// RUN: stablehlo-quant-opt %s -split-input-file -stablehlo-quantize -verify-each=false | FileCheck %s

// Tests if reduce_window op following quantized function is quantized.

module attributes {tf.versions = {bad_consumers = [], min_consumer = 12 : i32, producer = 1722 : i32}, tf_saved_model.semantics} {
  // CHECK-LABEL: main_00
  // CHECK-SAME: %[[ARG0:.*]]: tensor<2x3x1x1024xf32>

          GetAsVector(expected_tensor->shape()));
}

// Finds the match of the quantized tensor from the possible tensors. Each
// possible tensors can be used only once. It checks shape and name if the
// tensor is quantized and also checks buffer contents and tensor type if not
// quantized. For the quantized case, tensor type and quantizaction params are
// expected to be checked in the test body with the match.

                   QuantizationUnits& quantizable_ops) const {
    bool quantized = false;

    for (auto& quant_op : quantizable_ops) {
      if (quant_specs_.inference_type == tensorflow::DT_QINT8) {
        quantized |= quantizeOpAsInt8(rewriter, op, quant_op);
      }
    }
    return quantized;
  }

 protected:
  QuantizationSpecs quant_specs_;
  OpSet op_set_;

  // unquantized tensors are only inserted in the unquantized model
  // whereas `DumpTensor` ops for the quantized tensors are only inserted
  // in the quantized model. Both models are required to be able to dump
  // both quantized and unquantized tensors and compare them offline.
  if (quantization_options.has_debugger_config() &&
      quantization_options.debugger_config().debugger_type() ==

      std::min(quantized, static_cast<int>(std::numeric_limits<int8_t>::max()));
  quantized =
      std::max(quantized, static_cast<int>(std::numeric_limits<int8_t>::min()));
  return builder.getI32IntegerAttr(quantized);
}

// Decompose the TF ops with the registered composition library.
class DecomposeTFOpsPass
    : public PassWrapper<DecomposeTFOpsPass, OperationPass<func::FuncOp>> {
 public:

    // asymmetric range. For a state tensor, assigning correct quantization
    // parameters is sufficient, and for constants with asymmetric range it's
    // not correctly quantized by legacy quantizer so call the new Quantize.
    return Quantize(real_value, tensor_type);
  } else if (width == 16) {
    if (const auto uniform_type = dyn_cast<UniformQuantizedType>(q_type)) {
      const auto quantized_values =

      attrs.push_back(rewriter.getNamedAttr(
          attr_minmax, rewriter.getI64IntegerAttr(quant_val)));
    }
  }
  return success();
}

// This LogicalResult covers both the hybrid and fully quantized op cases.
LogicalResult FillAttributesForUniformQuantizedDotOp(
    PatternRewriter& rewriter, Operation* op,
    llvm::StringMap<Attribute>& identifier_to_attr,

    if (!mlir::isa<mlir::IntegerType>(raw_elem_type)) {
      return absl::InvalidArgumentError(
          "Quantized tensors must be stored as integers");
    }
    storage_type = mlir::cast<mlir::IntegerType>(raw_elem_type);
  }

  // TFlite uses narrow-range [u]int8 for constant buffers of quantized weights.
  // Since we don't know which ones are weights, we represent this optimization

    %mul = "tf.Mul"(%cast, %scale) : (tensor<*xf32>, tensor<*xf32>) -> tensor<*xf32>
    func.return %mul : tensor<*xf32>
  }

  // Requantizes and clips to the range of quantized type if there is no specific activation.
  func.func private @internal_requantize_no_activation_fn(%accumulation : tensor<*xi32>,
                         %input_scale : tensor<*xf32>, %input_zp : tensor<*xi32>,

            src_saved_model_path, signature_keys, tags, quantization_options);
        if (!exported_model.ok()) return exported_model.status();

        // Remove the `tpu` tag from the debug quantized saved model as it is
        // for CPU. Note the 'tpu' value should be the same as `TPU` defined in
        // tensorflow/python/saved_model/tag_constants.py.
        if (quantization_options.has_debugger_config()) {