TARGET=main
OBJECTS=util.o convolution.o main.o

CPPFLAGS=-std=c++11 -O3 -Wall -march=native -mavx2 -mfma -mavx512f -fopenmp
CPPFLAGS+=-I/usr/local/cuda/include/
LDFLAGS=-lm 
LDFLAGS+=-L/usr/local/cuda/lib64 -lOpenCL -lmpi -lmpi_cxx -lcudart

all: $(TARGET)

$(TARGET): $(OBJECTS)
	g++ $(CPPFLAGS) $^ -o $@ $(LDFLAGS)

clean:
	rm -rf $(TARGET) $(OBJECTS)

run:
	./run.sh

rv:
	rm -f rv_output.log
	./run_validate.sh > rv_output.log

rp:
	./run_performance.sh

submit: clean
	rm -f *.log
	cp -f * ~/submit/final/A/