Memória fixada com cudaMallocHost

Pinned memory is host memory the operating system promises never to move or swap out. Because its physical address is fixed, the GPU can read it directly.

You will see the term page-locked used as a synonym for pinned. Both mean the same thing: pages locked in place so no swapping can ever happen.

Ask CUDA for pinned memory with cudaMallocHost. It hands back an ordinary host pointer you can read and write just like any C++ array.

The returned pointer behaves like any host buffer, so your CPU code stays unchanged. You just fill it and read it exactly as before.

Since the pages cannot move, the driver skips its hidden staging copy. The GPU pulls straight from your buffer, so the doubled copy disappears.

The payoff is real bandwidth. Pinned transfers often run near the link's peak speed, sometimes roughly twice as fast as the pageable version.

Pinned memory needs its own matching release. Use cudaFreeHost, never the plain free, or you risk a crash and a leak.

Pinned pages cannot be swapped, so they hold real RAM hostage. Pin too much and you starve the rest of the system of usable memory.

Locking pages takes work, so cudaMallocHost is slower to allocate than malloc. Reuse one pinned buffer across many transfers instead of reallocating.

Beyond speed, pinned memory is the key that unlocks truly asynchronous copies. Without it, cudaMemcpyAsync cannot overlap with anything.

Pin buffers that you transfer often or that feed streams. For a one-off copy of a tiny array, plain malloc is perfectly fine. 🙂

You allocated a buffer with cudaMallocHost. How should you release it?

cudaMallocHost pins host pages so the GPU reads them directly, giving faster copies and enabling async overlap. Always release with cudaFreeHost. 🔒