TornadoMath.atan2 plugin has incorrect type

[andrii0lomakin]

Hi.
I noticed that TornadoMath.atan2 plugin is registered using double,double type parameters but in reality it uses float,float ones.
So my question is, does it work in the sense is this code executed on GPU?

[andrii0lomakin]

Ah, I found that plugins are registered both for double,double and float,float methods.
So, the second one is merely useless.

[andrii0lomakin]

If I got it correctly. And not second but first one of course.

[jjfumero]

Good catch. The atan function is registered for both float and double from the TornadoMath class. However we missed the double, double signature. We will provide those. Thanks!

[andrii0lomakin]

Do you need to have a double version ? I thought that GPUs do not widely support doubles. That will create false impression about the precision of the calculation for user. Do not you think so ?

[jjfumero]

Yes, most modern NVIDIA discrete GPUs support FP64, and we have some TornadoVM users doing HPC and data analytics that require FP64. For the precision, that's why we narrow to FP32 via the TornadoMath class instead of the Math class. But we are open to suggestions.

[andrii0lomakin]

I mean, in your response, you wrote that you will provide a double version. IMHO better to do not provide double,double version but remove plugin for double signature. In general, I suggest removing versions for methods that manipulate doubles or at least deprecate them because otherwise, it creates a false impression about the precision of the level of calculation.

[jjfumero]

If you call the TornadoMath with the double version, the TornadoVM JIT compiler generates fp64 operations. So, it is supported. As a developer, you can choose which one to invoke, the float or double. If you do not see the corresponding pairs double and float in the TornadoMath class, it is probably a miss from our side. TornadoVM never narrows down automatically, at least at the moment`.

[andrii0lomakin]

I see, got it. Thank you.