x86-64 linking CUDA object files into a .so Linux

Accelerated Computing CUDA CUDA Programming and Performance
1

I have a couple global functions I want to link into a shared library object on x86-64 linux with device emulation, but when linking I get the following error:

/usr/lib64/gcc/x86_64-suse-linux/4.1.2/…/…/…/…/x86_64-suse-linux/bin/ld: least_squares_gpu.o: relocation R_X86_64_PC32 against `__saturate’ can not be used when making a shared object; recompile with -fPIC
/usr/lib64/gcc/x86_64-suse-linux/4.1.2/…/…/…/…/x86_64-suse-linux/bin/ld: final link failed: Bad value

CFLAGS used to compile the .CU file least_squares_gpu.cu:
-Xcompiler -fPIC -DPIC -g -Xcompiler -std=gnu99 -Xopencc -fPIC -DPIC -I…/include -I./ -I…/include -I/usr/include -DCUDACC -DDEBUG -I/home/baver/NVIDIA_CUDA_SDK/common/inc -deviceemu

Compilation works fine without -deviceemu flag. All other non-CUDA files which are also being linked have -fPIC as part of the CFLAGS.

Suggestions? Using openSuSE 10.2

Thanks,
baver

2

I found a hack/workaround to this problem, if I’m using -O3 setting in gcc to compile .o file , the functions which make this problem appear like __saturate get inlined and linking is done.
Tis sort of not the best thing for debugging. I ended up with having my cuda code built into both .so and executable(for debugging).

3

Has anyone figured out any other workarounds for this? Reconfiguring my app at this stage isn’t much of an option for me.

Thanks,
Brian

4

aha, the easiest thing by far (if you have all the libraries to do so) is to build your project in 32 bit mode for emulation.

5

Never mind, thought I had it but haven’t tried for device emulation yet.

6

Did anyone ever manage to find a work-around for this in the end without using -O3? I have the same problem, also running openSuse 10.2.

7

Also, does anyone know if this problem has been fixed in 1.1? If so, I’ll probably just switch to Red Hat.

Edit:

Just thought I should include a quick example. I think I’m using nvcc/gcc in the right way, but I might be wrong. If anyone could try this out in 1.1 I would be very grateful.

#include <stdio.h>
extern “C” void test()
{
printf(“test\n”);
}

nvcc -o test.o -shared -c test.cu -Xcompiler -fPIC
gcc -o libtest.so -shared test.o

This works fine and the library can be linked into my main application. However, adding -deviceemu as below gives an error.

nvcc -o test.o -shared -c test.cu -Xcompiler -fPIC -deviceemu
gcc -o libtest.so -shared test.o

ld: test.o: relocation R_X86_64_PC32 against `__saturate’ can not be used when making a shared object; recompile with -fPIC
ld: final link failed: Bad value

8

tbrandvik,

you must be picking up some 1.0 header files.
in 1.1 we deprecated

double __saturate(double)

it should work when you’re picking up the right headers

9

Sorry, probably should have been clearer. I am actually using 1.0, but was wondering if it works in 1.1. From what you’re saying though, it sounds like it should do so I’ll have a go at upgrading. Thanks for your help.