NeuralAmpModeler-rs 3.0.0

An opinionated, high-performance Neural Amp Modeler (NAM) client and core implementation in Rust for Linux/PipeWire and CLAP plugins.
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

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

//! AVX-512 polynomial exp/tanh kernels (degree-6 Taylor).
//!
//! Depends on scalar fallback functions from `super::high_fidelity`
//! and polynomial constants from `crate::math::constants`.
//! Sigmoid polynomial AVX-512 kernels live in
//! `crate::math::activations::sigmoid::high_fidelity_avx512`.

use super::high_fidelity::scalar_tanh_poly;
use crate::math::activations::sigmoid::high_fidelity_avx512::simd_sigmoid_poly_avx512;
use crate::math::constants::*;
use core::arch::x86_64::*;

// ══════════════════════════════════════════════════════════════════════════════
// AVX-512 polynomial exp/tanh/sigmoid kernels
// ══════════════════════════════════════════════════════════════════════════════

/// Polynomial `exp(x)` for `__m512` — degree-6 Taylor polynomial with integer
/// range reduction `x = k·ln2 + r`.  AVX-512 counterpart of `simd_exp_poly_avx2`.
///
/// # Safety
/// The caller must guarantee AVX-512F and AVX-512VL support.  Input clamped to
/// [-20, 20] to prevent overflow.
#[inline]
#[target_feature(enable = "avx512f,avx512vl")]
unsafe fn simd_exp_poly_avx512(x: __m512) -> __m512 {
    let log2e = _mm512_set1_ps(POLY_LOG2_E);
    let ln2 = _mm512_set1_ps(POLY_LN2);
    let c6 = _mm512_set1_ps(POLY_EXP_C6);
    let c5 = _mm512_set1_ps(POLY_EXP_C5);
    let c4 = _mm512_set1_ps(POLY_EXP_C4);
    let c3 = _mm512_set1_ps(POLY_EXP_C3);
    let c2 = _mm512_set1_ps(POLY_EXP_C2);
    let one = _mm512_set1_ps(1.0f32);
    let bias = _mm512_set1_epi32(127);

    // k = round(x * log₂e),  r = x − k·ln2  (∈ [-ln2/2, ln2/2])
    let k_f = _mm512_roundscale_ps(
        _mm512_mul_ps(x, log2e),
        _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC,
    );
    let r = _mm512_fnmadd_ps(k_f, ln2, x);

    // P(r) = (((((c6·r + c5)·r + c4)·r + c3)·r + c2)·r + 1)·r + 1
    let p = _mm512_fmadd_ps(c6, r, c5);
    let p = _mm512_fmadd_ps(p, r, c4);
    let p = _mm512_fmadd_ps(p, r, c3);
    let p = _mm512_fmadd_ps(p, r, c2);
    let p = _mm512_fmadd_ps(p, r, one);
    let p = _mm512_fmadd_ps(p, r, one);

    // Scale by 2ᵏ: reinterpret ((k+127) << 23) as f32 and multiply
    let k_i = _mm512_cvtps_epi32(k_f);
    let exp_bits = _mm512_slli_epi32(_mm512_add_epi32(k_i, bias), 23);
    let scale = _mm512_castsi512_ps(exp_bits);
    _mm512_mul_ps(p, scale)
}

/// Polynomial `tanh(x)` for `__m512` — exp-based, single-division (AVX-512F/VL).
///
/// Formula: `tanh(x) = (e²ˣ − 1) / (e²ˣ + 1)`.
/// Input clamped to [-20, 20] for overflow safety, output clamped to [-1, 1].
///
/// # Safety
/// The caller must guarantee AVX-512F and AVX-512VL support.
#[inline]
#[target_feature(enable = "avx512f,avx512vl")]
pub unsafe fn simd_tanh_poly_avx512(x: __m512) -> __m512 {
    let clamp_lo = _mm512_set1_ps(-POLY_ACTIVATION_CLAMP);
    let clamp_hi = _mm512_set1_ps(POLY_ACTIVATION_CLAMP);
    let one = _mm512_set1_ps(1.0f32);
    let neg_one = _mm512_set1_ps(-1.0f32);

    let x = _mm512_max_ps(clamp_lo, _mm512_min_ps(clamp_hi, x));

    let exp_x = unsafe { simd_exp_poly_avx512(x) };
    let u2 = _mm512_mul_ps(exp_x, exp_x); // e²ˣ
    let num = _mm512_sub_ps(u2, one); // e²ˣ − 1
    let den = _mm512_add_ps(u2, one); // e²ˣ + 1
    let tanh_val = _mm512_div_ps(num, den);
    _mm512_max_ps(neg_one, _mm512_min_ps(one, tanh_val))
}

