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

//! Statistical operations for CUDA runtime
use crate::error::Result;
use crate::ops::{BinaryOps, ReduceOps, ScalarOps, StatisticalOps, UnaryOps};
use crate::runtime::cuda::{CudaClient, CudaRuntime};
use crate::tensor::Tensor;

// Import helper functions from statistics module
use crate::runtime::cuda::ops::statistics::{
    histogram_impl, kurtosis_impl, median_impl, mode_impl, percentile_impl, quantile_impl,
    skew_impl,
};

impl StatisticalOps<CudaRuntime> for CudaClient {
    fn var(
        &self,
        a: &Tensor<CudaRuntime>,
        dims: &[usize],
        keepdim: bool,
        correction: usize,
    ) -> Result<Tensor<CudaRuntime>> {
        // Variance implementation using existing ops
        // var(x) = mean((x - mean(x))^2) * N / (N - correction)

        let shape = a.shape();

        // When dims is empty, reduce over all dimensions
        let actual_dims: Vec<usize> = if dims.is_empty() {
            (0..shape.len()).collect()
        } else {
            dims.to_vec()
        };

        // Compute count of elements being reduced
        let count: usize = if dims.is_empty() {
            a.numel()
        } else {
            dims.iter().map(|&d| shape[d]).product()
        };

        // Compute mean (mean already handles empty dims internally)
        let mean_val = self.mean(a, dims, true)?;

        // Compute (x - mean)
        let diff = self.sub(a, &mean_val)?;

        // Compute (x - mean)^2
        let diff_squared = self.square(&diff)?;

        // Compute sum of squared differences over all dims when dims is empty
        let sum_sq = self.sum(&diff_squared, &actual_dims, keepdim)?;

        // Divide by (N - correction)
        let divisor = (count.saturating_sub(correction)).max(1) as f64;
        self.div_scalar(&sum_sq, divisor)
    }

    fn std(
        &self,
        a: &Tensor<CudaRuntime>,
        dims: &[usize],
        keepdim: bool,
        correction: usize,
    ) -> Result<Tensor<CudaRuntime>> {
        // Standard deviation is sqrt of variance
        let variance = self.var(a, dims, keepdim, correction)?;
        self.sqrt(&variance)
    }

    fn quantile(
        &self,
        a: &Tensor<CudaRuntime>,
        q: f64,
        dim: Option<isize>,
        keepdim: bool,
        interpolation: &str,
    ) -> Result<Tensor<CudaRuntime>> {
        quantile_impl(self, a, q, dim, keepdim, interpolation)
    }

    fn percentile(
        &self,
        a: &Tensor<CudaRuntime>,
        p: f64,
        dim: Option<isize>,
        keepdim: bool,
    ) -> Result<Tensor<CudaRuntime>> {
        percentile_impl(self, a, p, dim, keepdim)
    }

    fn median(
        &self,
        a: &Tensor<CudaRuntime>,
        dim: Option<isize>,
        keepdim: bool,
    ) -> Result<Tensor<CudaRuntime>> {
        median_impl(self, a, dim, keepdim)
    }

    fn histogram(
        &self,
        a: &Tensor<CudaRuntime>,
        bins: usize,
        range: Option<(f64, f64)>,
    ) -> Result<(Tensor<CudaRuntime>, Tensor<CudaRuntime>)> {
        histogram_impl(self, a, bins, range)
    }

    fn cov(&self, a: &Tensor<CudaRuntime>, ddof: Option<usize>) -> Result<Tensor<CudaRuntime>> {
        use crate::algorithm::LinearAlgebraAlgorithms;
        <Self as LinearAlgebraAlgorithms<CudaRuntime>>::cov(self, a, ddof)
    }

    fn corrcoef(&self, a: &Tensor<CudaRuntime>) -> Result<Tensor<CudaRuntime>> {
        use crate::algorithm::LinearAlgebraAlgorithms;
        <Self as LinearAlgebraAlgorithms<CudaRuntime>>::corrcoef(self, a)
    }

    fn skew(
        &self,
        a: &Tensor<CudaRuntime>,
        dims: &[usize],
        keepdim: bool,
        correction: usize,
    ) -> Result<Tensor<CudaRuntime>> {
        skew_impl(self, a, dims, keepdim, correction)
    }

    fn kurtosis(
        &self,
        a: &Tensor<CudaRuntime>,
        dims: &[usize],
        keepdim: bool,
        correction: usize,
    ) -> Result<Tensor<CudaRuntime>> {
        kurtosis_impl(self, a, dims, keepdim, correction)
    }

    fn mode(
        &self,
        a: &Tensor<CudaRuntime>,
        dim: Option<isize>,
        keepdim: bool,
    ) -> Result<(Tensor<CudaRuntime>, Tensor<CudaRuntime>)> {
        mode_impl(self, a, dim, keepdim)
    }
}