Changes to the routine CGETRS between MKL 11.1.3 and MKL 11.3.3

[Jens_E_]

Hi,

After an upgrade from MKL 11.1.3 to MKL 11.3.3 I suspect that the mkl routine CGETRS gives slightly different results. This sounds very strange (since I believe it just forwards the calculation to LAPACK), so I wanted to ask here if there has been any change to the implementation between these MKL versions before I dig deeper into the issue?

Best,
Jens

[Zhen_Z_Intel]

Hi Jens,

Could you please attach a screenshot of this difference or provided your sample code with calculation, I could tested on different version of MKL to see if this problem remains. Thank you so much.

Best regards,
Fiona

[Jens_E_]

Hi,

This turned out to be something unrelated to the abovementioned routine. The compiler update caused a roundoff-level difference (as expected) in some single-precision floating-point code, which through some other sensitive code led to a slighly different input to CGETRS which again gave a slightly (but correct) different output. The sensitive part of the code has been fixed now.

Best and thanks,
Jens