Newer
Older
CC=g++
ICC =icc
DEBUG ?= 1
CFLAGS=-O3 -lm -Wall -mavx -march=ivybridge -mtune=ivybridge -fopenmp
#CFLAGS=-O3 -lm -Wall -mavx2 -mfma -march=haswell -mtune=haswell
#CFLAGS=-O3 -Wall -xCORE-AVX-I
#CFLAGS=-O3 -Wall -xCORE-AVX2
#ICFLAGS=-O3 -Wall -qopenmp -axCORE-AVX2,CORE-AVX-I
GPU_CXX = nvcc
LDFLAGS ?=-L $(CUDA_PREFIX)/lib64 -lcudart -lcublas -lcusparse -lm -lrt
GPU_COMPILE = nvcc -I $(CUDA_PREFIX)/include -arch sm_35
GPU_MPI_COMPILE = $(GPU_MPI_CXX) -I $(CUDA_PREFIX)/include -arch sm_35
CPU_COMPILE = $(CC) $(CFLAGS)
ifeq ($(DEBUG), 1)
CPU_COMPILE += -D_DEBUG_
GPU_COMPILE += -D_DEBUG_
endif
CPU_COMPILE_OBJ= $(CPU_COMPILE) -c
GPU_COMPILE_OBJ= $(GPU_COMPILE) -c
EXT_DIR = /users/guest/petyros/Training/External_Functions/
OBJECTS = util.o matrix_op.o timer.o input.o gpu_util.o
cuBLAS_MultiGPU: $(OBJECTS) cuBLAS_MultiGPU.cu
$(GPU_MPI_COMPILE) -o $@ $(OBJECTS) $(LDFLAGS) $@.cu
cuBLAS: $(OBJECTS) cuBLAS.cu
$(GPU_COMPILE) -o $@ $(OBJECTS) $(LDFLAGS) $@.cu
gpu_util.o: $(EXT_DIR)gpu_util.cu
$(GPU_COMPILE_OBJ) -o $@ $<
%.o: $(EXT_DIR)%.c
$(CPU_COMPILE_OBJ) -o $@ $<
%.o: %.h
clean:
$(RM) $(PROGRAMS) $(OBJECTS)