Errors building CP2K 2022.2 for GPU Inbox

[Luis Cebamanos]

Hello,

I am trying to build CP2K 2022.2 for NVIDIA GPUs. I've built  ELPA with GPU support (no OpenMP, yes MPI), and COSMA 2.5 with GPU support too. The error I am getting is as follows:

mpifort -c -D__MKL -D__FFTW3 -D__LIBXSMM  -D__LIBINT -D__parallel -D__SCALAPACK -D__HAS_ISO_C_BINDING -D__ACC -D__DBCSR_ACC -D__PW_CUDA -D__GRID_CUDA -D__OFFLOAD_PROFILING -D__HAS_PATCHED_CUFFT_70 -D__COSMA -D__ELPA_NVIDIA_GPU  -D__LIBVORI -D__SPGLIB -O3 -fopenmp -ffree-form -ffree-line-length-none -ftree-vectorize -funroll-loops -std=f2008 -I/opt/intel/compilers_and_libraries_2019.5.281/linux/mkl/include  -I/home/COSMA/include -I/home/libxsmm-1.17.0/include   -I/home/elpa-2021.11.002/include/elpa-2021.11.002/modules -I/home/elpa-2021.11.002/include/elpa-2021.11.002/elpa -I/home/elpa-2021.11.002/include/elpa-2021.11.002/src/helpers -I/home/fftw-3.3.10/include -I/home/libint-v2.6.0-cp2k-lmax-4/include -I/cuda/cuda-11.2/include -m64 -I/opt/intel/compilers_and_libraries_2019.5.281/linux/mkl/include -D__COMPILE_ARCH="\"nvidiaGPU\""  -D__COMPILE_REVISION="\"git:a95ec40\"" -D__DATA_DIR="\"/home/cp2k-2022.2/data\"" -D__SHORT_FILE__="\"xc/xc_input_constants.F\"" -I'/home/cp2k-2022.2/src/xc/' -I'/home/cp2k-2022.2/obj/nvidiaGPU/psmp/exts/dbcsr' xc_input_constants.F90

/home/cp2k-2022.2/tools/build_utils/fypp -n /home/cp2k-2022.2/src/shg_int/constants_operator.F constants_operator.F90
cmp: /home/cp2k-2022.2/obj/git-ref: No such file or directory
/home/cp2k-2022.2/src/pw/gpu/pw_gpu_kernels.h(22): error: identifier "offloadStream_t" is undefined

and here you can see my arch file:

NVCC     = ${CUDA_HOME}/bin/nvcc

GPUVER   = V100

CXX      = mpicxx
CC       = mpicc
FC       = mpifort
LD       = mpifort
AR       = ar -r

OFFLOAD_CC = nvcc
OFFLOAD_FLAGS = -O3 -g -w --std=c++11

OPTFLAGS += -fopenmp -O3 -march=native

DFLAGS   += -D__MKL -D__FFTW3 -D__LIBXSMM  -D__LIBINT \
  -D__parallel -D__SCALAPACK -D__HAS_ISO_C_BINDING
 
DFLAGS += -D__ACC -D__DBCSR_ACC -D__PW_CUDA -D__GRID_CUDA -D__OFFLOAD_PROFILING
DFLAGS += -D__HAS_PATCHED_CUFFT_70 -D__COSMA -D__ELPA_NVIDIA_GPU

DFLAGS+= -D__LIBVORI -D__SPGLIB

# C++ flags
CXXFLAGS = -O3 $(OPTFLAGS) -g -I${CUDA_HOME}/include -std=c++11

# NVIDIA
NVFLAGS  = $(DFLAGS) $(OPTFLAGS) -g -O3 -arch sm_70 -Xcompiler='-fopenmp' --std=c++11

# Fortran flags
FCFLAGS  = $(DFLAGS) -O3 -fopenmp -ffree-form -ffree-line-length-none
FCFLAGS    += -ftree-vectorize -funroll-loops -std=f2008
INC_FLAGS    = -I${MKLROOT}/include
INC_FLAGS    += -I$(COSMA)/include
INC_FLAGS    += -I$(LIBXSMM)/include  
INC_FLAGS    += -I$(INCELPA)/modules -I$(INCELPA)/elpa -I$(INCELPA)/src/helpers
INC_FLAGS    += -I$(LIBFFTW3)/include
INC_FLAGS    += -I$(LIBINT)/include
INC_FLAGS    += -I${CUDA_HOME}/include
INC_FLAGS    += -m64 -I${MKLROOT}/include

