From 8d10fb37421355732d212955a2fe71e884024d0b Mon Sep 17 00:00:00 2001
From: "m.petschow" <m.petschow@googlemail.com>
Date: Wed, 15 May 2013 15:46:36 +0000
Subject: [PATCH] bug: if multiple blocks initial approx in parallel

---
 SRC/LAPACK/odebz.f | 16 ----------------
 SRC/mrrr.c         |  3 ++-
 SRC/mrrr_val.c     |  4 +++-
 3 files changed, 5 insertions(+), 18 deletions(-)

diff --git a/SRC/LAPACK/odebz.f b/SRC/LAPACK/odebz.f
index 921f921..ff259ce 100644
--- a/SRC/LAPACK/odebz.f
+++ b/SRC/LAPACK/odebz.f
@@ -251,7 +251,6 @@ SUBROUTINE ODEBZ( IJOB, NITMAX, N, MMAX, MINP, NBMIN, ABSTOL,
 *        Compute the number of eigenvalues in the initial intervals.
 *
          MOUT = 0
-*DIR$ NOVECTOR
          DO 30 JI = 1, MINP
             DO 20 JP = 1, 2
                TMP1 = D( 1 ) - AB( JI, JP )
@@ -406,21 +405,6 @@ SUBROUTINE ODEBZ( IJOB, NITMAX, N, MMAX, MINP, NBMIN, ABSTOL,
                   ITMP1 = 1
                   TMP2 = MIN( TMP2, -PIVMIN )
                END IF
-*
-*              A series of compiler directives to defeat vectorization
-*              for the next loop
-*
-*$PL$ CMCHAR=' '
-CDIR$          NEXTSCALAR
-C$DIR          SCALAR
-CDIR$          NEXT SCALAR
-CVD$L          NOVECTOR
-CDEC$          NOVECTOR
-CVD$           NOVECTOR
-*VDIR          NOVECTOR
-*VOCL          LOOP,SCALAR
-CIBM           PREFER SCALAR
-*$PL$ CMCHAR='*'
 *
                DO 90 J = 2, N
                   TMP2 = D( J ) - E2( J-1 ) / TMP2 - TMP1
diff --git a/SRC/mrrr.c b/SRC/mrrr.c
index b28ddfc..ec90a20 100644
--- a/SRC/mrrr.c
+++ b/SRC/mrrr.c
@@ -238,7 +238,8 @@ int mrrr(char *jobz, char *range, int *np, double *restrict D,
     tolstruct->rtol2 = fmax(tolstruct->rtol2, 4.0 * DBL_EPSILON);
   }
   /* LAPACK: tolstruct->bsrtol = sqrt(DBL_EPSILON); */
-  tolstruct->bsrtol = fmin(tolstruct->rtol1, sqrt(DBL_EPSILON));
+  //  tolstruct->bsrtol = fmin(tolstruct->rtol1, sqrt(DBL_EPSILON));
+  tolstruct->bsrtol = MIN_RELGAP;
   tolstruct->RQtol  = 2.0 * DBL_EPSILON;
  
   /*  Compute the desired eigenvalues */
diff --git a/SRC/mrrr_val.c b/SRC/mrrr_val.c
index 0f12a0a..2950460 100644
--- a/SRC/mrrr_val.c
+++ b/SRC/mrrr_val.c
@@ -501,7 +501,9 @@ int find_eigval_approx(int max_nthreads, char *range, in_t *Dstruct,
   while (nthreads > 1 && nvals/nthreads < MIN_BISEC_CHUNK)
     nthreads--;
   
-  if (nthreads > 1) {
+  /* Disabled the parallel execution if matrix splits into multiple blocks 
+     as it is not reliable as it is; need to be done with more care */
+  if (nthreads > 1 && nsplit == 1) {
     
     threads = (pthread_t *) malloc( nthreads * sizeof(pthread_t) );
     assert(threads != NULL);