On Fri, May 21, 2010 at 11:17 AM, Sergey Bochkanov < sergey.bochka...@alglib.net> wrote:
> correction to my previous message: > > > It leads to unnecessary allocation of temporaries and some > > performance penalty (about 25-30% for 128-bit precision). > > Sorry, I made a mistake when estimating MPLAKACK performance. "25-30%" > are ALGLIB's, not MPLAPACK's. > > > I've never measured MPLAPACK performance directly, but looking at its > RGEMM code and measuring allocation penalty for code like this: > > > mpfr_t tmp1, tmp2; > > mpfr_init2(tmp1, Precision); > > mpfr_init2(tmp2, Precision); > > mpfr_mul(tmp1, mp_b, mp_c, GMP_RNDN); > > mpfr_add(tmp2, mp_a, mp_tmp, GMP_RNDN); > > mpfr_set(mp_d, tmp1, GMP_RNDN); > > mpfr_clear(tmp1); > > mpfr_clear(tmp2); > > (which reproduces allocations done by MPLAPACK during "a=a+b*c" update > in the RGEMM call) and comparing it with > > > mpfr_mul(mp_tmp, mp_b, mp_c, GMP_RNDN); > > mpfr_add(mp_d, mp_a, mp_tmp, GMP_RNDN); > > which does the same thing, but more efficiently, I can conclude that > MPLAPACK allocation penalty will be 100% for 128-bit precision code > (i.e. it will run twice longer), 50% for 256-bit precision. > > > As for ALGLIB, it really has 25-30% allocation penalty for 128-bit > code. The reason behind such difference from MPLAPACK is that ALGLIB > tries to allocate new multiple precision variables from internal > cache, but still relies on OOP, which slows it down. > > So when I said that "ALGLIB ... suffers from the same OOP drawbacks", > I was a bit incorrect - it suffers from OOP drawbacks, but from > _other_ drawbacks. > > -- > With best regards, > Sergey mailto:sergey.bochka...@alglib.net > MPFR has a fused multiply-add function, so d = a+b*c can be written mpfr_fma(d,a,b,c, MPFR_RNDN). This is almost certainly what you'd want to base MPFR linear algebra around. Fredrik -- To post to this group, send an email to sage-devel@googlegroups.com To unsubscribe from this group, send an email to sage-devel+unsubscr...@googlegroups.com For more options, visit this group at http://groups.google.com/group/sage-devel URL: http://www.sagemath.org
