I like SIMD, but before super-optimizing your code with SIMD and the like, really consider your data structures and access patterns.
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>
}
Now my tree has all the bad characteristics of a linked list (* n nodes * m children), all the fragmentation of multiple heap vectors (* n nodes), and terrible set-up / tear-down time (in this case, Drop alone was taking up a good chunk of runtime).
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.
This is my eternal battle as a perf engineer. Performance starts with architecture and you can only squeeze so much out a hotpath with poor data layout.
The nice part is that data-oriented code almost always easily supports threading and SIMD.
Yeah, data layout/cache aware layouts are really key if you really want to unlock making something that ends up in a hot loop fast with SIMD.
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.
> Vectors (in C++) at least aren't necessarily the best fit either
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.
The active disdain a significant portion of this site has for actually understanding how computers work and how to make actually fast programs is kind of staggering.
Here's a helpful video about leveraging SIMD to solve a concrete performance problem for the dev team that made the game The Witness by Casey Muratori: https://www.youtube.com/watch?v=Ge3aKEmZcqY
I'd slightly rephrase the title to "everyone should know when SIMD didn't happen." Modern compliers are extremely good at vectorization until they suddenly aren't, an they'll often fall back to scalar code because if assumptions or a single-data dependent branch. Learning to check the compliers optimization reports is arguably more valuable.
99% of developers should just ignore SIMD. Most projects have a lot of low hanging fruit to increase performance, and still nobody finds the time to solve them.
That doesn't mean you should default to a slower implementation for new code. If you do, you're just creating more low-hanging fruit that nobody will find time to solve.
in many cases often its faster just to switch from debug to release - compilers are good to vectorise many loops. Worth to give it a try before rewriting clean loop/code into SIMD/NEON.
To bolster the argument, even if you do not plan to write the SIMD yourself or will "just get AI to do it", it is important to know what can be fast in SIMD (and on what hardware). That allows you to design your algorithms and structure your code so that the SIMD is possible.
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 think the big miss is that the former is achievable far more often than people imagine, which leads to settling for the later, aka pessimization. Reducing your allocations from millions per run to handfuls per run during initialization is effectively a "go fast" button, and is generally more reliable. It is very common for teams to spend big effort getting a 2-3x speedup on allocation-heavy code by disabling branches when a 100-1000x speedup can be had if restructuring allocations is a strategy under consideration (even before the extra 4-8x you might see if you go all the way to hand-tuned SIMD).
Everyone doesn't need to know SIMD. Mechanical sympathy is an important passive perk for software architects to cut down the number of reworks down the line, but I would rate benchmarking and being able to identify bottlenecks as more important everyday skills.
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...).
> Mechanical sympathy is an important passive perk for software architects to cut down the number of reworks down the line
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...
The Go language has long lacked official support for SIMD instructions, which means it has been at a disadvantage in terms of performance optimization. In recent years, with Go 1.26, an experimental version of the SIMD/ArchSIMD packages was introduced for AMD64 architecture. With Go 1.27, a portable version of the SIMD package was also added. Now, we can fully utilize native SIMD instructions to optimize go program performance.
Tangentially for Go programming, the last time I looked at optimising some Go code with SIMD there were a few different options available, but they were either not maintained any more or had incomplete support and required first writing your function in C++ with intrinsics and generating assembly, then converting it to go assembly with a tool [1]. I never got my function to work in go despite the C++ code working fine. In short, not really a production ready option for Go. This was a year or two ago, though.
i was having a conversation with a friend recently about simd in zig (which i have recently picked up and been having a pretty good time with). i find that simd writes decently well, though there's a few weird things:
- 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?
> mitchell, i know you hang around some of these comments sometimes
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?
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).
ahaaa, yeah, i don't personally do any runtime switching but i hear that as a deal-breaker from other folks.
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!
> 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).
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)
The loop dependency on a poly isn't all that hard for compilers to unroll, and most correctly rounded implementations are polys. You're often paying only a couple cycles' worth of stalls.
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.
i don't disagree – i have my own internal vector lib of approximations and whatnot for various tradeoffs of precision and speed, so i just use those. it's just that zig has a pretty strong stance of "no unexpected/obscured code execution" so it was surprising to see a vector-capable function that was just a bunch of scalar functions in a trench coat.
