### Improvements to the Xerox Alto Mandelbrot drop runtime from 1 hour to 9 minutes

Last week I wrote a Mandelbrot set program for the Xerox Alto, which took an hour to generate the fractal. The point of this project was to learn how to use the Alto's bitmapped display, not make the fastest Mandelbrot set, so I wasn't concerned that this 1970s computer took so long to run. Even so, readers had detailed suggestions form performance improvements, so I figured I should test out these ideas. The results were much better than I expected, dropping the execution time from 1 hour to 9 minutes.

The Mandelbrot set, generated by a Xerox Alto computer.

The Alto was a revolutionary computer designed at Xerox PARC in 1973 to investigate personal computing. It introduced high-resolution bitmapped displays, the GUI, Ethernet and laser printers to the world, among other things. For my program I used BCPL, the primary Alto programming language and a precursor to C. In the photo above, the Alto computer is in the lower cabinet. The black box is the 2.5 megabyte disk drive. The Alto's unusual portrait display and an early optical mouse are on top.

## Easy optimization: mirror the Mandelbrot set

The Mandelbrot set is a famous fractal, generated by a simple algorithm. Each point on the plane represents a complex number c. You repeatedly iterate the complex function f(z) = z^2 + c. If the value diverges to infinity, the point is outside the set. Otherwise, the point is inside the set and the pixel is set to black.

Since the Mandelbrot set is symmetrical around the X axis, a simple optimization is to draw both halves at the same time, cutting the time in half. (This only helps if you're drawing the whole set; this doesn't help if you're zoomed in.) I implemented this, cutting the time down to about 30 minutes. The image below shows the mirrored appearance midway through computation.

Drawing both halves of the Mandelbrot set simultaneously doubles the performance.

## Improving the code

Embedded programmer Julian Skidmore had detailed comments on how to speed up my original code. I tried his changes and the time dropped from 30 to 24 minutes. Some of his changes were straightforward - calculating the pixel address in memory incrementally instead of using multiplication, and simplifying the loop counting. But one of his changes illustrates how primitive the Alto's instruction set is.

Quick background: my Mandelbrot program is implemented in BCPL, a programming language that was a precursor to C. The program is compiled to Nova machine code, the instruction set used by the popular Data General Nova 16-bit minicomputer. The Alto implements the Nova instruction set through microcode.

Since the Xerox Alto doesn't support floating point, I used 16-bit fixed point arithmetic: 4 bits to the left of the decimal point and 12 bits to the right. After multiplying two fixed point numbers with integer multiplication, the 32-bit result must be divided by 2^12 to restore the decimal point location. Usually if you're dividing by a power of 2, it's faster to do a bit shift. That's what I originally did, in the code below. (In BCPL, % is the OR operation, not modulo. ! is array indexing.)

```let x2sp = (x2!0 lshift 4) % (x2!1 rshift 12)
```

Unfortunately this turns out to be inefficient for a couple reasons. Modern processors usually have a barrel shifter, so you can efficiently shift a word by as many bits as you want. The Alto's instruction set, however, only shifts by one bit. So to right shift by 12 bits, the compiled code calls an assembly subroutine (RSH) that does 12 separate shift instructions, much slower than the single instruction I expected. The second problem is the instruction set (surprisingly) doesn't have a bitwise-OR instruction, so the OR operation is implemented in another subroutine (IOR).1 I took Julz's suggestion and used the MulDiv assembly-language function to multiply two numbers and divide by 4096 instead of shifting. It's still not fast (since the Alto doesn't have hardware multiply and divide), but at least it reduces the number of instructions executed.

## Shutting off the display

One way to speed up the Alto is to shut off the display.2 I tried this and improved the time from 24 minutes to 9 minutes, a remarkable improvement. Why does turning off the display make such a big difference?

One of the unusual design features of the Alto is that it performed many tasks in software that are usually performed in hardware, giving the Alto more flexibility. (As Alan Kay put it, "Hardware is just software crystallized early.") For instance, the CPU is responsible for copying data between memory and the disk or Ethernet interface. The CPU also periodically scans memory to refresh the dynamic RAM. These tasks are implemented in microcode, and the hardware switches between tasks as necessary, preempting low priority tasks to perform higher-priority tasks. Executing a user program has the lowest priority and runs only when there's nothing more important to be done.

All this task management was done in hardware, not by the operating system. The Xerox Alto doesn't use a microprocessor chip, but instead has a CPU built out of three boards of simple TTL chips. The board below shows one of the CPU boards, the control board that implements the microcode tasks and controls what the CPU is doing. It has PROMs to hold the microcode, 64-bit RAM chips (yes, just 64 bits) to remember what each task is doing, and more chips to determine which task has the highest priority.

Control board for the Xerox Alto. Part of the CPU, this board holds microcode and handles microcode tasks.

The task that affects the Mandelbrot program is the display task: to display pixels on the screen, the CPU must move the pixels for each scan line from RAM to the display board, 30 times a second, over and over. During this time, the CPU can't run any program instructions, so there's a large performance impact just from displaying pixels on the screen. Thus, not using the display causes the program to run much, much faster. I still set the Mandelbrot pixels in memory though, so when the program is done, I update the display pointer causing the set to immediately appear on the display. Thus, the Mandelbrot set still appears on screen; you just don't see it as it gets drawn.

## Microcode: the final frontier

The hardest way to optimize performance on the Alto is to write your own microcode. The Alto includes special microcode RAM, letting you extend the instruction set with new instructions. This feature was used by programs that required optimized graphics such as the Bravo text editor and the Draw program. Games such as Missile Command and Pinball also used microcode for better performance. Writing the Mandelbrot set code in microcode would undoubtedly improve performance. However, Alto microcode is pretty crazy, so I'm not going to try a microcode Mandelbrot.

## Conclusion

After writing a Mandelbrot program for the Xerox Alto, I received many suggestions for performance improvements. By implementing these suggestions, the time to generate the Mandelbrot set dropped from one hour to 9 minutes, a remarkable speedup. The biggest speedup came from turning off the display during computation; just putting static pixels on the screen uses up a huge amount of the processing power. My improved Mandelbrot code is on github.

My goal with the Mandelbrot was to learn how to use the Alto's high-resolution display from a BCPL program. Using what I learned with the Mandelbrot, I wrote a program to display images; an example is below.3

The Xerox Alto displaying an image of the Xerox Alto displaying...

## Notes and references

1. The Alto has an AND machine instruction but not an OR instruction, so the OR operation is performed by complementing an argument, ANDing, and complement-adding the complement. I.e. ab plus b.

2. Strictly speaking, I left the display on; it just wasn't displaying anything. The Alto uses a complex display system with a display list pointing to arbitrarily-sized blocks of pixels. (For instance, when displaying text, each block is a separate text line. Scrolling the screen just involves updating the list pointers, not moving actual data.) Thus, I could set the display list to NULL while rendering the Mandelbrot into memory. Then when the Mandelbrot was done, I simply updated the display list to make the image appear.

3. The recursive picture-in-picture effect is known as the Droste effect. After making this picture, I learned that generating the Droste effect on old computers is apparently a thing

### One-hour Mandelbrot: Creating a fractal on the vintage Xerox Alto

I wrote a short program to generate the Mandelbrot set on the Xerox Alto, a groundbreaking minicomputer from the 1970s. The program, in the obsolete BCPL language, ran very slowly—taking almost exactly an hour—but the result below shows off the Alto's monochrome bitmapped display. (Bitmapped displays were a rarity at the time because memory was so expensive.)

The Xerox Alto took an hour to generate the Mandelbrot set.

The Alto was a revolutionary computer designed at Xerox PARC in 1973 to investigate personal computing. It introduced the GUI, Ethernet and laser printers to the world, among other things. In the photo above, the Alto computer itself is in the lower cabinet. The Diablo disk drive (with the 1970s orange stripe) uses a removable 14 inch disk pack that stores 2.5 megabytes of data. (A bunch of disk packs are visible behind the Alto.) The Alto's display is bitmapped with 606x808 pixels in an unusual portrait orientation, and the optical mouse is next to the display.

Last year Y Combinator received an Alto from computer visionary Alan Kay and I'm helping restore the system, along with Marc Verdiell, Luca Severini and Carl Claunch. My full set of Alto posts is here and Marc's videos are here. I haven't posted an update for a while, but now I can write new programs and download them to the Alto using the Living Computer Museum's Alto file system implementation and gateway to the Alto's 3Mb Ethernet. I decided to start with the Mandelbrot set to take advantage of the Alto's high resolution display.

Marc's latest video shows the Mandelbrot programming running on the Alto.

## The Mandelbrot program

The Mandelbrot set algorithm is fairly simple. Each point on the plane represents a complex number c. You repeatedly iterate the complex function f(z) = z^2 + c. If the value diverges to infinity, the point is outside the set. Otherwise, the point is inside the set and the pixel is set to black. Setting the pixel is tricky because the Alto doesn't have a graphics API; you need to determine which bit in memory to set.4

Since the Xerox Alto doesn't support floating point1, I needed a way to represent the numbers with its 16-bit word. I use fixed point arithmetic: 4 bits to the left of the decimal point and 12 bits to the right.2 For instance, the number 1.25 is represented in 16 bits as 1.25*2^12 = 0x1400. These fixed point numbers can be added with standard integer addition. After multiplying two fixed point numbers with integer multiplication, the 32-bit result must be divided by 2^12 (i.e. shifted right by 12) to restore the decimal point location.3

The code (above) is written in BCPL, the main language used on the Alto. BCPL is a precursor to C and many features of C are clearly visible in BCPL: everything from lvalues and rvalues to the ternary operator. You can think of BCPL as C without types; the only BCPL types are 16-bit words along with C-like structs, unions and bitfields. BCPL may look unfamiliar at first, but the code above should be clear if you consider the following syntax differences with C:

• Blocks are indicated with [ and ] instead of { and }.
• Indexing is with a!1 instead of a[1].
• And, Or, and Shift bit operations are &, %, and lshift/rshift.
• Variable definitions use let.
• Arrays are defined with vec.

More information on BCPL is in the BCPL Reference Manual and my earlier article on using BCPL with the Alto.

The Xerox Alto, a few minutes into generation of the Mandelbrot set.

## Why is the Alto so slow?

Running the Mandelbrot set illustrates the amazing improvement in computer speed since the Alto was created in 1973 and the huge changes in computer architecture. On a modern computer, a Javascript program can generate the Mandelbrot set in a fraction of a second, while the Alto took an hour. The first factor is the Alto's slow clock speed of 5.88 MHz, hundreds of times slower than a modern processor. In addition, the Alto doesn't execute machine instructions directly, but uses a relatively inefficient microcode emulator that takes many cycles to perform one machine instruction.

The ALU board from the Xerox Alto. The Alto doesn't use a microprocessor chip, but a CPU built from three boards of integrated circuits.

Unlike modern computers, the Alto doesn't use a microprocessor chip, but instead has a CPU built from three boards full of simple TTL chips. The photo above shows the arithmetic-logic unit (ALU) board, which uses four 4-bit 74181 ALU chips to perform addition, subtraction and logic operations. You can also see the CPU's registers on this board. The Alto doesn't include a hardware multiplier, but must perform multiplication by repeated shifts and adds. Thus, the Alto performs especially poorly on the Mandelbrot set, which is essentially repeated multiplications.

## Conclusion

The Mandelbrot set was a quick program to try out the Alto's graphics. Next I'll try some more complex projects on the Alto. If you want to run my code, it's on Github; you can run it on the ContrAlto simulator if you don't have an Alto available.

If you're interested in retrocomputing fractals, I also generated a Mandelbrot on a 50 year old IBM 1401 mainframe The 1401 generated the Mandelbrot set in 12 minutes—not because it's a faster machine than the Alto, but because the resolution on the line printer was very very low.

Mandelbrot generated on the IBM 1401 mainframe.

## Notes and references

1. There is a floating point library (source) for the Alto. I decided not use use it since the integer Mandelbrot was already very slow. But using floating point would make sense if you wanted to zoom in on the Mandelbrot.

2. Fixed-point arithmetic is a common trick for fast Mandelbrot calculation.

3. To multiply two 16-bit numbers efficiently, I use the double precision MulFull function (written in Nova assembler) in PressML.asm, part of the Computer History Museum's archived Alto software.

4. The hardest part of generating the Alto Mandelbrot was figuring out how to configure the display memory and update it correctly. The details on how the display works are in chapter 4 of the Xerox Alto Hardware Manual. To summarize, the display contents are defined by a linked list of display control blocks (DCBs), which define a rectangular region of pixels on the display. A microcode task reads 16 words of pixels from memory at a time and writes them to the display board, which shifts the pixels out to the monitor. Thus, as each scanline is being written to the CRT, the CPU is busily reading the pixels for that line from memory and feeding them to the display, another reason why the Alto is slow.

The Alto's Smalltalk environment has a simple graphics API, but we don't have Smalltalk running yet.