RewritePatternSet GetHardwareRewritePatterns(MLIRContext* context,
                                             const std::string& hardware);

// Convert quantized ops to float, this will essentially insert dequantize &
// quantize pair around the op.
void ConvertQuantizedOpToFloat(func::FuncOp func, OpBuilder* builder);

// This will optimize the quantized ops -> float graph.
void OptimizeQuantizedOpToFloat(func::FuncOp func, MLIRContext* context);

// RUN: tf-quant-opt %s -split-input-file -quant-prepare-quantize | FileCheck %s

module {
  func.func @same_scale_test(%arg0: tensor<*xf32>) -> tensor<*xf32> {
    %cst = arith.constant dense<[-1, 144]> : tensor<2xi32>
    %cst_1 = arith.constant dense<1.0> : tensor<144x10xf32>
    %cst_2 = arith.constant dense<0.1> : tensor<10xf32>
    %0 = "quantfork.qcast"(%arg0) : (tensor<*xf32>) -> tensor<*x!quant.uniform<i8:f32, 0.05:-10>>

// `tfl.quantize` or `tfl.dequantize` ops. ui8, i8 and i16 are supported.
bool IsSupportedByTfliteQuantizeOrDequantizeOps(IntegerType storage_type) {
  if (storage_type.getWidth() == 8 ||
      (storage_type.isSigned() && storage_type.getWidth() == 16)) {
    return true;
  }
  LLVM_DEBUG(llvm::dbgs()
             << "Uniform quantize / dequantize op only supports ui8, i8 or "

    enable_whole_model_verify,
    denylisted_ops,
    denylisted_nodes,
    enable_variable_quantization,
    disable_per_channel_for_dense_layers,
    debug_options_str,
):
  """Wraps experimental mlir quantize model."""
  return _pywrap_converter_api.ExperimentalMlirQuantizeModel(
      input_data_str,
      disable_per_channel,
      fully_quantize,
      inference_type,
      input_data_type,
      output_data_type,

        ),
        tags={tag_constants.SERVING},
        signature_keys=['serving_default'],
    )

    model = quantize_model.quantize(
        temp_path,
        quantization_options=quantization_options,
        representative_dataset=data_gen(),
    )
    return model

  @test_util.run_in_graph_and_eager_modes

  if (devce_hardware == nullptr) return {};
  return devce_hardware->IsOpSupported(op);
}

// ================== Convert Quantized Op ============================

// Walk through the func and convert the quantize ops to their float version.
void ConvertQuantizedOpToFloat(mlir::func::FuncOp func, OpBuilder* builder) {
  func.walk([&](Operation* op) {
    // TODO(renjieliu): Find a generic way to deal with const ops.

// going forward.
// NOLINTNEXTLINE
llvm::cl::list<std::string> custom_opdefs(
    "tf-custom-opdefs", llvm::cl::desc("List of custom opdefs when importing "
                                       "graphdef"));

// Quantize and Dequantize ops pair can be optionally emitted before and after
// the quantized model as the adaptors to receive and produce floating point
// type data with the quantized model. Set this to `false` if the model input is

import com.google.common.collect.ImmutableSet;
import com.google.common.collect.Range;
import java.util.Random;

/** Benchmarks some algorithms providing the same functionality as {@link Quantiles}. */
public class QuantilesBenchmark {

  private static final ContiguousSet<Integer> ALL_DECILE_INDEXES =
      ContiguousSet.create(Range.closed(0, 10), DiscreteDomain.integers());

  // It is known that `op` is `ModuleOp` when `pass` is
  // `QuantizeCompositeFunctionPass`, but the check is still performed to be
  // defensive.
  return pass != nullptr &&
         pass->getArgument() == "stablehlo-quantize-composite-functions" &&
         isa_and_nonnull<ModuleOp>(op);
}

// Report is saved only when:
// * After running `QuantizeCompositeFunctionPass`.
// * The pass is run on `ModuleOp`.

// RUN: tf-quant-opt %s -split-input-file -quant-prepare-quantize='post-training-quantize=true' | FileCheck %s

// -----

module {
  func.func @same_scale_ptq_test(%arg0: tensor<*xf32>) -> tensor<*xf32> {
    %cst = arith.constant dense<[-1, 144]> : tensor<2xi32>
    %cst_1 = arith.constant dense<1.0> : tensor<144x10xf32>
    %cst_2 = arith.constant dense<0.1> : tensor<10xf32>
    %0 = "quantfork.stats"(%arg0) {