How can I change the memory copy pattern with Unified Memory?

Accelerated Computing CUDA CUDA Programming and Performance
1

When I used Unified Memory with managed, and check visual profiler.

A lot of memory copying occurs(HtoD, DtoH both), So I want to control them.

I think there are some functions that automatically copy data according to the situation

in the Unified memory environment then I change those functions.

I found some header only. Is there any way to access the driver api functions?

2

If you are using a pascal device, the cudaMemPrefetchAsync call may be of interest:

[url]c++ - Why is NVIDIA Pascal GPUs slow on running CUDA Kernels when using cudaMallocManaged - Stack Overflow

3

this copying is perfromed by driver rather than user-controlled functions