Function TEST

test/singa/test_operation_benchmark.cc:36–64

#ifdef USE_CUDNN
TEST(OperationBenchmark, CrossEntropyFwd) {
  auto cuda = std::make_shared<singa::CudaGPU>();
  auto ctx = cuda->context(0);
  int bs = 64;
  int dim = 10;
  vector<DataType> dtypes = {kFloat16, kFloat32};

  Tensor t(Shape{bs}, cuda);
  t.SetValue(0.0f);
  t = t.AsType(kInt);

  for (auto dtype : dtypes) {
    Tensor p(Shape{bs, dim}, cuda, dtype);
    Uniform(0.0f, 1.0f, &p);

    high_resolution_clock::time_point t1 = high_resolution_clock::now();

    for (int i = 0; i < 1000; ++i) {
      auto l = CrossEntropyFwd(p, t);
      cudaStreamSynchronize(cuda->context(0)->stream);
    }

    high_resolution_clock::time_point t2 = high_resolution_clock::now();
    duration<double> time_span = duration_cast<duration<double>>(t2 - t1);
    cout << " dtype " << dtype;
    cout << " - " << time_span.count() << " sec";
    cout << endl;
  }
}

TEST(OperationBenchmark, Mult) {
  auto cuda = std::make_shared<singa::CudaGPU>();