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 LDFLAGS+=-L/usr/local/cuda/lib64 LDFLAGS+=-L/usr/local/cuda/lib64 -lOpenCL -lmpi -lmpi_cxx -lcudart 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) run: rm -f run_output.log ./run.sh > run_output.log # ./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/B/