Crate atomr_accel 

atomr-accel

Actor-shaped face for compute-acceleration backends, on top of the atomr actor runtime. NVIDIA CUDA is the first shipping implementation (atomr-accel-cuda); the same trait surface accommodates AMD ROCm, Apple Metal, Intel oneAPI, and Vulkan compute when those land.

[dependencies]
atomr-accel      = "0.1"
atomr-accel-cuda = "0.1"   # active backend

use atomr_accel::prelude::*;
use atomr_accel_cuda as cuda;

What this crate is

A thin core that names the abstractions every backend has to satisfy:

• AccelBackend — marker trait identifying a backend, with associated Device, Stream, Event, Error types.
• AccelDtype / DType — backend-agnostic numeric data-type trait + discriminant. Backends layer their own *Dtype trait on top with FFI mappings.
• AccelRef — generation-validated typed device pointer parametric over the backend.
• AccelError — typed error enum, #[non_exhaustive] so backends can add LibraryError variants without breaking core.
• CompletionStrategy — async wakeup contract for kernel completion (host-fn callback, sync, polled).
• KernelOp — marker trait for typed op envelopes (Sgemm, RngFillUniform, etc.).

The core deliberately ships no concrete actors. Each backend crate (atomr-accel-cuda, future atomr-accel-rocm, atomr-accel-metal, …) provides its own DeviceActor, KernelActor family, and library wrappers, and depends on this crate for the trait surface.

What this crate is not

• A least-common-denominator API. Backends expose more than the trait surface — atomr_accel_cuda::kernel::CudnnActor has a richer message set than KernelOp knows about, and that’s fine. The trait surface is for portable code; backend-specific work uses the concrete crate directly.
• A device-abstraction layer like wgpu or SYCL. We don’t try to compile one shader to many targets. We supervise the right library on the right hardware.

Re-exports

pub use backend::AccelBackend;

pub use backend::AccelDevice;

pub use backend::AccelStream;

pub use completion::CompletionStrategy;

pub use dtype::AccelDtype;

pub use dtype::DType;

pub use error::AccelError;

pub use gpu_ref::AccelRef;

pub use kernel::KernelOp;

Modules

backend

Backend identity traits.

completion

CompletionStrategy — async wakeup contract for kernel completion. Promoted to the core so every backend can plug into the same family (host-fn callback, sync block, polled query).

dtype

AccelDtype — backend-agnostic numeric data-type trait.

error

Backend-agnostic error taxonomy.

gpu_ref

AccelRef<T, B> — backend-agnostic typed device pointer.

kernel

KernelOp — marker trait for typed kernel-op envelopes.

prelude

Canonical re-exports. use atomr_accel::prelude::*;.