Request from Phil - Page 7

fatphil · 2006-12-05, 07:42

Quote:

Originally Posted by fatphil

My highest priority area is getting non-x86 working. Then it's mapping everything that works to +1. I can reeorder those if it's important to you.

I found a way of re-using the -1 functions for +1 with basically no code changes to the inner loop, so didn't need to create any new functions - which was what I was trying to avoid.

Unfortunately there's a very minor error in selecting what string to print out:

Code:

bash-3.2$ grep '\+1' phrot.base5 | sort -n -t '^' -k 2 | ./phrot.cygp3 -b=3 --
No ABC header found: 150526*5^88954+1

Actually testing 11759843750*1953125^9883+1 (9885/20480 limbs)
150526*5^88954+1 [] is composite LLR64=0000000000000001. (e=0.21300 (0.248697~5.61599e-16@1.029) t=490.93s)
Actually testing 16112000*1953125^9937+1 (9939/20480 limbs)
128896*5^89436+1 [] is composite LLR64=0000000000000001. (e=0.21355 (0.249502~5.61515e-16@1.016) t=485.53s)
Actually testing 159106*1953125^9998+1 (9999/20480 limbs)
159106*5^89982+1 [] is composite LLR64=0000000000000001. (e=0.21051 (0.250396~5.51868e-16@1.156) t=494.93s)
Actually testing 2375781250*1953125^10096+1 (10098/20480 limbs)
6082*5^90872+1 [] is composite LLR64=0000000000000001. (e=0.22501 (0.251864~5.86971e-16@0.937) t=498.32s)
Actually testing 408370*1953125^11184+1 (11185/24576 limbs)
81674*5^100657+1 [] is composite LLR64=0000000000000001. (e=0.26068 (0.352333~6.46147e-16@1.081) t=650.47s)
Actually testing 242120*1953125^11413+1 (11414/24576 limbs)
48424*5^102718+1 [] is composite LLR64=0000000000000001. (e=0.26060 (0.348278~6.39445e-16@1.063) t=667.15s)
Actually testing 37718*1953125^11611+1 (11612/24576 limbs)
37718*5^104499+1 [] is composite LLR64=0000000000000001. (e=0.28148 (0.351721~6.8475e-16@1.047) t=680.47s)
Actually testing 698920*1953125^12252+1 (12253/24576 limbs)
139784*5^110269+1 [] is composite LLR64=0000000000000001. (e=0.30305 (0.362694~7.17695e-16@0.997) t=724.25s)
Actually testing 2828900*1953125^12319+1 (12321/28672 limbs)
113156*5^110873+1 [] is composite LLR64=0000000000000001. (e=0.22356 (0.296094~5.27964e-16@0.914) t=872.32s)
Actually testing 220468750*1953125^13790+1 (13792/28672 limbs)
2822*5^124117+1 [] is composite LLR64=0000000000000001. (e=0.25000 (0.311898~5.58041e-16@1.010) t=994.43s)
Actually testing 8711093750*1953125^14889+1 (14891/32768 limbs)
111502*5^134008+1 [] is composite LLR64=0000000000000001. (e=0.25000 (0.330937~5.37054e-16@1.077) t=1167.37s)
Actually testing 48402343750*1953125^15140+1 (15142/32768 limbs)
123910*5^136268+1 [] is composite LLR64=0000000000000001. (e=0.28125 (0.334197~5.99157e-16@0.932) t=1190.62s)
Actually testing 5780000*1953125^15199+1 (15201/32768 limbs)
9248*5^136795+1 [] is composite LLR64=0000000000000001. (e=0.25000 (0.33496~5.3155e-16@0.993) t=1196.90s)
Actually testing 262600000*390625^19490+1 (19492/40960 limbs)
84032*5^155925+1 [] is composite LLR64=0000000000000001. (e=0.01316 (0.0148587~6.17776e-16@1.078) t=1834.18s)
Actually testing 12398125000*390625^20400+1 (20402/40960 limbs)
158696*5^163207+1 [] is composite LLR64=0000000000000001. (e=0.01328 (0.0152628~6.09376e-16@1.052) t=1935.85s)
Actually testing 3085150*390625^31703+1 (31705/65536 limbs)
123406*5^253626+1 [] is composite LLR64=0000000000000001. (e=0.01758 (0.0203066~6.46977e-16@0.921) t=4961.95s)
Actually testing 26886250*390625^34877+1 (34879/73728 limbs)
43018*5^279020+1 [] is composite LLR64=0000000000000001. (e=0.02083 (0.0263105~7.31055e-16@0.918) t=6407.82s)

[] and 0000000000000001 both mean 'PRP' - so the above can be considered passed tests. :-|

I'll fix that - it's a one-line change - and upload src + binary later.

