Skip to content
GitLab
Commit 51a8651d authored by Claus Kleinwort's avatar Claus Kleinwort
Browse files

Exploit sparisty of constraint matrix to save time and space 

git-svn-id: http://svnsrv.desy.de/public/MillepedeII/trunk@216 3547b9b0-65b8-46d3-b95d-921b3f43af62
parent 3bf83fd4
Expand all Hide whitespace changes
Inline Side-by-side
mpqldec.f90
View file @ 51a8651d
This diff is collapsed.
pede.f90
View file @ 51a8651d
......@@ -133,8 +133,8 @@
!! * 210301: New solution methods \ref ch-lapack "fullLAPACK" and \ref ch-lapack "unpackedLAPACK"
!! (matrix factorization) based on [LAPACK](http://www.netlib.org/lapack/) can be included optionally
!! (at compile time, <tt>-DLAPACK64=..</tt>).
!! * 210704: Exploit decomposition of constraints matrix into disjoint blocks for MINRES preconitioner
!! (with Lagrange multipliers) too.
!! * 210728: Exploit decomposition of constraints matrix into disjoint blocks for all
!! solution methods (e.g. QL decomposition, MINRES preconditioner).
!!
!! \section tools_sec Tools
!! The subdirectory \c tools contains some useful scripts:
......@@ -1664,9 +1664,9 @@ SUBROUTINE feasma
END IF
IF(matsto > 0) THEN ! True unless LAPACK
! QL decomposition
CALL qlini(nvgb,ncgb,npblck,monpg1)
CALL qlini(nvgb,ncgb,npblck,mszcon,monpg1)
! loop over parameter blocks
CALL qldecb(matConstraintsT,matParBlockOffsets,matConsBlocks)
CALL qldecb(matConstraintsT,matParBlockOffsets,matConsBlocks,matConsSort)
! check eignevalues of L
CALL qlgete(evmin,evmax)
#ifdef LAPACK64
......@@ -4784,37 +4784,45 @@ END SUBROUTINE prtstat ! print input statistics
!> Product symmetric (sub block) matrix times sparse vector.
!!
!! A(sym) * X => B. Used by \ref minresmodule::minres "MINRES" method (Is most CPU intensive part).
!! A(sym) * X => B. (Cumultative). Used by \ref minresmodule::minres "MINRES" method (Is most CPU intensive part).
!! The matrix A is the global matrix in full symmetric or (compressed) sparse storage.
!! In full symmetric storage it could be block diagonal (MATSTO=3) and only a
!! single block is used in the product.
!!
!! \param [in] n size of matrix
!! \param [in] l offset of (sub block) parameter range
!! \param [in] x vector X
!! \param [in] is start of range x(is:ie)
!! \param [in] ie end of range x(is:ie)
!! \param [out] b result vector B
!! \param [in] is sparsity structure of x (number of non-zero regions, {region start, end})
SUBROUTINE avprds(n,x,b,is)
SUBROUTINE avprds(n,l,x,is,ie,b)
USE mpmod
IMPLICIT NONE
INTEGER(mpi) :: i
INTEGER(mpi) :: ia
INTEGER(mpi) :: ia2
INTEGER(mpi) :: ib
INTEGER(mpi) :: ib2
INTEGER(mpi) :: in
INTEGER(mpi) :: ipg
INTEGER(mpi) :: iproc
INTEGER(mpi) :: ir
INTEGER(mpi) :: irgn
INTEGER(mpi) :: j
INTEGER(mpi) :: ja
INTEGER(mpi) :: ja2
INTEGER(mpi) :: jb
INTEGER(mpi) :: jb2
INTEGER(mpi) :: jn
INTEGER(mpi) :: jrgn
INTEGER(mpi) :: lj
INTEGER(mpi), INTENT(IN) :: n
INTEGER(mpl), INTENT(IN) :: l
REAL(mpd), INTENT(IN) :: x(n)
INTEGER(mpi), INTENT(IN) :: is
INTEGER(mpi), INTENT(IN) :: ie
REAL(mpd), INTENT(OUT) :: b(n)
INTEGER(mpi), INTENT(IN) :: is(*)
INTEGER(mpl) :: k
INTEGER(mpl) :: kk
INTEGER(mpl) :: ku
......@@ -4826,25 +4834,32 @@ SUBROUTINE avprds(n,x,b,is)
!$ INTEGER(mpi) OMP_GET_THREAD_NUM
SAVE
! ...
