Struct Gpu 

pub struct Gpu { /* private fields */ }

A GPU that can run the sweep.

Holds a device and queue. Creating one enumerates adapters, which is slow enough that it should happen once per process rather than once per sweep.

Implementations

impl Gpu

pub fn new() -> Option<Gpu>

Open the default adapter, or None if this machine exposes none.

None means not found on this machine, never “impossible”. A headless CI runner with no driver is the common case, which is why every test here skips rather than fails on it.

pub fn adapter(&self) -> &AdapterInfo

What the driver says this is.

Read it before quoting a speedup. DeviceType::Cpu is a software rasteriser — lavapipe, SwiftShader, WARP — which runs the shader correctly and tells you nothing about a GPU, and a benchmark that does not check this reports the wrong machine with full confidence.

pub fn is_hardware(&self) -> bool

True when the adapter is real silicon rather than a software rasteriser.

pub fn sweep( &self, m: &GpuModel, spins: &mut [i8], beta: f64, sweeps: u32, ) -> Result<(), String>

Run sweeps chromatic sweeps over spins, in place.

One dispatch per colour class per sweep, which is what makes the update correct: nodes in a class share no edge, so they can be resampled simultaneously without any of them reading a neighbour another lane is writing. Dispatching all nodes at once would be faster and wrong.