Intel® Math Kernel Library

memory increasing problem calling pardiso!

here is the exact description of my problem: In loops(up to more than 2,000,000),i need to call MKL Pardiso to calculate Ax=b, but when running my program, i find the memory is increasing all the time till the program has to stop.

Incidentally, i use MKL IA-32, static MKL, three libraries: libguide.lib mkl_c.lib mkl_solver.lib are linked.

mkl_custom.dll conflicting with DirectX libraries ?

I'm calling mkl_custom.dll (ia32, cdecl, sequential, predefined lapack exports) from within a C# class library via PInvoke. Everything works fine so far.
As soon as my application also loads the managed DirectX dependencies (loaded from within Direct3D libraries) and creates some plots with them, calling MKL functions afterwards give strange results.
I tested and confirmed this reproducable for 'dsyevr' with jobz = 'V'. On return, the info parameter gives -6, which means "lda - parameter was wrong" - but it's not.

Paths of MKL/icc on Mac OS X


I have (hopefully) a simple question.

I have just installed MKL (m_mkl_p_10.0.3.020) on my Mac Pro, but the header files and the libraries are not showing up at the usual convenient site (/usr/local/lib, or even /opt/intel/), but are in directories such as:

/Library/Frameworks/Intel_MKL.framework/Versions/Current/Libraries/em64t/ ... etc.

BUG for in-place forward DFTs ??

The following simple program does not produce the correct results. It's behavior actually seems to be some times correct, most of the times wrong, when you run it several times, consecutively!

It is a simple forward transform of two identical sets of datas (DFTI_NUMBER_OF_TRANSFORMS is > 1) stored consecutively in the same 1D array (TIME_DATA). The two data sets (8 time points each) correspond to a simple cosine signal.

The forward transform for thefirst data set is always properly calculated. The second transform is wrong most of the time.

pardiso crashed with no memory when only processing 1 million * 1 million sparse matrix


I use pardiso to solve a million cells problem, i.e. the matrix size is only 1 million * 1million, the nozero number is only 3.9 million numbers, in my pc with 2G RAM.

Unfortunately, pardiso crashes during LU factoration. Pardiso tells me during reordering in partition metis it produces nearly 800 million nozeros in L+U.

It seems to be a little ridiculous, right? In LU0 factoration, I think the nozero number should not exceed the original matrix's nozero.

Intel® Math Kernel Library abonnieren