===================== GPU support in Abinit =====================
IMPORTANT: GPU support is currently highly EXPERIMENTAL and should be used by experienced developers only. If you encounter any problem, please report it to Yann Pouillon before doing anything else. In particular, DO NOT TRY TO WILDLY AND DIRTILY HACK THE BUILD SYSTEM, EVEN IF YOU ARE A PHYSICIST!
GPU support is activated by the –enable-gpu option of configure. Another option of importance is the –with-gpu-flavor one, which selects the kind of GPU support that will be activated. A convenience option, codename –with-gpu-prefix, is also provided, in order to set automatically all relevant parameters whenever possible. A few other options are available as well, mainly for fine-tuning of the build parameters and testing purposes.
Full descriptions of all these options can be found in the ~abinit/doc/config/build-config.ac file. Do not hesitate to ask questions on https://forum.abinit.org/.
In addition, the permitted GPU-related preprocessiong options are:
- HAVE_GPU : generic use;
- HAVE_GPU_SERIAL : serial GPU support;
- HAVE_GPU_MPI : MPI-aware GPU support.
At present it is possible to ask for single- or double-precision Cuda support. The configure script will check that the Cuda libraries are properly working, but however not whether double-precision is actually supported by your version of Cuda (this might be added in the future).
All calls to Cuda routines should be carefully embedded within ‘#if defined HAVE_GPU_CUDA … #else … #endif’ preprocessing blocks. When a feature does require Cuda and will not work without it, the corresponding ‘#else’ part should display an error and cause Abinit to abort.
The permitted Cuda-related preprocessing options are :
- HAVE_GPU_CUDA : generic use;
- HAVE_GPU_CUDA_SP : single-precision calculations;
- HAVE_GPU_CUDA_DP : double-precision calculations.
All high-level routines directly accessing Cuda features have to be put in ~abinit/src/52_manage_cuda/, and low-level ones in ~abinit/src/15_gpu_toolbox/. All exceptions have to be approved by Xavier Gonze prior to any implementation.
All files belonging to nVidia must not be distributed with Abinit. Please discuss with Yann Pouillon if you need them inside the Abinit source tree during the build.
In any case, all Cuda-related developments should be done in good coordination with:
- Marco Mancini
- Marc Torrent
- Thierry Deutsch
- Damien Caliste
- Luigi Genovese
- Matthieu Ospici
- Yann Pouillon
To take advantage of the multiple FFT in cuda (FFT in batch), ABINIT have to be compiled with a Cuda version>=3.0. It is possible to build with previous vesions (>2.1 tested) but you make some changes. Contact MMancini. cuda implementation support devices with capabilty (revision)>1.0
The MAGMA project aims to develop a dense linear algebra library similar to LAPACK but for heterogeneous/hybrid architectures, starting with current “Multicore+GPU” systems. It is recommended to take advantage of MAGMA when using ABINIT with Cuda. Magma is not distributed within ABINIT package; it has to be preliminary installed. To activate MAGMA support during building process, use –wih-linalg-flavor=”…+magma” at configure level.
OpenCL support is currently under discussion. More info will come once decisions have been taken.
The S_GPU library provides higher performance and better load balancing when each GPU of a hybrid computer is shared by several processes, e.g. MPI tasks.
It will be supported in Abinit in the future, from its version 2.
See http://ligforge.imag.fr/projects/sgpu/ for details.