Currently three major operations in CP2K support CUDA-acceleration:

• Anything that uses dbcsr_multiply, i.e. sparse matrix multiplication, when compiled with -D__ACC -D__DBCSR_ACC. This benefits in particular the linear scaling DFT code. See also the DBCSR project.
• FFTs, when compiled with -D__PW_CUDA.
• If linked against an accelerated scalapack/blas library (in particular pdgemm/pdsyrk/dgemm) that executes these calls on the GPU. The impact of this is most visible for MP2 and RPA calculations. On the hybrid Cray XC30 linking against libsci_acc makes this happen.

To enable all CUDA acceleration options the following lines have to be added to the ARCH-file:

NVCC    = /path_to_cuda/bin/nvcc
DFLAGS += -D__ACC -D__DBCSR_ACC -D__PW_CUDA
LIBS   += -lcudart -lcublas -lcufft -lrt

As a prerequisite the Nvidia CUDA Toolkit has to be installed.

The acceleration of DBCSR is performed by libcusmm. This library provides a number of kernels. Each of these kernels can multiply blocks of specific blocksizes. The blocksizes of a simulation are determined by the employed basis-set. By default libcusmm is complied with about 200 common kernels. However, if an exotic basis set is used the particular blocksizes might be missing. This can be seen from the DBCSR Statistics, which is printed at the end of every CP2K-run.

In the following example the kernel for 13x13x15 was missing:

 -------------------------------------------------------------------------------
 -                                                                             -
 -                                DBCSR STATISTICS                             -
 -                                                                             -
 -------------------------------------------------------------------------------
 COUNTER                                      CPU                  ACC      ACC%
 number of processed stacks                   160                   64      28.6
 matmuls inhomo. stacks                     11880                    0       0.0
 matmuls total                             132360                53530      28.8
 flops  13 x   13 x   13                        0             33218640     100.0
 flops  13 x   13 x   15                 34810620                    0       0.0  <-- kernel missing
 flops  24 x   13 x   13                        0             55177824     100.0
...
 flops total                           1452705420            657928368      31.2
 marketing flops                       2048000000
 -------------------------------------------------------------------------------

There are over 2300 readily optimized kernel-parameters available in src/dbcsr/libsmm_acc/libcusmm/parameters.txt. If the desired kernel is already listed in parameters.txt, it can be included in libcusmm by editing the file src/dbcsr/libsmm_acc/libcusmm/generate.py. Otherwise new kernel parameters have to be optimized, which this howto explains in detail.

If you are interested in profiling CP2K with nvprof have a look at these remarks .