Help needed compiling GMP-ECM

[M0CZY]

I have a fresh, new installation of Ubuntu 8.10, and I have copies of gmp-4.2.4.tar.gz and ecm-6.2.1.tar.gz.
I don't know what, if any, other software I will need.
I need detailed instructions for compiling GMP-ECM, if anyone will help please?

[Jeff Gilchrist]

With Ubuntu you probably already have GMP installed. Just follow the detailed instructions that come with the ECM tar ball in the "INSTALL" file. It is pretty straight forward 5 step process.

[fivemack]

sudo apt-get install gmp-ecm gets you 6.1.2 without any effort.

[akruppa]

M0CZY,

which cpu does you computer have? There is no optimized assembly for x86_64 cpus in the official GMP 4.2.4, but patches exist that add at least some assembly for the most common functions for AMD64 and Core 2. If you want the fastest possible GMP-ECM, you'll need to hunt down bits and pieces of code and put together an optimized GMP on your own... it's a bit messy.

Alex

[henryzz]

Originally Posted by akruppa

M0CZY,

which cpu does you computer have? There is no optimized assembly for x86_64 cpus in the official GMP 4.2.4, but patches exist that add at least some assembly for the most common functions for AMD64 and Core 2. If you want the fastest possible GMP-ECM, you'll need to hunt down bits and pieces of code and put together an optimized GMP on your own... it's a bit messy.

Alex

i didnt know that
i still managed to get 2x speed by running in 64-bit though

i had to install GMP to get GMP-ECM to compile on the 64-bit version of Ubuntu
after i installed GMP 4.2.4 the first time the ./configure for GMP-ECM said i had mismatching versions of gmp.h which it said was 4.2.2 and the other part which it said was 4.2.4
it took quite a bit of fiddling to get it to work

i installed:
sudo apt-get install libc6-dev g++ gcc
sudo apt-get install build-essentials
sudo apt-get install m4

[Jeff Gilchrist]

Originally Posted by henryzz

i didnt know that
i still managed to get 2x speed by running in 64-bit though

I was just going to say the same thing. In Windows when I compile I get a 2x speedup with the 64bit version.

With the official GMP 4.2.4 in the Windows projects there are specific builds for Core2 64bit and Opteron 64bit. Would those have the right bits in them then? Do you have any idea of the 4.3 release will include these optimizations by default?

[henryzz]

before i compiled GMP-ECM on 64-bit linux for myself i had heard of a possible 3x speed benefit
i suspect that might have been with the extra bits
running 64-bit GMP-ECM in a virtual machine with a 32-bit host means ~1.5 times 32-bit speed

[bsquared]

Originally Posted by Jeff Gilchrist

I was just going to say the same thing. In Windows when I compile I get a 2x speedup with the 64bit version.

With the official GMP 4.2.4 in the Windows projects there are specific builds for Core2 64bit and Opteron 64bit. Would those have the right bits in them then? Do you have any idea of the 4.3 release will include these optimizations by default?

This post links to the core2 patch for GMP. http://www.mersenneforum.org/showpos...4&postcount=82

Geoff also has posts with the resulting core2 (and opteron) optimized gmp-ecm binary in the thread, which I've found to be the fastest on my core2 systems.

[akruppa]

If you configure on a 64 bit system with --enable-asm-redc, then for numbers of up to 20 words (385 digits), at least stage 1 uses GMP-ECM's own assembly code, so except for addition etc which still uses GMP functions, there's not much speed to gain (other than by writing better mulredc code of course).

Stage 2 uses GMP functions for computing the roots of the polynomials, and if you run with -no-ntt, for the polynomial multiplication, too (otherwise it uses NTT code which relies on GMP only little).

So getting a well optimized GMP matters most if you want to factor somewhat larger numbers, where stage 1 and/or stage 2 do all the work via GMP.

Pierrick Gaudry has a patch for AMD64 assembly (http://www.loria.fr/~gaudry/mpn_AMD64/), Jason Worth Martin has a patch for Core 2 (http://www.math.jmu.edu/~martin/). I found that in some cases, the AMD64 code is faster even on Core 2, but haven't looked into the issue much.

GMP 4.3 will (finally!) have asm code for x86_64 architectures. It's supposed to be out last week or so... but Torbjörn has always valued "when it's ready" over "when I said I would."

Alex

[henryzz]

Originally Posted by akruppa

GMP 4.3 will (finally!) have asm code for x86_64 architectures. It's supposed to be out last week or so... but Torbjörn has always valued "when it's ready" over "when I said I would."

if that's the case then it probably isn't worth fiddling around with the patches but waiting for GMP 4.3

[Jeff Gilchrist]

Originally Posted by akruppa

GMP 4.3 will (finally!) have asm code for x86_64 architectures. It's supposed to be out last week or so... but Torbjörn has always valued "when it's ready" over "when I said I would."

Is that why the 4.3 preliminary benchmarks are showing such a nice speedup or is that from something else?