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

/*
 * SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 * SPDX-License-Identifier: Apache-2.0
 */
use cutile;
use cutile_compiler::compiler::utils::CompileOptions;
use cutile_compiler::compiler::{CUDATileFunctionCompiler, CUDATileModules};
use cutile_compiler::cuda_tile_runtime_utils::get_gpu_name;

mod common;

#[cutile::module]
mod binary_math_ops_module {

    use cutile::core::*;

    #[cutile::entry()]
    fn minmax_kernel<const S: [i32; 1]>(output: &mut Tensor<f32, S>) {
        // Test min and max operations
        let x: Tile<f32, S> = load_tile_mut(output);
        let y: Tile<f32, S> = load_tile_mut(output);
        let max_result: Tile<f32, S> = maxf(x, y, nan::Disabled, ftz::Disabled);
        let min_result: Tile<f32, S> = minf(max_result, y, nan::Disabled, ftz::Disabled);
        output.store(min_result);
    }

    #[cutile::entry()]
    fn select_kernel<const S: [i32; 1]>(output: &mut Tensor<f32, S>) {
        // Test select operation
        let x: Tile<f32, S> = load_tile_mut(output);
        let y: Tile<f32, S> = load_tile_mut(output);

        let _zero: Tile<f32, S> = constant(0.0, output.shape());
        let _one: Tile<f32, S> = constant(1.0, output.shape());

        // Simplified to avoid bool literal issue
        let result: Tile<f32, S> = maxf(x, y, nan::Disabled, ftz::Disabled);
        output.store(result);
    }

    #[cutile::entry()]
    fn bf16_binary_arith_kernel<const S: [i32; 1]>(output: &mut Tensor<bf16, S>) {
        // Covers bf16 binary arithmetic lowering
        let x: Tile<bf16, S> = load_tile_mut(output);
        let y: Tile<bf16, S> = load_tile_mut(output);

        let sum: Tile<bf16, S> = x + y;
        let product: Tile<bf16, S> = sum * y;
        let result: Tile<bf16, S> = product / x;
        output.store(result);
    }
}

use binary_math_ops_module::_module_asts;

#[test]
fn compile_minmax() -> () {
    common::with_test_stack(|| {
        let modules =
            CUDATileModules::new(_module_asts()).expect("Failed to create CUDATileModules");
        let gpu_name = get_gpu_name(0);
        let compiler = CUDATileFunctionCompiler::new(
            &modules,
            "binary_math_ops_module",
            "minmax_kernel",
            &[128.to_string()],
            &[("output", &[1])],
            &[],
            &[],
            None,
            gpu_name,
            &CompileOptions::default(),
        )
        .expect("Failed.");
        let module_op_str = compiler
            .compile()
            .expect("Failed.")
            .to_string();
        println!("\n=== MIN/MAX MLIR ===\n{}", module_op_str);

        let expected_ops = ["maxf", "minf"];
        for op in expected_ops {
            assert!(
                module_op_str.contains(format!("= {}", op).as_str()),
                "Expected {} operation in MLIR output",
                op
            );
        }

        println!(
            "\n✓ All {} min/max operations verified in MLIR output",
            expected_ops.len()
        );
    });
}

#[test]
fn compile_select() -> () {
    common::with_test_stack(|| {
        let modules =
            CUDATileModules::new(_module_asts()).expect("Failed to create CUDATileModules");
        let gpu_name = get_gpu_name(0);
        let compiler = CUDATileFunctionCompiler::new(
            &modules,
            "binary_math_ops_module",
            "select_kernel",
            &[128.to_string()],
            &[("output", &[1])],
            &[],
            &[],
            None,
            gpu_name,
            &CompileOptions::default(),
        )
        .expect("Failed.");
        let module_op_str = compiler
            .compile()
            .expect("Failed.")
            .to_string();
        println!("\n=== SELECT MLIR ===\n{}", module_op_str);

        assert!(
            module_op_str.contains("maxf"),
            "Kernel compiled but select test needs comparison ops"
        );

        println!("\n✓ select operation verified in MLIR output");
    });
}

#[test]
fn compile_bf16_binary_arith() -> () {
    common::with_test_stack(|| {
        let modules =
            CUDATileModules::new(_module_asts()).expect("Failed to create CUDATileModules");
        let gpu_name = get_gpu_name(0);
        let compiler = CUDATileFunctionCompiler::new(
            &modules,
            "binary_math_ops_module",
            "bf16_binary_arith_kernel",
            &[128.to_string()],
            &[("output", &[1])],
            &[],
            &[],
            None,
            gpu_name,
            &CompileOptions::default(),
        )
        .expect("Failed.");
        let module_op_str = compiler
            .compile()
            .expect("Failed.")
            .to_string();
        println!("\n=== BF16 BINARY ARITH MLIR ===\n{}", module_op_str);

        for op in ["addf", "mulf", "divf"] {
            assert!(
                module_op_str.contains(format!("= {}", op).as_str()),
                "Expected {} operation in MLIR output",
                op
            );
        }
        assert!(
            module_op_str.contains("bf16"),
            "Expected bf16 type in MLIR output"
        );
    });
}