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

        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();
}

}

// 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

// 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

    "quantfork::QuantizationForkDialect"
  ];
  let options = [
      ListOption<"quantize_allowlist_", "quantize-allowlist", "std::string",
                 "comma separated list of allowlisted functions to be quantized. Only used in tests">,
      Option<"quantize_signed_", "quantize-signed", "bool", "false",
             "signed inference type. Only used in tests">,

  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_

class UsedBy<string op> : Constraint<
  CPred<"llvm::isa<mlir::TFL::" # op # "Op>(*$0.getUsers().begin())">>;

// When the op is passing-through, the output types of the quantized ops need
// to be updated as well. Since the quantize op manages its own type by the
// "qtype" attribute, we should update the type shape in this attribute.
def ReorderTransposeDequantQuant :
      Pat<(TF_TransposeOp:$old_value