IPP: cv::ipp::getIppTopFeatures() return value on 32-bit configurations

[alalek]

Windows 32-bit:

Intel(R) IPP version: ippIP SSE4.2 (p8) 2019.0.0 Gold (-) Jul 26 2018

cv::ipp::getIppTopFeatures() returns 0x8000 (ippCPUID_AVX2).
It should be 0x80 (ippCPUID_SSE42) because SSE4.2 backend is used.

relates #12877

[alalek]

/cc @maver1 @lupustr3

[alalek]

The same problem exists with IPPICV on MacOSX.
relates #18986

/cc @eplankin

[eplankin]

@alalek, what SSE4.2 backend means? Is it the machine or code that was dispatched?
which machine was used, avx2 or sse4.2?
I was able to reproduce the similar problem on avx2 machine - getCpuFeatures() results in avx2 that is right but getLibVersion() results in sse4.2 instead of avx2. Trying to find out if these issues are the same...

[alalek]

Nightly build reports no such issue anymore (workaround result is reported):

CPU features: SSE SSE2 *SSE4.1 *SSE4.2 *FP16 *AVX
Intel(R) IPP version: ippIP SSE4.2 (p8) 2020.0.0 Gold (-) Oct 21 2019
Intel(R) IPP features code: 0x80

Perhaps update to 2020.0.0 fixes it: #16639

[eplankin]

@alalek, shouldn't #18991 be reverted then?

[alalek]

Need to check....

Unfortunately reported value is after this workaround:

• Win32: [ INFO:0@0.002] IPP: replacing ippTopFeatures=00008000 => SSE4.2

So, this issue is not fixed yet.

[eplankin]

Then I need more information.
what SSE4.2 backend means? Is it the machine or code that was dispatched?
which machine was used, avx2 or sse4.2?
I was able to reproduce the similar problem on avx2 machine - getCpuFeatures() results in avx2 that is right but getLibVersion() results in sse4.2 instead of avx2. Trying to find out if these issues are the same...

[alalek]

Reproducible on:

• Windows 10 on Skylake machine. Build of 32-bit binaries.
• Or macOS X (64-bit).

---

what SSE4.2 backend means?

Not sure about IPP terminology, but I mean code path to execute SSE4.2 optimizations.
ippiGetLibVersion()->Name returns SSE4.2.
ippGetEnabledCpuFeatures() returns AVX2 - which is not available (IPPICV doesn't have AVX2 optimizations for 32-bit build to reduce code size).

[eplankin]

Due to an implementation of IPP dispatcher, ippGetEnabledCpuFeatures() returns all features that are supported on CPU except those which were manually disabled. Optimizations which the library contains (AVX2, SSE4.2, etc.) are not taken into account. As far as I understand, ippTopFeatures is used to define an optimization branch in IPP.
So the only way to know which optimization branch has been chosen is parsing of ippiGetLibVersion() ouput (ippiGetLibVersion()->Name).