That would be wierd. Currently configure defaults to sm_10 for single precision, and sm_13 for double precision. [I've verified this again] can you send configure.log [without the cuda-arch option] - to petsc-maint? Satish On Mon, 25 Apr 2011, Евгений Козлов wrote:

Yes, this was the problem. Now it works.

Thank you.

2011/4/25 Victor Minden <[email protected]>:

...

Eugene, Based off of  Configure options --prefix=/home/kukushkinav --with-blas-lapack-dir=/opt/intel/composerxe-2011.0.084/mkl --with-mpi-dir=/opt/intel/impi/4.0.1.007/intel64/bin --with-cuda=1 --with-cusp=1 --with-thrust=1 --with-thrust-dir=/home/kukushkinav/include --with-cusp-dir=/home/kukushkinav/include It looks like maybe you're not setting the configure flag --with-cuda-arch=XXX.  Without this, PETSc uses the nvcc default, which when I last looked was sm_10.  The problem with this is that sm_10 doesn't support double-precision.  Maybe this is your problem?  For example, I use -with-cuda-arch=sm_13, which corresponds to NVIDIA compute capability 1.3 Cheers, Victor --- Victor L. Minden

Tufts University School of Engineering Class of 2012

On Fri, Apr 22, 2011 at 12:22 PM, Евгений Козлов <[email protected]> wrote:

...

Hello,

I am interesting in using PETSc for iterative solving sparse linear system on multi-GPU systems.

First of all, I compiled PETSc-dev and tried to run some examples, which will run on GPU.

I found src/snes/examples/tutorials/ex47cu.cu. It was compiled and executed.

Out of original program src/snes/examples/tutorials/ex47cu.cu:

[0]PETSC ERROR: --------------------- Error Message ------------------------------------ [0]PETSC ERROR: Floating point exception! [0]PETSC ERROR: User provided compute function generated a Not-a-Number! [0]PETSC ERROR: ------------------------------------------------------------------------ [0]PETSC ERROR: Petsc Development HG revision: d3e10315d68b1dd5481adb2889c7d354880da362  HG Date: Wed Apr 20 21:03:56 2011 -0500 [0]PETSC ERROR: See docs/changes/index.html for recent updates. [0]PETSC ERROR: See docs/faq.html for hints about trouble shooting. [0]PETSC ERROR: See docs/index.html for manual pages. [0]PETSC ERROR: ------------------------------------------------------------------------ [0]PETSC ERROR: ex47cu on a arch-linu named cn03 by kukushkinav Fri Apr 22 18:33:32 2011 [0]PETSC ERROR: Libraries linked from /home/kukushkinav/lib [0]PETSC ERROR: Configure run at Thu Apr 21 19:18:22 2011 [0]PETSC ERROR: Configure options --prefix=/home/kukushkinav --with-blas-lapack-dir=/opt/intel/composerxe-2011.0.084/mkl --with-mpi-dir=/opt/intel/impi/4.0.1.007/intel64/bin --with-cuda=1 --with-cusp=1 --with-thrust=1 --with-thrust-dir=/home/kukushkinav/include --with-cusp-dir=/home/kukushkinav/include [0]PETSC ERROR: ------------------------------------------------------------------------ [0]PETSC ERROR: SNESSolve_LS() line 167 in src/snes/impls/ls/ls.c [0]PETSC ERROR: SNESSolve() line 2407 in src/snes/interface/snes.c [0]PETSC ERROR: main() line 38 in src/snes/examples/tutorials/ex47cu.cu application called MPI_Abort(MPI_COMM_WORLD, 72) - process 0

Then I tried to find the place in source with the problem, I changed the function in struct ApplyStencil to

void operator()(Tuple t) { thrust::get<0>(t) = 1; }

Result: [0]PETSC ERROR: --------------------- Error Message ------------------------------------ [0]PETSC ERROR: Floating point exception! [0]PETSC ERROR: Infinite or not-a-number generated in mdot, entry 0! [0]PETSC ERROR: ------------------------------------------------------------------------ [0]PETSC ERROR: Petsc Development HG revision: d3e10315d68b1dd5481adb2889c7d354880da362  HG Date: Wed Apr 20 21:03:56 2011 -0500 [0]PETSC ERROR: See docs/changes/index.html for recent updates. [0]PETSC ERROR: See docs/faq.html for hints about trouble shooting. [0]PETSC ERROR: See docs/index.html for manual pages. [0]PETSC ERROR: ------------------------------------------------------------------------ [0]PETSC ERROR: ex47cu on a arch-linu named cn11 by kukushkinav Fri Apr 22 18:58:04 2011 [0]PETSC ERROR: Libraries linked from /home/kukushkinav/lib [0]PETSC ERROR: Configure run at Thu Apr 21 19:18:22 2011 [0]PETSC ERROR: Configure options --prefix=/home/kukushkinav --with-blas-lapack-dir=/opt/intel/composerxe-2011.0.084/mkl --with-mpi-dir=/opt/intel/impi/4.0.1.007/intel64/bin --with-cuda=1 --with-cusp=1 --with-thrust=1 --with-thrust-dir=/home/kukushkinav/include --with-cusp-dir=/home/kukushkinav/include [0]PETSC ERROR: ------------------------------------------------------------------------ [0]PETSC ERROR: VecMDot() line 1146 in src/vec/vec/interface/rvector.c [0]PETSC ERROR: KSPGMRESClassicalGramSchmidtOrthogonalization() line 66 in src/ksp/ksp/impls/gmres/borthog2.c [0]PETSC ERROR: GMREScycle() line 161 in src/ksp/ksp/impls/gmres/gmres.c [0]PETSC ERROR: KSPSolve_GMRES() line 244 in src/ksp/ksp/impls/gmres/gmres.c [0]PETSC ERROR: KSPSolve() line 426 in src/ksp/ksp/interface/itfunc.c [0]PETSC ERROR: SNES_KSPSolve() line 3107 in src/snes/interface/snes.c [0]PETSC ERROR: SNESSolve_LS() line 190 in src/snes/impls/ls/ls.c [0]PETSC ERROR: SNESSolve() line 2407 in src/snes/interface/snes.cС уважением, Евгений [0]PETSC ERROR: main() line 38 in src/snes/examples/tutorials/ex47cu.cu

RedHat 5.5, Cuda 3.2

Question1: Is it my problem or a bug of the algorithm?

Question2: Where can I find simple doc or example, which describe how to solve sparse linear systems on multi-GPU systems using PETSc?

-- Best regards, Eugene