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 -lcudart -lmpi -lmpi_cxx
LDFLAGS+=-L/usr/local/cuda/lib64  

NVCC=/usr/local/cuda/bin/nvcc

all: $(TARGET)

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

convolution.o: convolution.cu
	$(NVCC) -c -o $@ $^
 
clean:
	rm -rf $(TARGET) $(OBJECTS)