rlx-fft 0.2.7

Learned FFT via butterfly networks — train for reference precision, run compiled on RLX backends
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

// 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.
//
// This program is distributed in the hope that it will be useful,
// but WITHOUT ANY WARRANTY; without even the implied warranty of
// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
// GNU General Public License for more details.
//
// You should have received a copy of the GNU General Public License
// along with this program. If not, see <https://www.gnu.org/licenses/>.

//! Compiled inference session for learned FFT / IFFT.

use crate::butterfly::{
    build_butterfly_forward_graph, build_butterfly_inverse_graph, butterfly_forward_real_batch,
    butterfly_inverse_complex_batch,
};
use crate::config::{FftLearnConfig, TransformDir};
use crate::reference::{fft_real_batch, ifft_complex_batch, max_abs_error};
use crate::twiddle::exact_twiddles;
use crate::weights::WeightStore;
use anyhow::{Result, bail};
use rlx_runtime::{CompiledGraph, Device};

pub struct FftLearnRunner {
    cfg: FftLearnConfig,
    direction: TransformDir,
    twiddles: Vec<f32>,
    compiled: Option<(Device, CompiledGraph)>,
}

impl FftLearnRunner {
    pub fn new(cfg: FftLearnConfig) -> Result<Self> {
        Self::new_dir(cfg, TransformDir::Forward)
    }

    pub fn new_ifft(cfg: FftLearnConfig) -> Result<Self> {
        Self::new_dir(cfg, TransformDir::Inverse)
    }

    pub fn new_dir(cfg: FftLearnConfig, direction: TransformDir) -> Result<Self> {
        cfg.validate()?;
        Ok(Self {
            twiddles: exact_twiddles(&cfg),
            cfg,
            direction,
            compiled: None,
        })
    }

    pub fn with_weights(cfg: FftLearnConfig, weights: &WeightStore) -> Result<Self> {
        Self::with_weights_dir(cfg, weights, TransformDir::Forward)
    }

    pub fn with_weights_ifft(cfg: FftLearnConfig, weights: &WeightStore) -> Result<Self> {
        Self::with_weights_dir(cfg, weights, TransformDir::Inverse)
    }

    pub fn with_weights_dir(
        cfg: FftLearnConfig,
        weights: &WeightStore,
        direction: TransformDir,
    ) -> Result<Self> {
        let mut this = Self::new_dir(cfg, direction)?;
        this.twiddles = weights.to_twiddles(this.cfg.n_fft)?;
        Ok(this)
    }

    pub fn load_compiled(&mut self, device: Device) -> Result<()> {
        let built = if self.direction.is_forward() {
            build_butterfly_forward_graph(&self.cfg)?
        } else {
            build_butterfly_inverse_graph(&self.cfg)?
        };
        let store = WeightStore::from_twiddles(&self.twiddles, self.cfg.n_fft);
        let mut compiled = crate::compile::try_compile_graph(device, built.graph)?;
        store.apply_butterfly(&mut compiled, self.cfg.batch, self.cfg.n_fft);
        self.compiled = Some((device, compiled));
        Ok(())
    }

    pub fn forward_eager(&self, input: &[f32]) -> Result<Vec<f32>> {
        if self.direction.is_forward() {
            butterfly_forward_real_batch(input, &self.twiddles, self.cfg.batch, self.cfg.n_fft)
        } else {
            butterfly_inverse_complex_batch(input, &self.twiddles, self.cfg.batch, self.cfg.n_fft)
        }
    }

    pub fn forward(&mut self, input: &[f32]) -> Result<Vec<f32>> {
        if self.compiled.is_some() {
            self.forward_compiled(input)
        } else {
            self.forward_eager(input)
        }
    }

    fn forward_compiled(&mut self, input: &[f32]) -> Result<Vec<f32>> {
        let expected = if self.direction.is_forward() {
            self.cfg.batch * self.cfg.n_fft
        } else {
            self.cfg.batch * self.cfg.n_fft * 2
        };
        if input.len() != expected {
            bail!("input len {} != expected {}", input.len(), expected);
        }
        let Some((_, ref mut exec)) = self.compiled else {
            bail!("compiled session not loaded");
        };
        let input_name = if self.direction.is_forward() {
            "signal"
        } else {
            "spectrum"
        };
        let outputs = exec.run(&[(input_name, input)]);
        outputs
            .into_iter()
            .next()
            .ok_or_else(|| anyhow::anyhow!("butterfly graph produced no outputs"))
    }

    pub fn compare_reference(&self, input: &[f32]) -> Result<(f32, f32)> {
        let pred = self.forward_eager(input)?;
        let target = if self.direction.is_forward() {
            fft_real_batch(input, self.cfg.batch, self.cfg.n_fft)?
        } else {
            ifft_complex_batch(input, self.cfg.batch, self.cfg.n_fft)?
        };
        Ok((
            crate::reference::mse(&pred, &target),
            max_abs_error(&pred, &target),
        ))
    }

    pub fn config(&self) -> &FftLearnConfig {
        &self.cfg
    }

    pub fn direction(&self) -> TransformDir {
        self.direction
    }
}