// RUN: odml-to-stablehlo-opt --compose-uniform-quantized-type \
// RUN:     --split-input-file --verify-diagnostics %s | FileCheck %s

module {
// CHECK-LABEL: quantized_conv_op
// CHECK-SAME: %[[ARG:.*]]: tensor<1x3x3x4xf32>
  func.func @quantized_conv_op(%arg0: tensor<1x3x3x4xf32>) -> tensor<1x3x3x4xf32> {
    %1 = stablehlo.constant dense<1.000000e+03> : tensor<1x1x1x1xf32>  // Input inverse scale.

          !IsConnectedWithCompsiteFunction(quantizing_op)) {
        continue;
      }

      // Same scale op is not supported for Uniform Quantized ops.
      if (target_opset_ == OpSet::UNIFORM_QUANTIZED) {
        continue;
      }

      // Collect all the quantized inputs and "clone" the matched op by these
      // inputs.
      SmallVector<Value, 4> inputs;
      inputs.reserve(quantizing_op->getNumOperands());

}

def QuantizeCompositeFunctionsPass : Pass<"stablehlo-quantize-composite-functions", "ModuleOp"> {
  let summary = "Quantize composite functions with QDQ input / outputs.";
  let options = [
    Option<"enable_per_channel_quantized_weight_",
        "enable-per-channel-quantized-weight",
        "bool", /*default=*/"true",
        "Whether to enable per-channel quantized weights.">,
    Option<"mlir_dump_file_name_", "mlir-dump-file-name",

//
//   1. Replaces quantized `TF::XlaCallModuleOp` with a `func::CallOp`.
//   2. Quantizes the callee function.
//
// The inputs of this pattern assumes an invalid IR, where even if a
// `TF::XlaCallModuleOp` is quantized the callee remains unquantized. Step (2)
// not only replaces the input and output tensor types into quantized ones, but
// also rewrites the body with a quantized equivalent.
//

};

// Converts quantized tensor type with signed integer type to quantized tensor
// type with unsigned integer type.
Type ConvertSignedQuantizedToUnsigned(Type signed_tensor_type, Location loc);

// Converts quantize ops with unsigned quantized types to these with signed
// quantized types and preserves the scales.
template <typename QuantizeOpT>

  bool verify_numeric = false;
  // Whether to add verification for layer by layer, or on whole model. When
  // disabled (per-layer) float and quantized ops will be run from same input
  // (output of previous quantized layer). When enabled, float and quantized ops
  // will run with respective float and quantized output of previous ops.
  bool whole_model_verify = false;

  // Whether to use fake quant attributes to calculate quantization parameters.

    // 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;
    // DEBUGGER_TYPE_FLOAT_PER_LAYER creates a debugging model with both

namespace mlir::quant::stablehlo {

// Checks whether an op is connected with a quantized composite function. If
// not, the same-scale op will not be quantized. This decision is based on the
// current assumption that the performance gain of the same-scale op itself
// could not beat the overhead of the quantize and dequantize routines need to
// be added around that op. When the assumption changes, this policy might

      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

  // The quantized call op without the _quantization_method attribute is not
  // captured as a `QuantizationResult`.
  ASSERT_THAT(results.results(), IsEmpty());
}

TEST_F(QuantizationReportTest, InitializeWithModuleOpWithInvalidCalleeName) {
  // A quantized dot_general op but the callee function has an invalid name. It
  // is expected to start with `quantized_`.