[SYCL][Driver] Support bfloat16 devicelib selection when multiple AOT targets specified

[jinge90]

User can specify multiple AOT targets when building sycl program in followings ways:
1). via -fsycl-targets=intel_gpu_pvc,intel_gpu_acm_g10,....
2). via -fsycl-targets=spir64_gen ... -Xs "-device pvc,dg2...."
3). via -fsycl-targets=spir64_gen..., -Xsycl-target-backend=spir64_gen "-device pvc"
We should select native bfloat16 devicelib when all AOT targets specified support native bfloat16 conversion. Currently, pvc, dg2, bmg devices support native bfloat16.
If user specifies JIT target together with AOT targets which all support native bfloat16 conversion, we still select native bfloat16 devicelib since bfloat16 devicelib is skipped in linking step for JIT target.

[srividya-sundaram]

What about the case where we can specify Intel GPUs via --offload-arch option?

clang++ --offload-new-driver -fsycl --offload-arch=bdw

[jinge90]

Hi, @srividya-sundaram
Thanks for pointing out this, this PR didn't consider new offload driver, I will update it.
Thanks very much.

[srividya-sundaram]

Please add some description to the PR in addition to the PR title.

[bashbaug]

This is fine as a short-term fix, but longer-term we should be thinking of a different way to identify architectures with bfloat16 support that doesn't require maintaining a list of devices. For example, is there a way to query the properties of the target device(s) to determine if bfloat16 is supported?

Note that Lunar Lake GPUs also support bfloat16 (I think), and it's not in this list.

[jinge90]

Hi, @bashbaug
The bf16 support can be queried in execution time because we can only know the real target the program is running during execution time. In compilation time for AOT mode, compiler driver can only decide according to target platform specified and we have to maintain the bf16 support information in compiler driver source code, otherwise compiler driver won't know whether a target platform supports bf16. The platform we are building the program may be different from the platform the program is going to run on.
Thanks very much.

[bashbaug]

Could we call into ocloc to do the query? For example, if we're AOT compiling for DG2, we could do something like (this example uses the ocloc command-line interface, but the same is supported for the library interface):

$ ocloc query CL_DEVICE_EXTENSIONS_WITH_VERSION -device dg2
<snip> cl_intel_bfloat16_conversions:1.0.0 <snip>

Because the cl_intel_bfloat16_conversions extension is supported, we can know that DG2 supports SPIR-V bfloat16 conversion instructions.

[jinge90]

Hi, @bashbaug
I just tried the ocloc command but found after running "ocloc query CL_DEVICE_EXTENSIONS_WITH_VERSION -device dg2" but found a file named with "CL_DEVICE_EXTENSIONS_WITH_VERSION" is generated in local directory, it seems to be ocloc's behavior, is there any way to get rid of this?
Thanks very much.

[jinge90]

Hi, @mdtoguchi
I am a little confused about the intel_gpu_* targets string, we can use:
clang++ -fsycl -fsycl-targets=intel_gpu_acm_g10 xxx to specify aot compilation for DG2 platform but we also have:
clang++ -fsycl -fsycl-targets=intel_gpu_dg2_g10 xxx for DG2 target. Is there duplicate here? Or is there any difference between intel_gpu_acm_* and intel_gpu_dg2_*

Thanks very much.

[mdtoguchi]

Hi, @mdtoguchi I am a little confused about the intel_gpu_* targets string, we can use: clang++ -fsycl -fsycl-targets=intel_gpu_acm_g10 xxx to specify aot compilation for DG2 platform but we also have: clang++ -fsycl -fsycl-targets=intel_gpu_dg2_g10 xxx for DG2 target. Is there duplicate here? Or is there any difference between intel_gpu_acm_* and intel_gpu_dg2_*

Thanks very much.

For intel_gpu_acm* and intel_gpu_dg2*, these are equivalent and map to using -device acm* for the ocloc call.