cyanea-gpu

GPU compute abstraction layer with CPU, CUDA, Metal, and WebGPU backends.

What's Inside

• Backend trait -- uniform API for compute operations across backends
• CPU backend -- always-available reference implementation
• CUDA backend -- NVIDIA GPU via cudarc + NVRTC runtime compilation, native f64
• Metal backend -- Apple GPU via metal-rs, MSL compute shaders, f32 with f64 boundary conversion
• WebGPU backend -- cross-platform via wgpu, WGSL shaders, works native and in-browser
• Buffers -- typed compute buffers backed by host memory or device memory
• Operations -- reduce (sum/min/max/mean), element-wise map, pairwise distance matrix, matrix multiply, batch z-score
• Tiled pairwise distances -- larger-than-memory matrices via tiling
• GPU k-mer counting -- 2-bit encoded parallel k-mer counting (k <= 14)
• GPU Smith-Waterman -- batch protein alignment with BLOSUM62, affine gaps
• GPU MinHash -- parallel bottom-k MinHash sketching with Jaccard estimation
• Auto-selection -- auto_backend() picks Metal > CUDA > WebGPU > CPU

Quick Start

[dependencies]
cyanea-gpu = { version = "0.1", features = ["metal"] }  # or "cuda" or "wgpu"

use cyanea_gpu::{auto_backend, ops};

let backend = auto_backend();
println!("Using: {}", backend.device_info().name);

let buf = backend.buffer_from_slice(&[1.0, 2.0, 3.0, 4.0]).unwrap();
let sum = ops::reduce_sum(&*backend, &buf).unwrap();
println!("Sum: {}", sum);

Feature Flags

No default features -- the crate is lightweight with only the CPU backend.

Modules

See Also

• API Reference
• Usage Guide
• Architecture
• Build Guide