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

//! Quantile, percentile, and median operations for CPU runtime.

use super::super::helpers::{dispatch_dtype, ensure_contiguous};
use super::super::sort::sort_impl;
use super::super::{CpuClient, CpuRuntime};
use super::{Interpolation, quantile_kernel};
use crate::error::{Error, Result};
use crate::ops::{compute_reduce_strides, reduce_dim_output_shape};
use crate::runtime::normalize_dim;
use crate::tensor::Tensor;

/// Compute quantile along a dimension.
///
/// # Arguments
///
/// * `client` - The CPU runtime client
/// * `a` - Input tensor
/// * `q` - Quantile to compute, must be in [0.0, 1.0]
/// * `dim` - Dimension to reduce along (None = flatten first)
/// * `keepdim` - Whether to keep the reduced dimension
/// * `interpolation` - Interpolation method ("linear", "lower", "higher", "nearest", "midpoint")
///
/// # Returns
///
/// Tensor containing the quantile values.
///
/// # Errors
///
/// - `InvalidArgument` if q is not in [0, 1]
/// - `InvalidArgument` if interpolation method is invalid
/// - `InvalidDimension` if dim is out of bounds
pub fn quantile_impl(
    client: &CpuClient,
    a: &Tensor<CpuRuntime>,
    q: f64,
    dim: Option<isize>,
    keepdim: bool,
    interpolation: &str,
) -> Result<Tensor<CpuRuntime>> {
    // Validate q is in [0, 1]
    if !(0.0..=1.0).contains(&q) {
        return Err(Error::InvalidArgument {
            arg: "q",
            reason: format!("Quantile q must be in [0, 1], got {}", q),
        });
    }

    let interp = Interpolation::parse(interpolation)?;
    let dtype = a.dtype();

    // Handle None dim: flatten to 1D first
    if dim.is_none() {
        let numel = a.numel();
        if numel == 0 {
            let out_shape = if keepdim { vec![1; a.ndim()] } else { vec![] };
            return Ok(Tensor::<CpuRuntime>::empty(
                &out_shape,
                dtype,
                &client.device,
            ));
        }

        let flat = a.reshape(&[numel])?;
        return quantile_impl(client, &flat, q, Some(0), keepdim, interpolation);
    }

    let dim_val = dim.unwrap();
    let shape = a.shape();
    let ndim = shape.len();

    if ndim == 0 {
        return Ok(a.clone());
    }

    let dim_idx = normalize_dim(dim_val, ndim)?;
    let dim_size = shape[dim_idx];

    if dim_size == 0 {
        let out_shape = reduce_dim_output_shape(shape, dim_idx, keepdim);
        return Ok(Tensor::<CpuRuntime>::empty(
            &out_shape,
            dtype,
            &client.device,
        ));
    }

    // Sort along the dimension
    let sorted = sort_impl(client, a, dim_val, false)?;

    // Compute output shape and strides
    let out_shape = reduce_dim_output_shape(shape, dim_idx, keepdim);
    let out = Tensor::<CpuRuntime>::empty(&out_shape, dtype, &client.device);
    let (outer_size, reduce_size, inner_size) = compute_reduce_strides(shape, dim_idx);

    let sorted_contig = ensure_contiguous(&sorted);
    let sorted_ptr = sorted_contig.ptr();
    let out_ptr = out.ptr();

    // Dispatch to typed kernel
    dispatch_dtype!(dtype, T => {
        unsafe {
            quantile_kernel::<T>(
                sorted_ptr as *const T,
                out_ptr as *mut T,
                outer_size,
                reduce_size,
                inner_size,
                q,
                interp,
            );
        }
    }, "quantile");

    Ok(out)
}

/// Compute percentile (quantile * 100) along a dimension.
///
/// # Arguments
///
/// * `client` - The CPU runtime client
/// * `a` - Input tensor
/// * `p` - Percentile to compute, must be in [0.0, 100.0]
/// * `dim` - Dimension to reduce along (None = flatten first)
/// * `keepdim` - Whether to keep the reduced dimension
///
/// # Returns
///
/// Tensor containing the percentile values.
pub fn percentile_impl(
    client: &CpuClient,
    a: &Tensor<CpuRuntime>,
    p: f64,
    dim: Option<isize>,
    keepdim: bool,
) -> Result<Tensor<CpuRuntime>> {
    if !(0.0..=100.0).contains(&p) {
        return Err(Error::InvalidArgument {
            arg: "p",
            reason: format!("Percentile p must be in [0, 100], got {}", p),
        });
    }

    quantile_impl(client, a, p / 100.0, dim, keepdim, "linear")
}

/// Compute median (50th percentile) along a dimension.
///
/// # Arguments
///
/// * `client` - The CPU runtime client
/// * `a` - Input tensor
/// * `dim` - Dimension to reduce along (None = flatten first)
/// * `keepdim` - Whether to keep the reduced dimension
///
/// # Returns
///
/// Tensor containing the median values.
pub fn median_impl(
    client: &CpuClient,
    a: &Tensor<CpuRuntime>,
    dim: Option<isize>,
    keepdim: bool,
) -> Result<Tensor<CpuRuntime>> {
    quantile_impl(client, a, 0.5, dim, keepdim, "linear")
}