> The other drawback of this method is that the optimizer won’t even touch anything involving floats (f32 and f64 types). It’s not permitted to change any observable outputs of the program, and reordering float operations may alter the result due to precision loss. (There is a way to tell the compiler not to worry about precision loss, but it’s currently nightly-only).
Ah - this makes a lot of sense. I've had zero trouble getting excellent performance out of Julia using autovectorization (from LLVM) so I was wondering why this was such a "thing" in Rust. I wonder if that nightly feature is a per-crate setting or what?
No it only uses the same LLVM compiler passes and you enable certain optimizations locally via macros if you want to allow reordering in a given expression.
For vectorizing, that quote is only true for loops with dependencies between iterations, e.g. summing a list of numbers (..that's basically the only case where this really matters).
For loops without such dependencies Rust should autovectorize just fine as with any other element type.
You just create f32x4 types, the wide crate does this. Then it autovectorizes just fine. But it still isn't the best idea if you are comparing values. We had a defect due to this recently.
LLVM autovectorizes many FP operations just fine, the article was a bit strange in that respect. Problem is, there are many other cases where it's unable to do so, not because it can't but because it isn't allowed.
In my experience, compiling C with -ffast-math will tremendously improve floating point autovectorization and optimizations to SIMD (C vector extensions, which are similar to Rust std::simd) code in general.
This obviously has a lot of caveats, and should only be enabled on a per function or per file basis.
Unfortunately Rust does not currently have options for adjusting per-function compiler optimization parameters. This is possible in some C compilers using function attributes.
No, because as I commented in another subthread, `-ffast-math` is:
1. dangerous assumptions hidden behind a simple, attractive-looking option [1]. It should be called -fwrong-math or -fdangerous-math or something (GCC does have the funnily named switch -funsafe-math-optimizations – what could go wrong with fun, safe math optimizations?!)
2. Translation-unit scoped, which means that dependencies not consented to "fast math" can break your code (as in UB land) or make the optimizations pointless, and your code can break your dependencies' semantics too via inlining. On the other hand, a library author must think very carefully what float opts to enable in order to be compatible with client code.
Deciding how the scoping of non-IEEE float math operations should work is a very nontrivial question. The scope could be a translation unit, a module, a type, a function, a block, or every individual operation, and none of those is without issues, particularly regarding questions like inlining and interprocedural and link-time-optimization, as well as ergonomics. In other ways, it's yet another function coloring problem.
There are currently-unstable "algebraic_add/mul/etc" methods for floats for letting LLVM treat those particular operations as if floats were real numbers [2]. They're the first step towards safe UB-free float optimizations, but of course those names are rather awkward to use in math-heavy code, and a wrapper type overloading the normal operators would be good to have.
No it doesn't. A global flag is a no-go as it breaks modularity. A local opt-in through dedicated types or methods is being designed but it's not stable.
Unfortunately SIMD in Rust tends to be pretty painful if you want to gracefully do runtime autodetection of a given SIMD extension (instead of it being a hard requirement for your program to even run).
The major problem is that Rust essentially requires you to annotate every (!) function in your whole call stack with e.g. `#[target_feature(enable = "avx2")]` to make sure that the SIMD intrinsics will actually get inlined (if they're not inlined then the performance is horrible, which makes using SIMD completely pointless). This makes it very hard to build any reasonable abstractions because you need to hardcode this all over your code. You can't have e.g. a `DataStructure<S>` where S is the SIMD ISA, so that you could do `DataStructure<AVX2>` or `DataStructure<SSE>` to get a nicely specialized version of it for a given instruction set. You need to copy-paste the whole thing with changed `target_feature` attributes (or use a procedural macro which does the copy-pasting) and have two entirely separate `DataStructureAVX2` and `DataStructureSSE` types.
That is, the `memchr` crate has a `Vector` trait that is generic over the vector type. Which is essentially your `DataStructure<S>` where `S` is the ISA. (Using the vector type isn't load bearing. I could do the ISA. But in `memchr`'s case, the vector type implies the ISA.)
This is a workaround, but it's still very painful, and there are still many problems with it:
1) you need to design the whole thing with this in mind; you can't just take an existing something which takes some T and get that to use SIMD,
2) `#[inline(always)]` can't be used with `#[target_feature]` at the same time,
3) you need to have `#[inline(always)]` on your whole call stack (I often don't want to inline everything, but I still want to propagate that we're in an SIMD-safe context, which you normally do with `#[target_feature]`, but that's incompatible with being able to abstract over things),
4) you have to sprinkle `unsafe` everywhere (which is especially annoying as recently we gained the ability to finally be able to call a lot of the intrinsics without any `unsafe`),
5) you have to manually make sure you're not calling something you should not be calling (`#[target_feature]` is great here because it prevents you from accidentally calling e.g. AVX2 intrinsics if you're not on the AVX2 codepath),
6) any mistake (e.g. missing an annotation somewhere) you make will just have the compiler silently not inline the intrinsics, completely killing your performance, and you're completely on your own debugging it.
I work with SIMD a lot in Rust, and unfortunately currently it's a miserable experience. It's not that bad when you're writing e.g. an algorithm which fits in a single function, but once you want to write something bigger with more sophisticated abstractions and/or you want to integrate SIMD into existing code then it becomes a world of pain.
