cubek-reduce 0.3.0-pre.2

CubeK: Reduce Kernels
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

use super::{ReduceFamily, ReduceInstruction};
use crate::components::{
    instructions::{
        Accumulator, AccumulatorFormat, Item, ReduceOutputMode, ReduceRequirements, ReduceStep,
        Value, normalize_to_flag,
    },
    precision::ReducePrecision,
};
use cubecl::prelude::*;

/// Logical-OR reduction: returns `1` if any element along the reduced axis is
/// non-zero, and `0` otherwise.
///
/// Each element is normalized to a `0/1` flag, then combined with `max`. On
/// `{0, 1}`, OR is exactly `max`, so the vectorized `plane_max` machinery is
/// reused unchanged.
#[derive(Debug, CubeType, Clone)]
pub struct Any;

impl ReduceFamily for Any {
    type Instruction<P: ReducePrecision> = Self;
    type Config = ();
}

#[cube]
impl<P: ReducePrecision> ReduceInstruction<P> for Any {
    type SharedAccumulator = Shared<[Vector<P::EA, P::SI>]>;
    type Config = ();

    fn requirements(_this: &Self) -> ReduceRequirements {
        ReduceRequirements { coordinates: false }
    }

    fn accumulator_format(_this: &Self) -> comptime_type!(AccumulatorFormat) {
        AccumulatorFormat::Single
    }

    fn from_config(_config: Self::Config) -> Self {
        Any {}
    }

    fn null_input(_this: &Self) -> Vector<P::EI, P::SI> {
        // 0 normalizes to flag 0, the identity for OR.
        Vector::empty().fill(P::EI::from_int(0))
    }

    fn null_accumulator(_this: &Self) -> Accumulator<P> {
        Accumulator::<P> {
            elements: Value::new_single(Vector::empty().fill(P::EA::from_int(0))),
            args: Value::new_None(),
        }
    }

    fn reduce(
        _this: &Self,
        accumulator: &mut Accumulator<P>,
        item: Item<P>,
        #[comptime] reduce_step: ReduceStep,
    ) {
        let accumulator_item = accumulator.elements.item();
        let flag = normalize_to_flag::<P::EI, P::SI>(item.elements);
        let elements = match reduce_step {
            ReduceStep::Plane => {
                let candidate_item = Vector::cast_from(plane_max(flag));
                select_many(
                    accumulator_item.greater_than(&candidate_item),
                    accumulator_item,
                    candidate_item,
                )
            }
            ReduceStep::Identity => {
                let flag = Vector::cast_from(flag);
                select_many(accumulator_item.greater_than(&flag), accumulator_item, flag)
            }
        };

        accumulator.elements.assign(&Value::new_single(elements));
    }

    fn plane_reduce_inplace(_this: &Self, accumulator: &mut Accumulator<P>) {
        // The accumulator already holds 0/1 flags, so max-of-flags is OR.
        let acc_item = accumulator.elements.item();
        let candidate_item = Vector::cast_from(plane_max(acc_item));
        let any = select_many(
            acc_item.greater_than(&candidate_item),
            acc_item,
            candidate_item,
        );
        accumulator.elements.assign(&Value::new_single(any));
    }

    fn fuse_accumulators(_this: &Self, accumulator: &mut Accumulator<P>, other: &Accumulator<P>) {
        let accumulator_item = accumulator.elements.item();
        let other_item = other.elements.item();

        accumulator.elements.assign(&Value::new_single(select_many(
            accumulator_item.greater_than(&other_item),
            accumulator_item,
            other_item,
        )));
    }

    fn output_mode(_this: &Self) -> comptime_type!(ReduceOutputMode) {
        ReduceOutputMode::Values
    }

    fn to_output_parallel<Out: Numeric, Idx: Numeric>(
        _this: &Self,
        accumulator: Accumulator<P>,
        _shape_axis_reduce: usize,
    ) -> (Value<Out>, Value<Idx>) {
        // Fold the vectorized flags from the OR identity (0).
        let mut any = P::EA::from_int(0);
        let accumulator = accumulator.elements.item();
        #[unroll]
        for k in 0..accumulator.size() {
            let candidate = accumulator.extract(k);
            any = select(candidate > any, candidate, any);
        }
        (Value::new_single(Out::cast_from(any)), Value::new_None())
    }

    fn to_output_perpendicular<Out: Numeric, Idx: Numeric>(
        _this: &Self,
        accumulator: Accumulator<P>,
        _shape_axis_reduce: usize,
    ) -> (Value<Vector<Out, P::SI>>, Value<Vector<Idx, P::SI>>) {
        (
            Value::new_single(Vector::cast_from(accumulator.elements.item())),
            Value::new_None(),
        )
    }
}