ggnfs on amd64

fivemack · 2007-02-27, 22:37

Simply compiling the distributed ggnfs with a 64-bit compiler fails, because there are lots of snippets of 32-bit assembly in the code, and nightmares of

Downloading gmp and building it with
./configure --build=i686-linux --prefix=/home/fivemack/maths/gmp-4.2.1-32bit
make
make check
make install

and then changing the ggnfs makefiles so that they use $(CC) rather than gcc as the compiler throughout, and link to $(LIBGMP) defined as /home/fivemack/maths/gmp-4.2.1-32bit/lib/libgmp.a rather than the system -lgmp, followed by defining CC as gcc -m32, produced binaries which at least appear to sieve ...

Has someone made ggnfs 64-bit-friendly? I suppose making lasieve other than 64-bit-hostile would be a significant piece of work, and the rest of the binaries are a marginal part of the calculation ... recoding blanczos64 as blanczos128 using SSE intrinsics would probably be worthwhile, but again that's a large chunk of assembler to grok and reabsorb.

frmky · 2007-02-28, 00:13

I'm not sure about the last source snapshot as it's quite old now, but the current CVS source, even the lattice sieve, is 64-bit friendly. It's a simple
make x86_64 or make nocona as appropriate. Although I would recommend using 32-bit lattice sieve binaries as they are about 20% faster than the 64-bit versions. The 64-bit matsolve and sqrt binaries are much faster as expected.

2007-02-27, 22:37	#1
fivemack (loop (#_fork)) Feb 2006 Cambridge, England 1913₁₆ Posts	ggnfs on amd64 Simply compiling the distributed ggnfs with a 64-bit compiler fails, because there are lots of snippets of 32-bit assembly in the code, and nightmares of Downloading gmp and building it with ./configure --build=i686-linux --prefix=/home/fivemack/maths/gmp-4.2.1-32bit make make check make install and then changing the ggnfs makefiles so that they use $(CC) rather than gcc as the compiler throughout, and link to $(LIBGMP) defined as /home/fivemack/maths/gmp-4.2.1-32bit/lib/libgmp.a rather than the system -lgmp, followed by defining CC as gcc -m32, produced binaries which at least appear to sieve ... Has someone made ggnfs 64-bit-friendly? I suppose making lasieve other than 64-bit-hostile would be a significant piece of work, and the rest of the binaries are a marginal part of the calculation ... recoding blanczos64 as blanczos128 using SSE intrinsics would probably be worthwhile, but again that's a large chunk of assembler to grok and reabsorb.

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
AMD64 on Solaris	Kyle	Software	9	2012-11-26 13:27
prime95 for amd64 on GNU/Linux OR alternative?	colo	Software	3	2008-02-28 13:19
AMD64 opcode timings	Prime95	Software	16	2005-03-04 17:48
AMD64 optimized version	LoKI.GuZ	Software	1	2005-01-02 06:28
llr on AMD64 ?	irzyxel	3*2^n-1 Search	4	2004-05-11 07:38

2007-02-28, 00:13	#2
frmky Jul 2003 So Cal 2×3⁴×13 Posts	I'm not sure about the last source snapshot as it's quite old now, but the current CVS source, even the lattice sieve, is 64-bit friendly. It's a simple make x86_64 or make nocona as appropriate. Although I would recommend using 32-bit lattice sieve binaries as they are about 20% faster than the 64-bit versions. The 64-bit matsolve and sqrt binaries are much faster as expected.