Basically I just don't want to hear about "the state of SIMD in Rust" unless it is about dramatic improvement in autovectorization in the rust compiler.
80%-90% or so of real life vectorization can be achieved in C or C++ just by writing code in a way that it can be autovectorized. Intrinsics get you the rest of the way on harder code. Autovectorization is essentially a solved problem for the vast majority of floating point code.
Not so with rust, because of a dogmatic approach to floating point arithmetic that assumes bitwise reproducibility is the "right" answer for everyone (actually, it's the right answer to almost nobody) to the point of not even allowing a user to even flag on these optimizations. and once you get to the point of writing intrinsics you have to handwrite code for every new architecture when autovectorizers could have gotten you 80%-90% of the way there with a single source and often this is just enough.
the contention with the above is that if a user needs SIMD they can just use some SIMD API and make their intention more clear. this is essentially an argument that we should handwrite intrinsics. well guess what. I'm a programmer and I use compilers because they _do this for me_ and indeed are able to do so very easily in C or C++ when I instruct it that I'm ok with with reordering operations and other "accuracy impacting" optimizations.
The huge joke on us is that these optimizations generally have the effect of _improving_ accuracy because it will reduce the number of rounding steps either by simply reducing the number of operations or by using fused multiply adds which round only once.
>80%-90% or so of real life vectorization can be achieved in C or C++ just by writing code in a way that it can be autovectorized.
Yep. I was pleasantly surprised by the autovectorization quality with recent clang at work a few days ago. If you write code that the compiler can infer to be multiples of 4, 8, etc the compiler goes off and emits pretty decent NEON/AVX code. The rest as you say is handled quite well by intrinsics these days.
Autovectorization was definitely poorer 5-10 years ago on older compiler toolchains.
Yikes. Sounds like we need this in rust ASAP. (I do a lot of parallizable code; GPU-centric, but CPU-SIMD is a good fallback for machines that don't have nvidia GPUs). I find the manual SIMD packing/unpacking clumsy, especially when managing this in addition to non-SIMD CPU, and GPU code.
Odd that c# has a better stable SIMD story than Rust! It has both generic vector types across a range of sizes and a good set of intrinsics across most of the common instruction sets
Why would that be odd? C# is an older and mature language backed by a corporation, while Rust is younger and has been run by a small group of volunteers for years now.
> hile Rust is younger and has been run by a small group of volunteers for years now
I thought Rust was getting financial support from Google, Microsoft and Mozilla? Or was the Rust Foundation just a convenient way for Mozilla to fire a large amount of developers and we are actually rapidly approaching the OpenSSL Heartbleed state. Where everyone is happily building on a secure foundation that is maintained by a half dead intern when he isn't busy begging for scraps on the street?
C# portable SIMD is very nice indeed, but it's also not usable without unsafety. On the other hand, Rust compiler (LLVM) has a fairly competent autovectorizer, so you may be able to simply write loops the right way instead of the fancy API.
Unsafety means different things. In C#, SIMD is possible via `ref`s, which maintains GC safety (no GC holes), but removes bounds safety (array length check). The API is called appropriately Vector.LoadUnsafe
Having worked in HPC a fair bit I'm not a fan of autovectorization. I prefer the compiled code's performance to be "unsuprising" based on the source and to use vectors etc where I know it's needed. I think in general it's better to have linting that points out performance issues (e.g. lift this outside the loop) rather than have compilers do it automatically and make things less predictable
You can write good autovectorized code in Rust today, but only for integers. Since Rust lacks --ffast-math, the results on most fp code are disappointing.
You are not "forced" into unsafe APIs with Vector<T>/Vector128/256/512<T>. While it is a nice improvement and helps with achieving completely optimal compiler output, you can use it without unsafe. For example, ZLinq even offers .AsVectorizable LINQ-style API, where you pass lambdas which handle vectors and scalars separately. It the user code cannot go out of bounds and the resulting logic even goes through (inlined later by JIT) delegates, yet still offers a massive speed-up (https://github.com/Cysharp/ZLinq?tab=readme-ov-file#vectoriz...).
Yeah, golang is a particular nightmare for SIMD. You have to write plan 9 assembly, look up what they renamed every instruction to, and then sometimes find that the compiler doesn't actually support that instruction, even though it's part of an ISA they broadly support. Go assembly functions are also not allowed to use the register-based calling convention, so all arguments are passed on the stack, and the compiler will never inline it. So without compiler support I don't believe there's any way to do something like intrinsics even. Fortunately compiler support for intrinsics seems to be on its way! https://github.com/golang/go/issues/73787
To be fair, Java's lack of support seems to have more to do with them needing to fix the whole primitive vs object mess rather than a lack of effort. It sounds like the Vector API will be stabilized shortly after they figure that out, but who knows how long it will take.
If anything, the Mono runtime is what prevents heavier (ab)use of SIMD types in the standard library, or at least it causes additional required effort to make it not regress in performance since it's nowhere near as capable as CoreCLR.
There really aren't that many people working on the compiler. It's mostly volunteers.
The structure is unlike a traditional company. In a traditional company, the managers decide the priorities and direct the employees what to work on while facilitating that work. While there are people with a more managerial type position working on rust compiler, their job is not to tell the volunteers what to work on (they cannot), but instead to help the volunteers accomplish whatever it is they want to do.
I don't know about std::simd specifically, but for many features, it's simply a case of "none of the very small number of people working on the rust compiler have prioritized it".
