tfhe-fft 0.10.1

tfhe-fft is a pure Rust high performance fast Fourier transform 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

use crate::{c64, fft_simd::*};
use pulp::{cast, x86::*};

impl FftSimd<c64x2> for V3 {
    #[inline(always)]
    fn try_new() -> Option<Self> {
        Self::try_new()
    }
    #[inline(always)]
    fn vectorize(self, f: impl pulp::NullaryFnOnce<Output = ()>) {
        self.vectorize(f)
    }

    #[inline(always)]
    fn splat_f64(self, value: f64) -> c64x2 {
        cast(self.splat_f64x4(value))
    }

    #[inline(always)]
    fn splat(self, value: c64) -> c64x2 {
        let f128 = cast(value);
        cast(self.avx._mm256_broadcast_pd(&f128))
    }

    #[inline(always)]
    fn xor(self, a: c64x2, b: c64x2) -> c64x2 {
        cast(self.xor_f64x4(cast(a), cast(b)))
    }

    #[inline(always)]
    fn swap_re_im(self, xy: c64x2) -> c64x2 {
        cast(self.avx._mm256_permute_pd::<0b0101>(cast(xy)))
    }

    #[inline(always)]
    fn add(self, a: c64x2, b: c64x2) -> c64x2 {
        cast(self.add_f64x4(cast(a), cast(b)))
    }

    #[inline(always)]
    fn sub(self, a: c64x2, b: c64x2) -> c64x2 {
        cast(self.sub_f64x4(cast(a), cast(b)))
    }

    #[inline(always)]
    fn real_mul(self, a: c64x2, b: c64x2) -> c64x2 {
        cast(self.mul_f64x4(cast(a), cast(b)))
    }

    #[inline(always)]
    fn mul(self, a: c64x2, b: c64x2) -> c64x2 {
        let xy = cast(b);
        let yx = cast(self.swap_re_im(b));
        let ab = cast(a);
        let aa = cast(self.avx._mm256_unpacklo_pd(ab, ab));
        let bb = cast(self.avx._mm256_unpackhi_pd(ab, ab));
        cast(self.mul_subadd_f64x4(aa, xy, self.mul_f64x4(bb, yx)))
    }
    #[inline(always)]
    fn catlo(self, a: c64x2, b: c64x2) -> c64x2 {
        cast(
            self.avx
                ._mm256_permute2f128_pd::<0b0010_0000>(cast(a), cast(b)),
        )
    }

    #[inline(always)]
    fn cathi(self, a: c64x2, b: c64x2) -> c64x2 {
        cast(
            self.avx
                ._mm256_permute2f128_pd::<0b0011_0001>(cast(a), cast(b)),
        )
    }
}

#[cfg(feature = "avx512")]
impl FftSimd<c64x4> for V4 {
    #[inline(always)]
    fn try_new() -> Option<Self> {
        Self::try_new()
    }
    #[inline(always)]
    fn vectorize(self, f: impl pulp::NullaryFnOnce<Output = ()>) {
        self.vectorize(f)
    }

    #[inline(always)]
    fn splat_f64(self, value: f64) -> c64x4 {
        cast(self.splat_f64x8(value))
    }

    #[inline(always)]
    fn splat(self, value: c64) -> c64x4 {
        let f128 = cast(value);
        cast(self.avx512f._mm512_broadcast_f32x4(f128))
    }

    #[inline(always)]
    fn xor(self, a: c64x4, b: c64x4) -> c64x4 {
        cast(self.xor_f64x8(cast(a), cast(b)))
    }

    #[inline(always)]
    fn swap_re_im(self, xy: c64x4) -> c64x4 {
        cast(self.avx512f._mm512_permute_pd::<0b0101_0101>(cast(xy)))
    }

    #[inline(always)]
    fn add(self, a: c64x4, b: c64x4) -> c64x4 {
        cast(self.add_f64x8(cast(a), cast(b)))
    }

    #[inline(always)]
    fn sub(self, a: c64x4, b: c64x4) -> c64x4 {
        cast(self.sub_f64x8(cast(a), cast(b)))
    }

    #[inline(always)]
    fn real_mul(self, a: c64x4, b: c64x4) -> c64x4 {
        cast(self.mul_f64x8(cast(a), cast(b)))
    }

    #[inline(always)]
    fn mul(self, a: c64x4, b: c64x4) -> c64x4 {
        let xy = cast(b);
        let yx = cast(self.swap_re_im(b));
        let ab = cast(a);
        let aa = cast(self.avx512f._mm512_unpacklo_pd(ab, ab));
        let bb = cast(self.avx512f._mm512_unpackhi_pd(ab, ab));
        cast(self.mul_subadd_f64x8(aa, xy, self.mul_f64x8(bb, yx)))
    }

    #[inline(always)]
    fn transpose(self, r0: c64x4, r1: c64x4, r2: c64x4, r3: c64x4) -> (c64x4, c64x4, c64x4, c64x4) {
        let t0 = self
            .avx512f
            ._mm512_shuffle_f64x2::<0b1000_1000>(cast(r0), cast(r1));
        let t1 = self
            .avx512f
            ._mm512_shuffle_f64x2::<0b1101_1101>(cast(r0), cast(r1));
        let t2 = self
            .avx512f
            ._mm512_shuffle_f64x2::<0b1000_1000>(cast(r2), cast(r3));
        let t3 = self
            .avx512f
            ._mm512_shuffle_f64x2::<0b1101_1101>(cast(r2), cast(r3));

        let s0 = cast(self.avx512f._mm512_shuffle_f64x2::<0b1000_1000>(t0, t2));
        let s1 = cast(self.avx512f._mm512_shuffle_f64x2::<0b1101_1101>(t0, t2));
        let s2 = cast(self.avx512f._mm512_shuffle_f64x2::<0b1000_1000>(t1, t3));
        let s3 = cast(self.avx512f._mm512_shuffle_f64x2::<0b1101_1101>(t1, t3));

        (s0, s2, s1, s3)
    }
}