What is the meaning of 'Input Reformatter 0'

AI & Data Science Deep Learning (Training & Inference) TensorRT
1

Description

Hello,

I am here to ask about reformatter on TensorRT.

When I ran trtexec, I found this phrase ‘input reformatter 0’ and ‘output reformatter 0’ from log.
What is the meaning of those phrases? What is the meaning(or role) of 0?

I found that
“TensorRT will always insert a reformat layer at the DLA boundary. Input reformats are inserted to synchronize tensors from a CUDA address space to an NvMedia address space.”.

However, what is the meaning of ‘synchronizing tensors from a CUDA address space to a NvMedia address space’?

Thanks!

Environment

TensorRT Version: 7.0.0

2

Hi @yjkim2,
Reformatting layers can be inserted by TensorRT for internal tensors to improve performance.
Please refer to the below link

Thanks!