I do wish there was a bounty system, where people could say "I really want std::simd so I'll pay $5,000 to the rust foundation if it gets stabilized". If enough people did that I'm sure they could find a way to make it happen. But I think realistically, very few people would be willing to put up even a cent for the features they want. I hear a lot of people wishing for better const generics, but only 27 people have set up a donation to boxy (lead of the const generics group https://github.com/sponsors/BoxyUwU ).
- It's a massive hard problem, to build a portable abstraction layer over the SIMD capabilities of various CPUs.
- It's a massive balance between performance and usability, and people care deeply about both.
- It's subject to Rust's stability guarantee for the standard library: once we ship it, we can't fix any API issues.
- There are already portable SIMD libraries in the ecosystem, which aren't subject to that stability guarantee as they can ship new semver-major versions. (One of these days, I hope we have ways to do that for the standard library.)
- Many people already use non-portable SIMD for the 1-3 targets they care about, instead.
> Many people already use non-portable SIMD for the 1-3 targets they care about, instead.
This is something a lot of people (myself included) have gotten tripped up by. Non-portable SIMD intrinsics have been stable under std::arch for a long time. Obviously they aren't nearly as nice to hold, but if you're in a place where you need explicit SIMD speed-ups, that probably isn't a killer.
Despite all of these issues you mention, std::simd is perfectly usable in the state it is in today in nightly Rust.
I've written thousands and thousands of lines of Rust SIMD code over the last ~4 years and it's, in my opinion, a pretty nice way of doing SIMD code that is portable.
I don't know about the specific issues in stabilization, but the API has been relatively stable, although there were some breaking changes a few years ago.
Maybe you can't extract 100% of your CPUs capabilities using it, but I don't find that a problem because there's a zero-cost fallback to CPU-specific intrinsics when necessary.
I recently wrote some computer graphics code and I could get really nice performance (~20x my scalar code, 5x from just a naive translation). And the same codebase can be compiled to AVX2, SSE2 and ARM NEON. It uses f32x8's (256b vector width), which are not available on SSE or NEON, but the compiler can split those vectors. The f32x8 version was faster than f32x4 even on 128b hardware. I would've needed to painstakingly port this codebase to each CPU, so it was at least a 3x reduction in lines of code (and more in programmer time).
There is a GitHub issue that details what's blocking stabilization for a each feature. I've read a few recently and noticed some patterns:
1. A high bar for quality in std
2. Dependencies on other unstable features
3. Known bugs
4. Conflicts with other unstable features
It seems anything that affects trait solving is very complicated and is more likely to have bugs or combine non-trivially with other trait-solving features.
I think there is also some sampling bias. Tons of features get stabilized, but you are much more likely to notice a nightly feature that is unstable for a long time and complex enough to be excited about.
> It seems anything that affects trait solving is very complicated and is more likely to have bugs or combine non-trivially with other trait-solving features.
Yep and this is why many features die or linger on forever. Getting the trait solving working correctly across types and soundly across lifetimes is complicated enough to have killed several features previously (like specialization/min_specialization). It was the reason async trait took so long and why GAT were so important.
Usually when I go and read the github and zulip threads the reason for paused work comes down to the fact that no one has come up with a design that maintains every existing promise the compiler has made. The most common ones I see are the feature conflicts with safety, semver/encapsulation, interacts weirdly with object safety, causes post post-monomorphization errors, breaks perfect type class coherence (see haskells unsound specialization).
Too many promises have been made.
Rust needs more unsafe opt outs. Ironically simd has this so it does not bother me.
std::arch::* intrinsics for SIMD are stable and you can use them today. The situation is only slightly worse than C/C++ because the rust compilers cares a lot about undefined behavior, so there's some safe-but-technically-unsafe/annoying cfg stuff to make sure the intrinsics are actually emitted as you intend.
There is nothing blocking high quality SIMD libraries on stable in Rust today. The bar for inclusion in std is just much higher than the rest of the ecosystem.
Given the “blazingly fast” branding, I too would have thought this would be in stable Rust by now.
However, like other commenters I assume it’s because it’s hard, not all that many users of Rust really need it, and the compiler team is small and only consists of volunteers.
Getting maximum performance out of SIMD requires rolling your own code with intrinsics. It is something a compiler can't do for you at a pretty fundamental level.
Most interesting performance optimizations from vector ISAs can't be done by the compiler.
> Given the “blazingly fast” branding, I too would have thought this would be in stable Rust by now.
It's the exact opposite. It's the portable SIMD abstraction that isn't stable yet. But the vendor specific SIMD intrinsics have been stable for quite some time already (x86-64 for many years for example). And indeed, those are necessary for some cases.
ripgrep wouldn't be as fast as it was if it weren't possible to use SIMD on stable Rust.
I do scientific computing, and even I rarely have a situation where CPU SIMD is a clear win. Usually it's either not worth the added complexity, or the problem is so embarrassingly parallel that you should use a GPU.
Don’t forget that autovectorization does a lot too. This is only for when you want to ensure you get exactly what you want, for many applications, they just kinda get it for free sometimes.
I would love generators too but I think the more features they add the more interactions with existing features they have to deal with, so it's not surprising that its slowing down.
