//     * Refer to "TensorFlow Versioning Strategy" section at
//       https://github.com/tensorflow/community/pull/257/files.
//     * Note that the API is still under active development and doesn't have
//       versioning guarantees yet.
//   * `void* ext` is a free-form field that can be populated by
//     a plugin in `TP_*` structs or potential future extension points .
//
// Example usage:
//

// Causes the calling thread to block till all ops dispatched in this executor
// have been executed. Note that "execution" here refers to kernel execution /
// scheduling of copies, etc. Similar to sync execution, it doesn't guarantee
// that lower level device queues (like GPU streams) have been flushed.
//
// This call may not block for execution of ops enqueued concurrently with this
// call.

#ifdef __cplusplus
extern "C" {
#endif

// Test only exports of the monitoring Cell Reader API which allows tests to
// read current values from streamz counters defined in other modules.
//
// The code under test will have created streamz counters like this:
// auto* streamz = tensorflow::monitoring::Counter<1>::New("name",
// "description", "label");
// and then incremented that counter for various values of label:

  // Blocks until the previous `StartExecute` has run `TFE_Execute` on each
  // device. If is_async=false (constructor argument) this means the ops have
  // run and have results. If is_async=true it means that all of the
  // device-specific executors have scheduled the op.
  //
  // Accepts inferred shapes for outputs (`expected_output_shapes`), which if
  // fully defined will avoid querying the shapes of the underlying

//
// Implicit copies onto `device_name` are allowed, replicating the value once
// per device in `underlying_devices`. Implicit copies off of the device throw
// an error.
//
// All component tensors must have the same dtype. Currently they must also have
// the same shape, although this requirement may be relaxed in the future.
//
// `device_name` must not name an existing physical or custom device (see

// inferred from an input and doesn't have a default value you wish to
// keep.

// `value` must point to a string of length `length` bytes.
TF_CAPI_EXPORT extern void TF_SetAttrString(TF_OperationDescription* desc,
                                            const char* attr_name,
                                            const void* value, size_t length);
// `values` and `lengths` each must have lengths `num_values`.

// things like the available devices, resource manager etc.
// TFE_Context must outlive all tensor handles created using it. In other
// words, TFE_DeleteContext() must be called after all tensor handles have
// been deleted (with TFE_DeleteTensorHandle).
//
// TODO(ashankar): Merge with TF_Session?
typedef struct TFE_Context TFE_Context;

TF_CAPI_EXPORT extern TFE_Context* TFE_NewContext(

// have all gradients available. This set usually is a subset of the set of
// targets (not all since targets which have outputs in the tape will not have
// gradients available initially).
//
// Then we repeatedly pop an entry from the stack, run its backprop, and update
// the gradients of its inputs. Once we have computed all gradients for a single

// `shape_inference::InferenceContext` constructor. Note the following:
//   - The inputs of the `op` are not used for shape inference. So, it is
//     OK to not have the inputs properly set in `op`. See `input_tensors`
//     if you want shape inference to consider the input tensors of the
//     op for shape inference.
//   - The types need not be set in `input_shapes` as it is not used.

/// TensorFlow but can be used by each filesystem plugin to represent internal
/// data.
///
/// We prefer to have these structures instead of passing `void*` into
/// method signatures to have some type of type safety: for example, operations
/// that are only valid on random access files have a `TF_RandomAccessFile`
/// argument.
///
/// Lifetime: The wrapper data structures are owned by core TensorFlow. The data