resampler 0.5.1

A small audio resampling library
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

use core::arch::x86_64::*;

use crate::fft::{
    Complex32,
    butterflies::ops::{complex_mul_i_sse2, complex_mul_sse2, load_neg_imag_mask_sse2},
};

/// Performs a single radix-4 Stockham butterfly stage for stride=1 (out-of-place, SSE2).
///
/// This is a specialized version for the stride=1 case (first stage) that uses the unpack-permute
/// pattern to enable sequential SIMD stores. For stride=1, output indices are sequential: j=4*i.
#[target_feature(enable = "sse2")]
pub(super) unsafe fn butterfly_radix4_stride1_sse2(
    src: &[Complex32],
    dst: &mut [Complex32],
    stage_twiddles: &[Complex32],
) {
    let samples = src.len();
    let quarter_samples = samples >> 2;
    let simd_iters = (quarter_samples >> 1) << 1;

    unsafe {
        let neg_imag_mask = load_neg_imag_mask_sse2();

        for i in (0..simd_iters).step_by(2) {
            let z0_ptr = src.as_ptr().add(i) as *const f32;
            let z0 = _mm_loadu_ps(z0_ptr);

            let z1_ptr = src.as_ptr().add(i + quarter_samples) as *const f32;
            let z1 = _mm_loadu_ps(z1_ptr);

            let z2_ptr = src.as_ptr().add(i + quarter_samples * 2) as *const f32;
            let z2 = _mm_loadu_ps(z2_ptr);

            let z3_ptr = src.as_ptr().add(i + quarter_samples * 3) as *const f32;
            let z3 = _mm_loadu_ps(z3_ptr);

            // Identity twiddles: t1 = (1+0i) * z1 = z1, t2 = z2, t3 = z3 (skip twiddle load and multiply)
            let t1 = z1;
            let t2 = z2;
            let t3 = z3;

            // Radix-4 butterfly
            let a0 = _mm_add_ps(z0, t2);
            let a1 = _mm_sub_ps(z0, t2);
            let a2 = _mm_add_ps(t1, t3);

            // a3 = i * (t1 - t3)
            let t1_sub_t3 = _mm_sub_ps(t1, t3);
            let a3 = complex_mul_i_sse2(t1_sub_t3, neg_imag_mask);

            // Final butterfly outputs
            let out0 = _mm_add_ps(a0, a2);
            let out2 = _mm_sub_ps(a0, a2);
            let out1 = _mm_add_ps(a1, a3);
            let out3 = _mm_sub_ps(a1, a3);

            // Apply unpack-permute pattern for sequential stores.
            // We have 4 outputs (out0, out1, out2, out3), each containing 2 complex numbers
            // Need to interleave them: [out0[0], out1[0], out2[0], out3[0], out0[1], ...]

            let out0_pd = _mm_castps_pd(out0);
            let out1_pd = _mm_castps_pd(out1);
            let out2_pd = _mm_castps_pd(out2);
            let out3_pd = _mm_castps_pd(out3);

            // Interleave pairs: out0 with out1, out2 with out3.
            // After unpack: out01_lo = [out0[0], out1[0]], out23_lo = [out2[0], out3[0]]
            let out01_lo = _mm_castpd_ps(_mm_unpacklo_pd(out0_pd, out1_pd));
            let out01_hi = _mm_castpd_ps(_mm_unpackhi_pd(out0_pd, out1_pd));
            let out23_lo = _mm_castpd_ps(_mm_unpacklo_pd(out2_pd, out3_pd));
            let out23_hi = _mm_castpd_ps(_mm_unpackhi_pd(out2_pd, out3_pd));

            let j = 4 * i;
            let dst_ptr = dst.as_mut_ptr().add(j) as *mut f32;
            _mm_storeu_ps(dst_ptr, out01_lo);
            _mm_storeu_ps(dst_ptr.add(4), out23_lo);
            _mm_storeu_ps(dst_ptr.add(8), out01_hi);
            _mm_storeu_ps(dst_ptr.add(12), out23_hi);
        }
    }

