scnc
/
quantum-espresso
mirror of https://gitlab.com/QEF/q-e.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Added workaround for a bug in MKL causing the code to crash in BFGS. Fixes #15

This commit is contained in:
Pietro Bonfa 2019-02-14 15:32:08 +01:00
parent a175a3fd53
commit 2221882c19
1 changed files with 65 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

66
Modules/invmat.f90
View File

@ -25,6 +25,9 @@ MODULE matrix_inversion
! if "da" is specified and if the matrix is dimensioned 3x3,
! it also returns the determinant in "da"
!
#if defined(_OPENMP)
USE omp_lib
#endif
USE kinds, ONLY : DP
IMPLICIT NONE
INTEGER, INTENT(in) :: n
@ -41,6 +44,27 @@ MODULE matrix_inversion
! more work space
INTEGER, SAVE :: lworkfact = 64
!
! WORKAROUND STARTS ==================================================
!
! Comment taken from v6.1 (https://github.com/fspiga/qe-gpu)
!
! There is a bug in several versions of MKL that will cause an hang in the multithreaded DGEMM for AVX2.
! To avoid the bug, we have two options, set the number of MKL threads to one or force to use AVX instead of AVX2.
! To force the single threads, we need to read the current number of threads with numt=mkl_get_max_threads(), set it
! temporarely to one with "call mkl_set_num_threads(1)" and then resetting it to the original numt at the end of the function.
! To force AVX, we can call mkl_cbwr_set(MKL_CBWR_AVX).
!
! There is currently no way to check if MKL is used for LA.
! Since the size of the matrix to be inverted in PW is generally small,
! we disable OpenMP in the calls below for the time being.
!
#ifdef _OPENMP
INTEGER :: num_threads
num_threads=omp_get_max_threads()
CALL omp_set_num_threads(1)
#endif
! WORKAROUND ENDS ====================================================
!
IF ( PRESENT(da) ) THEN
IF ( n == 3 ) THEN
da = a(1,1)*(a(2,2)*a(3,3)-a(2,3)*a(3,2)) + &
@ -72,13 +96,23 @@ MODULE matrix_inversion
!
lworkfact = INT (work(1)/n)
DEALLOCATE ( work, ipiv )
! WORKAROUND STARTS ==================================================
!
! ... and now restrore the previous value.
!
#ifdef _OPENMP
CALL omp_set_num_threads(num_threads)
#endif
! WORKAROUND ENDS ====================================================
END SUBROUTINE invmat_r
SUBROUTINE invmat_c (n, a, a_inv, da)
!-----------------------------------------------------------------------
! as invmat_r, for a complex matrix
!
#if defined(_OPENMP)
USE omp_lib
#endif
USE kinds, ONLY : DP
IMPLICIT NONE
INTEGER, INTENT(IN) :: n
@ -95,6 +129,27 @@ MODULE matrix_inversion
! more work space
INTEGER, SAVE :: lworkfact = 64
!
! WORKAROUND STARTS ==================================================
!
! Comment taken from v6.1 (https://github.com/fspiga/qe-gpu)
!
! There is a bug in several versions of MKL that will cause an hang in the multithreaded DGEMM for AVX2.
! To avoid the bug, we have two options, set the number of MKL threads to one or force to use AVX instead of AVX2.
! To force the single threads, we need to read the current number of threads with numt=mkl_get_max_threads(), set it
! temporarely to one with "call mkl_set_num_threads(1)" and then resetting it to the original numt at the end of the function.
! To force AVX, we can call mkl_cbwr_set(MKL_CBWR_AVX).
!
! There is currently no way to check if MKL is used for LA.
! Since the size of the matrix to be inverted in PW is generally small,
! we disable OpenMP in the calls below for the time being.
!
#ifdef _OPENMP
INTEGER :: num_threads
num_threads=omp_get_max_threads()
CALL omp_set_num_threads(1)
#endif
! WORKAROUND ENDS ====================================================
!
IF ( PRESENT(da) ) THEN
IF (n == 3) THEN
da = a(1,1)*(a(2,2)*a(3,3)-a(2,3)*a(3,2)) + &
@ -127,6 +182,15 @@ MODULE matrix_inversion
lworkfact = INT (work(1)/n)
DEALLOCATE ( work, ipiv )
!
! WORKAROUND STARTS ==================================================
!
! ... and now restrore the previous value.
!
#ifdef _OPENMP
CALL omp_set_num_threads(num_threads)
#endif
! WORKAROUND ENDS ====================================================
!
END SUBROUTINE invmat_c
END MODULE matrix_inversion