cutile 0.0.1

cuTile Rust lets programmers safely author and execute tile kernels directly in Rust.
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
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357

/*
 * SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 * SPDX-License-Identifier: Apache-2.0
 */
use std::panic::{self, AssertUnwindSafe};
use std::sync::Arc;

use cutile;
use cutile::api;
use cutile::tensor::{IntoPartition, Reshape, ToHostVec};
use cutile::tile_kernel::DeviceOp;

mod common;

#[cutile::module]
mod tensor_views_module {
    use cutile::core::*;

    #[cutile::entry()]
    fn passthrough_1d(output: &mut Tensor<f32, { [4] }>, input: &Tensor<f32, { [-1] }>) {
        let tile: Tile<f32, { [4] }> = load_tile_like_1d(input, output);
        output.store(tile);
    }

    #[cutile::entry()]
    fn passthrough_2d(output: &mut Tensor<f32, { [2, 2] }>, input: &Tensor<f32, { [-1, -1] }>) {
        let tile: Tile<f32, { [2, 2] }> = load_tile_like_2d(input, output);
        output.store(tile);
    }

    #[cutile::entry()]
    fn passthrough_1d_2(output: &mut Tensor<f32, { [2] }>, input: &Tensor<f32, { [-1] }>) {
        let tile: Tile<f32, { [2] }> = load_tile_like_1d(input, output);
        output.store(tile);
    }

    /// Copies a 4-element 1D slice to output. Used to verify sliced views
    /// with offset produce correct data.
    #[cutile::entry()]
    fn copy_4(output: &mut Tensor<f32, { [4] }>, input: &Tensor<f32, { [-1] }>) {
        let tile: Tile<f32, { [4] }> = load_tile_like_1d(input, output);
        output.store(tile);
    }

    /// Copies a 2x2 tile. Used to verify 2D sliced views with non-contiguous
    /// strides produce correct data.
    #[cutile::entry()]
    fn copy_2x2(output: &mut Tensor<f32, { [2, 2] }>, input: &Tensor<f32, { [-1, -1] }>) {
        let tile: Tile<f32, { [2, 2] }> = load_tile_like_2d(input, output);
        output.store(tile);
    }
}

#[test]
fn arc_views_are_zero_copy() {
    common::with_test_stack(|| {
        let base = Arc::new(api::arange::<f32>(8).sync().expect("Failed."));
        let view = base.reshape(&[2, 4]).unwrap();
        let dyn_view = base.reshape(&[4, 2]).unwrap();
        let flat = base.reshape(&[8]).unwrap();

        assert_eq!(
            base.device_pointer().cu_deviceptr(),
            view.device_pointer().cu_deviceptr()
        );
        assert_eq!(
            base.device_pointer().cu_deviceptr(),
            dyn_view.device_pointer().cu_deviceptr()
        );
        assert_eq!(
            base.device_pointer().cu_deviceptr(),
            flat.device_pointer().cu_deviceptr()
        );
        assert_eq!(view.shape(), vec![2, 4]);
        assert_eq!(view.strides(), vec![4, 1]);
        assert_eq!(dyn_view.shape(), vec![4, 2]);
        assert_eq!(dyn_view.strides(), vec![2, 1]);
        assert_eq!(flat.shape(), vec![8]);
        assert_eq!(flat.strides(), vec![1]);

        let base_host: Vec<f32> = (&base).to_host_vec().sync().expect("Failed.");
        let flat_host: Vec<f32> = flat.to_host_vec().sync().expect("Failed.");
        assert_eq!(base_host, flat_host);

        let copied = base.dup().sync().expect("Failed.");
        assert_ne!(
            base.device_pointer().cu_deviceptr(),
            copied.device_pointer().cu_deviceptr()
        );
    });
}

#[test]
fn invalid_view_shape_is_rejected() {
    common::with_test_stack(|| {
        let base = Arc::new(api::arange::<f32>(8).sync().expect("Failed."));
        assert!(base.reshape(&[5]).is_err());
        assert!(base.reshape(&[2, 2]).is_err());
    });
}

#[test]
fn shared_storage_blocks_mutable_partition() {
    common::with_test_stack(|| {
        let base = Arc::new(api::arange::<f32>(8).sync().expect("Failed."));
        let _view = base.reshape(&[2, 4]).unwrap();
        let owned = Arc::try_unwrap(base).expect("Expected unique outer Arc.");
        let result = panic::catch_unwind(AssertUnwindSafe(|| {
            let _ = owned.partition([8]);
        }));
        assert!(result.is_err(), "Expected mutable partition to be rejected");
    });
}

