ruPRIM 0.1.9

Parallel primitives, reductions, scans, and indexing for Ruda.
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
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183

use ruda_core::device::Device;
use ruda_kernel::dsl as kernel_dsl;
use ruda_kernel::dsl::{prelude::*, calculate_ruda_count_elemwise};
use ruda_kernel::library::tensor::layout::linear::LinearView;
use ruda_kernel::tensor::{RudaTensor, allocation::empty_device_dtype, element::TensorElement, layout::address_type};
use ruda_core::tensor::{DType, Shape};
use crate::collective::{RudaSum, RudaSumLaunch, radix::RudaRadixKey};
use super::{RudaPrimitiveError, check_type, scan_threads, scan, reduce};

#[ruda(launch_unchecked, explicit_define, address_type = "dynamic")]
pub(super) fn initialise(
    candidates: &mut LinearView<u64, ReadWrite>, selected: &mut LinearView<u64, ReadWrite>,
    state: &mut LinearView<u64, ReadWrite>, k: usize,
) {
    if ABSOLUTE_POS < candidates.shape() { candidates[ABSOLUTE_POS] = 1; selected[ABSOLUTE_POS] = 0; }
    if ABSOLUTE_POS == 0 { state[0] = k as u64; state[1] = 0; }
}

#[ruda(launch_unchecked, explicit_define, address_type = "dynamic")]
fn bit_flags<K: RudaRadixKey>(
    keys: &LinearView<K>, candidates: &LinearView<u64>, flags: &mut LinearView<u64, ReadWrite>,
    #[comptime] bit: u32, #[comptime] largest: bool,
) {
    let index = ABSOLUTE_POS;
    if index < keys.shape() {
        let mut flag = 0u64;
        if candidates[index] != 0 {
            let one = ((K::ordered_bits(keys[index]) >> bit) & 1u64) != 0;
            flag = u64::cast_from(one == largest);
        }
        flags[index] = flag;
    }
}

#[ruda(launch_unchecked, explicit_define, address_type = "dynamic")]
pub(super) fn decision(total: &LinearView<u64>, state: &mut LinearView<u64, ReadWrite>) {
    if ABSOLUTE_POS == 0 {
        let accept = total[0] <= state[0];
        state[1] = u64::cast_from(accept);
        if accept { state[0] = state[0] - total[0]; }
    }
}

#[ruda(launch_unchecked, explicit_define, address_type = "dynamic")]
pub(super) fn refine(
    flags: &LinearView<u64>, state: &LinearView<u64>,
    candidates: &mut LinearView<u64, ReadWrite>, selected: &mut LinearView<u64, ReadWrite>,
) {
    let index = ABSOLUTE_POS;
    if index < candidates.shape() {
        if candidates[index] != 0 {
            if state[1] != 0 {
                if flags[index] != 0 { selected[index] = 1; candidates[index] = 0; }
            } else {
                candidates[index] = flags[index];
            }
        }
    }
}

#[ruda(launch_unchecked, explicit_define, address_type = "dynamic")]
pub(super) fn resolve_ties(
    candidates: &LinearView<u64>, prefixes: &LinearView<u64>, state: &LinearView<u64>,
    selected: &mut LinearView<u64, ReadWrite>,
) {
    let index = ABSOLUTE_POS;
    if index < selected.shape() {
        if candidates[index] != 0 && prefixes[index] <= state[0] { selected[index] = 1; }
    }
}

#[ruda(launch_unchecked, explicit_define, address_type = "dynamic")]
fn pack_keys<K: Numeric>(
    keys: &LinearView<K>, selected: &LinearView<u64>, prefixes: &LinearView<u64>, output: &mut LinearView<K, ReadWrite>,
) {
    let index = ABSOLUTE_POS;
    if index < keys.shape() {
        if selected[index] != 0 { let rank = prefixes[index] as usize - 1; output[rank] = keys[index]; }
    }
}

#[ruda(launch_unchecked, explicit_define, address_type = "dynamic")]
fn pack_pairs<K: Numeric, V: Numeric>(
    keys: &LinearView<K>, values: &LinearView<V>, selected: &LinearView<u64>, prefixes: &LinearView<u64>,
    output: &mut LinearView<K, ReadWrite>, output_values: &mut LinearView<V, ReadWrite>,
) {
    let index = ABSOLUTE_POS;
    if index < keys.shape() {
        if selected[index] != 0 {
            let rank = prefixes[index] as usize - 1;
            output[rank] = keys[index];
            output_values[rank] = values[index];
        }
    }
}

