namespace mlir {
namespace TFL {
namespace tac {
namespace {

// We will caculate the total compute cost for each Func Op.
//
// The compute cost is simply an add-up of the costs of all the operations
// within the FuncOp. (Excluding const ops since they're just "data".)
// We will ignore quant/dequant/requant costs within the Func Op as well,
// intuition:
//

    validations:
      required: false

  - type: input
    id: perf-costs
    attributes:
      label: Performance Costs
      description: "What is the compile time cost? What is the run time cost? "
    validations:
      required: false

  - type: textarea
    id: prototype
    attributes:
      label: "Prototype"

// op1 = .... {tac.device = "CPU"}
// op2 = .... {tac.device = "GPU"}
// op3 = .... {tac.device = "GPU"}
// op4 = .... {tac.device = "CPU"}
//
// We will run a separate cost estimation for each op with each different
// hardwares.
//
// ======== IRs after cost-estimation ========
// op1 = .... {tac.device = "CPU"} {"CPU": 10, "GPU": -1.0}
// op2 = .... {tac.device = "GPU"} {"CPU": 20, "GPU": 5.0}

if [[ "$TFCI_DOCKER_PULL_ENABLE" == 1 ]]; then
  # Simple retry logic for docker-pull errors. Sleeps if a pull fails.
  # Pulling an already-pulled container image will finish instantly, so
  # repeating the command costs nothing.
  docker pull "$TFCI_DOCKER_IMAGE" || sleep 15
  docker pull "$TFCI_DOCKER_IMAGE" || sleep 30
  docker pull "$TFCI_DOCKER_IMAGE" || sleep 60
  docker pull "$TFCI_DOCKER_IMAGE"
fi 

apiVersion: release-notes/v2
kind: feature
area: networking
issue:
- 27749
releaseNotes:
- |
  **Updated** the default installation of gateways to not configure clusters for `AUTO_PASSTHROUGH`, reducing memory costs.

upgradeNotes:
- title: "`AUTO_PASSTHROUGH` Gateway mode"
  content: |
    Previously, gateways were configured with multiple Envoy `cluster` configurations for each Service in the cluster, even those

  ASSERT_TRUE(module);

  // Create a cost recorder with fake cost records.
  auto expected_op_cost_map = GetOpCostMap(module.get());
  EXPECT_EQ(expected_op_cost_map.size(), 1);
  unsigned int seed = 23579;
  for (auto& [op_key, cost] : expected_op_cost_map) {
    cost = rand_r(&seed) % 1000;
  }
  tensorflow::tfrt_stub::CostRecorder cost_recorder;
  for (const auto& [op_key, cost] : expected_op_cost_map) {

namespace tensorflow {
namespace tfrt_compiler {

// Analyze costs for tensorflow operations.
//
// The current heuristic used is quite simple, which is to calculate the total
// size of input tensors. The exception is that ops whose cost is irrelevant to
// input sizes, such as tf.Shape and tf.Reshape, are whitelisted to have cheap
// cost. This cost analysis is expected to be used conservatively (eg. use a low

	result := intPool.Get().(*big.Int)
	defer func() {
		intPool.Put(divisor)
		intPool.Put(exp)
		intPool.Put(result)
	}()

	// divisor = 10^(dif)
	// TODO: create loop up table if exp costs too much.
	divisor.Exp(bigTen, exp.SetInt64(int64(dif)), nil)
	// reuse exp
	remainder := exp

	// result = unscaled / divisor
	// remainder = unscaled % divisor
	result.DivMod(unscaled, divisor, remainder)

// If f panics, the returned function will panic with the same value on every call.
func OnceFunc(f func()) func() {
	var (
		once  Once
		valid bool
		p     any
	)
	// Construct the inner closure just once to reduce costs on the fast path.
	g := func() {
		defer func() {
			p = recover()
			if !valid {
				// Re-panic immediately so on the first call the user gets a
				// complete stack trace into f.
				panic(p)
			}
		}()

  double GetOpCost(mlir::Operation* op) const override {
    float cost = 0.0;
    int64_t count;
    if (ArithmeticCountUtilHelper::GetFirstOutputCount(op, &count)) {
      cost = kCPUArithmeticUnitCost * count;
    } else {
      cost = kCPUDefaultFixedValuedCost;
    }
    return cost * InferenceTypeEfficiency(GetInferenceType(op));
  }