This is the second article in a series on making our very own calculator.

Introduction

When I was in 13th grade, I almost flunked math. Test after test, I was getting poor grades and it seemed that I just could not get those answers straight. I was messing up a lot. As the semester progressed, things were getting worse. Luckily, my late dad, having been a teacher himself in the past, understood the problem, and hired an old lady, a former professor, to tutor me.

She was from the old school: strict, authoritative, and always with a proverbial ruler in her hands to smack me when I would not follow her directions. (Yes, she really did it a few times, but gently, I should say. Those were different times, a different country).

After a couple of months, my grades dramatically improved. In fact, I remember the day after we got one test back, our teacher said in front of the whole class just how impressive my improvement was and that I was reaching the top of the class. It felt surreal.

She did not teach me math or how to remember the formulas.

What she did teach me was the method of doing the math.

She made me slow down and do my work, pedantically, on the paper. “Don’t rush! Slow down! Step by step! Verify it on the side!”, she would keep repeating. I was trying to calculate too many steps in my head since I was getting them right. Almost … most of the time, at least.

This brings me to the importance of step-by-step testing and verification, and this article is about prototyping and the verification of basic algorithms.

Algorithm Verification

We need to investigate and pick algorithms that we will use and verify them against a known good, “golden”, model. That platform also needs to be easy to use and debug. It should be quick to prototype various models so we can research different ideas before committing to them.

Normally, for such tasks, Python comes to mind, but for this project, I have selected C++ and MS Visual Studio’s basic console C++ project to construct such a framework. In my opinion, Visual Studio is the best freely available debugging framework. I also checked in a Makefile to compile it under Linux. Using both, I have also found certain discrepancies between the two compilers which I will describe.

The goal of this step is to verify the basic operations of addition, subtraction, multiplication, and division and the parsing of the input buffer. We want to investigate various ideas on how to do that effectively on BCD values, with mantissa and exponent processed separately. The code should be flexible and not tied to a particular mantissa width so that we could change it and see the impacts of such changes on the overall precision.

This code, a functional-level simulator, does not have to be too fancy, or fast. In fact, quite the opposite: the goal is to start mimicking some kind of low-level implementation that could be done in hardware so that the eventual translation to it may be simpler. Still, nothing is set in stone yet – as the verification progresses, it should become more obvious which parts would belong to what implementation level.

I have checked this project in on github here: Calculator Project: Proof of Concept (github.com)

I have rewritten that code many times and spent many evening hours on it as I was trying and debugging various approaches and algorithms.

Gradually, some certainties about the internal architecture started to emerge almost as all by itself. As you take certain directions, many related variables seem to naturally fall in place to form an architecturally consistent, and arguably perfect, model (ex. the way scratch registers extend calculation precision by adding extra digits). Those are the easy ones to decide, they are obvious and represent a sound engineering conclusion. The harder ones require giving up some obvious benefits to satisfy other variables (ex. selecting the mantissa or exponent widths). They are an ever-present nemesis of these engineering tradeoffs.

Somewhere in that spectrum is a decision to use BCD as the internal number specification.

Why BCD?

While the binary floating point (or IEEE 754 standard) is overwhelmingly used in general and scientific computation, fixed point BCD calculations are still very relevant in banking, finance, and commercial calculations due to their precision, accuracy, and the immunity from implicit conversion errors. In many ways, doing arithmetic with BCD digits resembles calculating values as taught in school, on a paper, digit by digit – there are no conversions to other number systems. Also, at any point in the process, digits are readily visible to a debugger which makes the algorithms easier to debug.

Hence, mantissas will be stored in BCD format.

Exponents

Exponents, however, will be stored as 8-bit binary (integer) numbers. That format simplifies the kinds of operations that need to be done with them – exponents are never addressed as individual digits and are always used as integer numbers, so keeping them in a multi-digit BCD format would make their handling unnecessarily more complicated. Moreover, we will add a bias value of 128 to it. That will “move” their range to positive integers and simplify their comparisons.

Most calculators use a two-digit exponent, from -99 to +99, so we will limit ours to that range, too. As they get internally converted to 8 binary bits, they will “lose” values in the ranges [0-28] and [228-255] and that is ok. (Maybe we could allow intermediate results to extend into those ranges.) Since exponents are always whole numbers (integers), we are not losing any precision when we convert them from a format of two-digit input buffer to integers, and back when we need to display their values. The process of such a 2-digit conversion is quite simple either way.

Input parsing

This is the first practical operation that we need to code and verify.