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.
Love Mitchell’s writing and he’s one of the few people in the industry that I truly admire.
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.
"More importantly, when this loop matters enough for me to care about a 5x speedup, I want the vectorization to be explicit and predictable. I don't want an unrelated code change or compiler update to quietly turn it back into a scalar loop."
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 an interesting article. I don't really work with low level enough languages for this to matter (unless - does this ever show up in Javascript somehow?).
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"?
no macros in zig, but yes you could metaprogram it. types are first class values at compile time so you could do that sort of specialization if you wanted.
One of my favourite articles is "SIMD-friendly algorithms for substring searching" by Wojciech Muła [2]. If you were unfamiliar with SIMD and just jumped into the code, it'd be incomprehensible due to the intrinsics, but the generic algorithm description at the top is pretty simple if you take some time understand it.
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]
In the article, Mitchel mentions how this doesn’t always work. In fact, as someone who’s worked in compiler development, I can say it’s a small miracle when it does work.
Case-in-point, the example in my own post doesn't auto-vectorize with LLVM or GCC at highest optimization levels. Basically, compilers will never auto-vectorize loops with an early loop break afaik.
Most scalar-to-SIMD conversion requires changing the design of data structures and algorithms to be effective. Compilers are required to exactly reproduce the specified data structures in a deterministic way for obvious reasons.
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.
And -march=native or at least -march=x86-64-v3 or similar, alternatively identifying relevant functions and manually invoking FMV and uarch specialization via target_clones. Plus non-integer code can generally not be autovectorized in normal-math mode since FP is non-commutative.
I have no idea why you're being downvoted. HN has a fetish for SIMD, but if you are hand-rolling SIMD and you aren't writing an explicit acceleration library, you're doing it wrong. Like, 100% of the time.
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.
I was hand-rolling NEON SIMD 15 years ago, and in many cases the compiler (clang/llvm) simply out optimized me. I kept the attempts that were better than what the compiler could already do. That was ARM NEON, quite new at the time, not SSE, and again that was 15 years ago that the compiler could already beat me much of the time. I hate the naive cult coder adage concerning premature optimization, but this might be a situation where you peruse your compiler output before you start writing code in a manner the compiler can for you. In Clang, auto-vectorization is enabled by default at optimization levels -O2 and -O3.
I respect (and fear a little) those who intentionally utilize SIMD in their implementations, but I believe it's a bit too much of a semantic shift for 2-5x performance gain. Good news is that modern compilers are more than capable of emitting SIMD code even if original source is nothing but.
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.
> Good news is that modern compilers are more than capable of emitting SIMD code even if original source is nothing but.
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.
I like SIMD, but before super-optimizing your code with SIMD and the like, really consider your data structures and access patterns.
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:
Now my tree has all the bad characteristics of a linked list (* n nodes * m children), all the fragmentation of multiple heap vectors (* n nodes), and terrible set-up / tear-down time (in this case, Drop alone was taking up a good chunk of runtime).
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...
This is my eternal battle as a perf engineer. Performance starts with architecture and you can only squeeze so much out a hotpath with poor data layout.
The nice part is that data-oriented code almost always easily supports threading and SIMD.
Yeah, data layout/cache aware layouts are really key if you really want to unlock making something that ends up in a hot loop fast with SIMD.
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.
> Vectors (in C++) at least aren't necessarily the best fit either
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.
1 reply →
This. Tables are an efficient implementation of general graphs. It’s the best one I know of (unless your graph can be specialized).
The active disdain a significant portion of this site has for actually understanding how computers work and how to make actually fast programs is kind of staggering.
Here's a helpful video about leveraging SIMD to solve a concrete performance problem for the dev team that made the game The Witness by Casey Muratori: https://www.youtube.com/watch?v=Ge3aKEmZcqY
I'd slightly rephrase the title to "everyone should know when SIMD didn't happen." Modern compliers are extremely good at vectorization until they suddenly aren't, an they'll often fall back to scalar code because if assumptions or a single-data dependent branch. Learning to check the compliers optimization reports is arguably more valuable.
99% of developers should just ignore SIMD. Most projects have a lot of low hanging fruit to increase performance, and still nobody finds the time to solve them.
