This is really old school, but a lot of times we have users sending us
copy pasting of codes, and that is the only way to know the version of
the code.
This is mostly a long term maintenance improvement.
Many coding styles require elimination of trailing whitespace, and
many editors and source code management configurations automatically
gobble up whitespace. When these tools gobble up whitespace, it
complicates reviewing the meaningful code changes.
By removing whitespace on one patch, it makes future
code reviews much easier.
=SCRIPT====================================================================
if which tempfile &>/dev/null; then
TEMPMAKER=tempfile
elif which mktemp &>/dev/null; then
TEMPMAKER=mktemp
else
echo "Cannot find tempfile program." 2>&1
exit 1
fi
MYTEMP=$($TEMPMAKER)
trap 'rm -f $MYTEMP' SIGINT SIGTERM
stripit() {
echo "stripping $1"
sed 's/[ \t]*$//' "$1" > $MYTEMP
cp $MYTEMP "$1"
}
if [ $# -gt 0 ]; then
while [ "$1" != "" ]; do
stripit $1
shift
done
else
while read -t 2; do
stripit $REPLY
done
fi
rm $MYTEMP
=================================================
B (or X) being zero to skip a loop so as to enable better NaN (or Inf)
propagation.
Taking DGEMM as an example, if you use notation:
C(I,J) = C(I,J) + alpha * A(I,L) * B(L,J),
the reference BLAS DGEMM is using the JLI version of matrix matrix multiply and
is checking, for each J (from 1 to N) and for each L (from 1 to K), whether
B(L,J) is zero (or not) to save (or not) the 2M following operations.
(See the "IF (B(L,J).NE.ZERO) THEN" in the code below.)
The snippets of code is as follows
DO 90 J = 1,N
DO 80 L = 1,K
IF (B(L,J).NE.ZERO) THEN
TEMP = ALPHA*B(L,J)
DO 70 I = 1,M
C(I,J) = C(I,J) + TEMP*A(I,L)
70 CONTINUE
END IF
80 CONTINUE
90 CONTINUE
This induces some non NaN-propagation in a pretty ad-hoc way. For better NaN
propagation, this patch removes the above IF statement.
The snippet of code now becomes
DO 90 J = 1,N
DO 80 L = 1,K
TEMP = ALPHA*B(L,J)
DO 70 I = 1,M
C(I,J) = C(I,J) + TEMP*A(I,L)
70 CONTINUE
80 CONTINUE
90 CONTINUE
This enables correct NaN propagation for this piece of code.
Rationale: BLAS does not correctly propagate all NaNs (and Infs). We still have
no NaN propagation where for example ALPHA=0, etc. The goal of this commit is
to have correct NaN propagation no matter what the entries of the input
matrices/vectors (A, B, C, X, etc.) are. BLAS do not correctly propagate NaNs
and Infs based on some values of the scalars (ALPHA, BETA, etc.).
See below the email from Tom Callaway from RedHat, sent on July 9th to
lapack@cs.utk.edu.
Hello LAPACK people,
Martyn & Lejeczek (on CC) reported an issue to Fedora relating to R using our
system copy of BLAS (from LAPACK).
As noted in the R administration and Installation Manual, "R relies on ISO/IEC
60559 compliance of an external BLAS. This can be broken if for example the
code assumes that terms with a zero factor are always zero and do not need to
be computed - whereas x*0 can be NaN. This is checked in the test suite."
In the stock BLAS, DGBMV, DGEMM, and DGEMV fail this. R has been patching their
bundled BLAS to resolve this issue since 2010, but Fedora now uses the system
BLAS.
Attached is a patch (from upstream R) to fix this issue in the LAPACK BLAS.
Please consider applying it.
Thanks,
~tom
Use \par instead of \details for section.
add a Contributors Section and a Reference Section.
Remove (some) verbatim section when not needed.
Those changes have been done by hand so I am not sure I manage to catch them all.
Mark Gates