lak-kernels 0.1.0

BLAS-like linear algebra kernels in fully-safe Rust.
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

// gemv.rs 

use std::simd::num::SimdFloat;
use std::simd::{Simd, SimdElement};
use std::ops::{Add, AddAssign, Mul, MulAssign};

use crate::traits::Fma; 
use crate::types::{MatRef, VecMut, VecRef, Transpose};
use crate::{assert_length_eq_n_cols, assert_length_eq_n_rows}; 

use crate::l1::dot::dot; 
use crate::l1::scal::scal; 
use crate::fused::faxpy::faxpy; 
use crate::l1::dot::LANES; 
use crate::fused::faxpy::N_ROWS_PER_CHUNK;

pub(crate) const N_COLS_PER_PANEL: usize = 128; 

/// no transpose gemv 
pub(crate) fn gemv_n<T>( 
    alpha: T, 
    beta: T, 
    a: MatRef<'_, T>, 
    x: VecRef<'_, T>, 
    mut y: VecMut<'_, T>, 
) 
where  
    T: AddAssign 
        + Mul<Output=T> 
        + MulAssign 
        + SimdElement
        + Fma, 

    Simd<T, N_ROWS_PER_CHUNK>: SimdFloat<Scalar=T>
        + AddAssign 
        + Mul<Output=Simd<T, N_ROWS_PER_CHUNK>>
        + Fma,
{ 
    assert_length_eq_n_cols!(a, x); 
    assert_length_eq_n_rows!(a, y);

    scal(beta, y.reborrow()); 

    for (cols, a_panel) in a.col_panels(N_COLS_PER_PANEL) {
        let x_panel = VecRef::new(x.slice(cols)); 
        faxpy(alpha, a_panel, x_panel, y.reborrow());
    }
}

/// transpose gemv 
pub(crate) fn gemv_t<T>( 
    alpha: T, 
    beta: T, 
    a: MatRef<'_, T>, 
    x: VecRef<'_, T>, 
    mut y: VecMut<'_, T>, 
)
where 
    T: Copy 
        + Mul<Output=T>
        + MulAssign 
        + SimdElement 
        + Default
        + AddAssign
        + Fma 
        + Add<Output=T>, 

    Simd<T, LANES>: SimdFloat<Scalar=T>
        + Fma
        + AddAssign,      
{ 
    assert_length_eq_n_cols!(a, y); 
    assert_length_eq_n_rows!(a, x); 

    scal(beta, y.reborrow()); 

    let y_slice = y.as_slice_mut(); 

    for j in 0..a.n_cols() { 
        let aj = a.col(j); 
        y_slice[j] += alpha * dot(aj, x); 
    }
} 


/// general matrix-vector multiplication 
///
/// y <- alpha * Ax + beta * y 
///
/// args: 
/// * trans: [Transpose] - whether to use A or A^T 
/// * alpha: T - scales A * x 
/// * beta: T - scales y 
/// * a: [MatRef] - matrix A 
/// * x: [VecRef] - vector x 
/// * y: [VecMut] - vector y
///
/// example:
/// ```
/// use lak::l2::gemv;
/// use lak::types::{MatRef, VecMut, VecRef, Transpose};
///
/// // col-major 2 x 3 matrix:
/// // [1 3 5]
/// // [2 4 6]
/// let a = [1.0, 2.0, 3.0, 4.0, 5.0, 6.0];
/// let x = [1.0, 1.0, 1.0];
/// let mut y = [0.0, 0.0];
///
/// let alpha = 1.0;
/// let beta = 0.0;
///
/// let a = MatRef::new(&a, (2, 3));
/// let x = VecRef::new(&x);
/// let mut y = VecMut::new(&mut y);
/// 
/// // reborrow() used to allow y.as_slice() afterwards in the assert.
/// gemv(Transpose::NoTranspose, alpha, beta, a, x, y.reborrow());
///
/// assert_eq!(y.as_slice(), &[9.0, 12.0]);
/// ```
pub fn gemv<T>( 
    trans: Transpose, 
    alpha: T, 
    beta: T, 
    a: MatRef<'_, T>, 
    x: VecRef<'_, T>, 
    y: VecMut<'_, T>, 
)  
where  
    T: AddAssign 
        + Mul<Output=T> 
        + MulAssign 
        + SimdElement
        + Fma
        + Default
        + Add<Output=T>, 

    Simd<T, N_ROWS_PER_CHUNK>: SimdFloat<Scalar=T>
        + AddAssign 
        + Mul<Output=Simd<T, N_ROWS_PER_CHUNK>>
        + Fma,

    Simd<T, LANES>: SimdFloat<Scalar=T>
        + Fma
        + AddAssign,      
{
    match trans { 
        Transpose::NoTranspose => gemv_n(alpha, beta, a, x, y), 
        Transpose::Transpose   => gemv_t(alpha, beta, a, x, y), 
    }
}