SkylakeX teasers (aka prime95 29.5) - Page 15

kriesel · 2019-01-12, 16:19

Quote:

Originally Posted by pepi37

Where I can download latest version for linux64 bit?
And what is latest version beta6 or beta7?

At post 148, in pdf attachment, see links at bottom right.

kladner · 2019-01-12, 16:48

I have not had a hang with the new executable. I ran more 2560K benchmarks, since current work is using 2688K. All hangups, to my knowledge, have been with 2560K. I am now running benches from 2048K to 2560K to see if it hangs over a wider range of FFTs.

kriesel · 2019-01-12, 17:15

Quote:

Originally Posted by kladner

I have not had a hang with the new executable. I ran more 2560K benchmarks, since current work is using 2688K. All hangups, to my knowledge, have been with 2560K. I am now running benches from 2048K to 2560K to see if it hangs over a wider range of FFTs.

In my hang results, fft length varied. A lot. Part of that was where I chose to start (sometimes 1024K, sometimes 2048K, sometimes whatever length at which it hung previously), part was something else.

kriesel · 2019-01-13, 18:42

In prime95 29.5b7 for Win64, I tried the following in worktodo

PRP=N/A,1,2,51770007,-1,73,1
It resulted in the start of a P-1 factoring run.

Correcting the form to PRP=N/A,1,2,51770007,-1
and then Test, Status, it included the exponent in the count and the odds (as if it was a first test, as given in the worktodo line).

PRP or LL worker windows don't indicate what sort of calculation is being done in the worker title bar updates. Starting a PRP run, it brieflly flashes PRP there and is then quickly overwritten. The window contents will scroll out of view and then out of the buffer.

wantkitteh · 2019-01-13, 23:39

Seem to be having an issue with a specific FFT size when running the AVX routine in the torture test - 35K has an immediate rounding error. This doesn't occur on v29.4b8, although that does use a different Mersenne number to test. See attached image, v29.4b8 is on the left with local.txt open below showing manual instruction set config, v29.5b6 is on the right, instruction set config was performed through the torture test interface by checking the "Disable AVX2" box. CPU is 7600K, so no AVX-512 support anyway.

Prime95 · 2019-01-14, 00:38

Quote:

Originally Posted by wantkitteh

Seem to be having an issue with a specific FFT size when running the AVX routine in the torture test - 35K has an immediate rounding error.

Did you run the "fixed" version posted here (#152) https://www.mersenneforum.org/showpo...&postcount=152

kladner · 2019-01-14, 05:16

I have yet to see any problems with build 7.

Falkentyne · 2019-01-15, 21:21

Quote:

Originally Posted by Prime95

The new torture test feature to run a weaker test had a bug. It was selecting exponents that were too big for the FFT size. A new Windows executable for you to test:

https://www.dropbox.com/s/sc4ib5v4f4...ime95.zip?dl=0

Seems to be working great. Thank you.
I'll let you know if I run into any other problems. I only use basic stress testing though.
i9 9900K user.

Prime95 · 2019-01-21, 04:05

29.5 build 8

1) Fixed problems reading gwnum.txt.
2) Torture test dialog box rewritten.
3) Gwnum add and subtract are now multithreaded for AVX-512. Thus, faster multithread ECM and P-1.
4) Benchmark results are now written to results.bench.txt.
5) JSON results are now available for all work performed. JSON results are
written to results.json.txt.
6) The benchmarking hangs are fixed (fingers crossed).
7) Modular inverses in ECM now use GMP -- a small performance boost.

This is likely the last 29.5 build. My plan is for the next release to be 29.6 -- a release candidate. I have a few more bugs / features to investigate / add.

Linux 64-bit: ftp://mersenne.org/gimps/p95v295b8.linux64.tar.gz
Windows 64-bit: ftp://mersenne.org/gimps/p95v295b8.win64.zip

GP2 · 2019-01-21, 05:23

Quote:

Originally Posted by Prime95

29.5 build 8

For smallish exponents it doesn't print out res2048.

I only tested for b=10 so far. For example, for n=613 it only prints out res64; for n=641 it prints both res64 and res2048.

