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
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
//! Regression tests for LANE W1-K item 3: `src/axis_ops.rs`.
//!
//! `cumsum_axis` / `cumprod_axis` had two distinct bugs, both exercised
//! here:
//!
//! 1. **Correctness** (not merely a panic): the inner loop only ever wrote
//! to `stride` many flat positions per sequence (missing an `elem` loop
//! over the `stride` positions within each block), so accumulating along
//! any axis but the last silently produced wrong results for arrays with
//! more than one row along that axis's orthogonal dimensions. Verified
//! against NumPy (`np.cumsum` / `np.cumprod`) on a plain contiguous
//! array below -- this fails with the pre-fix code even though the array
//! is perfectly contiguous.
//! 2. **Panic hygiene**: `.as_slice_mut()` / `.as_slice()` were used
//! unconditionally with `.expect("Array must be contiguous ...")`,
//! reachable by calling these methods on a non-contiguous view (e.g. the
//! result of `transpose_axis`). Fixed by snapshotting elements in
//! logical order via `.iter()` instead.
use numrs2::array::Array;
use numrs2::prelude::*;
// ---------------------------------------------------------------------
// cumsum_axis: correctness on a plain contiguous array (bug #1 above).
// ---------------------------------------------------------------------
#[test]
fn cumsum_axis_0_on_contiguous_2x3_matches_numpy() {
// a = [[1,2,3],[4,5,6]]
// NumPy reference: np.cumsum(a, axis=0) -> [[1,2,3],[5,7,9]]
let a = Array::from_vec(vec![1, 2, 3, 4, 5, 6]).reshape(&[2, 3]);
let result = a.cumsum_axis(0).expect("cumsum_axis(0) should succeed");
assert_eq!(result.shape(), vec![2, 3]);
assert_eq!(result.to_vec(), vec![1, 2, 3, 5, 7, 9]);
}
#[test]
fn cumsum_axis_1_on_contiguous_2x3_matches_numpy() {
// NumPy reference: np.cumsum(a, axis=1) -> [[1,3,6],[4,9,15]]
let a = Array::from_vec(vec![1, 2, 3, 4, 5, 6]).reshape(&[2, 3]);
let result = a.cumsum_axis(1).expect("cumsum_axis(1) should succeed");
assert_eq!(result.shape(), vec![2, 3]);
assert_eq!(result.to_vec(), vec![1, 3, 6, 4, 9, 15]);
}
#[test]
fn cumsum_axis_0_on_larger_contiguous_matrix_matches_numpy() {
// a = [[1,2,3,4],[5,6,7,8],[9,10,11,12]] (3x4)
// NumPy reference: np.cumsum(a, axis=0)
// -> [[1,2,3,4],[6,8,10,12],[15,18,21,24]]
let a = Array::from_vec((1..=12).collect::<Vec<i64>>()).reshape(&[3, 4]);
let result = a.cumsum_axis(0).expect("cumsum_axis(0) should succeed");
assert_eq!(
result.to_vec(),
vec![1, 2, 3, 4, 6, 8, 10, 12, 15, 18, 21, 24]
);
}
/// The bug fixed in `cumsum_axis`/`cumprod_axis` (missing inner loop over
/// `stride`) and the one fixed in `argmin_axis`/`argmax_axis`
/// (over-allocated `result_data`) are only fully exercised when *both*
/// `n_sequences > 1` and `elements_per_sequence > 1` simultaneously -- a
/// plain 2D matrix only ever has one of those `> 1` for a given axis. This
/// uses a 3D array with the accumulation/reduction axis in the *middle*
/// (`axis=1` of a `[2,3,4]` array: `n_sequences=2`, `elements_per_sequence
/// (== stride) = 4`), which pins down the flat-index arithmetic in both
/// dimensions at once.
#[test]
fn cumsum_axis_1_on_contiguous_3d_matrix_matches_numpy() {
// a[i,j,k] = i*12 + j*4 + k, shape [2,3,4]:
// a[0] = [[0,1,2,3],[4,5,6,7],[8,9,10,11]]
// a[1] = [[12,13,14,15],[16,17,18,19],[20,21,22,23]]
// NumPy reference: np.cumsum(a, axis=1) ->
// [[[0,1,2,3],[4,6,8,10],[12,15,18,21]],
// [[12,13,14,15],[28,30,32,34],[48,51,54,57]]]
let a = Array::from_vec((0..24).collect::<Vec<i64>>()).reshape(&[2, 3, 4]);
let result = a.cumsum_axis(1).expect("cumsum_axis(1) should succeed");
assert_eq!(result.shape(), vec![2, 3, 4]);
assert_eq!(
result.to_vec(),
vec![
0, 1, 2, 3, 4, 6, 8, 10, 12, 15, 18, 21, 12, 13, 14, 15, 28, 30, 32, 34, 48, 51, 54,
57,
]
);
}
#[test]
fn argmin_axis_1_on_contiguous_3d_matrix_matches_numpy() {
// Hand-picked so the argmin along axis=1 (size 3) varies independently
// for every (i, k) combination -- an ascending-value array would give
// argmin == 0 everywhere and would not distinguish a correct
// implementation from one with broken index arithmetic.
