Crate baracuda_cufile 

Safe Rust wrappers for NVIDIA cuFile (GPUDirect Storage).

cuFile is Linux-only and requires a GDS-capable filesystem (ext4 / XFS on NVMe with the NVIDIA GDS kernel driver). On Windows and macOS every API returns baracuda_core::LoaderError::UnsupportedPlatform.

Workflow

1. Driver::open — initialize the driver (do this once per process).
2. FileHandle::register an open file descriptor.
3. BufRegistration::register a CUDA device buffer (optional but strongly recommended for performance).
4. FileHandle::read / FileHandle::write directly between the file and the device buffer — no bounce through host memory.

Modules

CUfileOpcode

Opcode selector for CUfileIOParams_t.

Structs

BatchIO

RAII handle for a cuFile batch-I/O request group. Supports up to capacity entries per submission cycle.

BufRegistration

A registered device-buffer region. Registration is optional — cuFile works without it — but it unlocks the fastest DMA path.

CUfileIOEvents_t

CUfileIOEvents_t — status for a single batched-I/O entry.

CUfileIOParams_t

CUfileIOParams_t — 1 entry in a batched-IO request.

Driver

RAII handle for the cuFile driver lifecycle.

FileHandle

A registered file descriptor.

StreamRegistration

A cuFile-aware CUDA stream. Register a stream once to use it with FileHandle::read_async / FileHandle::write_async — cuFile will queue the I/O behind prior stream work and signal completion as another stream op.

Functions

op_status_error_string

Human-readable string describing a CUfileOpError code.

probe

Verify cuFile is loadable on this host. Fails on non-Linux platforms.

version

cuFile runtime version as reported by cuFileGetVersion.

Type Aliases

CUfileBatchHandle_t

Opaque batch-handle (maps to CUfileBatchHandle_t).

Error

Error type for cuFile operations.

Result

Result alias.