dspk-ops 0.0.8

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
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
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227

//! Per-row reductions of a `[rows, cols]` buffer, one cube per row
//! ([`LaunchGeometry::per_row`]).
//!
//! Each unit folds a strided share of the row (`col = unit, unit + units, …`, so neighbouring units
//! read neighbouring values), then the cube merges the partial results pairwise in shared memory.
//! Pairwise merging keeps the error of long rows close to that of short sums.

use cubecl::prelude::*;
use cubecl::server::Handle;
use dspk_core::compute::LaunchGeometry;

use super::kernels::{row_kth_radix_kernel, row_mean_std_kernel, KEY_ABS, KEY_ABS_DEV, KEY_SIGNED};
use super::DspFloat;

/// Per-row mean and population standard deviation of a `[rows, cols]` buffer of `F` into `out_mean`
/// and `out_std` (`rows` values each).
pub fn row_mean_std<F: DspFloat>(
    client: &Client,
    input: &Handle,
    out_mean: &Handle,
    out_std: &Handle,
    rows: usize,
    cols: usize,
) {
    let geom = LaunchGeometry::per_row(client, rows, cols);
    unsafe {
        row_mean_std_kernel::launch::<F>(
            client,
            geom.cube_count,
            geom.cube_dim.clone(),
            BufferArg::from_raw_parts(input.clone(), rows * cols),
            BufferArg::from_raw_parts(out_mean.clone(), rows),
            BufferArg::from_raw_parts(out_std.clone(), rows),
            rows as u32,
            cols as u32,
            geom.cube_dim.x,
        );
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::core::buffer;

    fn reductions(client: &Client) {
        // Rows longer and shorter than a cube, one empty-ish row of a single value
        for cols in [1usize, 5, 1_000, 70_001] {
            let rows = 3;
            let data: Vec<f32> = (0..rows * cols).map(|i| ((i * 7919) % 1013) as f32 * 0.1 - 40.0 + (i / cols) as f32 * 100.0).collect();
            let input = buffer::upload(client, &data);
            let (mean, std) = (buffer::empty::<f32>(client, rows), buffer::empty::<f32>(client, rows));
            row_mean_std::<f32>(client, &input, &mean, &std, rows, cols);
            let (mean, std) = (buffer::download::<f32>(client, mean), buffer::download::<f32>(client, std));
            for r in 0..rows {
                let row: Vec<f64> = data[r * cols..(r + 1) * cols].iter().map(|v| *v as f64).collect();
                let m = row.iter().sum::<f64>() / cols as f64;
                let s = (row.iter().map(|v| (v - m).powi(2)).sum::<f64>() / cols as f64).sqrt();
                assert!((mean[r] as f64 - m).abs() < 1e-3 * m.abs().max(1.0), "{} cols {cols} row {r}: mean {} vs {m}", client.name(), mean[r]);
                assert!((std[r] as f64 - s).abs() < 1e-3 * s.max(1.0), "{} cols {cols} row {r}: std {} vs {s}", client.name(), std[r]);
            }
        }
    }
    runtime_test!(test_row_reductions_match_host, reductions);

    fn abs_kth(client: &Client) {
        // Odd / even lengths, repeated values, a column sub-range of a longer row
        let (rows, stride) = (3usize, 2_003usize);
        let data: Vec<f32> = (0..rows * stride).map(|i| ((i * 7919) % 211) as f32 * 0.25 - 26.0).collect();
        let input = buffer::upload(client, &data);
        for cols in [0..1usize, 0..2, 5..1_006, 0..stride] {
            for k in [0, cols.len() / 2, cols.len() - 1] {
                let out = buffer::empty::<f32>(client, rows);
                row_abs_kth::<f32>(client, &input, &out, rows, stride, cols.clone(), k);
                let got = buffer::download::<f32>(client, out);
                for r in 0..rows {
                    let mut abs: Vec<f32> = data[r * stride + cols.start..r * stride + cols.end].iter().map(|v| v.abs()).collect();
                    abs.select_nth_unstable_by(k, |a, b| a.total_cmp(b));
                    assert_eq!(got[r], abs[k], "{} cols {cols:?} k {k} row {r}", client.name());
                }
            }
        }
    }
    runtime_test!(test_row_abs_kth_matches_host, abs_kth);

