fastnn 0.3.3

A robust deep learning library in Rust with CUDA kernels
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

//! Strides, broadcasting and index arithmetic. Tensors are always contiguous, so
//! strides are derived from the shape instead of stored.

/// Row-major strides for `shape`, in elements.
pub fn strides_for(shape: &[usize]) -> Vec<usize> {
    let mut strides = vec![1usize; shape.len()];
    for i in (0..shape.len().saturating_sub(1)).rev() {
        strides[i] = strides[i + 1] * shape[i + 1];
    }
    strides
}

/// Total element count.
pub fn numel(shape: &[usize]) -> usize {
    shape.iter().product()
}

/// Fill in a single `-1` placeholder so the shape holds `numel` elements.
pub fn resolve(shape: &[i64], numel: usize) -> Vec<usize> {
    let mut placeholder = None;
    let mut known = 1usize;
    for (i, &d) in shape.iter().enumerate() {
        if d == -1 {
            assert!(
                placeholder.is_none(),
                "reshape: at most one -1, got {shape:?}"
            );
            placeholder = Some(i);
        } else {
            assert!(
                d > 0,
                "reshape: dimensions must be positive or -1, got {shape:?}"
            );
            known *= d as usize;
        }
    }

    match placeholder {
        Some(i) => {
            assert!(
                known > 0 && numel.is_multiple_of(known),
                "reshape: cannot fit {numel} elements into {shape:?}"
            );
            let mut out: Vec<usize> = shape.iter().map(|&d| d.max(0) as usize).collect();
            out[i] = numel / known;
            out
        }
        None => {
            assert_eq!(
                known, numel,
                "reshape: {shape:?} holds {known} elements, need {numel}"
            );
            shape.iter().map(|&d| d as usize).collect()
        }
    }
}

/// The shape two operands broadcast to, NumPy style.
pub fn broadcast(a: &[usize], b: &[usize]) -> Vec<usize> {
    let ndim = a.len().max(b.len());
    (0..ndim)
        .map(|i| {
            let da = dim_right_aligned(a, ndim, i);
            let db = dim_right_aligned(b, ndim, i);
            assert!(
                da == db || da == 1 || db == 1,
                "cannot broadcast shapes {a:?} and {b:?}"
            );
            da.max(db)
        })
        .collect()
}

/// Left-pad `shape` with 1s to `ndim` dimensions.
pub fn pad_left(shape: &[usize], ndim: usize) -> Vec<usize> {
    let mut out = vec![1usize; ndim - shape.len()];
    out.extend_from_slice(shape);
    out
}

/// Strides for reading the padded `shape` as `out`. Stretched dims get stride 0.
pub fn broadcast_strides(shape: &[usize], out: &[usize]) -> Vec<usize> {
    debug_assert_eq!(shape.len(), out.len());
    strides_for(shape)
        .iter()
        .zip(shape)
        .map(|(&s, &d)| if d == 1 { 0 } else { s })
        .collect()
}

/// Split `shape` around `axis` into (product before, size at, product after).
pub fn split_at_axis(shape: &[usize], axis: usize) -> (usize, usize, usize) {
    assert!(
        axis < shape.len(),
        "axis {axis} out of range for shape {shape:?}"
    );
    (
        numel(&shape[..axis]),
        shape[axis],
        numel(&shape[axis + 1..]),
    )
}

fn dim_right_aligned(shape: &[usize], ndim: usize, i: usize) -> usize {
    let offset = ndim - shape.len();
    if i < offset { 1 } else { shape[i - offset] }
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn strides_are_row_major() {
        assert_eq!(strides_for(&[2, 3, 4]), vec![12, 4, 1]);
        assert_eq!(strides_for(&[5]), vec![1]);
    }

    #[test]
    fn resolve_infers_placeholder() {
        assert_eq!(resolve(&[2, -1], 12), vec![2, 6]);
        assert_eq!(resolve(&[3, 4], 12), vec![3, 4]);
    }

    #[test]
    fn broadcast_right_aligns() {
        assert_eq!(broadcast(&[3, 1], &[4]), vec![3, 4]);
        assert_eq!(broadcast(&[2, 1, 4], &[3, 1]), vec![2, 3, 4]);
    }

    #[test]
    fn axis_split_matches_shape() {
        assert_eq!(split_at_axis(&[2, 3, 4], 1), (2, 3, 4));
        assert_eq!(split_at_axis(&[2, 3, 4], 0), (1, 2, 12));
    }
}