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
// change as well.

// CHECK:   %{{.*}} = "tfl.dequantize"(%{{.*}}) : (tensor<1x401408x!quant.uniform<u8:f32, 3.906250e-03>>) -> tensor<1x401408xf32>

  %cst = arith.constant dense<[1, 401408]> : tensor<2xi32>

// 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>

          return;
        }
      } else {
        result = quantized_op->getResult(i);
      }
      for (auto user : result.getUsers()) {
        // Skip the Requantize op and set the user to the following dequantize
        // op. This happens when the quantizer tries to match the scale conflict
        // with QuantizeOp - QuantizeOp(requant) - DequantizeOp triples. The

    // Finally, use the quantization parameter to create the quantize and
    // dequantize ops, and insert them between the tf.FakeQuantWithMinMaxVarsOp
    // and its users.
    auto quantize = rewriter.create<quantfork::QuantizeCastOp>(
        tf_op.getLoc(), qtype.getValue(), input);
    auto dequantize = rewriter.create<quantfork::DequantizeCastOp>(
        tf_op.getLoc(), res_type, quantize.getResult());

  // TODO: b/323478683 - Make the attribute being part of op definition.
  quantize->setAttr(kVolatileOpAttrName, builder_.getUnitAttr());

  // `original_result` has a use to `quantize`, so this will replace that use
  // by the result of `dequantize`. Remember to reset that use afterwards
  value.replaceAllUsesWith(dequantize);
  quantize.getOperation()->replaceUsesOfWith(dequantize, value);
}

      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++;
    }
  });

  // Check for  (Quant (Dequant $in), $qA) "qdq" pairs that couldn't be

    // 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

    assert(scales_.size() == zero_points_.size());
  }

  // Quantize an Attribute by the quantization parameters. Return nullptr if
  // the conversion fails or the input array isn't an ElementsAttr.
  ElementsAttr convert(Attribute real_value);

 private:
  // Quantize an DenseFPElementsAttr by the quantization parameters.
  DenseElementsAttr convert(DenseFPElementsAttr attr);

// The prepare-dynamic-range-quantize Pass.
//
namespace mlir {
namespace TFL {

namespace {
#define GEN_PASS_DEF_PREPAREDYNAMICRANGEQUANTIZEPASS
#include "tensorflow/compiler/mlir/lite/transforms/passes.h.inc"

// A boolean attribute used to describe whether input activations need to be
// asymmetrically quantized.
constexpr char kAsymmetricQuantizeInputsAttr[] = "asymmetric_quantize_inputs";