Does tflite utilize shared memory architecture?

[김민재]

Hi,

I am currently investigating the performance overhead of tensorflow lite.

In tflite, when the input model has some unsupported operators in GPUDelegate, then the operators fallback to the CPU.

AFAIK, recent mobile devices like the Qualcomm Adreno and ARM Mali GPU use a shared memory architecture, which can enable us to not copy the intermediate tensors when switching the target processor.

However, the OpenCL backend and OpenGL backend seem to not utilize this kind of optimization. Is there any plan to exploit this, or is there any reason that we need to copy the tensors?