Generators in particular has been blocked on the AsyncIterator trait. There are also open questions around consuming those (`for await i in stream`, or just keep to `while let Some(i) in stream.next().await`? What about parallel iteration? What about pinning obligations? Do that as part of desugaring or making it explicit?). It is a shame because it is almost orthogonal, but any given decision might not be compatible with different approaches for generators. The good news is that some people are working on it again.
Rust doesn’t have a BDFL so there’s nobody with the power to push things through when they’re good enough.
And since Rust basically sells itself on high standards (zero-cost abstractions, etc.) the devs go back and forth until it feels like the solution is handed down from the heavens.
As someone who used std::simd in an attempt for submitting to an academic conference CFP*, I have look deep into how std::simd and I would conclude that there are a couple of reasons it isn't stable yet (this is rather long and maybe need 10 minutes to read):
1. It is highly depending on LLVM intrinsics which itself can change quite a lot. Sometimes the intrinsic would even fail to instantiate and crashed the entire compilation. I for example met chronic ICE crashes for the same code in different nightly Rust version. Then I realize it is because the SIMD operation was too complicated and I need to simplify it, and sometimes need to stop recursing and expanding too much to prevent stack spilling and exhausting register allocation.
This happens from time to time especially when using std::simd with embedded target where registers are scarcity.
2. Some hardware design decisions making SIMD itself not ergonomic and hard to generalize, this is also reflected on the design of std::simd as well.
Recall that SIMD techniques stems from vector processors in supercomputers from the likes of Cray and IBM, that is from the 70s and back then computation and hardware design was primitive and simple, so they have fixed vector size.
The ancient design is very stable, and is still kept till this day, even with the likes of AVX2, AVX512, VFP and NEON, so this influenced the design of things like lane count (https://doc.rust-lang.org/std/simd/struct.LaneCount.html).
But the plot twist: as time goes on, it turns out that modern SIMD is now capable of doing variable sizes; RISC-V's SIMD extension is one such implementation for example.
So now we come to a dilemma on to keep the existing fixed lane count design, or allow it to extend further. If we allow it to extend further to cater for things like variable-SIMD vector length, then we need to wait for generic_const_exprs to be stable, and right now it is not only not stable but incomplete too (https://github.com/rust-lang/portable-simd/issues/416).
This is a hard design philosophical change and is not easy to deal with. Time will tell.
3. As an extension to #2, the way that thinking in SIMD is hard in the very first place, and to use it in production you even have to think about different situations. This come in the form of dynamic dispatch, and it is a pain to dealt with, although we have great helpers such as multiversion...it is still very hard to design an interface that scales. Take Google's highway (https://github.com/google/highway/blob/master/g3doc/quick_re...) for example, it is the library to write portable SIMD code with dynamic dispatch in C++, but in an esotheric and not so ergonomic way. How we could do better with std::simd is still a myth. How do you abstract the idea of scatter-gather operation? What the heck is swizzle? Why do we call it shuffle and not permutation. Lots of stuff to learn, that means lots of pain to go through.
4. Plus, when you think in SIMD, there could be multiple instructions and multiple ways to do the same thing, one maybe more efficient than the other.
For example, as I have to touch some finite field stuff in GF(2^8), there are few ways to do finite field multiplication:
a. Precomputed table lookup
b. Russian Peasant Multiplication (basically carryless Karatsuba multiplication, but oftenly reduce to the form of table lookups as well, can also seen as ripple counter with modulo arithmetic except carry has to be delivered in a different way)
d. Or just treat it as multiplcation over a polynominal power series but this essentially mean we treat it as a finite field convolution, which I suspect is highly related to fourier transform. (https://arxiv.org/pdf/1102.4772)
e. Use the somewhat new GF2P8AFFINEQB (https://www.felixcloutier.com/x86/gf2p8affineqb) from GFNI which, contrary to most people who think it is available for AVX512 only, but is actually available for SSE/AVX/AVX2 as well (this is called GFNI-SSE in gcc), so it works on my 13600KF too (except obviously I cannot use ZMM registers or I just get illegal instruction for any instructions that touches ZMM or uses the EVEX encoding).
I have an internal implementation of finite field multiplication using just that, but I need to use the polynomial of 0x11D rather than 0x11B so GF2P8MULB (https://www.felixcloutier.com/x86/gf2p8mulb) is out of question (which is supposed to be the fastest in the world theoretically if we can use arbitary polynomial), but this is rather hard to understand and explain in the first place. (by the way I used SIMDE for that: https://github.com/simd-everywhere/simde)
All of these can be done in SIMD, but each one of these methods have its pros and cons. Table lookup maybe fast and seemingly O(1) but you actually need to keep the table in cache, meaning we trade time with space, and SIMD would amplify the cache thrashing from multiple access. This could slow down the CPU pipeline although modern CPU are clever enough on cache management. If you want to do Russian Peasant Multiplication then you need a bunch of loops to go through the division and XOR chunk by chunk.
If you want Barrett reduction then you need to have efficient carryless multiplication such as PCLMULQDQ (https://www.felixcloutier.com/x86/pclmulqdq), to do the inner product and reduce the polynomial. Or a more primitive way find ways to do finite field Horner's method in SIMD...
How to think in SIMD is already hard as said in #3. How to balance in SIMD like this is even harder.
