values greater than quant_max are converted to 2^num_bits - 1.

    Args:
      quant_min: The minimum real value that can be represented by a quantized
        value.
      quant_max: The maximum real value that can be represented by a quantized
        value.

    Returns:
      (error, quant_min, quant_max): Tuple of weighted mean squared error.

limitations under the License.
==============================================================================*/
// This file contains operator definitions for TF uniform quantized ops.

include "tensorflow/compiler/mlir/tensorflow/ir/tf_op_base.td"
include "mlir/Interfaces/CallInterfaces.td"
include "mlir/Interfaces/InferTypeOpInterface.td"
include "mlir/IR/OpAsmInterface.td"

// RUN: tf-quant-opt %s -split-input-file -quant-convert-fake-quant-to-qdq -quant-lift-quantizable-spots-as-functions='target-opset=XLA' -quant-insert-quantized-functions -quant-quantize-composite-functions='target-opset=XLA' -symbol-dce -inline -tf-shape-inference -canonicalize -quant-replace-cast-hacks-with-tf-xla-ops -cse -quant-optimize | FileCheck %s

module attributes {tf.versions = {bad_consumers = [], min_consumer = 12 : i32, producer = 1219 : i32}, tf_saved_model.semantics} {

See the License for the specific language governing permissions and
limitations under the License.
==============================================================================*/
// Functions for quantization specifications of Uniform Quantized ops.

#ifndef TENSORFLOW_COMPILER_MLIR_QUANTIZATION_TENSORFLOW_OPS_UNIFORM_OP_QUANT_SPEC_H_
#define TENSORFLOW_COMPILER_MLIR_QUANTIZATION_TENSORFLOW_OPS_UNIFORM_OP_QUANT_SPEC_H_

#include <memory>

namespace mlir::quant::stablehlo {

// Performs post-calibration graph transformation as part of post-training
// static-range quantization.
//
// The resulting `ModuleOp` contains quantized StableHLO ops serialized in
// `TF::XlaCallModuleOp`s. They are quantized using the statistics collected
// after the calibration step, corresponding to each `TF::CustomAggregatorOp`s
// in the input module op.
//

  "32-bit quantized integer">;
def TF_Quint8  : AnyTypeOf<
  [TF_TensorFlowType<"Quint8", "quint8">, TF_Quint8Ref],
  "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">;

// RUN: tf-quant-opt %s -quant-insert-quantized-functions='quantization-method=weight_only target-opset=XLA' | FileCheck %s

// Empty module
module {
  func.func @simple_fn(%arg0: tensor<*xf32>) -> tensor<*xf32> {
    func.return %arg0 : tensor<*xf32>
  }
}

// CHECK-NOT: func private @internal_dequantize_f32
// CHECK-NOT: func private @internal_conv3d_fn
// CHECK-NOT: func private @internal_batch_matmul_fn

  static bool IsQuantizableCustomOp(Operation* op,
                                    const quant::CustomOpMap& custom_op_map) {
    // In some cases, ops may need to be quantized even though their op trait is
    // not quantizable. For example, for the case of custom op various ops can
    // be categorized as cusom ops despite each of them may require different

include "tensorflow/compiler/mlir/lite/ir/tfl_ops.td"

// Quantize attribute $0 by using quantization parameter from %1.
def QuantizeByQuantizedType : NativeCodeCall<"quant::Quantize($0, $1.getValue())">;
def F32ElementsAttr : ElementsAttrBase<
  CPred<"$_self.cast<ElementsAttr>().getShapedType().getElementType().isF32()">, "float constant tensor">;

// Squash tfl.dequantize and tfl.quantize pairs.

          clEnumValN(OpSet::XLA, "XLA", "Uses TF XLA ops"),
          clEnumValN(OpSet::UNIFORM_QUANTIZED, "UNIFORM_QUANTIZED",
                     "Uses TF Uniform Quantized ops"))};

  Option<QuantMethod> quantization_method_{
      *this, "quantization-method",
      llvm::cl::init(tensorflow::quantization::QuantizationMethod::
                         METHOD_STATIC_RANGE_INT8),