    /// The `f64` path (64-bit keys) is exact too, on runtimes with `f64`.
    fn abs_kth_f64(client: &Client) {
        if !client.properties().supports_type(f64::elem_type_native()) {
            return;
        }
        let (rows, cols) = (2usize, 1_001usize);
        let data: Vec<f64> = (0..rows * cols).map(|i| ((i * 7919) % 211) as f64 * 0.125 - 13.0 + 1e-9 * i as f64).collect();
        let input = buffer::upload(client, &data);
        for k in [0, cols / 2, cols - 1] {
            let out = buffer::empty::<f64>(client, rows);
            row_abs_kth::<f64>(client, &input, &out, rows, cols, 0..cols, k);
            let got = buffer::download::<f64>(client, out);
            for r in 0..rows {
                let mut abs: Vec<f64> = data[r * cols..(r + 1) * cols].iter().map(|v| v.abs()).collect();
                abs.select_nth_unstable_by(k, |a, b| a.total_cmp(b));
                assert_eq!(got[r], abs[k], "{} f64 k {k} row {r}", client.name());
            }
        }
    }
    runtime_test!(test_row_abs_kth_f64, abs_kth_f64);

    /// Signed and deviation selection, exact against a host sort, for `F` (negatives, repeats,
    /// odd / even lengths, a column sub-range).
    fn signed_and_dev_kth_of<F: DspFloat + std::fmt::Debug>(client: &Client) {
        let (rows, stride) = (3usize, 1_003usize);
        let data: Vec<F> = (0..rows * stride).map(|i| crate::core::cast::<F>(((i * 7919) % 211) as f64 * 0.25 - 26.0)).collect();
        let centers: Vec<F> = (0..rows).map(|r| crate::core::cast::<F>(r as f64 * 3.5 - 2.0)).collect();
        let (input, center) = (buffer::upload(client, &data), buffer::upload(client, &centers));
        let sorted_k = |mut v: Vec<F>, k: usize| {
            v.sort_by(|a, b| a.partial_cmp(b).expect("finite"));
            v[k]
        };
        for cols in [0..1usize, 0..2, 5..1_000, 0..stride] {
            for k in [0, cols.len() / 2, cols.len() - 1] {
                let (signed, dev) = (buffer::empty::<F>(client, rows), buffer::empty::<F>(client, rows));
                row_kth::<F>(client, &input, &signed, rows, stride, cols.clone(), k);
                row_abs_dev_kth::<F>(client, &input, &center, &dev, rows, stride, cols.clone(), k);
                let (signed, dev) = (buffer::download::<F>(client, signed), buffer::download::<F>(client, dev));
                for r in 0..rows {
                    let row = &data[r * stride + cols.start..r * stride + cols.end];
                    assert_eq!(signed[r], sorted_k(row.to_vec(), k), "{} signed cols {cols:?} k {k} row {r}", client.name());
                    // Host |d| (`.abs()` on `F` is CubeCL's kernel-only `abs`)
                    let zero = crate::core::cast::<F>(0.0);
                    let devs: Vec<F> = row.iter().map(|&v| v - centers[r]).map(|d| if d < zero { zero - d } else { d }).collect();
                    assert_eq!(dev[r], sorted_k(devs, k), "{} deviation cols {cols:?} k {k} row {r}", client.name());
                }
            }
        }
    }

    fn signed_and_dev_kth(client: &Client) {
        signed_and_dev_kth_of::<f32>(client);
        if client.properties().supports_type(f64::elem_type_native()) {
            signed_and_dev_kth_of::<f64>(client);
        }
    }
    runtime_test!(test_row_signed_and_deviation_kth_match_host, signed_and_dev_kth);
}

