parallelise descriptor computation.

[zhubonan]

It would be nice to use all those core of a PC.

At the moment ASAP is serial. It would be relatively easy to parallelise over the frames using multiprocessing.Pool in the following code:

ASAP/asaplib/data/xyz.py

Lines 197 to 205 in 2ccd298

	for i in sbs:
	frame = self.frames[i]
	# compute atomic descriptor
	desc_dict_now, atomic_desc_dict_now = global_desc.compute(frame)
	self.global_desc[i].update(desc_dict_now)
	if keep_atomic:
	self.atomic_desc[i].update(atomic_desc_dict_now)
	# we mark down that this descriptor has been computed
	self.computed_desc_dict['descriptors'][tag] = global_desc.desc_spec_dict

There can be some complications while combining this with having a progress bar, but solutions can be found at tqdm/tqdm#484.

Alternatively, there is already parallel implementation in dscribe, but one needs to modify the call stack and group the calls to SOAP.create together.

[BingqingCheng]

I prefer to do this at the asap level, as not all descriptors are parallelized, and this is really a trivial parrallelzation.

To do this properly, I'm thinking perhaps we first sort the frames according to sizes (as a proxy to the associated computing cost), and then start N processes for each N frames with similar sizes.

What do you think?

[zhubonan]

I agree. Sorting is probably not necessary as one can just use strides of the input frames? There might be some complications for avoiding using too much memory. For example, if the child process just inherent all the resources, the memory cost will increase N fold.

It may be better to use the joblib library, this is what dscribe uses.

https://joblib.readthedocs.io/en/latest/parallel.html#embarrassingly-parallel-for-loops

[BingqingCheng]

done