Why has simple threading using `threadid` become so complex in v1.12...?

General Usage Performance
30

See No more 1st class support for Threads? [thread-local storage] - #34 by foobar_lv2

Also note that if one does not need to store the task-local buffer and does not need the indices, this can be simplified to:

model = fit(Model, data)
# split work in n=nthreads() chunks
@threads for chunk in ChunkSplitters.chunks(1:1000; n=Threads.nthreads())
   m = deepcopy(model)
   for item in chunk
       predict!(m, newdata)
   end
end

Which makes pretty clear that the only n copies of model are created, and that there is not race condition among their uses.

2 Likes