Folding Forum

When I try to run it with CUDA I get an error: OpenMMCUDA.dll: 126

proteneer wrote:Ah yes, we've never tested much of OpenMM on IONs - it doesn't seem to support atomics.

According to CUDA-Z, mine has only CC 1.1, so support for atomics - or much of anything really - is indeed limited. At least the GT430 has CC 2.1.

Spongebob25 wrote:When I try to run it with CUDA I get an error: OpenMMCUDA.dll: 126

You need to install CUDA 5.0 most likely.

As long as we're beta testing, it sounds line folks are finding reasonable things to report. If you eventually plan to release this to the general public, there probably needs to be some better error messages for the novice. Napoleon's could be used as an example: maybe something like "Your ION has support for Compute capability 1.1; Minimum requirement is X.X"

The same sort of thing might belong in the FAHCores. I've heard that this may have been a problem in V6 with -forcegpu whereas V7 seems to do a pretty good job of partitioning the assignments.

I installed CUDA 5.0 and Microsoft Visual Studio 9.0. I get this error:

Error launching CUDA compiler: -1
nvcc : fatal error : Cannot find compiler 'cl.exe' in PATH

What do I need to fix this?

Find out where your vcvarsall.bat file is located (like C:\Program Files (x86)\Microsoft Visual Studio 9.0\VC) and run it in the console before running the benchmark. That should set up the environment variables (like PATH) so that cl.exe can be executed.

I did as you requested and get the same error. How do I set it manually?

Hmm, now I'm stumped. The batch file actually checks/sets a whole lot of variables, tweaking path alone isn't sufficient. How did you get the OpenCL results earlier on?

BTW, the environment variable modifications done by the batch file are temporary. They only apply to the console where you run the batch file in, and are lost as soon as you exit the Command Prompt. Anyway, can you see just the cl.exe outputting the help text if you try something along these lines?

Updated the DL. Now supports older GPUs slightly better (not sure about IONs still), but the compute_21 error should go away.

Make sure your PATH variable in ENV points to: C:\Program Files (x86)\Microsoft Visual Studio 9.0\VC\bin

I installed the CUDA 5 toolkit, and I have VS 10 (the 2010 version) and I used that.
The program produces the error:

Hi - you need to use the 2008 version (MSVC9)

Jesse_V wrote:

proteneer wrote:Hi - you need to use the 2008 version (MSVC9)

OK, I guess I discovered that the 2010 version doesn't work! After installing MS-VS 9 and changing my PATH variable, the program completed successfully and produced

Code: Select all

C:\Users\Admin\Downloads\bench>FAHBench.exe -deviceId 0 -platform CUDA -precisio
n single

                                          O              O
   P R O T E N E E R     C--N              \              \               N
                         |                  C              C=O           / \-C
                         C                 /               |          N-C     \
  .C-C                 C/                  C               C           |      C
 /    \          O     |                   |               /           N      |
C     C          |     |           O       C              C                 /-C
 \_N_/ \   N    _C_    C           |      /         O    /                 C
        C-/ \_C/   \N-/ \    N   /-C-\   C          |    |           O    /
        |     |           C-/ \C/     N-/ \_   N\  /C\  -C      N    |    |
        O     |           |    |            \C/  C/   N/  \_C__/ \   C-\  C
              C           O    |             |   |          |     C-/   N/ \-C
               \_C             C             O   |          O     |          |
                  \             \-O              C                C          O
                  |                               \                \
                  C    N         Folding@Home      C--N             C
                   \   |      Benchmark  (Beta)    |                |
                    N--C                           O                |
                        \        Yutong Zhao                       C=O
                         N    proteneer@gmail.com                 /
                                                                 O

               for official stats, please visit www.fahbench.com

Explicit:
Checking for accuracy...done
17.9127 ns/day
Implicit:
Checking for accuracy...done
67.6904 ns/day

C:\Users\Admin\Downloads\bench>

from my Nvidia GTX 560TI. If I could make two suggestions: 1) reduce the prerequisites and 2) add more output to the console. It sat on "Explicit:" for a good while, with 0% GPU usage, and I thought I had done something wrong.

Doing the math, 1.5 milliseconds (folding timescale for the NTL9 protein) divided by 67.7 nanoseconds gives 22,157 days, which is 60 years.

Haha - those numbers actually aren't too bad.

1) will be hard - we're working with/waiting on NVIDIA to add in some features
2) most of the wait time comes from checking for accuracy. maybe I can add some stuff in later.

Flaschie wrote:Managed to make it run fine on my PC after second download (first one got some DLL-error).

Anyways, a massively 2.93 ns/day Explicit and 20.26 ns/day Implicit for single precision on my AMD 7660D @ stock (OpenCL, naturally)

One question though; Is this bench related to the bench Anandtech used in his GPU test lately? I noticed his numbers were much higher.

Anandtech is using a much older version. We have contacted them about the new benchmark. They will (hopefully) be migrating to our new benchmark in the future.

Jesse_V wrote:

proteneer wrote:1) will be hard - we're working with/waiting on NVIDIA to add in some features

Good luck. Please forgive my ignorance, but could you please explain how this .exe is different than the FahCores which don't require all this software?

proteneer wrote:2) most of the wait time comes from checking for accuracy. maybe I can add some stuff in later.

All I'm saying is that you could try to put some additional std:cout's in to indicate that the program is making progress.
[/code]

Thanks for the numbers, they're useful to know. Seems like most of the people here got the benchmark to work.

The core will be OpenCL only (and no it won't be hackable to load in CUDA ) mainly for reasons you've witnessed above in people trying to get the CUDA version to work. It's way too much work to enforce people to DL CudaSDK and to get people to download the right version of visual studio.

I have started a new thread where people can post their results:

Might keep this thread clean for debugging/feedback purposes.

viewtopic.php?f=38&t=23440