];
}
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).
  }];

      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,

  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(

// `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 "

  // CHECK-DAG: %[[VAL2:.+]] = "tfl.dequantize"(%[[VAL0]])
  // CHECK-DAG: %[[VAL3:.+]] = "tfl.dequantize"(%[[VAL1]])
  // CHECK-DAG: %[[VAL4:.+]] = "tfl.conv_2d"(%arg0, %[[VAL2]], %[[VAL3]]) <{dilation_h_factor = 1 : i32, dilation_w_factor = 1 : i32, fused_activation_function = "NONE", padding = "SAME", stride_h = 1 : i32, stride_w = 1 : i32}>

    left as is for weight-only which means the weight is dequantized at runtime.

    For example, if the kernel does not support dynamic range quantization the
    graph will be converted into the following IR:

    %q_w = "tfl.pseudo_qconst"() {
         qtype = tensor<64x3x3x3x!quant.uniform<i8<-127:127>:f32, 1.000000e+00>>
    %w = "tfl.dequantize"(%q_w) :

// 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

// CHECK-DAG:       %[[VAL_8:.*]] = "tfl.pseudo_const"(){{.*}}dense<[384, 128]> : tensor<2xi32>
// CHECK:           %[[VAL_9:.*]] = "tfl.dequantize"(%[[VAL_0]]) {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<384x512x!quant.uniform<i8:f32, 1.000000e-01>>) -> tensor<384x512xf32>

    * A tensor is dequantized using a `func::FuncOp` whose name contains
      "uniform_dequantize". The first argument is the tensor to be quantized,
      the second argument is the zero point constant (element type: int) and
      the third argument is the inverse scale constant (element type: float).
    * Inputs to the target quantized op is quantized and the outputs are
      dequantized.

         llvm::make_early_inc_range(var_handle_op.getResult().getUsers())) {
      auto read_variable_op = dyn_cast_or_null<ReadVariableOp>(var_handle_user);
      if (!read_variable_op) continue;
      // Add dequantize.
      builder.setInsertionPointAfter(read_variable_op);
      auto new_read_variable_op =
          builder.create<ReadVariableOp>(read_variable_op.getLoc(), ref_qtype,