# installed into a virtual environment, and then that venv is used to run all
# bazel tests with a special flag "--define=no_tensorflow_py_deps=true", which
# drops all the bazel dependencies for each py_test; this makes all the tests
# use the wheel's TensorFlow installation instead of the one made available
# through bazel. This must be done in a different root directory, //bazel_pip/...,

  // Maps from name of an operation to the Node* in 'graph'.
  std::unordered_map<tensorflow::string, tensorflow::Node*> name_map
      TF_GUARDED_BY(mu);

  // The keys of this map are all the active sessions using this graph. Each
  // value records whether the graph has been mutated since the corresponding
  // session has been run (this is detected in RecordMutation function). If the
  // string is empty, no mutation has occurred. Otherwise the string is a

        break;
      case FetchState::FINISHED:
        return TF_SetStatus(status, TF_OK, "");
    }
  }
  return TF_SetStatus(
      status, TF_INTERNAL,
      "Control flow should never reach the end of RamFileBlockCache::Fetch.");
}

int64_t RamFileBlockCache::Read(const std::string& filename, size_t offset,
                                size_t n, char* buffer, TF_Status* status) {
  if (n == 0) {

  // The cache has space for `block_count` blocks. The loop with i = 0 should
  // fill the cache, and the loop with i = 1 should be all cache hits. The
  // fetcher checks that it is called once and only once for each offset (to
  // fetch the corresponding block).
  for (int i = 0; i < 2; i++) {
    for (int j = 0; j < block_count; j++) {
      TF_EXPECT_OK(ReadCache(&cache, "", block_size * j, block_size, &out));
    }
  }

namespace tensorflow {
namespace {

using ::tensorflow::error::Code;

// As we need to test multiple URI schemes we need a parameterized test.
// Furthermore, since each test creates and deletes files, we will use the same
// fixture to create new directories in `SetUp`. Each directory will reside in
// `::testing::TempDir()`, will use a RNG component and the test name. This
// ensures that two consecutive runs are unlikely to clash.

    TF_Tensor* f_t = TFE_TensorHandleResolve(handle, s);
    ASSERT_EQ(TF_OK, TF_GetCode(s)) << TF_Message(s);

    memcpy(&result_data[0], TF_TensorData(f_t), TF_TensorByteSize(f_t));

    // Verify results for each output
    for (int j = 0; j < 4; j++) {
      ASSERT_EQ(result_data[j], expected_outputs[idx][j]);
    }

    TF_DeleteTensor(f_t);
  }

  // Free memory associated with add and MatMul outputs

  EXPECT_EQ(X, dxy_dyValue);

  TF_DeleteTensor(fetchValues[0]);
  TF_DeleteTensor(fetchValues[1]);
}

// REGISTER_OP for CApiAttributesTest test cases.
// Registers two ops, each with a single attribute called 'v'.
// The attribute in one op will have a type 'type', the other
// will have list(type).
#define ATTR_TEST_REGISTER_OP(type)                           \

  // (as in the case of ConstOp kernel creation on GPU, which involves copying a
  // CPU tensor to GPU).
  // Setting a larger thread pool does not help with the Swift caller, as we use
  // a different TFE context for each thread of execution (for running graph
  // functions, and their send/recvs corountines).
  config.set_inter_op_parallelism_threads(1);

  TF_Buffer* ret = TF_NewBuffer();
  TF_CHECK_OK(MessageToBuffer(config, ret));

//
// First we filter the tape to just the subset of operations we want to
// differentiate. In the process of doing so we count how many times each Tensor
// is used as an input to an op (so we know when we're done computing gradients
// for that Tensor). We also count, for each tape entry, how many of its output
// Tensors need gradients to be computed (Tensors which are not used do not need
// any gradients to be computed).
//

        'device at: https://developer.nvidia.com/cuda-gpus. Each capability '
        'can be specified as "x.y" or "compute_xy" to include both virtual and'
        ' binary GPU code, or as "sm_xy" to only include the binary '
        'code.\nPlease note that each additional compute capability '
        'significantly increases your build time and binary size, and that '