Unless you want to have a certain edge, or want to shatter the benchmark, I would say SIMD is not a good investment. You need to use SIMD at the right scenario at the right time. SIMD is useful, but also kind of niche, and modern CPU is optimized well enough that the performance of general solutions without using SIMD, is good enough too, since all of which will eventually dump right down to the uops anyway, with deep pipeline, branch predictor, superscalar and speculative execution doing their magics altogether, and most of the time if you want to use SIMD, using the easiest SIMD methods is generally enough.
*: I myself used std::simd intensively in my own project, well it got refused that the paper was actually severely lacking in literature studies, but that I shouldn't have used LLM too much to generate the paper.
However, the code was here (https://github.com/stevefan1999-personal/sigmah). Now I have a new approach to this problem that is derived from my current work with finite field, error correction, divide and conquer and polynominal multiplication, and I plan to resubmit the paper once I have time to clear it, with a more careful approach next time too, although the problem of string matching with don't care can be seen as convolution and I doubt my approach would ended up something like that...making the paper still unworthy for acceptance.
> performance of general solutions without using SIMD, is good enough too, since all of which will eventually dump right down to the uops anyway, with deep pipeline, branch predictor, superscalar and speculative execution doing their magics altogether
A quick comment on this one point (personal opinion): from a hyperscalar perspective, scalar code is most certainly not enough. The energy cost from scheduling a MUL instruction is something like 10x of the actual operation it performs. It is important to amortize that cost over many elements (i.e. SIMD).
Of interest, I've written my own core::simd mimic so I don't have to make all my libs and programs use nightly. It started as me just making my Quaternion and Vec lib (lin-alg) have their own SoA SIMD variants (Vec3x16 etc), but I ended up implementing and publicly exposing f32x16 etc. Will remove those once core::simd is stable. Downside: These are x86 only; no ARM support.
I also added packing and unpacking helpers that assist with handling final lane 0 values etc. But there is still some subtly, as the article pointed out, compared to using Rayon or non-SIMD CPU code related to packing and unpacking. E.g. you should try to keep things in their SIMD form throughout the whole pipeline, how you pair them with non-SIMD values (Like you might pair [T; 8] with f32x8 etc) etc.
The nightly built-in core::simd makes use of a bunch of intrinsics to "implement" the SIMD ops (or, rather, directly delegate the implementation to LLVM which you otherwise cannot do from plain Rust), which are as much if not more volatile than core::simd itself (and also nightly-only).
simd was one I thought we needed. Then, i started benchmarking using iter with chunks and a nested if statement to check the chunk size. If it was necessary to do more, it was typically time to drop down to asm rather than worry about another layer in between the code and the machine.
I think you misinterpreted GP; he's saying that with some hints (explicit chunking with a branch on the chunk size), the compiler's auto-vectorization can handle the rest, inferring SIMD instructions in a manner that's 'good enough'.
Readit News
> The other drawback of this method is that the optimizer won’t even touch anything involving floats (f32 and f64 types). It’s not permitted to change any observable outputs of the program, and reordering float operations may alter the result due to precision loss. (There is a way to tell the compiler not to worry about precision loss, but it’s currently nightly-only).
Ah - this makes a lot of sense. I've had zero trouble getting excellent performance out of Julia using autovectorization (from LLVM) so I was wondering why this was such a "thing" in Rust. I wonder if that nightly feature is a per-crate setting or what?
The reason it’s a thing is from LLVM and I’m not sure you can “language design” your way out of this problem as it seems intrinsic to IEEE 754.
https://doc.rust-lang.org/std/intrinsics/index.html
Specifically the *_fast intrinsics.
For loops without such dependencies Rust should autovectorize just fine as with any other element type.
This obviously has a lot of caveats, and should only be enabled on a per function or per file basis.
Unfortunately Rust does not currently have options for adjusting per-function compiler optimization parameters. This is possible in some C compilers using function attributes.
I wonder if it could autovec the simd-ordered code.
Unfortunately it's a set of functions you have to use to perform arithmetic ops if you want the autovectorizer to touch them
1. dangerous assumptions hidden behind a simple, attractive-looking option [1]. It should be called -fwrong-math or -fdangerous-math or something (GCC does have the funnily named switch -funsafe-math-optimizations – what could go wrong with fun, safe math optimizations?!)
2. Translation-unit scoped, which means that dependencies not consented to "fast math" can break your code (as in UB land) or make the optimizations pointless, and your code can break your dependencies' semantics too via inlining. On the other hand, a library author must think very carefully what float opts to enable in order to be compatible with client code.
Deciding how the scoping of non-IEEE float math operations should work is a very nontrivial question. The scope could be a translation unit, a module, a type, a function, a block, or every individual operation, and none of those is without issues, particularly regarding questions like inlining and interprocedural and link-time-optimization, as well as ergonomics. In other ways, it's yet another function coloring problem.
There are currently-unstable "algebraic_add/mul/etc" methods for floats for letting LLVM treat those particular operations as if floats were real numbers [2]. They're the first step towards safe UB-free float optimizations, but of course those names are rather awkward to use in math-heavy code, and a wrapper type overloading the normal operators would be good to have.
---
[1] See, eg. https://simonbyrne.github.io/notes/fastmath/
[2] In terms of associativity and such, not in eg. assuming the nonexistence of NaNs, which would be very unsafe.
