Sr2sieve on PPC/Linux

[Joe O]

Originally Posted by jasong

Sorry, I should have clarified. It was a bad link. File not found. And I'm looking for the Linux 64-bit AMD one. I THINK SSE2, but I'm not sure.

AFAIK there was never a 64 bit version published.

By the way, the link works for me.http://81.233.34.242/proth_sieve.php

proth_sieve v0.42 windows regular

proth_sieve v0.42 windows cmov

proth_sieve v0.42 windows SSE2

proth_sieve v0.42 linux regular

proth_sieve v0.42 linux cmov

proth_sieve v0.42 linux SSE2

proth_sieve v0.42 linux regular statically linked

proth_sieve v0.42 linux cmov statically linked

proth_sieve v0.42 linux SSE2 statically linked

proth_sieve v0.42 bsd regular

proth_sieve v0.42 bsd cmov

proth_sieve v0.42 bsd regular statically linked

proth_sieve v0.42 bsd cmov statically linked


Older versions...

[tnerual]

Originally Posted by Joe O

AFAIK there was never a 64 bit version published.

By the way, the link works for me.http://81.233.34.242/proth_sieve.php

but the link for linux app don't work.

Jasong: download the "prothsieve 0.42 version, all versions" it's the only way to get the linux apps.

[jasong]

Originally Posted by tnerual

but the link for linux app don't work.

Jasong: download the "prothsieve 0.42 version, all versions" it's the only way to get the linux apps.

The answer is late coming(my answer, this one), but thanks. :)

[geoff]

Originally Posted by jasong

p=1000000032774601, 441084 p/sec, 0 factors, 0.66% done, ETA 30 Dec 00:11

Thanks for that, could you also also test the 32-bit i686 binary on the same machine and same range for comparison?

[jasong]

For base-2, sieving at 1,000,000 billion on Linux, sr2sieve 32-bit 398,800 p/sec.

Question: Would you be willing to work on it for the Riesel Sieve project as well. It may or may not benefit from the same optimizations.

[geoff]

Originally Posted by jasong

For base-2, sieving at 1,000,000 billion on Linux, sr2sieve 32-bit 398,800 p/sec.

That is interesting, the difference between 32 and 64 bit is not as great as it was for the base 5 sieve.

Question: Would you be willing to work on it for the Riesel Sieve project as well. It may or may not benefit from the same optimizations.

It already does, just run `sr2sieve -r' to sieve from riesel.dat, or `sr2sieve -rs' to sieve riesel.dat and SoB.dat together :-) [probably not a good idea unless both have the same n range].

[geoff]

In sr2sieve 1.4.18 I have incorporated rogue's improved ppc64 mulmod, and I have made corresponding changes to the inlined functions. I don't know how this affects the relative speed of inline vs non-inline, or the different inline variants.

In Makefile, there is a line that adds -DUSE_INLINE_MULMOD=<X> to the CPPFLAGS. The options are:
X=0: Use the externally linked code.
X=1,2,3: Use inline variant 1, 2 or 3. 1 is default. I think variant 2 should be a little faster in principle, provided the compiler does the right thing. Variant 3 might help the compiler do the right thing.

[BlisteringSheep]

geoff & rogue, you guys rock. I haven't completed testing with all variants yet, but glancing at the results I've got about a 15% improvement with SoB.dat. I'm also testing with riesel.dat. FYI, here is the matrix of the tests I've got:

Platforms:

• 2.0 GHz PowerMac, Gentoo
  • gcc-3.4.6
  • gcc-4.0.0
  • gcc-4.0.3
  • gcc-4.1.1
• 2.5 GHz 970MP, RedHat AS4
  • gcc-3.4.3
  • gcc-4.0.0
  • gcc-4.1.1

On each platform/compiler combination, I'm testing the following:

• HASH_MAX_DENSITY
  • 0.32
  • 0.65
• USE_INLINE_MULMOD
  • 0
  • 1
  • 2
  • 3

I'll post full results tonight.

[geoff]

Originally Posted by BlisteringSheep

On each platform/compiler combination, I'm testing the following:

• HASH_MAX_DENSITY

I forgot to mention that in version 1.4.18 the hashtable size is automatically doubled if the L1 cache is large enough (as determined by L1_CACHE_SHIFT), so it should be OK to leave HASH_MAX_DENSITY at 0.65.

[BlisteringSheep]

Finally, just a summary over all of the combinations listed above:

• 2.0 GHz G5 (Apple PowerMac)
  • gcc 4.1.1, DENSITY 0.32, INLINE 2, SoB.dat, 303294 p/sec
• 2.2 GHz 970FX (IBM JS20)
  • gcc 4.1.1, DENSITY 0.32, INLINE 2, SoB.dat, 205017 p/sec
  • gcc 4.1.1, DENSITY 0.65, INLINE 2, riesel.dat, 204986 p/sec
• 2.5 GHz 970MP (IBM JS21)
  • gcc 4.1.1, DENSITY 0.32, INLINE 2, SoB.dat, 380233 p/sec
  • gcc 4.1.1, DENSITY 0.32, INLINE 2, riesel.dat, 234616 p/sec

The JS20 numbers with the SoB.dat don't look right. I will re-run that set. I haven't run with riesel.dat on my PowerMac yet, but will soon.

[BlisteringSheep]

Updated with riesel.dat for the G5 and corrected SoB.dat for the JS20.

• 2.0 GHz G5 (Apple PowerMac)
  • gcc 4.1.1, DENSITY 0.32, INLINE 2, riesel.dat, 187283 p/sec
• 2.2 GHz 970FX (IBM JS20)
  • gcc 4.1.1, DENSITY 0.65, INLINE 2, SoB.dat, 331802 p/sec

I've attached a zip file of the sr2work.txt and checkpoint.txt along with the data collected.