That doesn't mean you should default to a slower implementation for new code. If you do, you're just creating more low-hanging fruit that nobody will find time to solve.
in many cases often its faster just to switch from debug to release - compilers are good to vectorise many loops. Worth to give it a try before rewriting clean loop/code into SIMD/NEON.
1 reply →
In most cases you should skip SIMD also for new code. Often a non-SIMD version is required for compatibility, just stick with that.
> Every developer should… most importantly, not be scared of 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.
[1] https://crates.io/crates/fearless_simd
To bolster the argument, even if you do not plan to write the SIMD yourself or will "just get AI to do it", it is important to know what can be fast in SIMD (and on what hardware). That allows you to design your algorithms and structure your code so that the SIMD is possible.
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.
> what can be fast
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 think the big miss is that the former is achievable far more often than people imagine, which leads to settling for the later, aka pessimization. Reducing your allocations from millions per run to handfuls per run during initialization is effectively a "go fast" button, and is generally more reliable. It is very common for teams to spend big effort getting a 2-3x speedup on allocation-heavy code by disabling branches when a 100-1000x speedup can be had if restructuring allocations is a strategy under consideration (even before the extra 4-8x you might see if you go all the way to hand-tuned SIMD).
Everyone doesn't need to know SIMD. Mechanical sympathy is an important passive perk for software architects to cut down the number of reworks down the line, but I would rate benchmarking and being able to identify bottlenecks as more important everyday skills.
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...).
> Mechanical sympathy is an important passive perk for software architects to cut down the number of reworks down the line
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...
The Go language has long lacked official support for SIMD instructions, which means it has been at a disadvantage in terms of performance optimization. In recent years, with Go 1.26, an experimental version of the SIMD/ArchSIMD packages was introduced for AMD64 architecture. With Go 1.27, a portable version of the SIMD package was also added. Now, we can fully utilize native SIMD instructions to optimize go program performance.
- https://pkg.go.dev/simd/archsimd@go1.26.5
Tangentially for Go programming, the last time I looked at optimising some Go code with SIMD there were a few different options available, but they were either not maintained any more or had incomplete support and required first writing your function in C++ with intrinsics and generating assembly, then converting it to go assembly with a tool [1]. I never got my function to work in go despite the C++ code working fine. In short, not really a production ready option for Go. This was a year or two ago, though.
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.
[1] https://github.com/minio/c2goasm
i was having a conversation with a friend recently about simd in zig (which i have recently picked up and been having a pretty good time with). i find that simd writes decently well, though there's a few weird things:
- 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?
> mitchell, i know you hang around some of these comments sometimes
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).
ahaaa, yeah, i don't personally do any runtime switching but i hear that as a deal-breaker from other folks.
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!
2 replies →
> 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).
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)
The loop dependency on a poly isn't all that hard for compilers to unroll, and most correctly rounded implementations are polys. You're often paying only a couple cycles' worth of stalls.
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...
[1] https://godbolt.org/z/rGGq8aG38
i don't disagree – i have my own internal vector lib of approximations and whatnot for various tradeoffs of precision and speed, so i just use those. it's just that zig has a pretty strong stance of "no unexpected/obscured code execution" so it was surprising to see a vector-capable function that was just a bunch of scalar functions in a trench coat.
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 bet there's a better way than unpacking, running sequentially and repacking. Even if the algorithm is very branchy you save a pack and unpack.
https://sourceware.org/glibc/wiki/libmvec ?
Love Mitchell’s writing and he’s one of the few people in the industry that I truly admire.
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.
"More importantly, when this loop matters enough for me to care about a 5x speedup, I want the vectorization to be explicit and predictable. I don't want an unrelated code change or compiler update to quietly turn it back into a scalar loop."
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.
Proving the point that the compilers aren't deterministic as some folks argue.
This is especially painful with dynamic languages, like in JavaScript's case.
However JIT also have positives hence their widespread use.
This is an interesting article. I don't really work with low level enough languages for this to matter (unless - does this ever show up in Javascript somehow?).
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"?
Yes. "`@reduce(.And, ...)` combines every boolean using `and` and returns a single boolean."
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.
Maybe I'm wrong, but should it be @clz instead?
