tenferro_gpu/cubecl/
identity.rs

1//! Provider-neutral GPU extension vocabulary (issue #1597).
2//!
3//! These types describe capabilities and stable device identity without
4//! forcing CUDA-, CubeCL-, or WebGPU-specific concepts into a shared surface.
5//! Backend-specific operations live in their provider namespaces
6//! (`cuda`, `webgpu`); the types here are shared vocabulary only.
7
8use std::fmt;
9
10/// Stable physical or MIG device identity as a 16-byte UUID.
11///
12/// This is the durable comparison identity for diagnostics and topology. It is
13/// distinct from the process-visible ordinal [`CudaDeviceId`](crate::cuda::CudaDeviceId);
14/// the two must never be conflated in equality semantics.
15///
16/// # Examples
17///
18/// ```
19/// use tenferro_gpu::cuda::CudaDeviceUuid;
20///
21/// let uuid = CudaDeviceUuid::from_bytes([7; 16]);
22/// assert_eq!(uuid.to_bytes(), [7; 16]);
23/// assert_eq!(uuid, CudaDeviceUuid::from_bytes([7; 16]));
24/// ```
25#[derive(Copy, Clone, Eq, PartialEq, Ord, PartialOrd, Hash)]
26pub struct CudaDeviceUuid([u8; 16]);
27
28impl CudaDeviceUuid {
29    /// Construct a device UUID from its 16 raw bytes.
30    ///
31    /// # Examples
32    ///
33    /// ```
34    /// use tenferro_gpu::cuda::CudaDeviceUuid;
35    ///
36    /// const UUID: CudaDeviceUuid = CudaDeviceUuid::from_bytes([1; 16]);
37    /// assert_eq!(UUID.as_bytes(), &[1; 16]);
38    /// ```
39    pub const fn from_bytes(bytes: [u8; 16]) -> Self {
40        Self(bytes)
41    }
42
43    /// Borrow the 16 raw UUID bytes.
44    ///
45    /// # Examples
46    ///
47    /// ```
48    /// use tenferro_gpu::cuda::CudaDeviceUuid;
49    ///
50    /// let uuid = CudaDeviceUuid::from_bytes([2; 16]);
51    /// assert_eq!(uuid.as_bytes(), &[2; 16]);
52    /// ```
53    pub const fn as_bytes(&self) -> &[u8; 16] {
54        &self.0
55    }
56
57    /// Return the 16 raw UUID bytes.
58    ///
59    /// # Examples
60    ///
61    /// ```
62    /// use tenferro_gpu::cuda::CudaDeviceUuid;
63    ///
64    /// let uuid = CudaDeviceUuid::from_bytes([3; 16]);
65    /// assert_eq!(uuid.to_bytes(), [3; 16]);
66    /// ```
67    pub const fn to_bytes(self) -> [u8; 16] {
68        self.0
69    }
70}
71
72impl fmt::Debug for CudaDeviceUuid {
73    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
74        f.debug_tuple("CudaDeviceUuid").field(&self.0).finish()
75    }
76}
77
78/// CUDA compute capability `major.minor` (e.g. 9.0 for Hopper-class parts).
79///
80/// # Examples
81///
82/// ```
83/// use tenferro_gpu::cuda::CudaComputeCapability;
84///
85/// let cc = CudaComputeCapability { major: 9, minor: 0 };
86/// assert_eq!(cc.major, 9);
87/// assert_eq!(cc.minor, 0);
88/// ```
89#[derive(Copy, Clone, Eq, PartialEq, Hash, Debug)]
90pub struct CudaComputeCapability {
91    /// Major component of the compute capability.
92    pub major: u32,
93    /// Minor component of the compute capability.
94    pub minor: u32,
95}
96
97/// Capability a GPU extension may query before using a provider-specific path.
98///
99/// [`CudaExecSession::supports`](crate::cuda::CudaExecSession::supports) reports
100/// whether the capability is plausibly available on the current session. This is
101/// orthogonal to the primitive [`OperationCapability`](tenferro_tensor::OperationCapability)
102/// matrix: extension capabilities describe what the extension seam itself can
103/// do, not which tensor operations the backend implements.
104///
105/// # Examples
106///
107/// ```
108/// use tenferro_gpu::cuda::GpuExtensionCapability;
109///
110/// let capability = GpuExtensionCapability::RawStream;
111/// assert!(matches!(capability, GpuExtensionCapability::RawStream));
112/// ```
113#[non_exhaustive]
114#[derive(Copy, Clone, Eq, PartialEq, Hash, Debug)]
115pub enum GpuExtensionCapability {
116    /// Launch external kernels through the CubeCL client.
117    CubeClKernel,
118    /// Load native (PTX/CUBIN/SPIR-V) module artifacts.
119    NativeModule,
120    /// Compile kernel source at runtime (NVRTC / shader compiler).
121    RuntimeCompilation,
122    /// Borrow a raw provider stream for library interop.
123    RawStream,
124    /// Asynchronous same-device copy across runtime/domain boundaries.
125    SameDeviceAsyncCopy,
126    /// Directional peer-to-peer copy between devices.
127    PeerCopy,
128}
tenferro_gpu/cubecl/identity.rs

tenferro_gpu/cubecl/
identity.rs