#[test]
fn arc_views_work_with_different_rank_kernels() {
    common::with_test_stack(|| {
        let base = Arc::new(api::arange::<f32>(4).sync().expect("Failed."));
        let input_1d = base.reshape(&[4]).unwrap();
        let input_2d = base.reshape(&[2, 2]).unwrap();

        let output_1d = api::zeros::<f32>(&[4]).sync().expect("Failed.");
        let (result_1d, _input_1d) =
            tensor_views_module::passthrough_1d(output_1d.partition([4]), input_1d)
                .sync()
                .expect("Failed.");
        let host_1d: Vec<f32> = result_1d
            .unpartition()
            .to_host_vec()
            .sync()
            .expect("Failed.");

        let output_2d = api::zeros::<f32>(&[2, 2]).sync().expect("Failed.");
        let (result_2d, _input_2d) =
            tensor_views_module::passthrough_2d(output_2d.partition([2, 2]), input_2d)
                .sync()
                .expect("Failed.");
        let host_2d: Vec<f32> = result_2d
            .unpartition()
            .to_host_vec()
            .sync()
            .expect("Failed.");

        assert_eq!(host_1d, vec![0.0, 1.0, 2.0, 3.0]);
        assert_eq!(host_2d, vec![0.0, 1.0, 2.0, 3.0]);
    });
}

// ── Slice tests ─────────────────────────────────────────────────────────────

#[test]
fn slice_1d_offset() {
    // arange(8) = [0, 1, 2, 3, 4, 5, 6, 7]
    // slice [4..8] = [4, 5, 6, 7]
    // Pass through a kernel and verify.
    common::with_test_stack(|| {
        let tensor = api::arange::<f32>(8).sync().expect("Failed.");
        let sliced = tensor.slice(&[4..8]).expect("Failed.");

        assert_eq!(sliced.shape(), &[4]);
        assert_eq!(sliced.strides(), &[1]);

        let output = api::zeros::<f32>(&[4]).sync().expect("Failed.");
        let (result, _) = tensor_views_module::copy_4(output.partition([4]), &sliced)
            .sync()
            .expect("Failed.");
        let host: Vec<f32> = result.unpartition().to_host_vec().sync().expect("Failed.");
        assert_eq!(host, vec![4.0, 5.0, 6.0, 7.0]);
    });
}

#[test]
fn slice_1d_middle() {
    // arange(8) = [0, 1, 2, 3, 4, 5, 6, 7]
    // slice [2..6] = [2, 3, 4, 5]
    common::with_test_stack(|| {
        let tensor = api::arange::<f32>(8).sync().expect("Failed.");
        let sliced = tensor.slice(&[2..6]).expect("Failed.");

        assert_eq!(sliced.shape(), &[4]);

        let output = api::zeros::<f32>(&[4]).sync().expect("Failed.");
        let (result, _) = tensor_views_module::copy_4(output.partition([4]), &sliced)
            .sync()
            .expect("Failed.");
        let host: Vec<f32> = result.unpartition().to_host_vec().sync().expect("Failed.");
        assert_eq!(host, vec![2.0, 3.0, 4.0, 5.0]);
    });
}

#[test]
fn slice_2d_rows_only() {
    // arange(16).view([4,4]):
    //   row 0: [0,  1,  2,  3]
    //   row 1: [4,  5,  6,  7]
    //   row 2: [8,  9, 10, 11]
    //   row 3: [12, 13, 14, 15]
    // slice [1..3, 0..2] -> [[4,5], [8,9]]
    // shape [2,2], strides [4,1] — non-contiguous (stride[0]=4, not 2).
    common::with_test_stack(|| {
        let tensor = api::arange::<f32>(16).sync().expect("Failed.");
        let tensor_2d = tensor.view(&[4, 4]).expect("Failed.");
        let sliced = tensor_2d.slice(&[1..3, 0..2]).expect("Failed.");

        assert_eq!(sliced.shape(), &[2, 2]);
        assert_eq!(sliced.strides(), &[4, 1]); // Non-contiguous.

        let output = api::zeros::<f32>(&[2, 2]).sync().expect("Failed.");
        let (result, _) = tensor_views_module::copy_2x2(output.partition([2, 2]), &sliced)
            .sync()
            .expect("Failed.");
        let host: Vec<f32> = result.unpartition().to_host_vec().sync().expect("Failed.");
        assert_eq!(host, vec![4.0, 5.0, 8.0, 9.0]);
    });
}

#[test]
fn slice_2d_columns_non_contiguous() {
    // arange(16).view([4,4]):
    //   row 0: [0,  1,  2,  3]
    //   row 1: [4,  5,  6,  7]
    //   row 2: [8,  9, 10, 11]
    //   row 3: [12, 13, 14, 15]
    // slice [0..2, 2..4] (first 2 rows, columns 2-3):
    //   [[2, 3], [6, 7]]
    // shape [2,2], strides [4,1] — non-contiguous.
    common::with_test_stack(|| {
        let tensor = api::arange::<f32>(16).sync().expect("Failed.");
        let tensor_2d = tensor.view(&[4, 4]).expect("Failed.");
        let sliced = tensor_2d.slice(&[0..2, 2..4]).expect("Failed.");

        assert_eq!(sliced.shape(), &[2, 2]);
        assert_eq!(sliced.strides(), &[4, 1]); // Non-contiguous.

        let output = api::zeros::<f32>(&[2, 2]).sync().expect("Failed.");
        let (result, _) = tensor_views_module::copy_2x2(output.partition([2, 2]), &sliced)
            .sync()
            .expect("Failed.");
        let host: Vec<f32> = result.unpartition().to_host_vec().sync().expect("Failed.");
        assert_eq!(host, vec![2.0, 3.0, 6.0, 7.0]);
    });
}