    super::butterfly_radix4_scalar::<2>(src, dst, stage_twiddles, 1, simd_iters);
}

/// Performs a single radix-4 Stockham butterfly stage for p>1 (out-of-place, SSE2).
///
/// Generic version for p>1 cases. Uses direct SIMD stores, accepting non-sequential
/// stores as the shuffle overhead isn't justified for larger strides.
#[target_feature(enable = "sse2")]
pub(super) unsafe fn butterfly_radix4_generic_sse2(
    src: &[Complex32],
    dst: &mut [Complex32],
    stage_twiddles: &[Complex32],
    stride: usize,
) {
    // We convince the compiler here that stride can't be 0 to optimize better.
    if stride == 0 {
        return;
    }

    let samples = src.len();
    let quarter_samples = samples >> 2;
    let simd_iters = (quarter_samples >> 1) << 1;

    unsafe {
        let neg_imag_mask = load_neg_imag_mask_sse2();

        for i in (0..simd_iters).step_by(2) {
            let k = i % stride;
            let k0 = k;
            let k1 = k + 1 - ((k + 1 >= stride) as usize) * stride;

            // Load z0 from first quarter.
            let z0_ptr = src.as_ptr().add(i) as *const f32;
            let z0 = _mm_loadu_ps(z0_ptr);

            // Load z1, z2, z3 from other quarters.
            let z1_ptr = src.as_ptr().add(i + quarter_samples) as *const f32;
            let z1 = _mm_loadu_ps(z1_ptr);

            let z2_ptr = src.as_ptr().add(i + quarter_samples * 2) as *const f32;
            let z2 = _mm_loadu_ps(z2_ptr);

            let z3_ptr = src.as_ptr().add(i + quarter_samples * 3) as *const f32;
            let z3 = _mm_loadu_ps(z3_ptr);

            // Load prepackaged twiddles directly (no shuffle needed).
            let tw_ptr = stage_twiddles.as_ptr().add(i * 3) as *const f32;
            let w1 = _mm_loadu_ps(tw_ptr); // w1[i], w1[i+1]
            let w2 = _mm_loadu_ps(tw_ptr.add(4)); // w2[i], w2[i+1]
            let w3 = _mm_loadu_ps(tw_ptr.add(8)); // w3[i], w3[i+1]

            // Complex multiply: t1 = w1 * z1, t2 = w2 * z2, t3 = w3 * z3
            let t1 = complex_mul_sse2(w1, z1);
            let t2 = complex_mul_sse2(w2, z2);
            let t3 = complex_mul_sse2(w3, z3);

            // Radix-4 butterfly.
            let a0 = _mm_add_ps(z0, t2);
            let a1 = _mm_sub_ps(z0, t2);
            let a2 = _mm_add_ps(t1, t3);

            // a3 = i * (t1 - t3)
            let t1_sub_t3 = _mm_sub_ps(t1, t3);
            let a3 = complex_mul_i_sse2(t1_sub_t3, neg_imag_mask);

            // Final butterfly outputs.
            let out0 = _mm_add_ps(a0, a2);
            let out2 = _mm_sub_ps(a0, a2);
            let out1 = _mm_add_ps(a1, a3);
            let out3 = _mm_sub_ps(a1, a3);

            // Calculate output indices.
            let j0 = 4 * i - 3 * k0;
            let j1 = 4 * (i + 1) - 3 * k1;

            // Direct SIMD stores using 64-bit operations.
            let out0_pd = _mm_castps_pd(out0);
            let out1_pd = _mm_castps_pd(out1);
            let out2_pd = _mm_castps_pd(out2);
            let out3_pd = _mm_castps_pd(out3);

            let dst_ptr = dst.as_mut_ptr() as *mut f64;

            _mm_storel_pd(dst_ptr.add(j0), out0_pd);
            _mm_storel_pd(dst_ptr.add(j0 + stride), out1_pd);
            _mm_storel_pd(dst_ptr.add(j0 + stride * 2), out2_pd);
            _mm_storel_pd(dst_ptr.add(j0 + stride * 3), out3_pd);

            _mm_storeh_pd(dst_ptr.add(j1), out0_pd);
            _mm_storeh_pd(dst_ptr.add(j1 + stride), out1_pd);
            _mm_storeh_pd(dst_ptr.add(j1 + stride * 2), out2_pd);
            _mm_storeh_pd(dst_ptr.add(j1 + stride * 3), out3_pd);
        }
    }

    super::butterfly_radix4_scalar::<2>(src, dst, stage_twiddles, stride, simd_iters);
}