One question about VTA runtime performance


#1

Hi,
I’m studying VTA now. From tutorial https://docs.tvm.ai/vta/tutorials/resnet.html#sphx-glr-vta-tutorials-resnet-py, we can see file graphlib.o is generated and upload for ARM running.
During runing, Instructions and Micro ops are generated (JIT compiling) and appended into two queues seperatedly. Instructions are run by synchronize.
My concern is: for one fix network like ResNet, we use it to for many figures. For each figure, are Instructions and Micro ops are generated and run each time ? Can we generate Instructions and Micro ops once, and run them multiple times for each input figure?
Thank you