Re: [Starpu-devel] [MACOSX] problem with examples

[tog <guillaume.alleon@gmail.com>]

Ok, this is better. We are now in the problem I was anticipating earlier ...
mpicc is generating 64 bits while nvcc is generationg 32 bits

Is there a way to trigger nvcc compiler flags ?

My automake version:
oggy:oce alleon$ automake --version
automake (GNU automake) 1.9.6
Written by Tom Tromey <tromey@redhat.com>.

Best Regards
Guillaume

....
  CCLD   tests/mpi_irecv
  CC     mpi_isend_detached.o
  CCLD   tests/mpi_isend_detached
  CC     mpi_irecv_detached.o
  CCLD   tests/mpi_irecv_detached
  CC     mpi_detached_tag.o
  CCLD   tests/mpi_detached_tag
  CC     ring.o
/usr/local/cuda/bin/nvcc tests/ring_kernel.cu -c -o
tests/ring_kernel.o --compiler-options -fno-strict-aliasing
-I/usr/local/cuda/include -I../include/ -I../include  -I../include/
-I../include/
  CCLD   tests/ring
ld: warning: in tests/ring_kernel.o, file was built for i386 which is
not the architecture being linked (x86_64)
Undefined symbols:
  "_increment_cuda", referenced from:
      _increment_cl in ring.o
ld: symbol(s) not found
collect2: ld returned 1 exit status
make[2]: *** [tests/ring] Error 1
make[1]: *** [all] Error 2
make: *** [all-recursive] Error 1




