struct CapabilitiesConforms to
OptionSet, Sendable, CustomStringConvertible, SetAlgebra, RawRepresentable, Equatable, ExpressibleByArrayLiteralCapability detection is exposed directly for diagnostics and tests:
import NumKong
// `capabilitiesEnabled` is what dispatch runs within: detected on this CPU AND compiled into the binary.
let enabled = try Device.cpu.capabilitiesEnabled
print(enabled) // like "serial,neon,neonhalf,neonfhm,neonsdot"
print(enabled.contains(.neon)) // `Capabilities` is an `OptionSet`
// The two raw axes, when you specifically mean one of them:
let onThisCpu = try Device.cpu.capabilitiesDetected
let inThisBinary = Device.cpu.capabilitiesCompiled
// Set up each thread that runs kernels:
try Device.cpu.configureThread(enabled)
// GPUs are devices too, by their runtime's own ordinal:
for ordinal in 0..<((try? Device.count(.metal)) ?? 0) {
print(try Device(kind: .metal, ordinal: ordinal).capabilitiesEnabled) // like "metal,apple9"
}Device.cpuis the host CPU, andDevice(kind:ordinal:)any deviceDevice.count(_:)counts, throwing anErrorpast the last one.capabilitiesDetectedis what the device can execute, from CPUID or HWCAP on the CPU.capabilitiesCompiledis what this binary contains for devices of its kind, from the ISA probes at build time.capabilitiesEnabledis what dispatch runs within, both axes at once, which the library settles for the CPU as it loads, and on the CPU always contains.serial. Every CPU kernel call passesCapabilities.cpus, which the library clamps to it, so the module keeps no mask of its own.configureThread(_:)sets per-thread state such as Intel AMX tiles for the given CPU capabilities, usually the enabled ones.
Reach for capabilitiesEnabled unless you specifically mean one of the raw axes. capabilitiesDetected describes the machine and says nothing about whether a kernel was compiled in, so a build whose ISA probes failed still reports your CPU’s full feature set while containing no SIMD kernels at all.
You usually do not need to branch on this in application code. The native layer still selects the best enabled kernel automatically.
API
StructCapabilitiesA set of capabilities of a CPU or a GPU, as a Device reports them.55
rawValue
let rawValue: UInt64init
init(rawValue: UInt64)serial
static let serial: Capabilitiesneon
static let neon: CapabilitiesneonHalf
static let neonHalf: CapabilitiesneonBFDot
static let neonBFDot: CapabilitiesneonFHM
static let neonFHM: CapabilitiesneonSDot
static let neonSDot: CapabilitiesneonFP8
static let neonFP8: Capabilitiessve
static let sve: CapabilitiessveHalf
static let sveHalf: CapabilitiessveSDot
static let sveSDot: CapabilitiessveBFDot
static let sveBFDot: Capabilitiessve2
static let sve2: Capabilitiessme
static let sme: CapabilitiessmeF64
static let smeF64: CapabilitiessmeBI32
static let smeBI32: Capabilitieshaswell
static let haswell: Capabilitiesalder
static let alder: Capabilitiessierra
static let sierra: Capabilitiesskylake
static let skylake: CapabilitiesiceLake
static let iceLake: Capabilitiesgenoa
static let genoa: Capabilitiesturin
static let turin: Capabilitiessapphire
static let sapphire: Capabilitiesdiamond
static let diamond: CapabilitiessapphireAMX
static let sapphireAMX: CapabilitiesgraniteAMX
static let graniteAMX: CapabilitiesdiamondAMX
static let diamondAMX: Capabilitiesrvv
static let rvv: CapabilitiesrvvBF16
static let rvvBF16: CapabilitiesrvvHalf
static let rvvHalf: CapabilitiesrvvBB
static let rvvBB: Capabilitiesv128
static let v128: Capabilitiesv128Relaxed
static let v128Relaxed: CapabilitiespowerVSX
static let powerVSX: CapabilitiesloongsonASX
static let loongsonASX: Capabilitiescuda
static let cuda: Capabilitiesampere
static let ampere: Capabilitiesada
static let ada: Capabilitieshopper
static let hopper: Capabilitiesblackwell
static let blackwell: CapabilitiesblackwellRTX
static let blackwellRTX: CapabilitiesblackwellUltra
static let blackwellUltra: Capabilitiesrocm
static let rocm: Capabilitiescdna3
static let cdna3: Capabilitiescdna4
static let cdna4: Capabilitiescdna5
static let cdna5: Capabilitiesmetal
static let metal: Capabilitiesapple9
static let apple9: Capabilitiesapple10
static let apple10: CapabilitiescpusEvery CPU capability, the bits below the first GPU vendor's, which every CPU kernel call passes, as the library clamps it to capabilitiesEnabled.
static let cpus: CapabilitiesgpusEvery GPU capability.
static let gpus: CapabilitiesanyEvery capability.
static let any: CapabilitiesdescriptionThe capability names, comma-separated, like "serial,haswell".
var description: String { get }EnumDeviceKindWhich runtime a device belongs to, as the nk_<kind>_* C functions name it.
enum DeviceKindValues
cpu | ||
|---|---|---|
cuda | ||
rocm | ||
metal |
Conforms to
Equatable, Hashable, SendableStructDeviceOne device NumKong can run kernels on: the host CPU, or a GPU by its runtime's own ordinal, the one cudaSetDevice or hipSetDevice takes, or the position in Metal's device list.9
struct DevicePrefer capabilitiesEnabled unless you specifically mean one of the raw axes:
capabilitiesDetected describes the device and says nothing about whether a kernel was
compiled into this binary, so selecting on it alone claims support for code that may not exist.
Conforms to
Sendable, Equatablekind
let kind: DeviceKindordinal
let ordinal: IntcpuThe host CPU, which every build has.
static let cpu: DeviceinitDevice ordinal of kind.
countHow many devices of kind the process sees: one CPU, or the GPUs its runtime counts.
capabilitiesDetectedWhat this device runs, whether or not this binary holds kernels for it.
var capabilitiesDetected: Capabilities { get throws }capabilitiesCompiledWhat this binary holds kernels for on devices of this kind, whether or not this one runs.
var capabilitiesCompiled: Capabilities { get }capabilitiesEnabledWhat this device's kernel calls run within: capabilitiesDetected and capabilitiesCompiled at once.
var capabilitiesEnabled: Capabilities { get throws }On the CPU the library settles it as it loads, and it always
contains serial.
configureThreadConfigures the current thread for capabilities, usually capabilitiesEnabled, e.g.
func configureThread(_ capabilities: Capabilities) throwsAMX tile state on x86. Must be called once per thread before using AMX operations.
Throws
Error on a GPU, which has no thread state to configure.