//
// a[0,:,:] = [[5,1,7,9],[2,9,3,8],[8,4,6,0]]
// a[1,:,:] = [[4,2,3,8],[6,0,1,5],[1,5,2,9]]
//
// Per (i, k), the column (j=0,1,2) minimum and its index:
// i=0: k0 [5,2,8]->j1 k1 [1,9,4]->j0 k2 [7,3,6]->j1 k3 [9,8,0]->j2
// i=1: k0 [4,6,1]->j2 k1 [2,0,5]->j1 k2 [3,1,2]->j1 k3 [8,5,9]->j1
// NumPy reference: np.argmin(a, axis=1) -> [[1,0,1,2],[2,1,1,1]]
let a = Array::from_vec(vec![
5, 1, 7, 9, 2, 9, 3, 8, 8, 4, 6, 0, 4, 2, 3, 8, 6, 0, 1, 5, 1, 5, 2, 9,
])
.reshape(&[2, 3, 4]);
let result = a.argmin_axis(1).expect("argmin_axis(1) should succeed");
assert_eq!(result.shape(), vec![2, 4]);
assert_eq!(result.to_vec(), vec![1, 0, 1, 2, 2, 1, 1, 1]);
}
#[test]
fn cumprod_axis_0_on_contiguous_2x3_matches_numpy() {
// NumPy reference: np.cumprod([[1,2,3],[4,5,6]], axis=0)
// -> [[1,2,3],[4,10,18]]
let a = Array::from_vec(vec![1, 2, 3, 4, 5, 6]).reshape(&[2, 3]);
let result = a.cumprod_axis(0).expect("cumprod_axis(0) should succeed");
assert_eq!(result.to_vec(), vec![1, 2, 3, 4, 10, 18]);
}
#[test]
fn cumprod_axis_1_on_contiguous_2x3_matches_numpy() {
// NumPy reference: np.cumprod([[1,2,3],[4,5,6]], axis=1)
// -> [[1,2,6],[4,20,120]]
let a = Array::from_vec(vec![1, 2, 3, 4, 5, 6]).reshape(&[2, 3]);
let result = a.cumprod_axis(1).expect("cumprod_axis(1) should succeed");
assert_eq!(result.to_vec(), vec![1, 2, 6, 4, 20, 120]);
}
// ---------------------------------------------------------------------
// cumsum_axis / cumprod_axis / argmin_axis / argmax_axis: must not panic
// on a non-contiguous (transposed) array, and must give logically-correct
// results (bug #2 above).
// ---------------------------------------------------------------------
/// Builds `b`, the transpose of `[[1,2,3],[4,5,6]]`, i.e. logically
/// `[[1,4],[2,5],[3,6]]` with shape `[3, 2]`, backed by a non-contiguous
/// memory layout (same buffer as the original 2x3 array, reinterpreted via
/// `transpose_axis`, so `.as_slice()`/`.as_slice_mut()` return `None`).
fn transposed_non_contiguous() -> Array<i64> {
let a = Array::from_vec(vec![1, 2, 3, 4, 5, 6]).reshape(&[2, 3]);
let b = a.transpose_axis(0, 1);
assert_eq!(b.shape(), vec![3, 2]);
assert_eq!(b.to_vec(), vec![1, 4, 2, 5, 3, 6]); // sanity check on b itself
assert!(
b.array().as_slice().is_none(),
"test fixture must actually be non-contiguous"
);
b
}
#[test]
fn cumsum_axis_on_non_contiguous_array_does_not_panic_and_matches_numpy() {
let b = transposed_non_contiguous();
// b = [[1,4],[2,5],[3,6]]
// NumPy reference: np.cumsum(b, axis=0) -> [[1,4],[3,9],[6,15]]
let result = b
.cumsum_axis(0)
.expect("must not panic on non-contiguous input");
assert_eq!(result.shape(), vec![3, 2]);
assert_eq!(result.to_vec(), vec![1, 4, 3, 9, 6, 15]);
}
#[test]
fn cumsum_axis_1_on_non_contiguous_array_matches_numpy() {
let b = transposed_non_contiguous();
// NumPy reference: np.cumsum(b, axis=1) -> [[1,5],[2,7],[3,9]]
let result = b
.cumsum_axis(1)
.expect("must not panic on non-contiguous input");
assert_eq!(result.to_vec(), vec![1, 5, 2, 7, 3, 9]);
}
#[test]
fn cumprod_axis_on_non_contiguous_array_does_not_panic_and_matches_numpy() {
let b = transposed_non_contiguous();
// NumPy reference: np.cumprod(b, axis=0) -> [[1,4],[2,20],[6,120]]
let result = b
.cumprod_axis(0)
.expect("must not panic on non-contiguous input");
assert_eq!(result.to_vec(), vec![1, 4, 2, 20, 6, 120]);
}
#[test]
fn argmin_axis_on_non_contiguous_array_does_not_panic_and_matches_numpy() {
let b = transposed_non_contiguous();
// NumPy reference: np.argmin(b, axis=0) -> [0, 0] (column mins: 1, 4)
let result = b
.argmin_axis(0)
.expect("must not panic on non-contiguous input");
assert_eq!(result.to_vec(), vec![0, 0]);
}
#[test]
fn argmax_axis_on_non_contiguous_array_does_not_panic_and_matches_numpy() {
let b = transposed_non_contiguous();
// NumPy reference: np.argmax(b, axis=0) -> [2, 2] (column maxes: 3, 6)
let result = b
.argmax_axis(0)
.expect("must not panic on non-contiguous input");
assert_eq!(result.to_vec(), vec![2, 2]);
}
#[test]
fn argmin_argmax_axis_1_on_non_contiguous_array_matches_numpy() {
let b = transposed_non_contiguous();
// b = [[1,4],[2,5],[3,6]]; NumPy reference:
// np.argmin(b, axis=1) -> [0, 0, 0] (each row's min is column 0)
// np.argmax(b, axis=1) -> [1, 1, 1] (each row's max is column 1)
let argmin = b
.argmin_axis(1)
.expect("must not panic on non-contiguous input");
let argmax = b
.argmax_axis(1)
.expect("must not panic on non-contiguous input");
assert_eq!(argmin.to_vec(), vec![0, 0, 0]);
assert_eq!(argmax.to_vec(), vec![1, 1, 1]);
}