1 #![feature(test)]
2 
3 extern crate test;
4 use test::Bencher;
5 
6 use ndarray::prelude::*;
7 use ndarray::NdProducer;
8 
9 #[bench]
chunk2x2_iter_sum(bench: &mut Bencher)10 fn chunk2x2_iter_sum(bench: &mut Bencher) {
11     let a = Array::<f32, _>::zeros((256, 256));
12     let chunksz = (2, 2);
13     let mut sum = Array::zeros(a.exact_chunks(chunksz).raw_dim());
14     bench.iter(|| {
15         azip!((a in a.exact_chunks(chunksz), sum in &mut sum) {
16             *sum = a.iter().sum::<f32>();
17         });
18     });
19 }
20 
21 #[bench]
chunk2x2_sum(bench: &mut Bencher)22 fn chunk2x2_sum(bench: &mut Bencher) {
23     let a = Array::<f32, _>::zeros((256, 256));
24     let chunksz = (2, 2);
25     let mut sum = Array::zeros(a.exact_chunks(chunksz).raw_dim());
26     bench.iter(|| {
27         azip!((a in a.exact_chunks(chunksz), sum in &mut sum) {
28             *sum = a.sum();
29         });
30     });
31 }
32 
33 #[bench]
chunk2x2_sum_get1(bench: &mut Bencher)34 fn chunk2x2_sum_get1(bench: &mut Bencher) {
35     let a = Array::<f32, _>::zeros((256, 256));
36     let chunksz = (2, 2);
37     let mut sum = Array::<f32, _>::zeros(a.exact_chunks(chunksz).raw_dim());
38     bench.iter(|| {
39         let (m, n) = a.dim();
40         for i in 0..m {
41             for j in 0..n {
42                 sum[[i / 2, j / 2]] += a[[i, j]];
43             }
44         }
45     });
46 }
47 
48 #[bench]
chunk2x2_sum_uget1(bench: &mut Bencher)49 fn chunk2x2_sum_uget1(bench: &mut Bencher) {
50     let a = Array::<f32, _>::zeros((256, 256));
51     let chunksz = (2, 2);
52     let mut sum = Array::<f32, _>::zeros(a.exact_chunks(chunksz).raw_dim());
53     bench.iter(|| {
54         let (m, n) = a.dim();
55         for i in 0..m {
56             for j in 0..n {
57                 unsafe {
58                     *sum.uget_mut([i / 2, j / 2]) += *a.uget([i, j]);
59                 }
60             }
61         }
62     });
63 }
64 
65 #[bench]
66 #[allow(clippy::identity_op)]
chunk2x2_sum_get2(bench: &mut Bencher)67 fn chunk2x2_sum_get2(bench: &mut Bencher) {
68     let a = Array::<f32, _>::zeros((256, 256));
69     let chunksz = (2, 2);
70     let mut sum = Array::<f32, _>::zeros(a.exact_chunks(chunksz).raw_dim());
71     bench.iter(|| {
72         let (m, n) = sum.dim();
73         for i in 0..m {
74             for j in 0..n {
75                 sum[[i, j]] += a[[i * 2 + 0, j * 2 + 0]];
76                 sum[[i, j]] += a[[i * 2 + 0, j * 2 + 1]];
77                 sum[[i, j]] += a[[i * 2 + 1, j * 2 + 1]];
78                 sum[[i, j]] += a[[i * 2 + 1, j * 2 + 0]];
79             }
80         }
81     });
82 }
83