timelockanalysis: use parfor instead of for #2068
Conversation
|
I am looking at a way where, for ~keeptrials, the size of covsig can stay at [nchan, nchan], instead of using [nrpt, nchan, nchan]. I found it is possible. For clarity, it should use a small helper function that contains the first half of the parfor loop, that calculates dat and numsmp. OLD: Due to how fieldtrip does the preamble, it cannot be a nested function nor a subfuction, but needs to be in its own file. |
With some small adaptations, it is possible to use a parfor loop in ft_timelockanalysis. This potentially speeds up the calculation, when the parallel computing toolbox is available.
Setting this limit to 0 will make the execution non-parallel
For the not-keeptrials case, it is possible to use less memory, by adding results within the loop. This was done with the regular for loop. Rework the parfor implementation to also do the reduction within the loop. Does require having two implementations of the parfor loop, because variables need to always be assigned to in the same way within a single parfor loop.
AljenU
force-pushed
the
parfor_timelockanalysis
branch
from
a6013d7
to
9af62bb
Compare
robertoostenveld
moved this from In progress
to Done
in FieldTrip performance improvements
With some small adaptations, it is possible to use a parfor loop in
ft_timelockanalysis. This potentially speeds up the calculation, when the
parallel computing toolbox is available.
Tested with https://github.com/AljenU/fieldtrip-parallel-support/tree/additional_tests, however no tests were found that take more than a few seconds on ft_timelockanalysis, making it hard to actually compare performance.