dspk-ops 0.0.3

Signal-processing operations for dspk, written once in CubeCL: filters, resampling, linear algebra, statistics, peak finding, spatial operators
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

use cubecl::prelude::*;
use dspk_core::compute::element_position;

/// `output[r, t] = input[offset + r · row_stride + t] − mean[r]` for `t < cols`: rows of a strided
/// buffer, centred, into a contiguous `[rows, cols]` buffer. The offset is applied in the kernel, so
/// the input binds at the buffer's start. One unit per output element ([`fn@element_position`]).
#[cube(launch)]
#[allow(clippy::too_many_arguments)]
pub fn center_rows_kernel<F: Float>(
    input: &[F],
    mean: &[F],
    output: &mut [F],
    offset: u32,
    row_stride: u32,
    cols: u32,
    total: u32,
) {
    let e = element_position() as u32;
    if e < total {
        let r = e / cols;
        let t = e - r * cols;
        output[e as usize] = input[(offset + r * row_stride + t) as usize] - mean[r as usize];
    }
}