numr 0.5.2

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
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
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326

//! Generic scalar tensor operation adapters
//!
//! Provides `apply_unary`, `apply_binary`, and `apply_ternary` helpers that
//! apply scalar functions element-wise over tensors with proper error handling.

use crate::dtype::DType;
use crate::error::{Error, Result};
use crate::runtime::cpu::{CpuDevice, CpuRuntime};
use crate::tensor::Tensor;

// ============================================================================
// Core dispatch helpers (all dtype variants delegate to these)
// ============================================================================

/// Internal: apply a unary f64→f64 function over any float tensor.
fn apply_unary_via_f64<F>(
    x: &Tensor<CpuRuntime>,
    device: &CpuDevice,
    f: F,
) -> Result<Tensor<CpuRuntime>>
where
    F: Fn(f64) -> f64,
{
    match x.dtype() {
        DType::F32 => {
            let data: Vec<f32> = x.to_vec();
            let result: Vec<f32> = data.iter().map(|&v| f(v as f64) as f32).collect();
            Ok(Tensor::from_slice(&result, x.shape(), device))
        }
        DType::F64 => {
            let data: Vec<f64> = x.to_vec();
            let result: Vec<f64> = data.iter().map(|&v| f(v)).collect();
            Ok(Tensor::from_slice(&result, x.shape(), device))
        }
        #[cfg(feature = "f16")]
        DType::F16 => {
            let data: Vec<half::f16> = x.to_vec();
            let result: Vec<half::f16> = data
                .iter()
                .map(|&v| half::f16::from_f64(f(v.to_f64())))
                .collect();
            Ok(Tensor::from_slice(&result, x.shape(), device))
        }
        #[cfg(feature = "f16")]
        DType::BF16 => {
            let data: Vec<half::bf16> = x.to_vec();
            let result: Vec<half::bf16> = data
                .iter()
                .map(|&v| half::bf16::from_f64(f(v.to_f64())))
                .collect();
            Ok(Tensor::from_slice(&result, x.shape(), device))
        }
        #[cfg(feature = "fp8")]
        DType::FP8E4M3 => {
            let data: Vec<crate::dtype::FP8E4M3> = x.to_vec();
            let result: Vec<crate::dtype::FP8E4M3> = data
                .iter()
                .map(|&v| crate::dtype::FP8E4M3::from_f32(f(v.to_f32() as f64) as f32))
                .collect();
            Ok(Tensor::from_slice(&result, x.shape(), device))
        }
        #[cfg(feature = "fp8")]
        DType::FP8E5M2 => {
            let data: Vec<crate::dtype::FP8E5M2> = x.to_vec();
            let result: Vec<crate::dtype::FP8E5M2> = data
                .iter()
                .map(|&v| crate::dtype::FP8E5M2::from_f32(f(v.to_f32() as f64) as f32))
                .collect();
            Ok(Tensor::from_slice(&result, x.shape(), device))
        }
        _ => unreachable!("dtype validated by caller"),
    }
}

/// Internal: apply a binary (f64,f64)→f64 function over any two float tensors.
fn apply_binary_via_f64<F>(
    a: &Tensor<CpuRuntime>,
    b: &Tensor<CpuRuntime>,
    device: &CpuDevice,
    f: F,
) -> Result<Tensor<CpuRuntime>>
where
    F: Fn(f64, f64) -> f64,
{
    if a.shape() != b.shape() {
        return Err(Error::ShapeMismatch {
            expected: a.shape().to_vec(),
            got: b.shape().to_vec(),
        });
    }

