ABINIT on GPU

[fabienfontaine]

Hi everybody,
Does anywone tried to compile ABINIT on CUDA-GPU ?
Which routines are GPU oriented ?
Thx

Fabien Fontaine-Vive
Institut de Chimie de Nice

[jbeuken]

Hello,

To make you wait, we can read in the release notes (abstract) of 6.12.1 version ( not yet released )

B. Most noticeable achievements

B.1 The Ground-state calculations can benefit from GPUs, with CUDA library.
See the input variable use_gpu_cuda, as well as the tests contained in the tests/gpu
subdirectory.
This functionnality is still in beta stage and has to be used carefully.
Note that to take benefit of the GPU efficiency the computation load has to be large enough.
Abinit has to be compiled with the NVidia Cuda package (v3.2+) as well as the Magma library (http://icl.cs.utk.edu/magma/).
Relevant configure options: --enable-gpu, --with-gpu-flavor, --with-gpu-prefix, NVCC, NVCC_FLAGS
and --with-linalg-flavor (for Magma activation).
When Abinit has been compiled with Cuda and an available Graphics card is detected, ground-states calculations
are automatically done on the GPU.
(F. Dahm and M. Torrent, with some testing by JM Beuken)

regards

jmb

[VasiaS]

Hello, then I use abinit with cuda for t11.in (for H2 from Examples) I get in log this: ERROR while fft work_gpu ==> fofr: CUFFT_EXEC_FAILED and the program does not complete the calculation. Abinit was successfully installed at options:
./configure --enable-mpi FC=mpif90 --enable-gpu --with-gpu-flavor=cuda-double --with-gpu-libs="-L/usr/local/cuda/lib64 -lcublas -lcufft -lcudart" --with-gpu-incs="-I/usr/local/cuda/include" NVCC="/usr/local/cuda/bin/nvcc" NVCC_CPPFLAGS="-DHAVE_CUDA_SDK" NVCC_CFLAGS="-O3 -arch=sm_13 -Xptxas=-v --use_fast_math --compiler-options -O3,-Wall,-fPIC,-g -I/home/vasia/NVIDIA_GPU_Computing_SDK/CUDALibraries/common/inc -I/usr/local/cuda/include -I${PWDloc}/../src/incs"
I tried yet with --with-fft-flavor="fftw3-mkl" and -arch=sm_20 but I get the same case. I have 560 gtx. Can You help me? please
ps: sorry for my english

[pouillon]

Cuda libraries are linked dynamically. Did you check that the LD_LIBRARY_PATH variable is properly set when running Abinit?

[VasiaS]

hmm... I tried this: export LD_LIBRARY_PATH=/usr/local/cuda/lib64 and export C_INCLUDE_PATH=/usr/local/cuda/include but it didn't help. Then I did ./configure... I get this:
checking cuda_runtime_api.h usability... yes
checking cuda_runtime_api.h presence... yes
checking for cuda_runtime_api.h... yes
checking cufft.h usability... yes
checking cufft.h presence... yes
checking for cufft.h... yes
checking cublas.h usability... yes
checking cublas.h presence... yes
checking for cublas.h... yes
checking whether Cuda programs can be compiled... yes
checking whether we have Cuda < 4... no
checking for the actual GPU support... cuda-double

I don't know much Linux and I do not know how else to check it.
I think the cause of my problem is the input file because input file from abinit/tests/gpu/input/t01.in works well.
Thank You for your help

[crischawo]

I would be very interested in some benchmarks vs. CPU - if You have the time to do some.
Chr.

[VasiaS]

I have made the tests (built-in, cpu, fast and gpu). But I could not upload my files (43Mb) here (The extension is not allowed.). Сan You give me your e-mail?

[VasiaS]

files containing my tests, can be found here https://www.yousendit.com/download/T2dk ... bUJqQThUQw File expiration date: February 15, 2012

[crischawo]

Thank You very much, I will have a look at it now.
Chr

[VasiaS]

That You think about my errors in some tests?

[nicspalla]

Hi VasiaS,
can you re-upload your tests?

thanks
Nicola

[VasiaS]

https://www.yousendit.com/dl?phi_action ... 0200000000

[roshan]

Hi VasiaS,
can you re-upload your tests?

thanks
Roshan

[pouillon]

@Roshan: this thread dates back to 2012. The tests were likely removed a long time ago.