A. Unique TensorFlower <******@****.***> 1718231025 -0700

        "@llvm-project//mlir:QuantOps",
        "@llvm-project//mlir:Support",
    ],
)

# Binary to apply quantization on the annotated files.
tf_cc_binary(
    name = "tfl_quantizer",
    srcs = [
        "tfl_quantizer.cc",
    ],
    deps = [
        ":quantize_model",
        "//tensorflow/compiler/mlir/lite/schema:schema_fbs",
        "//tensorflow/lite:framework",

// "qtype" attribute, we should update the type shape in this attribute.
def ReorderTransposeDequantQuant :
      Pat<(TF_TransposeOp:$old_value
              (TFL_DequantizeOp (TFL_QuantizeOp $input, $qtype)), $perm),
          (TFL_DequantizeOp (TFL_QuantizeOp
                                (TF_TransposeOp $input, $perm),
                                (UpdateShapeWithAxis<-1> $qtype, $old_value))),

// squashed to a requantize op if the scales are different.
def : Pat<(TFL_QuantizeOp (TFL_DequantizeOp $in), $qt), (replaceWithValue $in)>;

// If the tfl.dequantize op wasn't fused, we shouldn't quantize the floating
// point constant.
def : Pat<(TFL_DequantizeOp
             (TFL_QuantizeOp (Arith_ConstantOp F32ElementsAttr:$cst), $qt)),
          (TFL_ConstOp $cst)>;

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)]>;

// TODO(b/185915462): Compare the scale of input and output. This can also be
// squashed to a requantize op if the scales are different.

                             (TFL_ExpandDimsOp $input, $dim)>;

def LegalizeFakeQuantToDequantizeQuantize : Pat<
  (TF_FakeQuantWithMinMaxArgsOp $inputs, $min, $max, $num_bits, $narrow_range),
  (TFL_DequantizeOp
   (TFL_QuantizeOp $inputs,
                   (ConvertToQuantTypeFromAttrs $inputs, $min, $max,
                    $num_bits, $narrow_range)))>;

def LegalizeFill : Pat<(TF_FillOp $arg, $value), (TFL_FillOp $arg, $value)>;

  "8-bit quantized unsigned integer">;
def TF_Quint16 : AnyTypeOf<
  [TF_TensorFlowType<"Quint16", "quint16">, TF_Quint16Ref],
  "16-bit quantized unsigned integer">;

// Any quantized type
def TF_Quantized : AnyTypeOf<
  [TF_Qint8, TF_Qint16, TF_Qint32, TF_Quint8, TF_Quint16], "quantized">;

def TF_Qint8Tensor : TensorOf<[TF_Qint8]>;
def TF_Qint16Tensor : TensorOf<[TF_Qint16]>;

// incorrect placement in the quantization aware training.
def MatchHardSwishQuantized : Pat<
  (TFL_MulOp (TFL_DequantizeOp (TFL_QuantizeOp
    (TFL_MulOp
     $x, (TFL_DequantizeOp (TFL_QuantizeOp (TFL_AddOp
          $x,
          (Arith_ConstantOp ConstantAttr<RankedF32ElementsAttr<[]>, "3.0f">),
          TFL_AF_Relu6), $qattr2)),
     TFL_AF_None), $qattr1)),

// def UndoBroadcastFullyConnectedBiasAddWithQDQs : Pat<
//   (TFL_AddOp $lhs,
//     (TFL_DequantizeOp
//       (TFL_QuantizeOp
//         (Arith_ConstantOp:$const_op $bias),
//       $qparams)),
//   $act_fn),
//   (TFL_AddOp $lhs,
//     (TFL_DequantizeOp
//       (TFL_QuantizeOp
//         (Arith_ConstantOp:$const_op (FlattenTo1D $bias),
//       $qparams)),
//   $act_fn),