rlx-cuda 0.2.6

NVIDIA CUDA backend — cuBLAS for matmul + NVRTC-compiled kernels for everything else, via the pure-Rust `cudarc` crate.
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49

// RLX — versatile ML compiler + runtime.
// Copyright (C) 2026 Eugene Hauptmann, Nataliya Kosmyna.
//
// This program is free software: you can redistribute it and/or modify
// it under the terms of the GNU General Public License as published by
// the Free Software Foundation, version 3.

use std::sync::Arc;

use cudarc::driver::{CudaContext, CudaSlice, CudaStream, LaunchConfig, PushKernelArg};

use crate::kernels::{dispatch_grid_1d, welch_peaks_gpu_kernel};

/// Native GPU WelchPeaks on the device arena (f32 block-layout spectra).
pub fn run_welch_peaks_gpu(
    ctx: &Arc<CudaContext>,
    stream: &Arc<CudaStream>,
    buffer: &mut CudaSlice<f32>,
    spec_off: u32,
    dst_off: u32,
    welch_batch: u32,
    n_fft: u32,
    n_segments: u32,
    k: u32,
    n_bins: u32,
) {
    let kernel = welch_peaks_gpu_kernel(ctx);
    let (grid, block) = dispatch_grid_1d(welch_batch, 64);
    let cfg = LaunchConfig {
        grid_dim: (grid, 1, 1),
        block_dim: (block, 1, 1),
        shared_mem_bytes: 0,
    };
    let mut launcher = stream.launch_builder(&kernel.function);
    launcher
        .arg(&mut *buffer)
        .arg(&spec_off)
        .arg(&dst_off)
        .arg(&welch_batch)
        .arg(&n_fft)
        .arg(&n_segments)
        .arg(&k)
        .arg(&n_bins);
    unsafe {
        launcher
            .launch(cfg)
            .expect("rlx-cuda: welch_peaks_gpu launch failed");
    }
}