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
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

//! CUDA runtime operation helpers and TensorOps implementation.
//!
//! Trait implementations (CompareOps, ScalarOps, LogicalOps, etc.) live in `ops/cuda/`.
//! This module provides shared helpers, kernel launchers, and TensorOps.

pub(crate) mod helpers;
mod statistics;
mod tensor;

#[cfg(test)]
mod tests {
    use crate::ops::{
        ActivationOps, BinaryOps, IndexingOps, MatmulOps, NormalizationOps, ReduceOps,
    };
    use crate::runtime::Runtime;
    use crate::runtime::cuda::{CudaDevice, CudaRuntime, is_cuda_available};
    use crate::tensor::Tensor;

    #[test]
    fn test_cuda_tensor_add() {
        if !is_cuda_available() {
            return;
        }
        let device = CudaDevice::new(0);
        let client = CudaRuntime::default_client(&device);

        let a = Tensor::<CudaRuntime>::from_slice(&[1.0f32, 2.0, 3.0, 4.0], &[2, 2], &device);
        let b = Tensor::<CudaRuntime>::from_slice(&[5.0f32, 6.0, 7.0, 8.0], &[2, 2], &device);

        let c = client.add(&a, &b).unwrap();

        assert_eq!(c.shape(), &[2, 2]);
        let result: Vec<f32> = c.to_vec();
        assert_eq!(result, [6.0, 8.0, 10.0, 12.0]);
    }

    #[test]
    fn test_cuda_tensor_matmul_2x2() {
        if !is_cuda_available() {
            return;
        }
        let device = CudaDevice::new(0);
        let client = CudaRuntime::default_client(&device);

        let a = Tensor::<CudaRuntime>::from_slice(&[1.0f32, 2.0, 3.0, 4.0], &[2, 2], &device);
        let b = Tensor::<CudaRuntime>::from_slice(&[5.0f32, 6.0, 7.0, 8.0], &[2, 2], &device);

        let c = client.matmul(&a, &b).unwrap();

        assert_eq!(c.shape(), &[2, 2]);
        let result: Vec<f32> = c.to_vec();
        assert_eq!(result, [19.0, 22.0, 43.0, 50.0]);
    }

    #[test]
    fn test_cuda_tensor_matmul_3x2_2x4() {
        if !is_cuda_available() {
            return;
        }
        let device = CudaDevice::new(0);
        let client = CudaRuntime::default_client(&device);

        let a =
            Tensor::<CudaRuntime>::from_slice(&[1.0f32, 2.0, 3.0, 4.0, 5.0, 6.0], &[3, 2], &device);
        let b = Tensor::<CudaRuntime>::from_slice(
            &[1.0f32, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0],
            &[2, 4],
            &device,
        );

        let c = client.matmul(&a, &b).unwrap();

        assert_eq!(c.shape(), &[3, 4]);
        let result: Vec<f32> = c.to_vec();
        assert_eq!(
            result,
            [
                11.0, 14.0, 17.0, 20.0, 23.0, 30.0, 37.0, 44.0, 35.0, 46.0, 57.0, 68.0
            ]
        );
    }

    #[test]
    fn test_cuda_tensor_relu() {
        if !is_cuda_available() {
            return;
        }
        let device = CudaDevice::new(0);
        let client = CudaRuntime::default_client(&device);

        let a = Tensor::<CudaRuntime>::from_slice(&[-1.0f32, 0.0, 1.0, -2.0], &[4], &device);
        let b = client.relu(&a).unwrap();

        let result: Vec<f32> = b.to_vec();
        assert_eq!(result, [0.0, 0.0, 1.0, 0.0]);
    }

    #[test]
    fn test_cuda_tensor_sum() {
        if !is_cuda_available() {
            return;
        }
        let device = CudaDevice::new(0);
        let client = CudaRuntime::default_client(&device);

        let a =
            Tensor::<CudaRuntime>::from_slice(&[1.0f32, 2.0, 3.0, 4.0, 5.0, 6.0], &[2, 3], &device);

        let b = client.sum(&a, &[1], false).unwrap();

        assert_eq!(b.shape(), &[2]);
        let result: Vec<f32> = b.to_vec();
        assert_eq!(result, [6.0, 15.0]);
    }

