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

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::{TensorMetadata, element::Element};
use super::{RudaPrimitiveError, check_type};
use super::select::{RudaPredicate, RudaPredicateExpand};

pub use crate::collective::{RudaUnaryOp, RudaUnaryOpExpand};

#[ruda]
pub trait RudaBinaryTransform<T: RudaType, U: RudaType, V: RudaType>: RudaType {
    fn apply(&self, left: T, right: U) -> V;
}

#[ruda]
pub trait RudaGenerator<T: RudaType>: RudaType {
    fn generate(&mut self) -> T;
}

#[ruda(launch_unchecked, explicit_define, address_type = "dynamic")]
fn fill_kernel<T: Numeric>(output: &mut LinearView<T, ReadWrite>, value: InputScalar) {
    if ABSOLUTE_POS < output.shape() { output[ABSOLUTE_POS] = value.get::<T>(); }
}

#[ruda(launch_unchecked, explicit_define, address_type = "dynamic")]
fn generate_kernel<T: Numeric, G: RudaGenerator<T> + LaunchArg>(output: &mut LinearView<T, ReadWrite>, generator: &mut G) {
    if ABSOLUTE_POS < output.shape() { output[ABSOLUTE_POS] = generator.generate(); }
}

#[ruda(launch_unchecked, explicit_define, address_type = "dynamic")]
fn unary_if_kernel<T: Numeric, U: Numeric, O: RudaUnaryOp<T, U> + LaunchArg, P: RudaPredicate<T> + LaunchArg>(
    input: &LinearView<T>, output: &mut LinearView<U, ReadWrite>, op: &O, predicate: &P,
) {
    let index = ABSOLUTE_POS;
    if index < input.shape() {
        let value = input[index];
        if predicate.test(value) { output[index] = op.apply(value); }
    }
}

pub fn fill<R: Runtime, T: TensorElement>(output: &RudaTensor<R>, value: T) -> Result<(), RudaPrimitiveError> {
    check_type::<R, T>(output)?;
    let count = output.meta.num_elements();
    if count > 0 {
        let dim = RudaDim::new(output.client.properties(), count);
        let grid = calculate_ruda_count_elemwise(&output.client, count, dim);
        unsafe {
            fill_kernel::launch_unchecked::<T, R>(&output.client, grid, dim, address_type!(output),
                output.clone().into_linear_view(), InputScalar::new(value, output.dtype));
        }
    }
    Ok(())
}

pub fn generate<R, T, G>(output: &RudaTensor<R>, generator: G::RuntimeArg<R>) -> Result<(), RudaPrimitiveError>
where R: Runtime, T: TensorElement, G: RudaGenerator<T> + LaunchArg,
{
    check_type::<R, T>(output)?;
    let count = output.meta.num_elements();
    if count > 0 {
        let dim = RudaDim::new(output.client.properties(), count);
        let grid = calculate_ruda_count_elemwise(&output.client, count, dim);
        unsafe {
            generate_kernel::launch_unchecked::<T, G, R>(&output.client, grid, dim, address_type!(output),
                output.clone().into_linear_view(), generator);
        }
    }
    Ok(())
}

/// Transform selected input elements, leaving other output elements unchanged.
/// Input/output may alias exactly for identical element types, but must not
/// partially overlap.
pub fn unary_if_into<R, T, U, O, P>(
    input: &RudaTensor<R>, output: &RudaTensor<R>, op: O::RuntimeArg<R>, predicate: P::RuntimeArg<R>,
) -> Result<(), RudaPrimitiveError>
where R: Runtime, T: TensorElement, U: TensorElement,
    O: RudaUnaryOp<T, U> + LaunchArg, P: RudaPredicate<T> + LaunchArg,
{
    check_type::<R, T>(input)?;
    check_type::<R, U>(output)?;
    let count = input.meta.num_elements();
    if count != output.meta.num_elements() { return Err(RudaPrimitiveError::Length); }
    if input.device.to_id() != output.device.to_id() { return Err(RudaPrimitiveError::Device); }
    if count > 0 {
        let dim = RudaDim::new(input.client.properties(), count);
        let grid = calculate_ruda_count_elemwise(&input.client, count, dim);
        unsafe {
            unary_if_kernel::launch_unchecked::<T, U, O, P, R>(&input.client, grid, dim, address_type!(input, output),
                input.clone().into_linear_view(), output.clone().into_linear_view(), op, predicate);
        }
    }
    Ok(())
}

