CTS stream_api_core test segfault

[alexbatashev]

Describe the bug
stream_api_core test results in segmentation fault:

--- stream_api_core
hello world!hello world!c9999, 9999, 99, 9999, 99, 99, 99999999, 9999, 99, 999955  5  5  5  5  5  5 -5 +5 -5  5  55.00000e+005.00000e+00  5Segmentation fault (core dumped)

To Reproduce
Note: Due to a recent uplift of minimal supported C++ standard, one should modify this line to require C++17.

1. Get SYCL with CUDA support (GSG) and SYCL-CTS
2. Build CTS tests

cd SYCL-CTS
mkdir build && cd build
cmake -GNinja  -DSYCL_IMPLEMENTATION=Intel_SYCL -DINTEL_SYCL_ROOT=<path to built sycl> -Dopencl_platform_name=nvidia -Dopencl_device_name=opencl_gpu -DCMAKE_BUILD_TYPE=Release -DINTEL_SYCL_FLAGS="-Xsycl-target-backend;--cuda-gpu-arch=sm_50" -DCMAKE_EXE_LINKER_FLAGS=" -Wl,-no-relax " -DINTEL_SYCL_TRIPLE=nvptx64-nvidia-cuda-sycldevice -DSYCL_CTS_ENABLE_OPENCL_INTEROP_TESTS=Off -DSYCL_CTS_ENABLE_DOUBLE_TESTS=On -DSYCL_CTS_ENABLE_HALF_TESTS=On ..
ninja test_stream

3. Run tests

./bin/test_stream -p nvidia -d opencl_gpu  --test stream_api_core

Environment:

OS version: Ubuntu 18.04.5 LTS (Bionic Beaver)
CUDA version:

nvidia-smi

+-----------------------------------------------------------------------------+
| NVIDIA-SMI 435.21       Driver Version: 435.21       CUDA Version: 10.1     |
|-------------------------------+----------------------+----------------------+

Additional context
Backtrace:

#0  0x00007ffff533d10f in ?? () from /usr/lib/x86_64-linux-gnu/libcuda.so.1
#1  0x00007ffff5220ce1 in ?? () from /usr/lib/x86_64-linux-gnu/libcuda.so.1
#2  0x00007ffff538e9a0 in cuCtxSetCurrent () from /usr/lib/x86_64-linux-gnu/libcuda.so.1
#3  0x00007ffff62cf2a4 in (anonymous namespace)::ScopedContext::ScopedContext(_pi_context*) ()
   from lib/libpi_cuda.so
#4  0x00007ffff62d1fdb in cuda_piEventsWait ()
   from lib/libpi_cuda.so
#5  0x00007ffff7a73277 in cl::sycl::detail::event_impl::waitInternal() const ()
   from lib/libsycl.so.5
#6  0x00007ffff7ae9c58 in cl::sycl::detail::DispatchHostTask::operator()() const ()
   from lib/libsycl.so.5
#7  0x00007ffff7aca006 in std::thread::_State_impl<std::thread::_Invoker<std::tuple<cl::sycl::detail::ThreadPool::start()::{lambda()#1}> > >::_M_run() () from lib/libsycl.so.5
#8  0x00007ffff738b15f in std::execute_native_thread_routine (__p=0x12e38e0) at ../../../../../gcc-7.5.0/libstdc++-v3/src/c++11/thread.cc:83
#9  0x00007ffff65076db in start_thread (arg=0x7fffd4ffd700) at pthread_create.c:463
#10 0x00007ffff6a4471f in clone () at ../sysdeps/unix/sysv/linux/x86_64/clone.S:95

[steffenlarsen]

That definitely shouldn't happen! Do you have the specific commits you're working off of, for both intel/llvm and the CTS?

[alexbatashev]

@steffenlarsen I used SYCL build from 04/16 (should be around this tag https://github.com/intel/llvm/releases/tag/sycl-nightly%2F20210416) and the latest CTS, that I checked out today.

[steffenlarsen]

Thank you! I'm currently trying to build as instructed, but I'm getting some errors. Just to make sure I'm not trying to fix the wrong problems, are you on SYCL-1.2.1/master or SYCL-2020 of the CTS?

[alexbatashev]

Thank you! I'm currently trying to build as instructed, but I'm getting some errors. Just to make sure I'm not trying to fix the wrong problems, are you on SYCL-1.2.1/master or SYCL-2020 of the CTS?

I'm on SYCL-1.2.1/master. I just checked again. Due to a recent uplift of minimal requirements, some headers are no longer compatible with C++11. So, I modified this line some time ago and completely forgot about it. I'll update both reports.

BTW, the issue is also reproducible on March SYCL builds, but I didn't check with older versions.

[steffenlarsen]

That did the trick, thanks! I am able to reproduce the issue and will do some investigation.

[bader]

@alexbatashev, can you switch to SYCL-2020 branch, please?
I think recent DPC++ changes might make it incompatible with SYCL-1.2.1/master branch.

The C++ version has been updated in the SYCL-2020 branch https://github.com/KhronosGroup/SYCL-CTS/blob/SYCL-2020/CMakeLists.txt#L5.

[alexbatashev]

Switching to SYCL-2020 branch fixes C++17 issue, but the test still segfaults

[npmiller]

I'm not able to reproduce this, are you still seeing these crashes?

Looking at the stack trace I highly suspect this was fixed as part of: #4442

We might be able to close this ticket.

[bader]

@yuxianch, could you confirm that stream test is passing on Linux, please?

[yuxianch]

Yes, the stream test stream_api_core is passing on Linux with latest compiler(https://github.com/intel/llvm.git 7700edb) and CUDA backend.