Closed WrathfulSpatula closed 2 years ago
The problem might actually be due to a SEGFAULT, which NVIDIA apparently reports as -5
:
https://stackoverflow.com/questions/23742619/opencl-cl-out-of-resources-in-a-for-loop
It was a segmentation fault, due to recent changes to QEngineOCL::ShuffleBuffers()
. The need for these changes was predicated on ShuffleBuffers()
being made asynchronous, which is a feature currently on hold. However, with the method reverted, QPager
works again. Further, all deadlock conditions and failures on multi-page ket simulation widths have been fixed, (for real, this time,) so we'll want to iterate a PyQrack release immediately.
At this stage in development, it seems that
QPager
raises-5
(out of resources) in pretty much every nontrivial use case. This might have to do with genuine hardware limitations on my (and other) systems, but it should be debugged. IfQPager
cannot ultimately be used to attain qubit widths beyond max single allocation segment, it should be (temporarily) removed from the default layer stack.