File: test_codegen_triton.py

package info (click to toggle)
pytorch-cuda 2.6.0%2Bdfsg-7
  • links: PTS, VCS
  • area: contrib
  • in suites: forky, sid, trixie
  • size: 161,620 kB
  • sloc: python: 1,278,832; cpp: 900,322; ansic: 82,710; asm: 7,754; java: 3,363; sh: 2,811; javascript: 2,443; makefile: 597; ruby: 195; xml: 84; objc: 68
file content (87 lines) | stat: -rw-r--r-- 2,794 bytes parent folder | download | duplicates (3)
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
# Owner(s): ["module: inductor"]
import contextlib

import sympy

import torch
import torch._inductor.config as inductor_config
from torch._inductor.codegen import triton_utils
from torch._inductor.codegen.common import SizeArg
from torch._inductor.graph import GraphLowering
from torch._inductor.test_case import TestCase as InductorTestCase
from torch._inductor.virtualized import V
from torch.testing._internal.inductor_utils import HAS_CPU, HAS_GPU


class TestCodegenTriton(InductorTestCase):
    def setUp(self):
        super().setUp()

        class DummyModule(torch.nn.Module):
            def forward(self, x):
                return x * 2

        self._gm = torch.fx.symbolic_trace(DummyModule())
        self._graph = GraphLowering(self._gm)

        self._stack = contextlib.ExitStack()
        self._stack.enter_context(V.set_graph_handler(self._graph))

    def tearDown(self):
        self._stack.close()
        super().tearDown()

    @inductor_config.patch("triton.divisible_by_16", True)
    def test_config_of_sizearg(self):
        two = sympy.Integer(2)
        eight = sympy.Integer(8)
        sixteen = sympy.Integer(16)
        s0 = sympy.Symbol("s0", positive=True, integer=True)
        s1 = sympy.Symbol("s1", positive=True, integer=True)

        def _check_divisibility(config):
            try:
                from triton.backends.compiler import AttrsDescriptor  # noqa: F401

                return config.divisibility_16
            except ImportError:
                return config.divisible_by_16

        self.assertEqual(
            (2,),
            _check_divisibility(
                triton_utils.config_of(
                    [
                        SizeArg("A", two),  # no
                        SizeArg("B", eight),  # no
                        SizeArg("C", sixteen),  # yes
                        SizeArg("D", s0),  # no
                        SizeArg("E", s1),  # no
                    ]
                )
            ),
        )

        self.assertEqual(
            (0, 2, 4, 5, 6),
            _check_divisibility(
                triton_utils.config_of(
                    [
                        SizeArg("A", two * eight),  # 0: yes
                        SizeArg("B", eight * s0),  # 1: no
                        SizeArg("C", two * eight * s0),  # 2: yes
                        SizeArg("D", s0 * s1),  # 3: no
                        SizeArg("E", sixteen * s0),  # 4: yes
                        SizeArg("F", sixteen * eight * s0 * s1),  # 5: yes
                        SizeArg("G", two * eight * s0 * s1),  # 6: yes
                    ]
                )
            ),
        )


if __name__ == "__main__":
    from torch._inductor.test_case import run_tests

    if HAS_CPU or HAS_GPU:
        run_tests("sympy")