cuGL* v. cudaGL* functions

Hi,
I’m having some linker issues with the CUDA OpenGL interop functions. I’m using CUDA 6.5. In the cudaGL.h header file, there are the following definitions:

CUresult CUDAAPI cuGraphicsGLRegisterBuffer(CUgraphicsResource *pCudaResource, GLuint buffer, unsigned int Flags);
CUresult CUDAAPI cuGLUnmapBufferObject(GLuint buffer);

These are wrapped in an extern “C”, so I’d expect one of the supplied CUDA libs to contain the symbols _cuGraphicsGLRegisterBuffer and _cuGLUnmapBufferObject. However, this is not the case. All you find is:

0000000000029380 T _cudaGLGetDevices
0000000000029c40 T _cudaGLMapBufferObject
0000000000029690 T _cudaGLMapBufferObjectAsync
0000000000029db0 T _cudaGLRegisterBufferObject
0000000000029830 T _cudaGLSetBufferObjectMapFlags
0000000000029f00 T _cudaGLSetGLDevice
0000000000029af0 T _cudaGLUnmapBufferObject
0000000000029510 T _cudaGLUnmapBufferObjectAsync
00000000000299a0 T _cudaGLUnregisterBufferObject

When I grep for the cuGL functions, I see:

000000000003dc50 S __ZN6cudart14__fun_cuGLInitE
000000000003dc58 S __ZN6cudart20__fun_cuGLGetDevicesE
000000000003dc48 S __ZN6cudart22__fun_cuGLCtxCreate_v2E
000000000003dc78 S __ZN6cudart27__fun_cuGLUnmapBufferObjectE
000000000003dc68 S __ZN6cudart28__fun_cuGLMapBufferObject_v2E
000000000003dc60 S __ZN6cudart30__fun_cuGLRegisterBufferObjectE
000000000003dc80 S __ZN6cudart32__fun_cuGLUnmapBufferObjectAsyncE
000000000003dc88 S __ZN6cudart32__fun_cuGLUnregisterBufferObjectE
000000000003dc70 S __ZN6cudart33__fun_cuGLMapBufferObjectAsync_v2E
000000000003dc90 S __ZN6cudart33__fun_cuGLSetBufferObjectMapFlagsE
000000000003d6d0 b __ZN6cudartL17__funtmp_cuGLInitE
0000000000008c60 t __ZN6cudartL17__unimpl_cuGLInitEv
000000000003d6d8 b __ZN6cudartL23__funtmp_cuGLGetDevicesE
0000000000008c70 t __ZN6cudartL23__unimpl_cuGLGetDevicesEPjPij19CUGLDeviceList_enum
000000000003d6c8 b __ZN6cudartL25__funtmp_cuGLCtxCreate_v2E
0000000000008c50 t __ZN6cudartL25__unimpl_cuGLCtxCreate_v2EPP8CUctx_stji
000000000003d6f8 b __ZN6cudartL30__funtmp_cuGLUnmapBufferObjectE
0000000000008cb0 t __ZN6cudartL30__unimpl_cuGLUnmapBufferObjectEj
000000000003d6e8 b __ZN6cudartL31__funtmp_cuGLMapBufferObject_v2E
0000000000008c90 t __ZN6cudartL31__unimpl_cuGLMapBufferObject_v2EPyPmj
000000000003d6e0 b __ZN6cudartL33__funtmp_cuGLRegisterBufferObjectE
0000000000008c80 t __ZN6cudartL33__unimpl_cuGLRegisterBufferObjectEj
000000000003d700 b __ZN6cudartL35__funtmp_cuGLUnmapBufferObjectAsyncE
000000000003d708 b __ZN6cudartL35__funtmp_cuGLUnregisterBufferObjectE
0000000000008cc0 t __ZN6cudartL35__unimpl_cuGLUnmapBufferObjectAsyncEjP11CUstream_st
0000000000008cd0 t __ZN6cudartL35__unimpl_cuGLUnregisterBufferObjectEj
000000000003d6f0 b __ZN6cudartL36__funtmp_cuGLMapBufferObjectAsync_v2E
000000000003d710 b __ZN6cudartL36__funtmp_cuGLSetBufferObjectMapFlagsE
0000000000008ca0 t __ZN6cudartL36__unimpl_cuGLMapBufferObjectAsync_v2EPyPmjP11CUstream_st
0000000000008ce0 t __ZN6cudartL36__unimpl_cuGLSetBufferObjectMapFlagsEjj

What gives? These are even documented http://developer.download.nvidia.com/compute/cuda/4_1/rel/toolkit/docs/online/group__CUDA__GL_gd530f66cc9ab43a31a98527e75f343a0.html.

Here was my test code:

CUgraphicsResource testBufferCuda;

CUresult err = cuGraphicsGLRegisterBuffer(&testBufferCuda, buffer, CU_GRAPHICS_REGISTER_FLAGS_NONE);
err = cuGraphicsMapResources(1, &testBufferCuda, 0);

CUdeviceptr devPtr;
size_t devPtrSize;
err = cuGraphicsResourceGetMappedPointer(&devPtr, &devPtrSize, testBufferCuda);
err = cuGraphicsUnmapResources(1, &testBufferCuda, 0);

How am I supposed to map a buffer properly? All of these functions are marked deprecated and the examples I’m finding online to very different things depending on how old they are.

thanks!

cuXXXXX functions are provided by the driver API. They are not in the cuda runtime API, which provides functions like cudaXXXXX. cudart also means “cuda runtime”

To link driver API code correctly, its usually necessary to link against libcuda.so (on linux). This may get installed in various places depending on the distro, but may be in /usr/lib or /usr/lib64

The documentation you linked is for CUDA 4.1, quite old. Current CUDA docs are at docs.nvidia.com

On CUDA 6.5/RHEL 6.2:

grep cuGraphicsGLRegisterBuffer /usr/lib64/libcuda.so.340.29 Binary file /usr/lib64/libcuda.so.340.29 matches