Bug in ippsFIR_32f on Avx

[Nono_Dontcall]

There's a bug in ippsFIR_32f on Avx Cpu-s.

Symptom:
Numerical output of ippsFIR_32f is wrong.

Platform:
SandyBridge (i7-2600K), win7/64, 64b code, IPP7.0.4.

Reproducing the bug:
- Method 1: Compare output of ippsFIR_32f with numIters=1 with output of ippsFIROne_32f. They are different for most filter lengths that are bigger than or equal to 32 (ex: 32, or 100, or 1000, etc).
- Method 2: Use attached code (VS2010): IntelIppFirAvxBug.zip

Pls - what/when should I expect in terms of fixing it?

[levicki]

How much difference there is?

I am asking this obvious question because sometimes there are small numerical differences betwen scalar and vectorized code output and they are usually in favor of vectorized code -- it can be 4 times (or in case of AVX 8 times) more precise than scalar code. At least that is what I have observed with the code I developed.

Of course, it may also be a bug.

[Nono_Dontcall]

I'm not talking about small (and normal) differences caused by finite FP precision.
The output is plain wrong.

If you run attached code you'll see results such as:
>IppFir():
>32 (-15.700001, 21.400000), 33 (99.699997, 23.600000), 34 (-8.599999, -59.699997
>), 35 (-14.199997, -19.799999), 36 (0.800000, -3.800000), 37 (-90.899994, -8.700 ...
where the messages mean "filterLength (referenceResult, IppFirResult), "

Regards,
Teo

[levicki]

I do not have AVX capable CPU yet so I couldn't run the sample, that is why I asked for clarification.

That is indeed wrong. You will get faster response if you submit an issue on premier.intel.com.

[Gennady_F_Intel]

Igor, that's no completely true. We are tracking the forum and escalating the forum's problems also.

Particularly, this problem has been already escalated to the enginnering team.

[levicki]

I see...

That must be why my post with bug report fpr UIC was ignored for 5 days straight. I learned my lesson, deleted it, and submitted an issue instead. I still have an unanswered topic I created here two days ago, not to mention that I never got proper answer for deinterlacing functions in IPP and I asked about that some 2-3 years ago.

[Nono_Dontcall]

Thanks Igor,

I followed your advice, posted an issue on intel premier and then Gennady reproduced the bug and forwarded it to the engineering team.
I'm waiting with interest for engineering team's updates.

Regards,
Teo

[levicki]

You are welcome.