The deployed heuristic should be able to process any reasonably well-formatted number which a user might have typed in and produce a mantissa and exponent pair. Since the mantissa is stored in BCD format, a separate mantissa sign bit should also be extracted and stored.

We will allow two basic formats: with and without the exponent value. The test code contains a list of values that we consider valid and runs a verification algorithm on each. Note that the input parsing process does not check for incorrectly formatted numbers – it will be the job of the input editor routine to deliver us properly formatted buffers.

The input parser will have to normalize all input numbers. In this implementation, a normalized BCD register contains left-aligned digits, with an implied decimal point after the first digit. That decimal point is shifted by the value of the exponent.

All internal operations will be performed on such normalized values stored in registers.

Having specified the internal number representation, we can now tackle the four basic algorithms.

Addition/Subtraction

This implementation uses a “nibble-by-nibble” serial algorithm.

Usually, the subtraction is performed by negating the subtrahend and then using the addition. Most known BCD heuristics use 9’s or 10’s complement (subtract each BCD digit from 9 or 10, add one, and then add them). That is also similar to the way processors do subtraction of binary values.

After coding in several variations of popular algorithms, I have decided to create a slightly different one, one that is more suitable to the architecture that keeps the sign bits separately from their mantissas and does not store negative numbers as complements. Our implementation of subtraction does not negate the subtrahend. There are many ways to skin a cat, as they say.

Addition and subtraction operations are complementary and are used interchangeably depending on the signs of the addend and augend, or minuend and subtrahend (the former naming a base value, the latter naming a value to be added or subtracted).

First, we align the mantissas so that their digits represent equal decimal weights. The normalized form had them all left aligned while the exponent specified the position of their decimal points. We must shift the smaller of the two to the right by the difference of their exponents so that we can add ones to ones and tens to tens.  However, this could also result in a lesser number being shifted out of existence (ex. 1e20 – 1); we recognize such cases early and return the larger value unchanged.

Since the result may need a new significant digit (as in “9+9=18”), we need to have a nibble pre-allocated for it.

Our code splits up discrete addition and subtraction cases by this heuristic:

Another twist to subtraction is that we make sure to always subtract a lesser value from a greater one, avoiding the last (MSB digit) “borrow”. When the terms are swapped, the result is the same but negative (like in “10-3=7” but “3-10=-7”)

The implemented heuristic, although slightly different from the traditional ways to solve the problem, is well suited to this architectural implementation making the code not only readable and simple but more importantly, exact.

Multiplication

There are many methods of multiplying two BCD numbers. They range from traditional iterative partial-sum methods to intermediate conversion to binary value and back. There are also different approaches to implementing a multiplier in ASIC and FPGA. However, using an FPGA embedded multiplier would not be fun, would it?

The multiplication algorithm is using 2 extra scratch registers: one to store the intermediate product and the other one for a running total of the sums of products. In the outer loop, we multiply a multiplicand with each digit of a multiplier, while in the inner loop we multiply each pair of individual digits. Each result of these iterations is then being summed with the running total.

At this time, we assume that the hardware will be able to multiply a single (4-bit) BCD nibble with another one and return a 2-nibble wide result. This may be an oversimplification and may likely change as we start designing the hardware.

The sign of a result is an XOR of the signs of the input terms. The exponent is the sum of the exponents of those terms. Simple.

Division

We use a classic “Shift and Subtract” method to divide two numbers. This method mimics long division. Since both, the dividend and divisor, are already normalized, we do not need to align their decimal digits.

To obtain each digit of a quotient, we subtract the divisor from the dividend as many times as we can (for as long as the dividend is greater or equal to a divisor) and count. To get the next digit of a quotient, we shift the dividend one place (one digit) to the left (thus multiplying it by 10, decimal). Note that many classic “Division by Shift and Subtract” algorithms shift the divisor one bit to the right instead, a process that eventually “eats up” the bottom digits of the divisor, losing the precision.

Once we have found all the quotient digits, the dividend value (which is now numerically less than the divisor) represents the remainder. If we wanted to, this convenience leaves us with an option to implement a modulus or a remainder function in the future.

Not unlike multiplication, the sign of a result is simply an XOR of the signs of the input terms while the exponent is a difference of the term’s exponents.

With all four operations, we also need to normalize the result. In the end, the total number of digits might exceed the width of our scratch register causing a loss of accuracy.

Proof of Concept

To verify our calculations, the checked-in code carries along parallel computation using a standard C++ type “double”. This data type is convenient since, when printed, it provides up to 15 decimal digits of precision. Those are the “control” or “golden” values we compare our results with.

