cuda.bindings._v2.driver.MemCreateUsage

class cuda.bindings._v2.driver.MemCreateUsage(value: int)

Usage hints for MemAllocationProp_v1.alloc_flags (CU_MEM_CREATE_USAGE_*).

TILE_POOL: 1

This flag if set indicates that the memory will be used as a tile pool.

HW_DECOMPRESS: 2

This flag, if set, indicates that the memory will be used as a buffer for hardware accelerated decompression.

GPU_DIRECT_RDMA_OVER_PCIE: 4

Setting this flag forces GPUDirect RDMA on a locality-domain-localized allocation to use the PCIe (BAR1) path, allowing the allocation to remain locality-domain localized on platforms where the platform-coherent RDMA path does not support localized allocations. Because on some platforms the PCIe bandwidth is limited, using this flag may result in lower RDMA bandwidth than the default RDMA mapping link. Note that this flag does not itself force PCIe to be used, and that this must be done when creating the RDMA export. CU_DEVICE_ATTRIBUTE_GPU_DIRECT_RDMA_WITH_LOCALIZED_MEMORY_SUPPORTED indicates whether this flag is needed to create GPUDirect RDMA capable localized memory allocations. This flag is only valid if gpuDirectRDMACapable is set.