![]() I could imagine that other languages might not suffer so much from this if they have sleep(0) in the code to deferred to the OS at non critical points but the risk of register truncation still exists, it is just much less likely. This was in tensor flow with anaconda and numpy. However, running real time with an affinity of one processor will preserve the register in multiply accumulate operations such that the task scheduler is deferred. If so that problem is because inside the xx86 CPU is a extended precision floating point register, and on a task scheduling event it will truncate that value as the full bit precision is not externally accessible to be saved and restored. I think the issue may relate to multi core xx86 PC CPU's if it is the same effect as I saw, and in that case seeding the generator or turning the shuffle off will not fix the problem.
0 Comments
Leave a Reply. |