Bug in mkl_avx2.1.dll after a partial QR decomposition from LowRankApprox.jl ?

[BambOoxX]

Note : related to this Discource thread and JuliaLinearAlgebra/LowRankApprox.jl#35

After performing a partial QR decomposition from LowRankApprox I get a julia crash throwing this error on julia 1.5.4 with MKL.jl installed

The error disappears with julia 1.6.0 using OpenBLAS.

I could not extract an MWE showing the problem from my original code, but the errror seems to show up as soon as I try to access the factors of the partial QR.

I open this issue hoping for some pointers, as it doesn't seem to stem from LowRankApprox and is maybe a bit too specific for the Discource forum.

Note that the conventional QR works fine with ohterwise the same code however.

[jarlebring]

If you provide an MWE you are more likely to get help. Have you eliminated Julia version differences? Here is some start of a problem search.

If it happens during Q-access, you may want to first trace the code

julia> QRfact=qr(randn(5,4));
julia> @which(QRfact.Q)
getproperty(F::LinearAlgebra.QRCompactWY, d::Symbol) in LinearAlgebra at /home/jarl/archive/src/julia_latest_1.7/julia-29d5158d27/share/julia/stdlib/v1.7/LinearAlgebra/src/qr.jl:432

and then https://github.com/JuliaLang/julia/blob/fc02458492c60f6527245c6991f729c2a986f666/stdlib/LinearAlgebra/src/qr.jl#L437

Based on the stacktrace you may also want to look for where ztpsv (triangular system solve) is used . If the problem is due to memory allocation / deallocation (likely due to the "access violation" message), the problem is probably earlier than that.

[BambOoxX]

Well, thanks for the attention, but

1. As indicated, while I surely tried my best, I could not isolate an MWE from my original code
2. The problem does not show when using a standard QR decomposition, so I fail to see how studying the behavior with qr will help.

After subsequent research, though I cannot be entirely sure, this issue may occur due to a non-catched lapack error, so probably non specific to MKL.
This issue was posted here because the only obvious error message is related to MKL.

[jarlebring]

You need to provide more details in any case. What julia-function call exactly is generating this error message? What happens when you switch julia version? What do you mean by conventional QR? The julia function qr makes a blas call, so it also ends up in MKL.

In my experience this package works best on julia 1.7 - so nightly build.

[jarlebring]

Considering the error message is in ztpsv, I would start looking at ldiv!-calls with triangular matrices in LowRankApprox.jl assuming you get the error when calling a function from that package.

[BambOoxX]

You need to provide more details in any case. What julia-function call exactly is generating this error message? What happens when you switch julia version? What do you mean by conventional QR? The julia function qr makes a blas call, so it also ends up in MKL.

In my experience this package works best on julia 1.7 - so nightly build.

I am sorry to repeat myself, but as I said, I could not isolate more precisely this issue. When this appears, I only gets like 0.5 second to react before the terminal crashes with only this message printed on screen, before it closes. I will not invent information I do not have.
Running different versions do not change the issue either with 1.54 1.6.0 or 1.6.1.
"Conventional qr" is meant to clarify from the "partial qr" exposed in LowRankApprox.

[jarlebring]

That's a tough debugging environment. What command do you use to access the factors? I don't see why that would not be available even in your restricted setting. You seem to have been able to avoid it by replacing some code with "conventional QR". Which code is replaced?

Edit: I now saw you have provided more info on that in the discourse forum...

[jarlebring]

I know you said you tried your best to zoom in on the problem. This is how I would proceed. I would edit ~/.julia/packages/LowRankApprox/2wpw4/src/pqr.jl and add code to write a subset of parameters to file before and after every LAPACK-call.

[ViralBShah]

This package just replaces Julia's openblas with MKL under the hood. We have no way to address any issues in MKL ourselves - those need to be reported directly to Intel.

Perhaps some of this discussion and troubleshooting is better suited for Discourse.

[BambOoxX]

@ViralBShah yeah, sorry for that. I was a bit lost with this issue, and I opened it during my first days of julia. In the end and with some more insight into it, it seems it is not specific to MKL so it is better to close this issue. Thanks for your infos.

[ViralBShah]

No worries at all!