APL support in pytorch

[meghaanaaa]

I was going through the pytorch source code and wanted to use APL for blas and lapack operation however in FindAPL.cmake there is a check for library in the binary directory
“FIND_PATH(APL_BIN_DIR NAMES armpl_lp64.dll libarmpl_lp64.a PATHS ${APL_BIN_SEARCH_PATHS})”

I have installed arm performance library using the below link:

Arm Performance Libraries | Arm Learning Paths’%20to%20complete%20the%20installation

I don’t see any libraries in bin directory. Is this particular patch written in pytorch for a specific APL version. Currently I am using 24.10 version of Arm performance library.
It would be really helpful if someone gives an intuition on why this is happening.

cc @malfet @seemethere @peterjc123 @mszhanyi @skyline75489 @nbcsm @iremyux @Blackhex @snadampal @milpuz01

[malfet]

I think this script was added to support APL on Windows, please do not hesitate to propose a patch to extend it to Linux

[milpuz01]

@meghaanaaa Alternatively on Linux you can use ArmPL instead of OpenBLAS by putting in LD_PRELOAD path to where you have installed ArmPL libraries and it will pick up implementation of BLAS function from ArmPL instead of OpenBLAS as it implements the same interface

[meghaanaaa]

@malfet @milpuz01 Thanks for pointing it out, this is a relatively newer patch which caused errors when I was trying to link armpl as of now i have just commented things i felt was not necessary. I just wanted to be sure of this particular error.

There is also another thing I want to mention here, when I build with APL, lapack support is not provided from APL because of the below check.
check_function_exists("cheev_" APL_LAPACK_WORKS) in FindLapack.cmake.

However if i do an object dump of the static library (libarmpl_lp64.a) that is linked in FindACL.cmake I do see this particular symbol but I think it is undefined, should I link with shared library instead because doing that didnt cause any error.

Excuse me for any technical errors I may have made here, I am new to this space.

[milpuz01]

Hi @meghaanaaa,

ArmPL should have cheer_ function defined. This are the steps that I did to install ArmPL:

wget https://developer.arm.com/-/cdn-downloads/permalink/Arm-Performance-Libraries/Version_24.10/arm-performance-libraries_24.10_deb_gcc.tar
tar xvf arm-performance-libraries_24.10_deb_gcc.tar
cd arm-performance-libraries_24.10_deb/
sudo bash ./arm-performance-libraries_24.10_deb.sh

This installs library in /opt/arm and then if I check for that symbol I find it defined:

$ nm /opt/arm/armpl_24.10_gcc/lib/libarmpl_lp64.a | grep "T cheev_"
0000000000000000 T cheev_

Hope that helps.

Best,
Milos.

[meghaanaaa]

Thanks for the reply, I have checked it in object dump once again i do see that the function exists but the cmake file is not able to recognize it, the check_function_exists("cheev_" APL_LAPACK_WORKS) fails to pass. Is there something i should do like preload the library(if yes i cannot preload static library which the existing script is trying to link). As of now I am using the below commands to build pytorch:
export BLAS=APL
export ARMPL_DIR=/opt/arm/armpl_24.10_gcc/
python setup.py develop

[milpuz01]

Thanks @meghaanaaa for sharing the steps how to reproduce. There are actually two problems when running it on Linux platform. The first one is that presence of wrong file was checked in bin directory of ArmPL so it was failing to report that it found ArmPL. The second one is that in order to detect that function cheev_ is presented it has to be linked against math library -lm when using C compiler. The changes that should fix those issues are in #150031 so if you pick that up you should be able to see the following messages in the output of build process.

-- Trying to find preferred BLAS backend of choice: APL
-- Found APL header: /opt/arm/armpl_24.10_gcc/include
-- Found APL library: /opt/arm/armpl_24.10_gcc/lib
-- APL_LIBRARIES: /opt/arm/armpl_24.10_gcc/lib/libarmpl_lp64.a

-- Looking for cheev_
-- Looking for cheev_ - found
-- Looking for cgesdd_
-- Looking for cgesdd_ - found
-- Found a library with LAPACK API (apl).