Hi,
The diagnostics file creaeted by SNPE is very useful because it has each layer's running time.
I just want to know whether each layer's time is an average value of several infernece count.
Below is the diagnostics file:
Log File Created: Tue Jun 25 17:05:11 2019
Time Scale: 1e-06
Epoch Timestamp: 1561453511037503 Steady Clock Timestamp: 121377102436
Software library version: 1.23.1.245
Dnn Runtime Load/Deserialize/Create Statistics:
--------------------------------------------------
Load: 50432 us
Deserialize: 58137 us
Create: 534657 us
Average SNPE Statistics:
------------------------------
Total Inference Time: 23923 us
Forward Propagate Time: 23845 us
Layer Times:
---------------
0: 376 us : DSP
1: 0 us : DSP
2: 762 us : DSP
3: 0 us : DSP
4: 318 us : DSP
5: 0 us : DSP
6: 156 us : DSP
7: 150 us : DSP
8: 74 us : DSP
9: 0 us : DSP
10: 84 us : DSP
11: 0 us : DSP
12: 138 us : DSP
13: 132 us : DSP
14: 0 us : DSP
15: 15 us : DSP
16: 0 us : DSP
17: 134 us : DSP
18: 132 us : DSP
19: 0 us : DSP
...
Hi,
Yes, it is the average of all the inferences model made on a single execution.
To get individual count you can use the SNPE benchmarking tool, where you can get the summary of the inference time also the log file for each individual inference of the model.
Find the instructions on the usage of SNPE Benchmarking tool from the below link,
https://developer.qualcomm.com/docs/snpe/benchmarking.html
Thanks for your reply. I will try snpe_bench.py.
Another question: How to get each layer's time when model running on AIP mode ?
I tried to run resnet50 on HTA and got the SNPEDiag_0.log as following. Each layer's time is 0.
Log File Created: Tue Jun 25 16:45:51 2019
Time Scale: 1e-06
Epoch Timestamp: 1561452351506710 Steady Clock Timestamp: 120217493156
Software library version: 1.23.1.245
Dnn Runtime Load/Deserialize/Create Statistics:
--------------------------------------------------
Load: 38867 us
Deserialize: 44342 us
Create: 498144 us
Average SNPE Statistics:
------------------------------
Total Inference Time: 11158 us
Forward Propagate Time: 11037 us
Layer Times:
---------------
0: 0 us : AIP
1: 0 us : AIP
2: 0 us : AIP
3: 0 us : AIP
4: 0 us : AIP
5: 0 us : AIP
6: 0 us : AIP
7: 0 us : AIP
8: 0 us : AIP
9: 0 us : AIP
10: 0 us : AIP
...