Everyone should know SIMD
mitchellh.com
mitchellh.com
I just wouldn't start off with bold sentences as
> SIMD can be simple to understand
and
> writing SIMD is just about as easy as a for loop
and then the first example requires 12 lines to replace one line of scalar code.
Be honest and say SIMD is hard but the results are worth it!
(Another nitpick: if this article is for newbies, don't use SIMD-only words and concpts before explaining them. Step 5 is good: scalar tails are mentioned and described. Step 1 is bad: nobody is supposed to know what broadcast mean.)
And it turns out sometimes you cannot show it is simple, because it is in fact very complex. But every complex topic is made up of smaller, simpler ones. Good teachers then manage to find a good order of those smaller parts that makes the steep hill climbable. Then you only need to convince people it is actually worth climbing.
I try to ask myself whether things I write are (1) writing and (2) reading before writing. Sometimes that works.
Simple Mail Transfer Protocol
Lightweight Directory Access Protocol
Sometimes I think the RFC editors are trolling us.
In fact getting things to be simple can often be very hard and cumbersome.
(I've both played bridge actively and written SIMD code professionally, bridge rules are way simpler. Actually playing good bridge is probably harder.)
They covered that, in a very honest and blunt manner.
My favourite feature was
par(<start expression>; <end expression>; <loop expression>)
Everyone Should Know AI
AI has a reputation for being complex. I've met many very good software engineers who dismiss it as something too complex to learn or a niche meant for only highest-performance, not useful everyday. I think that's wrong. AI can be simple to understand, and common AI editorializing can speed up a blog, and almost always follows the same general shape. Once you learn the basics, editorializing with AI is just easy. And when it's not, it's usually a good sign to skip it for now. Every developer should know at least that much AI.> and then the first example requires 12 lines to replace one line of scalar code.
> Be honest and say SIMD is hard but the results are worth it!
I think SIMD, and certainly that first example is way more tedious than it is hard.
What makes it tedious are
- you have to figure out how many things your hardware can do in parallel
- you have to chop up the work in packets of that size
- once you have your results, you have to ‘unchop’ them
- if there is a chance ‘chopping up’ leaves you with remaining items, you have to handle that case separately
- if the code you want to apply SIMD to uses constants, you have to create vectors containing copies of them
Neither of those is particularly hard, but each adds work, making things tedious.
https://github.com/gitdepierre/cppshader
Not sure it will ever be useful, but it was a fun pet project with some interesting problems to solve.
I got the idea to write the library when I wanted a simplex noise function in C++, implemented with AVX2 for performance reasons (because why not). There were a lot of public HLSL/GLSL implementations that were concise and fast on GPU, some of which I had used for years for my own needs, but rewriting the whole code in plain C++ would have been a hassle, and i would have to do the same for every GPU code i came accross in the future, so building a wrapper seemed like the most efficient approach.
So in the end, the library is mostly aimed at people coming from the GPU world who want to keep most of their habits. But yeah, there is probably very few use cases for it.
<No annotation> – Gently suggests inlining
#[inline] - Really suggests inlining
#[inline(always)] – Really Really suggests inlining (still not guaranteed!)
More valuable than learning to write SIMD code? When writing SIMD code is itself the remedy to lousy autovectorisation?
If you can only identify the problem, you’re left with “well, that sucks”.
Where do I start?
I want to trust the compiler, but I don't always have time to feed every little piece into compiler explorer and interpret it. Is there a higher-level workflow?
Although as mentioned plenty of times, the largest wins tend to come from laying out your data correctly, or switching to a different algorithm. And heuristics. ;)
Internalizing things like how data dependencies matter, how expensive it is to increase the width of your vector elements (and how to avoid the need), how to turn conditions and branches into masks, or simply things like "division does not exist" becomes a lot easier when you have spent at least some time trying to use SIMD yourself.
I think this doesn't get talked about enough. If your input is a big run of data that is being checked/transformed in one shot, it works well. But, if you're likely to have to make a decision on several bytes of the input, SIMD will be the same or slower than the scalar method. It's not a magic "go fast" button.
I have been told SIMD is good for data-parallel loops, like the GPU is, and that is true, but it can also be used piecemeal, unlike the GPU. Because it is just the CPU, you can read 32 unaligned bytes, scan for the index of the first space, and take a branch based on that.
