CODEWITHSUNDEEP
c++memorycuda
YSC
YSC

Reputation: 5

passing a unified memory pointer to kernel slows down the program

In CUDA, does passing a unified memory pointer to a kernel slows down the program?

I got a kernel doing something,and then I cudaMallocManage some unified memory outside the kernel, then pass the pointer to the kernel as parameters. And I used cuda_Event to time. The kernels runs much slower from 1900ms to 3000ms.

Why passing a pointer takes too much ? A pointer?

How should I speed up the program?

Upvotes: 0

Views: 173

Answers (1)

talonmies
talonmies

Reputation: 72349

Passing a managed pointer to a kernel in itself has no effect on performance. But reading and writing to or from the managed memory which the pointer addresses may trigger a lot of ad-hoc memory transfer over the PCI-e bus. And that could be very slow.

Upvotes: 1

Related Questions