Function BM_AvgPool

tensorflow/core/kernels/nn_ops_test.cc:818–873

AvgPooling Op */

AvgPooling Op
*/
static void BM_AvgPool(int iters, int batch_size, int rows, int cols, int depth,
                       int kernel_rows, int kernel_cols, int stride,
                       Padding padding, int num_threads, const string& label) {
  tensorflow::testing::StopTiming();
  std::unique_ptr<Device> device(
      DeviceFactory::NewDevice("CPU", {}, "/job:a/replica:0/task:0"));

  thread::ThreadPool threadpool(Env::Default(), "test", num_threads);
  Eigen::ThreadPoolDevice eigen_cpu_device(threadpool.AsEigenThreadPool(),
                                           num_threads);
  device->set_eigen_cpu_device(&eigen_cpu_device);

  gtl::InlinedVector<TensorValue, 4> inputs;
  TensorShape shape1({batch_size, rows, cols, depth});
  Tensor input1(DT_FLOAT, shape1);
  test::FillIota<float>(&input1, 1.0);
  inputs.push_back({nullptr, &input1});

  // AvgPooling op.
  NodeDef avgpool_node_def;
  CHECK_EQ(kernel_rows, kernel_cols);
  Status status = NodeDefBuilder("avgpool_op", "AvgPool")
                      .Input(FakeInput(DT_FLOAT))
                      .Attr("ksize", {1, kernel_rows, kernel_cols, 1})
                      .Attr("strides", {1, stride, stride, 1})
                      .Attr("padding", padding == VALID ? "VALID" : "SAME")
                      .Finalize(&avgpool_node_def);
  TF_CHECK_OK(status);

  std::unique_ptr<OpKernel> op(CreateOpKernel(DEVICE_CPU, device.get(),
                                              cpu_allocator(), avgpool_node_def,
                                              TF_GRAPH_DEF_VERSION, &status));
  TF_CHECK_OK(status);
  OpKernelContext::Params params;
  params.device = device.get();
  params.frame_iter = FrameAndIter(0, 0);
  params.inputs = &inputs;
  params.op_kernel = op.get();
  std::vector<AllocatorAttributes> attrs;
  test::SetOutputAttrs(&params, &attrs);

  std::unique_ptr<OpKernelContext> avgpool_context(
      new OpKernelContext(&params));

  op->Compute(avgpool_context.get());
  testing::UseRealTime();
  tensorflow::testing::StartTiming();
  for (int i = 0; i < iters; ++i) {
    delete avgpool_context->release_output(0).tensor;
    op->Compute(avgpool_context.get());
  }
  tensorflow::testing::StopTiming();
  testing::ItemsProcessed(avgpool_context->mutable_output(0)->NumElements() *
                          iters);
  testing::SetLabel(label);
}

// BS: batch_size