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

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::{Shape, TensorMetadata};
use crate::collective::{RudaBinaryOp, RudaBinaryOpExpand};
use super::{RudaPrimitiveError, check_type, empty_like, scan_threads};

/// Caller-owned output, including exact in-place scans. Inclusive scans may
/// optionally be seeded; exclusive scans always require an initial value.
pub fn scan_into<R, T, O>(input: &RudaTensor<R>, output: &RudaTensor<R>, initial: Option<T>,
    exclusive: bool, op: O::RuntimeArg<R>, threads: u32) -> Result<(), RudaPrimitiveError>
where R: Runtime, T: TensorElement, O: RudaBinaryOp<T> + LaunchArg, O::RuntimeArg<R>: Clone,
{
    check_type::<R, T>(output)?;
    if input.meta.num_elements() != output.meta.num_elements() { return Err(RudaPrimitiveError::Length); }
    if input.device.to_id() != output.device.to_id() { return Err(RudaPrimitiveError::Device); }
    let scanned = match (exclusive, initial) {
        (true, None) => return Err(RudaPrimitiveError::Configuration("exclusive scan requires an initial value")),
        (true, Some(initial)) => exclusive_scan::<R, T, O>(input, initial, op, threads)?,
        (false, Some(initial)) => inclusive_scan_init::<R, T, O>(input, initial, op, threads)?,
        (false, None) => inclusive_scan::<R, T, O>(input, op, threads)?,
    };
    super::iteration::copy_into::<R, T>(&scanned, output)
}

#[ruda(launch_unchecked, explicit_define, address_type = "dynamic")]
fn tile_scan<T: Numeric, O: RudaBinaryOp<T> + LaunchArg>(
    input: &LinearView<T>,
    output: &mut LinearView<T, ReadWrite>,
    totals: &mut LinearView<T, ReadWrite>,
    op: &O,
    #[comptime] threads: usize,
) {
    let start = RUDA_POS as usize * threads;
    if start >= input.shape() { terminate!(); }
    let valid = min(threads, input.shape() - start);
    let index = start + UNIT_POS as usize;
    let mut local = Array::<T>::new(1usize);
    let mut result = Array::<T>::new(1usize);
    let mut scratch = SharedMemory::<T>::new(threads);
    let mut value = T::from_int(0);
    if index < input.shape() { value = input[index]; }
    local[0] = value;
    crate::block::inclusive_scan::<T, O>(&local, &mut result, &mut scratch, op, valid, threads, 1usize);
    if index < output.shape() { output[index] = result[0]; }
    if UNIT_POS == 0 { totals[RUDA_POS as usize] = scratch[valid - 1]; }
}

#[ruda(launch_unchecked, explicit_define, address_type = "dynamic")]
fn apply_prefix<T: Numeric, O: RudaBinaryOp<T> + LaunchArg>(
    output: &mut LinearView<T, ReadWrite>,
    totals: &LinearView<T>,
    op: &O,
    #[comptime] threads: usize,
) {
    let index = ABSOLUTE_POS;
    if index >= output.shape() { terminate!(); }
    let tile = index / threads;
    if tile > 0 { output[index] = op.combine(totals[tile - 1], output[index]); }
}

#[ruda(launch_unchecked, explicit_define, address_type = "dynamic")]
fn make_exclusive<T: Numeric, O: RudaBinaryOp<T> + LaunchArg>(
    inclusive: &LinearView<T>,
    output: &mut LinearView<T, ReadWrite>,
    initial: InputScalar,
    op: &O,
) {
    let index = ABSOLUTE_POS;
    if index >= output.shape() { terminate!(); }
    let mut value = initial.get::<T>();
    if index > 0 { value = op.combine(value, inclusive[index - 1]); }
    output[index] = value;
}

