• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1import numpy
2from pt import configs
3
4import operator_benchmark as op_bench
5
6import torch
7import torch.ao.nn.qat as nnqat
8from torch.ao.quantization import default_embedding_qat_qconfig
9
10
11"""
12Microbenchmarks for QAT Embedding + EmbeddingBag operators.
13"""
14
15
16class QATEmbeddingBagBenchmark(op_bench.TorchBenchmarkBase):
17    def init(
18        self,
19        embeddingbags,
20        dim,
21        mode,
22        input_size,
23        offset,
24        sparse,
25        include_last_offset,
26        device,
27    ):
28        qconfig = default_embedding_qat_qconfig
29        self.embedding = nnqat.EmbeddingBag(
30            num_embeddings=embeddingbags,
31            embedding_dim=dim,
32            mode=mode,
33            include_last_offset=include_last_offset,
34            sparse=sparse,
35            device=device,
36            qconfig=qconfig,
37        )
38        numpy.random.seed((1 << 32) - 1)
39        offsets = torch.LongTensor([offset], device=device)
40        input = torch.tensor(
41            numpy.random.randint(0, embeddingbags, input_size), device=device
42        ).long()
43        self.inputs = {
44            "input": input,
45            "offset": torch.cat(
46                (offsets, torch.tensor([input.size(0)], dtype=torch.long)), 0
47            ),
48        }
49        self.set_module_name("qatEmbeddingBag")
50
51    def forward(self, input, offset):
52        return self.embedding(input, offset)
53
54
55# Currently, EmbeddingBag QAT does not support sparse embeddings.
56embeddingbag_short_dense_configs = [
57    config
58    for config in configs.embeddingbag_short_configs
59    if {"sparse": True} not in config
60]
61
62op_bench.generate_pt_test(embeddingbag_short_dense_configs, QATEmbeddingBagBenchmark)
63op_bench.generate_pt_gradient_test(
64    embeddingbag_short_dense_configs, QATEmbeddingBagBenchmark
65)
66
67
68class QATEmbeddingBenchmark(op_bench.TorchBenchmarkBase):
69    def init(self, num_embeddings, embedding_dim, input_size, device):
70        qconfig = default_embedding_qat_qconfig
71        self.embedding = nnqat.Embedding(
72            num_embeddings=num_embeddings,
73            embedding_dim=embedding_dim,
74            qconfig=qconfig,
75            device=device,
76        )
77        self.embedding.qconfig = default_embedding_qat_qconfig
78        numpy.random.seed((1 << 32) - 1)
79        self.input = torch.tensor(
80            numpy.random.randint(0, num_embeddings, input_size), device=device
81        ).long()
82        self.inputs = {"input": self.input}
83        self.set_module_name("qatEmbedding")
84
85    def forward(self, input):
86        return self.embedding(input)
87
88
89op_bench.generate_pt_test(configs.embedding_short_configs, QATEmbeddingBenchmark)
90op_bench.generate_pt_gradient_test(
91    configs.embedding_short_configs, QATEmbeddingBenchmark
92)
93
94if __name__ == "__main__":
95    op_bench.benchmark_runner.main()
96