Struct GpuCapability 

pub struct GpuCapability {
    pub compute_major: i32,
    pub compute_minor: i32,
    pub has_tensor_cores: bool,
    pub has_fp64_tensor_cores: bool,
    pub has_async_copy: bool,
    pub has_cluster_launch: bool,
    pub has_tma: bool,
    pub min_warp_size: i32,
}

Fields

compute_major: i32

compute_minor: i32

has_tensor_cores: bool

has_fp64_tensor_cores: bool

has_async_copy: bool

has_cluster_launch: bool

has_tma: bool

min_warp_size: i32

Implementations

impl GpuCapability

pub const fn from_compute_capability(major: i32, minor: i32) -> Self

pub const fn nvrtc_arch(&self) -> &'static str

NVRTC --gpu-architecture virtual-arch string for this device’s compute capability (e.g. compute_80 for an A100 8.0).

Critical for NVRTC correctness, not just performance: with no --gpu-architecture, NVRTC defaults to a virtual arch below sm_60, where the atomicAdd(double*, double) overload (added in compute capability 6.0) does not exist. A kernel source using double atomics (the SAE arrow/Schur PCG kernels do) then fails to compile, the module load Errs, and the whole device path silently falls back to the CPU. Keying the arch to the real device capability admits those kernels.

Returns a &'static str because cudarc’s CompileOptions::arch is Option<&'static str>. Unknown/future capabilities round DOWN to the nearest known major to stay valid for the installed NVRTC, never up (an arch newer than the toolkit knows would itself fail to compile).

Trait Implementations

impl Clone for GpuCapability

fn clone(&self) -> GpuCapability

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for GpuCapability

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

Formats the value using the given formatter.

impl Eq for GpuCapability

impl PartialEq for GpuCapability

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

Tests for self and other values to be equal, and is used by ==.

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

Tests for !=. The default implementation is almost always sufficient, and should not be overridden without very good reason.

impl StructuralPartialEq for GpuCapability