Enter email address here to subscribe to B5500 blog posts...

Sunday, November 20, 2016

Emulator Release 1.04

Nigel and I are pleased to announce that version 1.04 of the retro-B5500 emulator was released on 4 September 2016. All changes have been posted to the repository for our GitHub project. The hosting site has also been updated with this release, and for those of you running your own web server, a zip file of the release can be downloaded from Google Drive.

This is a minor release containing a number of enhancements and corrections for issues we have discovered since version 1.03 was released a little over a year ago.


Enhancements in Release 1.04


This release of the emulator supports the following enhancements.

New Single-Precision Add/Subtract Implementation


The B5500 uses a unified numeric format, which means that it does not have the typical distinction between integer and floating-point numbers that most computer architectures do. Integers are simply values that have been normalized so that their exponent is zero. For exaple, octal 0000000000000001 (the integer representation of 1) and 1141000000000000 (the normalized floating-point representation of 1.0, i.e., octal 1000000000000 × 8-12) are the same value as far as the B5500 numeric operators are concerned.

There is a document informally referred to as "The Flows," which contains a schematic representation of the logic equations for each operator syllable. Alas, we did not have access to this document when initially developing the emulator, and had to rely on its companion, the Processor Training Manual, which is a good, but not-quite-complete narrative guide to The Flows. As a result, the first implementation of single-precision arithmetic was something of a best-guess effort, and as reported previously in this blog, some of my early numeric benchmarks were producing results that had at best one significant digit of agreement with results from a real B5500.

After much frustrating research, that problem turned out to be improper rounding in single-precision Add and Subtract, which are implemented as the same instruction with some small differences in the treatment of signs. I finally fixed the problem empirically by generating results from 64 "interesting" numeric values and comparing the octal results to those from a modern Unisys MCP system, which uses the same numeric word format. That revealed many cases where rounding was being done improperly -- mostly during scaling of the smaller operand to make exponents equal. Fixing those cases allowed the benchmarks to agree with the results from a real B5500.

By the time I got around to implementing the double-precision arithmetic operators about a year later, The Flows had become available on bitsavers.org, and I followed them closely in implementing the double-precision operators. That gave me the idea that at some point I should go back and redo single-precision Add/Subtract with more rigorous attention to implementing them the way The Flows said they worked.

Earlier this year, I had a lengthly and very interesting exchange of correspondence with Reinhard Meyer in Germany, who was contemplating design of a B5500 emulator of his own and perhaps eventually producing an FPGA implementation of the machine. That exchange finally galvanized me to go back and re-implement single-precision Add/Subtract according to The Flows.

Rigorously following The Flows turned out to be considerably more of a challenge than I had originally anticipated. Digital logic tends to work in parallel, with many states changing at each clock pulse. Most of that translates fairly easily to the sequentially-executed code of a traditional programming language, but some of it does not, and you continuously need to be on guard that states set during one clock cycle do not become effective until the start of the next one.

The most frustrating problem that arose during this effort was one where subtracting two numbers of equal value generated a result that was not completely zero. The mantissa would be zero, but the sign or exponent would not. There is explicit logic in The Flows to test for this condition in the J05L block under the secondary control of Q04F/ (see index 1.01.1 page 2 of 2 in the document) and exit the operator early if the sign or exponent do not need to be cleared to zero. That test is based on the logic term W03L, which is defined as "A[39->37] = B[39->37]", meaning the high-order octade of the A-register mantissa (bits 39 through 37) equals the high-order octade of the B-register mantissa. The corresponding narrative in the Processor Training Manual (page 3.17-11) states, "If the 13th octade of the A and B registers do not equal each other (W03L/) it is not possible for the resultant sum of the complement addition to equal all zeros and the operator is terminated."

I implemented the code exactly that way. It didn't work. After staring at that portion of the logic for days, it finally occurred to me that neither the definition of W03L nor the statement in the Training Manual made sense -- it's entirely possible for the two high-order octades to be unequal and still sum to zero -- 7+1 (ignoring the carry), for example. That made me begin to wonder whether "equal" and "not equal" meant what I thought they did. To answer that question, I had to look at the document upon which The Flows are based, the Processor Equation Book, where on page 422.00, I found the term W03L/ to be defined as:

A39F*B39F + A39F/*B39F/ + A38F*B38F + A38F/*B38F/ + A37F*B37F + A37F/*B37F/

This is somewhat simplified from the notation in the Equation Book, which appears to refer to physical circuit elements. For the uninitiated, "*" means logical-AND, "+" means logical-OR, and "/" used as a suffix means logical-NOT. Operator precedence is NOT-AND-OR, highest to lowest. The numbers refer to bits in the registers, with 48 being high-order and 01 being low-order. A suffix of "F" means flip-flop.