/// Polynomial `(tanh(x1), sigmoid(x2))` — dual gate (AVX-512F/VL).
///
/// Used in WaveNet gated activation: `tanh(zf) * sigmoid(zg)`.
/// Delegates sigmoid to `sigmoid::high_fidelity_avx512::simd_sigmoid_poly_avx512`.
///
/// # Safety
/// The caller must guarantee AVX-512F and AVX-512VL support.
#[inline]
#[target_feature(enable = "avx512f,avx512vl")]
pub unsafe fn simd_tanh_sigmoid_dual_poly_avx512(x1: __m512, x2: __m512) -> (__m512, __m512) {
    let t1 = unsafe { simd_tanh_poly_avx512(x1) };
    let s2 = unsafe { simd_sigmoid_poly_avx512(x2) };
    (t1, s2)
}

/// Polynomial `tanh(x)` - NR1, single Newton-Raphson on reciprocal (AVX-512).
///
/// Experimental variant (TC3). Replaces `_mm512_div_ps` with `_mm512_rcp14_ps` +
/// 1×Newton-Raphson refinement (~23-bit reciprocal precision).
///
/// # Safety
/// The caller must guarantee AVX-512F and AVX-512VL support.
#[inline]
#[target_feature(enable = "avx512f,avx512vl")]
pub unsafe fn simd_tanh_poly_nr1_avx512(x: __m512) -> __m512 {
    let clamp_lo = _mm512_set1_ps(-POLY_ACTIVATION_CLAMP);
    let clamp_hi = _mm512_set1_ps(POLY_ACTIVATION_CLAMP);
    let one = _mm512_set1_ps(1.0f32);
    let neg_one = _mm512_set1_ps(-1.0f32);
    let two = _mm512_set1_ps(2.0f32);

    let x = _mm512_max_ps(clamp_lo, _mm512_min_ps(clamp_hi, x));

    let exp_x = unsafe { simd_exp_poly_avx512(x) };
    let u2 = _mm512_mul_ps(exp_x, exp_x);
    let num = _mm512_sub_ps(u2, one);
    let den = _mm512_add_ps(u2, one);

    let mut r = _mm512_rcp14_ps(den);
    r = _mm512_mul_ps(r, _mm512_fnmadd_ps(den, r, two));

    let tanh_val = _mm512_mul_ps(num, r);
    _mm512_max_ps(neg_one, _mm512_min_ps(one, tanh_val))
}

/// Polynomial `tanh(x)` - NR2, double Newton-Raphson on reciprocal (AVX-512).
///
/// Experimental variant (TC3). Replaces `_mm512_div_ps` with `_mm512_rcp14_ps` +
/// 2×Newton-Raphson refinement (saturates f32 mantissa).
///
/// # Safety
/// The caller must guarantee AVX-512F and AVX-512VL support.
#[inline]
#[target_feature(enable = "avx512f,avx512vl")]
pub unsafe fn simd_tanh_poly_nr2_avx512(x: __m512) -> __m512 {
    let clamp_lo = _mm512_set1_ps(-POLY_ACTIVATION_CLAMP);
    let clamp_hi = _mm512_set1_ps(POLY_ACTIVATION_CLAMP);
    let one = _mm512_set1_ps(1.0f32);
    let neg_one = _mm512_set1_ps(-1.0f32);
    let two = _mm512_set1_ps(2.0f32);

    let x = _mm512_max_ps(clamp_lo, _mm512_min_ps(clamp_hi, x));

    let exp_x = unsafe { simd_exp_poly_avx512(x) };
    let u2 = _mm512_mul_ps(exp_x, exp_x);
    let num = _mm512_sub_ps(u2, one);
    let den = _mm512_add_ps(u2, one);

    let mut r = _mm512_rcp14_ps(den);
    r = _mm512_mul_ps(r, _mm512_fnmadd_ps(den, r, two));
    r = _mm512_mul_ps(r, _mm512_fnmadd_ps(den, r, two));

    let tanh_val = _mm512_mul_ps(num, r);
    _mm512_max_ps(neg_one, _mm512_min_ps(one, tanh_val))
}

/// Applies polynomial Tanh activation to a slice of f32 using AVX-512.
///
/// # Safety
/// Requires AVX-512F and AVX-512VL support.
#[inline]
#[target_feature(enable = "avx512f,avx512vl")]
pub unsafe fn tanh_poly_slice_avx512(slice: &mut [f32]) {
    let mut i = 0;
    let len = slice.len();

    unsafe {
        crate::activation_simd_avx512!(i, len, {
            let x = _mm512_loadu_ps(slice.as_ptr().add(i));
            let y = simd_tanh_poly_avx512(x);
            _mm512_storeu_ps(slice.as_mut_ptr().add(i), y);
        });
    }

    for item in slice.iter_mut().skip(i) {
        *item = scalar_tanh_poly(*item);
    }
}