I am speaking exactly of simdjson. If you look closely, you'll see that the high performance is really only achieved by skipping over large chunks of the input data. Which is great, if your use case is operating on a subset of data. But, if the application really needs to process every piece of the entire input, there's no performance win.
Think about it, JSON structure is almost entirely made up of single bytes ({ } [ ] , : "). If you're not skipping over stuff, you're not going to get away from having to examine those bytes and branch on them. For something like `{"a":1,"b":2}`, the SIMD scan has a bunch of overhead to find the interesting parts and then you go back and practically have to reprocess every single byte.
It's a great example of what I think of as vertical integration for performance. As you go through the talk you can understand why all these abstractions exist and why they have to be so generic. But when you have a specific use case, you can vertically integrate from the problem definition all the way down to SIMD and reap big rewards.
Make sure to link back to the source, of course, and don't pass it off as your own words.
https://github.com/seclorum/SIMDSynth
It has been a very rewarding experience, and the synth architecture - multitimbral polyphonic - provides a great stream of data for applying SIMD principles, i.e. multiple streams going through the same process.
Has been pretty hard to debug, though. I found myself wishing I had some sort of simulator to help me understand the state of things in each pipe. I suppose I should spend some time investigating SIMD tools next time I get into this - but I fear it'll require a lot more investment. If anyone has any tips, I'm all ears ..
I've been singing Data-Oriented Design's praises, so I'll just collect all my comments here [1], but I think it's a good approach to optimization. I played around with SIMD in my old code (in Zig), but my approach to modelling datastructures was so antithetical to optimization, it was like putting high-performance racing tires on a lemon with a broken engine.
It was the root-of-all-evil-type-premature-optimization, because I wasn't measuring performance, and I wasn't thinking about where the allocations were, etc. Now, I try to model my data as if it were SQL tables, see what my potential "primary keys" could be, and build my data structures around my access patterns.
For example, I used to model trees as structs pointing to other structs on the heap:
struct Tree {
tag: TreeTag,
children: Vec<&Tree>
}
But a tree can be represented a million ways, and can always be linearized. So now I really consider my access/insert patterns of the tree, whether it's really a tree or some other sort of graph, whether I can store it in a Vec or a Struct of Vecs, etc. Since really looking at things through their access patterns and "primary keys", my code has been much faster and simpler.
This has the added effect that a lot of your data ends up in homogeneous arrays / vecs, which means that the compiler can do its SIMD magic, the CPU can read it from your L1 cache a million times faster, etc. And then when you need to drop down into SIMD yourself, you can write some awesome branchless code.
1. https://hn.algolia.com/?dateRange=all&page=0&prefix=true&que...
Also, avoiding allocations or vtable lookups or a lot of indirection in the part of the code that's actually "hot" is really important. Vectors (in C++) at least aren't necessarily the best fit either, if you end up doing anything that can call an allocation unexpectedly.
I'm not sure if you use a different allocation strategy or if you're advocating allocating as much as possible up-front, but I'm curious if you have any thoughts on this:
I always end up using (Rust) vectors despite looking at a bunch of slab/arena allocation libraries. Preferably I'd know how much memory I need up front, but barring that I see three options for any allocation that needs to grow:
- Fail;
- Reallocate; or
- Put the overflow in a new allocation, keeping track of where all the "pages" are internally.
In 2, you can't use references/slices (anything with a pointer) because the potential reallocation invalidates those. In 3, it seems ideal because references can stay stable, but you wouldn't be able to have any array-like data cross the "page" barrier as the pointer jump would not be stable. (Although, am I correct in thinking the OS does something like this, and that's why pointer addresses are virtual?).
So, you can't really internally reference data in this sort of context by reference/pointer, it's preferable to use an integer index. In that case, what's the point of the allocator libraries at all? Your standard Vector would have the same reallocation/access characteristics, and you can set a reasonable initial capacity to try and avoid reallocations.
[0] https://man7.org/linux/man-pages/man2/mremap.2.html
[1] https://git.musl-libc.org/cgit/musl/tree/src/malloc/mallocng...
It feels dirty, but when your system has overcommit you can’t reliably check that the memory you want is actually there anyway, so you might as well reap the benefits.
Windows is cleaner, as it lets you reserve address space and then later commit it. The Linux equivalent to reserve is probably mapping with no permissions. Overcommit can be disabled on Linux and doesn't break control flow integrity or valgrind, so I know there is a way.
