Module registry 

Global kernel registry — load PTX once, cache modules/functions/streams.

Eliminates ~0.06-0.26ms overhead per kernel call from:

• CudaContext::new (driver init)
• CudaContext::load_module (PTX JIT compilation)
• CudaContext::new_stream (stream creation)
• cuModuleGetFunction (function lookup) — now cached after first call

Usage: let reg = AtlasRegistry::get_or_init(ordinal, &[(“gemm”, PTX_SRC), …])?; let func = reg.function(“gemm”, “dense_gemm_tc_bf16”)?; unsafe { reg.stream.launch_builder(&func).arg(&ptr).launch(cfg)?; } reg.stream.synchronize()?;

Re-exports

pub use crate::cuda_host::CudaHost;

pub use crate::cuda_host::host;

pub use crate::cuda_host::release;

Structs

AtlasRegistry

The PTX/CUBIN modules for one loaded model.

RawCudaFunc

Wrapper for raw CUfunction handle (Send+Sync safe — handles are context-wide).

Constants

CUDA_ERROR_DEINITIALIZED

CUDA_ERROR_DEINITIALIZED. The driver tears the primary context down in its own atexit handler, which can run before our Drop impls do. Every module unload and every host free then reports this code.

Functions

cuda_error_text

Resolve a CUresult status code into "<NAME>: <description>" via cuGetErrorName + cuGetErrorString. Returns “CUDA_UNKNOWN” / “(no message)” if the driver doesn’t recognize the code.

is_teardown_noop

Whether a CUresult means “the context is already gone, nothing to do”.