Enum ComputeBackend 

pub enum ComputeBackend {
    Serial,
    CpuMulti,
    Gpu,
}

Which execution backend a numerical kernel should use.

Variants and what they cost

• Serial — one thread, scalar. Always available. The deterministic trusted reference; bit-for-bit reproducible run to run.
• CpuMulti — rayon across CPU cores. Requires the parallel feature. Pays thread-pool overhead, so it loses on small problems; see CPU_MULTI_MIN_WORK_ITEMS. Reduction order is not generally reproducible unless a kernel says otherwise.
• Gpu — wgpu compute. Requires the gpu feature, a non-Android target, and an actual adapter at run time. Pays a host-device transfer per dispatch, so it needs a much larger problem to win; see GPU_MIN_WORK_ITEMS. WGSL has no f64, so a GPU kernel that computes in f32 will NOT match the serial f64 path to full precision — each kernel must document its own measured deviation.

Choosing one

Prefer select_backend over hand-picking. If you do name a backend, route it through resolve so an unavailable choice degrades instead of failing.

Units

Dimensionless — this is a mode selector, not a quantity.

Variants

Serial

Single-threaded scalar execution — always present, always the oracle.

CpuMulti

Multi-threaded CPU execution via rayon (parallel feature).

Gpu

GPU compute via wgpu (gpu feature, non-Android, adapter present).

Implementations

impl ComputeBackend

pub fn is_available(self) -> bool

Whether this backend can actually run in this build, on this target, right now.

• Serial — always true.
• CpuMulti — true when the parallel feature is on. Compile-time only; rayon needs no run-time probe.
• Gpu — true only when the gpu feature is on, the target is not Android, and a GPU adapter was found by gpu_adapter_present. This one genuinely queries the machine.

pub fn resolve(self) -> Self

Degrade this backend to the best one that is actually available, without erroring.

The order is Gpu -> CpuMulti -> Serial: a requested GPU with no adapter (or no gpu feature) tries multi-CPU, and failing that runs serially. Serial is always reachable, so this never fails and never panics.

This is the “graceful absence of a GPU” contract: a missing adapter is a normal outcome on a headless server, in CI, on Android, or on a laptop with no discrete GPU — not an error condition.

pub fn label(self) -> &'static str

A short human-readable label, for benchmark tables and log lines.

Trait Implementations

impl Clone for ComputeBackend

fn clone(&self) -> Self

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Copy for ComputeBackend

impl Debug for ComputeBackend

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for ComputeBackend

fn default() -> Self

Returns the “default value” for a type.

impl Eq for ComputeBackend

impl Hash for ComputeBackend

fn hash<__H: Hasher>(&self, state: &mut __H)

Feeds this value into the given Hasher.

fn hash_slice<H>(data: &[Self], state: &mut H)

where H: Hasher, Self: Sized,

Feeds a slice of this type into the given Hasher.

impl PartialEq for ComputeBackend

fn eq(&self, other: &Self) -> bool

Equality operator ==.

fn ne(&self, other: &Rhs) -> bool

Inequality operator !=.

impl StructuralPartialEq for ComputeBackend