I'm surprised there's no mention of store forwarding in that discussion. The -O3 codegen is bonkers, but the -O2 output is reasonable. In the case where one of the structs has just been computed, attempting to load it as a single 32-bit load can result in a store forwarding failure that would negate the benefit of merging the loads. In a non-inlined, non-PGO scenario the compiler doesn't have enough information to tell whether the optimization is suitable.
> In the case where one of the structs has just been computed, attempting to load it as a single 32-bit load can result in a store forwarding failure that would negate the benefit of merging the loads
Would that failure be significantly worse than separate loading?
Just negating the optimization wouldn't be much reason against doing it. A single load is simpler and in the general case faster.
The thing I like most about this is that the discussion isn't just 14 pages of "I'm having this issue as well" and "Any updates on when this will be fixed?" As a web dev, GitHub issues kinda suck.
Wonder if it's a poor interface issue... if people could just click a button that says "me too" but didn't add a full comment but rather just added some minimal notation at the bottom of the comment that indicated their username, 1) would people use it and 2) would that be not overly-busy enough to not be annoying? It could even mute notifications for the me-toos.
This just seems to illustrate the complexity of compiler authorship. I am very sure c compilers are wble to address this issue any better in the general case.
Keep in mind Rust is using the same backend as one of the main C compilers, LLVM. So if it is handling it any better that means the Clang developers handle it before it even reaches the shared LLVM backend. Well, or there is something about the way Clang structures the code that catches a pattern in the backend the Rust developers do not know about.
I just tried it, and the problem is even worse in gcc.
Given this C code:
typedef struct { uint16_t a, b; } pair;
int eq_copy(pair a, pair b) {
return a.a == b.a && a.b == b.b;
}
int eq_ref(pair *a, pair *b) {
return a->a == b->a && a->b == b->b;
}
Clang generates clean code for the eq_copy variant, but complex code for the eq_ref variant. Gcc emits pretty complex code in both variants.
For example, here's eq_ref from gcc -O2:
eq_ref:
movzx edx, WORD PTR [rsi]
xor eax, eax
cmp WORD PTR [rdi], dx
je .L9
ret
.L9:
movzx eax, WORD PTR [rsi+2]
cmp WORD PTR [rdi+2], ax
sete al
movzx eax, al
ret
I think that since the 1.5% one is only for aarch64 it's a bit unfair to claim the full number, more like 1/2 if you consider arm/x86 to be the majority of the (future) deployments
I suppose that’s fair, but I’d give credit for a 2.3% improvement in the test environment. For all we know it may be a net loss in other environments due to quirks (probably not, admittedly).
Thanks! Would be interesting to see if Rust/LLVM folks can get the compiler to apply this optimization whenever possible, as Rust can be much more accurate w.r.t memory initialization.
I think rust may be able to get it by adding a `freeze` intrinsic to the codegen here. that would force LLVM to pick a deterministic value if there was poison, and should thus unblock the optimization (which is fine here because we know the value isn't poison)
I think in this case Rust and C code aren't equivalent which maybe caused this slow down. Union trick also affects the alignment. C side struct is 32 bit aligned, but Rust struct only has 16bit alignment because it only contains fields with 16bit alignment. In practice the fields are likely anyway correctly aligned to 32bits, but compiler optimizations may have hard time verifying that.
Have you tried manually defining alignment of Rust struct?
It varies. New public APIs or language features may take a long time, but changes to internals and missed optimizations can be fixed in days or weeks, in both LLVM and Rust.
All being equal codecs ought to be in WUFFS† rather than Rust, but I can well imagine that it's a much bigger lift to take something as complicated as dav1d and write the analogous WUFFS than to clean up the c2rust translation, if you said a thousand times harder I'd have no trouble believing that. I just think it's worth it for us as a civilisation.
† Or an equivalent special purpose language, but WUFFS is right there
WUFFS would be great for parsing container files (Matroska, webm, mp4) but it does not seem at all suitable for a video decoder. Without dynamic memory allocation it would be challenging to deal with dynamic data. Video codecs are not simply parsing a file to get the data, they require quite a bit of very dynamic state to be managed.
Requiring dynamic state seems not obvious to me. At the end of the day you have a fixed number of pixels on the screen. If every single pixel changes from frame to frame that should constitute the most work your codec has to do, no? I'm not a codec writer but that's my intuition based on the assumption that codecs are basically designed to minimize the amount of 'work' being done from frame to frame.
