[SYCL] Don't add include/sycl/ to the system includes path

[aelovikov-intel]

Fixes #6770

Open question: I'm not sure if https://github.com/intel/llvm/tree/sycl/sycl/include/CL/__spirv should go into include/CL/__spirv or somewhere else. On the other hand, https://github.com/intel/llvm/blob/sycl/sycl/include/CL/sycl.hpp has to be installed into include/CL/sycl.hpp per the specification.

[sarnex]

esimd changes lgtm, will let others review the overall change

[mdtoguchi]

Driver changes look fine - seeing the number of header inclusion changes we have had to make internally - how do we minimize impact on users?

[bader]

Fixes #6770

Thanks for fixing this long-standing issue!

On the other hand, https://github.com/intel/llvm/blob/sycl/sycl/include/CL/sycl.hpp has to be installed into include/CL/sycl.hpp per the specification.

The spec says that

For compatibility with SYCL 1.2.1, SYCL provides another standard header file: <CL/sycl.hpp>, which can be included in place of <sycl/sycl.hpp>. In that case, all SYCL classes, constants, types and functions defined by this specification should exist within the ::cl::sycl C++ namespace.

It doesn't require implementation to put it into include directory. Moreover, it doesn't mandate the driver to add additional search paths for SYCL headers. It's DPC++ implementation choice.

I'm not sure if https://github.com/intel/llvm/tree/sycl/sycl/include/CL/__spirv should go into include/CL/__spirv or somewhere else.

What is your concern about the __spirv directory location?

[aelovikov-intel]

It doesn't require implementation to put it into include directory. Moreover, it doesn't mandate the driver to add additional search paths for SYCL headers. It's DPC++ implementation choice.

Yeah, I wasn't precise in what I said. What I meant is that #include <CL/sycl.hpp> must work and now that we don't put include/sycl in search path we have to move that header.

[aelovikov-intel]

What is your concern about the __spirv directory location?

It feels awkward to me that our implementation detail is located in OpenCL-specific folder. It might better be suited for sycl/__spirv.

[aelovikov-intel]

Driver changes look fine - seeing the number of header inclusion changes we have had to make internally - how do we minimize impact on users?

My personal opinion: we don't - it was a bug and now it's fixed.

[bader]

What is your concern about the __spirv directory location?

It feels awkward to me that our implementation detail is located in OpenCL-specific folder. It might better be suited for sycl/__spirv.

Agree. SYCL-1.2.1 implementation put all headers into CL directory. When we added support for SYCL-2020, we moved all headers from CL to sycl directory. It sounds like it's a leftover of this work. I expect CL to have only one file - sycl.hpp, which includes sycl/sycl.hpp and imports all SYCL declarations to cl namespace.

[aelovikov-intel]

What is your concern about the __spirv directory location?

It feels awkward to me that our implementation detail is located in OpenCL-specific folder. It might better be suited for sycl/__spirv.

Agree. SYCL-1.2.1 implementation put all headers into CL directory. When we added support for SYCL-2020, we moved all headers from CL to sycl directory. It sounds like it's a leftover of this work. I expect CL to have only one file - sycl.hpp, which includes sycl/sycl.hpp and imports all SYCL declarations to cl namespace.

Good! I'll address it separately after this PR is merged in.

[uditagarwal97]

SYCL RT changes LGTM

[AlexeySachkov]

This change makes me think that ur_api.h is our own header, but that's not true. I wonder if we need to put it into some _deps subfolder for clarity.