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
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264

//! Ops that rearrange elements without arithmetic. `reshape` shares storage,
//! the rest write a new buffer.

use crate::autograd::ops::view::{
    CatBackward, ExpandBackward, NarrowBackward, PermuteBackward, ReshapeBackward,
};
use crate::cuda::kernels;
use crate::tensor::Tensor;
use crate::tensor::shape;
use crate::tensor::storage::Storage;

impl Tensor {
    /// Same elements under a new shape, sharing storage. One dim may be `-1`.
    pub fn reshape(&self, dims: &[i64]) -> Tensor {
        let new_shape = shape::resolve(dims, self.numel());
        let old_shape = self.shape().to_vec();
        reshaped(self, new_shape).with_grad(&[self], || ReshapeBackward { shape: old_shape })
    }

    /// Collapse everything into one dimension.
    pub fn flatten(&self) -> Tensor {
        self.reshape(&[-1])
    }

    /// Collapse dimensions `from..` into one, e.g. `[N, C, H, W]` to `[N, C*H*W]` at 1.
    pub fn flatten_from(&self, from: usize) -> Tensor {
        let mut dims: Vec<i64> = self.shape()[..from].iter().map(|&d| d as i64).collect();
        dims.push(-1);
        self.reshape(&dims)
    }

    /// Insert a dimension of size 1 at `axis`.
    pub fn unsqueeze(&self, axis: usize) -> Tensor {
        assert!(
            axis <= self.ndim(),
            "unsqueeze: axis {axis} past shape {:?}",
            self.shape()
        );
        let mut dims: Vec<i64> = self.shape().iter().map(|&d| d as i64).collect();
        dims.insert(axis, 1);
        self.reshape(&dims)
    }

    /// Drop the dimension at `axis`, which must have size 1.
    pub fn squeeze(&self, axis: usize) -> Tensor {
        assert_eq!(
            self.dim(axis),
            1,
            "squeeze: axis {axis} has size {} in shape {:?}",
            self.dim(axis),
            self.shape()
        );
        let mut dims: Vec<i64> = self.shape().iter().map(|&d| d as i64).collect();
        dims.remove(axis);
        if dims.is_empty() {
            dims.push(1);
        }
        self.reshape(&dims)
    }

    /// Drop every dimension of size 1.
    pub fn squeeze_all(&self) -> Tensor {
        let dims: Vec<i64> = self
            .shape()
            .iter()
            .filter(|&&d| d != 1)
            .map(|&d| d as i64)
            .collect();
        self.reshape(if dims.is_empty() { &[1] } else { &dims })
    }

    /// Reorder dimensions: output dimension `i` is input dimension `order[i]`.
    pub fn permute(&self, order: &[usize]) -> Tensor {
        assert_eq!(
            order.len(),
            self.ndim(),
            "permute: {order:?} does not cover shape {:?}",
            self.shape()
        );
        let out_shape: Vec<usize> = order.iter().map(|&d| self.dim(d)).collect();
        let in_strides = shape::strides_for(self.shape());
        let order = order.to_vec();

        let out = gather(self, out_shape, &in_strides, &order);
        out.with_grad(&[self], || PermuteBackward { order })
    }

    /// Swap the last two dimensions.
    pub fn transpose(&self) -> Tensor {
        assert!(
            self.ndim() >= 2,
            "transpose needs 2+ dimensions, got {:?}",
            self.shape()
        );
        let n = self.ndim();
        let mut order: Vec<usize> = (0..n).collect();
        order.swap(n - 2, n - 1);
        self.permute(&order)
    }

    /// Alias for [`transpose`](Tensor::transpose).
    pub fn t(&self) -> Tensor {
        self.transpose()
    }

    /// Repeat size-1 dimensions out to `to`, which must have the same rank.
    pub fn expand(&self, to: &[usize]) -> Tensor {
        assert_eq!(
            to.len(),
            self.ndim(),
            "expand: {to:?} does not match rank of {:?}",
            self.shape()
        );
        let from = self.shape().to_vec();
        expanded(self, to).with_grad(&[self], || ExpandBackward { shape: from })
    }

    /// Tile each dimension `times[i]` times.
    pub fn repeat(&self, times: &[usize]) -> Tensor {
        let to: Vec<usize> = self
            .shape()
            .iter()
            .zip(times)
            .map(|(&d, &n)| d * n)
            .collect();
        self.expand(&to)
    }

    /// A `len`-long slice of dimension `axis`, starting at `start`.
    pub fn narrow(&self, axis: usize, start: usize, len: usize) -> Tensor {
        let (outer, size, inner) = shape::split_at_axis(self.shape(), axis);
        assert!(
            start + len <= size,
            "narrow: {start}..{} outside axis {axis} of shape {:?}",
            start + len,
            self.shape()
        );

        let src = self.to_vec();
        let mut data = Vec::with_capacity(outer * len * inner);
        for o in 0..outer {
            let row = (o * size + start) * inner;
            data.extend_from_slice(&src[row..row + len * inner]);
        }

        let mut out_shape = self.shape().to_vec();
        out_shape[axis] = len;
        let in_shape = self.shape().to_vec();

