ruPRIM 0.1.5

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

use ruda_kernel::dsl as kernel_dsl;
use ruda_kernel::dsl::Runtime;
use crate::elementwise::binary::numeric::AddOp;
use crate::elementwise::binary::numeric::AssignOp;
use crate::elementwise::binary::numeric::BinaryOp;
use crate::elementwise::binary::numeric::BinaryOpFamily;
use crate::elementwise::binary::numeric::OrOp;
use ruda_kernel::tensor::layout::address_type;
use ruda_kernel::tensor::layout::shape_divmod;
use ruda_kernel::tensor::RudaTensor;
use ruda_kernel::dsl::RudaDim;
use ruda_kernel::dsl::calculate_ruda_count_elemwise;
use ruda_kernel::dsl::prelude::*;
use ruda_kernel::library::FastDivmod;

#[ruda(launch_unchecked, address_type = "dynamic")]
fn scatter_kernel<T: Numeric, I: Int, Op: BinaryOpFamily>(
    input: &mut Tensor<T>,
    indices: &Tensor<I>,
    value: &Tensor<T>,
    in_shape: Sequence<FastDivmod<usize>>,
    #[comptime] dim: usize,
    #[define(T, I)] _dtypes: [StorageType; 2],
) {
    let rank = in_shape.len().comptime();
    let stride_input = input.stride(dim);
    let stride_value = value.stride(dim);
    let stride_indices = indices.stride(dim);
    let shape_value = value.shape(dim);

    let mut offset = ABSOLUTE_POS;
    let mut offset_input = 0;
    let mut offset_indices = 0;
    let mut offset_value = 0;
    let mut num_elems = 1;

    #[unroll]
    for i in 0..rank {
        let i = rank - i - 1;
        if i != dim {
            let shape_input_loop = input.shape(i);

            let (rem, local_pos) = in_shape[i].div_mod(offset);
            offset = rem;

            offset_input += local_pos * input.stride(i);
            offset_indices += local_pos * indices.stride(i);
            offset_value += local_pos * value.stride(i);

            num_elems *= shape_input_loop;
        }
    }

    let should_stop = ABSOLUTE_POS >= num_elems;
    if should_stop {
        terminate!();
    }

    for i in 0..shape_value {
        let value_idx = (stride_value * i) + offset_value;
        let index_idx = (stride_indices * i) + offset_indices;

        let value = value[value_idx];
        let index = usize::cast_from(indices[index_idx]);

        let input_idx = (stride_input * index) + offset_input;

        let value = Op::BinaryOp::<T, Const<1>>::execute(
            Vector::cast_from(input[input_idx]),
            Vector::cast_from(value),
        );
        input[input_idx] = value[0];
    }
}

pub fn scatter<R: Runtime>(
    dim: usize,
    tensor: RudaTensor<R>,
    indices: RudaTensor<R>,
    value: RudaTensor<R>,
    is_bool: bool,
) -> RudaTensor<R> {
    scatter_impl(dim, tensor, indices, value, is_bool, false)
}

pub fn scatter_assign<R: Runtime>(
    dim: usize,
    tensor: RudaTensor<R>,
    indices: RudaTensor<R>,
    value: RudaTensor<R>,
) -> RudaTensor<R> {
    scatter_impl(dim, tensor, indices, value, false, true)
}

fn scatter_impl<R: Runtime>(
    dim: usize,
    tensor: RudaTensor<R>,
    indices: RudaTensor<R>,
    value: RudaTensor<R>,
    is_bool: bool,
    assign: bool,
) -> RudaTensor<R> {
    if value.meta.num_elements() == 0 {
        return tensor;
    }

    let tensor = match tensor.can_mut() && tensor.is_nonoverlapping() {
        true => tensor,
        false => tensor.copy(),
    };

    let num_elems = tensor.meta.num_elements() / tensor.meta.shape()[dim];

    let working_units = num_elems;
    let ruda_dim = RudaDim::new(indices.client.properties(), working_units);
    let ruda_count = calculate_ruda_count_elemwise(&indices.client, working_units, ruda_dim);

    let launch = match (assign, is_bool) {
        (true, _) => scatter_kernel::launch_unchecked::<AssignOp, R>,
        (false, true) => scatter_kernel::launch_unchecked::<OrOp, R>,
        (false, false) => scatter_kernel::launch_unchecked::<AddOp, R>,
    };

    let (tensor_dtype, indices_dtype) = (tensor.dtype, indices.dtype);

    unsafe {
        launch(
            &tensor.client.clone(),
            ruda_count,
            ruda_dim,
            address_type!(tensor, indices, value),
            tensor.clone().into_tensor_arg(),
            indices.into_tensor_arg(),
            value.into_tensor_arg(),
            shape_divmod(&tensor),
            dim,
            [tensor_dtype.into(), indices_dtype.into()],
        )
    }
    tensor
}