Re: [Pw_forum] QE-GPU problems.
Hello, Well, my card is actually GTX1060 but I didn't bought it especially for calculations. Just from my built-in curiosity, and I know the game is not worth the candle, I want try to run it anyway and close that 'chapter' of my life :) I will post here some tests results - as I was looking for this type of data, and maybe this will save some time for other seekers. Mr Spiga, as the best expert, could You please give some guide, any informations what and where to change in source of QE-GPU? K. Gruszka W dniu 02.01.2017 o 08:07, Filippo SPIGA pisze: On Dec 30, 2016, at 3:48 AM, Oliver Generalaowrote: The only catch, which Mr Spiga always points out, is that the DP performance on GTX cards is very slow , which is like 1/32 (or 1/24) of the Single Precision So GTX 1070 is not worth using. A little bit of tuning of your input, better compile flags and the CPU will perform the same or faster than the CPU+GPU. Because you have a GTX 1070 it does mean you must use it. If you buy a GPU only for QE, do not buy that model. Very simple. -- Filippo SPIGA ~ Quantum ESPRESSO Foundation ~ http://www.quantum-espresso.org ___ Pw_forum mailing list Pw_forum@pwscf.org http://pwscf.org/mailman/listinfo/pw_forum ___ Pw_forum mailing list Pw_forum@pwscf.org http://pwscf.org/mailman/listinfo/pw_forum
Re: [Pw_forum] QE-GPU problems.
On Dec 30, 2016, at 3:48 AM, Oliver Generalaowrote: > The only catch, which Mr Spiga always points out, is that the DP performance > on GTX cards is very slow , which is like 1/32 (or 1/24) of the Single > Precision So GTX 1070 is not worth using. A little bit of tuning of your input, better compile flags and the CPU will perform the same or faster than the CPU+GPU. Because you have a GTX 1070 it does mean you must use it. If you buy a GPU only for QE, do not buy that model. Very simple. -- Filippo SPIGA ~ Quantum ESPRESSO Foundation ~ http://www.quantum-espresso.org ___ Pw_forum mailing list Pw_forum@pwscf.org http://pwscf.org/mailman/listinfo/pw_forum
Re: [Pw_forum] QE-GPU problems.
Hello Oliver, Many thanks for quick response. According to Your post it seems that You know well what to change and where to do those changes. Could You please provide more informations? Maybe something more specific? Konrad W dniu 30.12.2016 o 04:48, Oliver Generalao pisze: > Hi Konrad, > > Nvidia GTX 1060 is sm_61 and per https://github.com/fspiga/QE-GPU, it > supports up to sm_60, based on my experience QE-GPU will compile with > sm_61 however when you run pw-gpu.x it will throw errors, hence there > should be a slight modification on the QE-GPU source code to include > sm_61, it should be easy to do that. > I am using Nvidia GTX 1070 and have been working on it with QE-GPU for > 2 months now. The only catch, which Mr Spiga always points out, is > that the DP performance on GTX cards is very slow , which is like > 1/32 (or 1/24) of the Single Precision > > > > On 12/29/16, Konrad Gruszkawrote: >> Deer community, >> >> Recently I'm trying to compile and run QE-GPU version on my new Cuda >> capable card. Unfortunately after many attempts the result is poor. >> >> I have Nvidia GTX 1060 (with Pascal architecture, 1280 cuda units). I've >> managed to compile GPU PWSCF with sm_60 (pascal) support but when trying >> to run any calculation I get: >> >> Program received signal SIGFPE: Floating-point exception - erroneous >> arithmetic operation.Backtrace for this error: #0 0x7F84D59C4E08 >> >> #1 0x7F84D59C3F90 >> #2 0x7F84D4CB74AF >> #3 0x63B2C5 in newd_cuda_ >> #4 0x639B9F in newq_compute_gpu_ at newq_compute_gpu.f90:122 >> #5 0x597B1F in __dfunct_MOD_newd at newd.f90:262 >> #6 0x4E2B87 in init_run_ at init_run.f90:101 >> #7 0x4081DB in run_pwscf_ at run_pwscf.f90:78 >> #8 0x408049 in pwscf at pwscf.f90:30 #9 0x7F84D4CA282F >> >> The configure options were: >> >> ./configure --enable-parallel --enable-cuda --enable-openmp >> --with-cuda-dir=/usr/local/cuda-8.0/ --with-internal-blas >> --with-internal-lapack --without-magma --with-phigemm >> FFT_LIBS=/mnt/fast/fftw-3.3.4/.libs/libfftw3.a --with-gpu-arch=pascal >> >> For now I'm trapped here, not knowing what to do. Is it possible to run >> QE-GPU not only on specialized computing devices like e.g. K20 at all? >> How to manage this? >> >> Konrad Gruszka >> >> > -- dr inż. Konrad Gruszka Politechnika Częstochowska Instytut Fizyki ___ Pw_forum mailing list Pw_forum@pwscf.org http://pwscf.org/mailman/listinfo/pw_forum
Re: [Pw_forum] QE-GPU problems.
