Recently we had our application crashing on a single PC (having an Intel® Core™ i7-13700HX), while it was still performing good on other ones (Intel® Core™ i9-9900K, Intel® Core™ i7-11700 and Intel® Core™ i5-12500E) with the same input data.
We found a way to make it working on the failing one, just disabling HyperThreading in BIOS.
Analyzing the dump I found that an IPP operation (ippiPyrDownGetBufSize_Gauss5x5, pyramid decomposition, from IPP legacy) on a 32 bit floating point image wasn't carrying out the expected results (after four levels of processing, the fifth level produced an image with some -NaN inside, which in turn carried out to an access violation when trying to perform a LUT operation).
As we are using a quite old version of TBB (2020.2) and IPP (2020.0.166 and IPP legacy 9.0.0.009), I was wondering if these libraries can potentially have some issues when used on hybrid CPU like the i7-13700HX which has P and E cores (while the i5-12500E, despite being of the same architecture, doesn't have any E core as per specs).
