Running my inference python script with multiple images to process results in TIDL segfaulting on the host.
Debugging shows that the custom layer function TIDL_customProcess allocates memory via TIDL_getMemoryChunkFromSystem and does not check if the return value was NULL. It then does a memcpy into this result, which causes the segfault.
Further investigation shows that the reason NULL is returned is because TIDL_resetSysmem is only called during TIDL_init, but not during processing between layers (except for some reason conv2d_base calls it during processing, but if you don't have that layer that doesn't help you!)
Note, on one of the calls to TIDL_resetSysmem during init, there is a comment "System memories are expected to be reset for every layer", but that doesn't seem to be happening.
Thus with enough images supplied causing repeated calls to TIDL_customProcess, the memory is fully consumed and eventually it returns null, causing the problem.