CUDA compilation GCC 5.4.0

[schweitzer]

Hi,
PCL fails to compile from sources on release with gcc 5.4 and cuda 7.5.

Your Environment

• Operating System and version: Linux Mint 18.1
• Compiler: GCC 5.4
• PCL Version: Latest release pcl-1.8.0

Current Behavior

The compilation fails with this kind of errors:

/usr/include/string.h: In function ‘void* __mempcpy_inline(void*, const void*, size_t)’: /usr/include/string.h:652:42: error: ‘memcpy’ was not declared in this scope return (char *) memcpy (__dest, __src, __n) + __n; ^ CMake Error at cuda_compile_generated_bilateral_pyrdown.cu.o.Release.cmake:267 (message): Error generating file /home/mschweitzer/dev/f4s/deps/build/release/pcl/pcl-prefix/src/pcl-build/gpu/kinfu/CMakeFiles/cuda_compile.dir/src/cuda/./cuda_compile_generated_bilateral_pyrdown.cu.o

It seems that this kind of cuda error is common (like here in TensorFlow or also here )

Possible Solution

A possible walktrough is to add -D_FORCE_INLINES to PCL CUDA_NVCC_FLAGS (this works for me).

Maybe an update to the Cuda 8.0 version can also resolve this issue but I didn't try it.

Am I the only one getting this error ?

Thanks,

[taketwo]

From my experience this flag is also needed with CUDA 8.0. I had to use it in several of my projects, but I was never concerned with backward compatibility. If we don't break building with older GCC/CUDA by adding this flag, then we should add it I think.

[soulslicer]

I have never gotten any of the cuda libraries to work with the new pascal series of gpus. have you?

[taketwo]

I don't have a Pascal GPU, unfortunately.

[soulslicer]

Turns out CUDA 8.0 solves this. My bad