Troels Henriksen

There is no operator overloading in Futhark. In the long term, we have considered adding type classes, but we haven't considered a specific design yet. I think we are waiting...

In order to not complicate the code generator for this relatively niche use case, the best way to implement this would be to have a separate facility (which can be...

And it should probably be the same formatting language. The `#[trace]` attribute is a little trickier, because we need to generate C code for it.

Don't call entry points directly; use a wrapper function (it can be a trivial one, e.g. `def foo = bar`).

This is another symptom of the arguable misdesign that our entry points are too implicit. We should probably have made them a more distinguished thing requiring explicit descriptions of the...

I'm willing to help out with implementing the GradBench benchmarks in horde-ad. There is no real reason to use ADBench, as GradBench contains everything from ADBench (and with better documentation)....

I suggest taking the protocol implementation from [the existing haskell tool](https://github.com/gradbench/gradbench/blob/main/tools/haskell/src/Main.hs). There's not much to it - just a controlling loop and some Aeson instances. In fact, a horde-ad tool...

Validation is done by the eval, so the responses you send will be validated automatically unless you explicitly disable it. This does not depend on whether you use Docker. Regarding...

You're right, they're the same. I probably implemented that a bit too literally.

I do a lot of OpenCL debugging and oclgrind is completely indispensable. The data race detector (although memory hungry) is also a lifesaver. It may be that oclgrind does not...