MKLDNN in R4 cannot max all CPU cores (50% CPU utilization)

[nikos1]

When testing YOLO3 I can only utilize 50% of the CPU cores on a i7-8700K CPU ( Ubuntu 16.04 LTS linux kernel 4.15.0-33-generic )

./object_detection_demo_yolov3_async -i test.mp4 -m fp32/yolo_v3.xml -d CPU

[nikos1]

Same issue with SSD

./object_detection_demo_ssd_async -i test.mp4  -m SSD_fp32.xml  -d CPU

[Severine_H_Intel · Accepted solution]

Dear Nikos,

I could reproduce your issue. You need to run the following command before calling your code: export OMP_WAIT_POLICY=PASSIVE

And OpenMP will use all the cores when running the code. 

Best, 

Severine

[nikos1]

Thank you Severine, the following worked, fps still remain the same though.

export OMP_WAIT_POLICY=PASSIVE

[Afrath__Mohamed]

Hi Nikos, i would like to convert yolov3 daraknet weights to openvino ir format using model optimizer? can you give you model optimizer command options for the same? or the procedure that you used for  "fp32/yolo_v3.xml" ?

[Afrath__Mohamed]

Hi Severine, i face the same low fps issue with openvino R4 release. Where R3 release was faster. After migrated to R4 release the fps dropped by half. Even tried export OMP_WAIT_POLICY=PASSIVE command, there is minimal increase, while running inference my CPU loaded  upto 89%(4 cores).

My PC configurarion:

Model : Intel(R) Core(TM) i5-4460  CPU @ 3.20GHz

OS: Linux 4.10.0-28-generic 16.04.2-Ubuntu