Crate kaio 

KAIO

Rust-native GPU kernel authoring framework.

KAIO (καίω — to kindle, to ignite) lets developers write GPU compute kernels in Rust and lower them to PTX for execution on NVIDIA GPUs. A Rust alternative to OpenAI’s Triton, targeting Windows and Linux from day one, with automatic PTX generation and Rust’s type-safety guarantees.

Crates

• kaio_core — PTX IR types, instruction emitters, PtxWriter
• kaio_runtime — CUDA device management, buffers, PTX loading, kernel launch
• kaio_macros — #[gpu_kernel] proc macro (re-exported here)
• kaio_ops — pre-built GPU operations (matmul, more planned). Separate crate, not re-exported from kaio. Add with cargo add kaio-ops.

Status

Phase 4 complete. Tiled matmul (31% of cuBLAS sgemm), kaio-ops crate, 2D thread blocks, FMA, PTX inspection tools. See the repository README for the full feature table and roadmap.

Re-exports

pub use kaio_core as core;

pub use kaio_runtime as runtime;

Modules

gpu_builtins

Stub functions for GPU built-in operations.

prelude

The KAIO prelude — everything needed to write and launch GPU kernels.

Macros

shared_mem

Declare a shared memory buffer inside a #[gpu_kernel] function.

Attribute Macros

gpu_kernel

Re-export the #[gpu_kernel] attribute macro. Marks a function as a GPU kernel compiled to PTX.