qpmsblas.h File Reference

Naïve implementation of BLAS functions to workaround OpenBLAS bugs. More...

Go to the source code of this file.

Macros
#define	QPMS_BLAS_INDEX_T   long long int

Enumerations
enum	CBLAS_LAYOUT { CblasRowMajor =101 , CblasColMajor =102 }
enum	CBLAS_TRANSPOSE { CblasNoTrans =111 , CblasTrans =112 , CblasConjTrans =113 }
enum	CBLAS_UPLO { CblasUpper =121 , CblasLower =122 }
enum	CBLAS_DIAG { CblasNonUnit =131 , CblasUnit =132 }
enum	CBLAS_SIDE { CblasLeft =141 , CblasRight =142 }

Functions
void	qpms_zgemm (CBLAS_LAYOUT Order, CBLAS_TRANSPOSE TransA, CBLAS_TRANSPOSE TransB, const QPMS_BLAS_INDEX_T M, const QPMS_BLAS_INDEX_T N, const QPMS_BLAS_INDEX_T K, const _Complex double *alpha, const _Complex double *A, const QPMS_BLAS_INDEX_T lda, const _Complex double *B, const QPMS_BLAS_INDEX_T ldb, const _Complex double *beta, _Complex double *C, const QPMS_BLAS_INDEX_T ldc)
	Naïve serial reimplementation of cblas_zgemm.

Detailed Description

Naïve implementation of BLAS functions to workaround OpenBLAS bugs.

There is a bug in certain versions of OpenBLAS causing crashes if many BLAS functions are run from different threads in parallel. Typically, this happens when one calculates some large matrix (such as the global translation matrix) in several threads and cblas_zgemm() is called on relatively small submatrices. Because the submatrices are small, we can use a naïve, serial implementation of cblas_zgemm() as a workaround without a substantial impact on performance.

If included together with <cblas.h>, this must be include afterwards because of the typedefs!