Add test for successive armnn model loads with diff thread counts

[jishminor]

Signed-off-by: Josh Minor [email protected]

[jishminor]

This PR attempts to address #5, dealing with concurrent execution of ArmNN delegated Tflite models.

It adds a test that verifies when multiple armnn delegated models are loaded in sequence a segfault does not occur. It appears that requesting a load with fewer threads works correctly, whereas the inverse fails.

The design of the Arm compute library scheduler is a singelton, meaning all armnn delegated models share the same scheduler. Because of this, the runtime doesn't like when one models sets a number of threads, then another model follows up by setting a lesser number of threads.

[jishminor]

As a temporary hacky solution, I just simply maintain a global int for the min number of threads requested by an armnn delegated model so far, and if another model requests more threads, than it gets overwritten by current value of that global.

It's also important to note that the thread pool seems to be shared by all armnn models, rather than getting a unique threads per model for the execution of the armnn/acl op kernels.