Ciao,
Il 2022-02-23 12:35 Marco Bodrato ha scritto:
... then maybe I was wrong when I wrote that we should not trade a
factor 3 with a factor 2...
Yes, I think I was wrong. I pushed to the repository the use of the
_bknp1 (mod B^kn+1) code on the +1 side of the _bnm1 multiplication code
(mod B^2n-1).
And here are the measures by tune/speed on our development machine
"shell".
The _rounded measures use a larger size, rounded to some multiple of
some power of two. Actually, they are not faster than the non-rounded
size. Moreover, before the patch the _rounded cost was monotonous, now
it's not.
The rounding strategy needs a revision.
[bodrato@shell ~/gmp-repo]$ /var/tmp/bodrato/gmp/hg/build/tune/speed
-p100000000 -s 72-900 -rc -t36 mpn_mulmod_bnm1_rounded mpn_mulmod_bnm1
overhead 5.84 cycles, precision 100000000 units of 2.86e-10 secs, CPU
freq 3500.09 MHz
mpn_mulmod_bnm1_rounded mpn_mulmod_bnm1
72 4939.99 #0.9984
108 10049.80 #0.9249
144 12718.57 #0.9997
180 22397.62 #0.8529
216 23185.73 #0.9993
252 36533.91 #0.8669
288 34125.65 #0.9981
324 56029.22 #0.8225
360 48846.70 #0.9999
396 63004.78 #0.9731
432 61179.69 #0.9996
468 95840.50 #0.8255
504 #80506.97 1.0001
540 113637.14 #0.8538
576 #92016.26 1.0002
612 127849.99 #0.9169
648 #115918.53 1.0000
684 161355.14 #0.8673
720 130868.85 #0.9980
756 166164.06 #0.9613
792 197147.66 #0.7897
828 196738.21 #0.9304
864 212050.34 #0.9630
900 217971.55 #0.9548
Ĝis,
m
--
http://bodrato.it/papers/
_______________________________________________
gmp-devel mailing list
gmp-devel@gmplib.org
https://gmplib.org/mailman/listinfo/gmp-devel