Hi Konrad, Nvidia GTX 1060 is sm_61 and per https://github.com/fspiga/QE-GPU, it supports up to sm_60, based on my experience QE-GPU will compile with sm_61 however when you run pw-gpu.x it will throw errors, hence there should be a slight modification on the QE-GPU source code to include sm_61, it should be easy to do that. I am using Nvidia GTX 1070 and have been working on it with QE-GPU for 2 months now. The only catch, which Mr Spiga always points out, is that the DP performance on GTX cards is very slow , which is like 1/32 (or 1/24) of the Single Precision On 12/29/16, Konrad Gruszkawrote: > Deer community, > > Recently I'm trying to compile and run QE-GPU version on my new Cuda > capable card. Unfortunately after many attempts the result is poor. > > I have Nvidia GTX 1060 (with Pascal architecture, 1280 cuda units). I've > managed to compile GPU PWSCF with sm_60 (pascal) support but when trying > to run any calculation I get: > > Program received signal SIGFPE: Floating-point exception - erroneous > arithmetic operation.Backtrace for this error: #0 0x7F84D59C4E08 > > #1 0x7F84D59C3F90 > #2 0x7F84D4CB74AF > #3 0x63B2C5 in newd_cuda_ > #4 0x639B9F in newq_compute_gpu_ at newq_compute_gpu.f90:122 > #5 0x597B1F in __dfunct_MOD_newd at newd.f90:262 > #6 0x4E2B87 in init_run_ at init_run.f90:101 > #7 0x4081DB in run_pwscf_ at run_pwscf.f90:78 > #8 0x408049 in pwscf at pwscf.f90:30 #9 0x7F84D4CA282F > > The configure options were: > > ./configure --enable-parallel --enable-cuda --enable-openmp > --with-cuda-dir=/usr/local/cuda-8.0/ --with-internal-blas > --with-internal-lapack --without-magma --with-phigemm > FFT_LIBS=/mnt/fast/fftw-3.3.4/.libs/libfftw3.a --with-gpu-arch=pascal > > For now I'm trapped here, not knowing what to do. Is it possible to run > QE-GPU not only on specialized computing devices like e.g. K20 at all? > How to manage this? > > Konrad Gruszka > > -- Oliver B. Generalao M.S. Physics student Structure and Dynamics Group National Institue of Physics University of the Philippines Diliman, Quezon City Trunkline: +63-2-981-8500 Mobile: +63-927-4033966 ___ Pw_forum mailing list Pw_forum@pwscf.org http://pwscf.org/mailman/listinfo/pw_forum
Re: [Pw_forum] QE-GPU problems.
Hi Konrad, My experience with QE-GPU is that it works only with double precision supported GPUs but not single precision. I am using TitanZ and Intel compiler, it works fast. Here is my ./configure ./configure CC=icc F90=ifort F77=ifort MPIF90=mpiifort --enable-parallel --disable-openmp --without-scalapack --enable-cuda --with-gpu-arch=sm_35 --with-cuda-dir=/usr/local/cuda-6.5 --with-magma --with-phigemm --with-pinned-mem LDFLAGS="-L/usr/lib64/ -lstdc++" I doubt if GTX 1060 supports DP? If it is SP than it does not work for QE-GPU or it may be very slow. Rolly On 12/29/2016 03:56 PM, Konrad Gruszka wrote: Deer community, Recently I'm trying to compile and run QE-GPU version on my new Cuda capable card. Unfortunately after many attempts the result is poor. I have Nvidia GTX 1060 (with Pascal architecture, 1280 cuda units). I've managed to compile GPU PWSCF with sm_60 (pascal) support but when trying to run any calculation I get: Program received signal SIGFPE: Floating-point exception - erroneous arithmetic operation.Backtrace for this error: #0 0x7F84D59C4E08 #1 0x7F84D59C3F90 #2 0x7F84D4CB74AF #3 0x63B2C5 in newd_cuda_ #4 0x639B9F in newq_compute_gpu_ at newq_compute_gpu.f90:122 #5 0x597B1F in __dfunct_MOD_newd at newd.f90:262 #6 0x4E2B87 in init_run_ at init_run.f90:101 #7 0x4081DB in run_pwscf_ at run_pwscf.f90:78 #8 0x408049 in pwscf at pwscf.f90:30 #9 0x7F84D4CA282F The configure options were: ./configure --enable-parallel --enable-cuda --enable-openmp --with-cuda-dir=/usr/local/cuda-8.0/ --with-internal-blas --with-internal-lapack --without-magma --with-phigemm FFT_LIBS=/mnt/fast/fftw-3.3.4/.libs/libfftw3.a --with-gpu-arch=pascal For now I'm trapped here, not knowing what to do. Is it possible to run QE-GPU not only on specialized computing devices like e.g. K20 at all? How to manage this? Konrad Gruszka ___ Pw_forum mailing list Pw_forum@pwscf.org http://pwscf.org/mailman/listinfo/pw_forum -- PhD. Research Fellow, Dept. of Physics & Materials Science, City University of Hong Kong Tel: +852 3442 4000 Fax: +852 3442 0538 ___ Pw_forum mailing list Pw_forum@pwscf.org http://pwscf.org/mailman/listinfo/pw_forum