Thread local variables in prange.

[Daniil Pakhomov]

Hello.

I would be very glad if someone will help me to solve this problem.
I want to use thread-local variables in a prange loop.

Why it's not possible to use cython.parallel.threadlocal like it's mentioned here
https://github.com/cython/cython/wiki/enhancements-prange ?

Is this the only possible workaround by now:
https://groups.google.com/forum/#!topic/cython-users/Ady-DdWu6rE ?

Thank you.

[Jerome Kieffer]

I agree this is a limitation ...

I often use a temporary shared array of size (nthread, output_size), then perform the merge within that array.

here is an example, a kind of histogram.

cdef int[:,:] tmp = numpy.zeros((numthreads, size2), int)
cdef int[:] out = numpy.zeros(size2, int)
for i in parallel.prange(size1, nogil=True):
threadid = parallel.threadid()
j = <int> data[i]
tmp[threadid, j] += 1
for j in parallel.prange(size2, nogil=True):
s = 0
for i in range(numthreads):
s = s + tmp[i, j]
out[j] += s

Note
* "s = 0" and "s = s + x" to enforce the thread locality (i.e. not shared)
* "out[j] +=" to enforce a parallel reduction so out is shared and not local

This works when the number of core is a few: there is a large overhead
with 2 cores or less and requires a lot of memory when there are too
many cores.

Cheers,

--
Jérôme Kieffer
tel +33 476 882 445

[Daniele Nicolodi]

I may be missing something, but what is the difference in memory usage
with respect to a solution that uses thread local storage? Each thread
would have to allocate its own tmp array with shape (size2, ), resulting
in exactly the same memory consumption.

Cheers,
Daniele

[Jerome Kieffer]

On Thu, 02 Jul 2015 13:32:42 +0200
Daniele Nicolodi <dan...@grinta.net> wrote:

> > This works when the number of core is a few: there is a large overhead
> > with 2 cores or less and requires a lot of memory when there are too
> > many cores.
>
> I may be missing something, but what is the difference in memory usage
> with respect to a solution that uses thread local storage? Each thread
> would have to allocate its own tmp array with shape (size2, ), resulting
> in exactly the same memory consumption.

you are right, it is the same.