CUDA：零拷贝主机内存

The easy way to achieve copy/compute overlap!
1.Enable Host Mapping*

Runtime: cudaSetDeviceFlags() with cudaDeviceMapHost flag
Driver : cuCtxCreate() with CU_CTX_MAP_HOST

2.Allocate pinned CPU memory

Runtime: cudaHostAlloc(), use cudaHostAllocMapped flag
Driver : cuMemHostAlloc()use CUDA_MEMHOSTALLOC_DEVICEMAP

3.Get a CUDA device pointer to this memory

Runtime: cudaHostGetDevicePointer()
Driver : cuMemHostGetDevicePointer()

4.Just use that pointer in your kernels!

Zero-Copy Guidlines

•Data is transferred over the PCIe bus automatically, but it’s slow

•Use when data is only read/written once

•Use for very small amounts of data (new variables, CPU/GPU communication)

•Use when compute/memory ratio is very high and occupancy is high, so latency over PCIe is hidden
•Coalescing is critically important!