custos_math/ops/
max.rs

1use crate::Matrix;
2use custos::{number::Number, CDatatype, Device, MainMemory, CPU};
3
4#[cfg(feature = "cpu")]
5use custos::cache::Cache;
6
7#[cfg(feature = "cuda")]
8use crate::{cu_to_cpu_s, cu_to_cpu_scalar};
9#[cfg(feature = "cuda")]
10use custos::CUDA;
11
12#[cfg(feature = "opencl")]
13use super::{cl_to_cpu_s, cl_to_cpu_scalar};
14#[cfg(feature = "opencl")]
15use custos::OpenCL;
16
17impl<'a, T, D: MaxOps<T>> Matrix<'a, T, D> {
18    #[inline]
19    pub fn max(&self) -> T {
20        self.device().max(self)
21    }
22
23    #[inline]
24    pub fn max_rows(&self) -> Matrix<'a, T, D> {
25        self.device().max_rows(self)
26    }
27
28    #[inline]
29    pub fn max_cols(&self) -> Matrix<'a, T, D> {
30        self.device().max_cols(self)
31    }
32}
33
34pub trait MaxOps<T, D: Device = Self>: Device {
35    fn max(&self, x: &Matrix<T, D>) -> T;
36    fn max_rows(&self, x: &Matrix<T, D>) -> Matrix<T, Self>;
37    fn max_cols(&self, x: &Matrix<T, D>) -> Matrix<T, Self>;
38}
39
40// TODO: refactor this into own methods
41#[cfg(feature = "cpu")]
42impl<T: Copy + PartialOrd, D: MainMemory> MaxOps<T, D> for CPU {
43    fn max(&self, x: &Matrix<T, D>) -> T {
44        let mut max = x[0];
45
46        for value in x.iter() {
47            if *value > max {
48                max = *value;
49            }
50        }
51        max
52    }
53
54    fn max_rows(&self, x: &Matrix<T, D>) -> Matrix<T> {
55        let mut out = Cache::get(self, x.cols(), x.node.idx);
56
57        let data = x.as_slice();
58        let max_rows = out.as_mut_slice();
59
60        max_rows.copy_from_slice(&data[..max_rows.len()]);
61
62        for idx in 0..x.rows() {
63            let index = idx * x.cols();
64            let row = &data[index..index + x.cols()];
65
66            for (i, data) in row.iter().enumerate() {
67                if data > &max_rows[i] {
68                    max_rows[i] = *data;
69                }
70            }
71        }
72        (out, 1, x.cols()).into()
73    }
74
75    fn max_cols(&self, x: &Matrix<T, D>) -> Matrix<T> {
76        let data = x.as_slice();
77        let mut y = Cache::get(self, x.rows(), x.node.idx);
78
79        let max_cols = y.as_mut_slice();
80
81        for (idx, max_cols_val) in max_cols.iter_mut().enumerate().take(x.rows()) {
82            let index = idx * x.cols();
83            let row = &data[index..index + x.cols()];
84
85            let mut max = row[0];
86
87            for data in row {
88                if data > &max {
89                    max = *data;
90                }
91            }
92            *max_cols_val = max;
93        }
94        (y, x.rows(), 1).into()
95    }
96}
97
98#[cfg(feature = "opencl")]
99impl<T: CDatatype> MaxOps<T> for OpenCL {
100    fn max(&self, x: &Matrix<T, Self>) -> T {
101        cl_to_cpu_scalar(self, x, |device, x| device.max(x))
102    }
103
104    fn max_rows(&self, x: &Matrix<T, Self>) -> Matrix<T, Self> {
105        cl_to_cpu_s(self, x, |device, x| device.max_rows(x))
106    }
107
108    fn max_cols(&self, x: &Matrix<T, Self>) -> Matrix<T, Self> {
109        cl_to_cpu_s(self, x, |device, x| device.max_cols(x))
110    }
111}
112
113#[cfg(feature = "cuda")]
114impl<T: Number> MaxOps<T> for CUDA {
115    fn max(&self, x: &Matrix<T, CUDA>) -> T {
116        cu_to_cpu_scalar(x, |cpu, x| cpu.max(&x))
117    }
118
119    fn max_rows(&self, x: &Matrix<T, CUDA>) -> Matrix<T, CUDA> {
120        cu_to_cpu_s(self, x, |cpu, x| cpu.max_rows(&x))
121    }
122
123    fn max_cols(&self, x: &Matrix<T, CUDA>) -> Matrix<T, CUDA> {
124        cu_to_cpu_s(self, x, |cpu, x| cpu.max_cols(&x))
125    }
126}
custos_math/ops/max.rs

custos_math/ops/
max.rs
