Agner Fog | Hacker News Reader

Agner Fog

agner.org

165 pointsibobev35 comments

The link is presented without commentary, but for those who do not know, Agner Fog manuals are pretty much the bible on x86 microarchitectural details and optimization.

Why does the page look like a joke?

Why do you ask?

A sort of a filter, getting rid of the viewers who are only interested in the currently popular visual gimmicks instead of the actual content.

The animation led to me leaving the site without taking the time to look at the actual content.

The animation says "I hate Animations!".

I guess the filter worked then

If the filter is to keep out readers annoyed by animations, then yes, I suppose it worked perfectly!

Perhaps you'd be more comfortable if it loaded a few megs of JavaScript.

I guess working with x86 at the microarchitecture level requires a certain amount of humour.

If you were to ignore the man in the plane, all the content is presented very clearly and directly.

It really isn't. There's the weird link thing and the random space at the right you can scroll of the page.

To filter out people who would judge a book by its cover.

Also, this is an ancient web site, built back before everyone had to have glitzy junk. So it has a little whimsy. Did you notice how fast it loaded?

I noticed that it used <a> for fragment anchors instead of the id attribute. I changed over most of those on my WWW site some years ago.

Well, I just learned that the id attribute is what one should use for fragment anchors. All the HTML source code I've read (probably within the last year or so) used <a>. :/

Perhaps the moral here is that "newer" code was so obfuscated (minified; framework-infested) that I wasn't able to read it, so I only ever ended up looking at very dated HTML to fish examples out of. (Figuring out how to phrase something to Google can sometimes be slower and/or harder than simply looking in existing source code where you know you'll find something...)

FWIW, I started tinkering with HTML circa 2003 and have slowly (not especially concertedly) studied it over the years, which is my excuse^Hdefense. HTML5/CSS3 aren't especially intimidating to me (thanks to browser compatibility :D) and I could probably photocopy 60-80% of the designs I see here given significant time (this likely describes a lot of people here), but I don't really actively develop for the Web, so little oversights like this happen. Goes and reads how to include a CSS file again

I'm a bit ashamed to admit that you just taught me that anchor links work with id attributes. I just always used <a name> tags... which I rarely used, but still. I guess my excuse is that I started with HTML around 1996 :)

I guess it's a good thing that I don't do much frontend coding anymore!

Important clarification: it was JdeBP's comment that taught _me_ that one uses id attrs for anchors! That was the point I was making in my previous comment.

I think I followed a similar path to what you did, albeit 7 years later. IMO frontend dev is fine, given the time that seems to be required in order to stumble on these "things we didn't know we needed to know", whether by concerted tuition/learning or simple wandering.

Nice to see simple fast sites!

The page isn't to bad. Better than certain other pages I visit that has to fill half the screen with a notice about how they ask for my permission to ignore the GDPR.

Would it be better without an animation? Possibly.

But given the current state of the web I'd ve fine with a few extra "i hate animations"-animations if they replaced sites that spy on me, only work in Chrome and advertises dating sites I don't need.

Edit: good -> not too bad

Still clearer and easier to navigate than 95% of websites on the Internet today. Also, even with the animation it loads on my iPhone SE without causing Safari to restart, which is more than I can say for many websites.

Indeed. The instruction latency tables in particular [1] are great as a quick reference when considering the cost of each instruction.

[1] https://www.agner.org/optimize/instruction_tables.pdf

Agner is a legend. I had the mildly hilarious experience of, during our acquisition by Intel, seeing that one of Intel's most talented network engineers (the sadly departed Venky Venkatesan) was using Agner's material as his reference rather than some super-secret awesome Intel docs.

His work is an even-handed and well-presented coverage of modern architecture. I'm perennially amazed by the way that people will regularly spout off about what might or might not be expensive or cheap on modern IA when you could just go look it up in Agner's work.

Intel's documentation is unbelievably bad. A damning statement on their documentation is that Google staffs a project to parse their horrible PDF and compare the supposed instruction latency and throughput figures to real measurements. https://github.com/google/EXEgesis

[ ex-Intel, although not connected to docs, and I did wind up submitting a number of corrections which got taken up ]

I think "unbelievably bad" is unfair; specifically, compared to what? If you think Intel docs are bad you must not have read many equivalent docs for other processors. We worked on a range of processors for the Hyperscan project before version 4.x, and while Intel docs leave a lot to be desired, you should see the other guys.

That said, there are a lot of things that need to be fixed.

L/T numbers are always idiosyncratic as there are a bunch of ways of getting them and none perfectly predict behavior for all instructions and all cases.

L/T numbers in the ORM frequently 'go missing' which isn't great.

The ORM is a giant 'additive song' that gets a new chapter with each new extension, and tries to provide optimization direction for over 10 years worth of processors. There's still Pentium 4 material in there!

There's somewhat of a culture of secrecy about the architecture that Intel would actually benefit from relaxing. It's clear that other vendors know how to get close to IA in performance or even pass it on some metrics. So I think it would benefit Intel a great deal to help people build better mental models of how their processors work.

The reason it is "unbelievably bad" is because it's not machine-readable. It's a PDF that comes out of FrameMaker for crying out loud. Some parts of it are not even human-readable, such as this garbled chart in ORM 3.7.6.1, the otherwise rather confusing description of the possible benefits of REP MOVSB on Ivy Bridge or later, under certain circumstances, maybe.

I mean, the image is just totally unreadable. https://imgur.com/a/jqxpqAf

I agree that it could be improved, but I think "unbelievably bad" is hyperbole.

This is especially true in light of comparisons to other vendors; I'm not aware of any up-to-date machine-readable information that's the equivalent for Appendix C for any other vendor, and historically (pre Intel acquisition) we had trouble finding this information out at all for some of our platforms to which we had Hyperscan ports. It was either out of date, not available, or behind a paywall ("show you've bought our SDK to get basic processor docs").

didn't know where to start so I stumbled over a wealth of information here https://www.agner.org/cultsel/ and looks my weekend will be filled with study of this fascinating pdf linked:

https://www.agner.org/cultsel/warlike_peaceful.pdf

>Regality theory is a theory saying that people show a preference for strong leadership in times of war or collective danger, but a preference for an egalitarian political system in times of peace and safety. These psychological preferences in individuals are reflected in the political structure and culture of the society. A society in danger will develop in the direction called regal, which includes strong nationalism, discipline, strict religiosity, patriarchy, strict sexual morals, and perfectionist art. A society in peace will develop in the opposite direction called kungic, which includes egalitarianism and tolerance. This book is both theoretical and experimental. The theoretical explanation of the regal-versus-kungic dimension is based on evolutionary psychology and human ecology. Contributions from the social sciences and the humanities are added to further analyze historical examples of regal and kungic developments. The theory is tested on data from both contemporary and ancient societies. These tests confirm the predictions of the theory.

thanks OP

This is actually why Agner is famous: https://www.agner.org/optimize/

Hennessy and Patterson don't cite Fog and that's just crazy. H+P then get basic facts about x86 architecture and microarchitecture wrong "The length of 80x86 instructions varies between 1 and 17 bytes." CA 5th, p A-23. No, it's 15 bytes as per the Intel Software Developer Manual.

Seriously, any practitioner should be reading Fog. Of course, they should read the Intel + AMD optimization manuals, Dr. Bandwidth, Developer Zone, the patents, but they definitely should be reading Fog.

Thanks Agner.

The microarchitecture reference also has some educational illustrations of microprocessor design plays out in the real world. E.g. section 9.9:

> Register read stalls has been a serious, and often neglected, bottleneck in previous processors since the Pentium Pro. All Intel processors based on the P6 microarchitecture and its successors, the Pentium M, Core and Nehalem microarchitectures have a limitation of two or three reads from the permanent register file per clock cycle.

> This bottleneck has now finally been removed in the Sandy Bridge and Ivy Bridge. In my experiments, I have found no practical limit to the number of register reads.

This (and the preceding related discussion) gives you a ton of insight into the limitations Intel engineers were facing and the trade-offs they made. Here, ports on a register file have a cost--they require transistors, may limit cycle time, etc. Intel realized that in many cases, the operands to an instruction would be available on the bypass network, and you could get away with having fewer ports on the register file than you'd theoretically need given the number of functional units.

Just be aware that some of the manuals there are outdated on some of their content. For example Optimizing software in C++: An optimization guide for Windows, Linux and Mac platforms. A lot of optimizations from this manual are performed by compilers automatically now, in smart pointers section you got information about auto_ptr as most used smart pointer, while it was deprecated years ago, he writes that composite types will almost always be copied out of the function if returned by value, while now we have move semantics and guaranteed copy elision etc.

I had Agner as a teacher when studying IT-engineering. At the time I didn't know anything about him really, but always found him pleasant and knowledgable. Looking back now, I wish I had spent some more time picking his brain about his expertise.

The webpage is awesome in my opinion. It's functional, simple and with a good amount of humor :) It a good example of a content-first approach.

I see he has a design for "a new open instruction set architecture and the corresponding hardware and software standards" [1]. It occurred to me to wonder if it had any relation to RISC-V, another open-source ISA [2], but apparently none at all.

[1] https://forwardcom.info/

[2] https://en.wikipedia.org/wiki/RISC-V

I am incorrect, in a subordinate document [3] he compares his ISA to RISC-V, Open RISC, and others.

[3] https://www.forwardcom.info/comparison.html

Reply on news.ycombinator.com