xref: /aosp_15_r20/external/pytorch/benchmarks/fastrnns/test_bench.py (revision da0073e96a02ea20f0ac840b70461e3646d07c45)
1import pytest
2
3import torch
4
5from .fuser import set_fuser
6from .runner import get_nn_runners
7
8
9@pytest.fixture(scope="class")
10def modeldef(request, net_name, executor, fuser):
11    set_fuser(fuser, executor)
12
13    # Given a 'net_name' provided by generate_tests, build the thing
14    name, rnn_creator, context = get_nn_runners(net_name)[0]
15    creator_args = creator_args = {
16        "seqLength": 100,
17        "numLayers": 1,
18        "inputSize": 512,
19        "hiddenSize": 512,
20        "miniBatch": 64,
21        "device": "cuda",
22        "seed": None,
23    }
24    return rnn_creator(**creator_args)
25
26
27def cuda_sync(func, *args, **kwargs):
28    out = func(*args, **kwargs)
29    torch.cuda.synchronize()
30    return out
31
32
33@pytest.mark.benchmark(
34    warmup=True,
35    warmup_iterations=3,
36    disable_gc=True,
37    max_time=0.1,
38    group="fastrnns",
39)
40class TestBenchNetwork:
41    # See 'modeldef' fixture, which provides the things to benchmark
42    def test_forward(self, modeldef, benchmark):
43        forward_output = benchmark(cuda_sync, modeldef.forward, *modeldef.inputs)
44
45    def test_backward(self, modeldef, benchmark):
46        backward_input = modeldef.forward(*modeldef.inputs)
47        if modeldef.backward_setup is not None:
48            backward_input = modeldef.backward_setup(backward_input)
49
50        if modeldef.backward is not None:
51            benchmark(cuda_sync, modeldef.backward, *backward_input, retain_graph=True)
52
53            with torch.no_grad():
54                for param in modeldef.params:
55                    assert param.grad is not None
56                    param.grad.zero_()
57