Add SYCL build script

[ShengYang1]

This is a sub PR of #9042, it adds a script to build xla with --config=sycl.
Here is how to use:

workspace=$1
xla_path=$workspace/xla
cd $workspace
git clone -b yang/ci https://github.com/Intel-tensorflow/xla xla
bash $xla_path/build_tools/sycl/build.sh $workspace

[Zantares]

Hi @penpornk , any OpenXLA developer can use this script to check SYCL backend build if needed. It will automatically download the build environment including the necessary toolchain/lib, and clear it after build is done.

[kamaljeeti]

Hi @penpornk , can you please look into this? Thanks.

[ddunl]

I'm not totally convinced that this should be separate from the build_tools/sycl dir. Do you have a strong preference?

[ShengYang1]

Actually I'm not sure where should I put all these scripts. I'll move this file to build_tools/sycl if this is fine.

[penpornk]

@ddunl Does the new file path look good to you?

[abhaskumarsinha]

Unrelated,

I'm a bit confused about how XLA Backend works, glad you all are working with the SYCL backend. Can you please point out where sycl_compiler.cc exists? Like for NVIDIA this file nvptx_compiler.cc and for ROCM (AMD) this file amdgpu_compiler.cc exists.

If I see the compiler file writing SYCL backend of XLA, I'll at least gather some idea about the modules and way or writing it.

[Zantares]

Unrelated,

I'm a bit confused about how XLA Backend works, glad you all are working with the SYCL backend. Can you please point out where sycl_compiler.cc exists? Like for NVIDIA this file nvptx_compiler.cc and for ROCM (AMD) this file amdgpu_compiler.cc exists.

If I see the compiler file writing SYCL backend of XLA, I'll at least gather some idea about the modules and way or writing it.

Hi @abhaskumarsinha , actually we are using SPIRV target + SYCL runtime. I suppose you are asking this spirv_compiler.cc. It's a part of our upstream work and under working. You can check the whole work by this preview PR: #9042

[abhaskumarsinha]

Hi @abhaskumarsinha , actually we are using SPIRV target + SYCL runtime. I suppose you are asking this spirv_compiler.cc. It's a part of our upstream work and under working. You can check the whole work by this preview PR: #9042

Thank you @Zantares . I'm a bit noob here, do ASICs/FPGA support OpenCL/SYCL, can they support SPRIV?
Thank you for your PR. I'm glad to know there is some work going on here. It'll take some time for me to get some idea of how backends work in XLA. Thank you.

[Zantares]

do ASICs/FPGA support OpenCL/SYCL, can they support SPRIV?

It's up to your software stack. It usually needs 2 parts to support new hardware here:

• Code-gen, it's SPIRV for us here. You can see LLVM/MLIR... different code-gen in OpenXLA.
• Runtime, it's SYCL for us here. Same as above. you can see CPU/CUDA/ROCm... to manage different hardware resources.

So, you can choose the most suitable solution by combining these options.

[abhaskumarsinha]

@Zantares Okay, got it.

Can I see code-gen docs/library files in OpenXLA to understand what is happening here?

Can I use LLVM (code-gen) + SYCL (runtime) after PR #9042 ? I believe this should be sufficient for most of our hardware jobs.

[Zantares]

Can I see code-gen docs/library files in OpenXLA to understand what is happening here?

Hi @abhaskumarsinha , you can refer to the OpenXLA Arch Overview doc, especially the Compiler Backend section to understand it better. In addition, you can consult the community maintainer @penpornk or others for specific questions, since I'm also a hardware vendor developer and not clear about the roadmap or future tech design.

Can I use LLVM (code-gen) + SYCL (runtime) after PR #9042 ? I believe this should be sufficient for most of our hardware jobs.

You can reuse most works of #9042 once it's merged, but you still need to reorganize the bazel build dependency since you plan to use LLVM + SYCL rather than the current SPIRV + SYCL. Another point is that we used level_zero as the underlayer of SYCL runtime, you may need to change it to your own HW Driver.