The problem with overcommit, is that there is no reliable way to know in advance how much memory your program can actually use. That system with 16GB of RAM will allow you to "allocate" 128GB of memory. Just try it, call `malloc()` or the `with_capacity()` constructor, it will return a valid pointer or container. Try to write in that giant "buffer" at the beginning, at the end, somewhere in the middle at random… it will still work. Everything works exactly as if you really had a giant 128GB buffer, that you can write to and read back from…
…Until you hit somewhere below 5M different pages, where instead of getting a new page after the page fault, you get killed by the out of memory killer. With SIGTERM if you have the relevant privilege, so you have at least a chance of exiting cleanly, but user programs just get SIGKILL. No appeal, no way to check.
Well there is a way to check, kinda: allocate a fixed amount of memory up front, hit all the pages, and if your program didn't crash, it really has the amount of memory you just gave it. Then perform all your allocations within this real buffer you really have (this means a custom allocator). Any allocation success will be real, and bounds checks will actually work. It just doesn't play well with programs whose actual memory needs are highly unpredictable: most of the time you'll use much more memory than you need, and sometimes you won't have enough, forcing you to raise the threshold.
The nice part is that data-oriented code almost always easily supports threading and SIMD.
- Struct of Arrays (SoA) vs Array of Structs (AoS);
- Row-major order vs column-major order; and
- Row-based vs column based / columnar (in databases)
Most code has arrays of structs (or "lists of objects", the effect is the same), and most databases are row based (same thing). But many game engines use SoA and keep heterogeneous elements together. Some databases like DuckDB do this too, this is an article about the pros and cons of columnar storage in DBs [2].
In fact, I'm reasonably certain that R (known as S in the 70's) was the first real language designed around this concept.
[1]: https://en.wikipedia.org/wiki/APL_%28programming_language%29
> This [...] applies to any programming language. Support for SIMD instructions varies by programming language
This is a very pedantic nitpick because this article is good (& getting SIMD support across more languages would also be good), but the "every programmer should know" line feels a bit odd when neither of the 2 most popular languages natively support SIMD.
Seems like he should be recommending fearless_simd [1], the Rust crate by Raph Levian and the folks at Linebender :)
More seriously, if you’re looking to add SIMD to your Rust code, that’s the package to start with.
Edit, there's now an experimental official library at https://go.dev/pkg/simd/archsimd/ see https://go.dev/doc/go1.26#simd and at https://github.com/golang/go/issues/78902 so things have moved since I tried it last.
- some builtins purport to work on simd vectors but actually just unpack the vectors and do their work per-element (e.g. running `@sin()` on a `@Vector(4, f32)` will unpack the vector, run `@sin()` 4 times, and then pack it back into a vector).
- a lot of `std.math` is scalar-only (some functions support vectors, though, and i've got a pr open for one of them and plan to do more).
- i'm certainly missing some intrinsics that i get from xmmintrin.h (rcp, rsqrt, few others).
in general though i'm finding it pretty capable.
mitchell, i know you hang around some of these comments sometimes – i noticed that in ghostty you bring in some c++ libs to do the simd heavy lifting for you. any plans to port that to zig? anything missing from the language or libs that's preventing it?
hi im here
> i noticed that in ghostty you bring in some c++ libs to do the simd heavy lifting for you. any plans to port that to zig? anything missing from the language or libs that's preventing it?
No plans to port it. For others, this is referencing highway: https://github.com/google/highway
The major limitation of Zig's vectors is that they're compile-time only. So if you're building redistributed software that compiles for a baseline CPU target, it won't be as optimized as it could be for YOUR possible machine.
Highway compiles our SIMD modules for different hardware configurations and at startup does a CPUID fingerprint to figure out which to load. That way even baseline has AVX512 etc. implementations, and we just activate the right one at runtime.
We only use Highway for our hottest hot paths that we feel benefit from that specialization.
