Support CMAKE_CUDA_ARCHITECTURES

[yiranlus]

Hi, I wonder if DBCSR plans to support compile using CMAKE_CUDA_ARCHITECTURES.

As far as I know, currently DBCSR can only target single CUDA architecture at a time. It would be nice to support CMAKE_CUDA_ARCHITECTURES so multiple CUDA architectures can be considered during compilation.

[alazzaro]

It is not that easy with the current DBCSR implementation design.
The problem is that the GPU arch is used to select the kernel parameters during the compilation phase, which are then hard-coded (the code is generated with a python script). Then, the CUDA code is jitted. Having said that, this is not a problem of which flag to use during compilation in CUDA (for jitting we don't need that), but a priori choice of which kernels to use.

Can I ask what you are trying to achieve? If the problem is to use the same binary on multiple arch, then this is already possible. You just need to specify the lowest arch in the range to run on both (as I said, it will select the optimized kernels for a single GPU, but they are good enough for the newest too). For instance, I can ask A100 and support A100 and H100.

[yiranlus]

Thanks for answering.

I asked this question because we have both A100 and H100 GPUs and I wanted to get the best performance from them. Knowing that it is difficult to target more than one arch, I will just keep targeting A100.

[alazzaro]

Well, the performance is quite equivalent for the single kernel, so using A100 on H100 is OK.
A dynamic dispatch of the kernel for multiple architectures is possible in the long run, but I'm not fully convinced that it is worth the effort...