// and tfl.dequantize pairs before tf.FakeQuant* being foled.
LogicalResult ConvertFakeQuantOps(func::FuncOp func, MLIRContext* ctx,
                                  bool use_fake_quant_num_bits) {
  OpBuilder builder(func);
  if (failed(UnwrapTFCustomOps(func, builder))) {
    return failure();
  }

  // Insert the tfl.quantize/tfl.dequantize ops after the tf.FakeQuant* ops to

  ];
}
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).
  }];

include "mlir/IR/OpBase.td"
include "mlir/IR/PatternBase.td"
include "mlir/Dialect/Func/IR/FuncOps.td"
include "tensorflow/compiler/mlir/lite/ir/tfl_ops.td"

// Both Quantize and Dequantize ops have side effects, so we have to define
// patterns to remove dead ones after the quantization rewrite.
def : Pat<(TFL_QuantizeOp:$op $in, $qt), (replaceWithValue $in), [(HasNoUseOf:$op)]>;

      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,

}

def QuantConvertSimulatedQuant
    : Pass<"quant-convert-simulated-quantization", "func::FuncOp"> {
  let summary = "Converts training-time simulated quantization ops to "
                "corresponding quantize/dequantize casts";
  let constructor = "mlir::quantfork::createConvertSimulatedQuantPass()";
}

    // Check dequantize performed in bf16.
    func.return %0: tensor<1x2x2x1024xbf16>
  }

// CHECK-LABEL: func @quantize_xladotv2_bf16
// CHECK-DAG: %[[W:.*]] = "tf.Const"() <{value = dense<127> : tensor<2x1024xi8>
// CHECK: %[[IDENTITY:.*]] = "tf.Identity"(%[[W]]) : (tensor<2x1024xi8>) -> tensor<2x1024xi8>

  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(

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.

// `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}>