[gpaw-users] Install gpaw using intel parallel studio cluste version.

[Jussi Enkovaara]

On 27/05/2020 17.00, Christian Tuma via gpaw-users wrote:
> 
> Am 27.05.2020 um 14:49 schrieb Jens Jørgen Mortensen:
>> ...  We can always revisit this later if someone finds that this part
>> of the code is a bottleneck.
> 
> Good point. I was wondering why the SIMD pragmas were introduced at all
> in the code. Loop trip counts need to be large enough to compensate for
> the initial SIMD performance overhead. Would be interesting to see some
> benchmark runs proving that the code really benefits from enforced
> SIMD'fication of those loops. Also because there are some other, very
> similar loops in the same source file(s) which were not explicitly
> SIMD'fied ...

Hi,
I think the SIMD pragmas where introduced when optimizing the code for
Intel Xeon Phi, and at least in these CPUs there were some performance
benefits (although I do not have any exact numbers in my hand). In any
case, most likely more for the finite-difference operations than for the
symmetrization.

Also, the pragma does not enforce vectorization per se, they are just
additional information to compiler that the following loop is safe to
SIMD:ify (for cases where compiler cannot determine that automatically).
Compiler can still decide not to vectorize if it considers overhead too
large.

Best regards,
Jussi
-- 
Jussi Enkovaara
Application scientist, HPC support, CSC - IT Center for Science
jussi.enkovaara at csc.fi    tel. +358 50 3819 038