I transfered centernet to xavier.
there is a custom layer name dcnv2 in this model.
I copy plugin in github: TensorRT7-DCNv2-Plugin/DCNv2 at master · lesliejackson/TensorRT7-DCNv2-Plugin · GitHub.
It is ok in my server (2080TI, cuda11.0, TensorRT7.2, cudnn8.0)
But when I test the same code in my xavier, error happened
I have make it clear that the reason is that the initialize function has operation like cudamalloc, new .
but in enqueue function, the pointer head still equal to nullptr.
If I move memory operation to enqueue, the inference time should be double.
So how Can I solve this?
Need your help,