- Infos im HLRS Wiki sind nicht rechtsverbindlich und ohne Gewähr -
- Information contained in the HLRS Wiki is not legally binding and HLRS is not responsible for any damages that might result from its use -

Libraries(Hawk): Difference between revisions

From HLRS Platforms
Jump to navigationJump to search
 
(20 intermediate revisions by 3 users not shown)
Line 1: Line 1:
==AOCL==
== AOCL ==
 
The [https://developer.amd.com/amd-aocl/ AMD Optimizing CPU Libraries (AOCL)] are optimized to utilize AMD EPYC processors in an optimal way.
 
A module is available '''gcc''' and '''aocc''' with
<pre>module load aocl/2.2.0</pre>
Keep in mind this module only contains the proprietary non-MPI components like amd-libm.
 
The libraries utilizing MPI will be compiled separately with MPT and provided as their own modules at a later time.


The [https://developer.amd.com/amd-aocl/ AMD Optimizing CPU Libraries (AOCL)] are available via
<pre>module load numlib/aocl/2.0</pre>
They are optimized to utilize AMD EPYC processors in an optimal way.
{| class="wikitable"
{| class="wikitable"
|-
|-
Line 51: Line 56:
| http://developer.amd.com/amd-cpu-libraries/rng-library/#securerng
| http://developer.amd.com/amd-cpu-libraries/rng-library/#securerng
|}
|}
== Intel MKL ==
Intel MKL is available with
<pre>
module load mkl/<version>
</pre>
For '''linking flags''' for different compilers, there is an [https://software.intel.com/en-us/articles/intel-mkl-link-line-advisor Intel MKL Link Line Advisor]
'''Static linking''' and [https://software.intel.com/en-us/mkl-linux-developer-guide-using-the-ilp64-interface-vs-lp64-interface 32-bit integer interface] are recommended if applicable.
For '''distributed''' MKL functionality ('''ScaLAPACK''', etc.), select '''SGIMPT''' for the MPI library, if you use '''mpt'''.
Intel MKL seems to be using a '[https://www.extremetech.com/computing/302650-how-to-bypass-matlab-cripple-amd-ryzen-threadripper-cpus Cripple AMD CPU]' functionality that suppresses AVX/AVX2/FMA code paths for AMD CPUs even though these are capable.
This behavior can be bypassed with an undocumented environment variable set during runtime:
<pre>
export MKL_DEBUG_CPU_TYPE=5
./a.out
</pre>
In tests with [https://software.intel.com/en-us/product-code-samples Intels MKL sample codes] setting this environment variable has lead to a 2-3x better performance on a Zen 2 AMD CPU.
This environment variable is currently set as default for all users.
<font color="red">'''Attention'''</font>: It seems that with '''mkl/19.1.1''' this bypass (which worked on ''mkl/19.1.0''') has been '''removed''' thus this version is not made available.
Intel replaced the classic legacy compiler with the [https://www.intel.com/content/www/us/en/developer/articles/technical/adoption-of-llvm-complete-icx.html LLVM-based] oneAPI compiler, which should also support non-Intel CPUs and starting with MKL Release 2020.2 there seem to be specific functions for AMD Zen architectures. But it seems that the best possible performance can be achieved adding a function ''int mkl_serv_intel_cpu_true(){return 1;}'' to overwrite and fake an Intel CPU, which is just another hack.
== Further libraries ==
In order to get an up-to-date list, please refer to
<pre>module avail</pre>

Latest revision as of 16:19, 26 April 2023

AOCL

The AMD Optimizing CPU Libraries (AOCL) are optimized to utilize AMD EPYC processors in an optimal way.

A module is available gcc and aocc with

module load aocl/2.2.0

Keep in mind this module only contains the proprietary non-MPI components like amd-libm.

The libraries utilizing MPI will be compiled separately with MPT and provided as their own modules at a later time.

Library Purpose Source License URL
amd-blis/amd-blis-mt BLIS is a portable software framework for instantiating high-performance

BLAS-like dense linear algebra libraries.

https://github.com/amd/blis 3-clause BSD https://developer.amd.com/amd-aocl/blas-library/
amd-fftw An AMD optimized FFTW that includes selective kernels and routines

optimized for the AMD EPYC™ processor family.

https://github.com/amd/amd-fftw GPLv2 https://developer.amd.com/amd-aocl/fftw/
amd-libflame libFLAME is a portable library for dense matrix computations,

providing much of the functionality present in LAPACK.

https://github.com/amd/libflame 3-clause BSD http://developer.amd.com/amd-cpu-libraries/blas-library/#libflame
amd-libm amd-libm implements optimized trigonometric, logarithmic/exponential, power, etc.
functions that should perform better than the system libm.
Provides vector intrinsics of these as well (currently for __m128 __m128d types).
proprietary AMD LibM EULA http://developer.amd.com/amd-cpu-libraries/amd-math-library-libm/
amd-rng AMD Random Number Generator Library is a pseudorandom number generator library.
It provides a comprehensive set of statistical distribution functions and various uniform
distribution generators (base generators) including Wichmann-Hill and Mersenne Twister.
proprietary AMD RNG EULA http://developer.amd.com/amd-cpu-libraries/rng-library/
amd-securerng The AMD Secure Random Number Generator (RNG) is a library that provides APIs to access
the cryptographically secure random numbers generated by AMD’s hardware-based
random number generator implementation.
proprietary AMD SecureRNG EULA http://developer.amd.com/amd-cpu-libraries/rng-library/#securerng

Intel MKL

Intel MKL is available with

module load mkl/<version>

For linking flags for different compilers, there is an Intel MKL Link Line Advisor

Static linking and 32-bit integer interface are recommended if applicable.

For distributed MKL functionality (ScaLAPACK, etc.), select SGIMPT for the MPI library, if you use mpt.


Intel MKL seems to be using a 'Cripple AMD CPU' functionality that suppresses AVX/AVX2/FMA code paths for AMD CPUs even though these are capable.

This behavior can be bypassed with an undocumented environment variable set during runtime:

export MKL_DEBUG_CPU_TYPE=5
./a.out

In tests with Intels MKL sample codes setting this environment variable has lead to a 2-3x better performance on a Zen 2 AMD CPU.

This environment variable is currently set as default for all users.

'Attention: It seems that with mkl/19.1.1 this bypass (which worked on mkl/19.1.0) has been removed thus this version is not made available. Intel replaced the classic legacy compiler with the LLVM-based oneAPI compiler, which should also support non-Intel CPUs and starting with MKL Release 2020.2 there seem to be specific functions for AMD Zen architectures. But it seems that the best possible performance can be achieved adding a function int mkl_serv_intel_cpu_true(){return 1;} to overwrite and fake an Intel CPU, which is just another hack.

Further libraries

In order to get an up-to-date list, please refer to

module avail