ergoCubEmotions uses more than 100% of the cpu on ergocub-head

[GiulioRomualdi]

I noticed that ergoCubEmotions uses a really high amount of CPU on the ergocub-head.

In the following screenshot you can check the amount of used CPU (181%) only ergoCubEmotions was running on the ergocub-head

I was wondering if it is normal

cc @martinaxgloria

[traversaro]

If I remember correctly once we checked that, one problem may be that at every module update, the file are re-read (see

ergocub-software/src/modules/ergoCubEmotions/ergoCubEmotions.cpp

Line 119 in eaacbf1

bool isTransition_local;

). Perhaps we could read the images only once in the configure, and save them as cv::Mat attributes that can be used in the updateModule?

[martinaxgloria]

Hi @traversaro,

thank you for your suggestion. Actually, we are reading videos and not images (the static part is a 2 sec looped video and the transitions from one expression to another too). If I remember well, opencv shows videos one frame at the time after they are opened. I don't know if it's possible to save cv::VideoCapture elements as per cv::Mat ones.

Today I tried to use the configuration image -> transition -> image and I saw from jtop that the usage is reduced to ~30% but there's a spike when the transition is called.

Moreover, the % showed by jtop should correspond to the sum of the cpu usage for each core: in fact running top with IRIX mode set to OFF, this is the cpu consumption divided by the number of cores:

This means that the cores should not be saturated, I don't know if you see some delays.
I'm going to understand if it's possible to prevent the videos from being reopened at each update

cc @Nicogene @GiulioRomualdi

[martinaxgloria]

Sorry, I closed it by mistake.

[martinaxgloria]

I opened a draft PR with the suggestions of @traversaro.

• ergoCubEmotions: avoid reading videos at each update of the module #214

It's in draft since I have to test it on ergocub-head

cc @Nicogene @GiulioRomualdi

[Nicogene]

The @martinaxgloria fix improves significantly the performance, @GiulioRomualdi in case not sufficient feel free to reopen it.