tract-linalg 0.23.7

Tiny, no-nonsense, self contained, TensorFlow and ONNX inference
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

//! f32 ln kernels: the [`crate::generic::ln`] fit, eight or sixteen lanes at a time.

routine_ew_rust!(x86_64;
    f32,
    x86_64_fma_ln_f32_32n,
    32,
    8,
    #[inline(never)]
    fn run(buf: &mut [f32], _: ()) {
        debug_assert!(buf.len() % Self::nr() == 0);
        debug_assert!(buf.as_ptr() as usize % Self::alignment_bytes() == 0);
        unsafe { x86_64_fma_ln_f32_32n_run(buf) }
    },
    func(Ln),
    isa(X86_64Avx2, X86_64Fma)
);

#[cfg(target_arch = "x86_64")]
#[target_feature(enable = "avx2,fma")]
unsafe fn x86_64_fma_ln_f32_32n_run(buf: &mut [f32]) {
    use crate::generic::ln::{LN2_HI, LN2_LO, POLY, SPLIT, SUBNORMAL_SCALE, SUBNORMAL_SHIFT};
    use std::arch::x86_64::*;
    #[inline(always)]
    unsafe fn ln8(x: __m256) -> __m256 {
        unsafe {
            let one = _mm256_set1_ps(1.0);
            let zero = _mm256_setzero_ps();
            let subnormal = _mm256_cmp_ps::<_CMP_LT_OQ>(x, _mm256_set1_ps(f32::MIN_POSITIVE));
            let scaled =
                _mm256_blendv_ps(x, _mm256_mul_ps(x, _mm256_set1_ps(SUBNORMAL_SCALE)), subnormal);
            let bits = _mm256_castps_si256(scaled);
            let mut exponent =
                _mm256_sub_epi32(_mm256_srli_epi32::<23>(bits), _mm256_set1_epi32(127));
            exponent = _mm256_add_epi32(
                exponent,
                _mm256_and_si256(
                    _mm256_castps_si256(subnormal),
                    _mm256_set1_epi32(-SUBNORMAL_SHIFT),
                ),
            );
            let mantissa = _mm256_or_ps(
                _mm256_and_ps(scaled, _mm256_castsi256_ps(_mm256_set1_epi32(0x007fffff))),
                one,
            );
            let split = _mm256_cmp_ps::<_CMP_GT_OQ>(mantissa, _mm256_set1_ps(SPLIT));
            let mantissa =
                _mm256_blendv_ps(mantissa, _mm256_mul_ps(mantissa, _mm256_set1_ps(0.5)), split);
            exponent = _mm256_add_epi32(
                exponent,
                _mm256_and_si256(_mm256_castps_si256(split), _mm256_set1_epi32(1)),
            );
            let e = _mm256_cvtepi32_ps(exponent);
            let f = _mm256_sub_ps(mantissa, one);
            let mut p = _mm256_set1_ps(POLY[0]);
            for c in &POLY[1..] {
                p = _mm256_fmadd_ps(p, f, _mm256_set1_ps(*c));
            }
            let f2 = _mm256_mul_ps(f, f);
            let mut y = _mm256_mul_ps(_mm256_mul_ps(p, f2), f);
            y = _mm256_fmadd_ps(e, _mm256_set1_ps(LN2_LO), y);
            y = _mm256_add_ps(_mm256_fnmadd_ps(_mm256_set1_ps(0.5), f2, y), f);
            y = _mm256_fmadd_ps(e, _mm256_set1_ps(LN2_HI), y);
            // A NaN compares false whichever way round, so `not greater than zero` is what
            // gathers the negatives, the zeros and the NaNs in one mask.
            let outside = _mm256_cmp_ps::<_CMP_NGT_UQ>(x, zero);
            let special = _mm256_blendv_ps(
                _mm256_set1_ps(f32::NAN),
                _mm256_set1_ps(f32::NEG_INFINITY),
                _mm256_cmp_ps::<_CMP_EQ_OQ>(x, zero),
            );
            y = _mm256_blendv_ps(y, special, outside);
            _mm256_blendv_ps(
                y,
                _mm256_set1_ps(f32::INFINITY),
                _mm256_cmp_ps::<_CMP_EQ_OQ>(x, _mm256_set1_ps(f32::INFINITY)),
            )
        }
    }
    unsafe {
        let p = buf.as_mut_ptr();
        for i in (0..buf.len()).step_by(8) {
            _mm256_store_ps(p.add(i), ln8(_mm256_load_ps(p.add(i))));
        }
    }
}