I could have used one of many “Arbitrary Precision Math” packages instead-but taking the advantage of a C++ built-in type suffices since we don’t need our calculator to exceed its precision. We also don’t need to tie-in another, potentially large, package with our sources.

Creating a test bench, we add a set of tests to each function we implement. The tests run a function using statically defined values that include several common and corner case numbers (ex. 0, 1, 0.00000001, 0.99999999).

These values are used in combination with each other and with varying mantissa and exponent signs.

Additional test sections use purely random values. Here we use C++11 “std::minstd_rand” to generate the same pseudo-random sequence across different, compliant, C++ platforms. That way, our VS and gcc runs will use the same random sequence and we can simply compare their outputs as text files.

Although such test coverage is not exhaustive, it seems to be sufficient.

The validator code compares the output of our algorithms with the calculated control value and prints “OK” if the values match exactly, “NEAR” if the values differ by only a magnitude of the last digit (a rounding error), or “FAIL” otherwise.

Since we currently do not perform any rounding, there are many results marked as “NEAR”, as expected.

It turns out that VS and gcc libraries round numbers differently, so there is a slight mismatch in the control values between those two platforms.

For example, on VS:
123456789012345 / 0.1 = 1.2345678901234e+15 vs. +1.2345678901235e+15  NEAR (5e-14)

while, on Linux gcc:
123456789012345 / 0.1 = 1.2345678901234e+15 vs. +1.2345678901234e+15  OK

(The last number in each line is the control value calculated using a standard VS or gcc library).

I also cross-checked by hand many calculations using SpeedCrunch software, which in this case shows 1.23456789012345e15; and as with many/all other values, the difference is clearly in the rounding digit.

During the development of that code, algorithms, and tests, in the subconscious pursuit of a “perfect match”, I had to keep reminding myself that the goal of this validator was not to match a particular C++ library perfectly but to provide a well-defined and consistent accuracy of the result. At one point I have coded in the rounding function and had matched VS control values perfectly. However, the same values would mismatch when compiled with gcc (by the said rounding errors), so I decided to keep the code simpler and not do the rounding, for now, implementing only the truncation policy.

There are two major defines in the validator code: MAX_MANT specifies the number of mantissa digits for input values and the result, and MAX_SCRATCH specifies the number of mantissa digits used in the internal calculation. The latter I call, “Arithmetic Scratch Register” since it should likely map to an equivalent register in our final design.

You can change those two defines and see the impact they make on the precision of the results.

Here is the somewhat abbreviated result of the verification: Verification.txt

After getting basic algorithms to work and to match the control values, I could verify successive changes to the code by quickly re-running the tests and spot functional regressions. Such verification setup made the development process faster and less error-prone.

Finally, I iteratively rewrote the code, simplifying it into conceptually basic steps (no fancy C++ constructs!) so that we could write some kind of microcode that mimics it. The microcode will be for a simple processor we will design, and the design will greatly be influenced by the kind of operations and data structures (registers) we use in these basic algorithms.

Just like doing the basic math in 13th grade properly: we are doing it step by step, not rushing it and we are verifying it on the side. This time, though, without a proverbial ruler looming in the air…

Interesting Reads

• HP uses a faster, parallel BCD adder circuitry: http://patft1.uspto.gov/netacgi/nph-Parser?patentnumber=8539015
• An HP division patent: http://patft1.uspto.gov/netacgi/nph-Parser?patentnumber=5638314, where they perform multiple subtractions at the same time to determine the closest divisor that would fit into a dividend or an intermediate remainder.
• BCD Addition and Subtraction is nicely explained here: BCD or Binary Coded Decimal | BCD Conversion Addition Subtraction | Electrical4U, not that we use a different method.
• A really good desktop calculator: SpeedCrunch
• Ken Shirriff peeks into a Sinclair’s 1974 calculator: Reversing Sinclair's amazing 1974 calculator hack - half the ROM of the HP-35 (righto.com)
• Rounding modes nicely explained: DIY Calculator :: Introducing Different Rounding Algorithms (clivemaxfield.com)

Introduction

The idea for this project came about during a week of freezing winter arctic event here in Austin, Texas, with a failed power grid as well as government, while keeping close to a gas fireplace, the only source of heat and light, for a couple of days. With a weak internet over a phone data line, I could only do some preliminary searches and mainly work out various details on a writing pad, growing a feeling that the project may be personally exciting and practically doable.

I will try to post my progress over the coming weeks.

How do calculators work?

