Error compiling MAGMA 1.5.0 beta 2 in Debian Sid

[jgpallero]

Hello,

I have a Debian Sid (64 bits) installation with the GCC compilers 4.8 and 4.9 and CUDA 6.0. I'm trying to compile the MAGMA 1.5.0 beta 2 version and I obtain an error during make test.

First of all, as GCC 4.9 is not permitted in CUDA 6.0, I set the related variables in the make.inc file as

Code: Select all

CC        = gcc-4.8
NVCC      = nvcc --compiler-bindir /usr/bin/gcc-4.8
FORT      = gfortran-4.8

The make lib command works well, and libmagma.a is created in the lib/ folder. But in make test I obtain the error

Code: Select all

gcc-4.8 -fPIC -Xlinker -zmuldefs  testing_zhetrd_he2hb.o -o testing_zhetrd_he2hb \
libtest.a lin/liblapacktest.a -L../lib -lmagma \
-L/opt/openblas/lib -L/opt/cuda/lib64 -L/usr/lib64 \
-lopenblas  -lpthread -lcublas -lcudart -llapack -lm
/usr/bin/ld: ../lib/libmagma.a(zhetrd_hb2st.o): undefined reference to symbol '__gxx_personality_v0@@CXXABI_1.3'
//usr/lib/x86_64-linux-gnu/libstdc++.so.6: error adding symbols: DSO missing from command line
collect2: error: ld returned 1 exit status
Makefile:255: recipe for target 'testing_zhetrd_he2hb' failed
make[1]: *** [testing_zhetrd_he2hb] Error 1
make[1]: Leaving directory '/home/topo/Desktop/magma-1.5.0-beta2/testing'
Makefile:54: recipe for target 'test' failed
make: *** [test] Error 2

Searching in the web I've seen that the error __gxx_personality_v0@@CXXABI_1.3 is produced because I'm trying to compile C++ code with a C compiler, so I change

Code: Select all

CC        = gcc-4.8

by

Code: Select all

CC        = g++-4.8

in the make.inc file. This apparently solves the error, but a new one appears

Code: Select all

g++-4.8 -O3 -DADD_ -DHAVE_CUBLAS -DMIN_CUDA_ARCH=300 -DCUBLAS_GFORTRAN -I/opt/cuda/include -I../include -I../control -c /opt/cuda/src/fortran.c -o fortran.o
gfortran-4.8 -O3 -DADD_ -x f95-cpp-input -Dmagma_devptr_t="integer(kind=8)" -I/opt/cuda/include -I../include -I../control -c testing_zgetrf_gpu_f.F90 -o testing_zgetrf_gpu_f.o
gfortran-4.8 -fPIC -Xlinker -zmuldefs testing_zgetrf_gpu_f.o fortran.o -o testing_zgetrf_gpu_f \
libtest.a lin/liblapacktest.a -L../lib -lmagma \
-L/opt/openblas/lib -L/opt/cuda/lib64 -L/usr/lib64 \
-lopenblas  -lpthread -lcublas -lcudart -llapack -lm
/usr/bin/ld: ../lib/libmagma.a(ztrevc3_mt.o): undefined reference to symbol '__cxa_pure_virtual@@CXXABI_1.3'
//usr/lib/x86_64-linux-gnu/libstdc++.so.6: error adding symbols: DSO missing from command line
collect2: error: ld returned 1 exit status
Makefile:249: recipe for target 'testing_zgetrf_gpu_f' failed
make[1]: *** [testing_zgetrf_gpu_f] Error 1
make[1]: Leaving directory '/home/topo/Desktop/magma-1.5.0-beta2/testing'
Makefile:54: recipe for target 'test' failed
make: *** [test] Error 2

If I put the CC = g++-4.8 from the beginning of the compilation, i. e., I execute make lib and make test with CC = g++-4.8, I obtain also the last error.

Anyway, as libmagma.a was created, I've installed the library. But when I try to compile a program, I obtain

Code: Select all

In file included from /opt/magma/include/magmablas_z.h:14:0,
                 from /opt/magma/include/magmablas.h:12,
                 from /opt/magma/include/magma.h:15,
                 from /home/topo/Desktop/libgeoc/libgeoc/geocblas.h:72,
                 from geocblas.c:59:
/opt/magma/include/magma_types.h:160:6: error: #error "One of HAVE_CUBLAS, HAVE_clAmdBlas, or HAVE_MIC must be defined. This typically happens in Makefile.internal."
     #error "One of HAVE_CUBLAS, HAVE_clAmdBlas, or HAVE_MIC must be defined. This typically happens in Makefile.internal."
      ^

But apparently HAVE_CUBLAS was defined in the Makefile.internal (lines 61-52) as

Code: Select all

NVOPTS += -DHAVE_CUBLAS $(NV_SM) $(NV_COMP)
OPTS   += -DHAVE_CUBLAS -DMIN_CUDA_ARCH=$(MIN_ARCH)

Has anyone had a similar error?

Thanks

[hsahasra]

You need to define these preprocessor directives before you include magma.h, or while you compile your code:

#ifndef HAVE_CUBLAS
#define HAVE_CUBLAS
#endif

#ifndef ADD_
#define ADD_
#endif

#include "magma.h"

---or---

g++-4.8 -DHAVE_CUBLAS -DADD_

This is when you are using CUDA, and BLAS/LAPACK mangling adds an underscore.

[jgpallero]

You need to define these preprocessor directives before you include magma.h, or while you compile your code:

#ifndef HAVE_CUBLAS
#define HAVE_CUBLAS
#endif

#ifndef ADD_
#define ADD_
#endif

#include "magma.h"

---or---

g++-4.8 -DHAVE_CUBLAS -DADD_

This is when you are using CUDA, and BLAS/LAPACK mangling adds an underscore.

Mmmm, I don't know. In prior versions of MAGMA this was not neccessary

[mgates3]

Correct. For quite some time it has been required to define HAVE_CUBLAS when including magma.h. This is because very similar magma.h header files are used in CUDA MAGMA, OpenCL MAGMA, and Xeon Phi (MIC) MAGMA.

If you include magma_lapack.h, defining ADD_ has always been required. If you don't include the lapack functions, it shouldn't be required.

As for your compilation that fails, that particular case is using gfortran, so you probably need to add the C++ library to the link line, for instance -lstdc++.

Code: Select all

gfortran-4.8 -fPIC -Xlinker -zmuldefs testing_zgetrf_gpu_f.o fortran.o -o testing_zgetrf_gpu_f \
libtest.a lin/liblapacktest.a -L../lib -lmagma \
-L/opt/openblas/lib -L/opt/cuda/lib64 -L/usr/lib64 \
-lopenblas  -lpthread -lcublas -lcudart -llapack -lm \
-lstdc++

You can always just ignore the Fortran test cases (e.g., make -k) if you don't program in Fortran.

-mark

[jgpallero]

Thank you for your answer, Mark. And also thanks to hsahasra, who obviously was right and give a correct advice.