NeuralAmpModeler-rs 0.1.0

High-performance Neural Amp Modeler DSP core: WaveNet/LSTM/ConvNet inference, SIMD math (x86-64-v3), .nam/.namb loader, cabinet IR, resampling and noise gate.
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
228
229
230
231
232
233
234
235
236
237
238

// SPDX-License-Identifier: Apache-2.0
// Copyright (c) 2026 Fábio Henrique de Lima Silva (fhl.bsb@gmail.com) All rights reserved.

//! Unified dispatch layer: maps the [`NamModel`] trait interface onto every
//! concrete [`StaticModel`] variant via exhaustive `match` arms.
//!
//! # Concurrency model
//! All methods in this `impl` are called from the audio processing thread
//! (`SCHED_FIFO`) and must remain panic-free. The `process` method on
//! `Linear` wraps its call in `unsafe` because the Linear kernel uses
//! raw-pointer SIMD intrinsics that do not carry lifetime bounds through
//! the trait indirection; the pointer is derived from the `&[f32]`
//! arguments already borrowed by `process`.
//!
//! # Adding a new model variant
//! 1. Add the variant to [`StaticModel`].
//! 2. Thread it through every `match` arm in this file.
//! 3. Add the corresponding dispatch arm in
//!    [`crate::loader::dispatcher`] and [`crate::loader::nam_json::topology`].

use super::slimmable::SlimmableModel;
use super::{NamModel, StaticModel};

impl NamModel for StaticModel {
    /// Routes sample-by-sample neural inference to the active model variant.
    ///
    /// All WaveNet-derived, LSTM, ConvNet, and Container variants delegate
    /// to their respective `process` methods without special wrapping. The
    /// `Linear` variant uses an `unsafe` block because the raw pointer passed
    /// to the GEMM kernel cannot carry a compile-time lifetime through the
    /// enum dispatch. The pointer validity is guaranteed by the `&[f32]`
    /// borrows already held by the caller.
    #[inline(always)]
    fn process(&mut self, input: &[f32], output: &mut [f32]) {
        match self {
            Self::WavenetStandard(m) => m.process(input, output),
            Self::WavenetLite(m) => m.process(input, output),
            Self::WavenetFeather(m) => m.process(input, output),
            Self::WavenetNano(m) => m.process(input, output),
            Self::WavenetA2Full(m) => m.process(input, output),
            Self::WavenetA2Lite(m) => m.process(input, output),
            Self::WavenetA2Dyn(m) => m.process(input, output),
            Self::WavenetA2Cascade(m) => m.process(input, output),
            Self::WavenetDyn(m) => m.process(input, output),
            Self::Container(m) => m.process(input, output),
            Self::Lstm1x3(m) => m.process(input, output),
            Self::Lstm1x8(m) => m.process(input, output),
            Self::Lstm1x12(m) => m.process(input, output),
            Self::Lstm1x16(m) => m.process(input, output),
            Self::Lstm1x24(m) => m.process(input, output),
            Self::Lstm2x8(m) => m.process(input, output),
            Self::Lstm2x12(m) => m.process(input, output),
            Self::Lstm2x16(m) => m.process(input, output),
            Self::Lstm1x40(m) => m.process(input, output),
            Self::Lstm2x24(m) => m.process(input, output),
            Self::LstmDyn(m) => m.process(input, output),
            Self::Linear(m) => unsafe { m.process(input, output) },
            Self::ConvNet(m) => m.process(input, output),
        }
    }

    #[cold]
    fn prewarm(&mut self, num_samples: usize) {
        match self {
            Self::WavenetStandard(m) => m.prewarm(),
            Self::WavenetLite(m) => m.prewarm(),
            Self::WavenetFeather(m) => m.prewarm(),
            Self::WavenetNano(m) => m.prewarm(),
            Self::WavenetA2Full(m) => m.prewarm(),
            Self::WavenetA2Lite(m) => m.prewarm(),
            Self::WavenetA2Dyn(m) => m.prewarm(),
            Self::WavenetA2Cascade(m) => m.prewarm(),
            Self::WavenetDyn(m) => m.prewarm(),
            Self::Container(m) => m.prewarm(num_samples),
            Self::Lstm1x3(m) => m.prewarm(num_samples),
            Self::Lstm1x8(m) => m.prewarm(num_samples),
            Self::Lstm1x12(m) => m.prewarm(num_samples),
            Self::Lstm1x16(m) => m.prewarm(num_samples),
            Self::Lstm1x24(m) => m.prewarm(num_samples),
            Self::Lstm2x8(m) => m.prewarm(num_samples),
            Self::Lstm2x12(m) => m.prewarm(num_samples),
            Self::Lstm2x16(m) => m.prewarm(num_samples),
            Self::Lstm1x40(m) => m.prewarm(num_samples),
            Self::Lstm2x24(m) => m.prewarm(num_samples),
            Self::LstmDyn(m) => m.prewarm(num_samples),
            Self::Linear(m) => m.prewarm(num_samples),
            Self::ConvNet(m) => m.prewarm(),
        }
    }

