[SYCL] [Driver] Add offload-arch support for SYCL offloading

[srividya-sundaram]

Implement --offload-arch option to enable SYCL offloading to Intel CPUs, Intel GPUs, NVidia and AMD GPUs.

--offload-arch implementation is currently supported in the new driver model ( i.e. with --offload-new-driver option)

Example usage:

clang++   --offload-new-driver  -fsycl  --offload-arch=bdw         // Offload SYCL code to Intel GPU
clang++   --offload-new-driver  -fsycl  --offload-arch=broadwell   // Offload SYCL code to Intel CPU
clang++   --offload-new-driver  -fsycl  --offload-arch=sm_80       // Offload SYCL code to NVidia GPU
clang++   --offload-new-driver  -fsycl  --offload-arch=gfx700      // Offload SYCL code to AMD GPU

[GeorgeWeb]

Overall, despite not being a pinged reviewer, I wanted to say I like the direction of these changes. I am aware this is still a Draft but is there a plan to address the --offload-arch with no -fsycl-targets case in

llvm/clang/lib/Driver/Driver.cpp

Line 6397 in 35b926a

} else if (HasValidSYCLRuntime) {

either as part of this PR or a follow-up? I am asking this because the effect of these changes will also require extra handling in OffloadingActionBuilder::SYCLActionBuilder::initialize(). Thanks.

[GeorgeWeb]

Hello, AFAIK, OffloadingActionBuilder::SYCLActionBuilder::initialize() is the path taken when we do not pass --offload-new-driver option. Currently, --offloard-arch option is supported only in the new driver model ( i.e with --offload-new-driver option passed) Hence the --offload-arch implementation is not handled in OffloadingActionBuilder::SYCLActionBuilder::initialize()

I see, I must've missed the chain of changes wrt the new offload driver for SYCL.
I only pointed the above, likely due to confusion, because for me clang++ -fsycl --offload-new-driver --offload-arch=sm_75 test.cpp still hits OffloadingActionBuilder::SYCLActionBuilder::initialize() but does nothing really as it returns early as no SYCL Toolchain has been found as of yet. Since the implementation for --offload-arch with the new offload driver logic is handled elsewhere, I guess that's okay then. Thanks for explaining :)!

[GeorgeWeb]

LGTM (mainly commenting about the handling for Nvidia and AMD targets). Thanks again for your patience, explanations, and addressing review comments.

[srividya-sundaram]

Ping @intel/dpcpp-cfe-reviewers

[srividya-sundaram]

@intel/llvm-gatekeepers Hello, this PR is ready to be merged. Thanks!