Size Limit on the Feature Maps for Batch Processing

[latifisalar]

Hello,

I am trying to generate the artifacts for TDA4 inference from tflite models. While trying to invoke the tflite_runtime interpreter with the tidl_delegate options on a batched input with the shape of [2, 224, 224, 8], I faced the following error in the network compiler traces:

ERROR: Batch processing is not supported with large feature map case

I can go through the process without any issues if the input shape is changed to [2, 16, 16, 8]

I was wondering what is the size/shape limit for batch processing if we want to still offload the operations to TIDL rather than using ARM processor.

Thanks