That equation clearly does not determine inequality of the high-order A and B octades -- it determines the logical equivalence of the three pairs of bits in the octades and ORs those equivalences together. Logical equivalence is the complement of exclusive-OR. In other words, W03L/ is true if any of the three pairs of bits match. Equivalently, W03L is true if none of the three pairs of bits match. After making that change in the code, results with zero mantissas finally generated words of all zeroes.

After all the hard work to re-implement single-precision Add/Subtract from The Flows, it produced very little improvement over the results of the previous, empirically-corrected implementation. There were a few differences in the way that results are normalized, and a rounding difference if one of the values is octal 0161000000000000 and the other value is one of 0004444444444444, 0005555555555555, 0006666666666666, or 0007777777777777. On that basis, the effort invested in the new implementation was not worth the very slight improvement in results. On the other hand, it was a very interesting experience, and I learned a lot about digital logic and how the B5500 worked at the electronic level.

There is one other significant difference -- the new implementation runs noticeably slower, in terms of emulated clock time, than the prior one. For one heavily numeric benchmark that runs about ten minutes, the new implementation increases the emulated run time by 48 seconds -- a little over 8%.

Card Reader Enhancement


The implementation of the B129 card reader in the emulator allows you to load one or more text files into the reader as if they are card decks. Each line in a text file is treated as an 80-column card image. Thus, in order to read cards, you first must have a text file on your local system to hold the necessary card images.

This has proven to be a little inconvenient. If you have a file of source code you wish to compile, for example, you either need to modify that file and add the necessary B5500 control cards, or create a copy of the file and add the control cards to that.

Starting with this release, the user interface for the card reader now supports the ability to enter card images directly into the reader from your keyboard without creating a text file first. Card images entered this way can be interspersed with card images from text files. This allows you, among other things, to "wrap" existing files with control cards, compiler $-option cards, or anything else you need but don't have (or don't want) in a text file on disk.


The card reader's panel has a new button in the upper right-hand corner labeled KEY IN DECK. This button is disabled unless the reader is in a not-ready state. Clicking the button opens a sub-window that allows you to enter card images directly into the reader's buffer:


While the key-in window is open, the reader's START button is disabled.

The bulk of this window is occupied by a standard text area. You can enter text into this area from your keyboard or copy/paste text from another source. You can edit the text using the normal controls of your operating system. Clicking the Insert button at the bottom of the window appends the contents of the text area to the card reader's buffer and closes the window. Clicking the Cancel button simply closes the window without affecting the reader's buffer. Once text is inserted into the buffer, it cannot be modified or deleted, just as with inserting a text file into the buffer.

There are three white buttons along the bottom of the window with commonly-used control card images. Clicking one of these simply inserts the respective card image into the text area at the current cursor position. From there it can be appended to the reader's buffer by clicking the Insert button.

Miscellaneous Improvements


1. The size of the terminal adapter buffer in B5500DatacomUnit has been reduced from 112 to 56 characters to accommodate the requirements of the R/C interactive editor. Users of the Timesharing MCP (TSMCP) and CANDE will need to update their SYSTEM/DISK file with the new buffer size, e.g.,
?RUN SYSDISK/MAKER
?FILE LINE = LINE BACKUP DISK
?DATA CARD
LINE,0,0,56,0,0,7,0,
LINE,1,0,56,0,0,0,0,
STA,0,0,0,0,"0","0",0,0,
?END
2. We have discovered that the emulator now works with Apple Macintosh OS X using Safari versions 8 and above. Nothing has changed in the emulator to support this -- the issue was waiting for Safari to implement the IndexedDB API used by the emulated disk subsystem.

We continue to be unable to get the emulator to run in either Microsoft Windows Explorer or Edge. If anyone has tried the emulator in Opera or any other GUI browser, please let us know how that works. 

3. Despite a statement in the project wiki that the emulator must be loaded over HTTP from a web server, it appears to run just fine when loaded directly from your local file system, e.g., using a URL like this:
file:///C:/Projects/B5500/WIP-S1/webUI/B5500Console.html
The wiki has been corrected.

