Msieve v1.47 feedback

[jasonp]

Just starting things up here. This version includes some tuning for Nvidia's latest graphics cards, so if you have a Fermi card and know how to run NFS polynomial selection, then please give it a try. The precompiled binary on sourceforge now has two sets of PTX files, and if you have a Fermi card and version 3.0 or later of Nvidia's CUDA toolkit, then use the ptx files in the 'fermi' subdirectory.

[em99010pepe]

Thank you for the new version.
About MPI code, can it be used in a local area network? I want to take advantage of all computers I have in my LAN. If it is a stupid question please say so.

[Jeff Gilchrist]

Originally Posted by em99010pepe

Thank you for the new version.
About MPI code, can it be used in a local area network? I want to take advantage of all computers I have in my LAN. If it is a stupid question please say so.

Yes, you can install MPI on all your LAN machines and spread out the work via your LAN.

[Jeff Gilchrist]

When I compile msieve with VS 2010 for 32bit Release mode using MPIR 2.1.2, it seems to work when doing QS. The 64bit version works fine and the pre-compiled 32bit from Jason. Can anyone else reproduce this with VS 32bit binary?

It crashes for me with this number:
432717521981854371284364875772580616114001895598914408712561227

Sun Sep 19 13:56:15 2010  
Sun Sep 19 13:56:15 2010  
Sun Sep 19 13:56:15 2010  Msieve v. 1.47
Sun Sep 19 13:56:15 2010  random seeds: 2d89fa40 345674ae
Sun Sep 19 13:56:15 2010  factoring 432717521981854371284364875772580616114001895598914408712561227 (63 digits)
Sun Sep 19 13:56:16 2010  searching for 15-digit factors
Sun Sep 19 13:56:16 2010  commencing quadratic sieve (63-digit input)
Sun Sep 19 13:56:16 2010  using multiplier of 43
Sun Sep 19 13:56:16 2010  using VC8 32kb sieve core
Sun Sep 19 13:56:16 2010  sieve interval: 10 blocks of size 32768
Sun Sep 19 13:56:16 2010  processing polynomials in batches of 21
Sun Sep 19 13:56:16 2010  using a sieve bound of 101281 (4788 primes)
Sun Sep 19 13:56:16 2010  using large prime bound of 5064050 (22 bits)
Sun Sep 19 13:56:16 2010  using trial factoring cutoff of 22 bits
Sun Sep 19 13:56:16 2010  polynomial 'A' values have 8 factors
Sun Sep 19 13:56:28 2010  5073 relations (2328 full + 2745 combined from 23261 partial), need 4884
Sun Sep 19 13:56:28 2010  begin with 25589 relations
Sun Sep 19 13:56:28 2010  reduce to 7450 relations in 2 passes
Sun Sep 19 13:56:28 2010  attempting to read 7450 relations
Sun Sep 19 13:56:28 2010  recovered 7450 relations
Sun Sep 19 13:56:28 2010  recovered 5951 polynomials
Sun Sep 19 13:56:28 2010  attempting to build 5073 cycles
Sun Sep 19 13:56:28 2010  found 5073 cycles in 1 passes
Sun Sep 19 13:56:28 2010  distribution of cycle lengths:
Sun Sep 19 13:56:28 2010     length 1 : 2328
Sun Sep 19 13:56:28 2010     length 2 : 2745
Sun Sep 19 13:56:28 2010  largest cycle: 2 relations
Sun Sep 19 13:56:28 2010  matrix is 4788 x 5073 (0.6 MB) with weight 139188 (27.44/col)
Sun Sep 19 13:56:28 2010  sparse part has weight 139188 (27.44/col)
Sun Sep 19 13:56:28 2010  filtering completed in 4 passes
Sun Sep 19 13:56:28 2010  matrix is 4427 x 4491 (0.5 MB) with weight 119972 (26.71/col)
Sun Sep 19 13:56:28 2010  sparse part has weight 119972 (26.71/col)
Sun Sep 19 13:56:28 2010  commencing Lanczos iteration
Sun Sep 19 13:56:28 2010  memory use: 0.6 MB

At the Lanczos stage it crashes with this error:
Unhandled exception at 0x0046b060 in msieve32.exe: 0xC0000005: Access violation reading location 0x00b4e000.

