Pré-busca com cudaMemPrefetchAsync

Instead of waiting for slow first-touch faults, you can move pages early. cudaMemPrefetchAsync sends managed data to a device before the kernel needs it.

You name the pointer, the byte count, and the destination device. This one prefetch migrates the whole range up front in a single efficient move.

The third argument is the device id. Pass a GPU number to stage data on that GPU, ready for the kernel you are about to launch.

Use the special id cudaCpuDeviceId to pull results back to the host. Do it before CPU code reads them to avoid a wave of faults.

The Async in the name is real: the call returns immediately and runs in a stream. Your CPU keeps working while pages migrate in the background.

Because it rides a stream, a prefetch can overlap with other kernels. Stage the next chunk while the current one is still being processed.

A single bulk prefetch is far cheaper than thousands of tiny faults. You trade scattered overhead for one contiguous high-bandwidth transfer.

Only stage what the kernel actually touches. Prefetching a huge buffer the kernel barely reads just wastes bandwidth and GPU memory.

A clean rhythm emerges: prefetch to the GPU, launch the kernel, prefetch results back. This keeps migration off the critical path on both ends.

Add a prefetch, then check Nsight for fewer faults and tighter timelines. Let profiling confirm the win rather than trusting intuition.

Prefetching keeps the single-pointer ease of managed memory while giving you back control over timing. You get the best of both styles.

Let us confirm what prefetching buys you.

You learned to stage pages early with cudaMemPrefetchAsync, picking a GPU or cudaCpuDeviceId. It overlaps in streams and beats faulting. Great job! ✨