@@ -3,6 +3,11 @@
    backported.  */
 #define _ISOMAC 1
 
+#define __GLIBC_USE(F)	__GLIBC_USE_ ## F
+
+# define __attribute_alloc_size__(params) \
+  __attribute__ ((__alloc_size__ params))
+
 #include <iostream>
 
 using namespace std;
diff --git a/sysdeps/aarch64/dl-machine.h b/sysdeps/aarch64/dl-machine.h
index 185402f..bbdeae0 100644
--- a/sysdeps/aarch64/dl-machine.h
+++ b/sysdeps/aarch64/dl-machine.h

                 absl::Span<AbstractTensorHandle*> grad_inputs) override {
    /* Given upstream grad U and a BiasAdd: A + bias, the gradients are:
     *
     *    dA = U
     *    dbias = reduceSum(U, dims = channel_dim)
     */

    AbstractTensorHandle* upstream_grad = grad_outputs[0];
    DCHECK(upstream_grad);

    // Recover data format from forward pass for gradient.
    std::string data_format;

        "//tensorflow:android": [
            "//tensorflow/c/experimental/stream_executor:stream_executor_hdrs",
            "//tensorflow/core:portable_tensorflow_lib_lite",
        ],
        "//conditions:default": [
            ":tf_tensor",
            "//tensorflow/c/experimental/stream_executor",
            "//tensorflow/c/experimental/stream_executor:stream_executor_internal",
            "//tensorflow/core:framework",

    RunMetadata run_metadata_proto;
    result = session->Run(run_options_proto, input_pairs, output_tensor_names,
                          target_oper_names, &outputs, &run_metadata_proto);

    // Serialize back to upstream client, who now owns the new buffer
    if (run_metadata != nullptr) {
      status->status = MessageToBuffer(run_metadata_proto, run_metadata);
      if (!status->status.ok()) return;
    }
  } else {

    /* Given upstream grad U and a mul op A*B, the gradients are:
     *
     *    dA = U * B
     *    dB = A * U
     *
     */

    AbstractTensorHandle* upstream_grad = grad_outputs[0];

    // Gradient for A
    std::string name = "Mul_Grad_A";
    TF_RETURN_IF_ERROR(Mul(ctx, upstream_grad, forward_inputs_[1],

        otherwise = [
            "@local_xla//xla/stream_executor/cuda:all_runtime",
            "@local_xla//xla/stream_executor/cuda:cuda_driver",
            "@local_xla//xla/stream_executor/cuda:cuda_platform",
            "@local_xla//xla/stream_executor/cuda:cudnn_plugin",
            "@local_xla//xla/stream_executor/cuda:cufft_plugin",
            "@local_xla//xla/stream_executor:cuda_platform",
        ],
    ),

  if (TF_GetCode(status) != TF_OK) return -1;
  auto stream = gcs_file->gcs_client.ReadObject(
      bucket, object, gcs::ReadRange(offset, offset + buffer_size));
  TF_SetStatusFromGCSStatus(stream.status(), status);
  if ((TF_GetCode(status) != TF_OK) &&
      (TF_GetCode(status) != TF_OUT_OF_RANGE)) {
    return -1;
  }
  int64_t read;
  auto content_length = stream.headers().find("content-length");

  // Load the library.
  TF_Status* status = TF_NewStatus();
  string lib_path =
      tensorflow::GetDataDependencyFilepath(tensorflow::io::JoinPath(
          "tensorflow", "c", "experimental", "stream_executor", "test",
          "test_pluggable_device.so"));
  TF_Library* lib = TF_LoadPluggableDeviceLibrary(lib_path.c_str(), status);
  TF_Code code = TF_GetCode(status);
  string status_msg(TF_Message(status));

  * Added function `experimental.extension_type.as_dict()`, which converts an instance of `tf.experimental.ExtensionType` to a `dict` representation.
* `stream_executor`
  * Top level `stream_executor` directory has been deleted, users should use equivalent headers and targets under `compiler/xla/stream_executor`.
* `tf.nn`

    python3 -m auditwheel show "$TF_WHEEL" > audit.txt
    grep --quiet -zoP 'is consistent with the following platform tag:\n"manylinux_2_17_(aarch|x86_)64"\.' audit.txt
}

@test "Wheel conforms to upstream size limitations" {
    WHEEL_MEGABYTES=$(stat --format %s "$TF_WHEEL" | awk '{print int($1/(1024*1024))}')
    # Googlers: search for "test_tf_whl_size"
    case "$TF_WHEEL" in
        # CPU: