Does `zero padding` in CNN increase the inference time due to frequent data transfer between CPU and GPU?

28 views Asked by At

I think the zero-padding in CNN is a very annoying operation when designing a DNN accelerator or an NPU. So I want to know that if zero-padding is executed on CPU or GPU in modern ML framework like Pytorch/TF. If it is done on CPU, wouldn't this opperation greatly increase the total inference time due to the frequent data movement between CPU and GPU when there are continuous layers that need padding? Otherwise, how does GPU finish the inefficient padding operation?

0

There are 0 answers