cuda shared library linking: undefined reference t

Question

Goal:

1. create a shared library containing my CUDA kernels that has a CUDA-free wrapper/header.
2. create a test executable for the shared library.

Problem

1. shared library MYLIB.so seems to compile fine. (no problem).
2. Error in linking:

./libMYLIB.so: undefined reference to __cudaRegisterLinkedBinary_39_tmpxft_000018cf_00000000_6_MYLIB_cpp1_ii_74c599a1

simplified makefile:

libMYlib.so :  MYLIB.o
    g++  -shared  -Wl,-soname,libMYLIB.so  -o libMYLIB.so    MYLIB.o  -L/the/cuda/lib/dir  -lcudart


MYLIB.o : MYLIB.cu   MYLIB.h
    nvcc  -m64   -arch=sm_20 -dc  -Xcompiler '-fPIC'  MYLIB.cu  -o  MYLIB.o  -L/the/cuda/lib/dir  -lcudart


test : test.cpp  libMYlib.so
        g++   test.cpp  -o test  -L.  -ldl -Wl,-rpath,.   -lMYLIB  -L/the/cuda/lib/dir  -lcudart

---

indeed

nm libMYLIB.so shows that all CUDA api functions are "undefined symbols":

         U __cudaRegisterFunction
         U __cudaRegisterLinkedBinary_39_tmpxft_0000598c_00000000_6_CUPA_cpp1_ii_74c599a1
         U cudaEventRecord
         U cudaFree
         U cudaGetDevice
         U cudaGetDeviceProperties
         U cudaGetErrorString
         U cudaLaunch
         U cudaMalloc
         U cudaMemcpy

So CUDA somehow did not get linked to the shared library MYLIB.so What am I missing?

---

CUDA did not even get linked to the object file somehow:

nm MYLIB.o

         U __cudaRegisterFunction
         U __cudaRegisterLinkedBinary_39_tmpxft_0000598c_00000000_6_CUPA_cpp1_ii_74c599a1
         U cudaEventRecord
         U cudaFree
         U cudaGetDevice
         U cudaGetDeviceProperties
         U cudaGetErrorString
         U cudaLaunch
         U cudaMalloc
         U cudaMemcpy

(same as above)

Answer 1

Have you tried explictly disabling relocatable device code? i.e. -rdc=false? I got this undefined reference to __cudaRegisterLinkedBinaryWhatever with -rdc=true and it went away when I removed it. Although I'm not enough of an expert to explain what exactly is going on with that.

Answer 2

Here's an example linux shared object creation along the lines you indicated:

1. create a shared library containing my CUDA kernels that has a CUDA-free wrapper/header.
2. create a test executable for the shared library.

First the shared library. The build commands for this are as follows:

nvcc -arch=sm_20 -Xcompiler '-fPIC' -dc test1.cu test2.cu
nvcc -arch=sm_20 -Xcompiler '-fPIC' -dlink test1.o test2.o -o link.o
g++ -shared -o test.so test1.o test2.o link.o -L/usr/local/cuda/lib64 -lcudart

It seems you may be missing the second step above in your makefile, but I haven't analyzed if there are any other issues with your makefile.

Now, for the test executable, the build commands are as follows:

g++ -c main.cpp
g++ -o testmain main.o test.so

To run it, simply execute the testmain executable, but be sure the test.so library is on your LD_LIBRARY_PATH.

These are the files I used for test purposes:

test1.h:

int my_test_func1();

test1.cu:

#include <stdio.h>
#include "test1.h"

#define DSIZE 1024
#define DVAL 10
#define nTPB 256

#define cudaCheckErrors(msg) \
    do { \
        cudaError_t __err = cudaGetLastError(); \
        if (__err != cudaSuccess) { \
            fprintf(stderr, "Fatal error: %s (%s at %s:%d)\n", \
                msg, cudaGetErrorString(__err), \
                __FILE__, __LINE__); \
            fprintf(stderr, "*** FAILED - ABORTING\n"); \
            exit(1); \
        } \
    } while (0)

__global__ void my_kernel1(int *data){
  int idx = threadIdx.x + (blockDim.x *blockIdx.x);
  if (idx < DSIZE) data[idx] =+ DVAL;
}

int my_test_func1(){

  int *d_data, *h_data;
  h_data = (int *) malloc(DSIZE * sizeof(int));
  if (h_data == 0) {printf("malloc fail\n"); exit(1);}
  cudaMalloc((void **)&d_data, DSIZE * sizeof(int));
  cudaCheckErrors("cudaMalloc fail");
  for (int i = 0; i < DSIZE; i++) h_data[i] = 0;
  cudaMemcpy(d_data, h_data, DSIZE * sizeof(int), cudaMemcpyHostToDevice);
  cudaCheckErrors("cudaMemcpy fail");
  my_kernel1<<<((DSIZE+nTPB-1)/nTPB), nTPB>>>(d_data);
  cudaDeviceSynchronize();
  cudaCheckErrors("kernel");
  cudaMemcpy(h_data, d_data, DSIZE * sizeof(int), cudaMemcpyDeviceToHost);
  cudaCheckErrors("cudaMemcpy 2");
  for (int i = 0; i < DSIZE; i++)
    if (h_data[i] != DVAL) {printf("Results check failed at offset %d, data was: %d, should be %d\n", i, h_data[i], DVAL); exit(1);}
  printf("Results check 1 passed!\n");
  return 0;
}

test2.h:

int my_test_func2();

test2.cu:

