Makefile_hip

ifeq (x$(HIP_PATH),x)
	HIP_PATH=
endif

HIPCC=hipcc
HIPCCOPT=\
		-O3 \
		-fno-strict-aliasing \
		-fopenmp \
		-Wall \
		-mfma \
		-fvectorize \
		-fgpu-sanitize \
		-Rpass-analysis=kernel-resource-usage \
		-fhip-fp32-correctly-rounded-divide-sqrt \
		-Wno-unused-result \
		-Wno-unused-variable

# AMD Instinct MI100
#HIPCCOPT:= $(HIPCCOPT) \
#		--offload-arch=gfx908 -D__GFX9__
# AMD Instinct MI2X0
#HIPCCOPT:= $(HIPCCOPT) \
#		--offload-arch=gfx90a -D__GFX9__
# AMD Instinct MI300x
#HIPCCOPT:= $(HIPCCOPT) \
#		--offload-arch=gfx942 -D__GFX9__
# AMD Instinct RDNA2
#HIPCCOPT:= $(HIPCCOPT) \
#		--offload-arch=gfx1030,gfx1031,gfx1032,gfx1033 -D__GFX10__
# AMD Instinct RDNA3
HIPCCOPT:= $(HIPCCOPT) \
		--offload-arch=gfx1100,gfx1101,gfx1102,gfx1103 -D__GFX11__


OBJS = main.o eigen_GPU_check.o
OBJSS = $(OBJS) eigen_GPU_batch.o
LIBS = libeigenGbatch.a
LIBOPT = -leigenGbatch

all: a.out $(LIBS)
a.out : $(OBJSS) $(LIBS)
	$(HIPCC) -o $@ $(OBJSS) $(HIPCCOPT) -L./ -lhipsolver -lm -lgomp
	cp a.out a.out-hip

main.o: main.cpp
	$(HIPCC) -c -o $@ $< $(HIPCCOPT) -DPRINT_DIAGNOSTIC=0
libeigenGbatch.a: eigen_GPU_batch.o
	ar cr libeigenGbatch.a $<
	ranlib libeigenGbatch.a 
eigen_GPU_batch.o: eigen_GPU_batch.cu
	$(HIPCC) -c -o $@ $< $(HIPCCOPT)
	$(HIPCC) -S $< $(HIPCCOPT)
eigen_GPU_check.o: eigen_GPU_check.cu
	$(HIPCC) -c -o $@ $< $(HIPCCOPT)

clean:
	-\rm a.out a.out-* *.o *.cu_o *.ptx lib*.a