spirv translator assert failure building opencl-neo targets via ocloc

[dvrogozh]

Assuming that IGC and common_clang were built against LLVM-8 per this instruction: https://github.com/dvrogozh/intel-graphics-compiler/blob/rfc_readme/README.md#building

Then building https://github.com/intel/compute-runtime the following build failure could be observed:

cd /home/neo/compute-runtime/runtime/scheduler && LD_LIBRARY_PATH=/home/neo/compute-runtime/build/bin /home/neo/compute-runtime/build/bin/ocloc -q -file scheduler.cl -device skl -cl-intel-greater-than-4GB-buffer-required -64 -out_dir /home/neo/compute-runtime/build/bin/scheduler/x64/gen9 -cpp_file -options -I/home/install.llvm8.igc/include/igc\ -I/home/install.llvm8.igc/include/igc/cif\ -I/home/install.llvm8.igc/include/igc/ocl_igc_shared/executable_format\ -I/home/install.llvm8.igc/include/igc/ocl_igc_shared/device_enqueue\ -I\ ../gen9\ -cl-kernel-arg-info\ -cl-std=CL2.0

ocloc: /home/SPIRV-LLVM-Translator/lib/SPIRV/SPIRVWriter.cpp:685: SPIRV::SPIRVInstruction* SPIRV::LLVMToSPIRV::transUnaryInst(llvm::UnaryInstruction*, SPIRV::SPIRVBasicBlock*): Assertion `Cast->getSrcTy()->getPointerAddressSpace() == SPIRAS_Generic' failed.

That's an assert coming from here: https://github.com/KhronosGroup/SPIRV-LLVM-Translator/blob/0650ff887b5ef00066e6a0527bab592dafd59c83/lib/SPIRV/SPIRVWriter.cpp#L685

Also, I have tried to apply patches mentioned by common_clang here: https://github.com/intel/opencl-clang/tree/ocl-open-80/patches. These patches did not fix the above issue.

[dvrogozh]

Assertion is:
assert(Cast->getSrcTy()->getPointerAddressSpace() == SPIRAS_Generic)

I just checked that Cast->getSrcTy()->getPointerAddressSpace() returns 0. Which is SPIRAS_Private per definition below:

https://github.com/KhronosGroup/SPIRV-LLVM-Translator/blob/0650ff887b5ef00066e6a0527bab592dafd59c83/lib/SPIRV/SPIRVInternal.h#L181

enum SPIRAddressSpace {
  SPIRAS_Private,
  SPIRAS_Global,
  SPIRAS_Constant,
  SPIRAS_Local,
  SPIRAS_Generic,
  SPIRAS_Count,
};

[AlexeySotkin]

The issue is caused by casting q from private address space to global. I'm not sure it is legal. Need to investigate it further. BTW, there is no __builtin_astype in the OpenCL C spec.

__global IGIL_CommandHeader* TEMP_IGIL_GetNthCommandHeader( queue_t q, uint initialOffset, uint number )
{

    __global uchar *pQueueRaw = __builtin_astype( q, __global uchar* );

    __global IGIL_CommandHeader* pCommand = ( __global IGIL_CommandHeader* )( pQueueRaw + initialOffset );

    uint Offset = initialOffset;
    //Traverse queue_t unless nth command packet is found
    while( number > 0 )
    {
        Offset  += pCommand->m_commandSize;
        pCommand = TEMP_IGIL_GetCommandHeader( q, Offset );
        number--;
    }

    return pCommand;
}

[dvrogozh]

Just posting a link to the place in sources pointed by @AlexeySotkin:
https://github.com/intel/compute-runtime/blob/cb39644c694455ff4d78571aeb15f5ede6c17e0f/runtime/scheduler/scheduler.cl#L167

[AlexeySotkin]

I think the easiest way to fix the issue is to modify scheduler.cl

The kernel gets pQueue as argument of type __global IGIL_CommandQueue*
__kernel __attribute__((intel_reqd_sub_group_size(PARALLEL_SCHEDULER_COMPILATION_SIZE_20))) void SchedulerParallel20(
__global IGIL_CommandQueue* pQueue,
__global uint* commandsStack,

Then it passes pQueue to the following function, but for some reason casted to queue_t
TEMP_IGIL_GetNthCommandHeaderOffset( __builtin_astype( pQueue, queue_t ), InitialOffset, CurrentPacket );

But what we do with it in this function is casting to __global uchar*
__global IGIL_CommandHeader* TEMP_IGIL_GetNthCommandHeader( queue_t q, uint initialOffset, uint number ) { __global uchar *pQueueRaw = __builtin_astype( q, __global uchar* );

I believe if we change TEMP_IGIL_GetNthCommandHeaderOffset to take __global IGIL_CommandQueue* as argument and cast it directly to __global uchar* skipping the queue_t representation it should be possible to convert it to SPIR-V.
There are a few other the same cases which should be modified in the same way though.

[dvrogozh]

@AlexeySotkin: I made a change you've suggested, I pushed it here intel/compute-runtime#143. Unfortunately it seems that some earlier versions of clang don't like the change, but I hope it should be ok for LLVM-8.

I have also stepped into another issue on IGC side building the same command line. intel/intel-graphics-compiler#77. And here is a fix: intel/intel-graphics-compiler#78

[AlexeySotkin]

@dvrogozh Can we close this issue ?

[dvrogozh]

Fix in neo did land yet. Can we wait few days till it will be merged? or we will need to submit bug directly against neo.