    #[test]
    fn test_cuda_tensor_silu() {
        if !is_cuda_available() {
            return;
        }
        let device = CudaDevice::new(0);
        let client = CudaRuntime::default_client(&device);

        let a = Tensor::<CudaRuntime>::from_slice(&[-2.0f32, -1.0, 0.0, 1.0, 2.0], &[5], &device);
        let b = client.silu(&a).unwrap();

        let result: Vec<f32> = b.to_vec();
        // SiLU(x) = x / (1 + exp(-x))
        // SiLU(0) = 0
        // SiLU(1) ≈ 0.731
        // SiLU(-1) ≈ -0.269
        assert!((result[2] - 0.0).abs() < 1e-5); // SiLU(0) = 0
        assert!((result[3] - 0.7310586).abs() < 1e-4); // SiLU(1) ≈ 0.731
        assert!((result[1] - (-0.2689414)).abs() < 1e-4); // SiLU(-1) ≈ -0.269
    }

    #[test]
    fn test_cuda_tensor_gelu() {
        if !is_cuda_available() {
            return;
        }
        let device = CudaDevice::new(0);
        let client = CudaRuntime::default_client(&device);

        let a = Tensor::<CudaRuntime>::from_slice(&[-2.0f32, -1.0, 0.0, 1.0, 2.0], &[5], &device);
        let b = client.gelu(&a).unwrap();

        let result: Vec<f32> = b.to_vec();
        // GELU(0) = 0
        // GELU is approximately x for large positive x
        // GELU is approximately 0 for large negative x
        assert!((result[2] - 0.0).abs() < 1e-5); // GELU(0) = 0
        assert!((result[3] - 0.8413).abs() < 0.01); // GELU(1) ≈ 0.841
        assert!((result[4] - 1.9545).abs() < 0.01); // GELU(2) ≈ 1.955
    }

    #[test]
    fn test_cuda_tensor_rms_norm() {
        if !is_cuda_available() {
            return;
        }
        let device = CudaDevice::new(0);
        let client = CudaRuntime::default_client(&device);

        // Input: 2 rows, 4 features each
        let input = Tensor::<CudaRuntime>::from_slice(
            &[1.0f32, 2.0, 3.0, 4.0, 2.0, 4.0, 6.0, 8.0],
            &[2, 4],
            &device,
        );
        let weight = Tensor::<CudaRuntime>::from_slice(&[1.0f32, 1.0, 1.0, 1.0], &[4], &device);

        let out = client.rms_norm(&input, &weight, 1e-5).unwrap();
        let result: Vec<f32> = out.to_vec();

        // Row 1: [1, 2, 3, 4], RMS = sqrt((1+4+9+16)/4) = sqrt(30/4) = sqrt(7.5) ≈ 2.739
        let rms1 = (30.0f32 / 4.0 + 1e-5).sqrt();
        assert!((result[0] - 1.0 / rms1).abs() < 1e-3); // Wider tolerance for GPU
        assert!((result[1] - 2.0 / rms1).abs() < 1e-3);
        assert!((result[2] - 3.0 / rms1).abs() < 1e-3);
        assert!((result[3] - 4.0 / rms1).abs() < 1e-3);

        // Row 2: [2, 4, 6, 8]
        let rms2 = (120.0f32 / 4.0 + 1e-5).sqrt();
        assert!((result[4] - 2.0 / rms2).abs() < 1e-3);
    }

    #[test]
    fn test_cuda_tensor_layer_norm() {
        if !is_cuda_available() {
            return;
        }
        let device = CudaDevice::new(0);
        let client = CudaRuntime::default_client(&device);

        // Input: 2 rows, 4 features each
        let input = Tensor::<CudaRuntime>::from_slice(
            &[1.0f32, 2.0, 3.0, 4.0, 2.0, 4.0, 6.0, 8.0],
            &[2, 4],
            &device,
        );
        let weight = Tensor::<CudaRuntime>::from_slice(&[1.0f32, 1.0, 1.0, 1.0], &[4], &device);
        let bias = Tensor::<CudaRuntime>::from_slice(&[0.0f32, 0.0, 0.0, 0.0], &[4], &device);

        let out = client.layer_norm(&input, &weight, &bias, 1e-5).unwrap();
        let result: Vec<f32> = out.to_vec();

