ruPRIM 0.1.1

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

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::{RudaKeyEqual, RudaSum, RudaSumLaunch};
use super::{RudaPrimitiveError, empty_like, segmented, scan, transform::{self, RudaCast, RudaCastLaunch}};

/// Device-resident run records. Only the prefix specified by `count` is valid.
/// Offsets, lengths and the single run count are U64 tensors.
pub struct RudaRuns<R: Runtime> {
    pub values: RudaTensor<R>,
    pub offsets: RudaTensor<R>,
    pub lengths: RudaTensor<R>,
    pub count: RudaTensor<R>,
}

#[ruda(launch_unchecked, explicit_define, address_type = "dynamic")]
fn scatter_heads<T: Numeric>(
    input: &LinearView<T>, heads: &LinearView<u32>, prefixes: &LinearView<u64>,
    values: &mut LinearView<T, ReadWrite>, offsets: &mut LinearView<u64, ReadWrite>, count: &mut LinearView<u64, ReadWrite>,
) {
    let index = ABSOLUTE_POS;
    if index == 0 {
        let mut total = 0u64;
        if input.shape() > 0 { total = prefixes[input.shape() - 1]; }
        count[0] = total;
    }
    if index < input.shape() {
        if heads[index] != 0 {
            let rank = prefixes[index] as usize - 1;
            values[rank] = input[index];
            offsets[rank] = index as u64;
        }
    }
}

#[ruda(launch_unchecked, explicit_define, address_type = "dynamic")]
pub(super) fn lengths_kernel(offsets: &LinearView<u64>, count: &LinearView<u64>, lengths: &mut LinearView<u64, ReadWrite>, input_length: usize) {
    let run = ABSOLUTE_POS;
    let runs = count[0] as usize;
    if run < runs {
        let mut end = input_length as u64;
        if run + 1 < runs { end = offsets[run + 1]; }
        lengths[run] = end - offsets[run];
    }
}

#[ruda(launch_unchecked, explicit_define, address_type = "dynamic")]
pub(super) fn nontrivial_flags(lengths: &LinearView<u64>, count: &LinearView<u64>, flags: &mut LinearView<u64, ReadWrite>) {
    let index = ABSOLUTE_POS;
    if index < flags.shape() {
        let mut flag = 0u64;
        if index < count[0] as usize { flag = u64::cast_from(lengths[index] > 1); }
        flags[index] = flag;
    }
}

#[ruda(launch_unchecked, explicit_define, address_type = "dynamic")]
fn compact_runs<T: Numeric>(
    values: &LinearView<T>, offsets: &LinearView<u64>, lengths: &LinearView<u64>,
    flags: &LinearView<u64>, prefixes: &LinearView<u64>,
    output_values: &mut LinearView<T, ReadWrite>, output_offsets: &mut LinearView<u64, ReadWrite>,
    output_lengths: &mut LinearView<u64, ReadWrite>, output_count: &mut LinearView<u64, ReadWrite>,
) {
    let index = ABSOLUTE_POS;
    if index == 0 {
        let mut total = 0u64;
        if flags.shape() > 0 { total = prefixes[flags.shape() - 1]; }
        output_count[0] = total;
    }
    if index < flags.shape() {
        if flags[index] != 0 {
            let rank = prefixes[index] as usize - 1;
            output_values[rank] = values[index];
            output_offsets[rank] = offsets[index];
            output_lengths[rank] = lengths[index];
        }
    }
}

fn allocate<R: Runtime>(input: &RudaTensor<R>) -> RudaRuns<R> {
    let size = input.meta.num_elements();
    let buffer = |count| empty_device_dtype(input.client.clone(), input.device.clone(), Shape::new([count]), DType::U64);
    RudaRuns { values: empty_like(input), offsets: buffer(size), lengths: buffer(size), count: buffer(1) }
}

/// Encode adjacent equal values. Output values are the first input value of
/// each run, including when a custom equality relation groups unequal bit patterns.
pub fn encode<R, T, E>(input: &RudaTensor<R>, equal: E::RuntimeArg<R>, threads: u32) -> Result<RudaRuns<R>, RudaPrimitiveError>
where R: Runtime, T: TensorElement, E: RudaKeyEqual<T> + LaunchArg,
{
    let heads = segmented::key_heads::<R, T, E>(input, equal)?;
    let flags = transform::unary::<R, u32, u64, RudaCast>(&heads, RudaCastLaunch::new())?;
    let prefixes = scan::inclusive_scan::<R, u64, RudaSum>(&flags, RudaSumLaunch::new(), threads)?;
    let output = allocate(input);
    let work = input.meta.num_elements().max(1);
    let dim = RudaDim::new(input.client.properties(), work);
    let grid = calculate_ruda_count_elemwise(&input.client, work, dim);
    unsafe {
        scatter_heads::launch_unchecked::<T, R>(
            &input.client, grid, dim, address_type!(input, heads, prefixes, (output.values), (output.offsets), (output.count)),
            input.clone().into_linear_view(), heads.into_linear_view(), prefixes.into_linear_view(),
            output.values.clone().into_linear_view(), output.offsets.clone().into_linear_view(), output.count.clone().into_linear_view(),
        );
        let grid = calculate_ruda_count_elemwise(&input.client, work, dim);
        lengths_kernel::launch_unchecked::<R>(
            &input.client, grid, dim, address_type!((output.offsets), (output.count), (output.lengths)),
            output.offsets.clone().into_linear_view(), output.count.clone().into_linear_view(),
            output.lengths.clone().into_linear_view(), input.meta.num_elements(),
        );
    }
    Ok(output)
}

/// Encode only runs longer than one element, preserving their input order.
pub fn non_trivial_runs<R, T, E>(input: &RudaTensor<R>, equal: E::RuntimeArg<R>, threads: u32) -> Result<RudaRuns<R>, RudaPrimitiveError>
where R: Runtime, T: TensorElement, E: RudaKeyEqual<T> + LaunchArg,
{
    let runs = encode::<R, T, E>(input, equal, threads)?;
    let size = input.meta.num_elements();
    let flags = empty_like(&runs.lengths);
    let dim = RudaDim::new(input.client.properties(), size.max(1));
    if size > 0 {
        let grid = calculate_ruda_count_elemwise(&input.client, size, dim);
        unsafe {
            nontrivial_flags::launch_unchecked::<R>(
                &input.client, grid, dim, address_type!((runs.lengths), (runs.count), flags),
                runs.lengths.clone().into_linear_view(), runs.count.clone().into_linear_view(), flags.clone().into_linear_view(),
            );
        }
    }
    let prefixes = scan::inclusive_scan::<R, u64, RudaSum>(&flags, RudaSumLaunch::new(), threads)?;
    let output = allocate(input);
    let grid = calculate_ruda_count_elemwise(&input.client, size.max(1), dim);
    unsafe {
        compact_runs::launch_unchecked::<T, R>(
            &input.client, grid, dim, address_type!((runs.values), (runs.offsets), (runs.lengths), flags, prefixes, (output.values), (output.offsets), (output.lengths), (output.count)),
            runs.values.into_linear_view(), runs.offsets.into_linear_view(), runs.lengths.into_linear_view(),
            flags.into_linear_view(), prefixes.into_linear_view(), output.values.clone().into_linear_view(),
            output.offsets.clone().into_linear_view(), output.lengths.clone().into_linear_view(), output.count.clone().into_linear_view(),
        );
    }
    Ok(output)
}