Failed to Load GPU Sorting Engine (sort_engine_sm20.dll)
 

Hello Everyone,

I've followed the Beginners Guide to NFS.. things work with my i7, but I really want to use cuda. I have msieve152_svn942_win64_cuda.zip setup; however, when I run msieve w/ CUDA (either alone or from factmsieve.py) I receive the following:

.\msieve.exe -np1 -nps -t 8
[COLOR="DarkRed"]cannot load library 'sort_engine_sm20.dll', error 126
error: failed to load GPU sorting engine[/COLOR]

I have a GTX 660, i7, CUDA 6.0 ToolKit installed...

I also tried msieve150_win64_cuda.zip and ran with msieve.gpu.exe and it looks like it uses my GPU and runs through poly selection fast, but also states "crap" after each number, and the dat.p file doesnt grow.

I'm also interested in getting my VC10 environment setup to properly build any of the msieve branches from SVN as well... I'm running into dependency hell (ecmlib/libecm? maybe gmp?).. any great guides that anyone knows would be appreciated.

Thanks!

Here's a compilation of the latest MSieve SVN using CUDA 5.5. It won't currently build using CUDA 6.0. This latest version also has some bug fixes that JasonP made to address issues that were found with larger datasets.

[QUOTE=wombatman;380920]Here's a compilation of the latest MSieve SVN using CUDA 5.5. It won't currently build using CUDA 6.0. This latest version also has some bug fixes that JasonP made to address issues that were found with larger datasets.[/QUOTE]
It builds for me with CUDA 6.5 and I don't recall problems with CUDA 6.0. Do you remember what wouldn't build?

Brian

It built with 6.0, but when I actually ran it, I got the following:

[CODE][c:\dorkdom\svn_checkout\msieve-code\trunk\b40c\b40c\radix_sort\../radix_sort/tex_ref.cuh, 71] cudaBindTexture failed (CUDA error 20: invalid channel descriptor)
[C:/Dorkdom/SVN_Checkout/msieve-code/trunk/b40c/sort_engine.cu, 95] sort engine: (CUDA error 20: invalid channel descriptor)[/CODE]

Edit: And now I see that wrote "wouldn't build" above.:bangheadonwall: So, yeah, it'll build absolutely fine, but won't run at all using CUDA 6.0. I'll try with CUDA 6.5 and see if that's still the case for me.


Yeah, same issue with CUDA 6.5. Builds successfully (including the sort engine dll) but gives this error when actually running it:

[CODE][c:\dorkdom\svn_checkout\msieve-code\trunk\b40c\b40c\radix_sort\../radix_sort/tex_ref.cuh, 71] cudaBindTexture failed (CUDA error 20: invalid channel descriptor)
[C:/Dorkdom/SVN_Checkout/msieve-code/trunk/b40c/sort_engine.cu, 95] sort engine: (CUDA error 20: invalid channel descriptor)[/CODE]

Could I bug someone to compile a version of msieve with cuda 5.5 and mpir compiled for core 2. I get the above bug when compiling myself with a later cuda version. My version of VS2012 is completely messed up with header files missing etc so I can't use an earlier version of cuda than 6.5 myself.

It looks like msieve needs to move to [url]http://nvlabs.github.io/cub/[/url] which replaces b40c.

CUB is really nice and has support for more modern GPUs. Maybe someday...