optimizations. Setting the variable to 0 or unsetting it will
            disable the optimizations.
        *   These optimizations can yield slightly different numerical results
            from when they are off due to floating-point round-off errors from
            different computation approaches and orders.
        *   To verify that the optimizations are on, look for a message with

#### Debug builds

When [building Tensorflow](https://www.tensorflow.org/install/source), passing
`--config=dbg` to Bazel will build with debugging information and without
optimizations, allowing you to use GDB or other debuggers to debug C++ code. For
example, you can build the pip package with debugging information by running:

```bash

// as if passed in XLA_FLAGS. This has global effect.
TF_CAPI_EXPORT void TF_SetXlaAutoJitMode(const char* mode);

// Returns whether the single GPU or general XLA auto jit optimizations are
// enabled through MarkForCompilationPassFlags.
TF_CAPI_EXPORT unsigned char TF_GetXlaAutoJitEnabled();

// Sets XLA's minimum cluster size. This has global effect.

// --------------------------------------------------------------------------
TF_SessionOptions* TF_NewSessionOptions() {
  TF_SessionOptions* out = new TF_SessionOptions;
  // Disable optimizations for static graph to allow calls to Session::Extend.
  out->options.config.mutable_experimental()
      ->set_disable_optimize_for_static_graph(true);
  return out;
}

  // Add a config_proto attr, to trigger grappler graph rewrites in the current
  // eager runtime.
  if (enable_grappler) {
    tensorflow::ConfigProto config;
    // Do not skip grappler optimization even for small graphs.
    config.mutable_graph_options()
        ->mutable_rewrite_options()
        ->set_min_graph_nodes(-1);
    string serialized_config;

build:cuda --@local_config_cuda//:enable_cuda

# CUDA: This config refers to building CUDA op kernels with clang.
build:cuda_clang --config=cuda
# Enable TensorRT optimizations https://developer.nvidia.com/tensorrt
build:cuda_clang --config=tensorrt
build:cuda_clang --action_env=TF_CUDA_CLANG="1"
build:cuda_clang --@local_config_cuda//:cuda_compiler=clang

-       | ((const unsigned char *) (const char *) (src))[idx + 2]) << 8	      \
-      | ((const unsigned char *) (const char *) (src))[idx + 3])
-# endif
 #else
 /* These are a few types we need for the optimizations if we cannot
    use unaligned memory accesses.  */
@@ -94,148 +72,11 @@ __STRING2_COPY_TYPE (8);
 
 /* Set N bytes of S to C.  */
 #if !defined _HAVE_STRING_ARCH_memset
-# if !__GNUC_PREREQ (3, 0)

         * All you need is a good internet connection and you are all set.
      * Try to build TF from sources by changing CPU optimization flags.
   
   *Please let us know if this helps.*
   
windows_comment: >
   From the stack trace it looks like you are hitting windows path length limit.
      * Try to disable path length limit on Windows 10.

};

// Returns the number of elements in the gradient tensor.
int64_t TapeVSpace::NumElements(AbstractTensorHandle* tensor) const {
  // TODO(srbs): It seems like this is used only for performance optimization
  // and not for correctness. The only downside of keeping this 1 seems to be
  // that the gradient accumulation is unbounded and we will never
  // aggressively aggregate accumulated gradients to recover memory.

#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