[Mac OS X] n-body taking 100% of the CPU

[adinapoli]

Hi all,

probably this is related to #67, but if I run the accelerate-nbody I get an unresponsive graphical app which sucks up 100% of the CPU.

I believe is not even using the CUDA backend. Any thoughts/debugging methods/solutions?

[tmcdonell]

100% CPU indeeds sounds like you are using the interpreter backend. A couple of things spring to mind:

• when compiling the examples package, make sure you include the -fcuda flag to cabal. I have found that occasionally cabal silently decides to not use the CUDA backend even if accelerate-cuda is already installed.
• run with accelerate-nbody --cuda to force the cuda backend: you'll get an error if it is not available.

Let me know if either of those help.

[adinapoli]

Uhm, tried to reinstall the example, got this (inside accelerate-example folder):

cabal install -fcuda
Resolving dependencies...
cabal: Could not resolve dependencies:
trying: accelerate-examples-0.13.0.0
trying: accelerate-examples-0.13.0.0:+opencl

whilst if I try to run nbody with that flag I get:

accelerate-nbody --cuda
Error: unrecognized option `--cuda'
Run "accelerate-nbody --help" for usage information

Yesterday I've also tried this command, forcing the interpreter:

accelerate-noby --interpreter cuda

But apparently nothing changed, the program used 100% of the CPU nevertheless o.O

Cheers,
Alfredo

[tmcdonell]

Try cabal install -fcuda -f-opencl

I also pushed some patches that will display what backends are available and indicate which is currently being used, so it should be more obvious what is going on.

[adinapoli]

Using cabal configure -fgui -fcuda does succeed, but the outcome is the same. I'll pull your juicy patches and keep you posted, thanks :)

A.

[adinapoli]

Ok, interesting bits here. Running this fails (inside accelerate-examples):

cabal: Could not resolve dependencies:
trying: accelerate-examples-0.13.0.0
rejecting: bmp-1.2.4.1/installed-772... (conflict: accelerate-cuda =>
binary==0.5.1.1/installed-a36..., bmp => binary==0.6.4.0/installed-5bd...)
rejecting: bmp-1.2.4.1 (conflict: accelerate-cuda =>
binary==0.5.1.1/installed-a36..., bmp => binary==0.6.*)
trying: bmp-1.2.3.4
rejecting: accelerate-io-0.13.0.0/installed-fd4... (conflict: bmp==1.2.3.4,
accelerate-io => bmp==1.2.4.1/installed-772...)
rejecting: accelerate-io-0.12.1.0, 0.12.0.0 (conflict: accelerate-examples =>
accelerate-io>=0.13)

Usual run-of-the-mill cabal-hell errors I believe :) Tracking down the problem I've updated the accelerate-cuda pulling the latest and it does update binary but now I get:

Data/Array/Accelerate/CUDA/Analysis/Launch.hs:145:14:
    Not in scope: data constructor `Foreign'
Failed to install accelerate-cuda-0.13.0.0

..which solved itself updating the latest accelerate (sorry for the stream-of-thoughts-like, but I'm trying to document the process for whoever will have the same problem in the future :) )

Running a plain cabal install inside accelerate-examples succeed, but as we were suspecting nbody is using the sequential interpreter:

accelerate-nbody --interpreter cuda
accelerate-nbody (c) [2012..2013] The Accelerate Team

Usage: accelerate-nbody [OPTIONS]

Available backends:
  *  interpreter  reference implementation (sequential)

[adinapoli]

Ok, new updates:

• Installed latests version of accelerate, accelerate-cuda, accelerate-fft and accelerate-io
• Reinstalled accelerate-examples with cabal install -fgui -fcuda -f-opencl
• Run n-body with accelerate-nbody --cuda

Now finally the cuda implementation shows up, but...

accelerate-nbody (c) [2012..2013] The Accelerate Team

Usage: accelerate-nbody [OPTIONS]

Available backends:
     interpreter  reference implementation (sequential)
  *  cuda         implementation for NVIDIA GPUs (parallel)

accelerate-nbody:
*** Internal error in package accelerate-cuda ***
*** Please submit a bug report at https://github.com/AccelerateHS/accelerate/issues
./Data/Array/Accelerate/CUDA.hs:145 (unhandled): CUDA Exception: invalid argument

[tmcdonell]

AH! You have made progress!
This is an instance of #80. If you update to the latest CUDA package on hackage (0.5.0.1) it should (hopefully, finally) work.

[adinapoli]

Yes! Success!

Thans a lot :)
For who will come next, in case he might ever need it, this is the Mac OS X cuda lib directory, so this makes installing cuda a breeze. Posting here since I forgot about this path regularly :D

cabal install cuda --reinstall --force-reinstalls --extra-lib-dirs=/usr/local/cuda/lib/

A.

[tmcdonell]

Great! Thanks also for the excellent write-up. I've added a link here from the troubleshooting section of the wiki as others might find this useful.

-T