#
# common Makefile
#
# gets included after the local Makefile in the example sub-directory
#

#CC=gcc
CC=icc

ifeq ($(CC), icc)
	OPT_FLAGS=-fast -I/usr/include
else
	# for gcc
	OPT_FLAGS :=-fopenmp -O3 -ftree-vectorize -msse3 #-ftree-parallelize-loops=4;
endif

CFLAGS = -DTIME -lm
PLCFLAGS +=
TILEFLAGS += 

#PERFCTR=perfctr

ifdef PERFCTR
	CFLAGS += -DPERFCTR -L/usr/local/lib64 -lpapi
endif

PLC=../../polycc

all: orig opt tiled par

orig: $(SRC).c decls.h papi_defs.h util.h
	$(CC) $(OPT_FLAGS) -lm $(SRC).c -o orig $(CFLAGS)

orig_par: decls.h papi_defs.h util.h $(SRC).c
	$(CC) -parallel $(OPT_FLAGS) -lm $(SRC).c -o orig_par $(CFLAGS)

$(SRC).opt.c: 
	$(PLC) $(SRC).c $(PLCFLAGS) 


$(SRC).tiled.c: 
	$(PLC) $(SRC).c --tile $(TILEFLAGS) $(PLCFLAGS) 

$(SRC).par.c: 
	$(PLC) $(SRC).c --tile --parallel $(TILEFLAGS) $(PLCFLAGS) 

opt: $(SRC).opt.c decls.h papi_defs.h util.h
	$(CC) $(OPT_FLAGS) -lm $(SRC).opt.c -o opt $(CFLAGS)

tiled: $(SRC).tiled.c decls.h papi_defs.h util.h
	$(CC) $(OPT_FLAGS) -lm $(SRC).tiled.c -o tiled  $(CFLAGS)

par: $(SRC).par.c decls.h papi_defs.h util.h
	$(CC) $(OPT_FLAGS) -openmp -lm $(SRC).par.c -o par  $(CFLAGS)


orig_test: $(SRC).c decls.h papi_defs.h util.h
	$(CC) $(OPT_FLAGS) -lm $(SRC).c -o orig_test -DTEST $(CFLAGS)

opt_test: $(SRC).opt.c decls.h papi_defs.h util.h
	$(CC) $(OPT_FLAGS) -lm $(SRC).opt.c -o opt_test -DTEST $(CFLAGS)

tiled_test: $(SRC).tiled.c decls.h papi_defs.h util.h
	$(CC) $(OPT_FLAGS) -lm $(SRC).tiled.c -o tiled_test -DTEST $(CFLAGS)

par_test: $(SRC).par.c decls.h papi_defs.h util.h
	$(CC) $(OPT_FLAGS) -openmp -lm $(SRC).par.c -o par_test -DTEST $(CFLAGS)

test: opt_test tiled_test orig_test par_test
	./orig_test 2> out_orig
	./opt_test 2> out_opt
	diff -q out_orig out_opt
	./tiled_test 2> out_tiled
	diff -q out_orig out_tiled
	export OMP_NUM_THREADS=4; ./par_test 2> out_par4
	diff -q out_orig out_par4
	@echo Success!

clean:
	rm -f out_* *.tiled.c *.opt.c *_test opt orig tiled  sched_test sched hopt hopt_test *.par.c par *.par2d.c orig_par *.out.* *.kernel.* a.out \
		$(EXTRA_CLEAN) tags tmp* gmon.out *~ .unroll .vectorize par2d

exec-clean:
	rm -f out_* *_test opt orig tiled  sched_test sched hopt hopt_test par orig_par *.out.* *.kernel.* a.out \
		$(EXTRA_CLEAN) tags tmp* gmon.out *~ par2d