NAME=cuda

include ../../Makefile_comp_tests.target

CUDA_DIR ?= /opt/cuda

CUPTI_DIR ?= $(CUDA_DIR)/extras/CUPTI

CUDRV_DIR ?= $(CUDA_DIR)

TESTS = HelloWorld simpleMultiGPU simpleMultiGPU_no_counters 

cuda_tests: $(TESTS) 

CUDA_DIR ?= $(CUDA_PATH)

NVCC = $(CUDA_DIR)/bin/nvcc 

NVCFLAGS = -g -ccbin='$(CC)'

INCLUDE += -I$(CUDA_DIR)/include -I$(CUPTI_DIR)/include 

CUDALIBS = -L$(CUDRV_DIR)/lib64 -L$(CUDA_DIR)/lib64 -L$(CUPTI_DIR)/lib64 -lcudart -lcupti -lcuda 

PAPILIB += -L../../../libpfm4/lib -lpfm

default: $(TESTS)

%.o:%.cu

	$(NVCC) $(INCLUDE) $(NVCFLAGS) -c -o $@ $<

HelloWorld: HelloWorld.o $(UTILOBJS)

	$(NVCC) $(NVCFLAGS) -o HelloWorld HelloWorld.o $(UTILOBJS) $(PAPILIB) $(CUDALIBS) $(LDFLAGS)

simpleMultiGPU: simpleMultiGPU.cu $(UTILOBJS)

	$(NVCC) $(NVCFLAGS) $(INCLUDE) -DPAPI -g -o $@ $+ $(UTILOBJS) $(PAPILIB) $(CUDALIBS) $(LDFLAGS)

simpleMultiGPU_no_counters: simpleMultiGPU.cu $(UTILOBJS)

	$(NVCC) $(NVCFLAGS) $(INCLUDE) -DNO_COUNTERS -g -o $@ $+ $(UTILOBJS) $(PAPILIB) $(CUDALIBS) $(LDFLAGS)

simpleMultiGPU_cupti_only: simpleMultiGPU.cu $(UTILOBJS)

	$(NVCC) $(NVCFLAGS) $(INCLUDE) -DCUPTI_ONLY -g -o $@ $+ $(UTILOBJS) $(PAPILIB) $(CUDALIBS) $(LDFLAGS)

clean:

	rm -f $(TESTS) *.o cuda_ld_preload_example.so

# Extra example to show LD_PRELOAD can be used to insert PAPI CUDA into a pre-existing binary

cuda_ld_preload_example.so: simpleMultiGPU_no_counters cuda_ld_preload_example.c

	gcc -Wall -fPIC -shared -o cuda_ld_preload_example.so $(LFLAGS) $(INCLUDE) cuda_ld_preload_example.c ../../../libpapi.so

run: 

	( cd ~/icl/papi/papi/src/ && make )

	( make HelloWorld )

	./HelloWorld

	( cd ~/icl/papi/papi/src/ && make )

	( make simpleMultiGPU )

	./simpleMultiGPU