-Armnn --compute=VsiNpu --data-dir=data --model-dir=models
ArmNN v20190801
W [query_hardware_caps:53]Unsupported evis version
Warn-Start NN execution
Warn-Start NN execution
= Prediction values for test #0
Top(1) prediction is 3694 with value: 0.899821
Top(2) prediction is 3197 with value: 0.000630801
Top(3) prediction is 2411 with value: 0.000116962
Top(4) prediction is 1081 with value: 4.72108e-05
Top(5) prediction is 567 with value: 1.84702e-05
Total time for 1 test cases: 2.073 seconds
Average time per test case: 2073.426 ms
Overall accuracy: 1.000
In the above example with GPU acceleration ,we can see the Total time for the test cases taken is more
than the one computed with CPU.. and we can see this error when the demo is run with GPU acceleration:
W [query_hardware_caps:53]Unsupported evis version
Warn-Start NN execution
Warn-Start NN execution
and Total Time taken is more when run with GPU acceleration compared to CPU?
Can anyone please guide me what could be the issue?
Thanks..
