xref: /aosp_15_r20/external/pytorch/benchmarks/operator_benchmark/pt/qcomparators_test.py (revision da0073e96a02ea20f0ac840b70461e3646d07c45) 
1import operator_benchmark as op_bench
2
3import torch
4
5
6qcomparators_configs = op_bench.cross_product_configs(
7    N=(8, 64),
8    dtype=(torch.quint8, torch.qint8, torch.qint32),
9    contig=(False, True),
10    other_scalar=(False, True),
11    out_variant=(False, True),
12    tags=("short",),
13)
14
15qcomparators_ops = op_bench.op_list(
16    attrs=(
17        ("eq", torch.eq),
18        ("ne", torch.ne),
19        ("lt", torch.lt),
20        ("gt", torch.gt),
21        ("le", torch.le),
22        ("ge", torch.ge),
23    ),
24    attr_names=("op_name", "op_func"),
25)
26
27
28class QComparatorBenchmark(op_bench.TorchBenchmarkBase):
29    def init(self, N, dtype, contig, other_scalar, out_variant, op_func):
30        # TODO: Consider more diverse shapes
31        f_input = (torch.rand(N, N) - 0.5) * 256
32        scale = 1.0
33        zero_point = 0
34
35        q_input_a = torch.quantize_per_tensor(
36            f_input, scale=scale, zero_point=zero_point, dtype=dtype
37        )
38        q_input_b = q_input_a.clone()
39
40        if not contig:
41            permute_dims = list(range(f_input.ndim))[::-1]
42            q_input_a = q_input_a.permute(permute_dims)
43
44        self.qop = op_func
45        self.inputs = {
46            "q_input_a": q_input_a,
47            "q_input_b": q_input_b,
48            "out_variant": out_variant,
49            "other_scalar": other_scalar,
50        }
51
52    def forward(self, q_input_a, q_input_b, out_variant: bool, other_scalar: bool):
53        if out_variant:
54            if other_scalar:
55                return self.qop(q_input_a, 42, out=torch.tensor(True, dtype=torch.bool))
56            else:
57                return self.qop(
58                    q_input_a, q_input_b, out=torch.tensor(True, dtype=torch.bool)
59                )
60        else:
61            if other_scalar:
62                return self.qop(q_input_a, 42)
63            else:
64                return self.qop(q_input_a, q_input_b)
65
66
67op_bench.generate_pt_tests_from_op_list(
68    qcomparators_ops, qcomparators_configs, QComparatorBenchmark
69)
70
71
72if __name__ == "__main__":
73    op_bench.benchmark_runner.main()
74