CUDA 5.0 separate compilation of library with cmake

Question

The buildtime of my cuda library is increasing and so I thought that separate compilation introduced in CUDA 5.0 might help me. I couldn\'t figure out how to achieve separat

Answer 1

EDIT (2016-03-15): Yes, it is confirmed as a bug in FindCUDA: https://cmake.org/Bug/view.php?id=15157

---

TL;DR: This seems to be a bug in FindCUDA, which makes objects lose info on external definitions before the final linking.

The problem is that, even if separable compilation is enabled, a linking step is still performed for all the targets individually before the final linking.

For instance, I have module.cu with:

#include "module.h"
#include 

double arr[10] = {1,2,3,4,5,6,7,8,9,10};
__constant__ double carr[10];

void init_carr() {
  cudaMemcpyToSymbol(carr,arr,10*sizeof(double));
}

__global__ void pkernel() {
  printf("(pkernel) carr[%d]=%g\n",threadIdx.x,carr[threadIdx.x]);
}

void print_carr() {
  printf("in print_carr\n");
  pkernel<<<1,10>>>();
}

and module.h with:

extern __constant__ double carr[10];
extern double arr[10];

void print_carr();
void init_carr();

and finally main.cu with:

#include "module.h"

#include 

__global__ void kernel() {
  printf("(kernel) carr[%d]=%g\n",threadIdx.x,carr[threadIdx.x]);
}


int main(int argc, char *argv[]) {
  printf("arr: %g %g %g ..\n",arr[0],arr[1],arr[2]);

  kernel<<<1,10>>>();
  cudaDeviceSynchronize();
  print_carr();
  cudaDeviceSynchronize();
  init_carr();
  cudaDeviceSynchronize();
  kernel<<<1,10>>>();
  cudaDeviceSynchronize();
  print_carr();
  cudaDeviceSynchronize();

  return 0;
}

This then works fine with the following Makefile:

NVCC=nvcc
NVCCFLAGS=-arch=sm_20
LIB=libmodule.a
OBJS=module.o main.o
PROG=extern

$(PROG): main.o libmodule.a
    $(NVCC) $(NVCCFLAGS) -o $@ $^

%.o: %.cu
    $(NVCC) $(NVCCFLAGS) -dc -c -o $@ $^

$(LIB): module.o
    ar cr $@ $^

clean:
    $(RM) $(PROG) $(OBJS) $(LIB)

But then I try to use the following CMakeLists.txt:

CMAKE_MINIMUM_REQUIRED(VERSION 2.8.8)

PROJECT(extern)

FIND_PACKAGE(CUDA REQUIRED)
SET(CUDA_SEPARABLE_COMPILATION ON)

SITE_NAME(HOSTNAME)

SET(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS} -arch=sm_20)

cuda_add_library(module module.cu)

CUDA_ADD_EXECUTABLE(extern main.cu)
TARGET_LINK_LIBRARIES(extern module)

When then compiling, what then happens is that the following:

$ cmake ..
-- The C compiler identification is GNU 4.9.2
...
$ make VERBOSE=1
...
[ 25%] Building NVCC (Device) object CMakeFiles/module.dir//./module_generated_module.cu.o
...
-- Generating <...>/build/CMakeFiles/module.dir//./module_generated_module.cu.o
/usr/local/cuda/bin/nvcc <...>/module.cu -dc -o <...>/build/CMakeFiles/module.dir//./module_generated_module.cu.o -ccbin /usr/bin/cc -m64 -Xcompiler ,\"-g\" -arch=sm_20 -DNVCC -I/usr/local/cuda/include
[ 50%] Building NVCC intermediate link file CMakeFiles/module.dir/./module_intermediate_link.o
/usr/local/cuda/bin/nvcc -arch=sm_20 -m64 -ccbin "/usr/bin/cc" -dlink <...>/build/CMakeFiles/module.dir//./module_generated_module.cu.o -o <...>/build/CMakeFiles/module.dir/./module_intermediate_link.o
...
/usr/bin/ar cr libmodule.a  CMakeFiles/module.dir/./module_generated_module.cu.o CMakeFiles/module.dir/./module_intermediate_link.o
/usr/bin/ranlib libmodule.a
...
[ 50%] Built target module
[ 75%] Building NVCC (Device) object CMakeFiles/extern.dir//./extern_generated_main.cu.o
...
-- Generating <...>/build/CMakeFiles/extern.dir//./extern_generated_main.cu.o
/usr/local/cuda/bin/nvcc <...>/main.cu -dc -o <...>/build/CMakeFiles/extern.dir//./extern_generated_main.cu.o -ccbin /usr/bin/cc -m64 -Xcompiler ,\"-g\" -arch=sm_20 -DNVCC -I/usr/local/cuda/include -I/usr/local/cuda/include
...
[100%] Building NVCC intermediate link file CMakeFiles/extern.dir/./extern_intermediate_link.o
/usr/local/cuda/bin/nvcc -arch=sm_20 -m64 -ccbin "/usr/bin/cc" -dlink <...>/build/CMakeFiles/extern.dir//./extern_generated_main.cu.o -o <...>/build/CMakeFiles/extern.dir/./extern_intermediate_link.o
nvlink error   : Undefined reference to 'carr' in '<...>/build/CMakeFiles/extern.dir//./extern_generated_main.cu.o'

Clearly, the problem are the nvcc -dlink obj.o -o obj_intermediate_link.o lines. Then, I guess, the info on external definitions are lost. So, the question is, it is possible to make CMake/FindCUDA not do this extra linking step?

Otherwise, I would argue that this is a bug. Do you agree? I can file a bug report with CMake.