Code:

{"status":"C", "k":1, "b":10, "n":613, "c":-1, "known-factors":"9", "worktype":"PRP-3", "res64":"B4258816CA6260F2", "residue-type":1, "fft-length":128, "error-code":"00000000", "security-code":"04CA04CA", "program":{"name":"Prime95", "version":"29.5", "build":8, "port":8}, "timestamp":"2019-01-21 04:58:08", ...}
{"status":"C", "k":1, "b":10, "n":641, "c":-1, "known-factors":"9", "worktype":"PRP-3", "res64":"C9DAD4B06AD50DC4", "residue-type":1, "res2048":"...", "fft-length":128, "error-code":"00000000", "security-code":"05020502", "program":{"name":"Prime95", "version":"29.5", "build":8, "port":8}, "timestamp":"2019-01-21 04:58:08", ...}

ATH · 2019-01-21, 13:07

1 Round off error during 29.5b6 I noticed now when I was changing to 29.5b8:

[Work thread Jan 13 02:29:21] Resuming Gerbicz error-checking PRP test of M88479649 using AVX-512 FFT length 4704K, Pass1=1344, Pass2=3584, clm=1

[Work thread Jan 17 17:51:17] Iteration: 54083596/88479649, Possible error: round off (0.4366392078) > 0.42188

2019-01-13, 18:42	#158
kriesel "TF79LL86GIMPS96gpu17" Mar 2017 US midwest 2×5×503 Posts	In prime95 29.5b7 for Win64, I tried the following in worktodo PRP=N/A,1,2,51770007,-1,73,1 It resulted in the start of a P-1 factoring run. Correcting the form to PRP=N/A,1,2,51770007,-1 and then Test, Status, it included the exponent in the count and the odds (as if it was a first test, as given in the worktodo line). PRP or LL worker windows don't indicate what sort of calculation is being done in the worker title bar updates. Starting a PRP run, it brieflly flashes PRP there and is then quickly overwritten. The window contents will scroll out of view and then out of the buffer. Attached Thumbnails

2019-01-13, 23:39	#159
wantkitteh Jan 2019 1 Posts	Seem to be having an issue with a specific FFT size when running the AVX routine in the torture test - 35K has an immediate rounding error. This doesn't occur on v29.4b8, although that does use a different Mersenne number to test. See attached image, v29.4b8 is on the left with local.txt open below showing manual instruction set config, v29.5b6 is on the right, instruction set config was performed through the torture test interface by checking the "Disable AVX2" box. CPU is 7600K, so no AVX-512 support anyway. Attached Thumbnails

2019-01-12, 16:48	#156
kladner "Kieren" Jul 2011 In My Own Galaxy! 2·3·1,693 Posts	I have not had a hang with the new executable. I ran more 2560K benchmarks, since current work is using 2688K. All hangups, to my knowledge, have been with 2560K. I am now running benches from 2048K to 2560K to see if it hangs over a wider range of FFTs.

2019-01-14, 05:16	#161
kladner "Kieren" Jul 2011 In My Own Galaxy! 2·3·1,693 Posts	I have yet to see any problems with build 7.

2019-01-21, 04:05	#163
Prime95 P90 years forever! Aug 2002 Yeehaw, FL 7,411 Posts	29.5 build 8 1) Fixed problems reading gwnum.txt. 2) Torture test dialog box rewritten. 3) Gwnum add and subtract are now multithreaded for AVX-512. Thus, faster multithread ECM and P-1. 4) Benchmark results are now written to results.bench.txt. 5) JSON results are now available for all work performed. JSON results are written to results.json.txt. 6) The benchmarking hangs are fixed (fingers crossed). 7) Modular inverses in ECM now use GMP -- a small performance boost. This is likely the last 29.5 build. My plan is for the next release to be 29.6 -- a release candidate. I have a few more bugs / features to investigate / add. Linux 64-bit: ftp://mersenne.org/gimps/p95v295b8.linux64.tar.gz Windows 64-bit: ftp://mersenne.org/gimps/p95v295b8.win64.zip

2019-01-21, 13:07	#165
ATH Einyen Dec 2003 Denmark 17·181 Posts	1 Round off error during 29.5b6 I noticed now when I was changing to 29.5b8: [Work thread Jan 13 02:29:21] Resuming Gerbicz error-checking PRP test of M88479649 using AVX-512 FFT length 4704K, Pass1=1344, Pass2=3584, clm=1 [Work thread Jan 17 17:51:17] Iteration: 54083596/88479649, Possible error: round off (0.4366392078) > 0.42188