        Tensor::from_vec(data, &out_shape)
            .to(self.device())
            .with_grad(&[self], || NarrowBackward {
                shape: in_shape,
                axis,
                start,
            })
    }

    /// Join tensors along an existing dimension. All other dimensions must match.
    pub fn cat(parts: &[&Tensor], axis: usize) -> Tensor {
        assert!(!parts.is_empty(), "cat: needs at least one tensor");
        let first = parts[0];

        let mut out_shape = first.shape().to_vec();
        out_shape[axis] = parts.iter().map(|t| t.dim(axis)).sum();
        for t in parts {
            assert_eq!(
                t.ndim(),
                first.ndim(),
                "cat: rank mismatch {:?} vs {:?}",
                t.shape(),
                first.shape()
            );
            for d in 0..t.ndim() {
                assert!(
                    d == axis || t.dim(d) == first.dim(d),
                    "cat: shapes {:?} and {:?} differ outside axis {axis}",
                    t.shape(),
                    first.shape()
                );
            }
        }

        let (outer, _, inner) = shape::split_at_axis(&out_shape, axis);
        let sources: Vec<Vec<f32>> = parts.iter().map(|t| t.to_vec()).collect();
        let mut data = Vec::with_capacity(shape::numel(&out_shape));
        for o in 0..outer {
            for (part, src) in parts.iter().zip(&sources) {
                let chunk = part.dim(axis) * inner;
                data.extend_from_slice(&src[o * chunk..(o + 1) * chunk]);
            }
        }

        let sizes: Vec<usize> = parts.iter().map(|t| t.dim(axis)).collect();
        let refs: Vec<&Tensor> = parts.to_vec();
        Tensor::from_vec(data, &out_shape)
            .to(first.device())
            .with_grad(&refs, || CatBackward { axis, sizes })
    }

    /// Join tensors along a new dimension inserted at `axis`.
    pub fn stack(parts: &[&Tensor], axis: usize) -> Tensor {
        let lifted: Vec<Tensor> = parts.iter().map(|t| t.unsqueeze(axis)).collect();
        Tensor::cat(&lifted.iter().collect::<Vec<_>>(), axis)
    }
}

/// Same shape change as [`Tensor::reshape`], with no graph node.
pub(crate) fn reshaped(x: &Tensor, shape: Vec<usize>) -> Tensor {
    debug_assert_eq!(shape::numel(&shape), x.numel());
    Tensor::raw_shared(x.storage_arc(), shape, x.device())
}

/// Same expansion as [`Tensor::expand`], with no graph node.
pub(crate) fn expanded(x: &Tensor, to: &[usize]) -> Tensor {
    if x.shape() == to {
        return x.detach();
    }
    // stride 0 on stretched dims lets the permute gather double as expand
    let in_strides: Vec<usize> = shape::strides_for(x.shape())
        .iter()
        .enumerate()
        .map(|(d, &s)| if x.dim(d) == 1 { 0 } else { s })
        .collect();
    let identity: Vec<usize> = (0..x.ndim()).collect();
    gather(x, to.to_vec(), &in_strides, &identity)
}

/// Reshape then expand, so a lower-rank tensor can broadcast against `to`.
pub(crate) fn broadcast_to(x: &Tensor, to: &[usize]) -> Tensor {
    if x.shape() == to {
        return x.detach();
    }
    let padded = shape::pad_left(x.shape(), to.len());
    expanded(&reshaped(x, padded), to)
}

/// Read each output position through `in_strides[order[d]]`. Serves both permute and expand.
fn gather(x: &Tensor, out_shape: Vec<usize>, in_strides: &[usize], order: &[usize]) -> Tensor {
    let numel = shape::numel(&out_shape);
    let out_strides = shape::strides_for(&out_shape);

    if let Storage::Cuda(buf) = x.storage() {
        let out = kernels::permute_nd(buf, &out_strides, in_strides, order, numel)
            .expect("cuda permute_nd");
        return Tensor::raw(Storage::Cuda(out), out_shape, x.device());
    }

    let src = x.to_vec();
    let ndim = out_shape.len();
    let data: Vec<f32> = (0..numel)
        .map(|flat| {
            let (mut rest, mut src_index) = (flat, 0usize);
            for d in 0..ndim {
                let coord = rest / out_strides[d];
                rest %= out_strides[d];
                src_index += coord * in_strides[order[d]];
            }
            src[src_index]
        })
        .collect();

    Tensor::from_vec(data, &out_shape)
}