Errors Using Sandybridge (AVX) enabled OpenBLAS with UMFPACK

[grisuthedragon]

I detect a strange behavior with OpenBLAS 0.2.6 and the current development branch in combination with complex arithmetics in UMFPACK. I provide a minimal working example at http://pastebin.com/098eLzTZ

I compiled OpenBLAS using

  make USE_OPENMP=1 

then I compiled UMFPACK 5.6.1 / Suitesparse 4.2.1 without any changes in their default setting accept of disabling METIS and compiled my example by:

gcc -o umfpack_openblas_error -I ../Include -I ../../SuiteSparse_config/ -I ../../AMD/Include  umfpack_openblas_error.c ../Lib/libumfpack.a ../../CHOLMOD/Lib/libcholmod.a ../../COLAMD/Lib/libcolamd.a  ../../AMD/Lib/libamd.a ../../SuiteSparse_config/libsuitesparseconfig.a  /path/to/libopenblas.a -lgomp -lgfortran -l

from inside the UMFPACK/Demo folder. I have chosen the vector b such that the solution must contain only 1's in every component, but I get

x [0] = 0.999998 + 1.49251e-08
x [1] = 0.999995 + 4.04891e-08
x [2] = 0.999987 + 9.50786e-08
x [3] = 0.999966 + 2.1776e-07
x [4] = 0.999916 + 4.83075e-07
...

If I recompile OpenBLAS with NO_AVX=1, the error is gone and I get the correct result:

x [0] = 1 + 1.57946e-20
x [1] = 1 + 1.7528e-20
x [2] = 1 + -4.79149e-20
x [3] = 1 + -7.36116e-20
x [4] = 1 + -1.26716e-19
...

So it seems that there must be an error inside the OpenBLAS AVX code and it is reproduceable with other complex data. The real version does not have this problem.

Used Hardware:

• Xeon E3-1245

Software: Ubuntu 12.04, 64bit. GCC 4.6.3

[xianyi]

Hi @grisuthedragon ,

Thank you for the report.

Could you try to disable multi-threading in OpenBLAS. For example, build OpenBLAS with USE_THREAD=0

Xianyi

[grisuthedragon]

Hi @xianyi
I compiled OpenBLAS again with USE_THREAD=0 and I recognized exactly the same behavoir. The variant employing AVX is wrong and turning it off gives me the right result.

Additionally I found out, that umfpack only relies on the following blas symbols:

• zgemm_
• zgemv_
• zgeru_
• ztrsm_
• ztrsv_
• dgemm_
• dgemv_
• dger_dtrsm_
• dtrsv_

where obviously the real(double) case seems to work correctly.

[xianyi]

I think this is a bug in zgemm or ztrsm. @wangqian will debug it.

[wangqian]

Hi @grisuthedragon,
I'm sorry that I didn't reply timely. I compiled the UMFPACK 5.6.1 / Suitesparse 4.2.1 packages using make CC=gcc, and link to the OpenBLAS library. However, I could open the link http://pastebin.com/098eLzTZ of the provided minimal working example .
Could you provide me a other link or a copy of your working example? I'll try to debug this error as soon as possible.
Thanks very much.
Qian

[xianyi]

Hi Qian,

I will send you the example

Xianyi

2013/7/17 wangqian notifications@github.com

Hi @grisuthedragon https://github.com/grisuthedragon,
I'm sorry that I didn't reply timely. I compiled the UMFPACK 5.6.1 /
Suitesparse 4.2.1 packages using make CC=gcc, and link to the OpenBLAS
library. However, I could open the link http://pastebin.com/098eLzTZ of
the provided minimal working example .
Could you provide me a other link or a copy of your working example? I'll
try to debug this error as soon as possible.
Thanks very much.
Qian

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/237#issuecomment-21099185
.

[grisuthedragon]

Hi @wangqian
the working example is exactly the same but is uses OpenBLAS compiled with the NO_AVX=1 flag. All surrounding things are the same. Or use the netlib blas implementations this you give the right result too.

regards

[wangqian]

Hi @grisuthedragon,
I think I fixed this bug on develop branch. Could you test it?
Qian

[grisuthedragon]

Sorry for the late reply, I was out of office for two weeks. I tested the current development branch with the minimal example and other examples and it seems to work correctly. Thanks for you work.