PROG=volume_cu
ISPC_SRC=volume1.ispc
CXX_SRC=volume_cu.cpp 

CXX=g++
CXXFLAGS=-O3 -I$(CUDATK)/include
LD=g++
LDFLAGS=-lcuda

ISPC=ispc
ISPCFLAGS=-O3 --math-lib=default --target=nvptx64,avx

LLVM32 = $(HOME)/usr/local/llvm/bin-3.2
LLVM   = $(HOME)/usr/local/llvm/bin-3.3
PTXGEN = $(HOME)/ptxgen

LLVM32DIS=$(LLVM32)/bin/llvm-dis

.SUFFIXES: .bc .o .ptx .cu _ispc_nvptx64.bc


ISPC_OBJ=$(ISPC_SRC:%.ispc=%_ispc.o)
ISPC_BC=$(ISPC_SRC:%.ispc=%_ispc_nvptx64.bc)
PTXSRC=$(ISPC_SRC:%.ispc=%_ispc_nvptx64.ptx)
CXX_OBJ=$(CXX_SRC:%.cpp=%.o)

all: $(PROG)


$(CXX_OBJ) : kernel.ptx
$(PROG): $(CXX_OBJ) kernel.ptx
	/bin/cp kernel.ptx __kernels.ptx
	$(LD) -o $@ $(CXX_OBJ) $(LDFLAGS)

%.o: %.cpp
	$(CXX) $(CXXFLAGS)  -o $@ -c $<


%_ispc_nvptx64.bc: %.ispc
	$(ISPC) $(ISPCFLAGS) --emit-llvm -o `basename $< .ispc`_ispc.bc -h `basename $< .ispc`_ispc.h $< --emit-llvm

%.ptx: %.bc
	$(LLVM32DIS) $<
	$(PTXGEN)  `basename $< .bc`.ll > $@

kernel.ptx: $(PTXSRC)
	cat $^ > kernel.ptx

clean: 
	/bin/rm -rf *.ptx *.bc *.ll $(PROG)