Thread: new yafu stuff View Single Post
 2018-06-04, 14:46 #7 bsquared     "Ben" Feb 2007 329010 Posts r372 has new AVX512 code for SIQS, plus a couple tweaks for more efficiently running on large numbers of threads. RSA-100 in 6 minutes on a 256-thread KNL 7210. Even without the AVX512 code the KNL is pretty fast (about 9 minutes). I've also tested on a skylake 7800x but the performance gains aren't as nice there. Maybe 10% and only with larger (80 digits or more) inputs. This is linux only at this point... I have not yet been able to find a compiler capable of avx512 for windows and so I haven't been able to test there. The Intel compiler can, I know, but I don't have a windows version of it. Code: starting SIQS on c100: 1522605027922533360535618378132637429718068114961380688657908494580122963258952897654000350692006139 ==== sieve params ==== n = 100 digits, 330 bits factor base: 115520 primes (max prime = 3204611) single large prime cutoff: 464668595 (145 * pmax) double large prime range from 44 to 52 bits double large prime range from 10269531661321 to 3988944358351073 allocating 9 large prime slices of factor base buckets hold 2048 elements large prime hashtables have 1179648 bytes using KNL-AVX2 enabled 32k sieve core sieve interval: 8 blocks of size 32768 polynomial A has ~ 13 factors using multiplier of 1 using SPV correction of 19 bits, starting at offset 30 trial factoring cutoff at 100 bits ==== sieving in progress (256 threads): 115584 relations needed ==== ==== Press ctrl-c to abort and save state ==== 119777 rels found: 29067 full + 90710 from 1703942 partial, (4709.08 rels/sec) sieving required 6419308 total polynomials (4705 'A' polynomials) trial division touched 258498542 sieve locations out of 3365566152704 squfof: 0 failures, 1386863 attempts, 121153992 outside range, 6726584 prp, 1281923 useful total reports = 258498542, total surviving reports = 129718525 total blocks sieved = 102712976,avg surviving reports per block = 1.26 large prime scan failures = 0 QS elapsed time = 368.4496 seconds.