Awesome

Analytical Engine Emulator

This is an Ada translation of the Java emulator at Fourmilab.

In due course, it is hoped to create a web-based emulator (using Gnoga). In the mean time, there is a command line based version.

Copying

The Java source of the Fourmilab emulator is included in this repository for reference (in the fourmilab/ subdirectory); the Analytical Engine code's author states Emulator source code is intended for experienced Java developers, and is utterly unsupported. The program is in the public domain and you can do anything you like with it, but you're entirely on your own., while BigInt.java is copyrighted by the Massachusetts Institute of Technology.

The Ada source is licensed under the GNU General Public License (version 3) with the GCC Runtime Library exception.

Building

The Ada emulator is written using Ada2012, and uses the GNU Multiple Precision Arithmetic library, via the GMP binding in GNATCOLL.

If you have one of the FSF GCC binaries for OS X (6.1 or later), a suitable library is provided. Otherwise, GMP and GNATCOLL are reasonably straightforward to build.

Running

There are three switches:

• -h or --help: outputs help.
• -t or --trace: trace execution (like card T1)
• -z or --overwrite-nonzero: allow storing to a non-zero column

and, as usual, run by

./aes [switches] [card-chain-file]

(uses standard input if no card chain file is supplied).

The "storing to a non-zero column" issue is referenced in this paper by Rainer Glaschick, section 3.5, Memory peculiarities. It would have been necessary to ensure that a column that was to be written to was zero beforehand; the operation of storing a value involved racks transferring the value on the source to the destination as the source digits were rotated back to zeros.

If -z is given, overwriting will be allowed; otherwise, an error will be reported and execution will halt.

Programming Cards

The design here is as in the Fourmilab Java implementation. To date only the card types below have been implemented:

• Number Cards
• Operation Cards
• Variable Cards
• Stepping Up and Down Cards
• Combinatorial Cards
• Action Cards
• Comment Cards
• Attendant Request Cards:
  • Calculation Trace Cards
  • Numeric Output Format Cards
  • Output Annotation Cards

These card types have not been implemented (yet):

• Curve Drawing Cards
• Attendant Request Cards:
  • Advancing and Backing Block Cards
  • Card Library Inclusion Requests
  • Decimal Place Expansion Cards

As in the Fourmilab Java implementation, multiplication can be indicated by * or ×, division by / or ÷.

When counting for Combinatorial Cards (conditional and unconditional jumps), remember that comment cards (blank cards, or cards starting with a period or white space) need to be included!

Changes

• As a minor change, lower case can be used: n001 42 is acceptable (it stores 42 into column 1).
• Text after the required content is ignored, so comments can be included: n001 42 the answer is acceptable.

Examples

The file bernouilli.ae is an implementation of the Lovelace design in Note G of the Sketch of the Analytical Engine (the last section), with two errors corrected:

• in operation 4, the operands are reversed (it should be V4 / V5, not V5 / V4).
• in operation 24, the result placed in V24 should be **-**V13.

AAL's notes don't discuss the fact that the Engine only implemented integer arithmetic, which means that a naive attempt at dividing 7 by 9 at operation 4 will result in 0 remainder 7, not the 0.777777777 one might have hoped for. To deal with this, all real values are scaled by 10 decimal places;

• the precalculated value of B1, 0.1666666666, is stored as 1666666666, and similarly for B3, B5.
• dividends are scaled up by 10 decimal places.
• products are scaled down by 10 decimal places.

For more on this, see the notes on Stepping Up and Down Cards in the Fourmilab Java implementation.

The file bernouilli5.ae adds another "iteration" to the above program; the previously computed B7 is stored on column 24, and operations 13 to 23 are repeated once more (now using column 24 as input at operation 21).

The file check_for_prime.ae determines whether a number is prime. It's set to check 203 (AAL's 203rd birthday was in 2018).

Performance

The Deep Learning with the Analytical Engine project, with the Deep Learning with the Analytical Engine repository, reports on an

implementation of a convolutional neural network as a program for Charles Babbage's Analytical Engine, capable of recognising handwritten digits to a high degree of accuracy (98.5% if provided with a sufficient amount of training data and left running sufficiently long).

The program consists of over 400,000 cards! (it was generated by Python scripts from a high-level description). It requires -z to run. As a timing test, the program was run for the first 10 images only: in Java, this took 18 seconds, of which 10 seconds was initial processing, in Ada (where the initial processing was about a second) with -O0 -gnata 44 seconds, with -O2 -gnatp 22 seconds. So the Ada implementation is about 3 times slower than the Java one. More work to be done!