hermes-simd 0.6.0

High-performance, zero-overhead SIMD abstraction 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

//! Runtime-dispatched masked SIMD operations.

use hermes_simd_core::arch::SimdArch;
use hermes_simd_core::kernel::SimdKernel;
use hermes_simd_core::{view::SimdError, Scalar as ScalarTrait};
use hermes_simd_macros::runtime_dispatch;

// ---------------------------------------------------------------------------
// Internal generic kernel implementations
// ---------------------------------------------------------------------------

/// Generic masked sum using `leading_k_mask` for the scalar tail.
#[inline]
unsafe fn masked_sum_impl<T, Arch>(data: &[T], bool_mask: &[bool]) -> T
where
    T: ScalarTrait,
    Arch: SimdArch + SimdKernel<T>,
{
    assert_eq!(
        data.len(),
        bool_mask.len(),
        "data and mask lengths must match"
    );
    let len = data.len();
    let lane_count = Arch::LANE_COUNT;
    let simd_len = (len / lane_count) * lane_count;

    let mut total = T::ZERO;
    let mut i = 0usize;

    while i < simd_len {
        let v = Arch::load_unaligned(data.as_ptr().add(i));
        let msk = Arch::mask_from_bools(&bool_mask[i..i + lane_count]);
        total += Arch::masked_sum_reduce(v, msk);
        i += lane_count;
    }

    // Scalar tail
    while i < len {
        if bool_mask[i] {
            total += data[i];
        }
        i += 1;
    }

    total
}

/// Generic masked elementwise add: `out[i] = if mask[i] { a[i] + b[i] } else { a[i] }`.
#[inline]
unsafe fn masked_add_impl<T, Arch>(
    a: &[T],
    b: &[T],
    bool_mask: &[bool],
    out: &mut [T],
) -> Result<(), SimdError>
where
    T: ScalarTrait,
    Arch: SimdArch + SimdKernel<T>,
{
    if a.len() != b.len() || a.len() != bool_mask.len() || a.len() > out.len() {
        return Err(SimdError::LengthMismatch);
    }
    let len = a.len();
    let lane_count = Arch::LANE_COUNT;
    let simd_len = (len / lane_count) * lane_count;

    let mut i = 0usize;
    while i < simd_len {
        let va = Arch::load_unaligned(a.as_ptr().add(i));
        let vb = Arch::load_unaligned(b.as_ptr().add(i));
        let msk = Arch::mask_from_bools(&bool_mask[i..i + lane_count]);
        let src = va;
        let result = Arch::masked_add(va, vb, msk, src);
        Arch::store_unaligned(out.as_mut_ptr().add(i), result);
        i += lane_count;
    }

    // Scalar tail
    while i < len {
        out[i] = if bool_mask[i] { a[i] + b[i] } else { a[i] };
        i += 1;
    }

    Ok(())
}

/// Generic masked dot product: sum of `a[i] * b[i]` where `mask[i]`.
#[inline]
unsafe fn masked_dot_impl<T, Arch>(a: &[T], b: &[T], bool_mask: &[bool]) -> Result<T, SimdError>
where
    T: ScalarTrait,
    Arch: SimdArch + SimdKernel<T>,
{
    if a.len() != b.len() || a.len() != bool_mask.len() {
        return Err(SimdError::LengthMismatch);
    }
    let len = a.len();
    let lane_count = Arch::LANE_COUNT;
    let simd_len = (len / lane_count) * lane_count;

    let mut acc = Arch::zero();
    let mut i = 0usize;

    while i < simd_len {
        let va = Arch::load_unaligned(a.as_ptr().add(i));
        let vb = Arch::load_unaligned(b.as_ptr().add(i));
        let msk = Arch::mask_from_bools(&bool_mask[i..i + lane_count]);
        acc = Arch::masked_fmadd(va, vb, acc, msk);
        i += lane_count;
    }

    let mut total = Arch::sum_reduce(acc);

    // Scalar tail
    while i < len {
        if bool_mask[i] {
            total += a[i] * b[i];
        }
        i += 1;
    }

    Ok(total)
}

// ---------------------------------------------------------------------------
// Public dispatcher functions
// ---------------------------------------------------------------------------

#[runtime_dispatch(avx512f, avx2, neon, scalar)]
pub(super) fn dispatch_masked_sum_kernel<T, A>(data: &[T], mask: &[bool]) -> T
where
    T: ScalarTrait,
    A: SimdArch + SimdKernel<T>,
{
    unsafe { masked_sum_impl::<T, A>(data, mask) }
}

#[runtime_dispatch(avx512f, avx2, neon, scalar)]
pub(super) fn dispatch_masked_dot_kernel<T, A>(
    a: &[T],
    b: &[T],
    mask: &[bool],
) -> Result<T, SimdError>
where
    T: ScalarTrait,
    A: SimdArch + SimdKernel<T>,
{
    unsafe { masked_dot_impl::<T, A>(a, b, mask) }
}

#[runtime_dispatch(avx512f, avx2, neon, scalar)]
pub(super) fn dispatch_masked_add_kernel<T, A>(
    a: &[T],
    b: &[T],
    mask: &[bool],
    out: &mut [T],
) -> Result<(), SimdError>
where
    T: ScalarTrait,
    A: SimdArch + SimdKernel<T>,
{
    unsafe { masked_add_impl::<T, A>(a, b, mask, out) }
}