Expand description
The CUDA backend: one engine over t4a-cuda-runtime and the t4a-cuda-kernels
catalogue (issue #2208).
One [CudaRuntime] per engine owns the stream, the allocation domain, transfers and
pending handles; every operation reaches the kernel catalogue through the engine’s typed
dispatchers, and the vendor libraries through [raw::Session].
§Examples
use tenferro_gpu::cuda::{gpu_available, CudaDeviceId, CudaEngine};
if gpu_available() {
let engine = CudaEngine::open(CudaDeviceId::from_ordinal(0))?;
let _ = engine.runtime();
}Re-exports§
pub use crate::gpu;pub use runtime::PinnedStagingStats;pub use runtime::DEFAULT_PINNED_STAGING_LIMIT_BYTES;pub use ::cudarc;
Modules§
- raw
- The owner-scoped raw session an extension crate runs its vendor calls in.
Structs§
- Cuda
Compute Capability - CUDA compute capability
major.minor(e.g. 9.0 for Hopper-class parts). - Cuda
Device Id - Provider-qualified identity of a CUDA device ordinal.
- Cuda
Device Info - Immutable metadata describing one CUDA device.
- Cuda
Device Uuid - Stable physical or MIG device identity as a 16-byte UUID.
- Pinned
Staging Stats - What a runtime’s download staging retains, and the bound it retains it under.
Enums§
- Cuda
Device Error - Structured failures from CUDA device discovery and initialization.
- GpuExtension
Capability - Capability a GPU extension may query before using a provider-specific path.
Constants§
- DEFAULT_
PINNED_ STAGING_ LIMIT_ BYTES - Default bound on the retained download staging: 8 MiB.
Functions§
- cuda_
capabilities - Return the CUDA backend operation capability descriptor table.
- cuda_
devices - Discover CUDA devices visible to the current process.
- cuda_
runtime_ engine_ registration - Build the runtime registration for
engine. - cuda_
runtime_ hardware_ class - Return the canonical CUDA runtime hardware class.