ruDNN 0.1.43

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
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

use super::bicubic::{bicubic_coordinate, cubic_coefficient};
use ruda_kernel::dsl as kernel_dsl;
use ruda_kernel::dsl::calculate_ruda_count_elemwise;
use ruda_kernel::dsl::num_traits::Zero;
use ruda_kernel::dsl::prelude::*;
use ruda_kernel::library::FastDivmod;
use ruda_kernel::library::tensor::layout::linear::LinearLayout;
use ruda_kernel::library::tensor::layout::*;
use ruda_kernel::tensor::RudaTensor;
use ruda_kernel::tensor::layout::{address_type, linear_layout, max_vector_size_many, shape_divmod};

#[ruda]
fn contributing_range(
    input_index: usize,
    input_size: usize,
    output_size: usize,
    #[comptime] align_corners: bool,
) -> (usize, usize) {
    let last = (input_size - 1) as f32;
    let mut left = 0usize;
    let mut right = output_size;
    while left < right {
        let middle = left + (right - left) / 2;
        let coordinate = bicubic_coordinate(middle, input_size, output_size, align_corners);
        let upper = clamp(coordinate.floor() + 2.0, 0.0, last) as usize;
        if upper < input_index {
            left = middle + 1;
        } else {
            right = middle;
        }
    }
    let start = left;
    right = output_size;
    while left < right {
        let middle = left + (right - left) / 2;
        let coordinate = bicubic_coordinate(middle, input_size, output_size, align_corners);
        let lower = clamp(coordinate.floor() - 1.0, 0.0, last) as usize;
        if lower <= input_index {
            left = middle + 1;
        } else {
            right = middle;
        }
    }
    (start, left)
}

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

    let vector_size = grad.vector_size();
    let out_idx = out_layout.to_source_pos(ABSOLUTE_POS);
    let (rem, channel) = shape_out[3].div_mod(ABSOLUTE_POS * vector_size);
    let (rem, input_x) = shape_out[2].div_mod(rem);
    let (batch, input_y) = shape_out[1].div_mod(rem);
    let height = output.shape(1);
    let width = output.shape(2);
    let last_y = (height - 1) as f32;
    let last_x = (width - 1) as f32;
    let grad_height = grad.shape(1);
    let grad_width = grad.shape(2);
    let (y_start, y_end) = contributing_range(input_y, height, grad_height, align_corners);
    let (x_start, x_end) = contributing_range(input_x, width, grad_width, align_corners);
    let index_base = batch * grad.stride(0) + channel * grad.stride(3);
    let mut sum = Vector::zero();

    for y in y_start..y_end {
        let coordinate = bicubic_coordinate(y, height, grad_height, align_corners);
        let y_floor = coordinate.floor();
        let yw = Vector::<F, N>::new(F::cast_from(coordinate - y_floor));

        for x in x_start..x_end {
            let coordinate = bicubic_coordinate(x, width, grad_width, align_corners);
            let x_floor = coordinate.floor();
            let xw = Vector::<F, N>::new(F::cast_from(coordinate - x_floor));
            let index = index_base + y * grad.stride(1) + x * grad.stride(2);
            let value = grad[index / vector_size];

            #[unroll]
            for ky in 0..4usize {
                let offset_y = comptime![ky as f32 - 1.0];
                let y_index = clamp(y_floor + offset_y, 0.0, last_y) as usize;
                if y_index == input_y {
                    let row_grad = value * cubic_coefficient(yw, ky);
                    #[unroll]
                    for kx in 0..4usize {
                        let offset_x = comptime![kx as f32 - 1.0];
                        let x_index = clamp(x_floor + offset_x, 0.0, last_x) as usize;
                        if x_index == input_x {
                            sum += row_grad * cubic_coefficient(xw, kx);
                        }
                    }
                }
            }
        }
    }

    output[out_idx] = sum;
}

pub fn interpolate_bicubic_backward_launch<R: Runtime>(
    out_grad: RudaTensor<R>,
    output: RudaTensor<R>,
    align_corners: bool,
) -> RudaTensor<R> {
    if output.meta.num_elements() == 0 {
        return output;
    }
    let vector_size = max_vector_size_many(&[&out_grad, &output], 3);
    let out_shape = shape_divmod(&output);
    let out_layout = linear_layout(&output, vector_size);
    let working_units = output.meta.num_elements() / vector_size as usize;
    let ruda_dim = RudaDim::new(out_grad.client.properties(), working_units);
    let ruda_count = calculate_ruda_count_elemwise(&out_grad.client, working_units, ruda_dim);

    interpolate_bicubic_backward_kernel::launch(
        &output.client,
        ruda_count,
        ruda_dim,
        address_type!(out_grad, output),
        vector_size,
        out_grad.into_tensor_arg(),
        output.clone().into_tensor_arg(),
        out_shape,
        out_layout,
        align_corners,
        output.dtype.into(),
    );

    output
}