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

use ruda_core::tensor::{DType, TensorMetadata, element::Scalar as TensorScalar};
use ruda_kernel::dsl as kernel_dsl;
use ruda_kernel::dsl::{calculate_ruda_count_elemwise, prelude::*};
use ruda_kernel::library::tensor::layout::linear::LinearView;
use ruda_kernel::tensor::{
    RudaTensor,
    allocation::empty_device_dtype,
    layout::{address_type, broadcast_shape, max_vector_size},
};

#[ruda(launch_unchecked, address_type = "dynamic")]
fn tensor_kernel<F: Float, I: Int, N: Size>(
    lhs: &LinearView<Vector<F, N>>,
    rhs: &LinearView<Vector<I, N>>,
    out: &mut LinearView<Vector<F, N>, ReadWrite>,
    #[define(F, I)] _dtypes: [StorageType; 2],
) where F: Powi<I> {
    if !out.is_in_bounds(ABSOLUTE_POS) {
        terminate!();
    }
    out[ABSOLUTE_POS] = Vector::powi(lhs[ABSOLUTE_POS], rhs[ABSOLUTE_POS]);
}

#[ruda(launch_unchecked, address_type = "dynamic")]
fn scalar_kernel<F: Float, I: Int, N: Size>(
    input: &LinearView<Vector<F, N>>,
    exponent: InputScalar,
    out: &mut LinearView<Vector<F, N>, ReadWrite>,
    #[define(F, I)] _dtypes: [StorageType; 2],
) where F: Powi<I> {
    if !out.is_in_bounds(ABSOLUTE_POS) {
        terminate!();
    }
    out[ABSOLUTE_POS] = Vector::powi(input[ABSOLUTE_POS], Vector::new(exponent.get::<I>()));
}

pub fn tensor<R: Runtime>(lhs: RudaTensor<R>, rhs: RudaTensor<R>) -> RudaTensor<R> {
    let vector_size = max_vector_size(&lhs).min(max_vector_size(&rhs));
    let shape = broadcast_shape(&[&lhs, &rhs]);
    let dtype = lhs.dtype;
    let dtypes = [dtype.into(), rhs.dtype.into()];
    let client = lhs.client.clone();
    let working_units = shape.num_elements() / vector_size as usize;
    let ruda_dim = RudaDim::new(client.properties(), working_units);
    let ruda_count = calculate_ruda_count_elemwise(&client, working_units, ruda_dim);

    unsafe {
        if lhs.can_mut_broadcast(&rhs) {
            tensor_kernel::launch_unchecked::<R>(
                &client,
                ruda_count,
                ruda_dim,
                address_type!(lhs, rhs),
                vector_size,
                lhs.clone().into_linear_view(),
                rhs.into_linear_view_like(&lhs),
                lhs.as_linear_view_alias(0),
                dtypes,
            );
            lhs
        } else {
            let output = empty_device_dtype(client.clone(), lhs.device.clone(), shape, dtype);
            tensor_kernel::launch_unchecked::<R>(
                &client,
                ruda_count,
                ruda_dim,
                address_type!(lhs, rhs, output),
                vector_size,
                lhs.into_linear_view_like(&output),
                rhs.into_linear_view_like(&output),
                output.clone().into_linear_view(),
                dtypes,
            );
            output
        }
    }
}

pub fn scalar<R: Runtime>(input: RudaTensor<R>, exponent: TensorScalar) -> RudaTensor<R> {
    let (exponent, exponent_dtype) = match exponent {
        TensorScalar::UInt(value) => {
            let dtype = if u32::try_from(value).is_ok() { DType::U32 } else { DType::U64 };
            (TensorScalar::UInt(value), dtype)
        }
        value => {
            let value = value.elem::<i64>();
            let dtype = if i32::try_from(value).is_ok() { DType::I32 } else { DType::I64 };
            (TensorScalar::Int(value), dtype)
        }
    };
    let exponent = InputScalar::new(exponent, exponent_dtype);
    let vector_size = max_vector_size(&input);
    let client = input.client.clone();
    let working_units = input.meta.num_elements() / vector_size as usize;
    let ruda_dim = RudaDim::new(client.properties(), working_units);
    let ruda_count = calculate_ruda_count_elemwise(&client, working_units, ruda_dim);
    let dtypes = [input.dtype.into(), exponent_dtype.into()];

    unsafe {
        if input.can_mut() && input.is_nonoverlapping() {
            scalar_kernel::launch_unchecked::<R>(
                &client,
                ruda_count,
                ruda_dim,
                address_type!(input),
                vector_size,
                input.clone().into_linear_view(),
                exponent,
                input.as_linear_view_alias(0),
                dtypes,
            );
            input
        } else {
            let output = empty_device_dtype(
                client.clone(), input.device.clone(), input.shape(), input.dtype,
            );
            scalar_kernel::launch_unchecked::<R>(
                &client,
                ruda_count,
                ruda_dim,
                address_type!(input, output),
                vector_size,
                input.into_linear_view(),
                exponent,
                output.clone().into_linear_view(),
                dtypes,
            );
            output
        }
    }
}