!$ DO i=1,n
!$ b(i)=0.0_mpd ! reset 'global' B()
!$ END DO
ichunk=MIN((n+mthrd-1)/mthrd/8+1,1024)
ichunk=MIN((n+mthrd-1)/mthrd/8+1,128)
IF(matsto /= 2) THEN
! full symmetric matrix
! full or unpacked (block diagonal) symmetric matrix
! parallelize row loop
! private copy of B(N) for each thread, combined at end, init with 0.
! slot of 1024 'I' for next idle thread
! slot of 128 'I' for next idle thread
!$OMP PARALLEL DO &
!$OMP PRIVATE(J,IJ) &
!$OMP REDUCTION(+:B) &
!$OMP SCHEDULE(DYNAMIC,ichunk)
DO i=1,n
ij=globalRowOffsets(i)
b(i)=globalMatD(ij+i)*x(i)
ij=globalRowOffsets(i+l)+l
DO j=is,min(i,ie)
b(i)=b(i)+globalMatD(ij+j)*x(j)
END DO
END DO
!$OMP END PARALLEL DO
!$OMP PARALLEL DO &
!$OMP PRIVATE(J,IJ) &
!$OMP REDUCTION(+:B) &
!$OMP SCHEDULE(DYNAMIC,ichunk)
DO i=is,ie
ij=globalRowOffsets(i+l)+l
DO j=1,i-1
b(j)=b(j)+globalMatD(ij+j)*x(i)
b(i)=b(i)+globalMatD(ij+j)*x(j)
END DO
END DO
!$OMP END PARALLEL DO
......@@ -4858,7 +4873,7 @@ SUBROUTINE avprds(n,x,b,is)
! slot of 1024 'I' for next idle thread
!$OMP PARALLEL DO &
!$OMP PRIVATE(I,IR,K,KK,LL,KU,INDID,INDIJ,J,JN,LJ) &
!$OMP PRIVATE(IA,IB,IN,JA,JB,IRGN,JRGN) &
!$OMP PRIVATE(IA,IB,IN,JA,JB,IA2,IB2,JA2,JB2) &
!$OMP REDUCTION(+:B) &
!$OMP SCHEDULE(DYNAMIC,ichunk)
DO ipg=1,napgrp
......@@ -4868,18 +4883,11 @@ SUBROUTINE avprds(n,x,b,is)
ia=globalAllIndexGroups(ipg) ! first (global) row
ib=globalAllIndexGroups(ipg+1)-1 ! last (global) row
in=ib-ia+1 ! number of rows
! check x region
irgn=1 ! non-zero region in x
DO WHILE (ia > is(2*irgn+1).AND.irgn < is(1))
irgn=irgn+1
END DO
!