routine_ew_rust!(x86_64;
    f32,
    x86_64_avx512_ln_f32_64n,
    64,
    16,
    #[inline(never)]
    fn run(buf: &mut [f32], _: ()) {
        debug_assert!(buf.len() % Self::nr() == 0);
        debug_assert!(buf.as_ptr() as usize % Self::alignment_bytes() == 0);
        unsafe { x86_64_avx512_ln_f32_64n_run(buf) }
    },
    func(Ln),
    isa(X86_64Avx512f)
);

#[cfg(target_arch = "x86_64")]
#[target_feature(enable = "avx512f")]
unsafe fn x86_64_avx512_ln_f32_64n_run(buf: &mut [f32]) {
    use crate::generic::ln::{LN2_HI, LN2_LO, POLY, SPLIT, SUBNORMAL_SCALE, SUBNORMAL_SHIFT};
    use std::arch::x86_64::*;
    #[inline(always)]
    unsafe fn ln16(x: __m512) -> __m512 {
        unsafe {
            let one = _mm512_set1_ps(1.0);
            let zero = _mm512_setzero_ps();
            let subnormal = _mm512_cmp_ps_mask::<_CMP_LT_OQ>(x, _mm512_set1_ps(f32::MIN_POSITIVE));
            let scaled = _mm512_mask_mul_ps(x, subnormal, x, _mm512_set1_ps(SUBNORMAL_SCALE));
            let bits = _mm512_castps_si512(scaled);
            let mut exponent =
                _mm512_sub_epi32(_mm512_srli_epi32::<23>(bits), _mm512_set1_epi32(127));
            exponent = _mm512_mask_sub_epi32(
                exponent,
                subnormal,
                exponent,
                _mm512_set1_epi32(SUBNORMAL_SHIFT),
            );
            let mantissa = _mm512_or_ps(
                _mm512_and_ps(scaled, _mm512_castsi512_ps(_mm512_set1_epi32(0x007fffff))),
                one,
            );
            let split = _mm512_cmp_ps_mask::<_CMP_GT_OQ>(mantissa, _mm512_set1_ps(SPLIT));
            let mantissa = _mm512_mask_mul_ps(mantissa, split, mantissa, _mm512_set1_ps(0.5));
            exponent = _mm512_mask_add_epi32(exponent, split, exponent, _mm512_set1_epi32(1));
            let e = _mm512_cvtepi32_ps(exponent);
            let f = _mm512_sub_ps(mantissa, one);
            let mut p = _mm512_set1_ps(POLY[0]);
            for c in &POLY[1..] {
                p = _mm512_fmadd_ps(p, f, _mm512_set1_ps(*c));
            }
            let f2 = _mm512_mul_ps(f, f);
            let mut y = _mm512_mul_ps(_mm512_mul_ps(p, f2), f);
            y = _mm512_fmadd_ps(e, _mm512_set1_ps(LN2_LO), y);
            y = _mm512_add_ps(_mm512_fnmadd_ps(_mm512_set1_ps(0.5), f2, y), f);
            y = _mm512_fmadd_ps(e, _mm512_set1_ps(LN2_HI), y);
            let outside = _mm512_cmp_ps_mask::<_CMP_NGT_UQ>(x, zero);
            let special = _mm512_mask_blend_ps(
                _mm512_cmp_ps_mask::<_CMP_EQ_OQ>(x, zero),
                _mm512_set1_ps(f32::NAN),
                _mm512_set1_ps(f32::NEG_INFINITY),
            );
            y = _mm512_mask_blend_ps(outside, y, special);
            _mm512_mask_blend_ps(
                _mm512_cmp_ps_mask::<_CMP_EQ_OQ>(x, _mm512_set1_ps(f32::INFINITY)),
                y,
                _mm512_set1_ps(f32::INFINITY),
            )
        }
    }
    unsafe {
        let p = buf.as_mut_ptr();
        for i in (0..buf.len()).step_by(16) {
            _mm512_store_ps(p.add(i), ln16(_mm512_load_ps(p.add(i))));
        }
    }
}