#[ruda(launch_unchecked, explicit_define, address_type = "dynamic")]
fn seed_inclusive<T: Numeric, O: RudaBinaryOp<T> + LaunchArg>(
    output: &mut LinearView<T, ReadWrite>, initial: InputScalar, op: &O,
) {
    if ABSOLUTE_POS < output.shape() { output[ABSOLUTE_POS] = op.combine(initial.get::<T>(), output[ABSOLUTE_POS]); }
}

pub fn inclusive_scan_init<R, T, O>(
    input: &RudaTensor<R>, initial: T, op: O::RuntimeArg<R>, threads: u32,
) -> Result<RudaTensor<R>, RudaPrimitiveError>
where R: Runtime, T: TensorElement, O: RudaBinaryOp<T> + LaunchArg, O::RuntimeArg<R>: Clone,
{
    let output = inclusive_scan::<R, T, O>(input, op.clone(), threads)?;
    let count = input.meta.num_elements();
    if count > 0 {
        let dim = RudaDim::new(input.client.properties(), count);
        let grid = calculate_ruda_count_elemwise(&input.client, count, dim);
        unsafe {
            seed_inclusive::launch_unchecked::<T, O, R>(&input.client, grid, dim, address_type!(output),
                output.clone().into_linear_view(), InputScalar::new(initial, input.dtype), op);
        }
    }
    Ok(output)
}

/// Inclusive scan of the tensor's logical flattened order. Each hierarchy
/// level is a separate kernel dispatch, requiring no inter-block spin waits.
/// `threads` selects the tile width and must meet the device's block limits.
pub fn inclusive_scan<R, T, O>(
    input: &RudaTensor<R>,
    op: O::RuntimeArg<R>,
    threads: u32,
) -> Result<RudaTensor<R>, RudaPrimitiveError>
where
    R: Runtime,
    T: TensorElement,
    O: RudaBinaryOp<T> + LaunchArg,
    O::RuntimeArg<R>: Clone,
{
    check_type::<R, T>(input)?;
    scan_threads::<R, T>(input, threads)?;
    let output = empty_like(input);
    let count = input.meta.num_elements();
    if count == 0 { return Ok(output); }
    let tile_count = count.div_ceil(threads as usize);
    let totals = empty_device_dtype(input.client.clone(), input.device.clone(), Shape::new([tile_count]), input.dtype);
    let dim = RudaDim::new_1d(threads);
    let grid = calculate_ruda_count_elemwise(&input.client, count, dim);
    unsafe {
        tile_scan::launch_unchecked::<T, O, R>(
            &input.client, grid, dim, address_type!(input, output, totals),
            input.clone().into_linear_view(), output.clone().into_linear_view(),
            totals.clone().into_linear_view(), op.clone(), threads as usize,
        );
    }
    if tile_count > 1 {
        let prefixes = inclusive_scan::<R, T, O>(&totals, op.clone(), threads)?;
        let grid = calculate_ruda_count_elemwise(&input.client, count, dim);
        unsafe {
            apply_prefix::launch_unchecked::<T, O, R>(
                &input.client, grid, dim, address_type!(output, prefixes),
                output.clone().into_linear_view(), prefixes.into_linear_view(), op,
                threads as usize,
            );
        }
    }
    Ok(output)
}

/// Exclusive scan seeded by `initial`, without mutating input or reading
/// intermediate results back to the host.
pub fn exclusive_scan<R, T, O>(
    input: &RudaTensor<R>,
    initial: T,
    op: O::RuntimeArg<R>,
    threads: u32,
) -> Result<RudaTensor<R>, RudaPrimitiveError>
where
    R: Runtime,
    T: TensorElement,
    O: RudaBinaryOp<T> + LaunchArg,
    O::RuntimeArg<R>: Clone,
{
    let inclusive = inclusive_scan::<R, T, O>(input, op.clone(), threads)?;
    let output = empty_like(input);
    let count = input.meta.num_elements();
    if count == 0 { return Ok(output); }
    let dim = RudaDim::new_1d(threads);
    let grid = calculate_ruda_count_elemwise(&input.client, count, dim);
    unsafe {
        make_exclusive::launch_unchecked::<T, O, R>(
            &input.client, grid, dim, address_type!(inclusive, output),
            inclusive.into_linear_view(), output.clone().into_linear_view(),
            InputScalar::new(initial, input.dtype), op,
        );
    }
    Ok(output)
}