    match a.dtype() {
        DType::F32 => {
            let a_data: Vec<f32> = a.to_vec();
            let b_data: Vec<f32> = b.to_vec();
            let result: Vec<f32> = a_data
                .iter()
                .zip(b_data.iter())
                .map(|(&av, &bv)| f(av as f64, bv as f64) as f32)
                .collect();
            Ok(Tensor::from_slice(&result, a.shape(), device))
        }
        DType::F64 => {
            let a_data: Vec<f64> = a.to_vec();
            let b_data: Vec<f64> = b.to_vec();
            let result: Vec<f64> = a_data
                .iter()
                .zip(b_data.iter())
                .map(|(&av, &bv)| f(av, bv))
                .collect();
            Ok(Tensor::from_slice(&result, a.shape(), device))
        }
        #[cfg(feature = "f16")]
        DType::F16 => {
            let a_data: Vec<half::f16> = a.to_vec();
            let b_data: Vec<half::f16> = b.to_vec();
            let result: Vec<half::f16> = a_data
                .iter()
                .zip(b_data.iter())
                .map(|(&av, &bv)| half::f16::from_f64(f(av.to_f64(), bv.to_f64())))
                .collect();
            Ok(Tensor::from_slice(&result, a.shape(), device))
        }
        #[cfg(feature = "f16")]
        DType::BF16 => {
            let a_data: Vec<half::bf16> = a.to_vec();
            let b_data: Vec<half::bf16> = b.to_vec();
            let result: Vec<half::bf16> = a_data
                .iter()
                .zip(b_data.iter())
                .map(|(&av, &bv)| half::bf16::from_f64(f(av.to_f64(), bv.to_f64())))
                .collect();
            Ok(Tensor::from_slice(&result, a.shape(), device))
        }
        #[cfg(feature = "fp8")]
        DType::FP8E4M3 => {
            let a_data: Vec<crate::dtype::FP8E4M3> = a.to_vec();
            let b_data: Vec<crate::dtype::FP8E4M3> = b.to_vec();
            let result: Vec<crate::dtype::FP8E4M3> =
                a_data
                    .iter()
                    .zip(b_data.iter())
                    .map(|(&av, &bv)| {
                        crate::dtype::FP8E4M3::from_f32(
                            f(av.to_f32() as f64, bv.to_f32() as f64) as f32
                        )
                    })
                    .collect();
            Ok(Tensor::from_slice(&result, a.shape(), device))
        }
        #[cfg(feature = "fp8")]
        DType::FP8E5M2 => {
            let a_data: Vec<crate::dtype::FP8E5M2> = a.to_vec();
            let b_data: Vec<crate::dtype::FP8E5M2> = b.to_vec();
            let result: Vec<crate::dtype::FP8E5M2> =
                a_data
                    .iter()
                    .zip(b_data.iter())
                    .map(|(&av, &bv)| {
                        crate::dtype::FP8E5M2::from_f32(
                            f(av.to_f32() as f64, bv.to_f32() as f64) as f32
                        )
                    })
                    .collect();
            Ok(Tensor::from_slice(&result, a.shape(), device))
        }
        _ => unreachable!("dtype validated by caller"),
    }
}

// ============================================================================
// Public API
// ============================================================================

/// Apply a unary scalar function element-wise over a tensor.
pub fn apply_unary<F>(
    x: &Tensor<CpuRuntime>,
    device: &CpuDevice,
    f: F,
) -> Result<Tensor<CpuRuntime>>
where
    F: Fn(f64) -> f64,
{
    apply_unary_via_f64(x, device, f)
}

/// Apply a binary scalar function element-wise over two tensors.
///
/// Both tensors must have matching shapes (broadcasting not supported).
pub fn apply_binary<F>(
    a: &Tensor<CpuRuntime>,
    b: &Tensor<CpuRuntime>,
    device: &CpuDevice,
    f: F,
) -> Result<Tensor<CpuRuntime>>
where
    F: Fn(f64, f64) -> f64,
{
    apply_binary_via_f64(a, b, device, f)
}

