luminal 0.2.0

Deep learning at the speed of light.
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
265
266
267
268

use crate::op;
use crate::prelude::*;
use std::ops::AddAssign;
use std::ops::DivAssign;
use std::ops::MulAssign;
use std::ops::RemAssign;
use std::ops::SubAssign;
use std::ops::{Add, Div, Mul, Rem, Sub};

use self::symbolic::ExpressionStorage;
use self::symbolic::GenericExpression;
use self::symbolic::Term;

impl<S: Shape> Add for GraphTensor<S> {
    type Output = GraphTensor<S>;

    fn add(mut self, mut rhs: GraphTensor<S>) -> Self::Output {
        resolve_local_dyn_dims(&mut self.shape, &mut rhs.shape, false);
        let mut new_shape = ShapeTracker::new(&S::realized_shape());
        resolve_local_dyn_dims(&mut new_shape, &mut rhs.shape, false);
        let new_id = self
            .graph()
            .add_op(op::Add)
            .input(self.id, 0, self.shape)
            .input(rhs.id, 0, rhs.shape)
            .finish();
        GraphTensor::from_id(new_id, new_shape, self.graph_ref)
    }
}

impl<S: Shape> AddAssign for GraphTensor<S> {
    fn add_assign(&mut self, rhs: Self) {
        *self = *self + rhs;
    }
}

impl<S: Shape> Sub for GraphTensor<S> {
    type Output = GraphTensor<S>;

    fn sub(self, rhs: GraphTensor<S>) -> Self::Output {
        self + -rhs
    }
}

impl<S: Shape> SubAssign for GraphTensor<S> {
    fn sub_assign(&mut self, rhs: Self) {
        *self = *self - rhs;
    }
}

impl<S: Shape> Mul for GraphTensor<S> {
    type Output = GraphTensor<S>;

    fn mul(mut self, mut rhs: GraphTensor<S>) -> Self::Output {
        resolve_local_dyn_dims(&mut self.shape, &mut rhs.shape, false);
        let new_id = self
            .graph()
            .add_op(op::Mul)
            .input(self.id, 0, self.shape)
            .input(rhs.id, 0, rhs.shape)
            .finish();
        GraphTensor::from_id(new_id, self.shape.contiguous(), self.graph_ref)
    }
}

impl<S: Shape> MulAssign for GraphTensor<S> {
    fn mul_assign(&mut self, rhs: Self) {
        *self = *self * rhs;
    }
}

#[allow(clippy::suspicious_arithmetic_impl)]
impl<S: Shape> Div<GraphTensor<S>> for GraphTensor<S> {
    type Output = GraphTensor<S>;

    fn div(self, rhs: GraphTensor<S>) -> Self::Output {
        self * rhs.recip()
    }
}

impl<S: Shape> DivAssign for GraphTensor<S> {
    fn div_assign(&mut self, rhs: Self) {
        *self = *self / rhs;
    }
}

impl<S: Shape> Rem<GraphTensor<S>> for GraphTensor<S> {
    type Output = GraphTensor<S>;

    fn rem(mut self, mut rhs: GraphTensor<S>) -> Self::Output {
        resolve_local_dyn_dims(&mut self.shape, &mut rhs.shape, false);
        let mut new_shape = ShapeTracker::new(&S::realized_shape());
        resolve_local_dyn_dims(&mut new_shape, &mut rhs.shape, false);
        let new_id = self
            .graph()
            .add_op(op::Mod)
            .input(self.id, 0, self.shape)
            .input(rhs.id, 0, rhs.shape)
            .finish();
        GraphTensor::from_id(new_id, new_shape, self.graph_ref)
    }
}

impl<S: Shape> RemAssign for GraphTensor<S> {
    fn rem_assign(&mut self, rhs: Self) {
        *self = *self % rhs;
    }
}

impl<S: Shape> Add<f32> for GraphTensor<S> {
    type Output = GraphTensor<S>;

    fn add(self, rhs: f32) -> Self::Output {
        self + self.graph().constant(rhs).expand()
    }
}

impl<S: Shape, St: ExpressionStorage> Add<GenericExpression<St>> for GraphTensor<S>
where
    GenericExpression<Vec<Term>>: From<GenericExpression<St>>,
{
    type Output = GraphTensor<S>;

    fn add(self, rhs: GenericExpression<St>) -> Self::Output {
        self + self.graph().constant_expr(rhs).expand()
    }
}

impl<S: Shape> Sub<f32> for GraphTensor<S> {
    type Output = GraphTensor<S>;

    fn sub(self, rhs: f32) -> Self::Output {
        self - self.graph().constant(rhs).expand()
    }
}

impl<S: Shape, St: ExpressionStorage> Sub<GenericExpression<St>> for GraphTensor<S>
where
    GenericExpression<Vec<Term>>: From<GenericExpression<St>>,
{
    type Output = GraphTensor<S>;

