Hey Andreas, On Thu, Nov 6, 2014 at 12:50 PM, Andreas Tille <andr...@an3as.eu> wrote:
> > KMC said: > > > > Asmlib is essential in performance sense. In KMC there is a lot of work > > with buffers, sometimes these buffers or its part must be copied and > asmlib > > provide quicker copying than standard implementation. > > I'd keep on asking two things: > > 1. How quick means "quicker". What factor of speed does the usage of > asmlib gain? > > 2. Is there any chance to provide a compile time option > > --without-amslib > > or whatever name which simply uses the native implementation? > > Brilliant. I'll do that. > > > In the sense what I wrote above I personally would be happy if KMC > > > authors would use plain C/C++ code and do some serious testing what > > > speed they really gain and how maintainable their code would be. > > > > > > > I will see if I can get some evidence form the KMC authors as to why they > > choose this implementation, rather than native. > > Fine. At least an optional usage should come pretty cheap with some > preprocessor statements. > > yeap! > > I also thought this task would be much simpler than what it's turning out > > to be. > > ... as always. ;-) > :) > > > > Here we are facing another drawback of asmlib usage: KMC authors are > > > excluding promising architectures like arm64 and ppc64el which might in > > > the not so distant future could become relevant for tasks in > > > bioinformatics. > > > > Very valid point. I will bring this point accross on the next > communication. > > I can cc you if you want. If you want to talk to them directly, that > could > > also help. > > I'd consider it the best idea to CC debian-med list. I'm definitely no > expert about assembly code and we could easily ask for further help on > other Debian lists if the discussion is archived online. > Sounds good. Will CC debian-med on the next communication. > > BTW, I had a *very* short sneak into the code and by doing one pretty > simple > experiment with > > kmc-2.0/kmer_counter $ diff -u mem_disk_file.cpp.orig mem_disk_file.cpp > --- mem_disk_file.cpp.orig 2014-10-23 16:22:21.000000000 +0200 > +++ mem_disk_file.cpp 2014-11-06 13:47:09.834995768 +0100 > @@ -10,7 +10,10 @@ > */ > > #include "mem_disk_file.h" > -#include "libs/asmlib.h" > + > +// #include "libs/asmlib.h" > +#define A_memcpy memcpy > +#define nullptr NULL > > > > //---------------------------------------------------------------------------------- > // Constructor > > > The compiler revealed only issues which are probably caused by something > else than asmlib. So wrapping this into #ifdef statements could help > here - but please consider this as the most simple experiment to look > how our chances to avoid asmlib are and not as real code. > > I understand. We had a chat around the office and were thinking of building kmc without using asmlib to understand the trade off. But also here we put this as plan b for now. Thanks for having a look at this. Regards, Jorge
