gpu & user-cuda packages problem

LAMMPS LAMMPS Mailing List Mirror
1

Dear lammps-users,

I’ve encountered difficulties with building LAMMPS (31Aug2011) with gpu support:

  1. Package USER-CUDA fails to install with multiple errors in fft3d_cuda.cpp (please see the attached file “user-cuda errors”)
  2. Package GPU installs successfully, but when I go to examples/gpu only in.melt_x.x.gpu runs successfully (indeed showing improved productivuty). Rhodo and phosphate examples stops at the line where kspace_style is defined (whichever precision I choose) with ERROR: Invalid kspace style (please see the attached file lammps.log)

Both libgpu.a and liblammpscuda.a compiled successfully. I attach makefiles as well.

My system is:
Intel Core 2 duo + nvidia gtx 480
OpenSuSe 11.4 x86_64
LAMMPS was built using FFTW 3.3, openMPI 1.4.3, gnu compilers v. 4.4.6
Tried both CUDA Toolkit 4.0 and 3.2 (but with the same driver 270.41.19 - from 4.0 version) all SDK samples are working.
nvcc was invoced with parameter --compiler-bindir “path to the folder where gcc-4.4.6 is installed” (as the default ones 4.5.1 are not supported by current CUDA version)

I attach

If anybody knows the solution to the problem - I’d like to know it two.

Kind regards,

Sergey

user-cuda errors (5.39 KB)

log.lammps (741 Bytes)

Makefile.gfct (2.41 KB)

[Makefile.common (user-cuda)|attachment](upload://2Q0HDj1eGSQXSx5TmGBIP30qzul.common (us) (3.05 KB)

[Makefile.lammps (user-cuda)|attachment](upload://frDVQ5iyXoPkIjxxXhvcCFl1q9b.lammps (us) (348 Bytes)

[Makefile.fermi (gpu)|attachment](upload://bKkbo1hkTVftkgWjet8hhigmOC2.fermi (gpu) (1.88 KB)

[Makefile.lammps (gpu)|attachment](upload://u4BASC7BcI9l1ocBrQySUU1DFKa.lammps (gp) (170 Bytes)

2

Dear lammps-users,

I've encountered difficulties with building LAMMPS (31Aug2011) with gpu
support:

1. Package USER-CUDA fails to install with multiple errors in fft3d_cuda.cpp
(please see the attached file "user-cuda errors")

2. Package GPU installs successfully, but when I go to examples/gpu only
in.melt_x.x.gpu runs successfully (indeed showing improved productivuty).
Rhodo and phosphate examples stops at the line where kspace_style is defined
(whichever precision I choose) with ERROR: Invalid kspace style (please see
the attached file lammps.log)

as for 2.) the examples have not been updated to match the documentation.
the /<precision> postfix to pppm/gpu is no longer supported. you
have to use plain pppm/gpu. the precision is implicitly determined
by the precision used for the FFT library. i.e. when compiling with
-DFFT_SINGLE you get a (mostly, depending on how you compiled
the GPU library) single precision pppm/gpu and otherwise a
double precision pppm/gpu.

as for 1.) this may be related to the same reason as 2.)

cheers,
    axel.

3

Hi Sergey

where were some errors introduced at a recent change of the make system regarding the USER-CUDA package. I just finished to fix these bugs and are sending the changes to Steve so that he can include those in LAMMPS.

Best regards
Christian

-------- Original-Nachricht --------