fn selected_prefixes<R: Runtime, K: TensorElement + RudaRadixKey>(
    keys: &RudaTensor<R>, k: usize, largest: bool, threads: u32,
) -> Result<(RudaTensor<R>, RudaTensor<R>), RudaPrimitiveError> {
    let count = keys.meta.num_elements();
    let allocate = |len| empty_device_dtype(keys.client.clone(), keys.device.clone(), Shape::new([len]), DType::U64);
    let candidates = allocate(count);
    let selected = allocate(count);
    let flags = allocate(count);
    let state = allocate(2);
    let dim = RudaDim::new(keys.client.properties(), count);
    let grid = calculate_ruda_count_elemwise(&keys.client, count, dim);
    unsafe {
        initialise::launch_unchecked::<R>(&keys.client, grid, dim, address_type!(candidates, selected, state),
            candidates.clone().into_linear_view(), selected.clone().into_linear_view(), state.clone().into_linear_view(), k);
    }
    for bit in (0..core::mem::size_of::<K>() as u32 * 8).rev() {
        let grid = calculate_ruda_count_elemwise(&keys.client, count, dim);
        unsafe {
            bit_flags::launch_unchecked::<K, R>(&keys.client, grid, dim, address_type!(keys, candidates, flags),
                keys.clone().into_linear_view(), candidates.clone().into_linear_view(), flags.clone().into_linear_view(), bit, largest);
        }
        let total = reduce::reduce::<R, u64, RudaSum>(&flags, 0, RudaSumLaunch::new(), threads)?;
        unsafe {
            decision::launch_unchecked::<R>(&keys.client, RudaCount::Static(1, 1, 1), RudaDim::new_1d(1), address_type!(total, state),
                total.into_linear_view(), state.clone().into_linear_view());
            let grid = calculate_ruda_count_elemwise(&keys.client, count, dim);
            refine::launch_unchecked::<R>(&keys.client, grid, dim, address_type!(flags, state, candidates, selected),
                flags.clone().into_linear_view(), state.clone().into_linear_view(), candidates.clone().into_linear_view(), selected.clone().into_linear_view());
        }
    }
    let prefixes = scan::inclusive_scan::<R, u64, RudaSum>(&candidates, RudaSumLaunch::new(), threads)?;
    let grid = calculate_ruda_count_elemwise(&keys.client, count, dim);
    unsafe {
        resolve_ties::launch_unchecked::<R>(&keys.client, grid, dim, address_type!(candidates, prefixes, state, selected),
            candidates.into_linear_view(), prefixes.into_linear_view(), state.into_linear_view(), selected.clone().into_linear_view());
    }
    let prefixes = scan::inclusive_scan::<R, u64, RudaSum>(&selected, RudaSumLaunch::new(), threads)?;
    Ok((selected, prefixes))
}

/// Select the largest or smallest K keys by radix refinement. K is capped at
/// input length. Results are not sorted; selected items and cutoff ties retain
/// input order. Selection state and counts remain on the device.
pub fn keys<R: Runtime, K: TensorElement + RudaRadixKey>(
    input: &RudaTensor<R>, k: usize, largest: bool, threads: u32,
) -> Result<RudaTensor<R>, RudaPrimitiveError> {
    check_type::<R, K>(input)?;
    scan_threads::<R, u64>(input, threads)?;
    let count = input.meta.num_elements();
    let k = k.min(count);
    let output = empty_device_dtype(input.client.clone(), input.device.clone(), Shape::new([k]), input.dtype);
    if k == 0 { return Ok(output); }
    let (selected, prefixes) = selected_prefixes::<R, K>(input, k, largest, threads)?;
    let dim = RudaDim::new(input.client.properties(), count);
    let grid = calculate_ruda_count_elemwise(&input.client, count, dim);
    unsafe {
        pack_keys::launch_unchecked::<K, R>(&input.client, grid, dim, address_type!(input, selected, prefixes, output),
            input.clone().into_linear_view(), selected.into_linear_view(), prefixes.into_linear_view(), output.clone().into_linear_view());
    }
    Ok(output)
}

pub fn pairs<R: Runtime, K: TensorElement + RudaRadixKey, V: TensorElement>(
    input: &RudaTensor<R>, values: &RudaTensor<R>, k: usize, largest: bool, threads: u32,
) -> Result<(RudaTensor<R>, RudaTensor<R>), RudaPrimitiveError> {
    check_type::<R, K>(input)?;
    check_type::<R, V>(values)?;
    scan_threads::<R, u64>(input, threads)?;
    let count = input.meta.num_elements();
    if count != values.meta.num_elements() { return Err(RudaPrimitiveError::Length); }
    if input.device.to_id() != values.device.to_id() { return Err(RudaPrimitiveError::Device); }
    let k = k.min(count);
    let output = empty_device_dtype(input.client.clone(), input.device.clone(), Shape::new([k]), input.dtype);
    let output_values = empty_device_dtype(input.client.clone(), input.device.clone(), Shape::new([k]), values.dtype);
    if k == 0 { return Ok((output, output_values)); }
    let (selected, prefixes) = selected_prefixes::<R, K>(input, k, largest, threads)?;
    let dim = RudaDim::new(input.client.properties(), count);
    let grid = calculate_ruda_count_elemwise(&input.client, count, dim);
    unsafe {
        pack_pairs::launch_unchecked::<K, V, R>(
            &input.client, grid, dim, address_type!(input, values, selected, prefixes, output, output_values),
            input.clone().into_linear_view(), values.clone().into_linear_view(), selected.into_linear_view(), prefixes.into_linear_view(),
            output.clone().into_linear_view(), output_values.clone().into_linear_view(),
        );
    }
    Ok((output, output_values))
}