The actual violation happens in the last line of this code snippet:

#elif defined(MSC_ASM32A)
    i = 0;
    ASM_M
    {
        push    ebx
        mov    edi,y
        lea    ebx,c
        mov    esi,x
        mov    ecx,i
        align 16
    L0:    movq    mm0,[edi+ecx*8]

So at the L0: stage. Here is what the memory browser is showing for values:
L0 0x0046b060 L0 void *
ecx 15355 unsigned long
edi 11730984 unsigned long
i 0 unsigned int
mm0 10956266787831142436 unsigned __int64

Any ideas?

[jasonp]

Originally Posted by Jeff Gilchrist

Yes, you can install MPI on all your LAN machines and spread out the work via your LAN.

Note that while it's possible to use a collection of machines on a LAN to run an arbitrary MPI program, in practice it will help a great deal if all the machines are similar speed and the switch you use is as fast as possible. Don't even bother with MPI on a LAN for the LA if you don't have a gigabit switch.

Jeff, does the crash happen with v1.46 too? (I would predict yes). It's possible something in the LA is not initialized properly when compiling with MSVC. Actually, a small matrix like this would use the unpacked solver code, and the crash happens in the vector-vector multiply. An array index of 15355 is a lot bigger than the matrix dimension, so is something wrapping around?

[Jeff Gilchrist]

Originally Posted by jasonp

Jeff, does the crash happen with v1.46 too? (I would predict yes). It's possible something in the LA is not initialized properly when compiling with MSVC. Actually, a small matrix like this would use the unpacked solver code, and the crash happens in the vector-vector multiply. An array index of 15355 is a lot bigger than the matrix dimension, so is something wrapping around?

Yes it does crash in v1.46 as well when compiled with MSVC. Not sure about the wrapping stuff, is there something in particular I can check for you? Brian might be able to debug this more easily.

[Jeff Gilchrist]

New 64bit Windows binaries now available for 1.47 including both default settings and a second binary compiled with LARGEBLOCKS and TD=80

http://gilchrist.ca/jeff/factoring/

Jeff.

[Phil MjX]

Thanks Jeff,

Your binaries are always helpfull to me and I'll be happy to test the LARGEBLOCK version of your binaries with the c154 I am dealing with.

A question : I am very interested in the latest changes of the gpu enabled version of the polynomial selection, because of the introduction of a 12 multiplier instead of the "classical" 60 for the smallest coefficient.

This is the multiplier I used with ggnfs (amongst 144 and 720 to cover different search intervals).

My problem is that I am totally unable to compile gpu version of msieve with cygwin (I fear this isn't possible), and that I cannot test sources changes by myself...

I'd have a request :
Do you have the tools, and the time, to compile a 64 bits version of the gpu_enabled source ?

I currently do np1 with sourceforge 32 bits gpu'able binaries and plan to switch to np2 with yours...should it works, should the step 2 be faster this way ?

Thanks again for your work, and also a great thanks to Jason for improving msieve !

Regards,
Philippe

[Brian Gladman]

Originally Posted by Jeff Gilchrist

Yes it does crash in v1.46 as well when compiled with MSVC. Not sure about the wrapping stuff, is there something in particular I can check for you? Brian might be able to debug this more easily.

I can see this bug in win32, which I suspect is a compiler bug because the registers are not being loaded with the correct values.

Brian

[jasonp]

Philippe, I thought of you when jrk committed this patch :) It should be possible to compile the GPU version of msieve in cygwin, but you'd need to find a way to link to Nvidia's precompiled DLLs, and must have Microsoft's compiler to run Nvidia's compiler (the free MSVC Express is what I use). Note that poly selection really does not benefit from using a 64-bit binary, though optimizations are possible in the CPU branch that can significantly improve performance.

Oh, and you're welcome :)

[Brian Gladman]

Originally Posted by Brian Gladman

I can see this bug in win32, which I suspect is a compiler bug because the registers are not being loaded with the correct values.

Brian

Sadly, this was a bug in my inline assembler code, which I have now corrected this in the msieve SVN repository. I have also corrected a minor win32 build configuration issue.

Brian