Memory management when using amd blis with block2

[floatingCatty]

Hi!

I recently tried to use AMD-blis and AMD-libflame as the linear algebra backend for the numerical calculation. I bind the package with block2 software, while during the iterative computation, the memory keeps growing. The issue in pyblock2 package is: block-hczhai/block2-preview#163

The developer identified that the problem might caused by the AMD-blis library, can you help me check why this is the case, and how to address it? Thanks!

Best,

Zhanghao

[kvaragan]

Hi, Can you write a mail to toolchainsupport@amd.com.
If its direct issue on AOCL BLAS, we can address it here.
Thanks!!

[kvaragan]

Thanks for reporting, the details are mentioned here: block-hczhai/block2-preview#163
Will check !!

[floatingCatty]

Hi, Can you write a mail to toolchainsupport@amd.com. If its direct issue on AOCL BLAS, we can address it here. Thanks!!

Thanks for the quick reply!

Yes, it might be a direct issue on AOCL BLAS. DO I need to write an email?

[kvaragan]

Please do write.
In the meantime - can you try out: https://github.com/amd/blis/tree/dev
We update this branch biweekly.

$./configure --enable-cblas -t openmp amdzen --prefix=
$ make -j
$ make install

sup disabled will work - the code gets build, but we don't call those functions.

[floatingCatty]

Please do write. In the meantime - can you try out: https://github.com/amd/blis/tree/dev We update this branch biweekly.

$./configure --enable-cblas -t openmp amdzen --prefix= $ make -j $ make install

sup disabled will work - the code gets build, but we don't call those functions.

Hi I just tried to build the dev branch using <./configure --prefix=$PWD/install --disable-sup-handling --enable-cblas -t openmp auto>. The code takes a while to run to see the memory growth, hope it works. Thanks.

[floatingCatty]

Hi, this new dev branch wouldn't help. I switch to OpenBLAS for now. I tested with MKL and OpenBLAS, both worked well without the problem before. Hope this report helps to identify some possible issues.

[sireeshasanga]

Thank you for sharing the issue details. We will analyze further and keep you posted.

[kvaragan]

Hi, this new dev branch wouldn't help. I switch to OpenBLAS for now. I tested with MKL and OpenBLAS, both worked well without the problem before. Hope this report helps to identify some possible issues.

Hi, how do we reproduce this issue ?

[floatingCatty]

Hi, this new dev branch wouldn't help. I switch to OpenBLAS for now. I tested with MKL and OpenBLAS, both worked well without the problem before. Hope this report helps to identify some possible issues.

Hi, how do we reproduce this issue ?

Hi, I just sent a mail for the toolchains support for the detail of reproduction. I copied the reproduce info here also:

<
I installed the latest block2 in the repository. Located in block-hczhai/block2-preview: Efficient parallel quantum chemistry DMRG in MPO formalism
I am using Python 3.10.13, the system info is "Linux frea 4.19.87-amd64-nvm #4 SMP Thu Dec 12 07:06:36 EST 2019 x86_64 GNU/Linux".

I use this command to build the block2 "cmake .. -DBUILD_LIB=ON -DLARGE_BOND=ON -DMPI=ON -DUSE_MKL=OFF -DF77UNDERSCORE=ON -DUSE_COMPLEX=ON -DUSE_SG=ON -DBLAS_LIBRARIES=Path_to_AMD_BLIS -DLAPACK_LIBRARIES=path_to_amd_libflame -DCMAKE_BUILD_TYPE=Release -DCMAKE_CXX_FLAGS="-O3 -march=native".

To reproduce, there is an example I give in the block2 issue to reproduce the memory usage:
block-hczhai/block2-preview#163 (comment)

The example is in the params.zip file, the reply contains a code that can execute it once the block2_preview is installed.

Thanks!