build:mkl --define=build_with_mkl=true --define=enable_mkl=true
build:mkl --define=tensorflow_mkldnn_contraction_kernel=0
build:mkl --define=build_with_openmp=true
build:mkl -c opt

# config to build OneDNN backend with a user specified threadpool.
build:mkl_threadpool --define=build_with_mkl=true --define=enable_mkl=true
build:mkl_threadpool --define=tensorflow_mkldnn_contraction_kernel=0

    CHECK_EQ(TF_OK, TF_GetCode(s)) << TF_Message(s);
    LOG(INFO) << "Device " << i << " has name " << device_name << ", type "
              << device_type;
    if (string(device_type) == DEVICE_GPU) {
      return device_name;
    }
  }
  // No GPU device found.
  return "";
}

string GPUDeviceName() {
  std::unique_ptr<TF_Status, decltype(&TF_DeleteStatus)> status(

#include "tensorflow/core/util/command_line_flags.h"

#if defined(PLATFORM_WINDOWS)
// Make mkdir resolve to _mkdir to create the test temporary directory.
#include <direct.h>
#define mkdir(name, mode) _mkdir(name)

// Windows defines the following macros to convert foo to fooA or fooW,
// depending on the type of the string argument. We don't use these macros, so
// undefine them here.
#undef CopyFile
#undef DeleteFile

};

TEST_F(CApiFunctionTest, OneOp_ZeroInputs_OneOutput) {
  /*
   *                constant
   *                   |
   *                   v
   */
  // Define
  TF_Operation* c = ScalarConst(10, func_graph_, s_, "scalar10");
  Define(-1, {}, {}, {c}, {});

  // Use, run, and verify
  TF_Operation* func_op = Use({});
  Run({}, func_op, 10);
  VerifyFDef({"scalar10_0"}, {}, {{"scalar10", DT_INT32}},

  // Remote device on `worker1`.
  const char remote_device[] = "/job:localhost/replica:0/task:1/device:CPU:0";
  // `ctx_0`, `ctx_1`, `ctx_2` contains `remote_device`.
  {
    const std::vector<std::string>& device_names = ListDeviceNames(ctx_0);
    ASSERT_TRUE(std::find(device_names.begin(), device_names.end(),
                          remote_device) != device_names.end());
  }

  {

config_setting(
    name = "dynamic_loaded_kernels",
    define_values = {
        "dynamic_loaded_kernels": "true",
        "framework_shared_object": "true",
    },
    visibility = ["//visibility:public"],
)

config_setting(
    name = "using_rocm_hipcc",
    define_values = {"using_rocm_hipcc": "true"},
)

config_setting(
    name = "override_eigen_strong_inline",
    define_values = {"override_eigen_strong_inline": "true"},

    TF_ImportGraphDefOptions* opts, const char* prefix);

// Set the execution device for nodes in `graph_def`.
// Only applies to nodes where a device was not already explicitly specified.
// `device` is copied and has no lifetime requirements.
TF_CAPI_EXPORT extern void TF_ImportGraphDefOptionsSetDefaultDevice(
    TF_ImportGraphDefOptions* opts, const char* device);

// Set whether to uniquify imported operation names. If true, imported operation

  Returns:
    string of native cuda compute capabilities, separated by comma.
  """
  device_query_bin = os.path.join(
      environ_cp.get('CUDA_TOOLKIT_PATH'), 'extras/demo_suite/deviceQuery')
  if os.path.isfile(device_query_bin) and os.access(device_query_bin, os.X_OK):
    try:
      output = run_shell(device_query_bin).split('\n')
      pattern = re.compile('[0-9]*\\.[0-9]*')

/// SECTION 4. Plugin registration and initialization
/// ----------------------------------------------------------------------------
///
/// In this section we define the API used by core TensorFlow to initialize a
/// filesystem provided by a plugin. That is, we define the following:
///   * `TF_InitPlugin` function: must be present in the plugin shared object as
///     it will be called by core TensorFlow when the filesystem plugin is