The major problem is that Rust essentially requires you to annotate every (!) function in your whole call stack with e.g. `#[target_feature(enable = "avx2")]` to make sure that the SIMD intrinsics will actually get inlined (if they're not inlined then the performance is horrible, which makes using SIMD completely pointless). This makes it very hard to build any reasonable abstractions because you need to hardcode this all over your code. You can't have e.g. a `DataStructure<S>` where S is the SIMD ISA, so that you could do `DataStructure<AVX2>` or `DataStructure<SSE>` to get a nicely specialized version of it for a given instruction set. You need to copy-paste the whole thing with changed `target_feature` attributes (or use a procedural macro which does the copy-pasting) and have two entirely separate `DataStructureAVX2` and `DataStructureSSE` types.
That is, the `memchr` crate has a `Vector` trait that is generic over the vector type. Which is essentially your `DataStructure<S>` where `S` is the ISA. (Using the vector type isn't load bearing. I could do the ISA. But in `memchr`'s case, the vector type implies the ISA.)
It relies on `#[inline(always)]` to work. But you can write the algorithm generically once: https://github.com/BurntSushi/memchr/blob/1230fc5c638a4d922f...
And the entry point into a specific instantiation of the generic algorithm is where you apply `#[target_feature(enable = "foo")]`: https://github.com/BurntSushi/memchr/blob/1230fc5c638a4d922f...
This is a workaround, but it's still very painful, and there are still many problems with it:
1) you need to design the whole thing with this in mind; you can't just take an existing something which takes some T and get that to use SIMD,
2) `#[inline(always)]` can't be used with `#[target_feature]` at the same time,
3) you need to have `#[inline(always)]` on your whole call stack (I often don't want to inline everything, but I still want to propagate that we're in an SIMD-safe context, which you normally do with `#[target_feature]`, but that's incompatible with being able to abstract over things),
4) you have to sprinkle `unsafe` everywhere (which is especially annoying as recently we gained the ability to finally be able to call a lot of the intrinsics without any `unsafe`),
5) you have to manually make sure you're not calling something you should not be calling (`#[target_feature]` is great here because it prevents you from accidentally calling e.g. AVX2 intrinsics if you're not on the AVX2 codepath),
6) any mistake (e.g. missing an annotation somewhere) you make will just have the compiler silently not inline the intrinsics, completely killing your performance, and you're completely on your own debugging it.
I work with SIMD a lot in Rust, and unfortunately currently it's a miserable experience. It's not that bad when you're writing e.g. an algorithm which fits in a single function, but once you want to write something bigger with more sophisticated abstractions and/or you want to integrate SIMD into existing code then it becomes a world of pain.
Basically I just don't want to hear about "the state of SIMD in Rust" unless it is about dramatic improvement in autovectorization in the rust compiler.
80%-90% or so of real life vectorization can be achieved in C or C++ just by writing code in a way that it can be autovectorized. Intrinsics get you the rest of the way on harder code. Autovectorization is essentially a solved problem for the vast majority of floating point code.
Not so with rust, because of a dogmatic approach to floating point arithmetic that assumes bitwise reproducibility is the "right" answer for everyone (actually, it's the right answer to almost nobody) to the point of not even allowing a user to even flag on these optimizations. and once you get to the point of writing intrinsics you have to handwrite code for every new architecture when autovectorizers could have gotten you 80%-90% of the way there with a single source and often this is just enough.
the contention with the above is that if a user needs SIMD they can just use some SIMD API and make their intention more clear. this is essentially an argument that we should handwrite intrinsics. well guess what. I'm a programmer and I use compilers because they _do this for me_ and indeed are able to do so very easily in C or C++ when I instruct it that I'm ok with with reordering operations and other "accuracy impacting" optimizations.
The huge joke on us is that these optimizations generally have the effect of _improving_ accuracy because it will reduce the number of rounding steps either by simply reducing the number of operations or by using fused multiply adds which round only once.
Yep. I was pleasantly surprised by the autovectorization quality with recent clang at work a few days ago. If you write code that the compiler can infer to be multiples of 4, 8, etc the compiler goes off and emits pretty decent NEON/AVX code. The rest as you say is handled quite well by intrinsics these days.
Autovectorization was definitely poorer 5-10 years ago on older compiler toolchains.
I thought Rust was getting financial support from Google, Microsoft and Mozilla? Or was the Rust Foundation just a convenient way for Mozilla to fire a large amount of developers and we are actually rapidly approaching the OpenSSL Heartbleed state. Where everyone is happily building on a secure foundation that is maintained by a half dead intern when he isn't busy begging for scraps on the street?
Another example, note how these implementations, one in unsafe C# and another in safe F# have almost identical performance: https://benchmarksgame-team.pages.debian.net/benchmarksgame/..., https://benchmarksgame-team.pages.debian.net/benchmarksgame/...
Eg https://tirania.org/blog/archive/2008/Nov-03.html
You can find the overview of relevant high-level types here: https://learn.microsoft.com/en-us/dotnet/standard/simd
For low-level types and historical context: https://devblogs.microsoft.com/dotnet/dotnet-8-hardware-intr...
I’m sure more people than ever are working on the compiler. What’s going on?
The structure is unlike a traditional company. In a traditional company, the managers decide the priorities and direct the employees what to work on while facilitating that work. While there are people with a more managerial type position working on rust compiler, their job is not to tell the volunteers what to work on (they cannot), but instead to help the volunteers accomplish whatever it is they want to do.
