) -> 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,

import java.util.HashMap;
import java.util.Map;
import java.util.Set;
import java.util.function.Function;

/**
 * This factory is responsible for caching merging queries. It delegates computations
 * to another factory, so if the delegate returns the same instances for the same
 * queries, caching will be faster.
 */
public class CachingExcludeFactory extends DelegatingExcludeFactory {

    computationLatch.countDown();
    // let some computations complete
    while (computedCount.get() < nThreads) {
      Thread.yield();
    }
    cache.invalidateAll();
    tasksFinished.await();

    // Check all of the removal notifications we received: they should have had correctly-associated
    // keys and values. (An earlier bug saw removal notifications for in-progress computations,
    // which had real keys with null values.)

      if (!operand_type.hasStaticShape() || !indices_type.hasStaticShape() ||
          !updates_type.hasStaticShape()) {
        return failure();
      }

      // Match the scatter computation against computations supported by TF.
      if (failed(MatchBinaryReduceFunction<BinaryOp>(
              scatter_op.getUpdateComputation()))) {
        return failure();
      }

* The cache has a builtin test, enabled with `UNSAFE_PILOT_ENABLE_RUNTIME_ASSERTIONS=true`, that runs in CI. This will panic if any key is written to with a different value.

#### Partial Computations

Along with caching, partial computations are a critical performance optimization to ensure that we do not need to build (or send) every resource to every proxy on every change. This is discussed more in the Config Serving section.

// Copyright 2020 The Go Authors. All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.

// This file implements helper functions for scope position computations.

package syntax

// StartPos returns the start position of n.
func StartPos(n Node) Pos {
	// Cases for nodes which don't need a correction are commented out.
	for m := n; ; {
		switch n := m.(type) {
		case nil:

    mlir::TF::RuntimeDevices devices, std::string* host_device);

// Parses XLA compilation and execution devices from a tf_device.cluster and
// returns the host device for the head and tail computations. For TPU device,
// if the computation is replicated, GetDeviceAliasForHostOfLogicalCore(0) is
// returned instead.
mlir::LogicalResult GetHostDeviceOutsideComputation(
    mlir::TF::RuntimeDevices devices, mlir::tf_device::ClusterOp cluster,

// suggest a representation, namely using 1-1-1 00:00:00 UTC as the
// epoch, and that's what we do.
//
// The Add and Sub computations are oblivious to the choice of epoch.
//
// The presentation computations - year, month, minute, and so on - all
// rely heavily on division and modulus by positive constants. For
// calendrical calculations we want these divisions to round down, even

      location, output_type, concat_dimension_op.getOutput(), inputs);
}

// For tile sharded inputs to TPU computation, inject split op between the
// input values and TPU computation so that tiled input values are passed in
// as inputs to TPU computations. If more than one dimension is sharded, then
// a tree of connected split ops are added before tf_device.parallel_execute op.