On Thu, Apr 7, 2011 at 3:17 PM, Nathalie Furmento
<nathalie.furmento@labri.fr> wrote:
> I tried with autoconf 2.63, and i was not able to reproduce the problem.
> Which version of automake are you using ?
>
> There might be another solution to allow you to use StarPU. You could --
> instead of using the Subversion code -- use the latest nightly snapshot
> available here :
>
> http://starpu.gforge.inria.fr/testing/
>
> It means you do not  need to run autogen.sh, just run configure and make,
> and you will get the Makefiles as they have been generated on our system.
>
> Regards,
>
> Nathalie
>
> On 07/04/2011 11:05, tog wrote:
>>
>> Hi,
>>
>> I am using autotools 2.63.
>>
>> I have applied your second patch but I am sorry to report that I am
>> still having a problem:
>>
>> ...
>>
>> libtool: link: /usr/bin/mpicc -Wall -I/usr/local/cuda/include -O3
>> -gdwarf-2 -g3 -o tests/.libs/mpi_detached_tag mpi_detached_tag.o
>> -L/usr/local/cuda/lib ./.libs/libstarpumpi.dylib
>> /Users/alleon/PROJECTS/starpu/src/.libs/libstarpu.dylib
>> ../src/.libs/libstarpu.dylib -lm -lcudart -lcublas -lcuda -lcufft
>> -lgfortran -lcurand -lstdc++ -lblas -lpthread
>> if /usr/bin/mpicc -DHAVE_CONFIG_H -I. -I. -I../src/common -I../include
>> -I../gcc-plugin/src  -I../include/ -I../mpi/ -I../src/  -I../examples/
>> -I../src -I../include -D_GNU_SOURCE    -Wall -I/usr/local/cuda/include
>>  -O3  -gdwarf-2 -g3  -MT ring.o -MD -MP -MF ".deps/ring.Tpo" -c -o
>> ring.o `test -f 'tests/ring.c' || echo './'`tests/ring.c; \
>>        then mv -f ".deps/ring.Tpo" ".deps/ring.Po"; else rm -f
>> ".deps/ring.Tpo"; exit 1; fi
>> make[2]: *** No rule to make target `ring_kernel.o', needed by
>> `tests/ring'.  Stop.
>> make[1]: *** [all] Error 2
>> make: *** [all-recursive] Error 1
>>
>>
>> Regards
>> Guillaume
>>
>> On Thu, Apr 7, 2011 at 2:05 PM, Nathalie Furmento
>> <nathalie.furmento@labri.fr>  wrote:
>>
>>> Hi,
>>>
>>> I looked at your Makefile, autotools on your machine clearly has a
>>> different
>>> behaviour when handling object files. I am not  sure if the problem comes
>>> from the way we wrote mpi/Makefile.am or from autotools. In the meantime,
>>> could you please try the attached patch (after reverting the previous
>>> ones).
>>>
>>> Thanks,
>>>
>>> Nathalie
>>>
>>>
>>> On 06/04/2011 16:50, tog wrote:
>>>
>>>> Here is the file.
>>>> Best Regards
>>>> Guillaume
>>>>
>>>> On Wed, Apr 6, 2011 at 6:31 PM, Nathalie Furmento
>>>> <nathalie.furmento@labri.fr>    wrote:
>>>>
>>>>
>>>>> I think i just overlooked the fact it is looking for ring_kernel.o
>>>>> instead
>>>>> of tests/ring_kernel.o
>>>>>
>>>>> Could you please send me your build/mpi/Makefile file ?
>>>>>
>>>>> Regards,
>>>>>
>>>>> Nathalie
>>>>>
>>>>> On 06/04/2011 14:53, tog wrote:
>>>>>
>>>>>
>>>>>> I know from make -C examples that we will have 64 bits related issues.
>>>>>>
>>>>>> I did apply your patch and have the following behavior:
>>>>>>
>>>>>> obelix:starpu alleon$ VERBOSE=1 make -C mpi tests/ring_kernel.o
>>>>>> /usr/local/cuda/bin/nvcc tests/ring_kernel.cu -c -o
>>>>>> tests/ring_kernel.o --compiler-options -fno-strict-aliasing
>>>>>> -I/usr/local/cuda/include -I../include/ -I../include  -I../include/
>>>>>> -I../include/
>>>>>>
>>>>>> which is fine but ...
>>>>>>
>>>>>> obelix:starpu alleon$ VERBOSE=1 make
>>>>>> Making all in src
>>>>>> make[2]: Nothing to be done for `all-am'.
>>>>>> Making all in mpi
>>>>>> make  all-am
>>>>>> make[2]: *** No rule to make target `ring_kernel.o', needed by
>>>>>> `tests/ring'.  Stop.
>>>>>> make[1]: *** [all] Error 2
>>>>>> make: *** [all-recursive] Error 1
>>>>>> obelix:starpu alleon$
>>>>>>
>>>>>>
>>>>>> If your machine is running Snow Leopard the pressing the 6&      4
>>>>>> keys at
>>>>>> boot time will load the 64 bits kernel extensions.
>>>>>>
>>>>>> Best Regards
>>>>>> Guillaume
>>>>>>
>>>>>>
>>>>>> On Wed, Apr 6, 2011 at 6:03 PM, Nathalie Furmento
>>>>>> <nathalie.furmento@labri.fr>      wrote:
>>>>>>
>>>>>>
>>>>>>
>>>>>>> Could you please try to recompile with the attached patch. If it is
>>>>>>> still
>>>>>>> failing, could you then try from your build directory to start:
>>>>>>>
>>>>>>> % make -C mpi V=1 tests/ring_kernel.o
>>>>>>>
>>>>>>> And then run make again.
>>>>>>>
>>>>>>> If all of that fails, please send me the file build/mpi/Makefile
>>>>>>>
>>>>>>> Concerning OpenCL, configure looks for it in different locations
