Adding Support for Intel GPUs with sycl

[AlexFierro9]

I would like to drop an enhancement proposal for Intel's SYCL standard that can allow better performance for Intel GPU's(Intel Arc and the DataCenter Series), it's the Intel Equivalent to CUDA and ROCm.

If providing manpower for this proposal is not possible, I request to be assigned this issue and I'll be more than happy to work on this

[shibe2]

I'm working on OpenCL back-end. Would be interested to collaborate on SYCL as well, if time permits.

[AlexFierro9]

Actually Intel has created a tool called Syclomatic, it can port pretty much all of the CUDA backend all by itself, all that would be left is integrating the new files into current methods, and any other optimizations we might want to do.

Here take a look: SYCLomatic

[itlackey]

I have an A770 and would love to see this PR merged. I am will to help test and provide feedback if it would help push this forward.

Currently many tools use llama.cpp under the hood. Those tools currently require either Nvidia or Mac GPUs. Adding proper support for Intel GPUs would make these tools more accessible and ready for the future generation of Intel hardware.

[shibe2]

I have an A770

What are your current speeds (prompt and eval) with CPU build and CLBlast build? How does ngl parameter affect performance?

[koech-v]

Patiently waiting for this. Any progress?

[AlexFierro9]

As it turns out there is an already open issue for this