cuTile Rust 0.5.0

[elibol]

cuTile Rust and its shared host-side crates are now available at v0.5.0.

You can now fuse inter-GPU communication with computation in tile kernels using unsafe unicast/multicast APIs. For example, read a tile from another GPU, combine it with local data, and keep computing in the same kernel. Conceptually:

peer_tile = read_from_peer_gpu(...)
combined = local_tile + peer_tile
result = compute(combined)

You can now call math library functions directly from tile kernels through CUDA 13.5's tile-library preview. With a matching solver::cholesky binding declared and tile library configured on the host:

// Inside a tile kernel; a is a Tile<f64, { [1, 32, 32] }>.
let (factor, info) = unsafe {
    solver::cholesky::<1, 32, f64>("potrf", "lower", a)
};

You can also write your own Rust SIMT functions in CUDA Oxide and package them as tile libraries to call directly from cuTile Rust kernels.

The shared host runtime lets Tile and per-thread (SIMT) kernels use the same GPU allocation without copying between them. For example, pass a SIMT buffer to a Tile stage, then hand the same storage back:

use cutile::prelude::Tensor;

// buffer: cuda_core::DeviceBuffer<f32> with 32 elements.
let tensor = Tensor::from_buffer(buffer, vec![32]);
// ... run a Tile kernel using tensor, then wait for completion ...
let buffer = tensor.into_buffer(); // Same allocation for SIMT code.

The new communication and tile-library APIs require compatible CUDA 13.5 tooling. Communication needs supported NVLink hardware (plus NVLink SHARP-capable NVSwitch for multicast), explicit mappings, and synchronization.

Release and migration notes · Communication guide · Tile-library guide · Crate