Closed chpatrick closed 4 years ago
Would very much appreciate this being upstreamed.
CI failure seems unrelated:
/tmp/ccilBvEP.s: Assembler messages:
/tmp/ccilBvEP.s:15569: Error: operand type mismatch for `vxorps'
It pains me that I didn't manage to look into the cause of the undeterministic behavior. I would like to rephrase the description to clarify that setting the number of threads to one ensures deterministic behavior but that the algorithm itself is deterministic and that it is a bug in the implementation that causes the undeterministic behavior.
It seems like the only way to get bit-identical output is to disable multithreading. This PR adds a command-line flag to set the number of threads for both OpenMP and TBB.