No plans to port that (although, I spent hundreds of dollars and slop-forked it into Zig with the help of this good boy GPT and it worked great actually, but I didn't want to maintain it).
it's interesting – i've found that zig tends to extend my vectors to the native width of the platform and then operate on them there. e.g. i had a `@Vector(2, f32)` that i was using as a demo and the generated assembly was promoting it to 256 bits and using avx2 instructions on it!
oh interesting. though i suspect that isn't zig and thats llvm.
this is testing in isolation as well, could be that in the midst of other vector code it changes things. llvm definitely does a great job optimising tightly-written vector code to be even faster.
this is reasonable because there isn't really a generalizable "good way" to unroll trig functions for simd. if you really care about speed youre better off implementing to the precision you care about (you might not want full precision)
maybe functions that don't actually support actual vector execution just shouldn't work on vector arguments. i also wouldn't expect `@sin()` to expand in-place out to a full cephes-like sin implementation. maybe a function call.
i think that's reasonable.
We can see this in action. LLVM implements sin() with a correctly rounded double poly [0]. Let's ignore range reduction and throw the core into compiler explorer to be autovectorized [1]. uiCA estimates a theoretical latency for the inner loop of 15 cycles, and the code achieves 18.
I suspect most custom implementations would do worse than this.
[0] https://github.com/llvm/llvm-project/blob/165c472d65cd62eb33...
Only tangentially related but this is by far the most painful part about optimizing code for JIT compilers like V8. Even changing a constant from 1 to 1.0 somewhere else can change the optimizations performed and lead to an unexpected performance decrease.
This is especially painful with dynamic languages, like in JavaScript's case.
However JIT also have positives hence their widespread use.
https://github.com/kelindar/simd
However, having it officially supported is definitely much more convenient.
All of these provide a similar set of features and you can use normal arithmetic operations (+, -, *, etc) for SIMD vectors. Together with templates/generics you can also write code that can deal with any vector width. These get compiled to LLVM vector types and will generally give you pretty good generated code.
This is a very good way of writing basic SIMD code and has the benefit that your code can be compiled to multiple instruction sets. I've been working on a project that can compile down to SSE2, AVX2, AVX-512 and NEON, with just a change of compiler options. Somewhat surprisingly I get the best performance by using 2x the native vector width (ie. f32x16 = 512 bits on 256 bit AVX2), which is kinda like unrolling the loop once.
There are some caveats, though. You will need to keep an eye on the generated assembly code to make sure you're on the happy path. You will inevitably need to drop down to ISA specific intrinsics every now and then (for that fast reciprocal square root with `__mm_rsqrt_ps` etc).
As an example I needed to do a gather load from an array of fp16's on AVX2, which does not do 16 bit loads. Rust's `Simd::gather_select` takes 64 bit usize as the index but AVX2 doesn't do 64 bit indices. But as long as I did all the index arithmetic in 32 bits and cast to usize at the last second, the compiler did what I wanted. But you need to kinda know what is available in the ISA to stay on the happy path. Not really an issue with arithmetic.
I'm sure that an experienced SIMD programmer can get better performance by writing intrinsics manually (say 5-20% better) but I'm already at 3-6x better than the scalar implementation I started with. And you'd have to write (and benchmark) the code for each ISA separately, meaning that you'd spend at least five times more time with it (and have 5x more code to maintain).
[0] https://gcc.gnu.org/onlinedocs/gcc-4.6.1/gcc/Vector-Extensio... [1] https://doc.rust-lang.org/nightly/std/simd/index.html [2] https://en.cppreference.com/cpp/numeric/simd
I guess I don't understand the "reduce" step. It seems like you have to be careful not to "undo" all the benefit from SIMD. Sure, it can compare 8 values in parallel, but then if you have to look at each of the 8 answers in turn you're back to where you began. Is the `@reduce()` function in the example a special Vector one that tells you if all the values are true or not in one "step"?
If it's true (the common case here) then you proceed to look at the next 8 bytes.
If it's false, you apply a @bitcast (turn the booleans into bits) and @ctz (find the first 0) to get the index of where it was false.
Yes, it is. In SIMD you can do a reduce for a commutative function in O(log N) steps. Sum, product, min, max, all, any, etc.
Although in this case it might be better if you just take the mask (vector of booleans), convert it to a bitmask and check if it is (non-) zero.
If the language provides it, you might also use .all() or .any() for a mask.
I don’t need it to be optimal, just … handy as an option!
The last time I brought this up here, folks offered a bunch of options that don’t quite do this, and the best candidate was this 15 year old compiler project that is Intel specific!
Could some programming language nerd build this?
(While you are at it give me a clear idiomatic way to pay the cost to switch from array of structs to struct of arrays)
It was bad 11 years ago. It's worse now.
