MakefileBaseNVHPC

NVCC:=nvcc
CXX:=nvc++
#SM=70
#NGPU=1
# NUMA nodes
NSOCKETS=2

graph_ft_load=1
use_par_exec_policy=1
check=0
debug=0

ifeq ($(debug), 1)
	CUFLAGS = -g -G -Ddebug --std=c++14 --gpu-architecture=compute_${SM} --gpu-code=sm_${SM},compute_${SM} \
                  -Xcompiler -fopenmp -DMULTIPHASE -DNGPU=1 -DPRINT
	CXXFLAGS = -g -Ddebug -mcmodel=medium -Wextra -Wall -DMULTIPHASE -DNGPU=1 -DPRINT 
else 
	CUFLAGS = -O3 -Xptxas -O3 --std=c++14 --gpu-architecture=compute_${SM} --gpu-code=sm_${SM},compute_${SM} \
                  -Xcompiler -O3 -Xcompiler -fopenmp -DMULTIPHASE -DPRINT -DNGPU=${NGPU}
	CXXFLAGS = -O3 -mcmodel=medium -fopenmp -Wextra -Wall -DMULTIPHASE -DPRINT -DNGPU=${NGPU}
endif

ifeq ($(use_par_exec_policy), 1)
	 CXXFLAGS += -DUSE_PAR_EXEC_POLICY
endif

ifeq ($(check),1)
	CUFLAGS += -DCHECK
	CXXFLAGS  += -DCHECK
endif

CUFLAGS += -DGRAPH_FT_LOAD=${NSOCKETS}
CXXFLAGS  += -DGRAPH_FT_LOAD=${NSOCKETS}

ifeq ($(bit),32)
	CUFLAGS += -DUSE_32BIT
	CXXFLAGS  += -DUSE_32BIT
else ifeq ($(bit),64)
	CUFLAGS += -DUSE_64BIT
	CXXFLAGS  += -DUSE_64BIT
else
	CUFLAGS += -DUSE_64BIT
	CXXFLAGS  += -DUSE_64BIT
endif

EXE:= run_${NGPU}_${SM}

all: ${EXE}

graph.o: graph.cpp graph.hpp types.hpp
	${CXX} -cuda ${CXXFLAGS} -c $< -o $@

graph_gpu.o: graph_gpu.cpp graph_gpu.hpp cuda_wrapper.hpp types.hpp
	${CXX} -cuda ${CXXFLAGS} -c $< -o $@

louvain_gpu.o: louvain_gpu.cpp louvain_gpu.hpp cuda_wrapper.hpp types.hpp
	${CXX} -cuda ${CXXFLAGS} -c $< -o $@

main.o: main.cpp graph.hpp graph_gpu.hpp cuda_wrapper.hpp types.hpp louvain_gpu.hpp
	${CXX} -cuda ${CXXFLAGS} -c $< -o $@

graph_cuda.o: graph_cuda.cu graph_cuda.hpp graph.cpp graph.hpp
	${NVCC} ${CUFLAGS} -c $< -o $@

heap.o : heap.cpp heap.hpp types.hpp
	${CXX} -cuda ${CXXFLAGS} -c $< -o $@

clustering.o: clustering.cpp clustering.hpp types.hpp
	${CXX} -cuda ${CXXFLAGS} -c $< -o $@

ifeq ($(check),1)
graph_cpu.o: graph_cpu.cpp graph_cpu.hpp
	${CXX} -cuda ${CXXFLAGS} -c $< -o $@
endif

ifeq ($(check),1)
${EXE}: graph_gpu.o graph_cuda.o graph.o louvain_gpu.o main.o heap.o clustering.o graph_cpu.o
	${CXX} -cuda ${CXXFLAGS} $^ -o $@
else
${EXE}: graph_gpu.o graph_cuda.o graph.o louvain_gpu.o main.o heap.o clustering.o
	${CXX} -cuda ${CXXFLAGS} $^ -o $@
endif

clean:
	rm -f *.o ${EXE}