# C++ flags
CXXFLAGS += $(INC_FLAGS)

FCFLAGS +=  $(INC_FLAGS)

# C flags
CFLAGS   = $(DFLAGS) $(OPTFLAGS) $(INC_FLAGS)

# Libraries
LDFLAGS  = $(FCFLAGS)
LDFLAGS_C = $(FCFLAGS)

# ELPA libs
LIBS     = $(LIBELPA)/lib/libelpa.a
# Libvori lib
LIBS     += $(LIBVORI)/lib/libvori.a
# LIBXSMM libs
LIBS    += -L$(LIBXSMM)/lib -lxsmmf -lxsmm -lxsmmext  
# LIBINT libs
LIBS    += -L$(LIBINT)/lib -lint2 -lstdc++
# FFTW libs
LIBS    += -L$(LIBFFTW3)/lib -lfftw3 -lfftw3_threads
#SPGLIB
LIBS    += -L$(LIBSPG)/lib64 -lsymspg
#LIBXC
LIBS += -L$(LIBXC)/lib -lxcf03 -lxc

# Scalapack libs
LIBS    += -L/usr/lib64

LIBS    += -L$(MKLROOT)/lib/intel64 -Wl,--start-group -lmkl_scalapack_lp64 -lmkl_gf_lp64 -lmkl_sequential -lmkl_core -l
mkl_blacs_openmpi_lp64 -Wl,--end-group -lpthread -lz -ldl
 
#COSMA
LIBS    += -L$(COSMA_LIB)/ -lcosma_prefixed_pxgemm -lcosma -lcosta_prefixed_scalapack -lcosta

#CUDA
LIBS   += -L${CUDA_HOME}/lib64 -lstdc++ -lcudart -lnvrtc -lcuda -lcufft -lcublas -lnvToolsExt

# Required due to memory leak that occurs if high optimisations are used
mp2_optimize_ri_basis.o: mp2_optimize_ri_basis.F
$(FC) -c $(subst O2,O0,$(FCFLAGS)) $<

Could anyone help me with this please?

Regards

LC

[Ole Schütt]

Hi Luis,

the flags for GPU support have recently changed a bit:

  https://github.com/cp2k/cp2k/blob/master/INSTALL.md#2i-cuda-optional-improved-performance-on-gpu-systems

 Try adding -D__OFFLOAD_CUDA

-Ole

[Luis Cebamanos]

Thanks Ole,

I've replaced my DFLAGS  with the new ones:

 -D__MKL -D__FFTW3 -D__LIBXSMM  -D__LIBINT -D__parallel -D__SCALAPACK -D__HAS_ISO_C_BINDING -D__DBCSR_ACC -D__OFFLOAD_CUDA -D__OFFLOAD_PROFILING -D__HAS_PATCHED_CUFFT_70 -D__COSMA -D__ELPA_NVIDIA_GPU -D__LIBVORI -D__SPGLIB

Unfortunately, I am still getting the same error. Could the order of the CUDA DFLAGS elements be relevant?

Regards

--
You received this message because you are subscribed to the Google Groups "cp2k" group.
To unsubscribe from this group and stop receiving emails from it, send an email to cp2k+uns...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/cp2k/359554dc-fd75-466e-b155-2b14606e9ca9n%40googlegroups.com.

[Ole Schütt]

Hi Luis,

> Could the order of the CUDA DFLAGS elements be relevant?

No, the order of the DFLAGS does not matter. 

> /home/cp2k-2022.2/src/pw/gpu/pw_gpu_kernels.h(22): error: identifier "offloadStream_t" is undefined

This is really strange!

The "offloadStream_t" is defined here: https://github.com/cp2k/cp2k/blob/support/v2022.2/src/offload/offload_runtime.h#L27

... which "pw_gpu_kernels.h" includes here: https://github.com/cp2k/cp2k/blob/support/v2022.2/src/pw/gpu/pw_gpu_kernels.h#L11

I'm afraid you'll have to do some digging on your local machine to get to the bottom of this.

-Ole