automatic dependencies with make and nvcc

Accelerated Computing CUDA CUDA Programming and Performance
1

Since nvcc can’t inline device functions, I’ve taken to #including .cu files in other .cu files. This led me to try to get rebuilding in make working with auto-dependencies via the -M flag. (Also, nvcc is slow to begin with.) How are you implementing automatic dependency generation with make and nvcc?

2

nvcc can’t inline device functions that are in separate compilation units

Nsight eclipse edition uses auto dependency generation with -M

You could study how it works.

3

Yes, forgot to say that.

I got it to work by following Automatic Prerequisites (GNU make)

Namely, what works for me is

$(OBJ)/%.d: $(SRC)/%.cu
	@ set -e; rm -f $@; \
	$(NVCC) -M $(NVCCFLAGS) $< > $@.$$$$; \
	sed 's,\($*\)\.o[ :]*,.o $@ : ,g' < $@.$$$$ > $@; \
	rm -f $@.$$$$

and including the .d files as a prerequisite for the corresponding .o. $(OBJ) is the object directory and $(SRC) the source directory.