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

//! CUDA implementation of distance operations.

use crate::error::Result;
use crate::ops::distance_common::*;
use crate::ops::{DistanceMetric, DistanceOps};
use crate::runtime::cuda::{CudaClient, CudaRuntime, kernels};
use crate::tensor::Tensor;

impl DistanceOps<CudaRuntime> for CudaClient {
    fn cdist(
        &self,
        x: &Tensor<CudaRuntime>,
        y: &Tensor<CudaRuntime>,
        metric: DistanceMetric,
    ) -> Result<Tensor<CudaRuntime>> {
        let x_shape = x.shape();
        let y_shape = y.shape();

        // Validate inputs using shared validators
        validate_2d_tensor(x_shape, "x", "cdist")?;
        validate_2d_tensor(y_shape, "y", "cdist")?;
        validate_same_dimension(x_shape, y_shape, "cdist")?;

        let dtype = x.dtype();
        validate_float_dtype(dtype, "cdist")?;
        validate_same_dtype(dtype, y.dtype(), "cdist")?;

        let n = x_shape[0];
        let m = y_shape[0];
        let d = x_shape[1];

        // Handle empty tensors
        if n == 0 || m == 0 {
            return Ok(Tensor::<CudaRuntime>::empty(&[n, m], dtype, &self.device));
        }

        // Ensure contiguous
        let x = x.contiguous();
        let y = y.contiguous();

        let out = Tensor::<CudaRuntime>::empty(&[n, m], dtype, &self.device);

        unsafe {
            kernels::launch_cdist(
                &self.context,
                &self.stream,
                self.device.index,
                dtype,
                x.ptr(),
                y.ptr(),
                out.ptr(),
                n,
                m,
                d,
                metric,
            )?;
        }

        Ok(out)
    }

    fn pdist(
        &self,
        x: &Tensor<CudaRuntime>,
        metric: DistanceMetric,
    ) -> Result<Tensor<CudaRuntime>> {
        let x_shape = x.shape();

        // Validate inputs using shared validators
        validate_2d_tensor(x_shape, "x", "pdist")?;

        let n = x_shape[0];
        let d = x_shape[1];

        validate_min_points(n, 2, "x", "pdist")?;

        let dtype = x.dtype();
        validate_float_dtype(dtype, "pdist")?;

        // Output size: n*(n-1)/2
        let out_size = n * (n - 1) / 2;

        // Ensure contiguous
        let x = x.contiguous();

        let out = Tensor::<CudaRuntime>::empty(&[out_size], dtype, &self.device);

        unsafe {
            kernels::launch_pdist(
                &self.context,
                &self.stream,
                self.device.index,
                dtype,
                x.ptr(),
                out.ptr(),
                n,
                d,
                metric,
            )?;
        }

        Ok(out)
    }

    fn squareform(&self, condensed: &Tensor<CudaRuntime>, n: usize) -> Result<Tensor<CudaRuntime>> {
        let cond_shape = condensed.shape();

        // Validate inputs using shared validators
        validate_1d_tensor(cond_shape, "condensed", "squareform")?;
        validate_condensed_length(cond_shape[0], n, "condensed", "squareform")?;

        let dtype = condensed.dtype();
        validate_float_dtype(dtype, "squareform")?;

        // Handle edge cases
        if n == 0 {
            return Ok(Tensor::<CudaRuntime>::empty(&[0, 0], dtype, &self.device));
        }
        if n == 1 {
            return Ok(Tensor::<CudaRuntime>::zeros(&[1, 1], dtype, &self.device));
        }

        // Ensure contiguous
        let condensed = condensed.contiguous();

        let out = Tensor::<CudaRuntime>::empty(&[n, n], dtype, &self.device);

        unsafe {
            kernels::launch_squareform(
                &self.context,
                &self.stream,
                self.device.index,
                dtype,
                condensed.ptr(),
                out.ptr(),
                n,
            )?;
        }

        Ok(out)
    }

    fn squareform_inverse(&self, square: &Tensor<CudaRuntime>) -> Result<Tensor<CudaRuntime>> {
        let sq_shape = square.shape();

        // Validate inputs using shared validators
        validate_2d_tensor(sq_shape, "square", "squareform_inverse")?;
        validate_square_matrix(sq_shape, "square", "squareform_inverse")?;

        let n = sq_shape[0];
        let dtype = square.dtype();
        validate_float_dtype(dtype, "squareform_inverse")?;

        // Handle edge cases
        if n == 0 || n == 1 {
            return Ok(Tensor::<CudaRuntime>::empty(&[0], dtype, &self.device));
        }

        // Ensure contiguous
        let square = square.contiguous();

        let out_size = n * (n - 1) / 2;
        let out = Tensor::<CudaRuntime>::empty(&[out_size], dtype, &self.device);

        unsafe {
            kernels::launch_squareform_inverse(
                &self.context,
                &self.stream,
                self.device.index,
                dtype,
                square.ptr(),
                out.ptr(),
                n,
            )?;
        }

        Ok(out)
    }
}