Which device are the tasks running on?

[muziqaz]

Ask nVidia, not AMD
AMD is a hippy here

[Peter_Hucker]

https://massedcompute.com/faq-answers/? ... plications

[arisu]

nVidia does not support HIP. But HIP is CUDA code recompiled for AMD hardware I believe

It is a distinct runtime API from CUDA, but hipify can do a very efficient one-to-one conversion from CUDA to HIP because it supports all the same features. Unlike OpenCL where there are automatic transpilers but they are not smart enough to rewrite CUDA-isms in OpenCL efficiently.

From the AMD website:

"The Heterogeneous-computing Interface for Portability (HIP) is a C++ runtime API and kernel language that lets you create portable applications for AMD and NVIDIA GPUs from a single source code."

https://rocm.docs.amd.com/projects/HIP/ ... index.html

It works fine on Nvidia but FAH does not use it.

[Peter_Hucker]

"transpiler" and "CUDA-isms" - I must remember those terms I like.

[muziqaz]

I don't see the point of anyone using HIP over CUDA.
If you already have Nvidia GPU, your code is CUDA, unless you are some sort of masochistic enthusiast, or someone planning to ditch Nvidia, which never happens these days

[Peter_Hucker]

Because someone like Folding could write in HIP only and not have to bother with AMD/Nvidia considerations.

[muziqaz]

Because someone like Folding could write in HIP only and not have to bother with AMD/Nvidia considerations.

Yeah, no.
CUDA has been in fah waaay before HIP. Hell, HIP is still not out for AMD. So, no one will ditch perfectly working CUDA, which is constantly improved and looked after by well funded Nvidia software team, and go with HIP, which I had to beg OpenMM Devs to support, and which is developed and looked after hopes and dreams of upper management of AMD.
Let's be real here, please.

[Peter_Hucker]

When someone comes to write a new program, they will obviously want to write it only once. So they will obviously choose something which works on as many platforms as possible to get the most work units done. This has to be HIP. Sure, Nvidia will run OpenCL, but apparently not as efficiently.

[muziqaz]

When someone comes to write a new program, they will obviously want to write it only once. So they will obviously choose something which works on as many platforms as possible to get the most work units done. This has to be HIP. Sure, Nvidia will run OpenCL, but apparently not as efficiently.

Sigh, no one in their right mind in current market will choose HIP over CUDA for Nvidia.
They'd rather implement CUDA first, ignoring AMD completely or they gonna implement CUDA ignoring AMD and leaving it for later when they have some spare time.
CUDA has been in development for over a decade, it is feature rich and well supported by Nvidia who seem to have a plan in place for it. It is being developed by software guys who have been with CUDA for very very long time.
HIP has been developed for 3 years, it is still in development and is trying to catch up to CUDA and is still quite some distance away from it. It is being developed a small group of developers who have just been hired for it. AMD upper management only very recently realised that HIP needs extra attention.
Now looking at these two scenarios, do you still think anyone (besides diehard fanboys or mental patients) will pick HIP over CUDA for nVidia GPUs?

[Peter_Hucker]

If I was working for a project and I was writing a new program for crowd computing, I'd want to write it only once to save time and money, and get as many volunteers crunching as possible. There are lots of AMDs out there, why would I want to ignore them and get only half the work achieved? I want to use a program which will run on everything. So my choices are simply OpenCL or HIP. From what you've said, I gather HIP is more efficient than OpenCL.

[arisu]

If I was working for a project and I was writing a new program for crowd computing, I'd want to write it only once to save time and money, and get as many volunteers crunching as possible. There are lots of AMDs out there, why would I want to ignore them and get only half the work achieved? I want to use a program which will run on everything. So my choices are simply OpenCL or HIP. From what you've said, I gather HIP is more efficient than OpenCL.

Writing it in CUDA once and then using hipify to convert it into HIP would be better. Then if HIP goes tits up the only thing of value you lose is AMD devices. But if you write in HIP and HIP goes tits up, you lose everything.

[Peter_Hucker]

If the conversion is still just as efficient as writing it natively, then yes, good idea.

[arisu]

Yes, the hipify conversion is just as efficient as using the native CUDA. Unlike a direct conversion to OpenCL, where lots of performance is lost.

[bobdarnall]

I know this thread went a few directions, this is all I was trying to say, which sounded like the original ask: