A great old-timey game-programming hack (2013)
blog.moertel.comSomewhat related [1]:
> Years ago, I was working at Video Seven, a now-vanished video adapter manufacturer, helping to develop a VGA clone. The fellow who was designing Video Seven’s VGA chip, Tom Wilson, had worked around the clock for months to make his VGA run as fast as possible, and was confident he had pretty much maxed out its performance. As Tom was putting the finishing touches on his chip design, however, news came fourth-hand that a competitor, Paradise, had juiced up the performance of the clone they were developing by putting in a FIFO.
> That was all he knew; there was no information about what sort of FIFO, or how much it helped, or anything else. Nonetheless, Tom, normally an affable, laid-back sort, took on the wide-awake, haunted look of a man with too much caffeine in him and no answers to show for it, as he tried to figure out, from hopelessly thin information, what Paradise had done. Finally, he concluded that Paradise must have put a write FIFO between the system bus and the VGA, so that when the CPU wrote to video memory, the write immediately went into the FIFO, allowing the CPU to keep on processing instead of stalling each time it wrote to display memory.
> Tom couldn’t spare the gates or the time to do a full FIFO, but he could implement a one-deep FIFO, allowing the CPU to get one write ahead of the VGA. He wasn’t sure how well it would work, but it was all he could do, so he put it in and taped out the chip.
> The one-deep FIFO turned out to work astonishingly well; for a time, Video Seven’s VGAs were the fastest around, a testament to Tom’s ingenuity and creativity under pressure. However, the truly remarkable part of this story is that Paradise’s FIFO design turned out to bear not the slightest resemblance to Tom’s, and didn’t work as well. Paradise had stuck a read FIFO between display memory and the video output stage of the VGA, allowing the video output to read ahead, so that when the CPU wanted to access display memory, pixels could come from the FIFO while the CPU was serviced immediately. That did indeed help performance—but not as much as Tom’s write FIFO.
[1] https://www.jagregory.com/abrash-black-book/#the-challenge-o...
this alleged fifo sounds like a register with a two-phase handshake bit, but they lived in jargon deserts in those days
It's exact nature is kinda irrelevant, the important part is that it was a cheap solution that gave a damn good return on investment just by pondering on possible improvements based on a rumor.
Back in the MS-DOS days, my friend Ward Christensen worked at IBM in Technical Sales Support. They had a file informally maintained that listed every single product they had for sale, in a format that could be read on a PC. He was really good at searching it with his text editor, but he wanted a MS-DOS based stand alone program to make it easy to search for his co-workers. Thus was born, PCxREF which was a Turbo Pascal based program I wrote, that included search.
My first pass at it was a naïve search using the built in string matching. Because of the size of the file, and the speed of PCs, it would sometimes take more than 30 seconds to do a search. Ward wanted it faster, much faster.
I eventually came up with an algorithm that took the text you wanted to search for, and changed it into a special kind of bit mask, table. It was a table 256 bytes long, initialized to zeros. If the first character of the string you wanted to search for was a D, then the first bit of the $44 and $64 mask would be set, to handle case sensitivity. A wildcard * was only 1 character long, but would set all bits for that character. I then used the 8088 XLAT instruction and a very small loop ingest a character of the text being searched, use that as an index into the mask table, and AND that with the existing flag in a different register, then shifted it, checking for CARRY (a match!) and looping if not zero.
It took search down to less than 3 seconds if the whole file was loaded in RAM. The thing we used as a benchmark was searching for "DOS TECH REF", the IBM DOS Technical Reference Manual.
I can't find the code, lost it over the years, but it was one of my coolest hacks.
These kinds of things are still important these days. In my previous employment we had to create security guardrails for content distribution resources in AWS (API gateways, s3 buckets, instances and cloudfront distros) for a standalone CDN. For cloudfront you'd be right to question why put a CDN behind a CDN and the answer is technical debt, so we had to deal with it.
The problem is that cloudfront functions are very limited in code size and execution cycles, and the list of cidr blocks from the external CDN was so long it meant we couldn't do a simple comparison, so we did something similar, representing all cidr blocks and the incoming source IP to a binary representation and used binary search to see if the source IP matched any cidr block range. It's also very hard to estimate correctly if the resulting function would in fact fit the compute budget, because AWS isn't very clear about what a cf function cycle is in practice, so we tested it extensively, with real and synthetic cases to simulate worse conditions.
Awesome article. I guess grass is always greener but coding in those days seems really cool. Today it's all package managers this and toolchains that. We're so far abstracted away from what's actually running our code. Articles like this make me here the CPU calling "come back to meeeeee"
It depends a lot on what you work with. I am a much younger generation of game programmer (only 15 years of professional experience), but a lot of the work I do has a similar feel.
Not saying that most game programming is like that, but if you specialize in graphics and performance, preferably in large productions, you get to work close to the hardware and solve a lot of really fun problems in interesting ways.
You can do that now - and id argue its even easier to get into it.
WebGPU in a browser is worth the learning curve - and gives you the same interface then in Rust!
When you sniffed that sweet sweet shader code and memory layouting - why not get into WebAssembly?
You can have sth very similar to an OS in your browser if you want. It's a lot of fun and knowledge actually useful instead of the next hypecycle of frameworks and dialects.
These days if you want to get esoteric you can do it in GPU shaders; https://www.shadertoy.com/ provides lots of fun examples with code.
Or there's MCUs with programmable IO for bit-bashing protocols.
GPU shading languages are about as abstract as CPU programming languages. I would rather suggest bare-metal programming, preferably on some obscure hardware.
I guess my equivalent these days is optimizing distributed services - a small change here, a realization there and you can save whole percentages of performance in affected services when effectively DDoSing yourself.
The 6502 has a more narrow and diabolical version of the 6809's PSHS trick. The BRK instruction pushes the PC and status register to the stack in three consecutive cycles, and is pretty much the only non-interrupt instruction that writes more than twice.
The Atari 2600 has horizontal object counters that you can't set directly, but they latch when you write to them, and they set the corresponding object to the X position of the beam at that cycle. Pole Position requires three of these objects to draw the left/right side of the road and center line, and they are supposed to meet at the horizon.
But the beam moves to the right 3 pixels for every CPU cycle, and writing with STA would result in each being separated by 9 pixels. What's the fix? Move the stack pointer to where the object pointers live (the 2600 maps the stack to zero page, where the RAM and registers live) and invoke BRK, making sure the BRK vector returns you to the next instruction. The object pointers are adjacent, so BRK writes to each object pointer on successive cycles, and now each of them is 1 cycle = 3 pixels apart.
Great story! We may write little assembler (or, increasingly, code) these days but, optimization always has this same feeling of adventure, serial discovery, and of learning what parts it’s ok to throw away.
Trying to work out which computer actually had the 6509 in; candidates seem to be Dragon or Tandy?
This sort of thing is why the Amiga was so massively successful by adding a collection of co-processors: https://en.wikipedia.org/wiki/Amiga_Original_Chip_Set
Each of which had limited but dedicated functionality for moving and transforming bytes in RAM. The blitter in particular was dedicated to moving memory around for tile and sprite rendering.
The author stated in comments that it was a Tandy Color Computer 3: https://news.ycombinator.com/item?id=6913467
> Here’s the quick version: The graphics hardware didn’t do jack squat.
On the PC in question, if you wanted to draw something onto the screen, you had to do that yourself, byte by byte. No texture mappers, no blitters. Just bytes. Bytes that you had to move yourself.
That's not entirely correct: the graphics hardware took the video memory and built an image from that (frame buffer graphics). Which is already more than the earliest game consoles did.
A ZX81 works like one of those pre-framebuffer machines:
At some point, the Z80 cpu jumps into an array that holds character info for each line on the screen. Its ULA chip intercepts that instruction fetch, and feeds the Z80 a NOP instruction.
Character itself is read by the ULA, and output onto the address bus. The ROM then outputs the corresponding bit pattern. ULA reads that, and clocks it into a shift register: 8 bits parallel in, serial out produces the pixels on screen.
But here's the nifty bit: that ULA stuff is timed to take exactly 8 cycles of a 6.5 MHz pixel clock. The Z80 runs on 1/2 that, so it takes the exact same time to execute the NOP (4 cycles). Done -> next instruction fetch, ULA ready to grab bit pattern for the next 8 pixels.
That's how screen lines are built on a ZX81. 50/60 Hz screen refresh is done by varying the # of lines displayed. Downside is that the (already 'slow') CPU is executing NOPs ~3/4th of the time, making for an effective CPU speed of ~0.8 MHz.
But the scheme was somewhat flexible, and programmers found ways to produce high-resolution graphics by (ab)using ROM data for non-text graphics. And with a RAM pack you could enjoy full hi-res graphics. All of this in b&w only.
The follow-up ZX Spectrum added a regular framebuffer with color (and even hardware flashing!) in 8x8 pixel blocks. This required a much more complex ULA chip though.
This process is what’s missing in the ai era.
That's the secret, everyone who's done nasty tricks have become donors of tricks to be used.
That said, modern machines have enough oomph that tricks of that kind haven't really been the major contributions on the table for the past 25 years (sure, there's some tricks with PS2 VU's/memory, register swaps,etc but they're small in comparison to the total product sizes).
idk about that. people will always find a way of working together to solve problems. that "solving" the AI era means "pulling the lever on the claude machine" is sad, I agree
Such sweeping generalisations make me smile, as they can’t possibly be true.
There’s a guy who has got Command & Conquer running on Atari ST at ~6fps (it started out at ~2fps). Months and months of thinking up new tricks and approaches by himself and with his agents. It’s only missing of you choose not to do it.
Monster Max on Game Boy happens to use unrolled "ld de,NNNN / push de" loops constructed in main memory to write to video ram.
For those unfamiliar with Monster Max (1994, Game Boy), it's basically the last of the British Isometric Platformers, a genre started by Knight Lore on ZX spectrum.
Ooh, might steal that for my gameboy experiments if I get stuck in some perf-problem (right now I have asynchronous upload queueus that works most of the time since things are queued well ahead of time so it's updates outside of the screen).
For reference, unlike the PC,C64,etc f.ex. the Gameboy (and NES) only allows users to write memory during VBlanks/HBlanks (or when the screen is turned off), so instead of delaying and adding complicated logic for that due to memory usage contention the CPU is simply cut-off from the memory (thus reads/writes at the wrong time are just nonsense/discarded iirc).
This type of story is always so cool to read. In this type of community there's often a bit of wistful dreaming back to these days and the type of challenges they presented.
I'm young enough that I wasn't there, but I've been programming since I was a kid and working with graphics for the last 15 years.
This type of excitement is the feeling I got from learning C and C++ to begin with. Trying to really understand how things were working under the hood.
I still get that sense when working on low level graphics and performance in my current job. In fact, when the texture memory corruption was introduced the article felt eerily similar to the events which inspired me to write my own recent article about texture memory layouts ( https://agentlien.github.io/texture )
The clever 6809 PSHU/PULS memory copy trick was probably independently discovered several times before Tom Moertel discovered it. The game Tom worked on was for the Coco 3, which was released in 1986.
My first computer was the original 4K Color Computer (1981). I learned 6809 assembler on it and worked full time programming it for several years, while also running one of the largest users groups for it. Traveling to Coco trade shows around the country, I met many of the devs writing the most popular Coco games and the trick was definitely known by around Summer 1983 as that's when the first Rainbowfest was held and I heard about there.
Around 1983, Steve Bjork (MegaBug / Zaxxon Coco) told me the devs at DataSoft, the company behind some of the popular 1st party cartridge games Radio Shack sold, were using it. I also heard about it in 1984 from Tim Jenison (CocoMax). The earliest usage of the trick in a commercial game was probably in a Williams arcade game, either Defender (1981) or Joust (1982). A few years ago at the Hackers Conference Tim told me he'd once asked Eugene Jarvis (Joust, Defender) about it and Eugene told him after discovering it they'd nicknamed the trick "PUSH-U / PUL-Me." I confirmed this with Sam Dicker (Sinistar (1983), Defender) last year during the Amiga 040th anniversary at the Computer History Museum.
While the basic trick to do a fast software block image transfer was discovered separately by several 6809 programmers, I can't recall ever seeing the Direct Page (DP) register used with it before today. That and some of the other refinements in TFA are quite clever additions. I think the Coco's lack of any hardware sprites, tiles, scrolling or acceleration made Coco devs especially desperate, so we had to become masters of software blits with PSHU/PULS and software sprites with XOR. And since all this was happening before Usenet and dial-up BBSes were commonly available, there was no way to share cool tricks widely except printed books and magazines.
Fortunately, the 6809 made up for Radio Shack's penny-pinching design by being much faster than the 6502 and Z80, which were microcoded while the 6809 was implemented directly in logic gates. It's widely considered the most advanced of all the 8-bit CPUs with an especially elegant, orthogonal ISA supporting position independent, re-entrant, multitasking code and powerful indirect addressing modes. One little known bit of history is that Jef Raskin's team at Apple used the 6809 for the initial Macintosh prototypes before Steve Jobs managed to cut a steep discount deal with Motorola on the workstation-priced 68000 CPU. By 1981, Moto was probably starting to realize they'd been pricing their CPUs far too high, having lost the Apple II, Commodore PET, VIC-20, C64 and then IBM PC designs, despite having a better chip. Waking up just in time to win, arguably, the LAST big desktop computer CPU design... ever? Way to go Moto!
2013 Hacker News comments here: https://news.ycombinator.com/item?id=6913467
> A few years ago at the Hackers Conference Tim told me he'd once asked Eugene Jarvis (Joust, Defender) about it and Eugene told him after discovering it they'd nicknamed the trick "PUSH-U / PUL-Me."
It's interesting to think about the traditional attitude towards documentation. Back in the day, people could discover incredibly clever things, but when it came to writing stuff down and distributing the information in an organized way, it seems like you only got the "reference" documentation. The books you could get would only list the opcodes; to learn a trick like this, you'd either have to talk to someone who knew it, or figure it out again.
The 6502 was most definitely not microcoded! It's way simple.
On the Vectrex, also a 6809 machine, the fastest vector draw routines use the stack to thread draw commands, not quite the same thing but those sixteen bit stack pointers sure are useful.
Doh! You're correct - and I should have known that since the 6502 design was based on a greatly cost-reduced Motorola 6800, the 6809's dad. Somehow I bit-flipped between the 6502/Z80 (not microcoded) and the 8088 (microcoded). Thanks for the correction.
Re: the Vectrex, I never knew that but it makes sense since chasing the beam requires feeding a massive queue of vector pairs.
I started programming by making games in TI-BASIC, for the TI-83+ graphing calculators. There were some weird tricks like omitting all ending parentheses. Good preparation for weird hacks in real languages.
oh that's a miserable environment. i remember discovering that the label search starts from the beginning of your program text each time you branch (there's no table of label positions) so branching is much slower in later parts of your code.
the ti-81 was somehow worse. the If conditional only had the implicit else form (no Then/Else/End), there were no looping constructs (just Lbl, Goto, DS< and IS>), you had two lists for array-like data, and i can't recall ever figuring out how to manipulate strings—it may not have been possible.
Wonderful. What an extraordinary effort and great results. Makes my 6502 efforts (using page boundaries) seem rather simple. Those were the days!
damn that is quite literally the lore behind this leetcode problem - Rotate Image: https://leetcode.com/problems/rotate-image/
Terry Davis in the comments plugging Temple OS is a bonus
i've written a bit of HolyC and at one point knew how to get around in that environment. it's pretty neat actually. it's easy to make fun of his clunky demos and mental illness, but he built something incredibly entertaining to work with.
as he emphasized endlessly, you can modify the running kernel, so it's easy to rip out his fanatical religious stuff.
Seeing that actually shocked me a bit. I remember when I was first on HN I'd see occasional messages of him either promoting his TempleOS or going off on crazy aggressive rants completely out of context. A lot of comments were saying never mind Terry, he isn't well. It's very sad how his story ended.
this feels like precursor of SIMD mechanism, instead of byte processing, process the capacity (max). I remember my image processing class with Ian R. Greenshild. We had to dig a lot of tricks for matrix calculations. But the most important thing is that you have to rule the hardware (IRQs, Registers, PCs, etc). Thank you for the article, it made me smile and remember my 80s...
I remember doing software rendered truecolor 3d graphics in the mid/later 90s on my computer (that was probably among the last Intel machines without MMX).
Shading was expensive since you had 3 color components to scale, but one trick if you had monochrome color was that you could do it with 1(16bit) or 2 multiplications(32bit) by moving bits around.
If you had RGB, you could do red and blue at once since you'd easily mask by 0xff00ff , then multiply by the intensity (0-0x100), since the result of 0xff*0x100 is at most 0xff00 you could just mask the result by 0xff00ff00 and shift it down 8 bits (the shifting could be with the green component combined back in).
1 multiply for 16bit color (5R 6G 5B) was basically the same thing by shifting green upwards, or'ing, masking and multiply a 5bit intensity once on the 32bit pattern 00000gggg gggg00000 rrrrr000 000bbbbb before shifting down the green part again and combining the result.
I'm sure others had even more similarly clever tricks back in that era.
On modern CPUs such tricks are irrelevant. They are able to process multiple instructions in the same time, so that the main bottleneck is no longer executing instructions, but accessing memory.
Kind of, but not really? At its core this article is basically a hack to do a bit of SIMD on an instruction set that doesn't have SIMD. Obviously SIMD is still highly beneficial on modern CPUs otherwise compilers would stop emitting it.
Eh, you can always move the performance frontier into something else, and the SIMD capabilities give you lots of strangely shaped instructions to play with. We have absolutely not seen the end of clever hand written code, it's just very rare.
Those specific tricks are irrelevant mostly due to compiler optimisations, even in compute heavy workloads.
I love reading about this stuff, even if I understand very little of it. Similar to the demo scene and code golf stuff, I'm amazed at the knowledge that some people have.
I was already impressed by the stack pointer trick, but then the interrupts started corrupting memory and I thought, well, that's the end of that. Somehow the solution was to deliberately corrupt a different part of memory and change the drawing order. That's the kind of thing you'd never believe if you hadn't seen it work.
obligatory: The Story of Mel
It is definitely in the spirit of the story of Mel, since Tom was exploiting an intimate knowledge of the machine to achieve the impossible by using instructions in unintended ways, but always remember that Tom is not a real programmer:
That said, I prefer Tom's story. It's more about the joy of understanding the machine than it is about the framing of a legend.Real Programmers wrote in machine code. Not FORTRAN. Not RATFOR. Not, even, assembly language. Machine Code. Raw, unadorned, inscrutable hexadecimal numbers. Directly.
Man seeing Terry Davis commenting and promoting TempleOS "13 years ago"... Absolutely crazy stuff! This was an awesome read!