They are back to pre 2022 levels now! https://fred.stlouisfed.org/series/APU0000708111
I doubt GPU is included by most runtimes yet, but for the rest of that have you tried SQL?
The big problem with databases, in my opinion, is the horrible API friction between them and your code (not even SQL per se). It makes sense if you're calling out to a database server and transferring data, but for the small, intermediate values we see in code every day, the relational model is amazing and yet so painful to use within a given programming language.
I've been envying the C# people and their LINQ, and the Java people and their jOOQ, because I'm either making a half-assed database in my own code with structs, arrays, and hashmaps, or I'm constructing some SQL monstrosity, shoveling it out to SQLite or DuckDB through a library, and marshalling the types back and forth.
Why can't I just have everything I want all the time?
The reason we don’t have this is that it’s a holy grail, an unsolved problem, for quite fundamental reasons.
The other comment about SQL hints at why: SQL is largely declarative and has complex semantics built into the language, which allows for analysis and optimization that go beyond what’s possible for lower-level, general purpose languages, especially imperative ones.
For code in those languages, even just determining whether “arbitrary loop like code” is parallelizable is undecidable in general.
The challenge is that as you make a language expressive enough to describe arbitrary algorithms, you also make it progressively harder for a compiler to infer safe and useful parallel execution automatically.
Another big issue is that the various forms of parallelism are only similar at a very high level. They have fundamentally different execution models and constraints. Translating arbitrary imperative code to handle that essentially involves first inferring the intent of the code, then rewriting the code, including how data structures are organized, to fit the target architecture. This is far more than what ordinary compilers do.
There are also a lot of choices involved. Parallelism isn’t always free, so you’d need to make sure that the costs don’t outweigh the benefits - and you’d need to do that for many different decisions, like whether to use threads or not. Now you’d have a compiler building cost models to try to not make dumb choices - and without actually restarting and comparing alternatives, it’ll make mistakes.
In many ways, you’d be better off using an LLM for this, because that’s the level of understanding you need to have a hope of getting a good result.
That all said, you can do much better with more constrained languages or frameworks. SQL is the most successful example of that. Java’s streams and Rust’s Rayon only target multicore CPUs, but similar approaches could be used to do more. (Although you still potentially run into issues with optimal data shape across paradigms.) Languages like APL, J, and Futhark are all relevant.
The other family of solutions to this are the frameworks like Apache Spark, Apache Beam, and the ML frameworks like Pytorch. The latter lets you describe (tensor) computations at a high level, leaving the framework free to figure out how to implement them - much like with SQL.
It's called ParaSail. The next closest thing to ParaSail is ...
... literally just Rust.
Why? Because ParaSail has completely eliminated pointers, thereby preventing pointer aliasing. It can't be understated that pointer aliasing and alignment are the two biggest bottlenecks preventing autovectorization.
The people talking about better compilers, etc, just don't get it. It's not a compiler problem, it's a language semantics problem.
Pointer aliasing prevents parallelism full stop. If there is a single memory region and you perform a write to it, you have to assume that the write invalidates all data loaded from the pointers. If you make pointer aliasing illegal, then you have guaranteed that each pointer points to a distinct subset of the global memory, turning each pointer into a pointer to an isolated region. This means you have multiple regions you can write to in parallel. This is crucial, if you do not understand this you don't get parallel programming at all.
I mean think about it, this is the difference between having a four toilet bathroom with a single door or four doors.
The thing about alignment is not as easy to explain, but here is my attempt at it: If you allow the array to be misaligned at the front, then you have to run scalar code at the front. Same problem if you misalign at the end.
If you have a naked pointer and a for loop (think C), then the alignment problem alone precludes autovectorization without a complex scalar preamble. Autovectorization has to assume that the loop length could be anything, meaning it could be less than 8 elements to begin with. If the loop is always a multiple of 8 and always aligned, then the loop can be autovectorized even if the loop only does a single iteration.
Of course, after these critical blockages are gone you're still stuck with the problem of having to write branchless/non-diverging code.
> All of the objects declared in a given scope are associated with a storage region, essentially a local heap. As an object grows, all new storage for it is allocated out of this region. As an object shrinks, the old storage can be immediately released back to this region. When a scope is exited, the entire region is reclaimed. There is no need for asynchronous garbage collection, as garbage never accumulates. Objects may grow in a highly irregular fashion without losing their locality of reference.
