DGESVD function sometimes doesn't return on AMD Processors (Ryzen 9 5950X & Ryzen 7 PRO 7730U)

[KaiserCurry]

Dear Intel people,

first time posting, so please let me know there is something wrong with the format.

This occurs on MSVC v143 C++.

Our four layers are:
mkl_intel_lp64_dll.lib
mkl_intel_thread_dll.lib
mkl_core_dll.lib
libiomp5md.lib

In certain situations which are reproducible in our software (but I am not sure yet what is fundamentally different to other situations where it works) the DGESVD call will fill the input buffer with results up to the requested amount but then simply won't return, which eventually leads to a freeze of our program as soon as the thread calling DGESVD needs to be waited on.

In other situations the rval isn't 0 and the result makes no sense.

But also in many other situations it works as intended.

Luckily it works on our intel based test stations but that still leaves potentially half our customers in dire straits.

Has anyone here ever seen something similar?

Thanks so much!

PS.: I have a small sample C++ project to show the issue, let me know if I should supply it.

PPS.: MKL version 2023.1.0 works, it's 2024.2.1.0 that exhibits this behavior for us.

[Ruqiu_C_Intel]

Hi,

Thank you for contacting to us.

Yes, please provide out a simple reproducer, as well as compiler, compiler options, working logs on Intel vs failing logs on AMD etc.

Regards,

Ruqiu

[KaiserCurry]

Hi Ruqiu,

please find attached a short VS 2022 project that does an SVD decomposition of "real world" data. We've tried it on 3 AMD computers here in our office and all of them show the same behavior: After calling DGESVD, the program hangs and does not return.

VS version 17.11.4
MKL version 2024.2.1

If you replace the mkl_def.2.dll with version 2023.1.1, it seems to work (at least DGESVD comes back and returns "0")

Regards,
Fabian

[Ruqiu_C_Intel]

Hi Fabian,

We reproduced the issue with your reproducer on AMD CPU. We will further investigating it and update this thread when there is improvement.

Regards,

Ruqiu

[Ruqiu_C_Intel · Accepted solution]

The error happens on some AMD processors has been fixed, the fix will be available in the future release.

[Jörn]

Thanks for the information.

For the time being, is it sufficient to replace the mkl_def.2.dll on our customers' computers with an older version from last year? Or do we have to rollback all libraries?

Regards,

Jörn