diff --git a/Makefile b/Makefile index 7ced09e7a5f6e9553a025486cdd6fb6b4999a35a..431692f8df07e1d3a36b260320aecfb1e67ea3a5 100644 --- a/Makefile +++ b/Makefile @@ -2,16 +2,20 @@ GPU=NVIDIA GPU=AMD -OBJ = blas.o +OBJ_blas = blas.o +OBJ_fftw = fftw.o ifeq ($(GPU), AMD) - LDFLAGS = -L/opt/rocm/lib -lhipblas -lrocblas -fopenmp + LDFLAGS = -L/opt/rocm/lib -fopenmp + LDFLAGS_blas = -lhipblas -lrocblas + LDFLAGS_fftw = -lrocfft CFLAGS = -g -Wall -O3 -fopenmp -I/opt/rocm/include -I/opt/rocm/hip/include -DROC CC = hipcc else ifeq ($(GPU), NVIDIA) CC = nvcc - LDFLAGS = -lcublas -lm -lgomp + LDFLAGS = -lm -lgomp + LDFLAGS_blas = -lcublas INCLUDE= -I/usr/lib/x86_64-linux-gnu/openmpi/include/ CFLAGS = ${INCLUDE} --compile -O3 -pg -Xcompiler -fopenmp -DCUDA CUDAFLAGS = --Werror cross-execution-space-call --Wno-deprecated-gpu-targets @@ -19,9 +23,13 @@ else unknown_HW: endif +all: blas fftw -all: ${OBJ} - ${CC} -o blas ${OBJ} ${LDFLAGS} ${CUDAFLAGS} +blas: ${OBJ_blas} + ${CC} -o blas ${OBJ_blas} ${LDFLAGS} ${LDFLAGS_blas} ${CUDAFLAGS} + +fftw: ${OBJ_fftw} + ${CC} -o fftw ${OBJ_fftw} ${LDFLAGS} ${LDFLAGS_fftw} ${CUDAFLAGS} %.o: %.c ${HEADER} ${CC} ${CFLAGS} -c $<