> Note that pointers are still used behind the scenes in the ParaSail implementation, but eliminating them from the surface syntax and semantics eliminates the complexity associated with pointers.
This approach seems to come up in many contexts, where a pointer-based address (raw pointer, reference, slice, etc.) is abstracted into a higher-level address key, usually an index integer (essentially a higher-level virtual pointer). The implementation might reallocate under the surface, or manage chunks of data through some sort of paging where the underlying pointers don't change (I was musing about this here [2]).
I guess I'm thinking out loud here, but most dynamic languages (eg. Python) don't expose the pointers or care about invalidation of the addresses, they just happily reallocate. I've barely written parallelized code, is pointer aliasing really one of the biggest roadblocks? It seems like it can be abstracted away fairly easily, even in a pointer-exposing language.
1. https://www.adacore.com/uploads/papers/parasail-pointer-free...
* If you aren't profiling, don't bother
Otherwise, you are going to be wasting plenty of time on "optimizations" that don't actually do anything useful.
Just come up with at least a few test that represent important cases and time them. Dig in, see where the time is being spent, and focus on areas where significant time is spent, especially ones that look ripe for optimization.
Also:
* Think about laying out your data in ways that accommodate your access patterns and are cache-friendly. SIMD would typically follow from this, not be a starting point on its own.
Does zig have auto vectorisation? I'm thinking that if you write the code in a vector friendly way, then the compiler can do the boiler plate for you
https://llvm.org/docs/Vectorizers.html
https://inside.java/2025/08/16/jvmls-hotspot-auto-vectorizat...
I guess what I'm looking for is something that addresses the common idioms in SIMD. For example, want to do concurrent data look up? This is how you do it in SIMD; this is how you search; this is how to prepare your data in a manner conducive to SIMD operations; these are the data types you typically find in programming languages, like __mm128; so on and so forth.
Eh. It's both. You don't have one without the other. SIMD is pretty much entirely extensions to base processor ISA; there might be a few architectures that have SIMD as a basic part of them (GPUs are one of them, taken to an extreme order), but for the most part you have to know which instruction set you're working with.
SIMD is basically "pack multiple data points into a single CPU register, then to another, and perform some operation on them as if you ran that op on all of the pairwise data points individually". Some ops are binary (arithmetic, bitwise, etc) and some are unary.
Some have "gates" whereby you can do a comparison, the boolean output of which is stored in a bit packed integer. Then you can run conditional instructions after that that only perform the instruction if the bit in that variable is set, creating "constant time" SIMD instructions with what amounts to branching. Really depends on the instruction set's capabilities.
AVX512 is far and away one of the most extensive extensions, with a ton of super niche instructions meant for enterprise number crunching. It has weird stuff, like swapping bytes, collating them, doing all sorts of weird manipulations. But the number of x86 CPUs that support that instruction set are small.
You can't emit code that has e.g. AVX512 and just run it on a CPU that doesn't have that extension. You get a CPU exception and it crashes the process (or, if this is in kernel/driver land, your machine).
So they're kind of tied together.
Anyway if you want to see a list of them, Intel's SIMD intrinsics site has always been really nice to browse IMO.
https://www.intel.com/content/www/us/en/docs/intrinsics-guid...
imo this is one of the greatest libs ever written. it handles dynamic dispatching of correct simd instructions / lane widths for various hardware with just one simd loop written (handling NEON/AVX/AVX2/AVX512/extensions) with comparable performance to handwritten native intrinsics
I'm working on a voxel space renderer homebrew for the PlayStation. I only have so many cycles to spend on rendering before it becomes a slideshow, so I count them in my hot rendering loop and parallelize work as much stuff as I can, even across memory load stalls from main RAM.
I've worked on a basic network accessory card with a STM32 MCU that is extremely overkill for what it needs to do. We haven't bothered making any performance or memory optimizations whatsoever, writing plain C++ almost as if we were on server-class hardware because we had such egregious margins.
The first question to ask is not whether something can leverage SIMD, it's whether the performance requirements are met or not (although it's far too easy to not care when it's not your hardware that's struggling...).
We might be talking on different levels but when it comes to, on an opposite end of 'Should I use SIMD'... a databases a level of mechanical sympathy at a 'base' level is still important. e.x. row-by-row updates vs batching or bad logic where a 21k entry in clause forgot about unicode rules on columns and breaks an index [0]... is still super important.
