// is okay, as stat requests are typically fast, and concurrent requests are
    // often for the same file. Future work can split this up into one lock per
    // key if this proves to be a significant performance bottleneck.
    absl::MutexLock lock(&mu_);
    if (LookupLocked(key, value)) {
      return TF_SetStatus(status, TF_OK, "");
    }
    compute_func(key, value, status);
    if (TF_GetCode(status) == TF_OK) {

#include "tensorflow/c/c_api_macros.h"
#include "tensorflow/c/eager/c_api.h"

#ifdef __cplusplus
extern "C" {
#endif

// Resets `op_to_reset` with `op_or_function_name` and `raw_device_name`. This
// is for performance optimization by reusing an exiting unused op rather than
// creating a new op every time. If `raw_device_name` is `NULL` or empty, it
// does not set the device name. If it's not `NULL`, then it attempts to parse

TF_CAPI_EXPORT extern void TF_DefaultThreadOptions(TF_ThreadOptions* options);

// Returns a new thread that is running work_func and is identified
// (for debugging/performance-analysis) by thread_name.
//
// The given param (which may be null) is passed to work_func when the thread
// starts. In this way, data may be passed from the thread back to the caller.
//

  DEVICE_PLACEMENT_EXPLICIT = 0,
  // Copy the tensor to the right device but log a warning.
  DEVICE_PLACEMENT_WARN = 1,
  // Silently copy the tensor, which has a performance cost since the operation
  // will be blocked till the copy completes. This is the default policy.
  DEVICE_PLACEMENT_SILENT = 2,
  // Placement policy which silently copies int32 tensors but not other dtypes.

  TFE_DEVICE_PLACEMENT_EXPLICIT = 0,
  // Copy the tensor to the right device but log a warning.
  TFE_DEVICE_PLACEMENT_WARN = 1,
  // Silently copy the tensor, which has a performance cost since the operation
  // will be blocked till the copy completes. This is the default placement
  // policy.
  TFE_DEVICE_PLACEMENT_SILENT = 2,

  if (!s.ok()) {
    return s;
  }

  std::unordered_map<int64_t, int64_t> gradients_size;
  // TODO(apassos) multiple threads could be dequeuing from op_stack at the same
  // time, for better CPU backprop performance.
  VLOG(1) << "Initial stack:";
  if (VLOG_IS_ON(1)) {
    for (auto t : op_stack) {
      VLOG(1) << "  " << t;
    }
  }
  while (!op_stack.empty()) {
    const int64_t op = op_stack.back();