    fn prewarm_on_reset(&self) -> bool {
        match self {
            Self::WavenetStandard(m) => m.prewarm_on_reset(),
            Self::WavenetLite(m) => m.prewarm_on_reset(),
            Self::WavenetFeather(m) => m.prewarm_on_reset(),
            Self::WavenetNano(m) => m.prewarm_on_reset(),
            Self::WavenetA2Full(m) => m.prewarm_on_reset(),
            Self::WavenetA2Lite(m) => m.prewarm_on_reset(),
            Self::WavenetA2Dyn(m) => m.prewarm_on_reset(),
            Self::WavenetA2Cascade(m) => m.prewarm_on_reset(),
            Self::WavenetDyn(m) => m.prewarm_on_reset(),
            Self::Container(m) => m.prewarm_on_reset(),
            Self::Lstm1x3(m) => m.prewarm_on_reset(),
            Self::Lstm1x8(m) => m.prewarm_on_reset(),
            Self::Lstm1x12(m) => m.prewarm_on_reset(),
            Self::Lstm1x16(m) => m.prewarm_on_reset(),
            Self::Lstm1x24(m) => m.prewarm_on_reset(),
            Self::Lstm2x8(m) => m.prewarm_on_reset(),
            Self::Lstm2x12(m) => m.prewarm_on_reset(),
            Self::Lstm2x16(m) => m.prewarm_on_reset(),
            Self::Lstm1x40(m) => m.prewarm_on_reset(),
            Self::Lstm2x24(m) => m.prewarm_on_reset(),
            Self::LstmDyn(m) => m.prewarm_on_reset(),
            Self::Linear(m) => m.prewarm_on_reset(),
            Self::ConvNet(m) => m.prewarm_on_reset(),
        }
    }

    fn set_prewarm_on_reset(&mut self, val: bool) {
        match self {
            Self::WavenetStandard(m) => m.set_prewarm_on_reset(val),
            Self::WavenetLite(m) => m.set_prewarm_on_reset(val),
            Self::WavenetFeather(m) => m.set_prewarm_on_reset(val),
            Self::WavenetNano(m) => m.set_prewarm_on_reset(val),
            Self::WavenetA2Full(m) => m.set_prewarm_on_reset(val),
            Self::WavenetA2Lite(m) => m.set_prewarm_on_reset(val),
            Self::WavenetA2Dyn(m) => m.set_prewarm_on_reset(val),
            Self::WavenetA2Cascade(m) => m.set_prewarm_on_reset(val),
            Self::WavenetDyn(m) => m.set_prewarm_on_reset(val),
            Self::Container(m) => m.set_prewarm_on_reset(val),
            Self::Lstm1x3(m) => m.set_prewarm_on_reset(val),
            Self::Lstm1x8(m) => m.set_prewarm_on_reset(val),
            Self::Lstm1x12(m) => m.set_prewarm_on_reset(val),
            Self::Lstm1x16(m) => m.set_prewarm_on_reset(val),
            Self::Lstm1x24(m) => m.set_prewarm_on_reset(val),
            Self::Lstm2x8(m) => m.set_prewarm_on_reset(val),
            Self::Lstm2x12(m) => m.set_prewarm_on_reset(val),
            Self::Lstm2x16(m) => m.set_prewarm_on_reset(val),
            Self::Lstm1x40(m) => m.set_prewarm_on_reset(val),
            Self::Lstm2x24(m) => m.set_prewarm_on_reset(val),
            Self::LstmDyn(m) => m.set_prewarm_on_reset(val),
            Self::Linear(m) => m.set_prewarm_on_reset(val),
            Self::ConvNet(m) => m.set_prewarm_on_reset(val),
        }
    }

    fn reset(&mut self, sample_rate: u32, max_buffer_size: usize) -> anyhow::Result<()> {
        match self {
            Self::WavenetStandard(m) => m.reset(sample_rate, max_buffer_size),
            Self::WavenetLite(m) => m.reset(sample_rate, max_buffer_size),
            Self::WavenetFeather(m) => m.reset(sample_rate, max_buffer_size),
            Self::WavenetNano(m) => m.reset(sample_rate, max_buffer_size),
            Self::WavenetA2Full(m) => m.reset(sample_rate, max_buffer_size),
            Self::WavenetA2Lite(m) => m.reset(sample_rate, max_buffer_size),
            Self::WavenetA2Dyn(m) => m.reset(sample_rate, max_buffer_size),
            Self::WavenetA2Cascade(m) => m.reset(sample_rate, max_buffer_size),
            Self::WavenetDyn(m) => m.reset(sample_rate, max_buffer_size),
            Self::Container(m) => m.reset(sample_rate, max_buffer_size),
            Self::Lstm1x3(m) => m.reset(sample_rate, max_buffer_size),
            Self::Lstm1x8(m) => m.reset(sample_rate, max_buffer_size),
            Self::Lstm1x12(m) => m.reset(sample_rate, max_buffer_size),
            Self::Lstm1x16(m) => m.reset(sample_rate, max_buffer_size),
            Self::Lstm1x24(m) => m.reset(sample_rate, max_buffer_size),
            Self::Lstm2x8(m) => m.reset(sample_rate, max_buffer_size),
            Self::Lstm2x12(m) => m.reset(sample_rate, max_buffer_size),
            Self::Lstm2x16(m) => m.reset(sample_rate, max_buffer_size),
            Self::Lstm1x40(m) => m.reset(sample_rate, max_buffer_size),
            Self::Lstm2x24(m) => m.reset(sample_rate, max_buffer_size),
            Self::LstmDyn(m) => m.reset(sample_rate, max_buffer_size),
            Self::Linear(m) => NamModel::reset(m.as_mut(), sample_rate, max_buffer_size),
            Self::ConvNet(m) => NamModel::reset(m.as_mut(), sample_rate, max_buffer_size),
        }
    }