#include <stdio.h>
#include "test2.h"

#define DSIZE 1024
#define DVAL 20
#define nTPB 256

#define cudaCheckErrors(msg) \
    do { \
        cudaError_t __err = cudaGetLastError(); \
        if (__err != cudaSuccess) { \
            fprintf(stderr, "Fatal error: %s (%s at %s:%d)\n", \
                msg, cudaGetErrorString(__err), \
                __FILE__, __LINE__); \
            fprintf(stderr, "*** FAILED - ABORTING\n"); \
            exit(1); \
        } \
    } while (0)

__global__ void my_kernel2(int *data){
  int idx = threadIdx.x + (blockDim.x *blockIdx.x);
  if (idx < DSIZE) data[idx] =+ DVAL;
}

int my_test_func2(){

  int *d_data, *h_data;
  h_data = (int *) malloc(DSIZE * sizeof(int));
  if (h_data == 0) {printf("malloc fail\n"); exit(1);}
  cudaMalloc((void **)&d_data, DSIZE * sizeof(int));
  cudaCheckErrors("cudaMalloc fail");
  for (int i = 0; i < DSIZE; i++) h_data[i] = 0;
  cudaMemcpy(d_data, h_data, DSIZE * sizeof(int), cudaMemcpyHostToDevice);
  cudaCheckErrors("cudaMemcpy fail");
  my_kernel2<<<((DSIZE+nTPB-1)/nTPB), nTPB>>>(d_data);
  cudaDeviceSynchronize();
  cudaCheckErrors("kernel");
  cudaMemcpy(h_data, d_data, DSIZE * sizeof(int), cudaMemcpyDeviceToHost);
  cudaCheckErrors("cudaMemcpy 2");
  for (int i = 0; i < DSIZE; i++)
    if (h_data[i] != DVAL) {printf("Results check failed at offset %d, data was: %d, should be %d\n", i, h_data[i], DVAL); exit(1);}
  printf("Results check 2 passed!\n");
  return 0;
}

main.cpp:

#include <stdio.h>

#include "test1.h"
#include "test2.h"

int main(){

  my_test_func1();
  my_test_func2();
  return 0;
}

When I compile according to the commands given, and run ./testmain I get:

$ ./testmain
Results check 1 passed!
Results check 2 passed!

Note that if you prefer, you may generate a libtest.so instead of test.so, and then you may use a modified build sequence for the test executable:

g++ -c main.cpp
g++ -o testmain main.o -L. -ltest

I don't believe it makes any difference, but it may be more familiar syntax.

I'm sure there is more than one way to accomplish this. This is just an example. You may wish to also review the relevant section of the nvcc manual and also review the examples.

EDIT: I tested this under cuda 5.5 RC, and the final application link step complained about not finding the cudart lib (warning: libcudart.so.5.5., needed by ./libtest.so, not found). However the following relatively simple modification (example Makefile) should work under either cuda 5.0 or cuda 5.5.

Makefile:

testmain : main.cpp  libtest.so
        g++ -c main.cpp
        g++ -o testmain  -L.  -ldl -Wl,-rpath,.   -ltest -L/usr/local/cuda/lib64 -lcudart main.o

libtest.so : link.o
        g++  -shared -Wl,-soname,libtest.so -o libtest.so    test1.o test2.o link.o  -L/usr/local/cuda/lib64  -lcudart

link.o : test1.cu test2.cu   test1.h test2.h
        nvcc  -m64   -arch=sm_20 -dc  -Xcompiler '-fPIC'  test1.cu test2.cu
        nvcc  -m64   -arch=sm_20 -Xcompiler '-fPIC' -dlink test1.o test2.o -o link.o

clean :
        rm -f testmain test1.o test2.o link.o libtest.so main.o

Answer 3

The other answers did not work for me (maybe because I’m using cuda 10). The solution that worked for me was compiling the cuda files as:

nvcc -dc -o cuda_file.o cuda_file.cu

Than compiling the c++ file as:

g++ -c -o cpp_file.o cpp_file.cpp

And finally linking all using nvcc:

nvcc -o my_prog cpp_file.o cuda_file.o -lcudart -lcuda -L<other stuff>

Don’t take this code literally. But the core of the solution to the error was using nvcc instead of g++ in the final linking step.