I don't know about std::simd specifically, but for many features, it's simply a case of "none of the very small number of people working on the rust compiler have prioritized it".
I do wish there was a bounty system, where people could say "I really want std::simd so I'll pay $5,000 to the rust foundation if it gets stabilized". If enough people did that I'm sure they could find a way to make it happen. But I think realistically, very few people would be willing to put up even a cent for the features they want. I hear a lot of people wishing for better const generics, but only 27 people have set up a donation to boxy (lead of the const generics group https://github.com/sponsors/BoxyUwU ).
Seems smart to put the language as a requirement for compiling the linux kernel and a bunch of other core projects then!
Leaving aside any specific blockers:
- It's a massive hard problem, to build a portable abstraction layer over the SIMD capabilities of various CPUs.
- It's a massive balance between performance and usability, and people care deeply about both.
- It's subject to Rust's stability guarantee for the standard library: once we ship it, we can't fix any API issues.
- There are already portable SIMD libraries in the ecosystem, which aren't subject to that stability guarantee as they can ship new semver-major versions. (One of these days, I hope we have ways to do that for the standard library.)
- Many people already use non-portable SIMD for the 1-3 targets they care about, instead.
This is something a lot of people (myself included) have gotten tripped up by. Non-portable SIMD intrinsics have been stable under std::arch for a long time. Obviously they aren't nearly as nice to hold, but if you're in a place where you need explicit SIMD speed-ups, that probably isn't a killer.
I've written thousands and thousands of lines of Rust SIMD code over the last ~4 years and it's, in my opinion, a pretty nice way of doing SIMD code that is portable.
I don't know about the specific issues in stabilization, but the API has been relatively stable, although there were some breaking changes a few years ago.
Maybe you can't extract 100% of your CPUs capabilities using it, but I don't find that a problem because there's a zero-cost fallback to CPU-specific intrinsics when necessary.
I recently wrote some computer graphics code and I could get really nice performance (~20x my scalar code, 5x from just a naive translation). And the same codebase can be compiled to AVX2, SSE2 and ARM NEON. It uses f32x8's (256b vector width), which are not available on SSE or NEON, but the compiler can split those vectors. The f32x8 version was faster than f32x4 even on 128b hardware. I would've needed to painstakingly port this codebase to each CPU, so it was at least a 3x reduction in lines of code (and more in programmer time).
Can’t APIs be fixed between editions?
1. A high bar for quality in std
2. Dependencies on other unstable features
3. Known bugs
4. Conflicts with other unstable features
It seems anything that affects trait solving is very complicated and is more likely to have bugs or combine non-trivially with other trait-solving features.
I think there is also some sampling bias. Tons of features get stabilized, but you are much more likely to notice a nightly feature that is unstable for a long time and complex enough to be excited about.
Yep and this is why many features die or linger on forever. Getting the trait solving working correctly across types and soundly across lifetimes is complicated enough to have killed several features previously (like specialization/min_specialization). It was the reason async trait took so long and why GAT were so important.
AFAIK that’s not a blocker for Rust - the std library is allowed to use unstable at all times.
Too many promises have been made.
Rust needs more unsafe opt outs. Ironically simd has this so it does not bother me.
There is nothing blocking high quality SIMD libraries on stable in Rust today. The bar for inclusion in std is just much higher than the rest of the ecosystem.
However, like other commenters I assume it’s because it’s hard, not all that many users of Rust really need it, and the compiler team is small and only consists of volunteers.
Most interesting performance optimizations from vector ISAs can't be done by the compiler.
It's the exact opposite. It's the portable SIMD abstraction that isn't stable yet. But the vendor specific SIMD intrinsics have been stable for quite some time already (x86-64 for many years for example). And indeed, those are necessary for some cases.
ripgrep wouldn't be as fast as it was if it weren't possible to use SIMD on stable Rust.
And since Rust basically sells itself on high standards (zero-cost abstractions, etc.) the devs go back and forth until it feels like the solution is handed down from the heavens.
1. It is highly depending on LLVM intrinsics which itself can change quite a lot. Sometimes the intrinsic would even fail to instantiate and crashed the entire compilation. I for example met chronic ICE crashes for the same code in different nightly Rust version. Then I realize it is because the SIMD operation was too complicated and I need to simplify it, and sometimes need to stop recursing and expanding too much to prevent stack spilling and exhausting register allocation.
This happens from time to time especially when using std::simd with embedded target where registers are scarcity.
2. Some hardware design decisions making SIMD itself not ergonomic and hard to generalize, this is also reflected on the design of std::simd as well.
Recall that SIMD techniques stems from vector processors in supercomputers from the likes of Cray and IBM, that is from the 70s and back then computation and hardware design was primitive and simple, so they have fixed vector size.
The ancient design is very stable, and is still kept till this day, even with the likes of AVX2, AVX512, VFP and NEON, so this influenced the design of things like lane count (https://doc.rust-lang.org/std/simd/struct.LaneCount.html).
But the plot twist: as time goes on, it turns out that modern SIMD is now capable of doing variable sizes; RISC-V's SIMD extension is one such implementation for example.
So now we come to a dilemma on to keep the existing fixed lane count design, or allow it to extend further. If we allow it to extend further to cater for things like variable-SIMD vector length, then we need to wait for generic_const_exprs to be stable, and right now it is not only not stable but incomplete too (https://github.com/rust-lang/portable-simd/issues/416).
