Stefan Marsiske
Stefan Marsiske
i tested with /usr/bin/test-clkernel, /usr/bin/test-clenabled, /usr/bin/test-clfilter, /usr/bin/test-clxcorrelate, /usr/bin/test-clxengine and they all run correctly
i managed to get a backtrace in gdb: ``` Thread 65 "clMathOp6" received signal SIGSEGV, Segmentation fault. [Switching to Thread 0x7fff497fa700 (LWP 16741)] NEO::DrmAllocation::makeBOsResident (this=0x6ddaec0, osContext=0x5f15080, vmHandleId=vmHandleId@entry=0, bufferObjects=bufferObjects@entry=0x13de378, bind=bind@entry=false) at...
i dug down into this, and i'm convinced this is a different issue, the thing is, i can only test it with my patch in this pull request. anyway here...
would you be able to propose specific code to patch in, i'd be happy to run tests
i feed a sinus and a cosinus signal source in my gnuradiocompanion flowgraph into a cl-multiply and that into a nullsink. i don't really do buffers at all.
they are all complex.
i prepared testcases linked in my analysis under https://gist.github.com/stef/5e2d8a88f1d9d220f623a55531a9fb06
the intel people have posted some insights, does any of this ring a bell with you? https://github.com/intel/compute-runtime/issues/409#issuecomment-802685214
i created a much simpler c++ only testcase which still fails for me: https://gist.github.com/stef/e7818dc85b48c06d98a333e01f3d526f
there seems to be a violation of the opencl spec according to the intel engineers reproducing the testcase: https://github.com/intel/compute-runtime/issues/409#issuecomment-846040815