MemHostRegister #6
Description
I use your cuda5 on my Mac OSX 10.13.4 and cuda 9.1. It is the fastest FFT implementation I have found for GO FFTs. However the standard copy functions are the bottle neck of the Cuda FFT processing.
Could you provide an example of how to use 'MemHostRegister'? I'm trying to speed up the slow process of device to host copy and host to device copy. This is the slowest part of my Cufft GO code. I have a 'C'/cuda9 version that uses unified memory('cudaMallocManaged function') and results in a factor of 10 improvement in the memory copy. I would like to duplicate this in a GO implementation.
I will appreciate any help with this problem