Phil

fatphil · 2006-12-05, 08:20

Version 0.5 usable for both +1 and -1s.
For x86, has passed every test I've thrown at it.

masser · 2006-12-06, 20:23

Here are some PRP checks with version 0.47... sorry about the delay...

Code:

175268*5^360870-1 [] is PRP. (e=0.47433 (3.09116e-12@1.050) t=5369.85s)
339728*5^249588-1 [] is PRP. (e=0.42856 (3.35822e-12@1.010) t=2606.00s)
74528*5^218272-1 [] is PRP. (e=0.44403 (3.72065e-12@1.008) t=1882.28s)
346078*5^176421-1 [] is PRP. (e=0.31680 (2.95267e-12@0.939) t=1251.37s)
201590*5^165080-1 [] is PRP. (e=0.36111 (3.47933e-12@1.148) t=1026.64s)
256568*5^160920-1 [] is PRP. (e=0.36111 (3.52408e-12@1.003) t=998.12s)
217592*5^157128-1 [] is PRP. (e=0.33334 (3.29206e-12@0.978) t=968.43s)
294704*5^149576-1 [] is PRP. (e=0.33334 (3.37412e-12@1.151) t=920.75s)
292414*5^148601-1 [] is PRP. (e=0.33333 (3.38498e-12@1.099) t=910.53s)
197774*5^147396-1 [] is PRP. (e=0.31250 (3.18642e-12@0.945) t=783.46s)

Is it safe to use the code for first time PRP tests yet?

axn · 2006-12-06, 23:01

Quote:

Originally Posted by masser

Is it safe to use the code for first time PRP tests yet?

0.50 incorporates error prediction to better select the FFT, so it should be more reliable for bigger n's (albeit a bit slower). Plus, it handles + side also. I would suggest that you do some double checks for _really_ large exponents, to get an idea of how reliable it is. After that, you can probably use it for first time tests.

fatphil · 2006-12-07, 08:34

Quote:

Originally Posted by masser

Is it safe to use the code for first time PRP tests yet?

With -b=3 and residues cross-checkable with LLR, I'd be using it myself if I were contributing computationally to the project. However, I might be biased ;-).

Thanks all for contributing to testing, and helping steer the development direction! All I need to do now is get the blasted PPC client working as well as the x86 one does.

masser · 2006-12-08, 23:02

From phrot:

Code:

254*5^505462-1 [44689,43586,148187,-67751] is composite LLR64=5677D0B14E4DC294. (e=0.02344 (0.0302932~6.11065e-16@1.146) t=12302.92s)
254*5^508092-1 [-6538,-182600,-145325,-116902] is composite LLR64=CD6A3645DDDFF93A. (e=0.02539 (0.030384~6.60276e-16@1.031) t=12284.84s)
325922*5^510412-1 [119512,109695,103208,28031] is composite LLR64=F0D01D647861E556. (e=0.02734 (0.0304644~7.09443e-16@1.001) t=12233.63s)
27994*5^512629-1 [-141771,-27006,-85378,-153019] is composite LLR64=AB80E2FAD245996C. (e=0.02734 (0.0305407~7.07908e-16@1.006) t=12286.74s)
257708*5^515106-1 [-176949,-175831,67357,-73407] is composite LLR64=62F960171CEB9A48. (e=0.02539 (0.030626~6.55759e-16@1.015) t=12411.93s)
254*5^517492-1 [-14034,-132397,129767,-168487] is composite LLR64=BF35F924D429E08E. (e=0.02734 (0.0307075~7.04579e-16@1.019) t=12428.30s)
257708*5^519694-1 [190177,32208,-193729,61051] is composite LLR64=F58BF2834E44EFE5. (e=0.02734 (0.0307832~7.0308e-16@0.998) t=12590.07s)
254*5^522028-1 [185893,38135,189686,-68236] is composite LLR64=1EAA8B60D618D34C. (e=0.02734 (0.0308627~7.01511e-16@1.003) t=12620.02s)

