syom 0.8.0

AAC to planar f32. LC, HE-AAC, AAC-LD. ADTS, M4A, LATM. Zero deps.
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

//! Long-window overlap. Separate mul and add, never FMA, so each lane
//! matches the scalar rounding.

/// OnlyLong: `dst[i] = scratch[i] * left[i] + overlap[i]`,
/// `overlap[i] = scratch[1024 + i] * right_rev[i]`.
pub(crate) fn window_overlap(
    scratch: &[f32],
    overlap: &mut [f32],
    dst: &mut [f32],
    left: &[f32],
    right_rev: &[f32],
) {
    const N: usize = 1024;
    debug_assert!(scratch.len() >= 2 * N);
    debug_assert!(overlap.len() >= N && dst.len() >= N);
    debug_assert!(left.len() >= N && right_rev.len() >= N);
    #[cfg(target_arch = "x86_64")]
    if std::is_x86_feature_detected!("avx512f") {
        // SAFETY: AVX-512F is available and each slice covers 1024 lanes
        // (scratch covers 2048). The loop steps by 16.
        unsafe { window_overlap_avx512(scratch, overlap, dst, left, right_rev) }
        return;
    }
    for i in 0..N {
        let prod = scratch[i] * left[i];
        dst[i] = prod + overlap[i];
        overlap[i] = scratch[N + i] * right_rev[i];
    }
}

/// # Safety
///
/// AVX-512F is available. `scratch` has 2048 lanes; `overlap`, `dst`,
/// `left`, and `right_rev` have 1024.
#[cfg(target_arch = "x86_64")]
#[target_feature(enable = "avx512f")]
unsafe fn window_overlap_avx512(
    scratch: &[f32],
    overlap: &mut [f32],
    dst: &mut [f32],
    left: &[f32],
    right_rev: &[f32],
) {
    use std::arch::x86_64::{_mm512_add_ps, _mm512_loadu_ps, _mm512_mul_ps, _mm512_storeu_ps};
    let mut i = 0usize;
    while i < 1024 {
        // SAFETY: `i + 15 < 1024` and `1024 + i + 15 < 2048`.
        unsafe {
            let prod = _mm512_mul_ps(
                _mm512_loadu_ps(scratch.as_ptr().add(i)),
                _mm512_loadu_ps(left.as_ptr().add(i)),
            );
            let acc = _mm512_add_ps(prod, _mm512_loadu_ps(overlap.as_ptr().add(i)));
            _mm512_storeu_ps(dst.as_mut_ptr().add(i), acc);
            let next = _mm512_mul_ps(
                _mm512_loadu_ps(scratch.as_ptr().add(1024 + i)),
                _mm512_loadu_ps(right_rev.as_ptr().add(i)),
            );
            _mm512_storeu_ps(overlap.as_mut_ptr().add(i), next);
        }
        i += 16;
    }
}

/// `dst[i] = left[i] + overlap[i]`, then `overlap[i] = right[i]`.
pub(crate) fn overlap_add(dst: &mut [f32], overlap: &mut [f32], left: &[f32], right: &[f32]) {
    let n = dst
        .len()
        .min(overlap.len())
        .min(left.len())
        .min(right.len());
    #[cfg(target_arch = "x86_64")]
    if std::is_x86_feature_detected!("avx512f") {
        // SAFETY: AVX-512F is available. The vector loop stops 16 lanes
        // before `n`; the tail is scalar.
        unsafe { overlap_add_avx512(dst, overlap, left, right, n) }
        return;
    }
    for i in 0..n {
        dst[i] = left[i] + overlap[i];
        overlap[i] = right[i];
    }
}

/// # Safety
///
/// AVX-512F is available. The four slices each cover `n` lanes.
#[cfg(target_arch = "x86_64")]
#[target_feature(enable = "avx512f")]
unsafe fn overlap_add_avx512(
    dst: &mut [f32],
    overlap: &mut [f32],
    left: &[f32],
    right: &[f32],
    n: usize,
) {
    use std::arch::x86_64::{_mm512_add_ps, _mm512_loadu_ps, _mm512_storeu_ps};
    let mut i = 0usize;
    while i + 16 <= n {
        // SAFETY: 16 lanes at `i` are inside all four slices.
        unsafe {
            let sum = _mm512_add_ps(
                _mm512_loadu_ps(left.as_ptr().add(i)),
                _mm512_loadu_ps(overlap.as_ptr().add(i)),
            );
            _mm512_storeu_ps(dst.as_mut_ptr().add(i), sum);
            _mm512_storeu_ps(
                overlap.as_mut_ptr().add(i),
                _mm512_loadu_ps(right.as_ptr().add(i)),
            );
        }
        i += 16;
    }
    for j in i..n {
        dst[j] = left[j] + overlap[j];
        overlap[j] = right[j];
    }
}

#[cfg(test)]
mod tests {
    use super::{overlap_add, window_overlap};

    fn fill(n: usize, seed: u32) -> Vec<f32> {
        let mut state = seed;
        let mut out = Vec::with_capacity(n);
        for _ in 0..n {
            state = state.wrapping_mul(1_664_525).wrapping_add(1_013_904_223);
            let bits = state ^ (state >> 9);
            out.push(f32::from_bits(bits));
        }
        out
    }

    #[test]
    fn window_overlap_matches_scalar_mul_then_add() {
        let mut scratch = fill(2048, 0x1234_5678);
        scratch[0] = 0.0;
        scratch[1] = -0.0;
        scratch[2] = 1.0e-20;
        scratch[1024] = -1.0e20;
        let left = fill(1024, 0x1111);
        let right = fill(1024, 0x2222);
        let mut rev = vec![0.0f32; 1024];
        for i in 0..1024 {
            rev[i] = right[1023 - i];
        }
        let overlap0 = fill(1024, 0x3333);
        let mut z = scratch.clone();
        let mut overlap_ref = overlap0.clone();
        let mut dst_ref = vec![0.0f32; 1024];
        for i in 0..1024 {
            z[i] *= left[i];
            z[1024 + i] *= right[1023 - i];
        }
        for i in 0..1024 {
            dst_ref[i] = z[i] + overlap_ref[i];
            overlap_ref[i] = z[1024 + i];
        }
        let mut overlap = overlap0;
        let mut dst = vec![0.0f32; 1024];
        window_overlap(&scratch, &mut overlap, &mut dst, &left, &rev);
        for i in 0..1024 {
            assert_eq!(dst[i].to_bits(), dst_ref[i].to_bits(), "dst {i}");
            assert_eq!(
                overlap[i].to_bits(),
                overlap_ref[i].to_bits(),
                "overlap {i}"
            );
        }
    }

    #[test]
    fn overlap_add_matches_scalar_bits() {
        let left = fill(1024, 7);
        let right = fill(1024, 8);
        let overlap0 = fill(1024, 9);
        let mut dst_ref = vec![0.0f32; 1024];
        let mut ov_ref = overlap0.clone();
        for i in 0..1024 {
            dst_ref[i] = left[i] + ov_ref[i];
            ov_ref[i] = right[i];
        }
        let mut dst = vec![0.0f32; 1024];
        let mut ov = overlap0;
        overlap_add(&mut dst, &mut ov, &left, &right);
        for i in 0..1024 {
            assert_eq!(dst[i].to_bits(), dst_ref[i].to_bits(), "dst {i}");
            assert_eq!(ov[i].to_bits(), ov_ref[i].to_bits(), "ov {i}");
        }
    }
}