You press a few buttons, a function, and you get your result back. But what is the technology and the algorithms at work behind its computation? For sure we can read about that, but to fully internalize all the tradeoffs and corner cases, we need to embark on designing one. To create one from scratch, and by that, I mean it - not use emulators and pre-existing ROMs, or even a powerful microcontroller, will provide us with that innate knowledge. We should start with a blank drawing pad onto which we would sketch our design from the very first gate and build on top of it.

Let's dive in.

Architectural Decisions

We will have to make certain implementation decisions, many of which are exclusive, each one with its own tradeoff.

Internal number representation

The number format internally used for computation:

1. BCD (binary-coded decimal), where each nibble stores one decimal digit of mantissa and exponent
   • More cumbersome operations on individual digits.
2. Floating point representation
   • Needs conversion to and from decimal values.
   • Suffers from conversion errors.

The drawbacks of BCD are more complex routines and slightly lower compactness; the advantage is the perfect representation of decimal magnitudes, as opposed to pure binary, that cannot represent exactly some of them; also that there are very refined algorithms for both normal arithmetic and transcendental functions.

Digits of precision

How many digits of precision do we want to carry? Here we need to be very careful not to show more digits than we can compute exactly (with the exception of the last, rounding, digit). A calculator that displays incorrect values is much worse than one that displays only a few, but correct, digits.

1. BCD may need internally wider registers to compute trailing, rounding digit, so it uses more digits for computation that it shows in the result (ex. HP41:10 digits, 13 internally; HP71: 12 digits, 15 internally)
2. Binary (floating point) may be off when rounding certain values that can't be represented exactly in a decimal system. This notation may follow a standard like IEEE 754 or implement it’s own partitioning of mantissa and exponent values

A major decision to be made is whether to use one of many rounding modes or simply truncate the result. Each rounding mode has its share of advantages and drawbacks, or it’s suitable for a particular use (financial, scientific, etc.)

Calculation notation

1. RPN (Reverse Polish Notation)
   • Does not have parenthesis, uses stack, computes intermediate values immediately.
   • Does not have "=" key but have "push to stack" key (often called "ENTER")
   • Quicker, less keystrokes.
2. Algebraic (or infix notation)
   • Uses parenthesis and knows about operator precedence.
   • In practice, requires more cognitive effort.
   • Have "(", ")" and "=" keys

Programmable vs. fixed function

1. Programmable calculators have larger non-volatile storage where they store user input sequence (a program) and provide additional functions for execution control (value testing, looping,...)
2. More sophisticated machines are more akin to full-blown computers and have a high-level language interpreter (like BASIC)

Gates vs Microcode

How much is done in hardware (in gates) vs. in microcode (ROM). Here, by gates, I mean strictly specialized digital logic dedicated to computing arithmetic functions.

1. Everything is done in gates (no ROM) (early, or very simple, calculators).
2. Very minimal basic operations in gates; most functions is in code (ex. using HP's Saturn CPU).
3. Most is done in gates, only the most complex functions in code.
4. Everything is done in code, minimal necessary functionality in gates (ex. some Z80-based calculators).

Algorithms

What algorithms are used to compute higher level functions?

1. Taylor series
2. CORDIC
3. Chebyshev polynomials
4. Other series and algorithms

Keypad

Keypad, or keyboard, is the primary interface to the user.

Arguably the best keypads are those made by HP for their line of calculators. They use double molded plastic and a proprietary spring system.

On the other end of the keyboard technology are super-cheap membrane keypads, such are those used in budget home computers by Sinclair and many $1 calculators from China.

A small step up is the membrane sheets pushed by extruded rubber/molded keyboard overlays. They provide more tactile feedback. They are very common since they are inexpensive and provide a good balance of price vs how they feel. Your TV remote controls, thermostats, and most of the appliances around would have those.

Occasionally, a frequently used button, after being pressed harder and harder, would stop working. That happened to our thermostat: two buttons setting up the temperature up and down stopped responding to our increasingly frustrating attempts to set the temperature. The remedy was to open the box and scratch the back of a rubber contact, a small graphite pad called “the contact pill”, with a soft black drawing pencil. Since the pencil leaves graphite marks, those contacts are now fully conducive and will continue working.

Since we cannot make plastic molds easily (perhaps they could be 3d printed, but I don’t have a 3d printer), the membrane-style keypad looks to be the most promising alternative for our calculator. We need to have lettering on the actual keys (numbers “0” through “9” and major functions like “+” and “-“), and also text for secondary functions accessed by some kind of a shift key, and those are normally expected to be shown next to the primary functions (above or below). We could print those on a regular piece of paper (printing using a laser color printer for nicely colored labels if we want to get fancy), and then laminate that sheet. The lamination would add extra stiffness to the paper.