#[derive(Clone, Copy, Debug, RudaType, RudaLaunch)]
pub struct RudaCast;

impl<R: Runtime> Clone for RudaCastLaunch<R> {
    fn clone(&self) -> Self { Self::new() }
}

#[ruda]
impl<T: Numeric, U: Numeric> RudaUnaryOp<T, U> for RudaCast {
    fn apply(&self, value: T) -> U { U::cast_from(value) }
}

#[ruda(launch_unchecked, explicit_define, address_type = "dynamic")]
fn unary_kernel<T: Numeric, U: Numeric, O: RudaUnaryOp<T, U> + LaunchArg>(
    input: &LinearView<T>, output: &mut LinearView<U, ReadWrite>, op: &O,
) {
    let index = ABSOLUTE_POS;
    if index < output.shape() { output[index] = op.apply(input[index]); }
}

#[ruda(launch_unchecked, explicit_define, address_type = "dynamic")]
fn binary_kernel<T: Numeric, U: Numeric, V: Numeric, O: RudaBinaryTransform<T, U, V> + LaunchArg>(
    left: &LinearView<T>, right: &LinearView<U>, output: &mut LinearView<V, ReadWrite>, op: &O,
) {
    let index = ABSOLUTE_POS;
    if index < output.shape() { output[index] = op.apply(left[index], right[index]); }
}

pub fn unary<R, T, U, O>(input: &RudaTensor<R>, op: O::RuntimeArg<R>) -> Result<RudaTensor<R>, RudaPrimitiveError>
where R: Runtime, T: TensorElement, U: TensorElement, O: RudaUnaryOp<T, U> + LaunchArg,
{
    check_type::<R, T>(input)?;
    let output = empty_device_dtype(input.client.clone(), input.device.clone(), input.shape(), <U as Element>::dtype());
    let count = input.meta.num_elements();
    if count == 0 { return Ok(output); }
    let dim = RudaDim::new(input.client.properties(), count);
    let grid = calculate_ruda_count_elemwise(&input.client, count, dim);
    unsafe {
        unary_kernel::launch_unchecked::<T, U, O, R>(
            &input.client, grid, dim, address_type!(input, output),
            input.clone().into_linear_view(), output.clone().into_linear_view(), op,
        );
    }
    Ok(output)
}

pub fn binary<R, T, U, V, O>(left: &RudaTensor<R>, right: &RudaTensor<R>, op: O::RuntimeArg<R>) -> Result<RudaTensor<R>, RudaPrimitiveError>
where R: Runtime, T: TensorElement, U: TensorElement, V: TensorElement, O: RudaBinaryTransform<T, U, V> + LaunchArg,
{
    check_type::<R, T>(left)?;
    check_type::<R, U>(right)?;
    let count = left.meta.num_elements();
    if count != right.meta.num_elements() { return Err(RudaPrimitiveError::Length); }
    if left.device.to_id() != right.device.to_id() { return Err(RudaPrimitiveError::Device); }
    let output = empty_device_dtype(left.client.clone(), left.device.clone(), left.shape(), <V as Element>::dtype());
    if count == 0 { return Ok(output); }
    let dim = RudaDim::new(left.client.properties(), count);
    let grid = calculate_ruda_count_elemwise(&left.client, count, dim);
    unsafe {
        binary_kernel::launch_unchecked::<T, U, V, O, R>(
            &left.client, grid, dim, address_type!(left, right, output),
            left.clone().into_linear_view(), right.clone().into_linear_view(), output.clone().into_linear_view(), op,
        );
    }
    Ok(output)
}