In the case of legacy projects, ask the developers to submit new code only if they reduce the number of findings.
In the case of coding from scratch, allow no findings before uploading new code in your repository.

When enabling the check (-diag-enable sc3) and compiling the code, a new folder will be created where the findings will be stored using a custom XML format.

$ file rXsc/data.X/rXsc.pdr
rXsc/data.X/rXsc.pdr: XML document text

$ xml sel -t -m /diags/diag -v "concat(message/thread/stacktrace/loc/file, ':', message/thread/stacktrace/loc/line, ':', sc_verbose)" -n rXsc/data.0/rXsc.pdr
/home/$USER/work/$PROD/src/pool.c:157:pool.c(157): warning #12178: this value of "ret" isn't used in the program
/home/$USER/work/$PROD/src/pool.c:186:pool.c(186): error #12192: unreachable statement
/home/$USER/work/$PROD/src/pool.c:216:pool.c(216): warning #12135: procedure "pool_done" is never caled

Forthcoming Intel® MKL 11.0 (Expected date of release: Autumn 2012)

•  Intel® MKL 11.0 will have backward incompatibility with Intel® MKL 10.2 update 3

•  Open MP static library on Microsoft* Windows

•  GMP* Arithmetic functions

•  Intel® MKL Reference Manual will be removed from product package

•  Intel® Pentium III will no longer be supported

•  Red Hat* EL4 and PGI* Fortran 10.x support will be dropped

•  PGI* Fortran 77 support will be removed

• a powerful N-dimensional array object
• sophisticated (broadcasting) functions
• tools for integrating C/C++ and Fortran code
• useful linear algebra, Fourier transform, and random number capabilities.

Besides its obvious scientific uses, NumPy can also be used as an efficient multi-dimensional container of generic data.

For more information on NumPy, please visit http://NumPy.scipy.org/

Version Information

This application note was created to help NumPy users to make use of the latest versions of Intel MKL on Linux platforms.

The instructions given in this articles apply to Intel MKL 10.3 and above and Intel Compiler 11.0 and above.

 
Step 2 - Downloading NumPy Source Code

The NumPy source code can be downloaded from:

http://NumPy.scipy.org/

Prerequisites

Intel MKL can be obtained from the following options:

Download a FREE evaluation version of the Intel MKL product.
Download the FREE non-commercial* version of the Intel MKL product.

All of these can be obtained at: Intel® Math Kernel Library product web page.

Intel® MKL is also bundled with the following products

Intel® Parallel Studio XE 2011
Intel Parallel Composer XE 2011
Intel Cluster Studio 2011

Step 3 - Configuration

Use the following commands to extract the NumPy tar files from the downloaded NumPy-x.x.x.tar.gz.

1.	$gunzip numpy-x.x.x.tar.gz 
2.	$tar -xvf numpy-x.x.x.tar 

 
The above will create a directory named numpy-x.x.x

Make sure that C++ and FORTRAN compilers are installed and they are in PATH. Also set LD_LIBRARY_PATH to your compiler (C++ and FORTRAN), and MKL libraries.

Step 4 - Building NumPy

Change directory to numpy-x.x.x
Create a site.cfg from the existing one
 
Edit site.cfg as follows:

Add the following lines to site.cfg in your top level NumPy directory to use Intel® MKL, if you are building on Intel 64 platform:

[mkl]
library_dirs = /opt/intel/composer_xe_2011_sp1.6.233/mkl/lib/intel64
include_dirs = /opt/intel/composer_xe_2011.sp1.6.233/mkl/include
mkl_libs = mkl_rt
lapack_libs =

If you are building NumPy for 32 bit, please add as the following

[mkl]
library_dirs = /opt/intel/composer_xe_2011_sp1.6.233/mkl/lib/ia32
include_dirs = /opt/intel/composer_xe_2011_sp1.6.233/mkl/include
mkl_libs = mkl_rt
lapack_libs = 

Modify cc_exe in numpy/distutils/intelccompiler.py to be something like:

self.cc_exe = 'icc -O3 -g -fPIC -fp-model strict -fomit-frame-pointer -openmp -xhost'

Here we use, -O3, optimizations for speed and enables more aggressive loop transformations such as Fusion, Block-Unroll-and-Jam, and collapsing IF statements, -openmp for OpenMP threading and -xhost option tells the compiler to generate instructions for the highest instruction set available on the compilation host processor. If you are using the ILP64 interface, please add -DMKL_ILP64 compiler flag.

