Runtime Dispatch

Every call runs on the CPU and dispatches to the best kernel of its enabled capabilities, those it executes and this binary contains:

cpu := sz.CPU()
enabled, _ := cpu.CapabilitiesEnabled() // what dispatch uses: detected on this CPU and compiled in
fmt.Println(enabled)                    // like "serial,neon,neonaes,neonsha"
fmt.Println(enabled.Has(sz.CapNEON))    // test one capability

func CPU() Device
func CountDevices(kind DeviceKind) (int, error)
func NewDevice(kind DeviceKind, ordinal int) (Device, error)
func (d Device) CapabilitiesDetected() (Capability, error)
func (d Device) CapabilitiesCompiled() Capability
func (d Device) CapabilitiesEnabled() (Capability, error)
func (d Device) ConfigureThread(capabilities Capability) (func(), error)
func (c Capability) Has(capability Capability) bool
func (c Capability) String() string

A Device is the host CPU or one GPU of a runtime, DeviceCUDA, DeviceROCm or DeviceMetal, named by that runtime’s own ordinal. CountDevices counts them and fails without a GPU of that kind, and NewDevice fails past the last one. CapabilitiesDetected and CapabilitiesCompiled report the two raw axes, what the device executes and what this binary contains for its kind. CapabilitiesEnabled is both at once, and on the CPU always keeps CapSerial. ConfigureThread pins the goroutine to an OS thread and prepares it for the kernels of its argument, returning the function that unpins it; it fails on a GPU, which has no thread state to configure. Every capability is a typed Capability constant, like CapSerial, CapHaswell, CapNEON, CapSVE2 or CapCUDA, and CapCPUs, CapGPUs and CapAny group them. A call that reports a failure status returns it as an error where the function has one, and panics otherwise.