ruda-fft 0.1.8

Fast Fourier transforms for Ruda.
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

use super::*;
use crate::layout::BatchSignalLayout;
use ruda_kernel::library::tensor::{AsView as _, AsViewExpand, AsViewMut as _, AsViewMutExpand};

#[ruda(launch)]
pub(super) fn make_chirp(
    chirp_re: &mut Tensor<f32>, chirp_im: &mut Tensor<f32>,
    b_re: &mut Tensor<f32>, b_im: &mut Tensor<f32>,
    #[comptime] n: usize, #[comptime] m: usize, #[comptime] mode: FftMode,
) {
    let k = ABSOLUTE_POS;
    if k >= m { terminate!(); }
    let sign = comptime![mode.sign()];
    // Reduce the integer square modulo 2N *before* converting to f32. A
    // direct f32(k*k) loses phase accuracy even for moderate transform sizes.
    let j = if k < n { k } else { m - k };
    if j < n {
        let square = (j as u64 * j as u64) % (2u64 * n as u64);
        let phase = sign * core::f32::consts::PI * (square as f32 / n as f32);
        let re = phase.cos();
        let im = phase.sin();
        b_re[k] = re;
        b_im[k] = -im;
        if k < n { chirp_re[k] = re; chirp_im[k] = im; }
    } else { b_re[k] = 0.0; b_im[k] = 0.0; }
}

#[ruda(launch)]
pub(super) fn prepare_forward(
    input: &Tensor<f32>, chirp_re: &Tensor<f32>, chirp_im: &Tensor<f32>,
    work_re: &mut Tensor<f32>, work_im: &mut Tensor<f32>, total: u32, used: u32,
    #[comptime] n: usize, #[comptime] m: usize, #[comptime] dim: usize,
) {
    let pos = ABSOLUTE_POS;
    if pos >= total as usize { terminate!(); }
    let k = pos % m;
    let mut re = 0.0f32;
    let mut im = 0.0f32;
    if k < n && k < used as usize {
        let view = input.view(BatchSignalLayout::new(input, pos / m, dim));
        let x = view[k];
        re = x * chirp_re[k]; im = x * chirp_im[k];
    }
    work_re[pos] = re; work_im[pos] = im;
}

#[ruda(launch)]
pub(super) fn prepare_inverse(
    input_re: &Tensor<f32>, input_im: &Tensor<f32>,
    chirp_re: &Tensor<f32>, chirp_im: &Tensor<f32>,
    work_re: &mut Tensor<f32>, work_im: &mut Tensor<f32>, total: u32, used: u32,
    #[comptime] n: usize, #[comptime] m: usize, #[comptime] dim: usize,
) {
    let pos = ABSOLUTE_POS;
    if pos >= total as usize { terminate!(); }
    let k = pos % m;
    let mut re = 0.0f32;
    let mut im = 0.0f32;
    if k < n {
        let bin = if k <= n / 2 { k } else { n - k };
        if bin < used as usize {
            let rv = input_re.view(BatchSignalLayout::new(input_re, pos / m, dim));
            let iv = input_im.view(BatchSignalLayout::new(input_im, pos / m, dim));
            let xr = rv[bin];
            let mut xi = 0.0f32;
            // Never read discarded imaginary DC/Nyquist values, including NaN.
            if bin != 0 && !(n % 2 == 0 && bin == n / 2) {
                xi = if k <= n / 2 { iv[bin] } else { -iv[bin] };
            }
            let cr = chirp_re[k]; let ci = chirp_im[k];
            re = xr * cr - xi * ci; im = xr * ci + xi * cr;
        }
    }
    work_re[pos] = re; work_im[pos] = im;
}

#[ruda(launch)]
pub(super) fn multiply_spectrum(
    re: &mut Tensor<f32>, im: &mut Tensor<f32>,
    b_re: &Tensor<f32>, b_im: &Tensor<f32>, total: u32, #[comptime] m: usize,
) {
    let pos = ABSOLUTE_POS;
    if pos >= total as usize { terminate!(); }
    let k = pos % m;
    let ar = re[pos]; let ai = im[pos]; let br = b_re[k]; let bi = b_im[k];
    re[pos] = ar * br - ai * bi; im[pos] = ar * bi + ai * br;
}

#[ruda(launch)]
pub(super) fn finish_forward(
    re: &Tensor<f32>, im: &Tensor<f32>, cr: &Tensor<f32>, ci: &Tensor<f32>,
    output_re: &mut Tensor<f32>, output_im: &mut Tensor<f32>, total: u32,
    #[comptime] n: usize, #[comptime] m: usize, #[comptime] dim: usize,
) {
    let pos = ABSOLUTE_POS;
    if pos >= total as usize { terminate!(); }
    let bins = comptime![n / 2 + 1];
    let batch = pos / bins; let k = pos % bins; let src = batch * m + k;
    let mut rv = output_re.view_mut(BatchSignalLayout::new(output_re, batch, dim));
    let mut iv = output_im.view_mut(BatchSignalLayout::new(output_im, batch, dim));
    rv[k] = (re[src] * cr[k] - im[src] * ci[k]) / m as f32;
    if k == 0 || (n % 2 == 0 && k == n / 2) { iv[k] = 0.0; }
    else { iv[k] = (re[src] * ci[k] + im[src] * cr[k]) / m as f32; }
}

#[ruda(launch)]
pub(super) fn finish_inverse(
    re: &Tensor<f32>, im: &Tensor<f32>, cr: &Tensor<f32>, ci: &Tensor<f32>,
    output: &mut Tensor<f32>, total: u32,
    #[comptime] n: usize, #[comptime] m: usize, #[comptime] dim: usize,
) {
    let pos = ABSOLUTE_POS;
    if pos >= total as usize { terminate!(); }
    let batch = pos / n; let k = pos % n;
    let src = batch * m + k;
    let mut out = output.view_mut(BatchSignalLayout::new(output, batch, dim));
    // Sequential division avoids an overflow-prone integer M*N intermediate.
    out[k] = (re[src] * cr[k] - im[src] * ci[k]) / m as f32 / n as f32;
}

#[ruda(launch)]
pub(super) fn scalar_forward(input: &Tensor<f32>, re: &mut Tensor<f32>, im: &mut Tensor<f32>,
    rows: u32, used: u32, #[comptime] dim: usize) {
    let batch = ABSOLUTE_POS;
    if batch >= rows as usize { terminate!(); }
    let mut rv = re.view_mut(BatchSignalLayout::new(re, batch, dim));
    let mut iv = im.view_mut(BatchSignalLayout::new(im, batch, dim));
    let mut value = 0.0f32;
    if used != 0 { let x = input.view(BatchSignalLayout::new(input, batch, dim)); value = x[0]; }
    rv[0] = value; iv[0] = 0.0;
}

#[ruda(launch)]
pub(super) fn scalar_inverse(re: &Tensor<f32>, output: &mut Tensor<f32>,
    rows: u32, used: u32, #[comptime] dim: usize) {
    let batch = ABSOLUTE_POS;
    if batch >= rows as usize { terminate!(); }
    let mut out = output.view_mut(BatchSignalLayout::new(output, batch, dim));
    let mut value = 0.0f32;
    if used != 0 { let x = re.view(BatchSignalLayout::new(re, batch, dim)); value = x[0]; }
    out[0] = value;
}