Value &padding, int num_dims) {
  ShapedType input_shape = mlir::cast<ShapedType>(input.getType());
  SmallVector<int64_t> spatial_dims(num_dims - 2);
  absl::c_iota(spatial_dims, 1);
  bool has_dynamic_spatial_dim = absl::c_any_of(
      spatial_dims,
      [&input_shape](int64_t dim) { return input_shape.isDynamicDim(dim); });
  if (conv_padding.strref() == "SAME" && has_dynamic_spatial_dim) {

      // Prepare metadata indexed by spatial_dim for computing pad_before
      // and pad_after.
      int64_t input_size = input_shape[spatial_dim];
      if (input_size == ShapedType::kDynamic) return failure();
      int64_t output_size = out_backprop_ty.getDimSize(spatial_dim);
      if (output_size == ShapedType::kDynamic) return failure();
      int64_t filter_size = filter_ty.getDimSize(i);

      return false;

    if (mlir::cast<ShapedType>(conv_op.getWindowStrides().value().getType())
            .getRank() != 1)
      return false;

    auto spatial_dims =
        conv_op.getDimensionNumbers().getInputSpatialDimensions().size();
    if (spatial_dims != num_spatial_dims) return false;

    return true;
  }

 private:
  bool IsSamePadding(mhlo::ConvolutionOp conv_op, ArrayRef<int64_t> strides,