test with 1 thread, fix 1 threaded segfault #154
Conversation
|
With the newest releases of StrideArraysCore, VectorizationBase and LoopVectorization this fails due to rounding errors? Rest seems to work fine. (@v1.9) pkg> test SimpleChains Stacktrace: Stacktrace: |
|
And on multi threaded I get KILL julia -t auto (@v1.9) pkg> test SimpleChains |
|
Rounding errors are a little different with the curent Project.toml _ _ ()_ | Documentation: https://docs.julialang.org (@v1.9) pkg> test SimpleChains Stacktrace: Stacktrace: |
|
Unfortunately even with the older libraries multithreaded ends with a KILL julia -t auto (@v1.9) pkg> test SimpleChains |
|
Single threaded passed (@v1.9) pkg> test SimpleChains |
|
multi threaded still dies with a KILL (@v1.9) pkg> test SimpleChains |
|
this also happens when I use the newest release versions of LoopVectorization, StrideArraysCore and VectorizationBase (@v1.9) pkg> test SimpleChains |
| aligned_glen = align(glen) | ||
| g = _alloc_grad(Ptr{T}(pu), glen, numthreads, aligned_glen) | ||
| offset = static_sizeof(T) * aligned_glen * numthreads | ||
| train_unbatched_core!(c, pu + offset, g, pX, it, p, opt, mpt) | ||
| if numthreads == 1 | ||
| train_unbatched_core!(c, pu + offset, g[:, begin], pX, p, opt, it, mpt) |
There was a problem hiding this comment.
could you write a comment on why this is needed? can this branch be handled in train_unbatched_core! since all the other inputs are the same?
Fixes the problem reported in #153, and starts testing with a single thread to avoid such occurrences in the future.
@bendat78
Mind reviewing @mohamed82008 ?