        // Row 1: [1, 2, 3, 4], mean = 2.5, var = 1.25, std = 1.118
        let mean1 = 2.5f32;
        let var1 = ((1.0 - mean1).powi(2)
            + (2.0 - mean1).powi(2)
            + (3.0 - mean1).powi(2)
            + (4.0 - mean1).powi(2))
            / 4.0;
        let std1 = (var1 + 1e-5).sqrt();
        assert!((result[0] - (1.0 - mean1) / std1).abs() < 1e-3); // Wider tolerance for GPU
        assert!((result[1] - (2.0 - mean1) / std1).abs() < 1e-3);
        assert!((result[2] - (3.0 - mean1) / std1).abs() < 1e-3);
        assert!((result[3] - (4.0 - mean1) / std1).abs() < 1e-3);

        // Verify normalized outputs sum to approximately 0 (zero-centered)
        let row1_sum: f32 = result[0..4].iter().sum();
        assert!(row1_sum.abs() < 1e-3);
    }

    #[test]
    fn test_cuda_tensor_argmax() {
        if !is_cuda_available() {
            return;
        }
        let device = CudaDevice::new(0);
        let client = CudaRuntime::default_client(&device);

        // 2D tensor: [[1, 5, 3], [4, 2, 6]]
        let a =
            Tensor::<CudaRuntime>::from_slice(&[1.0f32, 5.0, 3.0, 4.0, 2.0, 6.0], &[2, 3], &device);

        // argmax along dim=1 (find max index in each row)
        let out = client.argmax(&a, 1, false).unwrap();
        let result: Vec<i64> = out.to_vec();
        assert_eq!(out.shape(), &[2]);
        assert_eq!(result, [1, 2]); // Row 0: max at index 1 (5.0), Row 1: max at index 2 (6.0)

        // argmax along dim=0 (find max index in each column)
        let out = client.argmax(&a, 0, false).unwrap();
        let result: Vec<i64> = out.to_vec();
        assert_eq!(out.shape(), &[3]);
        assert_eq!(result, [1, 0, 1]); // Col 0: max at 1 (4.0), Col 1: max at 0 (5.0), Col 2: max at 1 (6.0)

        // Test keepdim=true
        let out = client.argmax(&a, 1, true).unwrap();
        let result: Vec<i64> = out.to_vec();
        assert_eq!(out.shape(), &[2, 1]);
        assert_eq!(result, [1, 2]);
    }

    #[test]
    fn test_cuda_tensor_argmin() {
        if !is_cuda_available() {
            return;
        }
        let device = CudaDevice::new(0);
        let client = CudaRuntime::default_client(&device);

        // 2D tensor: [[1, 5, 3], [4, 2, 6]]
        let a =
            Tensor::<CudaRuntime>::from_slice(&[1.0f32, 5.0, 3.0, 4.0, 2.0, 6.0], &[2, 3], &device);

        // argmin along dim=1 (find min index in each row)
        let out = client.argmin(&a, 1, false).unwrap();
        let result: Vec<i64> = out.to_vec();
        assert_eq!(out.shape(), &[2]);
        assert_eq!(result, [0, 1]); // Row 0: min at index 0 (1.0), Row 1: min at index 1 (2.0)

        // argmin along dim=0 (find min index in each column)
        let out = client.argmin(&a, 0, false).unwrap();
        let result: Vec<i64> = out.to_vec();
        assert_eq!(out.shape(), &[3]);
        assert_eq!(result, [0, 1, 0]); // Col 0: min at 0 (1.0), Col 1: min at 1 (2.0), Col 2: min at 0 (3.0)

        // Test keepdim=true
        let out = client.argmin(&a, 1, true).unwrap();
        let result: Vec<i64> = out.to_vec();
        assert_eq!(out.shape(), &[2, 1]);
        assert_eq!(result, [0, 1]);
    }
}