cubecl-core 0.10.0-pre.3

CubeCL core create
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

use cubecl_ir::{ElemType, ManagedVariable, Type, Variable};

use crate::prelude::*;
use crate::{self as cubecl, unexpanded};

define_scalar!(ElemA);
define_size!(SizeA);

/// Change the meaning of the given cube primitive type during compilation.
///
/// # Warning
///
/// To be used for very custom kernels, it would likely lead to a JIT compiler error otherwise.
pub fn set_polyfill<E: Scalar, N: Size>(_elem: Type) {
    unexpanded!()
}

/// Expand module of [`set_polyfill()`].
pub mod set_polyfill {
    use super::*;

    /// Expand function of [`set_polyfill()`].
    pub fn expand<E: Scalar, N: Size>(scope: &mut Scope, ty: Type) {
        scope.register_type::<E>(ty.storage_type());
        scope.register_size::<N>(ty.vector_size());
    }
}

#[cube]
pub fn erf<F: Float, N: Size>(x: Vector<F, N>) -> Vector<F, N> {
    let erf = erf_positive(x.abs());
    select_many(x.less_than(Vector::new(F::new(0.0))), -erf, erf)
}

/// An approximation of the error function: <https://en.wikipedia.org/wiki/Error_function#Numerical_approximations>
///
/// > (maximum error: 1.5×10−7)
/// > All of these approximations are valid for x ≥ 0. To use these approximations for negative x, use the fact that erf x is an odd function, so erf x = −erf(−x).
#[cube]
fn erf_positive<F: Float, N: Size>(x: Vector<F, N>) -> Vector<F, N> {
    let p = Vector::new(F::new(0.3275911));
    let a1 = Vector::new(F::new(0.2548296));
    let a2 = Vector::new(F::new(-0.28449674));
    let a3 = Vector::new(F::new(1.4214137));
    let a4 = Vector::new(F::new(-1.453152));
    let a5 = Vector::new(F::new(1.0614054));
    let one = Vector::new(F::new(1.0));

    let t = one / (one + p * x);
    let tmp = ((((a5 * t + a4) * t) + a3) * t + a2) * t + a1;

    one - (tmp * t * (-x * x).exp())
}

#[allow(missing_docs)]
pub fn expand_erf(scope: &mut Scope, input: Variable, out: Variable) {
    scope.register_type::<ElemA>(input.ty.storage_type());
    scope.register_size::<SizeA>(input.vector_size());
    let res = erf::expand::<ElemA, SizeA>(scope, ManagedVariable::Plain(input).into());
    assign::expand_no_check(scope, res, ManagedVariable::Plain(out).into());
}

#[cube]
fn himul_i64<N: Size>(lhs: Vector<i32, N>, rhs: Vector<i32, N>) -> Vector<i32, N> {
    let shift = Vector::new(32);
    let mul = (Vector::<i64, N>::cast_from(lhs) * Vector::<i64, N>::cast_from(rhs)) >> shift;
    Vector::cast_from(mul)
}

#[cube]
fn himul_u64<N: Size>(lhs: Vector<u32, N>, rhs: Vector<u32, N>) -> Vector<u32, N> {
    let shift = Vector::new(32);
    let mul = (Vector::<u64, N>::cast_from(lhs) * Vector::<u64, N>::cast_from(rhs)) >> shift;
    Vector::cast_from(mul)
}

#[allow(missing_docs)]
pub fn expand_himul_64(scope: &mut Scope, lhs: Variable, rhs: Variable, out: Variable) {
    scope.register_size::<SizeA>(lhs.vector_size());
    match lhs.ty.elem_type() {
        ElemType::Int(_) => {
            let res = himul_i64::expand::<SizeA>(
                scope,
                ManagedVariable::Plain(lhs).into(),
                ManagedVariable::Plain(rhs).into(),
            );
            assign::expand_no_check(scope, res, ManagedVariable::Plain(out).into());
        }
        ElemType::UInt(_) => {
            let res = himul_u64::expand::<SizeA>(
                scope,
                ManagedVariable::Plain(lhs).into(),
                ManagedVariable::Plain(rhs).into(),
            );
            assign::expand_no_check(scope, res, ManagedVariable::Plain(out).into());
        }
        _ => unreachable!(),
    };
}

#[cube]
fn himul_sim<N: Size>(lhs: Vector<u32, N>, rhs: Vector<u32, N>) -> Vector<u32, N> {
    let low_mask = Vector::new(0xffff);
    let shift = Vector::new(16);

    let lhs_low = lhs & low_mask;
    let lhs_hi = (lhs >> shift) & low_mask;
    let rhs_low = rhs & low_mask;
    let rhs_hi = (rhs >> shift) & low_mask;

    let low_low = lhs_low * rhs_low;
    let high_low = lhs_hi * rhs_low;
    let low_high = lhs_low * rhs_hi;
    let high_high = lhs_hi * rhs_hi;

    let mid = ((low_low >> shift) & low_mask) + (high_low & low_mask) + (low_high & low_mask);
    high_high
        + ((high_low >> shift) & low_mask)
        + ((low_high >> shift) & low_mask)
        + ((mid >> shift) & low_mask)
}

#[allow(missing_docs)]
pub fn expand_himul_sim(scope: &mut Scope, lhs: Variable, rhs: Variable, out: Variable) {
    scope.register_size::<SizeA>(lhs.vector_size());
    let res = himul_sim::expand::<SizeA>(
        scope,
        ManagedVariable::Plain(lhs).into(),
        ManagedVariable::Plain(rhs).into(),
    );
    assign::expand_no_check(scope, res, ManagedVariable::Plain(out).into());
}