cutile 0.0.2

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

/*
 * SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 * SPDX-License-Identifier: Apache-2.0
 */

//! Tests that all float ElementTypes can:
//! 1. Create a 1D tensor of ones
//! 2. Add two such tensors in a kernel
//! 3. Convert the result to f32
//! 4. Verify correctness on the host

use cuda_core::{f8e4m3fn, f8e5m2};
use cutile;
use cutile::tensor::PartitionMut;
use cutile::tile_kernel::{DeviceOp, ToHostVecOp};
use half::{bf16, f16};
use std::sync::Arc;

mod common;

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

    #[cutile::entry()]
    fn add_f16<const B: i32>(
        out: &mut Tensor<f16, { [B] }>,
        a: &Tensor<f16, { [-1] }>,
        b: &Tensor<f16, { [-1] }>,
    ) {
        let pid = get_tile_block_id().0;
        let tile_a = a.load_tile(const_shape![B], [pid]);
        let tile_b = b.load_tile(const_shape![B], [pid]);
        out.store(tile_a + tile_b);
    }

    #[cutile::entry()]
    fn add_bf16<const B: i32>(
        out: &mut Tensor<bf16, { [B] }>,
        a: &Tensor<bf16, { [-1] }>,
        b: &Tensor<bf16, { [-1] }>,
    ) {
        let pid = get_tile_block_id().0;
        let tile_a = a.load_tile(const_shape![B], [pid]);
        let tile_b = b.load_tile(const_shape![B], [pid]);
        out.store(tile_a + tile_b);
    }

    #[cutile::entry()]
    fn add_f32<const B: i32>(
        out: &mut Tensor<f32, { [B] }>,
        a: &Tensor<f32, { [-1] }>,
        b: &Tensor<f32, { [-1] }>,
    ) {
        let pid = get_tile_block_id().0;
        let tile_a = a.load_tile(const_shape![B], [pid]);
        let tile_b = b.load_tile(const_shape![B], [pid]);
        out.store(tile_a + tile_b);
    }

    #[cutile::entry()]
    fn add_f64<const B: i32>(
        out: &mut Tensor<f64, { [B] }>,
        a: &Tensor<f64, { [-1] }>,
        b: &Tensor<f64, { [-1] }>,
    ) {
        let pid = get_tile_block_id().0;
        let tile_a = a.load_tile(const_shape![B], [pid]);
        let tile_b = b.load_tile(const_shape![B], [pid]);
        out.store(tile_a + tile_b);
    }

    // Note: FP8 types (f8e4m3fn, f8e5m2) don't support direct arithmetic
    // on GPU — they're storage/quantization formats. FP8 add kernels are
    // not included here. Use convert to f16/f32 for compute.
}

use float_add_module::{add_bf16, add_f16, add_f32, add_f64};

#[test]
fn add_ones_f16_and_convert_to_f32() {
    common::with_test_stack(|| {
        let a = cutile::api::ones::<f16>(&[128]).sync().expect("alloc a");
        let b = cutile::api::ones::<f16>(&[128]).sync().expect("alloc b");
        let mut out = cutile::api::zeros::<f16>(&[128]).sync().expect("alloc out");

        add_f16((&mut out).partition([128]), &a, &b)
            .sync()
            .expect("add_f16 failed");

        let result_f32: Vec<f32> = cutile::api::convert::<f16, f32>(Arc::new(out))
            .sync()
            .expect("convert failed")
            .dup()
            .to_host_vec()
            .sync()
            .expect("to_host");

        assert_eq!(result_f32.len(), 128);
        for (i, &v) in result_f32.iter().enumerate() {
            assert!(
                (v - 2.0).abs() < 1e-3,
                "f16: element {i} = {v}, expected 2.0"
            );
        }
    });
}

