ggml-sycl : use malloc_shared for UMA/integrated GPU devices

[vmartirosyan]

Overview

I have been executing the llama.cpp on my Intel Arc integrated platform and found out that it is not using shared memory advantage. I had already done something like this for my auto-parallelizer compiler and it really helps to boost the performance. The change helped me to achieve a speedup like: [ Prompt: 141.0 t/s | Generation: 10.7 t/s ].

Additional information

Tested on Intel Arc 140T iGPU. Should not affect any discrete GPU, as the newly added code is all in if blocks.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES. I designed the solution based on my prior OpenCL work; AI me assisted with SYCL API translation.

[arthw]

I want to test it.

Could you share the test result to show the benefit?
And the info of OS, oneAPI version, Intel CPU/GPU model, LLM gguf name.

Thank you for your contribution!