I've created a Haskell package that makes FFI calls to functions defined in CUDA code. I'd like to compile .cu file to an object (.o) file during package build and force linker to link it in.
So far, I tried to use a technique found this question. I've customized buildHook to:
- run nvcc
- run default buildHook
- create ar library file with nvcc compiled code.
Setup.hs is available here.
This solution has a major disadvantage in restricting this package to static linking. Although cabal produces a shared library, it won't work because it has no way of resolving symbols located in the object file.
Is there a simpler way to link custom code during building?
I do a similar thing. I have a Haskell file which calls CUDA code.
Here's how I compile CUDA libraries and link with Haskell:
$(NVCC) -c -E $(NVCC_OPTS) -o build/file.i file.cu
$(NVCC) -c $(NVCC_OPTS) -o build/file.o file.cu
I then link everything into a C++ Shared Library called LibSO
with Haskell options
$(CXX) -shared -Wl,-rpath=\$$$$ORIGIN $(CXX_LINK_LIBS) $(PACKAGE_RPATH) -Lbuild -rdynamic -L/usr/local/lib/ghc-7.6.3 -lHSrts-ghc7.6.3 -o build/LibSO.so build/file.o
where
CXX_LINK_LIBS = -Lbuild -lcudart -lcuda -lpthread -lcupti -lcurand -lnvidia-ml
NVCC_OPTS = --compiler-options -fPIC -maxrregcount=0 --machine 64 --DCUDA
I then take my Haskell files and compile them into o and hi files. (I compile twice because of TemplateHaskell)
ghc -v0 -Wall -rtsopts -threaded -stubdir build -ibuild/ -no-hs-main -o build/iop.o -ohi build/iop.hi -c haskell/iop.lhs
ghc -v0 -Wall -rtsopts -threaded -stubdir build -ibuild/ -no-hs-main -fPIC -dynamic -osuf dyn_o -hisuf dyn_hi -o build/iop.dyn_o -ohi build/iop.dyn_hi -c haskell/iop.lhs
So now we have haskell dynamic objects and a C++ shared library.
In the end, I link a main haskell file with everything:
ghc -optl "-Wl,-rpath=\$$ORIGIN" $(CXX_LINK_LIBS) -Lbuild -rtsopts -threaded -lstdc++ -lLibSO -o build/Main build/iop.dyn_o
Does this sort of help?