/// Per-row `k`-th smallest `|x|` (0-based, a sample value, exact) over columns `cols` of a buffer
/// of `rows` rows `row_stride` apart, into `out` (`rows` values of `F`). `k < cols.len()`. A radix
/// select over the bits of `|x|` ([`fn@row_kth_radix_kernel`]): 4 passes over each row for
/// `f32`, 8 for `f64`.
///
/// # Panics
/// If `F` is neither 32 nor 64 bits wide (no unsigned key of its width on every runtime).
pub fn row_abs_kth<F: DspFloat>(
    client: &Client,
    input: &Handle,
    out: &Handle,
    rows: usize,
    row_stride: usize,
    cols: std::ops::Range<usize>,
    k: usize,
) {
    row_kth_select::<F>(client, input, None, out, rows, row_stride, cols, k, KEY_ABS);
}

/// [`row_abs_kth`] of the signed values: the per-row `k`-th smallest `x` (a percentile, exact).
pub fn row_kth<F: DspFloat>(
    client: &Client,
    input: &Handle,
    out: &Handle,
    rows: usize,
    row_stride: usize,
    cols: std::ops::Range<usize>,
    k: usize,
) {
    row_kth_select::<F>(client, input, None, out, rows, row_stride, cols, k, KEY_SIGNED);
}

/// [`row_abs_kth`] of `|x − center[row]|` (`center`: `rows` values of `F`): the per-row `k`-th
/// smallest deviation from a centre, written as the deviation.
#[allow(clippy::too_many_arguments)]
pub fn row_abs_dev_kth<F: DspFloat>(
    client: &Client,
    input: &Handle,
    center: &Handle,
    out: &Handle,
    rows: usize,
    row_stride: usize,
    cols: std::ops::Range<usize>,
    k: usize,
) {
    row_kth_select::<F>(client, input, Some(center), out, rows, row_stride, cols, k, KEY_ABS_DEV);
}

#[allow(clippy::too_many_arguments)]
fn row_kth_select<F: DspFloat>(
    client: &Client,
    input: &Handle,
    center: Option<&Handle>,
    out: &Handle,
    rows: usize,
    row_stride: usize,
    cols: std::ops::Range<usize>,
    k: usize,
    mode: u32,
) {
    assert!(k < cols.len() && cols.end <= row_stride, "selection outside the rows");
    let geom = LaunchGeometry::per_row(client, rows, cols.len());
    let args = (rows as u32, row_stride as u32, cols.start as u32, cols.len() as u32, k as u32, (8 * size_of::<F>()) as u32, geom.cube_dim.x);
    // Modes without a centre get a one-value placeholder the kernel never reads (never the output
    // itself: one buffer bound twice, once written, is an aliased binding)
    let (center, center_len) = match center {
        Some(c) => (c.clone(), rows),
        None => (crate::core::buffer::empty::<F>(client, 1), 1),
    };
    // SAFETY: `input` holds `rows · row_stride`, `center` and `out` `rows` values of `F`
    unsafe {
        let (input, center, out) = (
            BufferArg::from_raw_parts(input.clone(), rows * row_stride),
            BufferArg::from_raw_parts(center, center_len),
            BufferArg::from_raw_parts(out.clone(), rows),
        );
        match size_of::<F>() {
            4 => row_kth_radix_kernel::launch::<F, u32>(client, geom.cube_count, geom.cube_dim, input, center, out, args.0, args.1, args.2, args.3, args.4, args.5, args.6, mode),
            8 => row_kth_radix_kernel::launch::<F, u64>(client, geom.cube_count, geom.cube_dim, input, center, out, args.0, args.1, args.2, args.3, args.4, args.5, args.6, mode),
            other => panic!("row_kth: no radix key for {other}-byte floats"),
        }
    }
}