Pascal,
Ipp vs. MKL: Depends on the vector or matrix sizes. MKL is designed
for large vectors/matrices. It ramps up to near peak for dgemm around
128-256 elements ( N=M=K ). So for smaller matrices you will probably
see as good/better/similar performance from IPP. The performance curves
vary by BLAS routine. Generally, though, MKL should have 128 element
vectors/matrices to amortize the startup costs.
Icc vs. ifort: depends on the program. If a C programmer avoids
pointer-based arrays and avoid aliasing, then the performance should be
similar. That is to say, if you program C like a Fortran programmer
then it should be close to as fast as ifort. Generally, though, Fortran
is faster. This can be a rather religious argument: you can program
well in either language, and you can write bad, slow programs in either
language. Fortran compilers generally are able to extract more
performance since aliasing is very limited whereas in C pointers cause
fits for optimizers.
For Core 2 Duo, try these options with both compilers: -O3 -xT -ipo as
a starting point.
Fortran matmul vs MKL. MKL will outperform matmul when matrices have
128 elements for more. Below that, hard to say. I'd be curious to find
out.
For MKL, set env var OMP_NUM_THREADS equal to the number of cores in
your system. It is internally threaded and will parallelize the
operations.
However, if your code is parallelized you will want to run MKL
single-threaded by setting OMP_NUM_THREADS=1. you do not want to
oversubscribe your cores.
ron
-----Original Message-----
From: fortran-dev-bounces+ronald.w.green=email@hidden
[mailto:fortran-dev-bounces+ronald.w.green=email@hidden] On
Behalf Of p666
Sent: Wednesday, December 20, 2006 4:58 PM
To: email@hidden
Subject: mkl versus ipp
hello,
does anyone know if the mkl is the same performance as ipp on
common functions, for instance matrix related operations ?
also does anyone know if a icc program is the same performance as a
ifort program
provided they link on the same mkl routines ?
and finally is the matmul fortran routine of ifort slower that the
mkl matrix multiply function ?
in short is intel fortran faster than intel c (on core 2 duo) with
the common mkl/ipp library ?
thanks in advance
pascal
_______________________________________________
Do not post admin requests to the list. They will be ignored.
Fortran-dev mailing list (email@hidden)
Help/Unsubscribe/Update your Subscription:
l.com
This email sent to email@hidden
_______________________________________________
Do not post admin requests to the list. They will be ignored.
Fortran-dev mailing list (email@hidden)
Help/Unsubscribe/Update your Subscription:
This email sent to email@hidden
