//go:build (!amd64 && !s390x && !ppc64 && !ppc64le && !arm64) || purego

package aes

import (
	"crypto/cipher"
)

// newCipher calls the newCipherGeneric function
// directly. Platforms with hardware accelerated
// implementations of AES should implement their
// own version of newCipher (which may then call
// newCipherGeneric if needed).
func newCipher(key []byte) (cipher.Block, error) {
	return newCipherGeneric(key)
}

#include "mlir/IR/BuiltinOps.h"  // from @llvm-project
#include "mlir/Pass/Pass.h"  // from @llvm-project

namespace mlir {
namespace TFDevice {

// For architectures that support accelerated embedding lookups, this pass will
// rewrite the graph to use pipelining for better device utilization.
std::unique_ptr<OperationPass<ModuleOp>> CreateEmbeddingSequencingPass();

#include "tensorflow/core/framework/device_base.h"
#include "tensorflow/core/platform/status.h"

namespace tensorflow {

// Helper class for managing data transfers between host and accelerator
// devices using PjRt.
class PjRtDeviceContext : public DeviceContext {
 public:
  explicit PjRtDeviceContext(
      XlaShapeLayoutHelpers::ShapeDeterminationFns shape_determination_fns,

  let summary = "Rewrite graph for embedding pipelining";
  let constructor = "TFDevice::CreateEmbeddingPipeliningPass()";
    let description = [{
    For architectures that support accelerated embedding lookups, this pass will
    rewrite the graph to use pipelining for better device utilization.
  }];
}

def EmbeddingSequencingPass : Pass<"tf-embedding-sequencing", "mlir::ModuleOp"> {

#include "tensorflow/core/common_runtime/local_device.h"
#include "tensorflow/core/framework/device_base.h"

namespace tensorflow {

// tensorflow::PjRtBaseDevice replaces the deprecated tensorflow::XlaDevice.
// This accelerator agnostic device is mainly used to store metadata.
class PjRtBaseDevice : public LocalDevice {
 public:
  // Stores metadata about the PjRtBaseDevice.
  class Metadata {
   public:

        uses: actions/setup-java@v4
        with:
          distribution: 'zulu'
          java-version: 17

      - name: Enable KVM group perms
        # https://github.blog/changelog/2023-02-23-hardware-accelerated-android-virtualization-on-actions-windows-and-linux-larger-hosted-runners/
        run: |
          echo 'KERNEL=="kvm", GROUP="kvm", MODE="0666", OPTIONS+="static_node=kvm"' | sudo tee /etc/udev/rules.d/99-kvm4all.rules

# is a request for 2 SCSI formatted and mounted SSDs and 1 NVMe block device SSD.
NODE_LOCAL_SSDS_EXT=${NODE_LOCAL_SSDS_EXT:-}
# Accelerators to be attached to each node. Format "type=<accelerator-type>,count=<accelerator-count>"
# More information on available GPUs here - https://cloud.google.com/compute/docs/gpus/
NODE_ACCELERATORS=${NODE_ACCELERATORS:-""}
export REGISTER_MASTER_KUBELET=${REGISTER_MASTER:-true}

- Introduced a new alpha feature gate, `SELinuxMount`, which can now be enabled to accelerate SELinux relabeling. ([#123157](https://github.com/kubernetes/kubernetes/pull/123157), [@jsafrane](https://github.com/jsafrane))

    std::function<bool(Operation*)> is_ignored_op) {
  // Iteratively find clusters of different targets within the `block`.
  // Whenever we see an operation that is assigned to an accelerator target
  // (ie. get_target(op) != ""), we try to merge it into the last cluster
  // of same target. If that is infeasible (say because of violating
  // def-before-use), create a new cluster with that operation and move on.

  let summary = "Lifting resource operations out of device computation";
  let description = [{
    This pass lifts resource variable operations outside of device computation.
    This is useful because a lot of accelerator devices can not interact with
    resource variables directly..

    Here is a simple example in TensorFlow where a device doubles the value of a
    TensorFlow resource variable and returns new value:

    ```mlir