! overlap
ia2=max(ia,is)
ib2=min(ib,ie)
! diagonal elements
IF (ia <= is(2*irgn+1).AND.ib >= is(2*irgn)) THEN
b(ia:ib)=globalMatD(ia:ib)*x(ia:ib)
ELSE
b(ia:ib)=0.0_mpd
END IF
IF (ia2 <= ib2) b(ia2:ib2)=b(ia2:ib2)+globalMatD(ia2:ib2)*x(ia2:ib2)
! off-diagonals double precision
ir=ipg
kk=sparseMatrixOffsets(1,ir) ! offset in 'd' (column lists)
......@@ -4889,26 +4897,21 @@ SUBROUTINE avprds(n,x,b,is)
indij=ll
IF (sparseMatrixColumns(indid+1) /= 0) THEN ! no compression
DO i=ia,ib
IF (i <= is(2*irgn+1).AND.i >= is(2*irgn)) THEN
IF (i <= ie.AND.i >= is) THEN
DO k=1,ku
j=sparseMatrixColumns(indid+k)
b(j)=b(j)+globalMatD(indij+k)*x(i)
END DO
END IF
jrgn=1 ! non-zero region in x
DO k=1,ku
j=sparseMatrixColumns(indid+k)
DO WHILE (j > is(2*jrgn+1).AND.jrgn < is(1))
jrgn=jrgn+1
END DO
IF (j <= is(2*jrgn+1).AND.j >= is(2*jrgn)) THEN
IF (j <= ie.AND.j >= is) THEN
b(i)=b(i)+globalMatD(indij+k)*x(j)
END IF
END DO
indij=indij+ku
END DO
ELSE
jrgn=1 ! non-zero region in x
! regions of continous column groups
DO k=2,ku-2,2
j=sparseMatrixColumns(indid+k) ! first group
......@@ -4916,21 +4919,19 @@ SUBROUTINE avprds(n,x,b,is)
lj=sparseMatrixColumns(indid+k-1) ! region offset
jn=sparseMatrixColumns(indid+k+1)-lj ! number of columns
jb=ja+jn-1 ! last (global) column
! check x region
DO WHILE (ja > is(2*jrgn+1).AND.jrgn < is(1))
jrgn=jrgn+1
END DO
IF (ja <= is(2*jrgn+1).AND.jb >= is(2*jrgn)) THEN
ja2=max(ja,is)
jb2=min(jb,ie)
IF (ja2 <= jb2) THEN
lj=1 ! index (in group region)
DO i=ia,ib
b(i)=b(i)+dot_product(globalMatD(indij+lj:indij+lj+jn-1),x(ja:jb))
b(i)=b(i)+dot_product(globalMatD(indij+lj+ja2-ja:indij+lj+jb2-ja),x(ja2:jb2))
lj=lj+jn
END DO
END IF
IF (mextnd == 0.AND.ia <= is(2*irgn+1).AND.ib >= is(2*irgn)) THEN
IF (mextnd == 0.AND.ia2 <= ib2) THEN
lj=1
DO j=ja,jb
b(j)=b(j)+dot_product(globalMatD(indij+lj:indij+jn*in:jn),x(ia:ib))
b(j)=b(j)+dot_product(globalMatD(indij+lj+jn*(ia2-ia):indij+jn*(ib2-ia):jn),x(ia2:ib2))
lj=lj+1
END DO
END IF
......@@ -4947,15 +4948,21 @@ SUBROUTINE avprds(n,x,b,is)
indij=ll
IF (sparseMatrixColumns(indid+1) /= 0) THEN ! no compression
DO i=ia,ib
IF (i <= ie.AND.i >= is) THEN
DO k=1,ku
j=sparseMatrixColumns(indid+k)
b(j)=b(j)+globalMatF(indij+k)*x(i)
END DO
END IF
DO k=1,ku
j=sparseMatrixColumns(indid+k)
b(j)=b(j)+REAL(globalMatF(indij+k),mpd)*x(i)
b(i)=b(i)+REAL(globalMatF(indij+k),mpd)*x(j)
IF (j <= ie.AND.j >= is) THEN
b(i)=b(i)+globalMatF(indij+k)*x(j)
END IF
END DO
indij=indij+ku
END DO
ELSE
jrgn=1 ! non-zero region in x
