Reviewer 3

[tbetcke]

• [x] It needs to introduce the PDEs and range of applications at the beginning better, as well as more welcoming notation for the general reader. I was also hoping for more insight into the results (fraction of peak perf, fraction of time on host-to-device transfers, etc), and take-home messages about mistakes made and lessons learned.

[tbetcke]

It needs to introduce the PDEs and range of applications at the beginning better, as well as more welcoming notation for the general reader. I was also hoping for more insight into the results (fraction of peak perf, fraction of time on host-to-device transfers, etc), and take-home messages about mistakes made and lessons learned.

I understand this sentiment. There is a strict page limit for this article. Within this limit we decided to focus on the software design and less on mathematical and application aspects.

With regards to more detailed performance measurements, we are preparing a more dedicated paper on the OpenCL algorithms which will have more detail on this than we have scope in this paper.

With regards to take-home message about lessons learned we again would love to talk much more than is possible to fit on a one page summary at the end. Should the referee wish I'd be happy to discuss this topic much more in person than is possible through the refereeing format. There are a number of aspects of our journey from C++ to Python that I could not discuss here, and the journey is continuing. While the development model that we use in the paper works very well for Bempp-cl, we encountered limits of the Python focused model for other types of algorithms that are worth discussing in their own rights (and we are currently planning a corresponding publication).