sycl : Add support for non-release DPC++ & oneMKL

[joeatodd]

This PR enables llama.cpp to be built with the non-release (i.e. built from source) versions of DPC++ (clang++ instead of icpx) and oneMKL. Some duplicate/unneeded flags have been removed, and I have rearranged the logic slightly to keep things where they are needed.

• I have read the contributing guidelines
• Self-reported review complexity:
  • Low
  • Medium
  • High

[joeatodd]

@airMeng @NeoZhangJianyu hey folks - I tested that I can still compile on PVC with the release compiler (2024.2), but perhaps you'd like to double check I haven't broken anything for your builds?

[airMeng]

LGTM @luoyu-intel our architecture for awareness

[mudler]

This seems to have broken things here:

[100%] Linking CXX executable ../../bin/grpc-server
/usr/bin/ld: ../../ggml/src/libggml.a(ggml-sycl.cpp.o): undefined reference to symbol '_ZN4sycl3_V15queue6memcpyEPvPKvmRKNS0_6detail13code_locationE'
/usr/bin/ld: /opt/intel/oneapi/compiler/2024.2/lib/libsycl.so.7: error adding symbols: DSO missing from command line
icpx: error: linker command failed with exit code 1 (use -v to see invocation)

Is there anything I'm missing maybe?

[mudler]

Similarly, this seems to break the rpc server build when sycl is enabled:

#45 1217.2 gmake[6]: Entering directory '/build/backend/cpp/llama-grpc/llama.cpp/build'                                                                                                                                                       
#45 1217.2 [ 92%] Building CXX object examples/rpc/CMakeFiles/rpc-server.dir/rpc-server.cpp.o                                                                                                                                                 
#45 1217.5 [100%] Linking CXX executable ../../bin/rpc-server                                                                                                                                                                                 
#45 1217.7 /usr/bin/ld: ../../ggml/src/libggml.a(ggml-sycl.cpp.o): undefined reference to symbol '_ZN4sycl3_V15queue6memcpyEPvPKvmRKNS0_6detail13code_locationE'                                                                              
#45 1217.7 /usr/bin/ld: /opt/intel/oneapi/compiler/2024.2/lib/libsycl.so.7: error adding symbols: DSO missing from command line                                                                                                               
#45 1217.7 icpx: error: linker command failed with exit code 1 (use -v to see invocation)    

[joeatodd]

Hi @mudler, sorry to hear I've messed things up for you. I can reproduce this locally with a build of llama-bench, if I pass -DBUILD_SHARED_LIBS=OFF, which I think is how you're building.

It's missing libsycl, I think because I need to pass -fsycl as a linker flag, as well as a compilation flag.

I'll try a fix locally & push something ASAP.

[mudler]

Hi @mudler, sorry to hear I've messed things up for you. I can reproduce this locally with a build of llama-bench, if I pass -DBUILD_SHARED_LIBS=OFF, which I think is how you're building.

No worries, was trying to trace it back to provide more contextual info and created #8665!

It's missing libsycl, I think because I need to pass -fsycl as a linker flag, as well as a compilation flag.

I'll try a fix locally & push something ASAP.

Thank you! 👍 I'm testing now a very simple workaround here to confirm that's what all is missing:

add_executable(rpc-server rpc-server.cpp)
if ( DEFINED ENV{ONEAPI_ROOT})
target_link_libraries(rpc-server PRIVATE ggml llama sycl)
else()
target_link_libraries(rpc-server PRIVATE ggml llama)
endif()

I'll let keep you posted! I can also give a shot to the PR once is up