candela-tensor 0.2.0

A lazy, graph-based tensor engine in Rust
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 crate::tensor::definitions::NumberLike;
use crate::tensor::storage::{Storage, TensorData};
use crate::tensor::traits::Numeric;
use crate::{Dimension, Layout, branch_fast_iter};

#[inline]
fn compute_op_tensor<T: NumberLike, F: Fn(T, T) -> T>(
    inputs: &[TensorData<T>],
    output_buffer: &mut [T],
    start_value: T,
    op: F,
) {
    output_buffer[0] = start_value;

    branch_fast_iter!(inputs[0].fast_iter() => it, {
        for el in it {
            output_buffer[0] = op(output_buffer[0], *el);
        }
    });
}

#[inline]
pub(crate) fn compute_sum_tensor<T: Numeric>(
    inputs: &[TensorData<T>],
    mut output_buffer: Vec<T>,
    output_layout: &Layout,
) -> TensorData<T> {
    compute_op_tensor(inputs, &mut output_buffer, T::SUM_NEUTRAL, |a, b| a + b);

    TensorData::new(Storage::from_vec(output_buffer), output_layout.clone())
}

#[inline]
pub(crate) fn compute_max_tensor<T: Numeric, F: Fn(T, T) -> T>(
    inputs: &[TensorData<T>],
    mut output_buffer: Vec<T>,
    output_layout: &Layout,
    max: F,
) -> TensorData<T> {
    compute_op_tensor(inputs, &mut output_buffer, T::MIN, max);

    TensorData::new(Storage::from_vec(output_buffer), output_layout.clone())
}

#[inline]
pub(crate) fn compute_mean_tensor<T: Numeric, F: FnOnce(T, usize) -> T>(
    inputs: &[TensorData<T>],
    mut output_buffer: Vec<T>,
    output_layout: &Layout,
    div: F,
) -> TensorData<T> {
    compute_op_tensor(inputs, &mut output_buffer, T::SUM_NEUTRAL, |a, b| a + b);
    output_buffer[0] = div(output_buffer[0], inputs[0].layout().len());

    TensorData::new(Storage::from_vec(output_buffer), output_layout.clone())
}

// TODO: Confirm that the contiguous path can vectorize
// TODO: Benchmark if the contiguous path can be made faster by reducing striding code
fn compute_op_axis_tensor<T: NumberLike, F: Fn(T, T) -> T>(
    inputs: &[TensorData<T>],
    axis: usize,
    output_buffer: &mut [T],
    start_value: T,
    op: F,
) {
    let t = &inputs[0];
    let n_outer: usize = t.shape()[..axis].iter().product();
    let n_size: usize = t.shape()[axis];
    let n_inner: usize = t.shape()[axis + 1..].iter().product();

    let mut base_pos = 0;
    output_buffer.fill(start_value);

    branch_fast_iter!(
        t.fast_iter() => _it, {
            let mut it = _it;

            for _ in 0..n_outer {
                for _ in 0..n_size {
                    for inner in 0..n_inner {
                        let el = unsafe { it.next().unwrap_unchecked() };
                        let current = output_buffer[base_pos + inner];

                        output_buffer[base_pos + inner] = op(current, *el);
                    }
                }

                base_pos += n_inner;
            }
        }
    );
}

#[inline]
pub(crate) fn compute_sum_axis_tensor<T: Numeric>(
    inputs: &[TensorData<T>],
    axis: usize,
    mut output_buffer: Vec<T>,
    output_layout: &Layout,
) -> TensorData<T> {
    compute_op_axis_tensor(inputs, axis, &mut output_buffer, T::SUM_NEUTRAL, |a, b| {
        a + b
    });

    TensorData::new(Storage::from_vec(output_buffer), output_layout.clone())
}

#[inline]
pub(crate) fn compute_max_axis_tensor<T: Numeric, F: Fn(T, T) -> T>(
    inputs: &[TensorData<T>],
    axis: usize,
    mut output_buffer: Vec<T>,
    output_layout: &Layout,
    max: F,
) -> TensorData<T> {
    compute_op_axis_tensor(inputs, axis, &mut output_buffer, T::MIN, max);

    TensorData::new(Storage::from_vec(output_buffer), output_layout.clone())
}

#[inline]
pub(crate) fn compute_mean_axis_tensor<T: Numeric, F: Fn(T, usize) -> T>(
    inputs: &[TensorData<T>],
    axis: usize,
    mut output_buffer: Vec<T>,
    output_layout: &Layout,
    div: F,
) -> TensorData<T> {
    compute_op_axis_tensor(inputs, axis, &mut output_buffer, T::SUM_NEUTRAL, |a, b| {
        a + b
    });

    let n = inputs[0].shape()[axis];
    for el in output_buffer.iter_mut() {
        *el = div(*el, n);
    }

    TensorData::new(Storage::from_vec(output_buffer), output_layout.clone())
}