Memory Leak: Neural Network Allocation Not Freed

[Amit-Matth]

Description
The audiomark benchmark initialization routine allocates memory for the Neural Network (NN) component via th_nn_init but acts as a "fire and forget" allocation, failing to release it during the teardown phase.

I identified this issue while reviewing src/ee_audiomark.c, where I found an explicit TODO comment indicating that the de-initialization was intended but not implemented: // TODO: De-init NN allocs?.

Location

• Allocation: In ports/arm/th_api.c, the function th_nn_init() allocates memory to ctx.buf.
• Missing Deallocation: In src/ee_audiomark.c, the ee_audiomark_release() function frees other components (BMF, AEC, ANR, KWS) but skips the Neural Network memory.

Impact
This results in a memory leak every time the benchmark is run. While the OS may reclaim memory after the process terminates, this behavior is problematic for embedded environments or scenarios where the benchmark functions are called repeatedly within a larger application or test harness.

Code Snippets
src/ee_audiomark.c:

void
ee_audiomark_release(void)
{
    th_free(p_bmf_inst, COMPONENT_BMF);
    th_free(p_aec_inst, COMPONENT_AEC);
    th_free(p_anr_inst, COMPONENT_ANR);
    th_free(p_kws_inst, COMPONENT_KWS);
    // TODO: De-init NN allocs?
}

[Amit-Matth]

Hi @joseph-yiu,

I discovered this issue while exploring the codebase and would like to work on it.

I have a possible fix in mind that involves adding a proper destruction/cleanup function to the Neural Network API and ensuring it is called during the benchmark release phase.

Could you please validate this issue? Once confirmed, I can proceed with the implementation and submit a PR.

Thanks!

[joseph-yiu]

HI @Amit-Matth ,
Thanks for reporting the issue. Okay, look like this is specific to Arm's port. I will check with Arm's software team. But it seems this need to change ee_audiomark_release() to add a function call that is "ports" specific. (We can declare that as a weak function so that it can be overridden.)

Usually for benchmarking the device would be reset at each benchmark run. But I agree that it is better to release the heap at the end of the benchmark execution.

regards,
Joseph

P.S. I am out of office next few days so will be slower in responding to messages here.

[Amit-Matth]

HI @Amit-Matth , Thanks for reporting the issue. Okay, look like this is specific to Arm's port. I will check with Arm's software team. But it seems this need to change ee_audiomark_release() to add a function call that is "ports" specific. (We can declare that as a weak function so that it can be overridden.)

Usually for benchmarking the device would be reset at each benchmark run. But I agree that it is better to release the heap at the end of the benchmark execution.

regards, Joseph

P.S. I am out of office next few days so will be slower in responding to messages here.

Thanks for the clarification, Joseph.

That makes sense - I agree this looks specific to the Arm port. Declaring a weak, port-specific cleanup function and calling it from ee_audiomark_release() sounds like a clean approach.

I’m happy to implement this once you’ve had a chance to confirm with the Arm software team. No rush , I’ll wait for your guidance.

Thanks, and enjoy your time off.

[joseph-yiu]

Closing this issue as the AudioMark v1.0.4 release is completed.