Problems with using arm version on Android

[kurak38]

Hi,
I am using Download openblas-0.2.9-rc2-armv6.tar.gz (8.1 MB) (http://sourceforge.net/projects/openblas/files/v0.2.9-rc2-arm/) and I have problem with undefined references to:
shmget, shmat, shmdt, shctl, get_nprocs, pthread_atfork, stderr and _gfortran_compare_string.
I am using APP_STL := gnustl_static with LDLIBS : libgnustl_static.a and libsupc++.a. GNU STL from android-ndk-r9d version 4.8 of gcc.
The most problematic, I think, would be function with shared memory, while on Android there is no such thing to use.
Please give me any advice
Thanks in advance
Adam

[xianyi]

Hi @kurak38 ,

Sorry, we didn't provide 0.2.9.rc2 version for Android so far.

Please try our old version.
http://sourceforge.net/projects/openblas/files/v0.2.8-arm/openblas-v0.2.8-android-rc1.tar.gz/download

It's armv5tel ( single thread and without lapack ).

Xianyi

[kurak38]

hmm, but there is no lapacke.h and I need it

[xianyi]

@kurak38, Why do you need BLAS and LAPACK on Android?

[xianyi]

@kurak38, Why do you need BLAS and LAPACK on Android?

[kurak38]

I'm porting some other library that requires both

[xianyi]

LAPACKE is a C wrapper for LAPACK, which also calls LAPACK fortran code. Thus, I thinks it need a Fortran compiler for Android.
I found a blog about Fortran on Android.
http://danilogiulianelli.blogspot.com/2013/02/how-to-build-gcc-fortran-cross-compiler.html

[kurak38]

Is it possible that you send me your source code? (kurak38@gmail.com)

[xianyi]

@kurak38 , I think the source code is on github.com.

We didn't have other codes for ARM.

[soumith]

Interesting. i'd definitely want to cross-compile for android. We use it for running neural nets on Android.

[kurak38]

I have managed to compile it by taking advantage of https://github.com/simonlynen/android_libs to build only lapack. Then add it to openBlas for android mention before and it worked.

[soumith]

@xianyi Are there instructions to compile the latest HEAD on android? or is it too complicated

[soumith]

@xianyi I ask, because I want to compile for android with multithreads, rather than just a single-threaded version

[xianyi]

In OpenBLAS/GotoBLAS, it used some System V IPC functions in multi-threading implementation, which is not supported in Android NDK.
Thus, we need to eliminate these IPC callings.

@wernsaar , any suggestions?

[wernsaar]

On 02.05.2014 11:53, Zhang Xianyi wrote:

In OpenBLAS/GotoBLAS, it used some System V IPC functions in multi-threading implementation, which is not supported in Android NDK.
Thus, we need to eliminate these IPC callings.

@wernsaar , any suggestions?

---

Reply to this email directly or view it on GitHub:
#363 (comment)
Hi,

Android does not support System V IPC in the kernel, but has a mechanism
called ashmem. First we need a configuration option like
TARGET_OS = OS_ANDROID, where OS_ANDROID is a subtype of
OS_LINUX. Then we can implement ashmem in memory.c and some other
files.

Best regards

Werner