[Wien] more on pathscale and ifort64 : ACML

[Stefaan Cottenier]

To Stargmoon and others interested: benchmarks for the acml library, 
using both ifort64 and pathscale as compilers. Tests have run on an 
identical machine as in the previous mail, where ifort64+goto got 215 s.

pathscale and acml 2.1
===============
test_case : *segmentation fault* in lapw1c (real case might work, didn't 
test)

     O   Compiler options:        -freeform -march=opteron -mcpu=opteron 
-mtune=opteron -w -IPA
     L   Linker Flags:            -L/apps/prod/math-lib/acml/gnu64/lib 
-L/usr/lib64 -L../SRC_lib -pthread -Wl,-rpath 
/apps/prod/math-lib/acml/gnu64/lib -Wl,-rpath /usr/lib64 -IPA -static
     P   Preprocessor flags       '-DParallel'
     R   R_LIB (LAPACK+BLAS):     -lacml -llapack_lapw -lg2c -lacml 
-llapack_lapw -lg2c

Same general comments as for previous tests apply. /usr/lib64 was the 
place were g2c was located on our system. For some reason I could not 
get it run with dynamic libraries (probably trivial to solve), such that 
static linking was used. Although compilation went smoothly, a 
segmentation fault occurs when running the bench_mark (lapw1c). The same 
happened for pathscale with the goto and atlas libraries, only mkl works 
for me with pathscale (but with a bad timing). It would be interesting 
if somebody with debugging skills takes a look at this, because with 3 
failures out of 4 cases, I get suspicious pathscale reveals some bug in 
lapw1c which the other compilers miss.

ifort64 and acml 2.1
==============
test_case: 280 s

     O   Compiler options:        -FR -mp1 -w -prec_div -pc80 -pad -ip 
-DINTEL_VML
     L   Linker Flags:            -L/apps/prod/math-lib/acml/gnu64/lib 
-L/apps/prod/math-lib/mkl72/lib/em64t -L/usr/lib64 -L../SRC_lib -pthread 
-Wl,-rpath /apps/prod/math-lib/acml/gnu64/lib -Wl,-rpath /usr/lib64 
-Wl,-rpath /apps/prod/math-lib/mkl72/lib/em64t
     P   Preprocessor flags       '-DParallel'
     R   R_LIB (LAPACK+BLAS):     -lacml -llapack_lapw -lg2c -lvml 
-lacml -llapack_lapw -lg2c -lvml

It turned out I had to link with vml (apparantly part of mkl), in order 
to compile lapw1 completely. If not, 3 routines in hamilt.f were not 
recognized. The 280 s for the benchmark make acml the second-best 
library for ifort64, after goto but before atlas and mkl.

Stefaan