SkylakeX teasers (aka prime95 29.5) - Page 10

Chuck · 2018-12-09, 22:56

Quote:

Originally Posted by Prime95

This is a link to the Windows executable I built for Chuck:

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

It contains some of the new swizzle optimizations discussed in another thread that seemed to work OK.

I'll be available for email and forum conversations, but unable to upload or download big files.

Good results continue with this version. I successfully finished a PRP test and started another with no error messages.

ATH · 2018-12-16, 18:06

In build 5 I had a PRP test at 88.22M where it immediately chose 4704K FFT. It had SoftCrossoverAdjust=-0.004 in prime.txt

Then another instance had a PRP test at 88.26M where it started to test 4608K FFT and when that had the average roundoff error 0.30835 it chose 4800K FFT.

Granted that instance had SoftCrossover=1.0 and SoftCrossoverAdjust=-0.008 in prime.txt, but why did it not test 4704K FFT?

I now removed all SoftCrossover and SoftCrossoverAdjust from all instances, it was a leftover from when I added it to build 3 when I got so many roundoff > 0.4 errors.

ATH · 2018-12-16, 22:47

There might be a problem with choosing the best FFT parameters after restarting mprime, see here:
https://www.mersenneforum.org/showpo...&postcount=133

Prime95 · 2019-01-03, 21:37

29.5 build 6.

1) Fixed testing tiny numbers with AVX-512 FFTs, numbers less than about 5120 bits are forced to use FMA FFTs.
2) Torture test dialog box allows picking AVX-512, AVX2, AVX, SSE2 FFTs.
3) Changed FFT crossovers. ATH, let me know if this is better.
4) Eliminated soft crossovers. Instead, I hope to get the crossovers right in the gwnum code. Soft crossovers can be re-enabled (see undoc.txt).
5) Eliminated a memory leak when a benchmark is interrupted.
6) In linux a pid file is created (mprime.pid).
7) New swizzle code is used. FFTs should be about 0% faster. Well, maybe small FFTs that run in the L2 cache will see a tiny speed bump.

Not fixed:
1) The hangs reported in benchmarking. Please try again with this release. I cannot get it to happen in Linux. Is this a Windows-only issue? Does this only happen benchmarking multi-threaded FFTs? Is CPU usage 0% at the time of the hang? Hangs are usually due to a deadlock. Benchmarks use locks to sync the start of all the worker threads. Multi-threaded FFTs also use locks to coordinate threads.

Linux 64-bit: ftp://mersenne.org/gimps/p95v295b6.linux64.tar.gz
Windows 64-bit: ftp://mersenne.org/gimps/p95v295b6.win64.zip

GP2 · 2019-01-04, 01:27

Quote:

Originally Posted by Prime95

29.5 build 6.

Wait, does this build include the Chuck fix? You didn't mention anything about that (saving registers).

Quote:

Originally Posted by Prime95

I have an idea. Some routines were not saving and restoring xmm8 - xmm15 as per the Windows 64 ABI. If I can build an executable for you before I go on a cruise, you can test out that theory.

Quote:

Originally Posted by Prime95

I was wrong, xmm8-xmm15 arrived with SSE2 in 2004. Saving the registers fixed Chuck's problem.

Quote:

Originally Posted by Prime95

This is a link to the Windows executable I built for Chuck:

Prime95 · 2019-01-04, 01:47

Quote:

Originally Posted by GP2

Wait, does this build include the Chuck fix? You didn't mention anything about that (saving registers).

Yes, that is included.

LaurV · 2019-01-04, 02:24

Mark Rose · 2019-01-04, 03:24

Quote:

Originally Posted by Prime95

Not fixed:
1) The hangs reported in benchmarking. Please try again with this release. I cannot get it to happen in Linux. Is this a Windows-only issue? Does this only happen benchmarking multi-threaded FFTs? Is CPU usage 0% at the time of the hang? Hangs are usually due to a deadlock. Benchmarks use locks to sync the start of all the worker threads. Multi-threaded FFTs also use locks to coordinate threads.

I did have an mprime hang with build 5, during a 8000K benchmark. It was on a 48 core dual socket Epyc system. I was testing combinations of workers and threads. I seem to remember it happening during a several workers, several threads test. It wasn't during the single worker test.

kriesel · 2019-01-04, 10:00

Quote:

Originally Posted by Prime95

29.5 build 6.

1) Fixed testing tiny numbers with AVX-512 FFTs, numbers less than about 5120 bits are forced to use FMA FFTs.
2) Torture test dialog box allows picking AVX-512, AVX2, AVX, SSE2 FFTs.
3) Changed FFT crossovers. ATH, let me know if this is better.
4) Eliminated soft crossovers. Instead, I hope to get the crossovers right in the gwnum code. Soft crossovers can be re-enabled (see undoc.txt).
5) Eliminated a memory leak when a benchmark is interrupted.
6) In linux a pid file is created (mprime.pid).
7) New swizzle code is used. FFTs should be about 0% faster. Well, maybe small FFTs that run in the L2 cache will see a tiny speed bump.

