if (add.getType() != op.getType()) {
      add = rewriter.create<tensor::CastOp>(loc, op.getType(), add);
    }
    rewriter.replaceOp(op, {add});
    return success();
  }
};

// Conterts tf.Conv2D to mhlo.dynamic_conv.
// TODO(disc): To recover static special case's performance with adding folding,
// canonicalization func and removing ConvertConvOp.
template <typename OpT, int num_spatial_dims, bool depthwise_conv = false>

    return emitOptionalError(location, "invalid padding format provided");
  }

  // Output always have rank 4. All dimensions are initialized to
  // dynamic size and can be partially inferred.
  // TFL's conv2d is always NHWC format & the filter is OHWI.
  SmallVector<int64_t, 4> return_shape(4, ShapedType::kDynamic);
  return_shape[0] = input_ty.getDimSize(0);
  return_shape[3] = filter_ty.getDimSize(0);

w.r.t. the input of the convolution.}]>:$output
  );

  TF_DerivedOperandTypeAttr T = TF_DerivedOperandTypeAttr<0>;
}

def TF_Conv3DOp : TF_Op<"Conv3D", [InferTensorType, Pure]> {
  let summary = [{
Computes a 3-D convolution given 5-D `input` and `filter` tensors.
  }];

  let description = [{