Hello,
I am working on developing a modified stencil for the Einstein Toolkit and I am discovering that regression tests pass when I use the GNU compilers but fail when I use the Intel compilers, even with gentle optimization settings such as -O1. I suspect over-zealous optimization. Has anyone encountered this sort of problem before? And if so, does anybody know a good combination of flags that makes the intel compiler more cautious, without sacrificing too much speed? (For example, disabling fast math, which seems to be set to active by default.)
Thank you in advance for your help!
Best, Jonah Miller
On Tue, Nov 10, 2015 at 01:38:08PM -0500, Jonah Miller wrote:
I am working on developing a modified stencil for the Einstein Toolkit and I am discovering that regression tests pass when I use the GNU compilers but fail when I use the Intel compilers, even with gentle optimization settings such as -O1. I suspect over-zealous optimization. Has anyone encountered this sort of problem before? And if so, does anybody know a good combination of flags that makes the intel compiler more cautious, without sacrificing too much speed? (For example, disabling fast math, which seems to be set to active by default.)
The default for the Intel compiler is non-exact math. Try some of the options mentioned here:
https://software.intel.com/en-us/articles/consistency-of-floating-point-resu...
How much that influences effective speed likely depends a lot on your code.
Frank
Roland, I think you have been using such flags. Which flags exactly did you use?
-erik
On Tue, Nov 10, 2015 at 1:45 PM, Frank Loeffler knarf@cct.lsu.edu wrote:
On Tue, Nov 10, 2015 at 01:38:08PM -0500, Jonah Miller wrote:
I am working on developing a modified stencil for the Einstein Toolkit and I am discovering that regression tests pass when I use the GNU compilers but fail when I use the Intel compilers, even with gentle optimization settings such as -O1. I suspect over-zealous optimization. Has anyone encountered this sort of problem before? And if so, does anybody know a good combination of flags that makes the intel compiler more cautious, without sacrificing too much speed? (For example, disabling fast math, which seems to be set to active by default.)
The default for the Intel compiler is non-exact math. Try some of the options mentioned here:
https://software.intel.com/en-us/articles/consistency-of-floating-point-resu...
How much that influences effective speed likely depends a lot on your code.
Frank
Users mailing list Users@einsteintoolkit.org http://lists.einsteintoolkit.org/mailman/listinfo/users
Hello all,
uhh, let me actually read the question. Frank seems to have pointed you to what I assume (have not checked yet) to be the proper documentation that shows what can go wrong. Now for how I worked around the issues myself:
For the symmetric operators branch I use -fp-model precise to the intel compiler eg on zwicky I use the default zwicky machine defintion file in simfactory then add "-fp-model precise" to all CFLAGS/CXXFLAGS/F90FLAGS so that they read:
C_OPTIMISE_FLAGS = -O2 -xHost -fp-model precise CXX_OPTIMISE_FLAGS = -O2 -xHost -fp-model precise F77_OPTIMISE_FLAGS = -O2 -xHost -fp-model precise F90_OPTIMISE_FLAGS = -O2 -xHost -fp-model precise
for the gnu compiler on bluewaters I had to also disable some optimizations (I don't know why):
C_OPTIMISE_FLAGS = -O3 -fno-expensive-optimizations -fno-ipa-cp-clone
@Jonah: if you have access to the SXS wiki (say due to Spectre work), you can find this information on https://www.black-holes.org/wiki/center_of_mass_conserving_setup somewhere under the first big table.
Yours, Roland
On 2015-11-10 22:47, Erik Schnetter wrote:
Roland, I think you have been using such flags. Which flags exactly did you use?
-erik
On Tue, Nov 10, 2015 at 1:45 PM, Frank Loeffler knarf@cct.lsu.edu wrote:
On Tue, Nov 10, 2015 at 01:38:08PM -0500, Jonah Miller wrote:
I am working on developing a modified stencil for the Einstein Toolkit and I am discovering that regression tests pass when I use the GNU compilers but fail when I use the Intel compilers, even with gentle optimization settings such as -O1. I suspect over-zealous optimization. Has anyone encountered this sort of problem before? And if so, does anybody know a good combination of flags that makes the intel compiler more cautious, without sacrificing too much speed? (For example, disabling fast math, which seems to be set to active by default.)
The default for the Intel compiler is non-exact math. Try some of the options mentioned here:
https://software.intel.com/en-us/articles/consistency-of-floating-point-resu...
How much that influences effective speed likely depends a lot on your code.
Frank
Users mailing list Users@einsteintoolkit.org http://lists.einsteintoolkit.org/mailman/listinfo/users
Users mailing list Users@einsteintoolkit.org http://lists.einsteintoolkit.org/mailman/listinfo/users
Hi Roland, Frank,
Thanks very much. This was exactly what I was looking for.
Best, Jonah
On 15-11-11 04:46 AM, Roland Haas wrote:
Hello all,
uhh, let me actually read the question. Frank seems to have pointed you to what I assume (have not checked yet) to be the proper documentation that shows what can go wrong. Now for how I worked around the issues myself:
For the symmetric operators branch I use -fp-model precise to the intel compiler eg on zwicky I use the default zwicky machine defintion file in simfactory then add "-fp-model precise" to all CFLAGS/CXXFLAGS/F90FLAGS so that they read:
C_OPTIMISE_FLAGS = -O2 -xHost -fp-model precise CXX_OPTIMISE_FLAGS = -O2 -xHost -fp-model precise F77_OPTIMISE_FLAGS = -O2 -xHost -fp-model precise F90_OPTIMISE_FLAGS = -O2 -xHost -fp-model precise
for the gnu compiler on bluewaters I had to also disable some optimizations (I don't know why):
C_OPTIMISE_FLAGS = -O3 -fno-expensive-optimizations -fno-ipa-cp-clone
@Jonah: if you have access to the SXS wiki (say due to Spectre work), you can find this information on https://www.black-holes.org/wiki/center_of_mass_conserving_setup somewhere under the first big table.
Yours, Roland
On 2015-11-10 22:47, Erik Schnetter wrote:
Roland, I think you have been using such flags. Which flags exactly did you use?
-erik
On Tue, Nov 10, 2015 at 1:45 PM, Frank Loeffler knarf@cct.lsu.edu wrote:
On Tue, Nov 10, 2015 at 01:38:08PM -0500, Jonah Miller wrote:
I am working on developing a modified stencil for the Einstein Toolkit and I am discovering that regression tests pass when I use the GNU compilers but fail when I use the Intel compilers, even with gentle optimization settings such as -O1. I suspect over-zealous optimization. Has anyone encountered this sort of problem before? And if so, does anybody know a good combination of flags that makes the intel compiler more cautious, without sacrificing too much speed? (For example, disabling fast math, which seems to be set to active by default.)
The default for the Intel compiler is non-exact math. Try some of the options mentioned here:
https://software.intel.com/en-us/articles/consistency-of-floating-point-resu...
How much that influences effective speed likely depends a lot on your code.
Frank
Users mailing list Users@einsteintoolkit.org http://lists.einsteintoolkit.org/mailman/listinfo/users
Users mailing list Users@einsteintoolkit.org http://lists.einsteintoolkit.org/mailman/listinfo/users
Users mailing list Users@einsteintoolkit.org http://lists.einsteintoolkit.org/mailman/listinfo/users
users@lists.einsteintoolkit.org
-
Erik Schnetter -
Frank Loeffler -
Jonah Miller -
Roland Haas