Not fixed:
1) The hangs reported in benchmarking. Please try again with this release. I cannot get it to happen in Linux. Is this a Windows-only issue? Does this only happen benchmarking multi-threaded FFTs? Is CPU usage 0% at the time of the hang? Hangs are usually due to a deadlock. Benchmarks use locks to sync the start of all the worker threads. Multi-threaded FFTs also use locks to coordinate threads.

Linux 64-bit: ftp://mersenne.org/gimps/p95v295b6.linux64.tar.gz
Windows 64-bit: ftp://mersenne.org/gimps/p95v295b6.win64.zip

Benchmark tries on v29.5b6 on Windows X on i7-8750H;
set allowed memory to 8192M.
Set throughput benchmark to do 2048 - 32768K, 1-3,6 workers, hyperthreaded too.
First try, stalls early in 2048K. At this point there was an mfakto instance on the UHD630 IGP and mfaktc instance on the GTX1050Ti gpu. Halt mfakto on UHD630 for second try.
Another benchmark stall early in 2048k. Halt mfaktc and retry. Application crash. Retry; benchmark stall again. Try #4: nearly finished 2560K before stalling, 8 minutes is longest of 4 tries before stall or crash. Try 5: no hyperthreading, made it to 23040K before stalling. Try 6 starting from 23040K reached 32768K.

kladner · 2019-01-04, 11:38

Quote:

Not fixed:
1) The hangs reported in benchmarking. Please try again with this release. I cannot get it to happen in Linux. Is this a Windows-only issue? Does this only happen benchmarking multi-threaded FFTs? Is CPU usage 0% at the time of the hang? Hangs are usually due to a deadlock. Benchmarks use locks to sync the start of all the worker threads. Multi-threaded FFTs also use locks to coordinate threads.

I tried one additional time with 295b5. It hung fairly quickly. As before, it was on a 2560K FFT.
i7-6700K, 4300 MHz, 16 GiB RAM, 3200 MHz, Win 7 Pro

I will try the latest version when I am able to check on it sooner than after work (~10 hours.)

Prime95 · 2019-01-04, 16:06

Aaargh! Why can't I reproduce the hang! Makes debugging much harder.

For those that can reproduce it quickly, does a normal torture test or daily work ever hang? If you add "TortureTestThreads=<# of logical cores>" to prime.txt and then run a torture test with just one worker, does that hang?

In the meantime, I'll re-examine the bench code.

2018-12-16, 22:47	#102
ATH Einyen Dec 2003 Denmark 2×1,579 Posts	There might be a problem with choosing the best FFT parameters after restarting mprime, see here: https://www.mersenneforum.org/showpo...&postcount=133

2019-01-03, 21:37	#103
Prime95 P90 years forever! Aug 2002 Yeehaw, FL 2·3,767 Posts	29.5 build 6. 1) Fixed testing tiny numbers with AVX-512 FFTs, numbers less than about 5120 bits are forced to use FMA FFTs. 2) Torture test dialog box allows picking AVX-512, AVX2, AVX, SSE2 FFTs. 3) Changed FFT crossovers. ATH, let me know if this is better. 4) Eliminated soft crossovers. Instead, I hope to get the crossovers right in the gwnum code. Soft crossovers can be re-enabled (see undoc.txt). 5) Eliminated a memory leak when a benchmark is interrupted. 6) In linux a pid file is created (mprime.pid). 7) New swizzle code is used. FFTs should be about 0% faster. Well, maybe small FFTs that run in the L2 cache will see a tiny speed bump. Not fixed: 1) The hangs reported in benchmarking. Please try again with this release. I cannot get it to happen in Linux. Is this a Windows-only issue? Does this only happen benchmarking multi-threaded FFTs? Is CPU usage 0% at the time of the hang? Hangs are usually due to a deadlock. Benchmarks use locks to sync the start of all the worker threads. Multi-threaded FFTs also use locks to coordinate threads. Linux 64-bit: ftp://mersenne.org/gimps/p95v295b6.linux64.tar.gz Windows 64-bit: ftp://mersenne.org/gimps/p95v295b6.win64.zip

2019-01-04, 02:24	#106
LaurV Romulan Interpreter Jun 2011 Thailand 25B5₁₆ Posts

2019-01-04, 16:06	#110
Prime95 P90 years forever! Aug 2002 Yeehaw, FL 7534₁₀ Posts	Aaargh! Why can't I reproduce the hang! Makes debugging much harder. For those that can reproduce it quickly, does a normal torture test or daily work ever hang? If you add "TortureTestThreads=<# of logical cores>" to prime.txt and then run a torture test with just one worker, does that hang? In the meantime, I'll re-examine the bench code.