SVD multithreading bug in MKL

Reference Number : DPD200335246

Product : Intel® Math Kernel Library (Intel® MKL)

Version : Intel MKL 11.0 Update 3 and Intel MKL 11.1 beta for OS X*

Operating Systems affected: Windows* OS, Linux * OS, OS X*

Problem Description : 

Intel MKL Customers using SVD routines (which computes the singular value decomposition of a general rectangular matrix) can get the wrong results or even segmentation faults when the number of threads is greater than 4 and not a multiple of 4. This problem is caused by a bug int the function which splits the workload for certain bandwidth limited functions. The full list of functions affected is in the “function.txt” file attached to this article.

Workaround: Explicitly set the number of threads to a number less than 4 or a multiple of 4. In other words, 1, 2, 3, 4, or 4*i, where i is an integer

Resolution Status: The problem will be fixed in the next release and a patch for Intel MKL 11.0 update 3 is available here (Windows, Linux, OS X*)

For more complete information about compiler optimizations, see our Optimization Notice.

1 comment


version 11.0 update 4 has been released at May 29 2013.

Add a Comment

Have a technical question? Visit our forums. Have site or software product issues? Contact support.