! regions of continous column groups
DO k=2,ku-2,2
j=sparseMatrixColumns(indid+k) ! first group
......@@ -4963,21 +4970,19 @@ SUBROUTINE avprds(n,x,b,is)
lj=sparseMatrixColumns(indid+k-1) ! region offset
jn=sparseMatrixColumns(indid+k+1)-lj ! number of columns
jb=ja+jn-1 ! last (global) column
! check x region
DO WHILE (ja > is(2*jrgn+1).AND.jrgn < is(1))
jrgn=jrgn+1
END DO
IF (ja <= is(2*jrgn+1).AND.jb >= is(2*jrgn)) THEN
ja2=max(ja,is)
jb2=min(jb,ie)
IF (ja2 <= jb2) THEN
lj=1 ! index (in group region)
DO i=ia,ib
b(i)=b(i)+dot_product(REAL(globalMatF(indij+lj:indij+lj+jn-1),mpd),x(ja:jb))
lj=lj+jn
b(i)=b(i)+dot_product(REAL(globalMatF(indij+lj+ja2-ja:indij+lj+jb2-ja),mpd),x(ja:jb))
lj=lj+jn
END DO
END IF
IF (mextnd == 0.AND.ia <= is(2*irgn+1).AND.ib >= is(2*irgn)) THEN
IF (mextnd == 0.AND.ia2 <= ib2) THEN
lj=1
DO j=ja,jb
b(j)=b(j)+dot_product(REAL(globalMatF(indij+lj:indij+jn*in:jn),mpd),x(ia:ib))
b(j)=b(j)+dot_product(REAL(globalMatF(indij+lj+jn*(ia2-ia):indij+jn*(ib2-ia):jn),mpd),x(ia:ib))
lj=lj+1
END DO
END IF
......@@ -5164,6 +5169,7 @@ SUBROUTINE avprd0(n,l,x,b)
END SUBROUTINE avprd0
!> Analyse sparsity structure.
!!
SUBROUTINE anasps
......@@ -7484,8 +7490,10 @@ SUBROUTINE vmprep(msize)
CALL mpalloc(matPreCond,length,'default preconditioner matrix')
END IF
nwrdpc=nwrdpc+2*length
IF (nwrdpc > 250000) &
IF (nwrdpc > 250000) THEN
WRITE(*,*)
WRITE(*,*) 'Size of preconditioner matrix:',INT(REAL(nwrdpc,mps)*4.0E-6,mpi),' MB'
END IF
END IF
......@@ -7576,7 +7584,7 @@ SUBROUTINE minver
INTEGER(mpl) :: ioff1
REAL(mpd) :: matij
EXTERNAL avprd0
EXTERNAL avprds
SAVE
! ...
......@@ -7594,7 +7602,7 @@ SUBROUTINE minver
WRITE(lunlog,*) 'Shrinkage of global matrix (A->Q^t*A*Q)'
CALL monini(lunlog,monpg1,monpg2)
END IF
CALL qlssq(avprd0,globalMatD,globalRowOffsets,.true.) ! Q^t*A*Q
CALL qlssq(avprds,globalMatD,globalRowOffsets,.true.) ! Q^t*A*Q
IF(monpg1 > 0) CALL monend()
END IF
END IF
......@@ -7691,7 +7699,7 @@ SUBROUTINE mchdec
REAL(mpd) :: evmax
REAL(mpd) :: evmin
EXTERNAL avprd0
EXTERNAL avprds
SAVE
! ...
......@@ -7704,7 +7712,7 @@ SUBROUTINE mchdec
WRITE(lunlog,*) 'Shrinkage of global matrix (A->Q^t*A*Q)'
CALL monini(lunlog,monpg1,monpg2)
END IF
IF(nfgb < nvgb) CALL qlssq(avprd0,globalMatD,globalRowOffsets,.true.) ! Q^t*A*Q
IF(nfgb < nvgb) CALL qlssq(avprds,globalMatD,globalRowOffsets,.true.) ! Q^t*A*Q
IF(monpg1 > 0) CALL monend()
END IF
......@@ -7801,7 +7809,7 @@ SUBROUTINE mdptrf
INTEGER(mpi) :: infolp
REAL(mpd) :: matij
EXTERNAL avprd0
EXTERNAL avprds
SAVE
! ...
