func.return %add : tensor<3x2xf32>
    }
  }
)mlir";

// Quantizable ops: XlaCallModule op with "fully_quantizable" attribute and
// same-scale StableHLO ops
// Non-quantizable ops: quantize/dequantize ops
constexpr absl::string_view kModuleCompositeSameScale = R"mlir(
  module {
    func.func @same_scale_after_composite() -> tensor<3x1xf32> {

                                       int32_t qmax) {
  auto quantize = [scale, zero_point](float f) {
    return zero_point + static_cast<int32_t>(std::round(f / scale));
  };

  if (rmin.has_value() && rmax.has_value()) {
    return {std::max(qmin, quantize(rmin.value())),
            std::min(qmax, quantize(rmax.value()))};
  } else if (rmin.has_value()) {
    return {std::max(qmin, quantize(rmin.value())), qmax};
  } else if (rmax.has_value()) {

// RUN: tf-quant-opt %s -split-input-file -quant-lift-quantizable-spots-as-functions -quant-prepare-quantize-drq -quant-quantize='weight-quantization=true' -verify-each=false | FileCheck %s

// -----

module {
  func.func @matmul(%arg0: tensor<1x2x2x3xf32>) -> (tensor<*xf32>) {
    %cst_0 = "tf.Const"() {value = dense<0.000000e+00> : tensor<2x1024xf32>} : () -> tensor<2x1024xf32>

    Value operand = op->getOperand(index);
    auto inserted_value = values->insert(operand).second;
    if (inserted_value) continue;
    // We can only clone the constant op or const->dequantize combo. The latter
    // case is useful for float16 quantization. Since all ops have been
    // legalized to tflite ops, so we only care about ConstOp or QConstOp or
    // mlir constant op.

    // This is the argument used to refer to the pass in
    // the textual format (on the commandline for example).
    return "quant-quantize-weights";
  }

  StringRef getDescription() const final {
    // This is a brief description of the pass.
    return "Quantize weights used by quantizable ops.";
  }

  void getDependentDialects(DialectRegistry& registry) const override {

message QuantizationInfo {
  // min/max of the per axis value range. To quantize the value, the metadata
  // of the target properties should be specified or read from the ops
  // quantization specification.
  message MinMax {
    float min = 1;
    float max = 2;
  }

  // Affine parameters to quantize the per axis value. The metadata of the
  // target properties should be specified as well.

  // CHECK-DAG:  %[[CST_1:.*]] = "tfl.pseudo_const"() <{value = dense<0.000000e+00> : tensor<4x4xf16>}> : () -> tensor<4x4xf16>
  // CHECK-DAG:  %[[DQ_0:.*]] = "tfl.dequantize"(%[[CST_0]]) : (tensor<4x4xf16>) -> tensor<4x4xf32>
  // CHECK-DAG:  %[[DQ_1:.*]] = "tfl.dequantize"(%[[CST_1]]) : (tensor<4x4xf16>) -> tensor<4x4xf32>

// CHECK-NO-UNPACK: %[[DEQUANTIZE:.+]] = stablehlo.uniform_dequantize %[[QUANTIZE_1]] : (tensor<1x3x!quant.uniform<i8:f32, {{.*}}>>) -> tensor<1x3xf32>
// CHECK-NO-UNPACK: return %[[DEQUANTIZE]] : tensor<1x3xf32>

// -----

// Tests that a simple dot_general without CustomAggregators is not quantized.

    "Bitcast",
    "BroadcastArgs",
    "BroadcastTo",
    "CheckNumerics",
    "ConcatV2",
    "ConjugateTranspose",
    "DebugGradientIdentity",
    "DeepCopy",
    "DepthToSpace",
    "Dequantize",
    "EditDistance",
    "Empty",
    "EnsureShape",
    "ExpandDims",
    "ExtractImagePatches",
    "ExtractVolumePatches",
    "FakeQuantWithMinMaxArgs",
    "FakeQuantWithMinMaxArgsGradient",

    METHOD_UNSPECIFIED = 0;  // go/do-include-enum-unspecified

    // Apply default weight-only quantization. Weights are quantized during
    // conversion, then dequantized during inference.
    // Activation: f32, Weight: qi8, Bias: f32
    WEIGHT_ONLY = 1;

    // Apply default dynamic range quantization. Quantized tensor value's
    // ranges are determined during graph runtime.