scalar_mul(dilation_rate_value, scalar_sub(filter_size, one)), one);

  // output_size = (input_size + stride - 1) / stride
  Value output_size = scalar_div(
      scalar_add(input_size, scalar_sub(stride_value, one)), stride_value);
  // padding_needed = std::max(
  //     0,
  //     (output_size - 1) * stride + effective_filter_size - input_size)
  Value padding_needed = scalar_sub(
      scalar_add(effective_filter_size_op,

func.func @transposeConv2DBackpropInput_f32(
  %input_size:   tensor<4xi32>,
  %filter:       tensor<1x28x28x64xf32>,
  %out_backprop: tensor<1x28x28x64xf32>
) -> tensor<1x28x28x64xf32> {

  // CHECK: "tf.Conv2DBackpropInput"
  // CHECK-SAME: data_format = "NCHW"
  %0 = "tf.Conv2DBackpropInput"(%input_size, %filter, %out_backprop)
       {
         data_format = "NHWC",

func.func @transposeConv2DBackpropInput_f16(
  %input_size:   tensor<4xi32>,
  %filter:       tensor<1x28x28x64xf16>,
  %out_backprop: tensor<1x28x28x64xf16>
) -> tensor<1x28x28x64xf16> {

  // CHECK: "tf.Conv2DBackpropInput"
  // CHECK-SAME: data_format = "NCHW"
  %0 = "tf.Conv2DBackpropInput"(%input_size, %filter, %out_backprop)
       {
         data_format = "NHWC",
         padding = "VALID",

      }
    }
  }
  return found_dtype && found_value;
}

bool IsAddN(const tensorflow::NodeDef& node_def, int n) {
  if (node_def.op() != "AddN" || node_def.name() != "add" ||
      node_def.input_size() != n) {
    return false;
  }
  bool found_t = false;
  bool found_n = false;
  for (const auto& attr : node_def.attr()) {
    if (attr.first == "T") {

        int64_t output_size;
        int64_t pad_low_int64;
        int64_t pad_high_int64;
        int64_t input_size = input_ty.getDimSize(dim);
        if (input_size == ShapedType::kDynamic) return failure();
        absl::Status status = tensorflow::GetWindowedOutputSizeVerbose(
            input_size, filter_ty.getDimSize(i), dilation, stride, padding,
            &output_size, &pad_low_int64, &pad_high_int64);

    Location loc = op.getLoc();
    Value scalar_zero = CreateI32SplatConst(loc, &rewriter, {}, 0);

    // Compute the input tensorlist's length and store it in `input_size`.
    IntegerType shape_dtype = rewriter.getIntegerType(32);
    auto input_size = rewriter.create<TF::TensorListLengthOp>(
        loc, tensorflow::GetTypeFromTFTensorShape({}, shape_dtype),
        op.getOperand(0));

StatusOr<string> GetVarHandleName(
    const std::unordered_map<string, NodeDef*>& name_to_node_map,
    string node_name) {
  const NodeDef* node = name_to_node_map.at(node_name);
  while (node->input_size() > 0) {
    auto parent = name_to_node_map.find(node->input(0));
    if (parent == name_to_node_map.end()) break;
    node = parent->second;
    if (node->op() != "Identity") {

      for (int64_t i = 0; i < input_ty.getRank(); ++i) {
        int64_t slice_size = slice_sizes.getValues<APInt>()[i].getSExtValue();
        int64_t input_size = input_shape[i];
        if (slice_size != -1 && input_size != ShapedType::kDynamic &&
            slice_size > input_size) {
          return op.emitOpError() << "requires size[i] <= Di, even if begin[i] "
                                     "is unknown at compile time";

                        std::vector<const Edge*> control_deps, Output* size) {
  // If slice_size[i] >= 0 then slice_size[i] = slice_size[i].
  //
  // If slice_size[i] == -1 then slice_size[i] = input_size[i] -
  // begin[i].
  //
  // If slice_size[i] < -1 then executing the slice will throw an error, and we
  // don't do anything here.  We've already filtered these cases out in
  // IsRewritableSlice.

                           b.device());
    }
    return false;
  }
  if (a.input_size() != b.input_size()) {
    if (diff) {
      *diff = absl::StrCat(diff_preamble, " mismatch for node ", a.name(),
                           ", expected ", a.input_size(), " inputs got ",
                           b.input_size(), " expected:\n", a.DebugString(),
                           "\ngot:\n", b.DebugString());
    }