Function BM_MaxPoolBk

tensorflow/core/kernels/nn_ops_test.cc:1109–1146

BM_MaxPoolFwdCPU(32, 14, 14, 576, 3, 3, 2, SAME, 4, "maxpool10_SAME");

static void BM_MaxPoolBk(int iters, int batch_size, int rows, int cols,
                         int depth, int kernel_rows, int kernel_cols,
                         int stride, Padding padding, int num_threads,
                         bool use_gpu, const string& label) {
  auto root = Scope::NewRootScope().ExitOnError();

  int64 out_height, out_width, pad_rows, pad_cols;
  TF_CHECK_OK(GetWindowedOutputSize(rows, kernel_rows, stride, padding,
                                    &out_height, &pad_rows));
  TF_CHECK_OK(GetWindowedOutputSize(cols, kernel_cols, stride, padding,
                                    &out_width, &pad_cols));

  Tensor input_data(DT_FLOAT, TensorShape({batch_size, rows, cols, depth}));
  input_data.flat<float>().setRandom();

  Tensor output_data(DT_FLOAT,
                     TensorShape({batch_size, out_height, out_width, depth}));
  output_data.flat<float>().setRandom();

  Tensor output_diff(DT_FLOAT,
                     TensorShape({batch_size, out_height, out_width, depth}));
  output_diff.flat<float>().setRandom();

  CHECK_EQ(kernel_rows, kernel_cols);
  ops::internal::MaxPoolGrad(root, input_data, output_data, output_diff,
                             {1, kernel_rows, kernel_cols, 1} /* ksize */,
                             {1, stride, stride, 1} /* stride */,
                             padding == VALID ? "VALID" : "SAME");
  TF_CHECK_OK(root.status());
  Graph* g = new Graph(OpRegistry::Global());
  TF_CHECK_OK(root.ToGraph(g));
  string device = use_gpu ? "gpu" : "cpu";
  testing::UseRealTime();
  test::Benchmark(device, g).Run(iters);

  testing::ItemsProcessed(batch_size * rows * cols * depth * iters);
  testing::SetLabel(label);
}

// BS: batch_size
// IR: input_rows