Matt Pharr (author of the SIMD programming language ISPC) nailed it with the insight "Auto-vectorization is not a programming model."

As Matt writes ([1]) : " The problem with an auto-vectorizer is that as long as vectorization can fail (and it will), then if you’re a programmer who actually cares about what code the compiler generates for your program, you must come to deeply understand the auto-vectorizer. Then, when it fails to vectorize code you want to be vectorized, you can either poke it in the right ways or change your program in the right ways so that it works for you again. This is a horrible way to program; it’s all alchemy and guesswork and you need to become deeply specialized about the nuances of a single compiler’s implementation—something you wouldn’t otherwise need to care about one bit."

[1] https://pharr.org/matt/blog/2018/04/18/ispc-origins

Vectorization doesn't imply SIMD, of course. The first vectorizing compilers were for CDC(?) systems long before SIMD. Today you have SVE in Arm, for instance, distinct from SIMD Neon.

Anyway, I'm familiar with optimizing numeric loops in C (and Fortran) rather than Rust. I've rarely seen simply using SIMD intrinsics work where GCC auto-vectorization didn't with the same semantics (like numeric equivalence in reductions). In most cases you can get away with -fassociative-math, of course, and not sacrifice peak performance, e.g. BLIS passes its extensive tests with it on, but you should check, of course. (GCC also documents the option as necessary to get Arm (Neon?) to vectorize at all.) Most of the time when people tell you how much better the Itel compiler is, it's because it incorrectly defaults to something like -funsafe-math.

Regardless, GCC (like other compilers) will tell you about vectorization with the -fopt-info- options without examining assembler, and you can have some surprises. For instance, you use unsigned in C for loop indices that you know are positive, and see failed vectorization due to "loop not affine", because of C's overflow semantics; use signed types instead.

There's another reason for using properly-optimized numerical libraries (typically BLAS), is that, at least for level three (matrix-matrix) operations. Even if you get the blocking right for the memory hierarchy, you typically won't get peak performance just with vectorization because tricky preloading is needed for the inner loops.

I do this. I do this a lot. My job involves processing a significant amount of weather data and flight telemetry _very quicky_.

What I have found is that getting rust to auto-vectorize is a nightmare. The options available to me have always converged around: 1) use simd-like apis 2) frame it as matrix-vector operations.

(1) is touched on in TFA

(2) is way easier, and allows use of well-tested apis and crates, each of which (sensibly) call out to better-tested C libraries. Each of those can, should, might, or will use your CPU better than you will. If you can frame it as matrix-vector operations, you will go very fast, not least of which, by stacking the operations into a _big_ matrix/vector op, which your CPU will happily tear though.

However the article proposes a third, very cool option: use algebraic ops API! Worth a read.

   Rust used to not have any reasonable way to do anything like this on stable (for my own definition of reasonable), but now it does! Rust 1.98 stabilized algebraic operators for floats. These allow you to declare per-operation that you’re okay with the compiler making optimizations that may change the result as long as the optimized code is algebraically equivalent to the original. Yes, this includes potentially reordering operations.

It has been a while, but I was quite surprised by how good gcc/g++ was at explaining why it had failed to vectorize a certain loop. At that time clang was being positioned as the better-than-gcc at optimization and error messages and it turned out that on my code it was the other way round -- hence the surprise.

I had written a expressions template C++ helper library with sort of the same functionality as Python's itertools before I was familiar with itertools.

This was for my own consumption. I expected very little from g++ and it had me impressed. Would be around 2008 - 2010.

For a couple of days I experimented with auto-vectorization in my language which uses LLVM as its backend. In my example I had a simple function multiplying elements from two arrays and writing them in the third one. And it worked as expected, I did see autovectorization in the result code. But I also noticed something strange - there were a couple of pointer difference calculations and jumps prior to the unrolled/vectorized loop. Then I realized, what this means. LLVM isn't sure that it can use autovectorization, since input and output arrays may overlap. So, it just performs a runtime overlapping check and if no overlap is detected, a vectorized version of the loop is executed, otherwise it jumps to a non-vectorized loop version.

That's why C has restrict keyword. If it's provided, LLVM is sure that no runtime overlapping check is needed and can autovecrorize, sometimes even in places where such vectorization without restrict isn't possible at all.

Interesting, but would make for a much stronger article with a proper discussion of performance, backed by measurements.

This would allow the reader to better weigh the balance of code damage against any gains in performance.

Unless I missed it, it just seemed to be hinted at, but not presented.

This is cool, but I mostly wanna see the numbers.

How much faster did it actually get?

If I have to make the code kinda weird just so the compiler behaves, I wanna know it was worth it lol.

maybe unrelated by NumPy and Cython have made it incredibly easy to write vectorized code.