Crate baracuda_driver 

Safe Rust wrappers for the CUDA Driver API.

This crate takes the raw FFI in baracuda_cuda_sys and dresses it up with RAII handles, typed memory, lifetime-checked slices, and a kernel launch builder. It deliberately does not hide the Driver-API model: contexts are explicit, modules are explicit, streams are explicit.

Quickstart

use baracuda_driver::{Context, Device, DeviceBuffer, Module, Stream};

let device = Device::get(0)?;
let ctx = Context::new(&device)?;
let stream = Stream::new(&ctx)?;
let host_data: Vec<f32> = (0..1024).map(|i| i as f32).collect();
let device_data = DeviceBuffer::from_slice(&ctx, &host_data)?;
let mut back = vec![0.0f32; host_data.len()];
device_data.copy_to_host(&mut back)?;
stream.synchronize()?;
assert_eq!(host_data, back);

Modules

• device — Device enumeration and attributes.
• context — Context (explicit CUDA contexts + primary-context reuse).
• stream — Stream, ordered async work queues.
• event — Event, synchronization and timing.
• memory — DeviceBuffer<T>, DeviceSlice<'_, T>, DeviceSliceMut<'_, T>.
• module — Module, Function (PTX/CUBIN loading).
• launch — launch::LaunchBuilder for cuLaunchKernel.
• init() — init() helper and driver version queries.

Re-exports

pub use array::Array;

pub use array::ArrayFormat;

pub use array::SurfaceObject;

pub use array::TextureAddressMode;

pub use array::TextureDesc;

pub use array::TextureFilterMode;

pub use array::TextureObject;

pub use context::Context;

pub use context::PrimaryContext;

pub use device::Device;

pub use error::error_name;

pub use error::error_string;

pub use error::Error;

pub use error::Result;

pub use event::Event;

pub use graph::instantiate_flags;

pub use graph::CaptureMode;

pub use graph::Graph;

pub use graph::GraphExec;

pub use graph::GraphNode;

pub use init::init;

pub use init::version;

pub use launch::Dim3;

pub use launch::LaunchBuilder;

pub use memory::mem_get_info;

pub use memory::DeviceBuffer;

pub use memory::DevicePtr;

pub use memory::DevicePtrMut;

pub use memory::DeviceSlice;

pub use memory::DeviceSliceMut;

pub use memory::ManagedAttach;

pub use memory::ManagedBuffer;

pub use memory::MemAdvise;

pub use pinned::PinnedBuffer;

pub use pinned::PinnedRegistration;

pub use module::Function;

pub use module::Module;

pub use stream::Stream;

Modules

array

CUDA arrays + texture / surface objects.

context

CUDA contexts — both primary (shared with the Runtime API) and explicit.

coredump

CUDA GPU core-dump configuration (CUDA 12.1+).

device

Physical-GPU query and enumeration.

error

Error type for baracuda-driver.

event

CUDA events — lightweight synchronization objects you can record on a stream and later wait on, or use to measure elapsed device time.

external

External memory / semaphore interop — import buffers and sync primitives from Vulkan, D3D11, D3D12, NvSci, and OpaqueFd sources.

graph

CUDA Graphs — record a sequence of operations once, replay cheaply.

graphics

Graphics-API interop — register GL / D3D / VDPAU / EGL resources with CUDA for zero-copy compute on their memory.

green

Green contexts (CUDA 12.4+) — partition a GPU’s SMs into isolated “green” subsets that each run their own stream/kernel pipeline without contending for SMs with other green contexts on the same device.

init

Driver initialization helpers.

ipc

Inter-Process Communication for CUDA events and allocations.

launch

Kernel launch builder — the Rust equivalent of CUDA C’s triple-chevron syntax.

launch_attr

Typed builders for CUlaunchAttribute entries consumed by crate::LaunchBuilder::launch_ex (CUDA 12.0+).

library

Driver-API library + kernel management (CUDA 12.0+).

memcpy2d

Strided (2-D) memory copies via cuMemcpy2D, and pitched device allocations via cuMemAllocPitch.

memcpy3d

3-D arrays + 3-D memcpy + mipmapped arrays.

memory

Device-memory types.

mempool

Stream-ordered memory pools (CUDA 11.2+).

module

Compiled module loading (PTX, CUBIN, fatbin) and kernel entry-point lookup.

multicast

Multicast objects (CUDA 12.0+, NVSwitch systems only).

occupancy

Occupancy calculators — how many blocks can a kernel fit per SM, and what block size maximizes utilization.

pinned

Pinned (page-locked) host memory.

pointer

Pointer attribute queries (cuPointerGetAttribute).

profiler

Thin wrappers over cuProfilerStart / cuProfilerStop. These tell external profilers (Nsight, CUPTI) which sections of a program are interesting; they do not produce output themselves.

stream

CUDA streams — ordered queues of work on a device.

tensor_map

Hopper Tensor Memory Accelerator (TMA) descriptors.

user_object

CUDA Graph user objects (CUDA 12.0+).

vmm

CUDA Virtual Memory Management (VMM) — the fine-grained alternative to cuMemAlloc.