This is a hard design philosophical change and is not easy to deal with. Time will tell.
3. As an extension to #2, the way that thinking in SIMD is hard in the very first place, and to use it in production you even have to think about different situations. This come in the form of dynamic dispatch, and it is a pain to dealt with, although we have great helpers such as multiversion...it is still very hard to design an interface that scales. Take Google's highway (https://github.com/google/highway/blob/master/g3doc/quick_re...) for example, it is the library to write portable SIMD code with dynamic dispatch in C++, but in an esotheric and not so ergonomic way. How we could do better with std::simd is still a myth. How do you abstract the idea of scatter-gather operation? What the heck is swizzle? Why do we call it shuffle and not permutation. Lots of stuff to learn, that means lots of pain to go through.
4. Plus, when you think in SIMD, there could be multiple instructions and multiple ways to do the same thing, one maybe more efficient than the other.
For example, as I have to touch some finite field stuff in GF(2^8), there are few ways to do finite field multiplication:
a. Precomputed table lookup
b. Russian Peasant Multiplication (basically carryless Karatsuba multiplication, but oftenly reduce to the form of table lookups as well, can also seen as ripple counter with modulo arithmetic except carry has to be delivered in a different way)
c. Do an inner product and then do Barrett reduction (https://www.esat.kuleuven.be/cosic/publications/article-1115...)
d. Or just treat it as multiplcation over a polynominal power series but this essentially mean we treat it as a finite field convolution, which I suspect is highly related to fourier transform. (https://arxiv.org/pdf/1102.4772)
e. Use the somewhat new GF2P8AFFINEQB (https://www.felixcloutier.com/x86/gf2p8affineqb) from GFNI which, contrary to most people who think it is available for AVX512 only, but is actually available for SSE/AVX/AVX2 as well (this is called GFNI-SSE in gcc), so it works on my 13600KF too (except obviously I cannot use ZMM registers or I just get illegal instruction for any instructions that touches ZMM or uses the EVEX encoding). I have an internal implementation of finite field multiplication using just that, but I need to use the polynomial of 0x11D rather than 0x11B so GF2P8MULB (https://www.felixcloutier.com/x86/gf2p8mulb) is out of question (which is supposed to be the fastest in the world theoretically if we can use arbitary polynomial), but this is rather hard to understand and explain in the first place. (by the way I used SIMDE for that: https://github.com/simd-everywhere/simde)
All of these can be done in SIMD, but each one of these methods have its pros and cons. Table lookup maybe fast and seemingly O(1) but you actually need to keep the table in cache, meaning we trade time with space, and SIMD would amplify the cache thrashing from multiple access. This could slow down the CPU pipeline although modern CPU are clever enough on cache management. If you want to do Russian Peasant Multiplication then you need a bunch of loops to go through the division and XOR chunk by chunk.
If you want Barrett reduction then you need to have efficient carryless multiplication such as PCLMULQDQ (https://www.felixcloutier.com/x86/pclmulqdq), to do the inner product and reduce the polynomial. Or a more primitive way find ways to do finite field Horner's method in SIMD...
How to think in SIMD is already hard as said in #3. How to balance in SIMD like this is even harder.
Unless you want to have a certain edge, or want to shatter the benchmark, I would say SIMD is not a good investment. You need to use SIMD at the right scenario at the right time. SIMD is useful, but also kind of niche, and modern CPU is optimized well enough that the performance of general solutions without using SIMD, is good enough too, since all of which will eventually dump right down to the uops anyway, with deep pipeline, branch predictor, superscalar and speculative execution doing their magics altogether, and most of the time if you want to use SIMD, using the easiest SIMD methods is generally enough.
*: I myself used std::simd intensively in my own project, well it got refused that the paper was actually severely lacking in literature studies, but that I shouldn't have used LLM too much to generate the paper.
However, the code was here (https://github.com/stevefan1999-personal/sigmah). Now I have a new approach to this problem that is derived from my current work with finite field, error correction, divide and conquer and polynominal multiplication, and I plan to resubmit the paper once I have time to clear it, with a more careful approach next time too, although the problem of string matching with don't care can be seen as convolution and I doubt my approach would ended up something like that...making the paper still unworthy for acceptance.
A quick comment on this one point (personal opinion): from a hyperscalar perspective, scalar code is most certainly not enough. The energy cost from scheduling a MUL instruction is something like 10x of the actual operation it performs. It is important to amortize that cost over many elements (i.e. SIMD).
I also added packing and unpacking helpers that assist with handling final lane 0 values etc. But there is still some subtly, as the article pointed out, compared to using Rayon or non-SIMD CPU code related to packing and unpacking. E.g. you should try to keep things in their SIMD form throughout the whole pipeline, how you pair them with non-SIMD values (Like you might pair [T; 8] with f32x8 etc) etc.
Can't you just make a local copy of the existing package and use that? Did you need to re-implement?
Deleted Comment
Zig has @Vector. This is a builtin, so it gets resolved at comptime. Is the problem with Rust here too much abstraction?
This matches the conclusion we reached for Chromium. We were okay with nightly, so we're using `std::simd` but trying to avoid the least stable APIs. More details: https://docs.google.com/document/d/1lh9x43gtqXFh5bP1LeYevWj0...