From PRP:

Code:

254*5^505462-1 is not prime.  RES64: 5677D0B14E4DC294.  OLD64: 197E0390AA68694C
254*5^508092-1 is not prime.  RES64: CD6A3645DDDFF93A.  OLD64: 032ABC880B24B1CE
325922*5^510412-1 is not prime.  RES64: F0D01D647861E556.  OLD64: 8FD302FC87CDFC7D
27994*5^512629-1 is not prime.  RES64: AB80E2FAD245996C.  OLD64: 4AD8AC8CCC9F1C20
257708*5^515106-1 is not prime.  RES64: 62F960171CEB9A48.  OLD64: 28EC204556C2CED5
254*5^517492-1 is not prime.  RES64: BF35F924D429E08E.  OLD64: E841E3502A29626D
257708*5^519694-1 is not prime.  RES64: F58BF2834E44EFE5.  OLD64: 789B4A89AAB58921
254*5^522028-1 is not prime.  RES64: 1EAA8B60D618D34C.  OLD64: 5FE2EDD8687AAF27

I think this is good enough confirmation for me... I might begin using phrot for first time tests soon....

axn · 2006-12-08, 23:46

Quote:

Originally Posted by masser

I think this is good enough confirmation for me... I might begin using phrot for first time tests soon....

Go for it

It will be worthwhile to keep an eye out for biggish error values (say > 0.4) and doublecheck them with LLR/PRP.

btw, how does the runtimes compare with PRP? I would imagine that P4's won't see any gain at all, whereas everything else would see a 2x gain?

masser · 2006-12-09, 00:54

Quote:

Originally Posted by axn1

btw, how does the runtimes compare with PRP? I would imagine that P4's won't see any gain at all, whereas everything else would see a 2x gain?

I've only been testing on a P4 - there is a slight boost with phrot:

For the n = 505462 results, the timings were:

PRP: 213 minutes
PHROT: 205 minutes

The 4% boost IS significant, but I'm anxiously waiting for the YEAFFT code to be worked out for the SSE2.....

geoff · 2006-12-10, 23:13

If you want to make a Linux binary available for P2 or P3 machines I will test it.

Can your ideas be used to speed up P-1 and ECM factoring too?

Ken_g6 · 2006-12-11, 17:25

That does sound like a nice speedup!

Hmmm - Since this is using a generic FFT library, I wonder if something like GPUFFTW would be useable?

Ken_g6 · 2006-12-13, 00:37

Hm - I tried to compile the latest source here, but I seem to be missing some dependencies. Can anyone who's built it help me?

For one, I got and compiled GLucas, but compiling it didn't produce libyeafft.*.a (where I think * should be "athlon-pc-cygwin"? Or is it athlonxp or something?)

For another, I don't have /usr/local/include/Make.inc. What goes there? I assume ${HOSTTYPE}, and maybe ${PHILMATH}?

2006-12-11, 17:25	#76
Ken_g6 Jan 2005 Caught in a sieve 5×79 Posts	That does sound like a nice speedup! Hmmm - Since this is using a generic FFT library, I wonder if something like GPUFFTW would be useable? Last fiddled with by Ken_g6 on 2006-12-11 at 17:25

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Request	pinhodecarlos	Lounge	3	2017-10-26 18:58
Bug/request	Dubslow	YAFU	4	2012-03-31 03:07
Odd request?	Xyzzy	Lounge	23	2011-03-08 17:50
GMP-ECM Request	rogue	GMP-ECM	4	2009-11-23 15:07
A little request	10metreh	Aliquot Sequences	8	2009-10-07 18:13

2006-12-10, 23:13	#75
geoff Mar 2003 New Zealand 1157₁₀ Posts	If you want to make a Linux binary available for P2 or P3 machines I will test it. Can your ideas be used to speed up P-1 and ECM factoring too?

2006-12-13, 00:37	#77
Ken_g6 Jan 2005 Caught in a sieve 613₈ Posts	Hm - I tried to compile the latest source here, but I seem to be missing some dependencies. Can anyone who's built it help me? For one, I got and compiled GLucas, but compiling it didn't produce libyeafft..a (where I think should be "athlon-pc-cygwin"? Or is it athlonxp or something?) For another, I don't have /usr/local/include/Make.inc. What goes there? I assume ${HOSTTYPE}, and maybe ${PHILMATH}?