1#include <gtest/gtest.h>
2
3#include "test/cpp/jit/test_utils.h"
4#include "torch/csrc/jit/runtime/graph_executor.h"
5#include "torch/jit.h"
6#include "torch/script.h"
7#include "torch/torch.h"
8
9namespace torch {
10namespace jit {
11
12TEST(GraphExecutorTest, Basic_CUDA) {
13 constexpr int batch_size = 4;
14 constexpr int input_size = 256;
15
16 int hidden_size = 2 * input_size;
17
18 auto input = at::randn({batch_size, input_size}, at::kCUDA);
19 auto hx = at::randn({batch_size, hidden_size}, at::kCUDA);
20 auto cx = at::randn({batch_size, hidden_size}, at::kCUDA);
21 auto w_ih = t_def(at::randn({4 * hidden_size, input_size}, at::kCUDA));
22 auto w_hh = t_def(at::randn({4 * hidden_size, hidden_size}, at::kCUDA));
23
24 auto g = build_lstm();
25 GraphExecutor executor(g, "");
26 auto stack = createStack({input, hx, cx, w_ih, w_hh});
27 executor.run(stack);
28 ASSERT_EQ(stack.size(), 2);
29 at::Tensor r0, r1;
30 std::tie(r0, r1) = lstm(input, hx, cx, w_ih, w_hh);
31 ASSERT_TRUE(almostEqual(stack[0].toTensor(), r0));
32 ASSERT_TRUE(almostEqual(stack[1].toTensor(), r1));
33}
34
35TEST(GraphExecutorTest, runAsync_executor) {
36 /*
37 TODO: there are some problem with C++ parsing script program involving
38 fork. Use the test module below for now.
39 issue about this: github.com/pytorch/pytorch/issues/46368
40 The test module file is generated by following:
41 class DemoModule(torch.nn.Module):
42 def forward(self):
43 r1 = torch.jit.fork(torch.mm, torch.rand(100,100),torch.rand(100,100))
44 r2 = torch.jit.fork(torch.mm, torch.rand(100,100),torch.rand(100,100))
45 return r1.wait() + r2.wait()
46 demo = DemoModule()
47 torch.jit.save(torch.jit.script(demo), 'test_interpreter_async.pt')
48 */
49 std::string filePath(__FILE__);
50 auto testModelFile = filePath.substr(0, filePath.find_last_of("/\\") + 1);
51 testModelFile.append("test_interpreter_async.pt");
52 auto module = load(testModelFile);
53 auto graph = module.get_method("forward").graph();
54 GraphExecutor graphExecutor(graph, "");
55 auto asyncCounter = 0;
56 std::mutex mtx;
57 // a dummy executor which actually use at::launch, but add up a counter
58 auto launcher = [&](std::function<void()> f) {
59 mtx.lock();
60 ++asyncCounter;
61 mtx.unlock();
62 at::launch(std::move(f));
63 };
64 std::vector<IValue> stack;
65 // NOLINTNEXTLINE(modernize-use-emplace)
66 stack.push_back(module._ivalue());
67 graphExecutor.runAsync(stack, launcher)->wait();
68 ASSERT_TRUE(asyncCounter > 0);
69}
70
71} // namespace jit
72} // namespace torch
73