Crate ringkernel_cuda 

CUDA Backend for RingKernel

This crate provides NVIDIA CUDA GPU support for RingKernel using cudarc.

Features

• Persistent kernel execution (cooperative groups)
• Lock-free message queues in GPU global memory
• PTX compilation via NVRTC
• Multi-GPU support

Requirements

• NVIDIA GPU with Compute Capability 6.0+ (Pascal and newer) for the core persistent-actor/cooperative-groups path. Some features have higher floors (e.g. Thread Block Clusters/DSMEM/TMA/Green Contexts require Hopper, CC 9.0+) — see the top-level README’s “Feature to minimum compute capability” table for the full breakdown.
• CUDA Toolkit 11.0+
• Native Linux (persistent kernels) or WSL2 (event-driven fallback)

Example

use ringkernel_cuda::CudaRuntime;
use ringkernel_core::runtime::RingKernelRuntime;

#[tokio::main]
async fn main() -> Result<(), Box<dyn std::error::Error>> {
    let runtime = CudaRuntime::new().await?;
    let kernel = runtime.launch("vector_add", Default::default()).await?;
    kernel.activate().await?;
    Ok(())
}

Modules

stub 🔒

Structs

CudaRuntime

Stub runtime when the backend feature is disabled.

Constants

RING_KERNEL_PTX_TEMPLATEDeprecated

PTX kernel source template for persistent ring kernel — deprecated fixed-sm_75 fallback, kept only for API compatibility with existing callers that don’t have a device handle available.

Functions

compile_ptx

Stub compile_ptx when CUDA is not available.

cuda_device_count

Get CUDA device count.

is_cuda_available

Check if CUDA is available at runtime.

ring_kernel_ptx_template_for

PTX kernel source template for persistent ring kernel, generic over the target compute capability (major, minor).