"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(

      != _PresetMethod.METHOD_STATIC_RANGE_WEIGHT_ONLY_INT8
  ):
    raise ValueError(
        'StableHLO quantized opset currently only supports static range'
        ' quantization and weight-only quantizationvia TF Quantizer.'
    )

  # Set `force_graph_mode_calibration` to True to avoid skipping op execution,
  # which are not connected to return ops, during calibration execution.

  // The following two passes find specific uniform quantization patterns in
  // StableHLO and converts them to TFLite ops that accept or produce uniform
  // quantized types. They only target a specific set of models that contain
  // "decomposed" quantized ops produced from the framework level. This is why
  // they are placed right after the `LegalizeTFXlaCallModuleToStablehloPass`

    return success();
  }
};

// UniformDequantizeOp takes TF quantized types as input which would have been
// converted to the mhlo quantized types. Use OpConversionPattern in order to
// retrieve the operand type *after* conversion, using OpAdaptor operand
// accessor.
// Same for other Uniform Quant Ops that take TF quantized types as input.
class ConvertUniformDequantizeOp

}

// 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

        per_axis_type.getStorageTypeMin(), per_axis_type.getStorageTypeMax());
  }

  auto quantize = builder.create<quantfork::QuantizeCastOp>(
      q_op.getLoc(), new_value_type.clone(new_qtype), new_value);
  auto dequantize = builder.create<quantfork::DequantizeCastOp>(
      dq_op.getLoc(), new_value_type, quantize.getResult());
  return dequantize.getResult();
}

// 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);

#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

  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(