Run icc --help for more information on processor-specific options, and refer Intel Compiler documentation for more details on the various compiler flags.

Compile and install NumPy with the Intel compiler: (on 64-bit platforms replace "intel" with "intelem")

python setup.py config --compiler=intel build_clib --compiler=intel build_ext --compiler=intel install

If you build NumPY for Intel64 bit platforms:

$export LD_LIBRARY_PATH=/opt/intel/composer_xe_2011_sp1.6.233/mkl/lib/intel64:/opt/intel/composer_xe_2011_sp1.6.233/compiler/lib/intel64:$LD_LIBRARY_PATH

If you build NumPY for ia32 bit platforms:

$export LD_LIBRARY_PATH=/opt/intel/composer_xe_2011_sp1.6.233/mkl/lib/ia32:/opt/intel/composer_xe_2011_sp1.6.233/compiler/lib/ia32:$LD_LIBRARY_PATH

It is possible that LD_LIBRARY_PATH causes a problem, if you have installed MKL and Composer XE in other directories than the standard ones. The only solution I've found that always works is to build Python, NumPy and SciPy inside an environment where you've set the LD_RUN_PATH variable, e.g:

$export LD_RUN_PATH=~/opt/lib:~/intel/composer_xe_2011_sp1.6.233/compiler/lib:~/intel/composer_xe_2011_sp1.6.233/mkl/lib/ia32

Note: We recommend users to use arrays with 'C' ordering style which is row-major, which is default than Fortran Style which is column-major, and this is because NumPy uses CBLAS and also to get better performance.

Appendex A: Example:

Please see below an example Python script for matrix multiplication that you can use Numply installed with Intel MKL which has been provided for illustration purpose.

import numpy as np
import time

N = 6000
M = 10000

k_list = [64, 80, 96, 104, 112, 120, 128, 144, 160, 176, 192, 200, 208, 224, 240, 256, 384]

def get_gflops(M, N, K):
	return M*N*(2.0*K-1.0) / 1000**3

np.show_config()

for K in k_list:
	a = np.array(np.random.random((M, N)), dtype=np.double, order='C', copy=False)
	b = np.array(np.random.random((N, K)), dtype=np.double, order='C', copy=False)
	A = np.matrix(a, dtype=np.double, copy=False)
	B = np.matrix(b, dtype=np.double, copy=False)

	C = A*B

	start = time.time()

	C = A*B
	C = A*B
	C = A*B
	C = A*B
	C = A*B

	end = time.time()

	tm = (end-start) / 5.0

	print "{0:4}, {1:9.7}, {2:9.7}".format(K, tm, get_gflops(M, N, K) / tm)

Appendix B: Performance Comparison

Please click Examples.py to download the examples for LU, Cholesky and SVD.



 

$ cat testcase.cpp
#include <pthread.h>
#include <stdio.h>

int counter=0;

class Guard {
     public:
     Guard() {
         ++counter;
     }

     ~Guard() {
         --counter;
     }
};

void *handler_with_guard(void *)
{
    Guard l;
    pthread_exit(NULL);
}

int main(int argc, char *argv[])
{
    pthread_t t;
    int result = pthread_create(&t, 0, handler_with_guard, 0);
    printf("pthread_create result = %d\n", result);

    result = pthread_join(t, NULL);
    printf("pthread_join result = %d\n", result);
    printf("counter = %d\n", counter);

    return 0;
}

$ icc -c testcase.cpp
$ icc -lpthread testcase.o
$ ./a.out
pthread_create result = 0
terminate called without an active exception
Aborted (core dumped)

$ icc -c -O0 testcase.cpp
$ icc -lpthread testcase.o
$ ./a.out
pthread_create result = 0
pthread_join result = 0
counter = 0

Below each error is described in details:

0: no error

-1: Input inconsistency

• Incorrect stage number for PARDISO was called.
• Incorrect PARDISO calling sequence, e.g. run stage > 1 without initialization results in error reporting.
• Incorrect number of matrices to be solved was set (PARDISO can be used for solving several matrices with the same sparsity structure at once, taking into account that their maximum number was defined previously. Setting the number of matrices to be solved outside the range of 1≤ … ≤maxfct results in error).
• PARDISO checks the parameters at each stage for consistency with the parameters at the previous stages. Every disagreement results in error reporting.