[0] - That one is real, thanks lazy bodyshop having their people use copilot and yet, we get the same billable hours, nothing is done faster, and management is too stupid to pay attention...
This is what so many people miss when doing micro-benchmarks of language X vs Y, sure Y might win out in execution speed, however if X delivers within the performance requirements and has a lower development cost, it wins out while being slower than Y.
Naturally taken to the extreme, when it isn't our hardware is how we end up with Electron apps.
This will only be effective if the data you’re working on is smaller than 64 bits. If you’re working with bytes, for example, you might get 8x parallelism.
Once I got past that, it was fairly simple. mcyoung's articles were also super helpful
It blew my mind once I understood what was happening, because it's quite clever but one of those "I could've thought of that" algorithms. There was some pretty good discussion on it last year (feat. ripgrep). [2]
Later CPUs (updated skylake xeons and later, AMD Zen 4 and newer) don't have the issue
Thanks for the deeper dive
AMD Zen 4 and Zen 5, and also those Intel CPUs with AVX-512 support starting with Ice Lake, behave much better and there is no reason to avoid AVX-512, which has much better energy efficiency than the alternatives.
It's been at least a decade for me since SIMD was more than an implementation detail handled by the runtime. And even then it was "how do I avoid preventing the runtime from vectorizing this".
NumPy-like:
mean = np.mean(array)
maximum = np.max(np.abs(array)
return mean / maximum
Highway:
HWY_FULL(float) d;
using V = decltype(hn::Zero(d));
V sum = hn::Zero(d);
V max = hn::Zero(d);
hn::Foreach(d, values, N, hn::Zero(d), [&](auto d, auto v) HWY_ATTR {
sum = hn::Add(sum, v);
max = hn::Max(max, hn::Abs(v));
});
const float mean = hn::ReduceSum(d, sum) / N;
return mean / hn::ReduceMax(d, max);
When the actual computation is made so much faster by SIMD, memory bandwidth starts to be a significant bottleneck.
I bet will be easy to turn into a lib.
There might be a narrow class of problem where:
- SISD is the bottleneck
- Compiler won't autovectorize
- Data is small or weird enough, or the environment constrained enough that running it on an accelerator is not feasible
But you need a better color scheme for your light theme my guy. Greys on greys on whites with bright pinks and light blues…it’s really hard to just read.
If you work on a performance-critical application, that's fine, but most of the time we use some interpreted / VM / garbage-collected slow environment that performs so much slower than just using hardware directly, and we are fine with it. Under those circumstances, it is much better to have code that every developer can understand than to introduce a 5x improvement, with code that only the person who wrote it understands.
I mean, e.g. for Go, this simplicity was one of the explicit design goals. For C / Rust / Zig, it might be a different story, but I hope you choose those languages only if they fit your use-case.
Portable SIMD is (perma?) nightly.
Requires 'unsafe' everywhere.
To skip bounds-checking, you typically need to switch your writing style from loops to iterators.
No JIT, so you need multiversioning and/or target-cpu=native. Since Rust doesn't bring a compiler, you need to predict all your target architectures in advance.
Cannot inspect @code_llvm/@code_native at the function level like Julia, you need to compile the entire app.
Prioritizes Floating-Point strictness over --ffast-math. It's a genuine win for safety, but that's overkill in some domains (e.g. graphics, games, audio).
I rather let the compiler auto vectorise itself, or with AI help.
The best SIMD optimizations likely require changing your data format from AoS to SoA.
I read a book book about DoD [3] really which confused me at first with all its talk about database table design (in a book about a high-performance C++ game engine?), but when it finally clicked it was amazing. The point is that you want to think hard about your access patterns and what could constitute good "primary keys", then model it accordingly. SoA ends up being useful a lot of the time, because having your data in homogeneous arrays/vectors is great for cache locality and branch elimination. Even without SIMD you can get huge speedups from that, but that's also where your compiler (or you as a programmer) can get incredible SIMD gains.
SoA is not a silver bullet as it may not align well with your access patterns, but it can great to add to your toolkit.
