OpenCL assumes OpenCL C 2.0, ruling out many ICDs

[inducer]

In a similar setting to #787, but with the OpenCL thread model, I see a message

Kernel [gatherScatter]: Build Log
Build option -cl-std specified OpenCL C version 2.0,but device cpu-tbb doesn't support that OpenCL C version.
terminate called after throwing an instance of 'occa::exception'
  what():
---[ Error ]--------------------------------------------------------------------
    File     : /home/andreas/pack/libparanumal/occa/src/occa/internal/modes/opencl/utils.cpp
    Line     : 345
    Function : buildProgram
    Message  : Kernel [gatherScatter]: Building Program
               OpenCL Error [ -11 ]: CL_BUILD_PROGRAM_FAILURE
    Stack
      18 occa::error(std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, int, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&)
      17 occa::opencl::error(int, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, int, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&)
      16 occa::opencl::buildProgram(occa::opencl::info_t&, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&)
      15 occa::opencl::buildProgramFromSource(occa::opencl::info_t&, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, occa::json const&)
      14 occa::opencl::device::buildKernelFromProcessedSource(occa::hash_t, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, bool, occa::lang::sourceMetadata_t&, occa::lang::sourceMetadata_t&, occa::json const&)
      13 /home/andreas/pack/libparanumal/occa/lib/libocca.so(+0x2bb4dd)
      12 occa::io::stageFiles(std::vector<std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, std::allocator<std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > > > const&, bool, std::function<bool (std::vector<std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, std::allocator<std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > > > const&)>)
      11 occa::io::stageFile(std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, bool, std::function<bool (std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&)>)
      10 occa::launchedModeDevice_t::buildKernel(std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, occa::hash_t, bool, occa::json const&)
       9 occa::launchedModeDevice_t::buildKernel(std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, occa::hash_t, occa::json const&)
       8 occa::device::buildKernel(std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, occa::json const&) const
       7 ./cnsMain(+0x290bab)
       6 ./cnsMain(+0x228a49)
       5 ./cnsMain(+0x56b22)
       4 ./cnsMain(+0x392f1)
       3 /lib/x86_64-linux-gnu/libc.so.6(+0x29ca8)
       2 /lib/x86_64-linux-gnu/libc.so.6(__libc_start_main+0x85)
       1 ./cnsMain(+0x39e01)
================================================================================

Is there a specific reason why OpenCL C 2.0 is required? It does not seem widely implemented, and this rules out, for example, execution on https://github.com/pocl/pocl.

[noelchalmers]

Im not sure OCCA itself requires OpenCL v2.0. I know libParanumal enforces that here: https://github.com/paranumal/libparanumal/blob/main/libs/core/platformProperties.cpp#L96 but not sure thats a strict requirement for OCCA. Might be worth trying a simple test program

[kris-rowe]

IIRC there are some OpenCL features OCCA uses that are in the OpenCL 2.0 spec, but not previous specs, and became optional or changed somewhat in OpenCL 3.0.

The OpenCL backend hasn't received much love in a while. We should likely revisit this and add some runtime logic or preprocessor guards to handle the differences between specs.

[inducer]

Thanks!

IIRC there are some OpenCL features OCCA uses that are in the OpenCL 2.0 spec

IIRC, you are talking about the CL runtime. That's different from the CL C kernel language, which has its own version numbers. Searching for "2.0", in the standard, I wonder what features paranumal might rely on. Nothing really sticks out.

[noelchalmers]

Ah I think I remember. I think it has to do with passing pointers to device functions. OpenCL requires that the arguments to those functions have their address space declared, which OCCA does not determine and insert for you today. You need the Generic Address Space feature from OpenCL 2.0 for that. If you avoid device functions, then I think you're good