#[test]
fn add_ones_bf16_and_convert_to_f32() {
    common::with_test_stack(|| {
        let a = cutile::api::ones::<bf16>(&[128]).sync().expect("alloc a");
        let b = cutile::api::ones::<bf16>(&[128]).sync().expect("alloc b");
        let mut out = cutile::api::zeros::<bf16>(&[128])
            .sync()
            .expect("alloc out");

        add_bf16((&mut out).partition([128]), &a, &b)
            .sync()
            .expect("add_bf16 failed");

        let result_f32: Vec<f32> = cutile::api::convert::<bf16, f32>(Arc::new(out))
            .sync()
            .expect("convert failed")
            .dup()
            .to_host_vec()
            .sync()
            .expect("to_host");

        assert_eq!(result_f32.len(), 128);
        for (i, &v) in result_f32.iter().enumerate() {
            assert!(
                (v - 2.0).abs() < 1e-2,
                "bf16: element {i} = {v}, expected 2.0"
            );
        }
    });
}

#[test]
fn add_ones_f32_direct() {
    common::with_test_stack(|| {
        let a = cutile::api::ones::<f32>(&[128]).sync().expect("alloc a");
        let b = cutile::api::ones::<f32>(&[128]).sync().expect("alloc b");
        let mut out = cutile::api::zeros::<f32>(&[128]).sync().expect("alloc out");

        add_f32((&mut out).partition([128]), &a, &b)
            .sync()
            .expect("add_f32 failed");

        let host: Vec<f32> = out.dup().to_host_vec().sync().expect("to_host");
        assert_eq!(host.len(), 128);
        for (i, &v) in host.iter().enumerate() {
            assert!(
                (v - 2.0).abs() < 1e-5,
                "f32: element {i} = {v}, expected 2.0"
            );
        }
    });
}

#[test]
fn add_ones_f64_and_convert_to_f32() {
    common::with_test_stack(|| {
        let a = cutile::api::ones::<f64>(&[128]).sync().expect("alloc a");
        let b = cutile::api::ones::<f64>(&[128]).sync().expect("alloc b");
        let mut out = cutile::api::zeros::<f64>(&[128]).sync().expect("alloc out");

        add_f64((&mut out).partition([128]), &a, &b)
            .sync()
            .expect("add_f64 failed");

        let result_f32: Vec<f32> = cutile::api::convert::<f64, f32>(Arc::new(out))
            .sync()
            .expect("convert failed")
            .dup()
            .to_host_vec()
            .sync()
            .expect("to_host");

        assert_eq!(result_f32.len(), 128);
        for (i, &v) in result_f32.iter().enumerate() {
            assert!(
                (v - 2.0).abs() < 1e-5,
                "f64: element {i} = {v}, expected 2.0"
            );
        }
    });
}

// FP8 types don't support direct arithmetic (addf) on GPU — they're
// storage formats. Test the create → convert → check pipeline instead.

#[test]
fn create_ones_f8e4m3fn_and_convert_to_f32() {
    common::with_test_stack(|| {
        let ones = cutile::api::ones::<f8e4m3fn>(&[128]).sync().expect("alloc");

        let result_f32: Vec<f32> = cutile::api::convert::<f8e4m3fn, f32>(Arc::new(ones))
            .sync()
            .expect("convert failed")
            .dup()
            .to_host_vec()
            .sync()
            .expect("to_host");

        assert_eq!(result_f32.len(), 128);
        for (i, &v) in result_f32.iter().enumerate() {
            assert!(
                (v - 1.0).abs() < 0.5,
                "f8e4m3fn: element {i} = {v}, expected 1.0"
            );
        }
    });
}

#[test]
fn create_ones_f8e5m2_and_convert_to_f32() {
    common::with_test_stack(|| {
        let ones = cutile::api::ones::<f8e5m2>(&[128]).sync().expect("alloc");

        let result_f32: Vec<f32> = cutile::api::convert::<f8e5m2, f32>(Arc::new(ones))
            .sync()
            .expect("convert failed")
            .dup()
            .to_host_vec()
            .sync()
            .expect("to_host");

        assert_eq!(result_f32.len(), 128);
        for (i, &v) in result_f32.iter().enumerate() {
            assert!(
                (v - 1.0).abs() < 0.5,
                "f8e5m2: element {i} = {v}, expected 1.0"
            );
        }
    });
}