If you are doing something like a GIF or an MJPEG, sure. If you are doing forwards and backwards keyframes with a variable amount of deltas in between, with motion estimation, with grain generation, you start having a very dynamic amount of state. Granted, encoders are more complex than decoders in some of this. But still you might need to decode between 1 and N frames to get the frame you want, and you don't know how much memory it will consume once it is decoded unless you decode it into bitmaps (at 4k that would be over 8MB per frame which very quickly runs out of memory for you if you want any sort of frame buffer present).
I suspect the future of video compression will also include frame generation, like what is currently being done for video games. Essentially you have let's say 12 fps video but your video card can fill in the intermediate frames via what is basically generative AI so you get 120 fps output with smooth motion. I imagine that will never be something that WUFFS is best suited for.
> But still you might need to decode between 1 and N frames to get the frame you want, and you don't know how much memory it will consume...
All of these things are bounded for actual codecs. AV1 allows storing at most 8 reference frames. The sequence header will specify a maximum allowable resolution for any frame. The number of motion vectors is fixed once you know the resolution. Film grain requires only a single additional buffer. There are "levels" specified which ensure interoperability at common operating points (e.g., 4k) without even relying on the sequence header (you just reject sequences that fall outside the limits). Those are mostly intended for hardware, but there is no reason a software decoder could not take advantage of them. As long as codecs are designed to be implemented in hardware, this will be possible.
> I suspect the future of video compression will also include frame generation
That's how most video codecs work already. They try to "guess" what the next frame will be, based on past (for P-frames) and future (for B-frames) frames. The difference is that the codec encodes some metadata to help with the process and also the difference between the predicted frame and the real frame.
As for using AI techniques to improve prediction, it is not a new thing at all. Many algorithms optimized for compression ratio use neural nets, but these tend to be too computationally expensive for general use. In fact the Hutter prize considers text compression as an AI/AGI problem.
See this is interesting to me. I understand the desire to dynamically allocate buffers at runtime to capture variable size deltas. That's cool, but also still maybe technically unnecessary? Because like you say, at 4k and over 8MB per frame; you still can't allocate over a limit. So likely a codec would have some boundary set on that anyway. Why not just pre-allocate at compile time? For sure this results in a complex data structure. Functionally it could be the same and we would elide the cost of dynamic memory allocations. What I'm suggesting is probably complex, I'm sure.
In any case I get what you're saying and I understand why codecs are going to be dynamically allocating memory, so thanks for that.
> codecs are basically designed to minimize the amount of 'work' being done from frame to frame
But to do that they have to keep state and do computations on that state. If you've got frame 47 being a P frame, that means you need frame 46 to decode it correctly. Or frame 47 might be a B frame in which case you need frame 46 and possibly also frame 48 - which means you're having to unpack frames "ahead" of yourself and then keep them around for the next decode.
Memory usage can vary, but video codecs are designed to make it practical to derive bounds on those memory requirements because hardware implementations don't have the freedom to dynamically allocate more silicon.
Maybe you're not familiar with how long GOP encoding works with IPB frames? If all frames were I-frames, maybe what you're thinking might work. Everything you need is in the one frame to be able to describe every single pixel in that frame. Once you start using P-frames, you have to hold on to data from the I-frame to decode the P-frame. With B-frames, you might need data from frames not yet decoded as the are bi-direction references.
Still you don't necessarily need to have dynamic memory allocations if the number of deltas you have is bounded. In some codecs I could definitely see those having a varying size depending on the amount of change going on in the scene.
I'm not a codec developer, I'm only coming at this from an outside/intuitive perspective. Generally, performance concerned parties want to minimize heap allocations, so I'm interested in this as how it applies in codec architecture. Codecs seem so complex to me, with so much inscrutable shit going on, but then heap allocations aren't optimized out? Seems like there has to be a very good reason for this.
You're actually right about allocation - most video codecs are written with hardware decoders in mind which have fixed memory size. This is why their profiles hard limit the memory constraints needed for decode - resolution, number of reference frames, etc.
That's not quite the case for encoding - that's where things get murky since you have way more freedom at what you can do to compress better.
The very good reason is that there's simply not a lot of heap allocations going on. It's easy to check; run perf against e.g. ffmpeg decoding a big file to /dev/null, and observe the distinct lack of malloc high up in the profile.
There's a heck of a lot of distance from “not a lot” to “zero”, though.
Hey maybe we can discuss why I'm being downvoted? This is a technical discussion and I'm contributing. If you disagree then say why. I'm not stating anything as fact that isn't fact. I am getting downvoted for asking a question.
You know it's a good post when it starts with a funny meme. Seems related to the recent discussion: $20K Bounty Offered for Optimizing Rust Code in Rav1d AV1 Decoder (memorysafety.org) | 108 comments | https://news.ycombinator.com/item?id=43982238
Honestly its a little surprising the first optimization he found was something fairly obvious just by using perf. I thought they had discussed the zeroing buffers issue in the first post? The second optimization was definitely more involved/interesting but was still pointed at by perf. Don't underestimate that tool!
AFAICS, it wasn't “just perf”; it was doing a differential profile between the C and Rust versions, with manual matching up. (perf diff exists, but can't match across the differing symbol names, and few people seem to use it.)
He came from the aarch64 perspective on an Apple device. I often experience someone spotting an "obvious in hindsight" gap because they come from a different background.
I generally trust rbultje to benchmark correctly but the ravid tracking ticket has multithread numbers across multiple platforms that don't show that big a difference.
Reading the ffmpeg twitter account is enough to turn me off using ffmpeg. It's a shame there's no real alternative -- the devs seem very toxic.
I mean sure, max performance is great if you control every part of your pipeline, but if you're accepting untrusted data from users-at-large ffmpeg has at least a half-dozen remotely exploitable CVEs a year. Better make sure your sandbox is tight.
I've worked with ffmpeg for literally a decade and I've never found them particularly toxic.
What I have found that they (as many others who do great work) have very little tolerance of random junior language fanboys criticizing their decades of work without even understanding what they're talking about and constantly throwing out silly rewrite ideas.
The SQlite folks, half of Linux, and other maintainers have encountered the same kind of zealotry. Dealing with language supremacism is annoying and I don’t blame ffmpeg for venting.
In fact, I’d even say that twitter thread is informative, because it demonstrates out how big tech fund their own pet projects over the actual maintainers.
I'm not saying that they don't do great work, but that twitter thread (https://x.com/ffmpeg/status/1924137645988356437) is pretty obnoxious and reads like they are upset they didn't get funding. It's entirely possible that they are just difficult to work with and funders _don't_ want to fund them.
"Because substantial amounts of human and financial resources go into these rust ports that are inferior to the originals. Orders of magnitude more resources than the originals which remain extremely understaffed/underfunded." -- https://x.com/FFmpeg/status/1924149949949775980
It's kind of sad to see that snarky attitude. Clearly the corporate sponsors _want_ a more secure decoder. Maybe they should try and work _with_ the system instead of wasting energy on sarcasm on Twitter?
There is not much, unless you're working with AV1. rav1d is the alternative there but you've got to trade off some performance for security gains.
ffmpeg is a monopoly in the space which means that you either take the exact set of tradeoffs they offer, or... well, you have no alternatives, so take it.
Of course the alternatives are never going to be as good as the originals until they've had more effort put into them. It took _years_ until the Rust gzip/zip libraries surpassed the C ones while being more secure overall.
How many of those "remotely exploitable CVEs" have actually been exploited in the wild? Quite a few are denial-of-service and memory leak CVEs too, which Rust doesn't consider to be unsafe.
The healthier response might have been work to speed-up dav1d. If you refine the Olympic Record metrics and force them to retrospectively update previous records so that Bolt's 100m sprint record is revised to 9.64s rather than 9.63s nobody cares man, get a life, but if you can run an actual nine second 100 metre sprint that people care about†
† If you're a human. If you're an ostrich this is not impressive, but on the whole ostrichs aren't competing in the Olympic 100 metre sprint.
yeah. Proving that the zero initialization is useless requires proving that the rest of the program never reads one of the zeroed values. This is really difficult because compilers generally don't track individual array indices (since you often don't even know how big the array is)
There's something about real optimization stories that I find fascinating – particularly the detailed ones including step-by-step improvements and profiling to show how numbers got better. In some way, they are satisfying to read.
Nicholas Nethercote's "How to speed up the Rust compiler" writings[1] fall into this same category for me.
Since you seem to enjoy this kind of writing I'd love to get your feedback on something I've written a while back about branchless partitioning [1]. Despite it being content wise the most work to create of the things I've written about the topic, it found much less attention than other things I've written. So far I've wondered if it was maybe too technical? Would love to get an honest opinion.
I read an article a while ago where the goal is to process a file as fast as possible and the article talks about compressing the data chunks so they fit in L1 cache. The cache misses were slower than compressing and decompressing the data from L1 cache.
I've been trying to find that article ever since but I'm not able to. Anyone knows the article I'm talking about?
AV1 is an amazing codec. I really hope it replaces proprietary codecs like h264 and h265. It has a similar, if not better, performance to h265 while being completely free. Currently on an Intel-based Macbook it is only supported in some browsers, however it seems that newer video cards from AMD, Nvidia, and Intel do include hardware decoders.
RDNA3 cards also have AV1 encode. RDNA 2 only has decode.
With the bitrate set to 100MB/s it happily encodes 2160p or even 3240p, the maximum resolution available when using Virtual Super Resolution (which renders at >native res and downsamples, is awesome for titles without resolution scaling when you don't want to use TAA)
I'm not really well versed with codecs, but is it up to the devices or the providers (where you're uploading them) to handle playback or both? A couple of days ago, I tried to upload an Instagram Reel in AV1 codec, and I was struggling to preview it on my Samsung S20 FE Snapdragon version (before uploading and during preview as well). I then resorted to H.264 and it worked w/o any issues.
Playback is 100% handled by the device. The primary (and essentially only) benefit of H.264 is that almost every device in the entire world has an H.264 hardware decoder builtin to the chip, even extremely cheap devices.
AV1 hardware decoders are still rare so your device was probably resorting to software decoding, which is not ideal.
Instagram (the provider) will transcode for compatibility but likely the preview is before transcoding, the assumption being that the device that uploads the video is able to play it.
I don't know instagram, but I would expect any provider to be handle almost any container/codec/resolution combination going (they likely use ffmpeg underneath) and generate their different output formats at different bitrates for different playback devices.
Either instagram won't accept av1 (seems unlikely) or they just haven't processed it yet as you infer.
Yeah, but I think it has much higher CPU usage, at least when there is no native hardware decoder/encoder. Maybe this has more to do with adoption, since H264 has been an industry standard.
Codec selection is always a complex task. You've got to weigh quality/bitrate vs availability of hardware encode/decode, licensing, and overall resource usage.
The ITU standards have had a lot better record of inclusion in devices that people actually have; and often using hardware encode/decode takes care of licensing. But hardware encode doesn't always have the same quality/bitrate as software and may not be able to do fancier things like simulcast or svc. Some of the hardware decoders are pretty picky about what kinds of streams they'll accept too.
IMHO, if you're looking at software h.264 vs software vp9, I think vp9 is likely to give you better quality at a given bitrate, but will take more cpu to do it. So, as always, it depends.
> IMHO, if you're looking at software h.264 vs software vp9, I think vp9 is likely to give you better quality at a given bitrate, but will take more cpu to do it. So, as always, it depends.
That's a pretty messy way to measure. h.264 with more CPU can also beat h.264 with less CPU.
How does the quality compare if you hold both bitrate and CPU constant?
How does the CPU compare if you hold both bitrate and quality constant?
AV1 will do significantly better than h.264 on both of those tests. How does VP9 do?
AV1 is the outright winner in terms of compression efficiency (until you start comparing against VVC/H.266¹), with the advantage being even starker at high resolutions. The only current notable downside of AV1 is that client hardware support isn't yet universal.
Things like getting hardware to The Scene for encoding might help, but I'm not sure of the bottleneck, it might be bureaucratic or educational or cultural.
At higher quality/bitrates, the difference is much smaller and device support is universal for AVC and quite good for HEVC. Anything over 1.5GB for a single episode would probably only be farily similarly sized with AV1.
There is one large exception, but I don't know the current scene well enough to know if it matters: sources that are grainy. I have some DVD and blurays with high grain content and AV1 can work wonders with those thanks to the in-loop grain filter and synthesis -- we are talking half the size for a high-quality encode. If I were to encode them for AVC at any reasonable bitrate, I would probably run a grain-removal filter which is very finicky if you don't want to end up with something that is overly blurry.
This may be in part because people that automatized their media servers are using hardware acceleration for transcoding (from 4k for example), and hardware has only recently added decoding for AV1.
In my case, I get both 4k (h265) and 1080p (h264) blurays and let the client select.
transpiled code is rarely good. Rust is often better than C for SIMD code (it actually has useful SIMD instructions exposed, and aliasing guarantees make it a lot easier for the compiler to figure out obvious optimization. By transpiling, however you loose most of the structure of an idiomatic project and generally make a bit of a mess of things.
