Fix ARMv7 build by making recent ZIP NEON optimizations be ARMv8 (aarch64) only

[aras-p]

Should fix #1365. Recent PR (#1348) added NEON accelerated code paths for ZIP filtering. But that code uses several instructions that are ARMv8 (aarch64) only, and thus fail building on 32-bit ARM (armv7) platforms. Make these optimizations only kick in when building for 64-bit ARM platforms.

[lgritz]

Aside, not really to Aras per se, but more a "note to the rest of the OpenEXR team, while this topic is on our mind":

In OIIO/OSL, the approach I took was to make simd wrapper classes -- within each method of which are #if clauses for SSE, AVX, NEON, or no SIMD -- and that header is the only place where any simd intrinsics can be found in the entire code base. I had to do this for my own sanity, as even with the experience of having written this, I can never remember what the cryptic intrinsic names mean, so code littered with them is just gobbledygook to me unless I have the intrinsic man pages open and am looking them up line by line. And then if I find something like this issue (I've used the wrong intrinsic, etc.) it's usually isolated to a single place in one header to fix, even if it is ultimately called all over the place. It also means that if a new ISA comes out (e.g. a new set of AVX extensions, or porting to ARM), I can merely add some new #if clauses in that header and don't need to recode any actual "algorithms".

I'm not necessarily advocating copying/using any of my code, and in fact there's probably plenty to criticize about my specific implementation. (As well as a variety of alternate implementations open sourced, with varying levels of complexity.) Just saying that isolating it in one header behind some simple wrapper classes can have big maintainability benefits.

[cary-ilm]

I'm groping in the dark without an AArch64 machine, but don't you mean the #define IMF_HAVE_NEON_AARCH64 to be inside the #if defined(__ARM_NEON) above? Don't you want this only if __ARM_NEON and __aarch64__? Or does __aarch64__ sufficient itself?

[aras-p]

64-bit ARM makes NEON the standard feature that's "always there", so no need for a separate check for it. Very similar to how x64 made SSE2 be guaranteed.

[cary-ilm]

Got it, thanks.

[cary-ilm]

Doesn't this duplicate what's in ImfSimd.h above?

[aras-p]

Yes, but internal_zip.c already did this for all SIMD sets including SSE. I don't know the reasoning, just went with the flow.

[cary-ilm]

Right, this file doesn't include ImfSimd.h, so that's appropriate.

[aras-p]

was to make simd wrapper classes -- within each method of which are #if clauses for SSE, AVX, NEON, or no SIMD -- and that header is the only place where any simd intrinsics can be found in the entire code base

Yeah that makes a lot of sense. OpenEXR does not use SIMD extensively (yet?), and for a very limited use the "SIMD wrapper" can be really short and compact. E.g. for my recent investigations into floating point data compression, handing both SSE and NEON in a single wrapper was total under 100 lines of code (simd.h). ASTC encoder also has a much more exntesive and fairly nice wrapper, astcenc_vecmathlib.h and friends.