ni...@lysator.liu.se (Niels Möller) writes:
> So the gcd_22 code with all the branches needs about 11 cycles per input
> bit. gcd_11 is coreihwl/gcd_11.asm in this build.
I had to do a quick try with the masking version before leaving for work:
$ ./tune/speed -p 100000 -s 1-64 -t 3 -C mpn_gcd_11 mpn_gcd_22
overhead 4.01 cycles, precision 100000 units of 5.08e-10 secs, CPU freq 1966.75
MHz
mpn_gcd_11 mpn_gcd_22
1 #7.0459 11.0729
4 #2.8721 6.6946
7 #2.8415 7.3658
10 #3.3039 7.4346
13 #3.6084 7.8537
16 #4.0952 7.8193
19 #4.0620 8.0769
22 #3.9624 7.9849
25 #3.9169 8.0956
28 #4.0359 8.2084
31 #3.9816 8.0920
34 #3.9748 8.0933
37 #4.0041 8.0792
40 #3.9327 8.0656
43 #3.9239 8.1070
46 #3.9064 8.0289
49 #3.9143 8.0191
52 #4.0642 9.5547
55 #4.1251 9.8507
58 #4.2523 7.9096
61 #3.8440 7.8590
64 #3.8325 7.8398
So down to around 8 cycles per input bit.
Regards,
/Niels
--
Niels Möller. PGP-encrypted email is preferred. Keyid 368C6677.
Internet email is subject to wholesale government surveillance.
_______________________________________________
gmp-devel mailing list
gmp-devel@gmplib.org
https://gmplib.org/mailman/listinfo/gmp-devel
