Re: cuda compilation

[Ralf Wildenhues]

Hello Tomas,

* Tomas Oberhuber wrote on Tue, Dec 29, 2009 at 07:25:36PM CET:
> I am experiencing some troubles with cuda. I studied mailing lists about 
> autoconf, automake and libtool and from what I have read

> I have understood that from the linking point of view nvcc and gcc are 
> compatible. So one should only compile .cu files with nvcc and then link all 
> together with gcc. To compile .cu files with nvcc I did this in my 
> Makefile.am
> 
> include  ${top_srcdir}/cuda.mk
> 
> where cuda.mk is
> 
> SUFFIXES = .cu
> .cu.$(OBJEXT):
>       $(NVCC) -c $(DEFS) $(DEFAULT_INCLUDES) $(INCLUDES) $(AM_CPPFLAGS) $ 
> (CPPFLAGS) $(AM_CFLAGS) $(CFLAGS) -o $@ $<
> 
> libcore_tests_sources = tnlLongVectorCUDATester.cu \
>                         tnlLongVectorCUDATester.h
> 
> check_LTLIBRARIES = libcore-tests.la
> libcore_tests_la_SOURCES = $(libcore_tests_sources)

The .cu.$(OBJEXT) rule to create objects from sources only works for
objects not created with libtool.  For the latter, you'd need a .cu.lo
rule, and you also need libtool to know about nvcc.  The support that
was added to git Libtool in commit 741a9867e only works when $CC is set
to nvcc, so at this point your best bet might be to either use
  ./configure CC=nvcc

and compile all your sources with nvcc.

Alternatively, you could write a .cu.lo rule that looks like the
automake-generated .c.lo rule, has --tag=CC but uses $(NVCC); you'd then
still need a nvcc-wrapper that translates '-fPIC' to '-Xcompiler -fPIC'
for nvcc.  Ugly, yes, but I'm not sure how to do this any nicer at the
moment.

Cheers,
Ralf