MAGMA installation for mac 10.9 with NVIDIA GPU

[hirokiH]

Hi

Recently, I tried installing MAGMA on my mac OS 10.9, with NVIDIA GeForce GT 650M.
However, the installation failed.
I have few experience of installing libraries, so I really appreciate if you can help me.

The error says:
ld: symbol(s) not found for architecture x86_64
collect2: error: ld returned 1 exit status
make[1]: *** [testing_z_cublas_v2] Error 1
make: *** [test] Error 2

This is what I did:
-- installed LAPACK, ATLAS, and CUDA.
-- downloaded magma-1.4.1
-- copy make.inc.atlas into make.inc

This is how I modified in make.inc:
-- provided path to libraries
-- deleted -DMAGMA_SETAFFINITY and -lgfortran -lifcore (but I don't know if I should do this..)
-- changed -L$(CUDADIR)/lib64 into -L$(CUDADIR)/lib

make.inc looks like:
OPTS = -O3 -DADD_ -fopenmp
F77OPTS = -O3 -DADD_
FOPTS = -O3 -DADD_ -x f95-cpp-input
NVOPTS = -O3 -DADD_ -Xcompiler -fno-strict-aliasing
LDOPTS = -fopenmp

LIB = -llapack -lf77blas -latlas -lcblas -lcublas -lcudart -lstdc++ -lm

LAPACKDIR ?= /Users/hiroki/Desktop/lapack-3.5.0
ATLASDIR ?= /usr/local/atlas
CUDADIR ?= /usr/local/cuda
-include make.check-atlas
-include make.check-cuda

LIBDIR = -L$(LAPACKDIR) \
-L$(ATLASDIR)/lib \
-L$(CUDADIR)/lib

INC = -I$(CUDADIR)/include

I researched about the error, but honestly, I don't really understand.
Any help is much appreciated.

[mgates3]

It's okay to install LAPACK and ATLAS, but it is not needed. MacOS already comes with the BLAS and LAPACK in its Veclib (aka Accelerate) framework. See the make.inc.macos. Basically, add "-framework veclib" to LIB, instead of lapack, f77blas, cblas, and atlas.

MacOS supports both 32-bit and 64-bit binaries (and fat binaries that contain both). This error is saying it can't find some symbol (i.e. function) for 64-bit. So either you are missing some required library, or one of the libraries is 32-bit instead of 64-bit. Was there more to that error message, like *which* symbols it didn't find? That would tell what library it is missing. Or it should complain that a library is not for the architecture being linked if the library is 32-bit. CUDA and Veclib are both fat (32 and 64-bit), so you should be able to link either way.

HOWEVER, there is also a bug in Veclib that manifests itself in 64-bit, which causes some BLAS routines to return bad values. So I find it is easier to compile everything 32-bit, by adding -m32 to the various OPTS. See make.inc.macos.

-mark

[hirokiH]

Thank you for your quick reply Mark.

Okay, I understood that I can use make.inc.macos also.
But currently, I am still trying to use LAPACK and ATLAS.
I will try this way for a little bit more, and if this doesn't work, I will use make.inc.macos.
Yes there is a following error message in addition to "symbol(s) not found for architecture x86_64"

