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

//! Reduce operations for CUDA runtime
use crate::error::Result;
use crate::ops::{ReduceOps, ScalarOps};
use crate::runtime::cuda::kernels::AccumulationPrecision;
use crate::runtime::cuda::ops::helpers::native_reduce_op;
use crate::runtime::cuda::{CudaClient, CudaRuntime};
use crate::tensor::Tensor;

/// Normalize dims for reduction: empty means all dimensions
#[inline]
fn normalize_reduce_dims(dims: &[usize], ndim: usize) -> Vec<usize> {
    if dims.is_empty() {
        (0..ndim).collect()
    } else {
        dims.to_vec()
    }
}

impl ReduceOps<CudaRuntime> for CudaClient {
    fn sum(
        &self,
        a: &Tensor<CudaRuntime>,
        dims: &[usize],
        keepdim: bool,
    ) -> Result<Tensor<CudaRuntime>> {
        let dims = normalize_reduce_dims(dims, a.shape().len());
        native_reduce_op(self, a, "sum", &dims, keepdim, None)
    }

    fn sum_with_precision(
        &self,
        a: &Tensor<CudaRuntime>,
        dims: &[usize],
        keepdim: bool,
        precision: AccumulationPrecision,
    ) -> Result<Tensor<CudaRuntime>> {
        native_reduce_op(self, a, "sum", dims, keepdim, Some(precision))
    }

    fn mean(
        &self,
        a: &Tensor<CudaRuntime>,
        dims: &[usize],
        keepdim: bool,
    ) -> Result<Tensor<CudaRuntime>> {
        let count: usize = if dims.is_empty() {
            a.numel()
        } else {
            dims.iter().map(|&d| a.shape()[d]).product()
        };

        let dims = normalize_reduce_dims(dims, a.shape().len());
        let sum_result = self.sum(a, &dims, keepdim)?;
        self.div_scalar(&sum_result, count as f64)
    }

    fn max(
        &self,
        a: &Tensor<CudaRuntime>,
        dims: &[usize],
        keepdim: bool,
    ) -> Result<Tensor<CudaRuntime>> {
        let dims = normalize_reduce_dims(dims, a.shape().len());
        native_reduce_op(self, a, "max", &dims, keepdim, None)
    }

    fn max_with_precision(
        &self,
        a: &Tensor<CudaRuntime>,
        dims: &[usize],
        keepdim: bool,
        precision: AccumulationPrecision,
    ) -> Result<Tensor<CudaRuntime>> {
        native_reduce_op(self, a, "max", dims, keepdim, Some(precision))
    }

    fn min(
        &self,
        a: &Tensor<CudaRuntime>,
        dims: &[usize],
        keepdim: bool,
    ) -> Result<Tensor<CudaRuntime>> {
        let dims = normalize_reduce_dims(dims, a.shape().len());
        native_reduce_op(self, a, "min", &dims, keepdim, None)
    }

    fn min_with_precision(
        &self,
        a: &Tensor<CudaRuntime>,
        dims: &[usize],
        keepdim: bool,
        precision: AccumulationPrecision,
    ) -> Result<Tensor<CudaRuntime>> {
        native_reduce_op(self, a, "min", dims, keepdim, Some(precision))
    }

    fn prod(
        &self,
        a: &Tensor<CudaRuntime>,
        dims: &[usize],
        keepdim: bool,
    ) -> Result<Tensor<CudaRuntime>> {
        let dims = normalize_reduce_dims(dims, a.shape().len());
        native_reduce_op(self, a, "prod", &dims, keepdim, None)
    }

    fn prod_with_precision(
        &self,
        a: &Tensor<CudaRuntime>,
        dims: &[usize],
        keepdim: bool,
        precision: AccumulationPrecision,
    ) -> Result<Tensor<CudaRuntime>> {
        native_reduce_op(self, a, "prod", dims, keepdim, Some(precision))
    }

    fn any(
        &self,
        a: &Tensor<CudaRuntime>,
        dims: &[usize],
        keepdim: bool,
    ) -> Result<Tensor<CudaRuntime>> {
        let dims = normalize_reduce_dims(dims, a.shape().len());
        native_reduce_op(self, a, "any", &dims, keepdim, None)
    }

    fn all(
        &self,
        a: &Tensor<CudaRuntime>,
        dims: &[usize],
        keepdim: bool,
    ) -> Result<Tensor<CudaRuntime>> {
        let dims = normalize_reduce_dims(dims, a.shape().len());
        native_reduce_op(self, a, "all", &dims, keepdim, None)
    }
}