numr 0.5.1

High-performance numerical computing with multi-backend GPU acceleration (CPU/CUDA/WebGPU)
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

//! Macro-based code generation for var_* functions
//!
//! These macros eliminate boilerplate by generating var_* functions with
//! consistent patterns. Each macro handles a specific backward constructor
//! signature pattern.

/// Binary operation with shape-based backward (add, sub)
macro_rules! impl_var_binary_op_shapes {
    ($(#[$meta:meta])* $fn_name:ident, $op_method:ident, $backward_ty:ident) => {
        $(#[$meta])*
        pub fn $fn_name<R, C>(a: &Var<R>, b: &Var<R>, client: &C) -> Result<Var<R>>
        where
            R: Runtime,
            C: RuntimeClient<R> + TensorOps<R>,
            R::Client: TensorOps<R>,
        {
            let output = client.$op_method(a.tensor(), b.tensor())?;

            if a.requires_grad() || b.requires_grad() {
                let grad_fn = $backward_ty::<R>::new(
                    a.id(),
                    b.id(),
                    a.shape(),
                    b.shape(),
                    a.grad_fn().cloned(),
                    b.grad_fn().cloned(),
                );
                Ok(Var::from_op(output, std::sync::Arc::new(grad_fn)))
            } else {
                Ok(Var::new(output, false))
            }
        }
    };
}

/// Binary operation with tensor-based backward (mul, div, matmul)
macro_rules! impl_var_binary_op_tensors {
    ($(#[$meta:meta])* $fn_name:ident, $op_method:ident, $backward_ty:ident) => {
        $(#[$meta])*
        pub fn $fn_name<R, C>(a: &Var<R>, b: &Var<R>, client: &C) -> Result<Var<R>>
        where
            R: Runtime,
            C: RuntimeClient<R> + TensorOps<R>,
            R::Client: TensorOps<R>,
        {
            let output = client.$op_method(a.tensor(), b.tensor())?;

            if a.requires_grad() || b.requires_grad() {
                let grad_fn = $backward_ty::<R>::new(
                    a.id(),
                    b.id(),
                    a.tensor().clone(),
                    b.tensor().clone(),
                    a.grad_fn().cloned(),
                    b.grad_fn().cloned(),
                );
                Ok(Var::from_op(output, std::sync::Arc::new(grad_fn)))
            } else {
                Ok(Var::new(output, false))
            }
        }
    };
}

/// Unary operation with id-only backward (neg)
macro_rules! impl_var_unary_op_id {
    ($(#[$meta:meta])* $fn_name:ident, $op_method:ident, $backward_ty:ident) => {
        $(#[$meta])*
        pub fn $fn_name<R, C>(a: &Var<R>, client: &C) -> Result<Var<R>>
        where
            R: Runtime,
            C: RuntimeClient<R> + TensorOps<R>,
            R::Client: TensorOps<R>,
        {
            let output = client.$op_method(a.tensor())?;

            if a.requires_grad() {
                let grad_fn = $backward_ty::<R>::new(a.id());
                Ok(Var::from_op(output, std::sync::Arc::new(grad_fn)))
            } else {
                Ok(Var::new(output, false))
            }
        }
    };
}

/// Unary operation with output-based backward (exp, sqrt, tanh, recip, sigmoid)
macro_rules! impl_var_unary_op_output {
    ($(#[$meta:meta])* $fn_name:ident, $op_method:ident, $backward_ty:ident) => {
        $(#[$meta])*
        pub fn $fn_name<R, C>(a: &Var<R>, client: &C) -> Result<Var<R>>
        where
            R: Runtime,
            C: RuntimeClient<R> + TensorOps<R>,
            R::Client: TensorOps<R>,
        {
            let output = client.$op_method(a.tensor())?;

            if a.requires_grad() {
                let grad_fn = $backward_ty::<R>::new(a.id(), output.clone());
                Ok(Var::from_op(output, std::sync::Arc::new(grad_fn)))
            } else {
                Ok(Var::new(output, false))
            }
        }
    };
}

/// Unary operation with input tensor backward (log, sin, cos, square, abs, tan, relu)
macro_rules! impl_var_unary_op_input {
    ($(#[$meta:meta])* $fn_name:ident, $op_method:ident, $backward_ty:ident) => {
        $(#[$meta])*
        pub fn $fn_name<R, C>(a: &Var<R>, client: &C) -> Result<Var<R>>
        where
            R: Runtime,
            C: RuntimeClient<R> + TensorOps<R>,
            R::Client: TensorOps<R>,
        {
            let output = client.$op_method(a.tensor())?;

            if a.requires_grad() {
                let grad_fn = $backward_ty::<R>::new(a.id(), a.tensor().clone());
                Ok(Var::from_op(output, std::sync::Arc::new(grad_fn)))
            } else {
                Ok(Var::new(output, false))
            }
        }
    };
}

