Function TEST

test/cpp/eager/performance_tests/benchmark_eager_cuda.cc:40–79

#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)

TEST(Benchmark, EagerScaleCUDA) {
  eager_test::InitEnv(phi::GPUPlace());

  for (const std::string mode : {"Accuracy", "WarmUp", "Performance"}) {
    phi::DDim ddim = common::make_ddim({2, 4, 4, 4});
    paddle::Tensor tensor =
        eager_test::CreateTensorWithValue(ddim,
                                          phi::GPUPlace(),
                                          phi::DataType::FLOAT32,
                                          phi::DataLayout::NCHW,
                                          5.0 /*value*/,
                                          true /*is_leaf*/);
    RetainGradForTensor(tensor);

    if (mode == "Accuracy") {
      benchmark_eager_scale(tensor, true /* accuracy_check */);

    } else if (mode == "WarmUp") {
      benchmark_eager_scale(tensor);

    } else if (mode == "Performance") {
      auto t_start = std::chrono::high_resolution_clock::now();
#ifdef WITH_GPERFTOOLS
      ProfilerStart("eager_scale_cuda.out");
#endif
      benchmark_eager_scale(tensor);

#ifdef WITH_GPERFTOOLS
      ProfilerStop();
#endif
      auto t_end = std::chrono::high_resolution_clock::now();
      double elapsed_time_ms =
          std::chrono::duration<double, std::milli>(t_end - t_start).count();
      std::cout << "Duration: " << elapsed_time_ms << " ms" << std::endl;

    } else {
      PADDLE_THROW(common::errors::Fatal("Unknown benchmark mode"));
    }
  }
}

TEST(Benchmark, EagerMatmulCUDA) {
  phi::GPUPlace place;