cajeta.xpu.Device — the active XPU device, and host-side queries about it. The runtime selects one backend at the first device touch (CUDA → HIP → Vulkan → CPU, honoring CAJETA_XPU_BACKEND); Device answers questions about whatever it got, so an app can pick a path at run time. supports is a host query — call it before launching, to choose which kernel or which build to dispatch — not a device-side op.

if (Device.supports(Capability.RayQueryNative)) {
    // hardware inline ray query — take the native fast path
} else {
    // same source, portable software BVH path
}

Geometry follows the active backend

The geometry queries (multiprocessorCount, simdsPerMultiprocessor, waveSize, registersPerMultiprocessor, sharedBytesPerBlock, … and the dispatch law built on them) describe the backend that will run the kernel, which is what activeBackend() names. On the cpu backend that is the host: online cores as multiprocessors (one block per worker thread), one scheduler each, a wave of the host’s SIMD lanes in 32-bit words, physical RAM, integrated. Quantities a host has no counterpart for (register file, shared memory, blocks per multiprocessor, clocks) answer 0 = unknown, never an invented figure. A card that happens to be installed is not consulted, so a cpu-backend Autotune.deviceKey() varies with the CPU and not with the GPU beside it.

Methods

Signature
`static boolean supports(Capability cap)` ⚑	Whether the active device advertises `cap` natively; when false, fall to core (which floors to the portable software path)

⚑ = @EntryPoint

Device

Geometry follows the active backend

Methods

See also