4. The alternate B5500SyllableDebugger user interface for the emulator has been enhanced with an Execute Single button. This executes the instruction currently in the T register, but differs from the Step button in that it does not advance the C and L instruction counter registers. This makes Execute Single a better choice than Step for testing and debugging a specific operator.

5. The default height of the SPO and line printer windows has been reduced to about a third of the screen height.

6. The tools/B5500DeleteStorageDB.html script can be used to remove the IndexedDB database for an emulated disk subsystem. Previously, this script could only remove the default data base, named B5500DiskUnit. The script now takes a "?db=" query string parameter with the name of the data base to remove.

7. The Algol Glyphs option for the card punch (CPA) has been set to false by default. Some text editors were corrupting the Unicode characters produced in Cold- and Cool-start decks when this option was set to true, making the decks useless for loading under CARD LOAD SELECT.

8. The exponential moving average computations for Delay Delta and Processor Slack in the Processor object have been adjusted slightly.

Corrections


1. Rounding in Integer Store operators has been corrected to produce a word of all zeroes if the mantissa is zero. Previously, non-zero sign and exponent fields could have been present when mantissa was scaled to zero.

2. BCL translation for alpha-mode (even parity) tape files has been corrected. Previously, the translation was being done twice -- once by the I/O Control Unit and once again by the tape drive. It is now handled entirely within the tape drive.

3. Both EOF and parity error are now reported by the tape units for any attempt to read beyond the end of a tape image.

4. The way that "green-bar" formatting was being done on the line printer was causing a blank line to be inserted every third line when copying text directly from the "paper" area of the printer's window. A different method is now used that does not result in extraneous lines in the copied text.

5. When the emulator was powered on, the browser windows for the individual peripheral devices were being created twice. This behavior was intentional, but dated from very early in the emulator's development when crashes and aborts were a frequent occurrence. The act of opening a window, closing it, and then reopening it served to close any windows left open by a crashed emulator instance. We are long past the need for that now, so the screwy open/close behavior has been removed. As part of this effort, de-registration of event handlers for the various windows has been implemented.

6. After almost five years, we are still finding errors in the transcription of the Mark XVI Algol compiler listing. Most of the errors that have come to light in the past few years have been in comments, but a few months ago we found a serious error in line 06073850 -- a plus sign had been transcribed as a minus. I doubt we are ever going to get all of the errors out of this transcription -- or of any other large transcription like this.


The B5500 Source Code Archive


We are starting to amass a significant amount of source code for the B5500. Most of this has been recovered by transcribing it from listings. Up to this point, we have been including the transcriptions we've done or received from others in the emulator project's repository. Rich Cornwell in North Carolina (US) and Jim Fehlinger in New Jersey (US) have been especially prolific in turning scanned listings into text files.

There are now multiple B5500 emulators, however, either in existence or under development. In addition to Sid McHarg's C++ emulator and this project, Rich Cornwell completed an emulator earlier this year that runs in the SimH framework, and Mark Lloyd in the UK continues to work on his Lazarus/Pascal version.

With multiple emulator projects and multiple sources of recovered source code, it has becoming less and less appropriate for that code to be embedded in a specific project. I have been thinking about this for a while, and a few months ago decided to set up a separate project on GitHub to serve as a common repository for all B5500 emulators:
https://github.com/retro-software/B5500-software
Rather than set this up under someone's personal account, I've created a GitHub "organization" to be the owner. This will make it easier to share and ultimately transfer responsibility for the archive, and also to create additional repositories in the future to maintain recovered source code for other systems. Initially, Nigel and I are the administrators of this organization and the B5500 archive, but hopefully that can be expanded in the future.

To seed this new repository, I have exported the source code transcriptions, and their history, from the retro-b5500 project and made that the initial contribution to the B5500-software project. In the process, I rearranged the directory structure into something that hopefully will be more maintainable over time. The source code in the retro-b5500 project will probably be deleted at some point down the road.

In a future blog post, I will propose some standards and procedures for maintaining the archive, but will be happy to receive comments and recommendations on this from others. I will also be happy to receive new contributions and will consider pull requests for updates. Each transcription should be accompanied by some text that describes it provenance and who did the transcription. See the README.txt files in the repository for examples of what I'd like to have in this regard.

I would like for all of the transcriptions to be faithful to their original sources, so please include sequence numbers if they are present in the original, and please do not correct spelling in comments, or even bugs in the code. Any original errors can be corrected in separate projects or branches from the transcriptions of the originals.