ruDNN 0.1.108

Neural network operators 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

use ruda_kernel::dsl as kernel_dsl;
use ruda_kernel::library::FastDivmod;
use ruda_kernel::library::tensor::layout::linear::LinearLayout;
use ruda_kernel::library::tensor::layout::*;
use ruda_kernel::dsl::calculate_ruda_count_elemwise;
use ruda_kernel::dsl::prelude::*;

use ruda_kernel::dsl::Runtime;
use ruda_kernel::tensor::layout::address_type;
use ruda_kernel::tensor::layout::linear_layout;
use ruda_kernel::tensor::layout::shape_divmod;
use ruda_kernel::tensor::layout::max_vector_size;
use ruda_kernel::tensor::RudaTensor;
use ruda_core::ir::AddressType;
use crate::indexing::scaled_index_division;

#[ruda(launch_unchecked, address_type = "dynamic")]
fn interpolate_nearest_kernel<F: Float, N: Size>(
    input: &Tensor<Vector<F, N>>,
    output: &mut Tensor<Vector<F, N>>,
    shape_out: Sequence<FastDivmod<usize>>,
    out_layout: LinearLayout,
    #[comptime] max_value: usize,
    #[comptime] exact: bool,
    #[define(F)] _dtype: StorageType,
) {
    if ABSOLUTE_POS >= output.len() {
        terminate!();
    }

    let vector_size = input.vector_size();
    let out_idx = out_layout.to_source_pos(ABSOLUTE_POS);

    let out_pos = ABSOLUTE_POS * vector_size;

    let (h_in, w_in) = (input.shape(1), input.shape(2));
    let (h_out, w_out) = (output.shape(1), output.shape(2));

    let (rem, c) = shape_out[3].div_mod(out_pos);
    let (rem, x) = shape_out[2].div_mod(rem);
    let (b, y) = shape_out[1].div_mod(rem);

    let y = source_index(y, h_in, h_out, max_value, exact);
    let x = source_index(x, w_in, w_out, max_value, exact);

    let in_idx =
        b * input.stride(0) + y * input.stride(1) + x * input.stride(2) + c * input.stride(3);

    output[out_idx] = input[in_idx / vector_size];
}

#[ruda]
fn source_index(index: usize, input_size: usize, output_size: usize,
    #[comptime] max_value: usize, #[comptime] exact: bool) -> usize {
    let (quotient, remainder) = scaled_index_division(index, input_size, output_size, max_value);
    if exact {
        let half = input_size / 2;
        let shifted = quotient + half / output_size;
        if remainder >= output_size - half % output_size {
            shifted + 1
        } else {
            shifted
        }
    } else {
        quotient
    }
}

pub fn interpolate_nearest_launch<R: Runtime>(
    input: RudaTensor<R>,
    output: RudaTensor<R>,
) -> RudaTensor<R> {
    interpolate_nearest_launch_mode(input, output, false)
}

pub(super) fn interpolate_nearest_launch_mode<R: Runtime>(
    input: RudaTensor<R>, output: RudaTensor<R>, exact: bool,
) -> RudaTensor<R> {
    let client = input.client.clone();

    let vector_size = max_vector_size(&input);

    let working_units = output.meta.num_elements() / vector_size as usize;
    let ruda_dim = RudaDim::new(input.client.properties(), working_units);
    let ruda_count = calculate_ruda_count_elemwise(&input.client, working_units, ruda_dim);

    let shape_out = shape_divmod(&output);
    let out_layout = linear_layout(&output, vector_size);
    let address = address_type!(input, output);
    let max_value = if address == AddressType::U64 { usize::MAX } else { u32::MAX as usize };

    unsafe {
        interpolate_nearest_kernel::launch_unchecked(
            &client,
            ruda_count,
            ruda_dim,
            address,
            vector_size,
            input.into_tensor_arg(),
            output.clone().into_tensor_arg(),
            shape_out,
            out_layout,
            max_value,
            exact,
            output.dtype.into(),
        )
    };

    output
}