quant::QuantizationSpecs quant_specs_;
};

#include "tensorflow/compiler/mlir/lite/utils/generated_op_quant_spec_getters.inc"

// If the weight is applicable to dynamic range quantization, insert Quantize
// and Dequantize ops with either per-axis or per-tensor scale.
class PrepareDynamicRangeQuantizableOp
    : public OpRewritePattern<arith::ConstantOp> {
 public:
  explicit PrepareDynamicRangeQuantizableOp(

      llvm::cl::desc("Whether enable per-channel quantized weights.")};
};

// If the weight is applicable to dynamic range quantization, insert Quantize
// and Dequantize ops with per-tensor scale.
class PrepareDRQQuantizableOp : public OpRewritePattern<arith::ConstantOp> {
 public:
  explicit PrepareDRQQuantizableOp(MLIRContext* context,

  // If `value` is produced by tf.Dequantize op, then return the Dequantize op's
  // input. Otherwise return `value`.
  auto get_real_input_value = [](Value value) -> Value {
    Operation* defining_op = value.getDefiningOp();
    if (auto dequantize = dyn_cast_or_null<TF::DequantizeOp>(defining_op)) {
      return dequantize.getInput();
    } else if (auto dequantize =

        per_axis_type.getStorageTypeMin(), per_axis_type.getStorageTypeMax());
  }

  auto quantize = builder.create<quantfork::QuantizeCastOp>(
      q_op.getLoc(), new_value_type.clone(new_qtype), new_value);
  auto dequantize = builder.create<quantfork::DequantizeCastOp>(
      dq_op.getLoc(), new_value_type, quantize.getResult());
  return dequantize.getResult();
}

  ];
}
def DecomposeHybridQuantizationPass : Pass<"tfl-decompose-hybrid-quantization", "mlir::func::FuncOp"> {
  let summary = "Decomposes hybridge quantization to explicit quantize / dequantize";
  let description = [{
      Decomposes (with explicit quantize/dequantize ops) selected math
      operations which exist in the model with hybrid quantization
      (some arguments/results left in floating point).
  }];

      auto* inst = value.getDefiningOp();
      if (!inst) {
        continue;
      }

      // There could be a Dequantize op after the weight tensor in cases like
      // fp16 post-training quantization. We need to get the weight from the
      // input of the Dequantize op.
      if (isa<DequantizeOp>(inst)) {
        op = inst;
        value = inst->getOperand(0);
        inst = value.getDefiningOp();
        if (!inst) {

  }

  void rewrite(quantfork::DequantizeCastOp op,
               PatternRewriter& rewriter) const final {
    // Rewrite the floating-point ops to the quantized version, by fusing
    // preceding dequantize ops and succeding quantize ops.
    for (Operation* op_with_region : op.getResult().getUsers()) {
      // Collect all the quantized inputs and "clone" the matched op by these
      // inputs.
      SmallVector<Value, 4> inputs;

      func_name, rewriter, quant_type, val_to_dequantize, result_type,
      LogicsForUniformDequanization);

  return dequant_op;
}
}  // namespace

// Generate quantize and dequantize functions with uniform quantization.
std::optional<TF::PartitionedCallOp> ApplyUniformQuantization(
    PatternRewriter& rewriter, TF::ConstOp op,
    tensorflow::quantization::QuantizationComponentSpec& weight_spec) {

    find the quant_min and quant_max that best describe this distribution. To do
    this, we quantize hist_mids using quant_min and quant_max and dequantize
    them again. Then the difference between hist_mids and dequantized hist_mids
    equates to quantization error when using quant_min and quant_max.


    Args:

  // weights but will dequantize them back at runtime which is useful for
  // memory bound case without kernel support available in lower precisions.
  // Used in MLIR dynamic range quantizer.
  bool weight_only_quantization = false;

  // The minimum number of elements in a weights array required to apply
  // quantization. This is especially useful not to quantize small tensors as