#include "tensorflow/core/common_runtime/optimization_registry.h"

namespace tensorflow {

// PRE_PLACEMENT passes:

// EncapsulateXlaComputationsPass rewrites computations generated by the
// xla.compile() Python code into XlaLaunch nodes.
REGISTER_OPTIMIZATION(OptimizationPassRegistry::PRE_PLACEMENT, 36,
                      EncapsulateXlaComputationsPass);

// from

	metricComputationTotal = metrics.NewCounterVec(
		&metrics.CounterOpts{
			Subsystem:      hpaControllerSubsystem,
			Name:           "metric_computation_total",
			Help:           "Number of metric computations. The label 'action' should be either 'scale_down', 'scale_up', or 'none'. Also, the label 'error' should be either 'spec', 'internal', or 'none'. The label 'metric_type' corresponds to HPA.spec.metrics[*].type",

limitations under the License.
==============================================================================*/
//
// This file defines support utilities for interoperating with FakeQuant* based
// QAT (Quantized Aware Training) computations, as implemented by TFLite. Note
// that FakeQuant* operators mix multiple concerns specific to how TFLite
// originally implemented quantization. As such, utilities here enforce

// * create_while_ctx: if true, a WhileContext is created and populated for this
//     loop. See core/graph/while_context.h for more details on
//     WhileContexts. This is set to false for loops used as part of gradient
//     computations, since they're part of the gradient for a loop in the
//     forward-pass.
//     TODO(skyewm): revisit this. Should we create WhileContexts for all loops,
//     even if we don't need them?

          compilation_result->computation->proto(), xla::DebugOptions()));

  TF_ASSIGN_OR_RETURN(
      std::unique_ptr<xla::HloModule> hlo_module,
      xla::HloModule::CreateFromProto(compilation_result->computation->proto(),
                                      hlo_module_config));

  std::string all_computations;
  for (auto computation : hlo_module->computations()) {

	// false
}

// This example shows how to use big.Float to compute the square root of 2 with
// a precision of 200 bits, and how to print the result as a decimal number.
func Example_sqrt2() {
	// We'll do computations with 200 bits of precision in the mantissa.
	const prec = 200

	// Compute the square root of 2 using Newton's Method. We start with
	// an initial estimate for sqrt(2), and then iterate:

}

def TfDevice_ReplicateOp : TfDevice_Op<"replicate",
    [SingleBlockImplicitTerminator<"ReturnOp">, AttrSizedOperandSegments]> {
  let summary = "Wraps an N-way replicated computation.";

  let description = [{
The region held by this operation represents a computation that is replicated
across multiple devices. The number of replications is based on the `n`
attribute. Explicit devices can be populated in the `devices` attribute, and it

                            ) -> tensor<?x256xf32> {

  // This is a simplified ResNet layer that gets input in NHWC format, converts
  // it to NCHW before padding, and does all computations in NCHW (this is the
  // default setup for ResNet model trained in fp32 on GPU).
  //
  // To be able to use Tensor Cores on latest NVIDIA GPUs this model has to be
  // converted to NHWC data format.

// from TF-Quantizer to stableHLO quantization
bool IsOpWithDataMovementTrait(Operation* op) {
  // Supported data movement ops. These ops do not perform any computations and
  // has one result operand.
  return isa<TF::IdentityOp, TF::CastOp, TF::ReshapeOp, TF::XlaShardingOp,
             TF::GatherOp, TF::GatherV2Op, TF::XlaGatherOp, TF::ExpandDimsOp,

  public void reset() {
    internalReset(0L);
  }

  /**
   * Equivalent in effect to {@link #sum} followed by {@link #reset}. This method may apply for
   * example during quiescent points between multithreaded computations. If there are updates
   * concurrent with this method, the returned value is not guaranteed to be the final
   * value occurring before the reset.
   *
   * @return the sum
   */