Memory leak in NN SDK

VIM4
1

Which system do you use? Android, Ubuntu, OOWOW or others?

Armbian 24.2.1 Bookworm

Please describe your issue below:

The memory constantly grows when using functions aml_module_create followed by successive calls to aml_module_input_set and aml_module_output_get with aml_output_format_t set as AML_OUTDATA_FLOAT32 or aml_module_output_get_simple.
To solve this issue I used aml_module_output_get with aml_output_format_t set as AML_OUTDATA_RAW.

libadla version is this one https://github.com/numbqq/libadla_deb_aml/raw/main/jammy/arm64/libadla_0.4-202307_arm64.deb

My project where I needed to use the NN SDK is there : https://github.com/sousmangoosta/libadla_interface

2

Hi,

Thanks for reporting the issue. Armbian Images are in quite an early stage and hardware decoding and neural network support was not tested on the same.

Would it be possible to use fenix created images instead until Armbian images get sorted out? You can create bookworm images using GitHub - khadas/fenix: One-stop script set to build Ubuntu/Debian images. I have also built the same for you that you can download from here

3

Hi @viraniac,

Thanks for your answer, I didn’t see it before :face_with_raised_eyebrow:
I will try and report.

4

On Fenix 1.6.6 Debian GNU/Linux 12 (bookworm) Linux 5.15.119
I got same issue if I use the version before my fix, the memory constantly grow, with the fix the memory is stable, so definitely something in NN library.