......@@ -7821,7 +7829,7 @@ SUBROUTINE mdptrf
WRITE(lunlog,*) 'Shrinkage of global matrix (A->Q^t*A*Q)'
CALL monini(lunlog,monpg1,monpg2)
END IF
CALL qlssq(avprd0,globalMatD,globalRowOffsets,.true.) ! Q^t*A*Q
CALL qlssq(avprds,globalMatD,globalRowOffsets,.true.) ! Q^t*A*Q
IF(monpg1 > 0) CALL monend()
END IF
END IF
......@@ -8271,7 +8279,7 @@ SUBROUTINE mdiags
INTEGER(mpi) :: ntop
REAL(mpd) :: matij
!
EXTERNAL avprd0
EXTERNAL avprds
SAVE
! ...
......@@ -8287,7 +8295,7 @@ SUBROUTINE mdiags
!use elimination for constraints ?
IF(nfgb < nvgb) THEN
IF(icalcm == 1) CALL qlssq(avprd0,globalMatD,globalRowOffsets,.true.) ! Q^t*A*Q
IF(icalcm == 1) CALL qlssq(avprds,globalMatD,globalRowOffsets,.true.) ! Q^t*A*Q
! solve L^t*y=d by backward substitution
CALL qlbsub(vecConsResiduals,vecConsSolution)
! transform, reduce rhs
......@@ -8582,7 +8590,10 @@ SUBROUTINE mminrsqlp
IF(mbandw == 0) THEN ! default preconditioner
IF(icalcm == 1) THEN
IF(monpg1 > 0) CALL monini(lunlog,monpg1,monpg2)
WRITE(lun,*) 'MMINRS: QLPSSQ started'
IF(nfgb < nvgb) CALL qlpssq(avprds,matPreCond,1,.true.) ! transform preconditioner matrix
IF(monpg1 > 0) CALL monend()
IF(monpg1 > 0) CALL monini(lunlog,monpg1,monpg2)
WRITE(lun,*) 'MMINRS: PRECONS started', nprecond(2), nprecond(1)
CALL precons(nprecond(1),nprecond(2),nprecond(3),matPreCond,matPreCond, &
......@@ -8595,7 +8606,10 @@ SUBROUTINE mminrsqlp
x=globalCorrections, istop=istop, itn=itn)
ELSE IF(mbandw > 0) THEN ! band matrix preconditioner
IF(icalcm == 1) THEN
IF(monpg1 > 0) CALL monini(lunlog,monpg1,monpg2)
WRITE(lun,*) 'MMINRS: QLPSSQ started'
IF(nfgb < nvgb) CALL qlpssq(avprds,matPreCond,mbandw,.true.) ! transform preconditioner matrix
IF(monpg1 > 0) CALL monend()
IF(monpg1 > 0) CALL monini(lunlog,monpg1,monpg2)
WRITE(lun,*) 'MMINRS: EQUDECS started', nprecond(2), nprecond(1)
CALL equdecs(nprecond(2),nprecond(1),nprecond(3),lprecm,matPreCond,indPreCond,blockPreCond,nrkd,nrkd2)
......@@ -8750,7 +8764,7 @@ SUBROUTINE xloopn !
CHARACTER (LEN=7) :: cratio
CHARACTER (LEN=7) :: cfacin
CHARACTER (LEN=7) :: crjrat
EXTERNAL avprd0
EXTERNAL avprds
SAVE
! ...
......@@ -9260,7 +9274,7 @@ SUBROUTINE xloopn !
CALL monini(lunlog,monpg1,monpg2)
END IF
IF(matsto > 0) THEN
CALL qlssq(avprd0,globalMatD,globalRowOffsets,.false.) ! Q*A*Q^t
CALL qlssq(avprds,globalMatD,globalRowOffsets,.false.) ! Q*A*Q^t
#ifdef LAPACK64
ELSE ! unpack storage, use LAPACK
CALL lpavat(.false.)
......
Supports Markdown
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment

imprint | privacy | Conditions of Use