-
Notifications
You must be signed in to change notification settings - Fork 79
/
TODO
33 lines (27 loc) · 1.54 KB
/
TODO
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
TODO:
* Data parallelization
MrBayes currently uses MPI to run different chains in parallel
on different processors. However, data sets are becoming a lot
larger now with respect to number of sites, so it is becoming
important also to distribute likelihood computations on data across
processors. This would require some programming but it is not a
huge challenge.
* Native AVX and FMA code
I have added native AVX and FMA code for the 4by4 nucleotide models.
It would be straightforward to add native AVX and FMA code for the
other models. Particularly the AA-model would probably benefit from
this code (when not using the beagle library).
* Dynamic scaling
Scaling of likelihoods take a surprisingly large amount of time when
using the standard 4by4 nucleotide models. Maxim added dynamic
scaling code for the Beagle implementation but not for the native
likelihoood calculators, which are faster. The dynamic scaling
code adjusts the rescaling frequency so that you do not rescale
likelihoods more often than needed to avoid numerical exceptions and
achieve a reasonable level of precision. For typical-size trees
and the 4by4 model, no rescaling is often needed. I added the
infrastructure needed to support dynamic scaling for both Beagle and
native calculators in December 2015. In principle, it should be
straightforward now to rewrite Maxim's dynamic scaling code (which
is not so pretty C code) so that it works for both Beagle and native
likelihood calculators.