Undefined symbols for architecture x86_64:
"__gfortran_compare_string", referenced from:
_ilaenv_ in liblapack.a(ilaenv.o)
"__gfortran_st_write", referenced from:
_xerbla_ in liblapack.a(xerbla.o)
"__gfortran_st_write_done", referenced from:
_xerbla_ in liblapack.a(xerbla.o)
"__gfortran_stop_string", referenced from:
_xerbla_ in liblapack.a(xerbla.o)
"__gfortran_string_len_trim", referenced from:
_xerbla_ in liblapack.a(xerbla.o)
"__gfortran_transfer_character_write", referenced from:
_xerbla_ in liblapack.a(xerbla.o)
"__gfortran_transfer_integer_write", referenced from:
_xerbla_ in liblapack.a(xerbla.o)
"_cublasDzasum", referenced from:
_magma_dzasum in libmagma.a(blas_z.o)
"_cublasDznrm2", referenced from:
_magma_dznrm2 in libmagma.a(blas_z.o)
"_cublasIzamax", referenced from:
_magma_izamax in libmagma.a(blas_z.o)
"_cublasIzamin", referenced from:
_magma_izamin in libmagma.a(blas_z.o)
"_cublasSetKernelStream", referenced from:
_magmablasSetKernelStream in libmagma.a(stream.cu_o)
"_cublasZaxpy", referenced from:
_magma_zaxpy in libmagma.a(blas_z.o)
"_cublasZcopy", referenced from:
_magma_zcopy in libmagma.a(blas_z.o)
"_cublasZdotc", referenced from:
_magma_zdotc in libmagma.a(blas_z.o)
"_cublasZdotu", referenced from:
_magma_zdotu in libmagma.a(blas_z.o)
"_cublasZdrot", referenced from:
_magma_zdrot in libmagma.a(blas_z.o)
"_cublasZdscal", referenced from:
_magma_zdscal in libmagma.a(blas_z.o)
"_cublasZgemm", referenced from:
_magma_zgemm in libmagma.a(blas_z.o)
"_cublasZgemv", referenced from:
_magma_zgemv in libmagma.a(blas_z.o)
"_cublasZgerc", referenced from:
_magma_zgerc in libmagma.a(blas_z.o)
"_cublasZgeru", referenced from:
_magma_zgeru in libmagma.a(blas_z.o)
"_cublasZhemm", referenced from:
_magma_zhemm in libmagma.a(blas_z.o)
"_cublasZhemv", referenced from:
_magma_zhemv in libmagma.a(blas_z.o)
"_cublasZher", referenced from:
_magma_zher in libmagma.a(blas_z.o)
"_cublasZher2", referenced from:
_magma_zher2 in libmagma.a(blas_z.o)
"_cublasZher2k", referenced from:
_magma_zher2k in libmagma.a(blas_z.o)
"_cublasZherk", referenced from:
_magma_zherk in libmagma.a(blas_z.o)
"_cublasZrot", referenced from:
_magma_zrot in libmagma.a(blas_z.o)
"_cublasZscal", referenced from:
_magma_zscal in libmagma.a(blas_z.o)
"_cublasZswap", referenced from:
_magma_zswap in libmagma.a(blas_z.o)
"_cublasZsymm", referenced from:
_magma_zsymm in libmagma.a(blas_z.o)
"_cublasZsyr2k", referenced from:
_magma_zsyr2k in libmagma.a(blas_z.o)
"_cublasZsyrk", referenced from:
_magma_zsyrk in libmagma.a(blas_z.o)
"_cublasZtrmm", referenced from:
_magma_ztrmm in libmagma.a(blas_z.o)
"_cublasZtrmv", referenced from:
_magma_ztrmv in libmagma.a(blas_z.o)
"_cublasZtrsm", referenced from:
_magma_ztrsm in libmagma.a(blas_z.o)
"_cublasZtrsv", referenced from:
_magma_ztrsv in libmagma.a(blas_z.o)

Am I missing any libraries? or does this message say these functions above are only for 32 bits?

[mgates3]

You need to add -lgfortran for __gfortran_compare_string, etc.

The _cublasDzasum, etc. should be in -lcublas, which provides both 32-bit and 64-bit versions. Unless you provided the wrong path to cuda, I don't know why that would not be found.

-mark

[hirokiH]

Thank you very much for your help, Mark.

Currently, I have this error:
ld: library not found for -lgfortran

I already installed gfortran.
I would like to know which version of gfortran is required for MAGMA.

Thank you

[mgates3]

Can you find the libgfortran.a or libgfortran.so library file? That's what -lgfortran refers to. If it isn't in a standard library path, you need to add its directory to LIBDIR in make.inc. For instance, if it was installed in

/mnt/gfortran/lib/libgfortran.a

then add that directory to LIBDIR, something like:

LIBDIR = -L$(MKLROOT)/lib/intel64 \
-L$(CUDADIR)/lib64 \
-L/mnt/gfortran/lib

MAGMA should work with any reasonably recent version of gfortran. That is, don't ask about a 10 year old version, but something even a couple years old should work.

Note it is not MAGMA that needs -lgfortran, but your LAPACK library that needs it. With Veclib, it is not needed.
-mark

[clintpatty]

Have you tried this yet with Fortran, such as gfortran linked here:

http://hpc.sourceforge.net

I would like to get it working on the newer MacBook Pro with CUDA-capable graphics and OpenCL with the coprocessor, so I would like to compare the NVIDIA vs Iris Pro, but I cannot tell from reading if the OpenCL is only supporting AMD or will also support any OpenCL-capable device. Thanks.

[mgates3]

clintpatty,

I would recommend using cuda MAGMA 1.5, if you have a CUDA GPU.

However, it is possible to use clMAGMA and AMD's clBLAS with general OpenCL GPUs, not just AMD GPUs. Except the Iris Pro doesn't support double-precision, and it seems clBLAS requires double-precision. At least, on my Mac laptop I have to initialize the OpenCL context with only the NVIDIA card, skipping the Iris Pro card, to get clBLAS to run. Otherwise it crashes with a "Trace/BPT trap" error.

We are updating clMAGMA now, and using it on MacOS should be much easier with the next release, expected in November.

-mark