Does tvm fulsh the cache between the fused groups

Questions
#1

Hi all,

How does the tvm flush cache between any fused groups execution?

I ran a model with two variants where the tensor passed between the two fused groups is int8 and int16. But while profiling both have the same execution time. I think there must be some speedup during tensor load/store from int16 to int8.

thanks.

#2

I don’t think inserting stop_fusion op flush the cache between the ops.

I would like to know how the cache is organized in tvm, and which source file drives that.

thanks.

#3

It doesn’t clear the cache when executing the model. During auto tuning on cpu, there is an option to flush the cache. https://github.com/apache/tvm/blob/main/python/tvm/meta_schedule/testing/tune_onnx.py#L155