strided-perm 0.11.0

Cache-efficient tensor permutation / transpose (HPTT-inspired).
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

use super::*;
use strided_view::StridedArray;

#[test]
fn test_copy_into_contiguous() {
    let src = StridedArray::<f64>::from_fn_col_major(&[2, 3], |idx| (idx[0] * 10 + idx[1]) as f64);
    let mut dst = StridedArray::<f64>::col_major(&[2, 3]);
    copy_into(&mut dst.view_mut(), &src.view(), Exec::serial()).unwrap();
    assert_eq!(dst.get(&[0, 0]), 0.0);
    assert_eq!(dst.get(&[1, 2]), 12.0);
}

#[test]
fn test_copy_into_transposed() {
    // src is row-major [3,2], dst is col-major [3,2]
    let src = StridedArray::<f64>::from_parts(
        vec![0.0, 1.0, 2.0, 3.0, 4.0, 5.0],
        &[3, 2],
        &[2, 1], // row-major
        0,
    )
    .unwrap();
    let mut dst = StridedArray::<f64>::col_major(&[3, 2]);
    copy_into(&mut dst.view_mut(), &src.view(), Exec::serial()).unwrap();
    assert_eq!(dst.get(&[0, 0]), 0.0);
    assert_eq!(dst.get(&[0, 1]), 1.0);
    assert_eq!(dst.get(&[1, 0]), 2.0);
    assert_eq!(dst.get(&[2, 1]), 5.0);
}

#[test]
fn test_copy_into_col_major_basic() {
    let src = StridedArray::<f64>::from_fn_col_major(&[4, 3], |idx| (idx[0] * 10 + idx[1]) as f64);
    let mut dst = StridedArray::<f64>::col_major(&[4, 3]);
    copy_into_col_major(&mut dst.view_mut(), &src.view(), Exec::serial()).unwrap();
    assert_eq!(dst.get(&[0, 0]), 0.0);
    assert_eq!(dst.get(&[3, 2]), 32.0);
}

#[test]
fn test_copy_into_col_major_permuted_src() {
    // src has scattered strides, dst is col-major
    let data: Vec<f64> = (0..24).map(|i| i as f64).collect();
    let src = StridedArray::<f64>::from_parts(
        data,
        &[2, 3, 4],
        &[12, 4, 1], // row-major
        0,
    )
    .unwrap();
    let mut dst = StridedArray::<f64>::col_major(&[2, 3, 4]);
    copy_into_col_major(&mut dst.view_mut(), &src.view(), Exec::serial()).unwrap();
    // Verify element-by-element
    for i in 0..2 {
        for j in 0..3 {
            for k in 0..4 {
                assert_eq!(
                    dst.get(&[i, j, k]),
                    src.get(&[i, j, k]),
                    "mismatch at [{},{},{}]",
                    i,
                    j,
                    k
                );
            }
        }
    }
}

#[test]
fn test_copy_into_grouped_high_rank_binary_permutation() {
    let dims = vec![2usize; 12];
    let total: usize = dims.iter().product();
    let src = StridedArray::<u32>::from_fn_col_major(&dims, |idx| {
        idx.iter()
            .enumerate()
            .map(|(axis, &coord)| (coord as u32) << axis)
            .sum::<u32>()
    });
    let perm = vec![2, 4, 5, 6, 0, 3, 7, 1, 8, 9, 10, 11];
    let src_view = src.view().permute(&perm).unwrap();
    let mut dst = StridedArray::<u32>::col_major(&dims);

    copy_into(&mut dst.view_mut(), &src_view, Exec::serial()).unwrap();

    for flat in 0..total {
        let mut rem = flat;
        let mut src_offset = 0usize;
        for &stride in src_view.strides() {
            let coord = rem & 1;
            rem >>= 1;
            src_offset += coord * stride as usize;
        }
        assert_eq!(dst.data()[flat], src.data()[src_offset], "flat={flat}");
    }
}

#[test]
fn test_try_fuse_group_empty() {
    assert_eq!(try_fuse_group(&[], &[]), Some((1, 0)));
}

#[test]
fn test_try_fuse_group_single() {
    assert_eq!(try_fuse_group(&[5], &[2]), Some((5, 2)));
}

#[test]
fn test_try_fuse_group_contiguous_row_major() {
    assert_eq!(try_fuse_group(&[3, 4], &[4, 1]), Some((12, 1)));
}

#[test]
fn test_try_fuse_group_contiguous_col_major() {
    assert_eq!(try_fuse_group(&[3, 4], &[1, 3]), Some((12, 1)));
}

#[test]
fn test_try_fuse_group_non_contiguous() {
    assert_eq!(try_fuse_group(&[3, 4], &[8, 1]), None);
}

#[test]
fn parallel_transpose_splits_the_full_tile_grid_for_a_short_row_extent() {
    // The source's unit-stride extent is the short side of the transpose, so
    // the whole kernel would be a single row tile over `block`, and the budget is
    // two. The parallel entry must split the
    // column tiles too; a row-only split would leave the work on one worker.
    // The result is checked element by element.
    let rows = 20000usize;
    let cols = 3usize;
    let data: Vec<f64> = (0..rows * cols).map(|i| i as f64).collect();
    let src = StridedArray::<f64>::from_parts(data, &[rows, cols], &[cols as isize, 1], 0).unwrap();
    let pool = threadteam::Pool::builder().threads(2).build().unwrap();
    let mut dst = StridedArray::<f64>::col_major(&[rows, cols]);
    let exec = Exec::new(&pool, 2).unwrap();
    copy_into(&mut dst.view_mut(), &src.view(), exec).unwrap();
    for r in 0..rows {
        for c in 0..cols {
            assert_eq!(dst.get(&[r, c]), (r * cols + c) as f64);
        }
    }
}

#[test]
fn test_copy_shape_mismatch() {
    let src = StridedArray::<f64>::col_major(&[2, 3]);
    let mut dst = StridedArray::<f64>::col_major(&[3, 2]);
    let result = copy_into(&mut dst.view_mut(), &src.view(), Exec::serial());
    assert!(result.is_err());
}