cudaMallocHost

cudaMallocHost is a CUDA runtime API function that allocates page-locked (also known as pinned) host memory. The allocation enables faster host-to-device and device-to-host data transfers compared with pageable memory, and it can be used with asynchronous copies when combined with streams. The memory remains resident in physical RAM and is not swapped out by the operating system, which reduces transfer latency and can improve bandwidth.

The typical usage is cudaError_t cudaMallocHost(void** ptr, size_t size). On success, *ptr points to a host memory region of at least size bytes; on failure, a CUDA error code such as cudaErrorMemoryAllocation is returned. The allocated memory should be freed with cudaFreeHost(ptr) when it is no longer needed. Errors should be checked and handled in accordance with the CUDA runtime API.

Pinned memory allocated by cudaMallocHost is suitable for use as the source or destination in host-device transfers performed with cudaMemcpy or cudaMemcpyAsync. Transfers can be overlapped with computation when using streams, which can improve overall throughput. Note that kernels running on the device typically cannot directly dereference host memory allocated by cudaMallocHost unless the memory is specifically allocated with mapping capabilities (for example, via cudaHostAlloc with appropriate flags). Therefore, cudaMallocHost memory is primarily used for efficient data transfer rather than direct device access.

Common considerations include the higher cost of pinning memory, limited availability of pinned memory, and potential fragmentation from frequent allocations and deallocations. Developers should reuse allocated pinned memory when possible and release it promptly with cudaFreeHost. Related functions include cudaHostAlloc and cudaFreeHost.

---