    fn sub(self, rhs: GenericExpression<St>) -> Self::Output {
        self - self.graph().constant_expr(rhs).expand()
    }
}

impl<S: Shape> Mul<f32> for GraphTensor<S> {
    type Output = GraphTensor<S>;

    fn mul(self, rhs: f32) -> Self::Output {
        self * self.graph().constant(rhs).expand()
    }
}

impl<S: Shape, St: ExpressionStorage> Mul<GenericExpression<St>> for GraphTensor<S>
where
    GenericExpression<Vec<Term>>: From<GenericExpression<St>>,
{
    type Output = GraphTensor<S>;

    fn mul(self, rhs: GenericExpression<St>) -> Self::Output {
        self * self.graph().constant_expr(rhs).expand()
    }
}

#[allow(clippy::suspicious_arithmetic_impl)]
impl<S: Shape> Div<f32> for GraphTensor<S> {
    type Output = GraphTensor<S>;

    fn div(self, rhs: f32) -> Self::Output {
        self * self.graph().constant(rhs.recip()).expand()
    }
}

impl<S: Shape, St: ExpressionStorage> Div<GenericExpression<St>> for GraphTensor<S>
where
    GenericExpression<Vec<Term>>: From<GenericExpression<St>>,
{
    type Output = GraphTensor<S>;

    fn div(self, rhs: GenericExpression<St>) -> Self::Output {
        self / self.graph().constant_expr(rhs).expand()
    }
}

impl<S: Shape> Rem<f32> for GraphTensor<S> {
    type Output = GraphTensor<S>;

    fn rem(self, rhs: f32) -> Self::Output {
        self % self.graph().constant(rhs).expand()
    }
}

impl<S: Shape, St: ExpressionStorage> Rem<GenericExpression<St>> for GraphTensor<S>
where
    GenericExpression<Vec<Term>>: From<GenericExpression<St>>,
{
    type Output = GraphTensor<S>;

    fn rem(self, rhs: GenericExpression<St>) -> Self::Output {
        self % self.graph().constant_expr(rhs).expand()
    }
}

// Comparisons (based on https://github.com/tinygrad/tinygrad/blob/3e0c2d256fe9f4f5f85cd3e4d8733a51d7b4a984/tinygrad/tensor.py#L653)
impl<S: Shape> GraphTensor<S> {
    pub fn less_than(mut self, mut rhs: GraphTensor<S>) -> GraphTensor<S> {
        resolve_local_dyn_dims(&mut self.shape, &mut rhs.shape, false);
        let mut new_shape = ShapeTracker::new(&S::realized_shape());
        resolve_local_dyn_dims(&mut new_shape, &mut rhs.shape, false);
        let new_id = self
            .graph()
            .add_op(op::LessThan)
            .input(self.id, 0, self.shape)
            .input(rhs.id, 0, rhs.shape)
            .finish();
        GraphTensor::from_id(new_id, new_shape, self.graph_ref)
    }

    pub fn greater_than(self, rhs: GraphTensor<S>) -> GraphTensor<S> {
        rhs.less_than(self)
    }

    pub fn less_than_equal(self, rhs: GraphTensor<S>) -> GraphTensor<S> {
        -self.greater_than(rhs) + 1.0
    }

    pub fn greater_than_equal(self, rhs: GraphTensor<S>) -> GraphTensor<S> {
        -self.less_than(rhs) + 1.0
    }

    pub fn not_equals(self, rhs: GraphTensor<S>) -> GraphTensor<S> {
        self.less_than(rhs) + self.greater_than(rhs)
    }

    pub fn equals(self, rhs: GraphTensor<S>) -> GraphTensor<S> {
        -self.not_equals(rhs) + 1.0
    }
}

// Clipping ops (min, max, clip)
impl<S: Shape> GraphTensor<S> {
    /// Take the elementwise maximum of two tensors
    pub fn max(self, rhs: GraphTensor<S>) -> GraphTensor<S> {
        (self.less_than(rhs) * rhs) + (rhs.less_than_equal(self) * self)
    }

    /// Take the elementwise maximum of a tensor and a float
    pub fn max_f32(self, rhs: f32) -> GraphTensor<S> {
        self.max(self.graph().constant(rhs).expand())
    }

    /// Take the elementwise minimum of two tensors
    pub fn min(self, rhs: GraphTensor<S>) -> GraphTensor<S> {
        -(-self).max(-rhs)
    }

    /// Take the elementwise minimum of a tensor and a float
    pub fn min_f32(self, rhs: f32) -> GraphTensor<S> {
        -(-self).max_f32(-rhs)
    }

    /// Clip a tensor in a range
    pub fn clip(self, min: f32, max: f32) -> GraphTensor<S> {
        self.min_f32(min).max_f32(max)
    }
}