Prime95 30.7

[WraithX]

Hello Prime95, I just ran into a small issue when compiling 30.7b9 with MingW64 in MSYS. I just needed to make two small changes to the makemw64 file to correct these issues and get the gwnum.a library to build.

First, on line 29 I needed to add a line continuation character \.
Second, I needed to add a target to build radix.o.

Third, I needed to add "#include <sys/time.h>" to line 68 of gwutil.h to get rid of a warning during compile:

gcc -I.. -I../sqlite-amalgamation-3180000 -DX86_64 -DWINDOWS64 -O2 -o mw64/giants.o -c giants.c
In file included from giants.c:23:
gwutil.h:68:26: warning: 'struct timeval' declared inside parameter list will not be visible outside of this definition or declaration
 int gettimeofday (struct timeval *tp, void *tzp);
                          ^~~~~~~
gcc -I.. -I../sqlite-amalgamation-3180000 -DX86_64 -DWINDOWS64 -O2 -o mw64/radix.o -c radix.c
In file included from radix.c:17:
gwutil.h:68:26: warning: 'struct timeval' declared inside parameter list will not be visible outside of this definition or declaration
 int gettimeofday (struct timeval *tp, void *tzp);
                          ^~~~~~~

I have attached the updated makemw64.

[Prime95]

Originally Posted by WraithX

Third, I needed to add "#include <sys/time.h>" to line 68 of gwutil.h to get rid of a warning during compile

Which, of course, breaks the MSVC build. Is there a MINGW ifdef I can test?

[WraithX]

Originally Posted by Prime95

Which, of course, breaks the MSVC build. Is there a MINGW ifdef I can test?

Ah, sorry. Good catch! Yes, there are two you can use:

#define __MINGW32__ 1
#define __MINGW64__ 1

I'd recommend using the mingw64 target. If someone needs the mingw32 target, we can check at that time how to resolve any issues they have.

[S485122]

This is not relevant any more since there is no more GIMPS work with that FFT length, but it seems that 2940 KiB FFT's are slower than they should be.
That FFT size seems to be AVX512 specific, but shouldn't it be 2944 KiB instead ? (An odd multiple of 4096 while the "neighbouring" FFTs are multiples of 65536...)
Just curious.

[axn]

2944 = 2^7 * 23. Ignore the powers of 2. AFAIK, George hasn't implemented a 23 FFT.
2940 = 2^2 * 3 * 5 * 7^2, which looks better. Still... a multiple of 735 is unusual. Probably will be slower than expected. 3072 ought to be faster.

2916 & 2880 seems to be more "normal" smaller sizes.

[S485122]

Originally Posted by axn

2944 = 2^7 * 23. Ignore the powers of 2. AFAIK, George hasn't implemented a 23 FFT.
...

In other words I just showed my ignorance of how the FFT innards work. :-(

[Zhangrc]

Originally Posted by axn

2944 = 2^7 * 23. Ignore the powers of 2. AFAIK, George hasn't implemented a 23 FFT.
2940 = 2^2 * 3 * 5 * 7^2, which looks better. Still... a multiple of 735 is unusual. Probably will be slower than expected. 3072 ought to be faster.

2916 & 2880 seems to be more "normal" smaller sizes.

I wonder if there's a 5888k FFT (which handles 110M exponents).
(5888=2^8*23, which is an auspicious number in Chinese, because it sounds roghly like "I'm rich rich rich")

[kriesel]

Xeon Phi 7250 is a 68-core & x4 HT CPU with 8 channels MCDRAM (totaling 16GiB). A 7210 is similar but 64-core.
The motherboard Supermicro K1SPE has 6 DIMM slots. With all DIMM slots empty, prime95 shows activity of one thread per physical core in Task Manager, where the hyperthreads of a specific core appear in succession. Windows 10 Task Manager displays the correct numbers of physical and logical cores, correct cache amounts, etc.

When DIMMs are added (observed with 1 32 or 64GiB in slot A, 3 32GiB in A-C, 6 32GiB in A-F), Task Manager shows 234 physical cores, 272 logical for the 7250, or 208 physical, 256 logical for the 7210, and odd figures for cache amounts. Prime95 shows core numbers higher than are physically present per Intel spec.
With any DIMMs installed, Task Manager shows a different pattern of logical core use that appears to be using multiple hyperthreads in prime95 when intending not to. Iteration times are noticeably longer than before the DIMMs were added.

In Ubuntu 18.04 atop WSL1 on Win10, after the DIMMs are added, stream.c shows ~30% higher memory bandwidth.

[Prime95]

Run a prime95 benchmark (abort it). Results.bench.txt will contain hwloc's description of your machine.

[kriesel]

hwloc's portion looks ok to me.

[kriesel]

Hi,

On Win10 pro, Xeon Phi 7210, prime95 v30.7b9, PrimeNet connected, manual assignments, I'm seeing very disparate ETAs between status output and worker windows, that leads to months discrepancies for exponents currently in progress, and YEARS difference in ETAs for following work (which should all complete within ~1 year).
Also the CPU clock rate ("Speed") is misrepresented on my Computer Properties page, and I have not found a way to fix that. Should be 1.3 GHz, is 0.346 GHz. (3.76:1 ratio)

W1 332897017 ~39. d vs. ~204. d ~5.23 ratio;
W2 344587487 ~18.27 d vs 126. d ~6.90 ratio.

Help!