---
Mike Acton: Data-Oriented Design and C++: https://www.youtube.com/watch?v=rX0ItVEVjHc
Andrew Kelley: A Practical Guide to Applying Data Oriented Design: https://www.youtube.com/watch?v=IroPQ150F6c
Richard Fabian: Data-Oriented Design: https://www.dataorienteddesign.com/dodbook/
We do have gather load instructions in SIMD instruction sets these days (AVX2 and newer), so AoS vs SoA is not nearly as important as it was once.
Scatter stores are also available but only in newer CPUs.
You start the post with:
> There is an opportunity to use SIMD. SIMD turns those into this: > > for (8 byte chunk in bytes) { /* ... */ }
If you actually wrote that loop, there is a good chance the compiler (gcc specifically) will auto-vectorize.
In any case, the more manual SIMD optimizations I have seen require reworking the data altogether, not just processing N elements at a time. For example, instead of packing two 4-vectors into two registers to do a dot product, pack the XXXXs, YYYYs, etc. into 4 vectors and compute 4 dot products for the price of one. That not only requires having 4 vectors to process, but also thinking how exactly they are packed in registers.
I don't know why qurren is downvoted. You really should see if you can get the compiler to auto-vectorize first (possibly padding data structures and loops) before you write anything by hand.
Counterpoint: https://pharr.org/matt/blog/2018/04/18/ispc-origins
> I think that the fatal flaw with the approach the compiler team was trying to make work was best diagnosed by T. Foley, who’s full of great insights about this stuff: auto-vectorization is not a programming model.
> 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.
> And God help you when they release a new version of the compiler with changes to the auto-vectorizer’s implementation.
> With a proper programming model, then the programmer learns the model (which is hopefully fairly clean), one or more compilers implement it, the generated code is predictable (no performance cliffs), and everyone’s happy.
And what does the last statement in the quote mean anyway? When is performance "predictable"; do you freeze the entire toolchain?
And what is the alternative? Handroll manual SIMD code for every possible architecture you may target?
If you're writing C++, you're already rolling on decades of compiler optimization. You return by value because it makes code more readable and safer and rely on RVO. You write functions to abstract and rely on the compiler inlining. When it doesn't work for your specific target/toolchain, you may decide to handroll stuff. The proof that you're banking on the compiler is that if you run a debug build of any non-trivial program, it runs like absolute dogshit.
When it can be single-handedly responsible for an 8× speedup, you might not want to rely as much on the compiler as in the case of smaller, cumulative optimisations.
> And what is the alternative? Handroll manual SIMD code for every possible architecture you may target?
Use a library like Highway? https://github.com/google/highway
> compilers will never auto-vectorize loops with an early loop break afaik.
I think this is changing. https://godbolt.org/z/ea1E7dx9v. GCC 14 won't try to vectorize this because of the break statement. But GCC 15 does vectorize it. This got a callout in the "General Improvements" section of the release notes: https://gcc.gnu.org/gcc-15/changes.html
I don't think there's any equivalent in clang/LLVM.
Even if compilers were clever enough to transform your data structures and algorithms for SIMD (they're not), the data structures are a contract that can't be unilaterally modified.
Every modern language has a vectorization optimizing compiler, and through some fairly straightforward techniques this is automagic. And contrary to the various replies, unless you screwed something up compilers are really good at vectorizing on whatever hardware you're targeting, including SVE.
What if the “explicit acceleration library” for what you need to do doesn’t exist?
Most software's poor performance would be fixed long before SIMD comes into play. Reduce obvious DB/server round trips, bad data structure/cache locality, poor algorithm complexity, doing redundant computations, etc.
In my mental model of optimization, the above and SIMD are basically the same thing.
They're really not (I have a whole section on it in the blog post). This example in the post doesn't auto-vectorize, for example. And its a pretty big part of the overall throughput for plain text runs (ascii or unicode). Really, the point of that section is that almost nothing auto-vectorizes, backed up by LLVM docs and published research.
Instead, writing 12 lines for a 5x gain is way easier than crossing your fingers and hope someone else pays your bills.
Bigger picture, the real point is that this stuff isn't complicated. You wouldn't copy and paste 100 lines because you hope the compiler "lifts this into a for loop", you just write the for loop cause you know how and its simple.
Similarly, the common case of "process N values in parallel" is very simple. Write a dozen lines of code you're comfortable with. No need to pray the compiler people saved your bacon.
This is only true if you are intentionally writing code that the compiler can easily vectorise. Which is not most code.