-2: not enough memory

This error value is returned in the case of any problem with memory allocation inside PARDISO.

PARDISO messages:

"*** Error in PARDISO memory allocation: [STRUCTURE NAME], size to allocate: %d bytes"

"total memory wanted here: %d kbyte"

"symbolic (max): %d symbolic (permanent): %d"

"real(including 1 factor): %d"

It describes the issue that arises on allocation of STRUCTURE_NAME inside PARDISO. A additional information about current memory usages is also printed.

-3: reordering problem

Returned for any problem on a reordering stage (phase 11)

PARDISO messages:

"*** error PARDISO: reordering, symbolic factorization"

-4: zero pivot, numerical factorization or iterative refinement problem

Let us start with a citation of the Intel MKL manual:

Using phase =33 results in an error message (error =4, should be -4 * ) if the stopping criteria for the Krylow-Subspace iteration cannot be reached.

If phase= 23, then the factors L, U are recomputed for the matrix A and the error flag error=0 in case of a successful factorization. If phase =33, then error = -4 signals the failure

If the solver detects a zero or negative pivot for these matrix types, the factorization is stopped, PARDISO returns immediately with an error (error = -4) and iparm(30) contains the number of the equation where the first zero or negative pivot is detected.

The error returned in the case of any problem at the factorization stage (phase 22) or at the iterative refinement stage of solution.

PARDISO messages:

"*** Error in PARDISO: cgs error iparam(20) %d" – prints iparm(20) – see CG / CGS diagnostics in the Intel MKL manual

"*** error PARDISO: iterative refinement"

" contraction rate is greater than 0.9, interrupt" – rate of contraction is too small

"*** error PARDISO: iterative refinement"

" exceeds max. iteration number %d" - prints abs(iparm(8))

"*** Error in PARDISO: internal error, insufficient memory factorization" – looks like a problem at the factorization stage except for pivoting issues (see errors below)

"*** Error in PARDISO: zero or negative pivot, A is not SPD-matrix" – original matrix (almost) not SPD one (probably due to computer arithmetic inaccuracies)

"*** Error in PARDISO: zero pivot" – the same as above but for other matrix types

-5: unclassified (internal) error

This error value is not used currently and is reserved for the future use.

-6: preordering failed (matrix types 11, 13 only)

This error value is returned in the case of any problem at the stage of preparation for reordering (in matching algorithm).

PARDISO messages:

"*** Error in PARDISO: preordering failed after %d neqns out of %d"

"structure singular or input/parameter problem (matrix type 11,13)"

Looks like the message provides no meaningful information.

-7: diagonal matrix problem

PARDISO prints no messages.

-8: 32-bit integer overflow problem

This error value is returned on 32-bit architecture for big matrices when indices become greater than the maximal integer value on this platform.

PARDISO messages:

"*** error PARDISO: reordering, symbolic factorization"

-9: not enough memory for OOC

Let us start with a citation of the Intel MKL manual:

Note that if iparm(60) is equal to 1 or 2, and the total peak memory needed for strong local arrays is more than MKL_PARDISO_OOC_MAX_CORE_SIZE, the program stops with error -9. In this case, increase of MKL_PARDISO_OOC_MAX_CORE_SIZE is recommended.

This error value is returned when amount of memory available for PARDISO (defined by MKL_PARDISO_OOC_MAX_CORE_SIZE, by default 2000 Mb) is not enough to solve the current matrix. The issue can be resolved by increasing the value for available memory.

-10: problems with opening OOC temporary files

This error value is returned when PARDISO can’t create / open temporary files for storing OOC arrays, e.g. in the case of wrong permissions or when files were removed or blocked or not released after the previous steps.

-11: read/write problems with the OOC data file

This error value is returned when some problems appear in the process of working with files, e.g. in the case of no space left on device or problems with read / write operations because of algorithm issues.

* - the documentation error. Will be fixed in the version 10.3 Update3.

** - available memory means the RAM system's memory which is available at the moment of starting the calculations.