2 replies →
My compiler knows SIMD. However, knowing the limitations of SIMD might help avoid a calculation that can't be optimized to use it.
Your compiler can vectorise trivial things, once your code gets complicated it will no longer vectorise.
I don’t know zig syntax, but wouldn’t it be possible to put this common pattern into a macro and simplify it to mostly a lambda on V?
no macros in zig, but yes you could metaprogram it. types are first class values at compile time so you could do that sort of specialization if you wanted.
The biggest barrier I faced learning SIMD is the weird naming convention for intrinsics.
Once I got past that, it was fairly simple. mcyoung's articles were also super helpful
One of my favourite articles is "SIMD-friendly algorithms for substring searching" by Wojciech Muła [2]. If you were unfamiliar with SIMD and just jumped into the code, it'd be incomprehensible due to the intrinsics, but the generic algorithm description at the top is pretty simple if you take some time understand it.
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]
1. https://news.ycombinator.com/item?id=44274001
Easier than ever, with RISC-V Vector (RVV), which is part of RVA23.
Is interesting that this is how array langs work.
I bet will be easy to turn into a lib.
I think more developers would use SIMD if there were macros handling the details.
SIMD does not pay, speaking from 20yr exp in the field in various semis.
Not really, not as SIMD remains an esoteric art from packing matrix and vector operations in endless opcodes.
I rather let the compiler auto vectorise itself, or with AI help.
I just do gcc -O3 and get SIMD without having to learn it
In the article, Mitchel mentions how this doesn’t always work. In fact, as someone who’s worked in compiler development, I can say it’s a small miracle when it does work.
Case-in-point, the example in my own post doesn't auto-vectorize with LLVM or GCC at highest optimization levels. Basically, compilers will never auto-vectorize loops with an early loop break afaik.
1 reply →
Most scalar-to-SIMD conversion requires changing the design of data structures and algorithms to be effective. Compilers are required to exactly reproduce the specified data structures in a deterministic way for obvious reasons.
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.
auto-vectorization is not nearly as good as you would hope it to be.
The best SIMD optimizations likely require changing your data format from AoS to SoA.
The one feature in Jonathan Blow's Jai language I really envy is a a single keyword to switch AoS to SoA and visa-versa at comptime
1 reply →
Either this or you have to do special tricks like pairwise tree reductions and hand-unroll certain portions of loops.
What are AoS and SoA?
3 replies →
And -march=native or at least -march=x86-64-v3 or similar, alternatively identifying relevant functions and manually invoking FMV and uarch specialization via target_clones. Plus non-integer code can generally not be autovectorized in normal-math mode since FP is non-commutative.
Well, then I just prompt Claude and get SIMD without having to learn it /s
I have no idea why you're being downvoted. HN has a fetish for SIMD, but if you are hand-rolling SIMD and you aren't writing an explicit acceleration library, you're doing it wrong. Like, 100% of the time.
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.
I was hand-rolling NEON SIMD 15 years ago, and in many cases the compiler (clang/llvm) simply out optimized me. I kept the attempts that were better than what the compiler could already do. That was ARM NEON, quite new at the time, not SSE, and again that was 15 years ago that the compiler could already beat me much of the time. I hate the naive cult coder adage concerning premature optimization, but this might be a situation where you peruse your compiler output before you start writing code in a manner the compiler can for you. In Clang, auto-vectorization is enabled by default at optimization levels -O2 and -O3.
[flagged]
[flagged]
Everyone should know about SIMD, so you can know when to ask your good friend Al, who knows how to do it, to use it for you in the right places.
I respect (and fear a little) those who intentionally utilize SIMD in their implementations, but I believe it's a bit too much of a semantic shift for 2-5x performance gain. Good news is that modern compilers are more than capable of emitting SIMD code even if original source is nothing but.
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.
> Good news is that modern compilers are more than capable of emitting SIMD code even if original source is nothing but.
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.
>Good news is that modern compilers are more than capable of emitting SIMD code even if original source is nothing but.
This is only true if you are intentionally writing code that the compiler can easily vectorise. Which is not most code.
> Reduce obvious DB/server round trips
In my mental model of optimization, the above and SIMD are basically the same thing.
There is nothing to fear or respect about using SIMD. Don't lionize learning. You can learn too, if only you allow yourself to.