Illegal instruction on non-AVX processor

[kjkjava]

---

Required Info
Camera Model	Any
Firmware Version	Any
Operating System & Version	Ubuntu 14.04
Kernel Version (Linux Only)	4.4.0
Platform	Pentium
SDK Version	2.14.1
Language	C++

Issue Description

As mentioned in previous issues (#1428, #1443, #1464, #1484), there have been issues executing AVX instructions on CPUs that don't support AVX (causing SIGILL, "Illegal instruction (core dumped)"). Using SDK version 2.14.1, the issue is two-fold:

1. If I compile with -mavx, my application crashes at global initialization. I see that there is auto-selection code to bypass the AVX implementation if not supported on a CPU, but in debugging, I saw crashes in two different locations (one in librealsense and one in another library). Generally speaking, adding -mavx gives the compiler free-reign to optimize using AVX wherever it wishes, so I don't think this should always be enabled for all non-ARM builds.

2. If I remove -mavx, the build fails due to SSSE3 guards (which my processor supports) where AVX guards should be used instead. I have a patch for this, which is ultimately the solution to my issue. I'll send a pull request with this patch shortly.

[dorodnic]

Hi @kjkjava
You are right, and the patch makes sense.
I haven't seen any recent issues on the up-board lately which is our go to platform for non-AVX but SSE enabled CPU. That said, we certainly do not test on sufficient variety of ARM boards to cover all the corner cases, so your feedback helps a lot.

[dorodnic]

@kjkjava, with the PR merged into latest release, can the issue be resolved, or do you see additional problems related to AVX?

[kjkjava]

Hi @dorodnic,

Sorry for the delay and thanks for merging. I think my PR resolves issue 2 (doesn't compile with SSE and without AVX). Issue 1 may still be a problem, though. I think I read somewhere that AVX compilation is automatically enabled for non-ARM builds (and dynamically enabled/disabled), but this is not desirable for all applications; as I mentioned, other software may be AVX-optimized as the compiler sees fit.

[dorodnic]

Thanks @kjkjava
We compile with -maxv only image_avx.cpp (CMakeLists.txt:849)
So with your PR functions implemented in image_avx will never get called.
I don't think that the compiler will auto-generate AVX unless we explicitly allow it (after all machines without AVX are not that uncommon)
Do you think we are still missing some case?

[dorodnic]

I think we should be good now, but feel free to open a new issue / PR if you identify something.
Thank you!

[kjkjava]

Hey @dorodnic, sounds reasonable, especially since it's just image_avx.cpp. I think we had just enabled it for the entire build as a workaround for the failed build fixed in this bug.