NVCC     = ${CUDA_PATH}/bin/nvcc
CC       = gcc
CXX      = g++
FC       = mpif90
LD       = mpif90
AR       = ar -r
GPUVER   = V100

CUDAPATH  = /cineca/prod/opt/compilers/cuda/11.0/none

CXXFLAGS = -O3 -I$(CUDAPATH)/include -std=c++11 -fopenmp
DATA_DIR   = /m100_work/Ppp4x_5489/CP2K/cp2k-8.1/data
CP2K_ROOT  = /m100_work/Ppp4x_5489/CP2K



LIBINT_INC  = $(CP2K_ROOT)/libs/libint/include
LIBINT_LIB  = $(CP2K_ROOT)/libs/libint/lib
LIBXC_INC   = $(CP2K_ROOT)/libs/libxc/include
LIBXC_LIB   = $(CP2K_ROOT)/libs/libxc/lib





DFLAGS   = -D__FFTW3 -D__ACC -D__DBCSR_ACC  -D__SCALAPACK  -D__PW_CUDA -D__parallel -D__LIBINT -D__MPI_VERSION=3 -D__LIBXC -D__GFORTRAN

FCFLAGS  = -fopenmp  -std=f2008 -fimplicit-none -ffree-form  -fno-omit-frame-pointer -O3 -ftree-vectorize $(DFLAGS) $(WFLAGS)

FCFLAGS    += -I$(LIBINT_INC) -I$(LIBXC_INC)

LDFLAGS  =  -L$(CUDAPATH)/lib64 $(FCFLAGS)
NVFLAGS  = $(DFLAGS)  -O3 -arch sm_70 -Xcompiler='-fopenmp' --std=c++11
CFLAGS   = $(DFLAGS) -I$(LAPACK_INC) -I${FFTW_INC}  -fno-omit-frame-pointer -g -O3 -fopenmp
LIBS     =  -L${LAPACK_LIB} -L${BLAS_LIB} -L${FFTW_LIB} -L${CUDA_LIB} -L${SCALAPACK_LIB} -lscalapack   -llapack -lblas  -lstdc++ -lfftw3 -lfftw3_omp -lcuda -lcudart -lnvrtc -lcufft -lcublas -lrt

LIBS       += $(LIBINT_LIB)/libint2.a
LIBS       += $(LIBXC_LIB)/libxcf03.a $(LIBXC_LIB)/libxc.a