};

// Matches the pattern for quantized convolution op and rewrites it to use
// uniform quantized types.
//
// Currently assumes asymmetric per-tensor quantization for activations and
// symmetric per-channel quantization for filters.
//
// This pattern represents the following derived equation, where:
// * rn = real (expressed) value for tensor n
// * qn = quantized value for tensor n
// * sn = scale for tensor n

    METHOD_NO_QUANTIZE = 1;

    // Static range quantization. Quantized tensor values' ranges are statically
    // determined. The activation and weight are quantized to INT8 while bias is
    // quantized to INT32.
    METHOD_STATIC_RANGE_INT8 = 2;

    // Dynamic range quantization. Quantized tensor values' ranges are
    // determined in the graph executions. The weights are quantized during
    // conversion.
    METHOD_DYNAMIC_RANGE_INT8 = 3;

};

void QuantizeCompositeFunctionsPass::runOnOperation() {
  MLIRContext& ctx = getContext();

  PassManager pm(&ctx);
  // Intermediate output from QuantizePass will have quantized ops
  // (XlaCallModuleOps) with quantized input and output types, which are not
  // allowed in the TF dialect.
  pm.enableVerifier(false);

  PrepareQuantizePassOptions options;
  options.enable_per_channel_quantized_weight_ =

    // and unquantized layers. The unquantized layer's input come from the
    // previous quantized layer (Please note that this part is different part
    // from DEBUGGER_TYPE_FLOAT_PER_LAYER). Each layer in the debugging model
    // has a DumpTensor, and it is used to save the entire value of outputs from
    // both the quantized and unquantized layer.
    DEBUGGER_TYPE_INT_PER_LAYER = 2;

//      quantized representation may be acceptable.
//
// Especially early in transformation, it is common to have pairs of
// qcast/dcast at points where a transition to a quantized type is
// required. In addition, it is also common to have an identity qcast
// (where the operand and result type are not quantized) at all points where
// it is legal to use a quantized representation (but is not known to be

// (e.g. matmul) has both quantized and unquantized inputs by dequantizing
// the quantized inputs, performing the operation in the expressed type, then
// requantizing if a quantized output is required.
//
// The motivation behind these changes is for Dialects that assume only float
// or quantized computation, and do not support a mixture of these types on

// CHECK: -------- Quantization Summary --------
// CHECK: Number of quantized layers in the model
// CHECK: --------------------------------
// CHECK: Name    Count/Total
// CHECK: ================================
// CHECK: Conv2D  1/2

// CHECK: Number of quantized layers with quantized outputs: 1/1
// CHECK: Number of quantize layers added: 1
// CHECK: Number of dequantize layers added: 1
}

// -----

  bool SetConstantResultParams(Operation* op);

  // Inserts the Quantize and Dequantize ops after `op`'s `index`-th result. The
  // quantized element type for the result is `quantized_type`.
  void QuantizeOpResult(Operation* op, int result_index,
                        QuantizedType quantized_type);

  // Inserts the Quantize and Dequantize ops after `arg`. The quantized element
  // type for `arg` is `quantized_type`.

    return result;
  } else {
    return std::nullopt;
  }
}

// Populates quantized ops from `module_op` to `results`. After going through
// the quantization passes, quantized ops are represented as `func::CallOp` with
// a callee's prefix of `quantized_`.
void PopulateQuantizedResults(ModuleOp module_op,
                              QuantizationResults& results) {

      llvm::SmallVector<Value, 4> quantized;
      for (auto user : returned.getUsers()) {
        if (auto q = Quantized(user)) {
          quantized.push_back(q);
        }
      }
      if (quantized.size() == 1) {
        ret.setOperand(i, quantized.front());
      }
      i++;
    }
  });

  // We prefer to placing quantization emulation ops on the results of the