armintoepfer / aligner-testbed Goto Github PK

I saw someone forked miniwfa, so I followed that and came here :-) It is recommended to initialize km = km_init() once per thread and then free it with km_destroy() close to the end of the thread (should document this). In your code, call km_init() before for(i) at this line and free it after the loop. Miniwfa is doing frequent heap allocation and relying on kalloc to reduce the overhead. Without kalloc, it can be ~30% slower. The slowdown may be more noticeable when you align different pairs in multiple threads.

WFA2-lib uses heuristic by default (see smarco/WFA2-lib#7). You need to add attributes.heuristic.strategy = wf_heuristic_none if you want to use the exact algorithm. Miniwfa doesn't support any heuristic as I don't have the need for now. It is not hard to implement them.

BTW: 1) I saw you mentioned banding on twitter. I actually like fixed bandwidth. Fixed banding is predictable. We can often get a good sense whether the bandwidth is enough by looking at the cigar or even before getting the alignment. We can then double the bandwidth when we worry it is too small. Such a band doubling heuristic would be at most twice as slow in comparison to optimal banding. Bwa-mem has a version of this. With wavefront reduction, it is hard to tell when the heuristic fails. And when the heuristic fails, we may need more time to do alignment as the score would be higher. 2) I am not sure if WFA works for spliced alignment in its current form. WFA assumes non-exact-match incurs a positive penalty. In spliced alignment, however, intron extension doesn't increase the score. (EDIT: we can make intron extension cost, though I am not sure about the effect of that) I guess there should be a way to make WFA work, but even in that case, WFA may need to fill most part of the DP matrix and lose its key advantage over NW.