CUDA without nVidia

[Michał Wichrowski]

Dear all, 

Have anyone managed to compile and run CUDA code on non-nVidia device?
I've found HIPify: https://github.com/ROCm-Developer-Tools/HIPIFY 

that translates CUDA code to HIP code that can be then compiled and ran on both AMD and nVidia GPUs. 

The thing is, I'm looking for a new laptop. Since I'm planning switching to GPU I want to be able to compile, run and test my code locally. Choosing a hardware with nVidia chip narrows the choice, moreover AMD CPUs  seem to come with good enough built-in GPUs.

 Some years ago I've tried GPUOcelot and did not managed to make it work.

Best,

Michał

[Bruno Turcksin]

Michal,

Support for AMD through HIP is very high on my todo list but the compiler is not that great right now, i.e. the compiler can (and will) generate buggy code. I've tried to hipify the code a few months ago but it didn't go too well. My plan is to work on support for AMD after we branch the release so that we don't have to support buggy compilers. I am pretty confident that we will have support for HIP by the end of the Summer maybe earlier.

Best,

Bruno

[Michał Wichrowski]

Michal,

Support for AMD through HIP is very high on my todo list but the compiler is not that great right now, i.e. the compiler can (and will) generate buggy code.

Ok, so even if Deal.II will be ported I can assume so for other libraries.  Good to know that HIP does  require porting.

I've tried to hipify the code a few months ago but it didn't go too well. My plan is to work on support for AMD after we branch the release so that we don't have to support buggy compilers. I am pretty confident that we will have support for HIP by the end of the Summer maybe earlier.

Great news! 

 

Best,

Bruno

Thanks!
Michał