nvcc bug w/ templaes - compiles .cpp but not .cu

pbarletta · August 26, 2018, 4:05pm

Hi, I was using the CGAL libraries with my code and everything was fine until I tried to use this particular piece of code:
https://github.com/CGAL/cgal/blob/7180fe5bb3da062687af2a378a9d770248241cff/Triangulation_3/include/CGAL/Delaunay_triangulation_3.h
So I decided to compile a minimal example with nvcc:
https://doc.cgal.org/latest/Triangulation_3/Triangulation_3_2copy_triangulation_3_8cpp-example.html
and I get the same errors I got with my own code:
https://gist.github.com/pgbarletta/659e9fe1ca01c45e8e29ea8e03ea0fcc

As long as the source file is named “.cpp" everything works fine, but if I change the ending to ".cu”, these template errors pop out.

I know this is lot to look at, but I’m just wondering: is anyone aware if nvcc is having some difficulties with templates?

$nvcc --version
Cuda compilation tools, release 9.0, V9.0.176

errors when compiling as:

nvcc as.cu -lCGAL -lmpfr -lgmp -lboost_thread

everything ok when compiling like:

nvcc as.cpp -lCGAL -lmpfr -lgmp -lboost_thread

dmitrys98ha · August 30, 2018, 8:09pm

I didn’t look carefully through your error output so, I may take a wild guess that CGAL uses many c++14 features. To enable those, you can try

nvcc -std=c++14 as.cu -lCGAL -lmpfr -lgmp -lboost_thread

Apologies if it is not helpful.

pbarletta · August 31, 2018, 12:56pm

Oh, I forgot to say that I tried that too. In the end I just reordered my code, so these libraries only showed up in *cpp files.
Tahnks!

FangQ · September 1, 2018, 10:11pm

simple template does work fine for all nvcc’s that I’ve used,

github.com

fangq/mcx/blob/master/src/mcx_core.cu#L972


			      r=1.f-2.f*rand_uniform01(t);
			      s=1.f-2.f*rand_uniform01(t);
			      q=sqrt(1.f-v->x*v->x-v->y*v->y)*(rand_uniform01(t)>0.5f ? 1.f : -1.f);
			      *((float4*)v)=float4(v->y*q-v->z*s,v->z*r-v->x*q,v->x*s-v->y*r,v->nscat);
		      }
                      *rv=float3(rv->x+(gcfg->srcparam1.x)*0.5f,
		                 rv->y+(gcfg->srcparam1.y)*0.5f,
				 rv->z+(gcfg->srcparam1.z)*0.5f);
                      canfocus=(gcfg->srctype==MCX_SRC_SLIT);
		      break;
		}
	  }
          /**
           * If beam focus is set, determine the incident angle
           */
          if(canfocus){
	    if(isnan(gcfg->c0.w)){ // isotropic if focal length is nan
                float ang,stheta,ctheta,sphi,cphi;
                ang=TWO_PI*rand_uniform01(t); //next arimuth angle
                sincosf(ang,&sphi,&cphi);
                ang=acosf(2.f*rand_uniform01(t)-1.f); //sine distribution

but I know CGAL code is heavy on template, especially latest standards, portability has always been an issue. try an old CGAL release perhaps.