You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I stumbled on this bug when setting up an environment for debugging/profiling OpenCL code (Portable CL + Cuda JIT compilation to profile on nvidia GPU).
The error message printed before the program abortion is:
Aborted
After inspection, it turns out it is the probing for atomic operation on devices in 'silx.opencl.atomic' which causes the crash.
Deeper inspection with gdb shows that PoCL explicitely asked for the termination of the program by calling pocl_cuda_abort_on_error. My problem is that we are compiling some code to actually validate if those atomic operation at available or not. Here, the PTX code is apparently not useable. silx expected an exception, not a brutal program termination !
What needs to be done:
Warn the developers of PyOpenCL (probably nothing they can do on their side) and PortableCL.
Deactivate the probing of devices when driver=="ProtableCL" and devicetype=="GPU" since it crashes the program without being able to recover it at the Python level.
The text was updated successfully, but these errors were encountered:
I stumbled on this bug when setting up an environment for debugging/profiling OpenCL code (Portable CL + Cuda JIT compilation to profile on nvidia GPU).
The error message printed before the program abortion is:
After inspection, it turns out it is the probing for atomic operation on devices in 'silx.opencl.atomic' which causes the crash.
Deeper inspection with
gdb
shows that PoCL explicitely asked for the termination of the program by callingpocl_cuda_abort_on_error
. My problem is that we are compiling some code to actually validate if those atomic operation at available or not. Here, the PTX code is apparently not useable.silx
expected an exception, not a brutal program termination !What needs to be done:
The text was updated successfully, but these errors were encountered: