"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
// integer types.
// NOLINTNEXTLINE
opt<bool> emit_quant_adaptor_ops(

}

// CHECK-NOT: stablehlo.transpose
// CHECK: %[[CONV:.+]] = stablehlo.convolution
// CHECK-SAME{LITERAL}: [b, f, 0, 1]x[o, i, 0, 1]->[b, 0, 1, f]
// CHECK-NOT: stablehlo.transpose

// -----

// Tests that a quantized convolution does not match. No conversion occurs.

// CHECK-LABEL: quantized_convolution

#include "mlir/Support/LogicalResult.h"  // from @llvm-project

namespace mlir::quant::stablehlo {

// Decomposes a given floating point value num into a normalized and quantized
// fraction and an integral power of two.
LogicalResult QuantizeMultiplier(double double_multiplier,
                                 int32_t& quantized_fraction, int32_t& shift);

}  // namespace mlir::quant::stablehlo

// This basically assumes pure load/store. This is just fake data.
constexpr float kCPUCopyUnitCost = 0.5;

// Default values.
constexpr float kCPUDefaultFixedValuedCost = 10000.0;

// Quantized inference cost efficiency.
// For CPU, quantized inference is ~3x faster than the float alternative, this
// is just an estimation.
constexpr float kQuantizedInferenceEfficiency = 0.3;

                                RewritePatternSet* patterns);

// Populates TF to MHLO legalization for some of the quantization ops.
//
// TODO(hinsu): Remove this once we combine quantized and non quantized op
// legalization in the ODML conversion pipeline.
void PopulateLegalizeTfQuantizationPatterns(MLIRContext* context,
                                            RewritePatternSet* patterns);

  AddShapeLegalizationPasses(pm);
  pm.addNestedPass<func::FuncOp>(
      CreateConvertCustomAggregationOpToQuantStatsPass());
  pm.addPass(createQuantizeCompositeFunctionsPass(options));
  // Add an inliner pass to inline quantized StableHLO functions.
  pm.addPass(createInlinerPass());
  if (pipeline_config.unpack_quantized_types()) {
    AddStablehloQuantToIntPasses(pm);
  }
}

void AddWeightOnlyQuantizationPasses(

limitations under the License.
==============================================================================*/
// This header file defines common utils used when transforming TF ops to
// Uniform Quantized ops.

#ifndef TENSORFLOW_COMPILER_MLIR_QUANTIZATION_TENSORFLOW_UTILS_TF_TO_UNIFORM_ATTRIBUTE_UTILS_H_
#define TENSORFLOW_COMPILER_MLIR_QUANTIZATION_TENSORFLOW_UTILS_TF_TO_UNIFORM_ATTRIBUTE_UTILS_H_

absl::StatusOr<mlir::quant::QuantizedType> GetQuantizedType(
    const tflite::TensorT& tensor, mlir::Builder builder,
    bool is_constant = false, mlir::Type storage_type = {});

// Imports float tensor with calibration value into calibrated quantized type.
absl::StatusOr<mlir::quant::QuantizedType> GetCalibratedQuantizedType(
    const tflite::TensorT& tensor, mlir::Builder builder);

absl::StatusOr<mlir::TensorType> GetTensorType(const tflite::TensorT& tensor,

  quantized_fraction = static_cast<int32_t>(std::round(fraction * (1L << 15)));
  // Clip extreme values.  These are more than enough to overflow int8, the
  // storage type for quantized values, and the final values will be clamped
  // no matter what.
  if (quantized_fraction == (1L << 15)) {
    quantized_fraction /= 2;
    ++shift;
  }
  if (shift < -15) {
    shift = 0;

namespace mlir::quant::stablehlo {

// A class that manages information about `QuantizableUnit`s post-quantization,
// internally in the form of `QuantizationUnits`. It is used to collect
// quantization summary from a quantized `ModuleOp` and emit it in a human- and
// machine-readable format.
class QuantizationReport {
 public:
  QuantizationReport() = default;

  // Initializes `QuantizationReport` by collecting `QuantizationResults` from