Early Beta of version 24.11

[TheJudger]

Originally Posted by Prime95

I'm sure the binutils guys will make objcopy work eventually if they haven't done so already. The source is available if someone wants to try a 64-bit linux port.

So if binutils provite support for MASM -> ELF64 than you would give it a try? :)

Originally Posted by Prime95

My next task is more optimizations, especially making use of the extra SSE2 registers in 64-bit mode. Don't expect much - a few percent on AMD64, perhaps a little more on the P4.

Optimizations for ll-tests and/or factoring?

[Peter Nelson]

From what I have read, I believe the latest version of objcopy in binutils DOES have the ability to convert to ELF64 format.

Someone could try to do this from the source provided above.

Obviously this would not include the security module for primenet comms, but if successful, it would not be too much work for George to make an official build using the same process.

[Prime95]

Originally Posted by Peter Nelson

d) please would it be possible to include a short test of trial factoring speed in the benchmark (of the release version) because this would be very useful to know and compare the benefit of your optimisations.

Done. Look for it next time I upload an executable

[rainchill]

Hi, I have an Athlon64 3000+ Hp laptop, with 1Gb of ram. Here are my benchmark results with 23.8
Prime95 version 23.8, RdtscTiming=1
Best time for 384K FFT length: 26.055 ms.
Best time for 448K FFT length: 31.393 ms.
Best time for 512K FFT length: 35.744 ms.
Best time for 640K FFT length: 45.166 ms.
Best time for 768K FFT length: 54.970 ms.
Best time for 896K FFT length: 66.214 ms.
Best time for 1024K FFT length: 74.506 ms.
Best time for 1280K FFT length: 100.525 ms.
Best time for 1536K FFT length: 122.588 ms.
Best time for 1792K FFT length: 147.037 ms.
Best time for 2048K FFT length: 165.960 ms.

And here are my results with 24.11
Prime95 version 24.11, RdtscTiming=1
Best time for 512K FFT length: 32.064 ms.
Best time for 640K FFT length: 38.795 ms.
Best time for 768K FFT length: 46.873 ms.
Best time for 896K FFT length: 56.537 ms.
Best time for 1024K FFT length: 63.192 ms.
Best time for 1280K FFT length: 82.394 ms.
Best time for 1536K FFT length: 102.145 ms.
Best time for 1792K FFT length: 122.718 ms.
Best time for 2048K FFT length: 136.854 ms.

So it seems like a good little speed bump.

[Prime95]

New 24.11 executables are available. From whatsnew.txt:

5) Added timeouts to PrimeNet communications in hopes of avoiding rare hangs
when contacting the PrimeNet server.
6) Fixed rare bug where P-1's GCD could miss a factor.
7) Added trial factoring to the benchmark.
8) Fixed bug in ECM when using zero-padded FFTs.

[TheJudger]

9) crashes immediately on CPUs supporting sse2 when "advance/time" exponent is 78000000 :(

Happens with windows-client aswell as the linux-client...
Timing exponents works fine in older versions up to 79.xxxM

With "CPUSupportsSSE2=0" in local.ini timing exponents up to 79.xxxM works fine

Your choice: 8

Exponent to time (10000000): 78000000
Number of Iterations (10):

Accept the answers above? (Y):
Floating point exception

[Xyzzy]

Originally Posted by TheJudger

9) crashes immediately on CPUs supporting sse2 when "advance/time" exponent is 78000000 :(

http://www.mersenneforum.org/showpos...42&postcount=4

[TheJudger]

xyzzy: I know... but older versions (23.5, 23.9) automaticly switches to x87-code when leaving sse2-range... 24.11 just crashes...

[Dresdenboy]

I posted many prime95 results (incl. 64bit TF numbers vs. 32bit) in the benchmark thread.

[Dresdenboy]

Trial factoring speedup in 64bit mode:

24.11    24.11
32 bit   64 bit
  (ms)    (ms)  Speedup
 6,010   3,786    58,7%
 6,032   3,782    59,5%
 6,003   4,032    48,9%
 6,029   4,140    45,6%
10,961   4,843   126,3%
10,963   5,643    94,3%
13,932   6,816   104,4%
13,836   8,007    72,8%
13,849   7,951    74,2%
13,831   7,934    74,3%

[ET_]

Originally Posted by Dresdenboy

Trial factoring speedup in 64bit mode:

24.11    24.11
32 bit   64 bit
  (ms)    (ms)  Speedup
 6,010   3,786    58,7%
 6,032   3,782    59,5%
 6,003   4,032    48,9%
 6,029   4,140    45,6%
10,961   4,843   126,3%
10,963   5,643    94,3%
13,932   6,816   104,4%
13,836   8,007    72,8%
13,849   7,951    74,2%
13,831   7,934    74,3%

cOOOOl!

Luigi