if (!device.empty()) {
    LOG(INFO) << "Setting op Min on device " << device;
  }
  TF_Operation* min = MinWithDevice(feed, one, graph, device, s);
  ASSERT_EQ(TF_OK, TF_GetCode(s)) << TF_Message(s);

  // Run the graph.
  csession.SetInputs({{feed, Int32Tensor({3, 2, 5})}});
  csession.SetOutputs({min});
  csession.Run(s);
  ASSERT_EQ(TF_OK, TF_GetCode(s)) << TF_Message(s);

}

TEST(CAPI, Execute_MatMul_CPU_Type_Error) {
  Execute_MatMul_CPU_Type_Error(false);
}
TEST(CAPI, Execute_MatMul_CPU_Type_ErrorAsync) {
  Execute_MatMul_CPU_Type_Error(true);
}
TEST(CAPI, Execute_Min_CPU) {
  TF_Status* status = TF_NewStatus();
  TFE_ContextOptions* opts = TFE_NewContextOptions();
  TFE_Context* ctx = TFE_NewContext(opts, status);
  CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  write_action_env_to_bazelrc('ANDROID_NDK_VERSION', ndk_version)

  # Now grab the NDK API level to use. Note that this is different from the
  # SDK API level, as the NDK API level is effectively the *min* target SDK
  # version.
  meta = open(os.path.join(android_ndk_home_path, 'meta/platforms.json'))
  platforms = json.load(meta)
  meta.close()
  aliases = platforms['aliases']

# ARM64 PYCPP
# In Linux Arm64 presubmit/continuous build, we cross-compile the binaries on
# Linux x86 so that we can use RBE. Since tests still need to run on the single
# host Arm64 machine, the build becomes too slow (~30 min) to be a presubmit.
# For testing purposes, we want to see the runtime performance of an
# experimental job that is build-only, i.e, we only build the test targets and