Enum anira_memory_kind#

enum anira_memory_kind#

The physical memory a provider’s device computes in, against the host’s (section 7): the fact a device provider’s edge over host memory takes its class from. Asked of the device’s own API at run time, never inferred from a platform’s name: Metal’s hasUnifiedMemory, the Vulkan device type cross-checked with its memory heaps, the CUDA driver’s CU_DEVICE_ATTRIBUTE_INTEGRATED, HIP’s hipDeviceAttributeIntegrated, OpenCL’s CL_DEVICE_HOST_UNIFIED_MEMORY, D3D12’s ARCHITECTURE.UMA; a browser exposes none.

Values:

enumerator ANIRA_MEMORY_UNKNOWN#

No probe established it: the device’s API is missing or failed, the provider maps to no API anira can ask (an NPU, a custom provider, the web), or the adapters a self-choosing runtime picks among disagree. The edge is ANIRA_EDGE_UNKNOWN, which every limit of the edge cost refuses.

enumerator ANIRA_MEMORY_UNIFIED#

The host’s own physical memory: the CPU path, an integrated GPU, a CPU device. A crossing between the CPU’s domain and the device’s is a unified transfer: a copy plus a hand-over between the processors, no bus.

enumerator ANIRA_MEMORY_DEDICATED#

Memory of the device’s own across a bus (a discrete GPU’s VRAM): a crossing to or from it is a dedicated transfer over the bus.

enumerator ANIRA_MEMORY_FORCE32#