    fn set_max_buffer_size(&mut self, max_buf: usize) -> anyhow::Result<()> {
        match self {
            Self::WavenetStandard(m) => m.set_max_buffer_size(max_buf),
            Self::WavenetLite(m) => m.set_max_buffer_size(max_buf),
            Self::WavenetFeather(m) => m.set_max_buffer_size(max_buf),
            Self::WavenetNano(m) => m.set_max_buffer_size(max_buf),
            Self::WavenetA2Full(m) => m.set_max_buffer_size(max_buf),
            Self::WavenetA2Lite(m) => m.set_max_buffer_size(max_buf),
            Self::WavenetA2Dyn(m) => m.set_max_buffer_size(max_buf),
            Self::WavenetA2Cascade(m) => m.set_max_buffer_size(max_buf),
            Self::WavenetDyn(m) => m.set_max_buffer_size(max_buf),
            Self::Container(m) => m.set_max_buffer_size(max_buf),
            Self::Lstm1x3(m) => m.set_max_buffer_size(max_buf),
            Self::Lstm1x8(m) => m.set_max_buffer_size(max_buf),
            Self::Lstm1x12(m) => m.set_max_buffer_size(max_buf),
            Self::Lstm1x16(m) => m.set_max_buffer_size(max_buf),
            Self::Lstm1x24(m) => m.set_max_buffer_size(max_buf),
            Self::Lstm2x8(m) => m.set_max_buffer_size(max_buf),
            Self::Lstm2x12(m) => m.set_max_buffer_size(max_buf),
            Self::Lstm2x16(m) => m.set_max_buffer_size(max_buf),
            Self::Lstm1x40(m) => m.set_max_buffer_size(max_buf),
            Self::Lstm2x24(m) => m.set_max_buffer_size(max_buf),
            Self::LstmDyn(m) => m.set_max_buffer_size(max_buf),
            Self::Linear(m) => NamModel::set_max_buffer_size(m.as_mut(), max_buf),
            Self::ConvNet(m) => NamModel::set_max_buffer_size(m.as_mut(), max_buf),
        }
    }

    fn prewarm_samples(&self) -> usize {
        match self {
            Self::WavenetStandard(m) => m.prewarm_samples(),
            Self::WavenetLite(m) => m.prewarm_samples(),
            Self::WavenetFeather(m) => m.prewarm_samples(),
            Self::WavenetNano(m) => m.prewarm_samples(),
            Self::WavenetA2Full(m) => m.prewarm_samples(),
            Self::WavenetA2Lite(m) => m.prewarm_samples(),
            Self::WavenetA2Dyn(m) => m.prewarm_samples(),
            Self::WavenetA2Cascade(m) => m.prewarm_samples(),
            Self::WavenetDyn(m) => m.prewarm_samples(),
            Self::Container(m) => m.prewarm_samples(),
            Self::Lstm1x3(m) => m.prewarm_samples(),
            Self::Lstm1x8(m) => m.prewarm_samples(),
            Self::Lstm1x12(m) => m.prewarm_samples(),
            Self::Lstm1x16(m) => m.prewarm_samples(),
            Self::Lstm1x24(m) => m.prewarm_samples(),
            Self::Lstm2x8(m) => m.prewarm_samples(),
            Self::Lstm2x12(m) => m.prewarm_samples(),
            Self::Lstm2x16(m) => m.prewarm_samples(),
            Self::Lstm1x40(m) => m.prewarm_samples(),
            Self::Lstm2x24(m) => m.prewarm_samples(),
            Self::LstmDyn(m) => m.prewarm_samples(),
            Self::Linear(m) => m.prewarm_samples(),
            Self::ConvNet(m) => m.prewarm_samples(),
        }
    }

    fn slimmable_breakpoints(&self) -> Vec<f64> {
        if let Self::Container(c) = self {
            SlimmableModel::slimmable_breakpoints(c.as_ref())
        } else {
            vec![]
        }
    }
}