/// Scalar operation with id-only backward (add_scalar, sub_scalar)
macro_rules! impl_var_scalar_op_id {
    ($(#[$meta:meta])* $fn_name:ident, $op_method:ident, $backward_ty:ident) => {
        $(#[$meta])*
        pub fn $fn_name<R, C>(a: &Var<R>, scalar: f64, client: &C) -> Result<Var<R>>
        where
            R: Runtime,
            C: RuntimeClient<R> + ScalarOps<R>,
            R::Client: ScalarOps<R>,
        {
            let output = client.$op_method(a.tensor(), scalar)?;

            if a.requires_grad() {
                let grad_fn = $backward_ty::<R>::new(a.id(), a.grad_fn().cloned());
                Ok(Var::from_op(output, std::sync::Arc::new(grad_fn)))
            } else {
                Ok(Var::new(output, false))
            }
        }
    };
}

/// Scalar operation with scalar value backward (mul_scalar, div_scalar)
macro_rules! impl_var_scalar_op_scalar {
    ($(#[$meta:meta])* $fn_name:ident, $op_method:ident, $backward_ty:ident) => {
        $(#[$meta])*
        pub fn $fn_name<R, C>(a: &Var<R>, scalar: f64, client: &C) -> Result<Var<R>>
        where
            R: Runtime,
            C: RuntimeClient<R> + ScalarOps<R>,
            R::Client: ScalarOps<R>,
        {
            let output = client.$op_method(a.tensor(), scalar)?;

            if a.requires_grad() {
                let grad_fn = $backward_ty::<R>::new(a.id(), scalar, a.grad_fn().cloned());
                Ok(Var::from_op(output, std::sync::Arc::new(grad_fn)))
            } else {
                Ok(Var::new(output, false))
            }
        }
    };
}

/// Unary operation with output-based backward that requires ScalarOps (sqrt, tanh, recip, sigmoid)
///
/// These operations save the output tensor and require ScalarOps for backward computation.
macro_rules! impl_var_unary_op_output_scalar {
    ($(#[$meta:meta])* $fn_name:ident, $op_method:ident, $backward_ty:ident) => {
        $(#[$meta])*
        pub fn $fn_name<R, C>(a: &Var<R>, client: &C) -> Result<Var<R>>
        where
            R: Runtime<DType = crate::dtype::DType>,
            C: RuntimeClient<R> + TensorOps<R>,
            R::Client: TensorOps<R> + ScalarOps<R>,
        {
            let output = client.$op_method(a.tensor())?;

            if a.requires_grad() {
                let grad_fn = $backward_ty::<R>::new(a.id(), output.clone());
                Ok(Var::from_op(output, std::sync::Arc::new(grad_fn)))
            } else {
                Ok(Var::new(output, false))
            }
        }
    };
}

/// Unary operation with input-based backward that requires ScalarOps (square, tan)
///
/// These operations save the input tensor and require ScalarOps for backward computation.
macro_rules! impl_var_unary_op_input_scalar {
    ($(#[$meta:meta])* $fn_name:ident, $op_method:ident, $backward_ty:ident) => {
        $(#[$meta])*
        pub fn $fn_name<R, C>(a: &Var<R>, client: &C) -> Result<Var<R>>
        where
            R: Runtime,
            C: RuntimeClient<R> + TensorOps<R>,
            R::Client: TensorOps<R> + ScalarOps<R>,
        {
            let output = client.$op_method(a.tensor())?;

            if a.requires_grad() {
                let grad_fn = $backward_ty::<R>::new(a.id(), a.tensor().clone());
                Ok(Var::from_op(output, std::sync::Arc::new(grad_fn)))
            } else {
                Ok(Var::new(output, false))
            }
        }
    };
}

pub(crate) use impl_var_binary_op_shapes;
pub(crate) use impl_var_binary_op_tensors;
pub(crate) use impl_var_scalar_op_id;
pub(crate) use impl_var_scalar_op_scalar;
pub(crate) use impl_var_unary_op_id;
pub(crate) use impl_var_unary_op_input;
pub(crate) use impl_var_unary_op_input_scalar;
pub(crate) use impl_var_unary_op_output;
pub(crate) use impl_var_unary_op_output_scalar;