auto perm_values = perm_values_attr.getValues<APInt>();
  size_t idx = 0;
  for (; idx < perm_values_attr.size() - 2; ++idx) {
    if (perm_values[idx].getSExtValue() != idx) return false;
  }

  return (perm_values[idx].getSExtValue() == perm_values_attr.size() - 1) &&
         (perm_values[idx + 1].getSExtValue() == idx);
}

  // Transpose and constant-fold the weight if needed.
  if (adj_y.getValue()) {
    SmallVector<int32_t> perm_values(num_batch_dim);
    absl::c_iota(perm_values, 0);
    perm_values.push_back(num_batch_dim + 1);
    perm_values.push_back(num_batch_dim);
    Value perm = Create1DConstValue<int32_t>(builder, loc, perm_values);
    auto transpose_op = builder.create<TF::TransposeOp>(loc, weight, perm);

      return failure();
    }

    auto perm1_value = llvm::dyn_cast<DenseElementsAttr>(perm1.getValue())
                           .getValues<int32_t>();
    auto perm2_value = llvm::dyn_cast<DenseElementsAttr>(perm2.getValue())
                           .getValues<int32_t>();

    llvm::SmallVector<int32_t> perm1_arr(perm1_value.begin(),
                                         perm1_value.end());

  SmallVector<int64_t, 8> perm0_values;
  for (const auto &value : perm0.getValues<APInt>())
    perm0_values.push_back(value.getSExtValue());

  SmallVector<int64_t, 8> perm1_values;
  for (const auto &value : perm1.getValues<APInt>())
    perm1_values.push_back(value.getSExtValue());

  for (int i = 0; i < perm0_values.size(); ++i) {
    if (perm0_values[perm1_values[i]] != i) return false;
  }

  return true;

    $asymmetric_quantize_inputs_bmm
  ),
  [(AreLastTwoDimsTransposed $perm_value), (IsNoneType $bias)]>;

// Fuses TFL_FullyConnectedOp and TFL_TransposeOp Rhs to TFL_BatchMatMulOp
def FuseTransposeFCRhsToBatchMatmul : Pat<
  (TFL_FullyConnectedOp
    2DTensorOf<[F32]>:$lhs,
    (TFL_TransposeOp TensorOf<[F32]>:$rhs, (Arith_ConstantOp:$perm_value $p0)),
    $bias, $TFL_AF_None, $TFL_FCWO_Default,

def FuseTransposeIntoBatchMatMulLHS: Pat<
  (TFL_BatchMatMulOp
    (TFL_TransposeOp:$transposed_value $input, (Arith_ConstantOp:$perm_value $p0)),
    $rhs, $adj_x, $adj_y, $asymmetric_quantize_inputs),
  (TFL_BatchMatMulOp $input, $rhs, ConstBoolAttrTrue, $adj_y, $asymmetric_quantize_inputs),
  [(AreLastTwoDimsTransposed $perm_value),
   (IsBoolAttrEqual<"false"> $adj_x)]>;

// Fuses TFL_TransposeOp after TFL_BatchMatMulOp into the BMM.