>>>>>>> ("/usr/local/cuda" "/c/cuda" "/cygdrive/c/cuda" "/opt/cuda"). If your
>>>>>>> installation is in another directory, you can specify
>>>>>>> --with-opencl-dir=<directory>      when calling configure.
>>>>>>>
>>>>>>> I do not think your problem is related to 64 bits. I do not have
>>>>>>> right
>>>>>>> now a
>>>>>>> 64 bits MacOS machine to start. I should be able to get one shortly.
>>>>>>>
>>>>>>> Regards,
>>>>>>>
>>>>>>> Nathalie
>>>>>>>
>>>>>>> On 06/04/2011 14:06, tog wrote:
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>> Hi,
>>>>>>>>
>>>>>>>> Few things that might be different on my machine. It is running 64
>>>>>>>> bits i.e.  in the menu About This Mac ->        More info ... ->
>>>>>>>>  Software, I
>>>>>>>> have
>>>>>>>>   64-bit Kernel and Extensions:        Yes
>>>>>>>>
>>>>>>>> obelix:starpu alleon$ uname -a
>>>>>>>> Darwin oggy.lan 10.7.0 Darwin Kernel Version 10.7.0: Sat Jan 29
>>>>>>>> 15:16:10 PST 2011; root:xnu-1504.9.37~1/RELEASE_X86_64 x86_64
>>>>>>>>
>>>>>>>>
>>>>>>>> Then the answers to your questions:
>>>>>>>>
>>>>>>>> obelix:starpu alleon$ ls -lrt mpi/tests/ring_kernel.cu
>>>>>>>> -rw-r--r--  1 alleon  staff  1125 Mar 29 14:14
>>>>>>>> mpi/tests/ring_kernel.cu
>>>>>>>>
>>>>>>>>
>>>>>>>> The output is attached.
>>>>>>>>
>>>>>>>> Do you know how to enable OpenCL on the Mac ?
>>>>>>>>
>>>>>>>> Best Regards
>>>>>>>> Guillaume
>>>>>>>>
>>>>>>>> On Wed, Apr 6, 2011 at 5:09 PM, Nathalie Furmento
>>>>>>>> <nathalie.furmento@labri.fr>        wrote:
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>> Hi,
>>>>>>>>>
>>>>>>>>> I tested on a MacOS with mpi and cuda, it is all working for me.
>>>>>>>>>
>>>>>>>>> Could you please check the file mpi/tests/ring_kernel.cu is
>>>>>>>>> available.
>>>>>>>>> And
>>>>>>>>> could you also please send the output of "make V=1".
>>>>>>>>>
>>>>>>>>> Thanks,
>>>>>>>>>
>>>>>>>>> Nathalie
>>>>>>>>>
>>>>>>>>> On 05/04/2011 13:51, tog wrote:
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>> Hi,
>>>>>>>>>>
>>>>>>>>>> Yes similar problems are there.
>>>>>>>>>> For examples, the mismatch between i386 and x86_64 can be solved
>>>>>>>>>> by
>>>>>>>>>> adding the -m64 flag to nvcc (looks like mpicc generates 64 bits
>>>>>>>>>> objects, white the default for nvcc is 32 bits).
>>>>>>>>>> It might also be solved by adding -m32 to mpicc (not tested) -
>>>>>>>>>> however
>>>>>>>>>> my machine is running 64 bits.
>>>>>>>>>>
>>>>>>>>>> Best Regards
>>>>>>>>>> Guillaume
>>>>>>>>>>
>>>>>>>>>> On Tue, Apr 5, 2011 at 4:30 PM, Nathalie Furmento
>>>>>>>>>> <nathalie.furmento@labri.fr>          wrote:
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>> Hi,
>>>>>>>>>>>
>>>>>>>>>>> ring_kernel.o is build from ring_kernel.cu. The rule to build a
>>>>>>>>>>> .o
>>>>>>>>>>> from
>>>>>>>>>>> a
>>>>>>>>>>> .cu is in Makefile.am.
>>>>>>>>>>>
>>>>>>>>>>> Could you see if you also have the problems when building the
>>>>>>>>>>> tests
>>>>>>>>>>> or
>>>>>>>>>>> the
>>>>>>>>>>> examples (make -C tests or make -C examples)
>>>>>>>>>>>
>>>>>>>>>>> Thanks,
>>>>>>>>>>>
>>>>>>>>>>> Nathalie
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>> On 05/04/2011 12:35, tog wrote:
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>> Hi,
>>>>>>>>>>>>
>>>>>>>>>>>> I still have a problem with the build in the mpi tests.
>>>>>>>>>>>>
>>>>>>>>>>>> if /usr/bin/mpicc -DHAVE_CONFIG_H -I. -I. -I../src/common
>>>>>>>>>>>> -I../include
>>>>>>>>>>>> -I../gcc-plugin/src  -I../include/ -I../mpi/ -I../src/
>>>>>>>>>>>>  -I../examples/
>>>>>>>>>>>> -I../src -I../include -D_GNU_SOURCE    -Wall
>>>>>>>>>>>> -I/usr/local/cuda/include
>>>>>>>>>>>>  -O3  -gdwarf-2 -g3  -MT ring.o -MD -MP -MF ".deps/ring.Tpo" -c
>>>>>>>>>>>> -o
>>>>>>>>>>>> ring.o `test -f 'tests/ring.c' || echo './'`tests/ring.c; \
>>>>>>>>>>>>        then mv -f ".deps/ring.Tpo" ".deps/ring.Po"; else rm -f
>>>>>>>>>>>> ".deps/ring.Tpo"; exit 1; fi
>>>>>>>>>>>> make[2]: *** No rule to make target `ring_kernel.o', needed by
>>>>>>>>>>>> `tests/ring'.  Stop.
>>>>>>>>>>>> make[1]: *** [all] Error 2
>>>>>>>>>>>> make: *** [all-recursive] Error 1
>>>>>>>>>>>>
>>>>>>>>>>>>
>>>>>>>>>>>> Any idea of what is wrong ?
>>>>>>>>>>>>
>>>>>>>>>>>> Best Regards
>>>>>>>>>>>> Guillaume
>>>>>>>>>>>>
>>>>>>>>>>>>
>>>>>>>>>>>>
>>>>>>>>>>>>
>>>>>>>>>>>>
>>>>>>>>>>>>
>>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>
>>>>>
>>>>
>>>>
>>>>
>>>
>>>
>>
>>
>>
>
>



-- 
PGP KeyID: 2048R/EA31CFC9  subkeys.pgp.net