As far as I understood the current implementation, the SOFA file is accessed for each head orientation and for each loudspeaker position separately. This means, that an impulse response (having e.g. the index "42" ) is loaded from the hard disk drive into the RAM several times. If we really want to speed up things, then this architecture has to be reassessed. An alternative workflow could be:
iterate over the head orientations and the secondary source position in order to get the indices of the impulse responses, which are needed from the SOFA file. This can be done by evaluating the header of the SOFA file, i.e. using nearest_neighbor, etc. (as you have done it in the current implementation)
the indices are likely to be non-unique, as some head-orientations-ssd-position pairs might need the same impulse responses. However, functions like unique (http://de.mathworks.com/help/matlab/ref/unique.html) help us to remove those duplicates.
after loading the impulse responses, they have to be processed according to the head-orientations-ssd-position pairs they are need for, i.e. distance adjustment, angle interpolation, adding up all loudspeakers.
The major drawback of this approach is, that there might be a huge number of impulse responses which are loaded into the RAM at once. One could define a maximum number of irs loaded in one step and then iterate as long as there are impulse responses left to be processed. However, one would have ensure, that there is always at least one head-orientations-ssd-position pair which can be processed with the current impulse responses in the RAM.
If someone has an idea how to do this or some time left we can work on this, but the priority is not very high at the moment.
As @fietew wrote in #45:
If someone has an idea how to do this or some time left we can work on this, but the priority is not very high at the moment.