#[test]
fn slice_chained() {
    // arange(16).view([4,4]):
    //   [0,  1,  2,  3]
    //   [4,  5,  6,  7]
    //   [8,  9, 10, 11]
    //   [12, 13, 14, 15]
    // slice [1..3] -> rows 1-2: [[4,5,6,7], [8,9,10,11]]
    // then slice [:, 1..3] -> cols 1-2: [[5,6], [9,10]]
    common::with_test_stack(|| {
        let tensor = api::arange::<f32>(16).sync().expect("Failed.");
        let tensor_2d = tensor.view(&[4, 4]).expect("Failed.");
        let row_slice = tensor_2d.slice(&[1..3]).expect("Failed.");
        let final_slice = row_slice.slice(&[0..2, 1..3]).expect("Failed.");

        assert_eq!(final_slice.shape(), &[2, 2]);
        assert_eq!(final_slice.strides(), &[4, 1]);

        let output = api::zeros::<f32>(&[2, 2]).sync().expect("Failed.");
        let (result, _) = tensor_views_module::copy_2x2(output.partition([2, 2]), &final_slice)
            .sync()
            .expect("Failed.");
        let host: Vec<f32> = result.unpartition().to_host_vec().sync().expect("Failed.");
        assert_eq!(host, vec![5.0, 6.0, 9.0, 10.0]);
    });
}

#[test]
fn slice_out_of_bounds_rejected() {
    common::with_test_stack(|| {
        let tensor = api::arange::<f32>(8).sync().expect("Failed.");
        assert!(tensor.slice(&[0..9]).is_err());
        assert!(tensor.slice(&[5..3]).is_err());

        let tensor_2d = tensor.view(&[2, 4]).expect("Failed.");
        assert!(tensor_2d.slice(&[0..3]).is_err()); // axis 0 has dim 2
        assert!(tensor_2d.slice(&[0..2, 0..5]).is_err()); // axis 1 has dim 4
    });
}

#[test]
fn slice_preserves_view_reshape_rejection() {
    // A non-contiguous slice should reject reshape via view().
    common::with_test_stack(|| {
        let tensor = api::arange::<f32>(16).sync().expect("Failed.");
        let tensor_2d = tensor.view(&[4, 4]).expect("Failed.");
        let col_slice = tensor_2d.slice(&[0..4, 1..3]).expect("Failed.");

        // col_slice has strides [4, 1] but shape [4, 2] — non-contiguous.
        assert!(
            col_slice.view(&[8]).is_err(),
            "reshape on non-contiguous view should fail"
        );
    });
}

// ── Slice metadata (CPU-only, no kernel launch) ───────────────────────────
//
// Inspired by numpy: verify shape, strides, and offset for various
// slicing patterns without launching GPU kernels.

#[test]
fn slice_full_range_is_identity() {
    // numpy: a[0:n] is a view with same shape
    common::with_test_stack(|| {
        let a = api::arange::<f32>(16).sync().expect("Failed.");
        let v = a.slice(&[0..16]).expect("Failed.");
        assert_eq!(v.shape(), &[16]);
        assert_eq!(v.strides(), &[1]);
    });
}

#[test]
fn slice_single_element() {
    // numpy: a[5:6] → shape [1]
    common::with_test_stack(|| {
        let a = api::arange::<f32>(16).sync().expect("Failed.");
        let v = a.slice(&[5..6]).expect("Failed.");
        assert_eq!(v.shape(), &[1]);
    });
}

#[test]
fn slice_2d_partial_axes() {
    // numpy: a[1:3] on a 4x4 → rows 1-2, all columns preserved
    common::with_test_stack(|| {
        let a = api::arange::<f32>(16).sync().expect("Failed.");
        let a2d = a.view(&[4, 4]).expect("Failed.");
        let v = a2d.slice(&[1..3]).expect("Failed.");
        assert_eq!(v.shape(), &[2, 4]);
        assert_eq!(v.strides(), &[4, 1]);
    });
}

#[test]
fn slice_chained_accumulates_offset() {
    // numpy: a[2:10][3:5] == a[5:7]
    // Verify that chained slices compose offsets correctly.
    common::with_test_stack(|| {
        let a = api::arange::<f32>(16).sync().expect("Failed.");
        let s1 = a.slice(&[2..10]).expect("Failed.");
        assert_eq!(s1.shape(), &[8]);
        let s2 = s1.slice(&[3..5]).expect("Failed.");
        assert_eq!(s2.shape(), &[2]);

        let output = api::zeros::<f32>(&[2]).sync().expect("Failed.");
        let (result, _) = tensor_views_module::passthrough_1d_2(output.partition([2]), &s2)
            .sync()
            .expect("Failed.");
        let host: Vec<f32> = result.unpartition().to_host_vec().sync().expect("Failed.");
        // a[5:7] = [5.0, 6.0]
        assert_eq!(host, vec![5.0, 6.0]);
    });
}