/// Apply a ternary scalar function element-wise over three tensors.
///
/// All tensors must have matching shapes (broadcasting not supported).
pub fn apply_ternary<F>(
    a: &Tensor<CpuRuntime>,
    b: &Tensor<CpuRuntime>,
    c: &Tensor<CpuRuntime>,
    device: &CpuDevice,
    f: F,
) -> Result<Tensor<CpuRuntime>>
where
    F: Fn(f64, f64, f64) -> f64,
{
    if a.shape() != b.shape() {
        return Err(Error::ShapeMismatch {
            expected: a.shape().to_vec(),
            got: b.shape().to_vec(),
        });
    }
    if a.shape() != c.shape() {
        return Err(Error::ShapeMismatch {
            expected: a.shape().to_vec(),
            got: c.shape().to_vec(),
        });
    }

    match a.dtype() {
        DType::F32 => {
            let a_data: Vec<f32> = a.to_vec();
            let b_data: Vec<f32> = b.to_vec();
            let c_data: Vec<f32> = c.to_vec();
            let result: Vec<f32> = a_data
                .iter()
                .zip(b_data.iter())
                .zip(c_data.iter())
                .map(|((&av, &bv), &cv)| f(av as f64, bv as f64, cv as f64) as f32)
                .collect();
            Ok(Tensor::from_slice(&result, a.shape(), device))
        }
        DType::F64 => {
            let a_data: Vec<f64> = a.to_vec();
            let b_data: Vec<f64> = b.to_vec();
            let c_data: Vec<f64> = c.to_vec();
            let result: Vec<f64> = a_data
                .iter()
                .zip(b_data.iter())
                .zip(c_data.iter())
                .map(|((&av, &bv), &cv)| f(av, bv, cv))
                .collect();
            Ok(Tensor::from_slice(&result, a.shape(), device))
        }
        _ => unreachable!("dtype validated by caller"),
    }
}

/// Apply a unary scalar function with one extra i32 parameter.
pub fn apply_unary_with_int<F>(
    x: &Tensor<CpuRuntime>,
    device: &CpuDevice,
    n: i32,
    f: F,
) -> Result<Tensor<CpuRuntime>>
where
    F: Fn(i32, f64) -> f64,
{
    apply_unary_via_f64(x, device, |v| f(n, v))
}

/// Apply a unary scalar function with two extra i32 parameters.
pub fn apply_unary_with_two_ints<F>(
    x: &Tensor<CpuRuntime>,
    device: &CpuDevice,
    n: i32,
    m: i32,
    f: F,
) -> Result<Tensor<CpuRuntime>>
where
    F: Fn(i32, i32, f64) -> f64,
{
    apply_unary_via_f64(x, device, |v| f(n, m, v))
}

/// Apply a binary scalar function with two extra i32 parameters (for sph_harm).
pub fn apply_binary_with_two_ints<F>(
    a: &Tensor<CpuRuntime>,
    b: &Tensor<CpuRuntime>,
    device: &CpuDevice,
    n: i32,
    m: i32,
    f: F,
) -> Result<Tensor<CpuRuntime>>
where
    F: Fn(i32, i32, f64, f64) -> f64,
{
    apply_binary_via_f64(a, b, device, |av, bv| f(n, m, av, bv))
}

/// Apply a unary scalar function with three extra f64 parameters (for hyp2f1).
pub fn apply_unary_with_three_f64s<F>(
    z: &Tensor<CpuRuntime>,
    device: &CpuDevice,
    a: f64,
    b: f64,
    c: f64,
    f: F,
) -> Result<Tensor<CpuRuntime>>
where
    F: Fn(f64, f64, f64, f64) -> f64,
{
    apply_unary_via_f64(z, device, |v| f(a, b, c, v))
}

/// Apply a unary scalar function with two extra f64 parameters (for hyp1f1).
pub fn apply_unary_with_two_f64s<F>(
    z: &Tensor<CpuRuntime>,
    device: &CpuDevice,
    a: f64,
    b: f64,
    f: F,
) -> Result<Tensor<CpuRuntime>>
where
    F: Fn(f64, f64, f64) -> f64,
{
    apply_unary_via_f64(z, device, |v| f(a, b, v))
}