There are two kinds of actuators we could use underneath.

The first one could be traditional PCB mounted push buttons. Since those have a certain thickness, we would make another PCB sheet, acting as a spacer, with round holes where the buttons are, providing that clearance. If the clearance is not sufficient, we could raise that board up with some spacers.

As you press a spot on a laminated sheet, it would press a switch; the intermediate board would keep it all flat and level.

Another option would be to use small, tactile dome switches. Their use is rather uncommon among hobbyists. Those switches are made of tiny pieces of domed conductive material (sometimes gold plated) which bend as you press to make contact in the middle. They make a quite satisfying tactile “click”. The domes should not be soldered to the board since that would stiffen their elasticity and would make them prone to failure, so they are either physically held by some template or they are glued on top (frequently they are sold with an adhesive sheet).

Surprisingly, although these cost only a fraction of a cent to manufacture, US electronics shops like Mouser sell them for 50 cents each! If you shop from Alibaba (China), you can get a bag of 500 for $10. If you are making a large order, you should not pay more than 0.4 cents per button! I am sure there are quality differences, but we are not building a nuclear reactor control panel, either. It’s ok if they are not perfect.

A variation of those dome switches has a hole in the middle, where you can PCB mount an LED behind it and, for example, have it lit as you press a switch, or show a latched state.

Although I’ve never used those dome switches before, I just ordered a bag from China and am planning to give it a try with this project. I hope they will work as visualized.

There is another kind of “switches”, or better, a lack of them. With the advent of microcontrollers and their special purpose IO blocks, many now offer touch detection which does not need anything else besides access to an open trace. The controller registers even the slightest capacitive touch of a finger. Since we are not using one of those MCUs, we won’t pursue that idea here.

Modes of Operation

When we are already making our own calculator, being engineers, we should plan to support several different number bases. Besides decimal, we should also have hex and binary, and we may as well throw in octal. We could, of course, architect it to support any arbitrary base, but I don’t see any practical use for something like that.

We want to follow conventions and call these modes DEC (decimal), HEX, OCT and BIN.

Any number currently on the screen can be converted amongst these bases. Non-whole decimal numbers will be truncated to integers.

Consequently, we need a few additional keys: besides the standard “0”…”9” we also need “A”, “B”, “C”, “D”, “E”, and “F”. We also need to assign 4 mode functions to keys.

We should also add a couple of bitwise operations to go with these modes. At the very least, basic AND, OR, XOR, NOR, and SHIFT (left or right) operations come to mind.
 

The CPU

The heart of our calculator will be a small micro-sequencer implemented in FPGA. We will try to implement IO as discrete blocks to make our life easier.

For example, keypad scanning could be a separate and independent module that takes care of debouncing and simply queues key codes to an internal FIFO.

Similarly, the LCD driver could also be a separate module that initializes the display and implements some simple protocol to control writing the characters.

The exact details should be more obvious once we start writing a concrete architectural document.

Summary

In this first article, I tried to provide an overview of various options we can select from, trying not to decide anything in particular. Other major things are still being left out (like the power source) because we are still wrapping our heads around the complete project. As I was writing, some things did start to crystalize in terms of dependency (“if you decide to implement it this way, then you can’t do that”) and feasibility. Still, anything is on the table.

Interesting Reads

• HP Saturn processor:  HP Saturn - Wikipedia
• Multi-cycle sequencer:  Appendix D Mapping Control to Hardware (tufts.edu)
• The Spectrum ROM Disassembly (Uses Chebyshev polynomials for trig functions):www.primrosebank.net/computers/zxspectrum/docs/CompleteSpectrumROMDisassemblyThe.pdf

• How HP calculates CORDIC: www.jacques-laporte.org/TheSecretOfTheAlgorithms.htm
• How CORDIC works:  Cordic for Dummies (jacques-laporte.org)
• HP35 on Arduino (with source): www.jacques-laporte.org/TheSecretOfTheAlgorithms.htm
• Calculator forensic:  Calculator Forensics (rskey.org)

This set of articles explains how HP calculators calculate various functions:

• SQRT:  1977 , Volume , Issue May-1977 (hp.com)
• Trig:  1977 , Volume , Issue June-1977 (hp.com)
• Inv trig.:  1977 , Volume , Issue Nov-1977 (hp.com)
• Logs:  1978 , Volume , Issue April-1978 (hp.com)