Compiling CUDA Library: Works on Mac, not on Ubuntu

I'm making my first attempts at building a .so full of CUDA routines. I have matrix_vector_mult.cu file which currently does nothing:

#include <stdio.h>extern "C"double *matrix_vector_mult(const double ** const M,                   const double * const v,                   const size_t num_rows,                   const size_t num_cols){    printf("Hello!\n");    double * p = (double *) malloc(num_rows*sizeof(double));    return p;}

I also have a makefile, whose contents is as follows:

CC := clangUNAME := $(shell uname -s)ifeq ($(UNAME), Darwin)CUDA_PATH := /Developer/NVIDIA/CUDA-6.5CUDA_LIB := ${CUDA_PATH}/libendififeq ($(UNAME), Linux)CUDA_PATH := /usr/local/cuda-6.5CUDA_LIB := ${CUDA_PATH}/lib64endifLIBS := -L ${CUDA_LIB} -lcudart -lcudadevrtNVCC := ${CUDA_PATH}/bin/nvcc -ccbin ${CC}CFLAGS := -g -std=c11 -Wextra -Wall -I include -rpath ${CUDA_LIB}NVCCFLAGS := -g -m64 -D__STRICT_ANSI__vpath %.cu srcvpath %.h includeall: matrix_vector_mult.o        ${CC} ${CFLAGS} -o matrix_vector_mult.so -shared -fPIC $^ ${LIBS}matrix_vector_mult.o: matrix_vector_mult.cu        ${NVCC} ${NVCCFLAGS} -o $@ -c $^clean:        rm -f *.o *.so *.pyc

On Mac, this compiles just fine. However, on my Ubuntu box, I get the error message:

/usr/bin/ld: matrix_vector_mult.o: relocation R_X86_64_32S against `.rodata.str1.1' can not be used when making a shared object; recompile with -fPICmatrix_vector_mult.o: error adding symbols: Bad value

What could be the problem? (Adding -fPIC to the compile line doesn't work.) nvcc --version yields identical information on both boxes, clang --version gives

Apple LLVM version 6.0 (clang-600.0.56) (based on LLVM 3.5svn)

on the Mac, and

Ubuntu clang version 3.4-1ubuntu3 (tags/RELEASE_34/final) (based on LLVM 3.4)

on the Ubuntu box. I somewhat doubt the slight LLVM version difference is the problem, because I regard the makefile as a mess already. Any help appreciated.