path: root/README

$Id: README,v 1.2 2006-01-13 04:54:08 matju Exp $

PureUnity

Copyright 2006 by Mathieu Bouchard <matju à artengine point ca>

This program is free software; you can redistribute it and/or
modify it under the terms of the GNU General Public License
as published by the Free Software Foundation; either version 2
of the License, or (at your option) any later version.

See file ./COPYING for further informations on licensing terms.

This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
GNU General Public License for more details.

You should have received a copy of the GNU General Public License
along with this program; if not, write to the Free Software
Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.

+-+-+--+---+-----+--------+-------------+---------------------+
GOALS

  1. To provide a unit-test framework, which also provide benchmarking
     features, all made in Pd for use in Pd.

  2. To provide tests for functionality in internals, externals, abstractions, 
     etc., in a modularized way, in a DRY/OAOO fashion, thus abstracting out
     common features so that many objects share the same test patch for the 
     features that they have in common.

+-+-+--+---+-----+--------+-------------+---------------------+
REQUIREMENTS

  1. Pd 0.39 (PureMSP or Devel)

+-+-+--+---+-----+--------+-------------+---------------------+
TEST PROTOCOL

  new:
    create common (reusable) fixtures.

  inlet 0:
    bang:
      run all available tests in that class. individual tests don't have
      to be available through individual methods but may. If they do, the
      names of the methods must match those given in the test results. 

      each test should build its own non-reusable fixtures and reinitialize
      common fixtures, not assuming that the previous tests have left the
      common fixtures in a normal state.

  outlet 0:
    test results. a sequence of lists like:
      list $passed? $accuracy $elapsed $name1 ...

    where:
      $passed? is either 0 for failure or 1 for success
      $accuracy is a float proportional to relative error on math
         (if not applicable, use 0)
      $elapsed is a nonnegative float, the time elapsed in milliseconds
         or it is any negative float meaning the time hasn't been measured.
      $name1 and the rest are symbols and/or floats identifying the test

    for example:
      list 1 0 -1 commutative f + *

    Which means that the 1st test about commutativity passed ($2=1) because it
    was perfectly accurate ($3==0) and that we didn't measure the time ($4=-).

+-+-+--+---+-----+--------+-------------+---------------------+
SEVERITIES (in decreasing order)

 * crash: Segmentation Fault, Bus Error, Illegal Instruction, Infinite Loop, 
   etc.  You can't deal with those errors at the level of the tests. Maybe there 
   should be a way to tell a test object to skip certain tests, by name, in 
   order to be able to perform as many tests as possible while waiting for a 
   fix. It could become possible to rescue from some of those crashes if Pd
   supported exceptions (stack-unwinding).

 * corruption: this may cause future crashes and failures on innocent 
   objects/features. I have no solution for this except to be careful.

 * post(),error(),pd_error(): Gets printed in the console. The problem is that 
   those can't be handled by the test objects, so someone has to read them and 
   interpret them. Also they prevent test objects to ensure that error 
   conditions produce error messages.

 * pd_error2(): I wish this would exist. It would be sort of like pd_error() 
   but it would produce a pd message instead, whose selector would be an 
   error code, designed to be both localizable and [route]able. By default, that 
   message would be sent to the console, but there would be an internal class
   designed to catch those messages. (If stack-unwinding were possible, it would 
   be disabled by default on pd_error2 and could be enabled explicitly 
   by-selector).

 * failure: a test object reports a problem through outlet 0.

 * dropout: a failure in realtimeness... difficult for an object to detect.

 * inaccuracy: a test more or less succeeds but the test detected that the 
   epsilon sucks.

+-+-+--+---+-----+--------+-------------+---------------------+
PROTOCOL FOR [error]

new:
  optional argument which would either be a float
  (e.g. the $0 of the enclosing abstraction) or a pointer.

inlet 0:
  set $scapegoat:
    replaces the originator of the message by $scapegoat, which can be a
    float or a pointer

  error $1 ...:
    causes its arguments to be concatenated, space-separated (may include
    floats), and then sent through pd_error using the appropriate
    originator (scapegoat).

  list $1 ...:
    for future use. would use pd_error2() (see README or previous mail).
    $1 has to be a symbol.

+-+-+--+---+-----+--------+-------------+---------------------+
ACCURACY AND ERROR (in math-related unit tests)

The "absolute error" between a practical result and the expected value
is considered to be the distance between the two value. That is the
absolute value of the difference.

In the case of positions in 2D, 3D, etc., use the L2-Norm which is
a generalized Pythagoras' Theorem: dist^2 = x^2 + y^2 + z^2 + ...
A norm is a distance between something and zero.

Sometimes you have several practical results for one expected value
and must extract a single absolute error out of that. Then you should pick
the largest of the individual absolute errors.

Sometimes you don't have an expected value, you just have several
practical results that you expect to be quite the same. In that case,
the absolute error is the "diameter" of those results. The meaning
of diameter here is: the largest distance between any two results.

If in a single test you must compare 2D errors with 3D errors and 1D
errors, etc., you may have to adjust them by dividing the error by
the square root of N (N is the number of dimensions). In that case,
the resulting value is called a RMS (Root-Mean-Square).

The maximum error introduced by just representing a number as a float
(instead of an exact value) is at most proportional to the magnitude
of the number (e.g. usually 16 million times smaller: about 6 decimals).
Also, often we are only interested in relative error, which is absolute
error divided by the norm of the expected result, because small absolute
errors don't matter much with large results. This is the reason floats
exist in the first place. By default, use relative error as the $accuracy
in Pd tests.

If you don't have an expected result, then compute the relative error as
being the absolute error divided by the norm of the average of practical
results.

In the RMS case of relative error, the norms of expected results should also
be adjusted, but both adjustments cancel because they get divided by each
other. That means: don't divide by the sqrt(N) at all and you'll get an
appropriate result.

+-+-+--+---+-----+--------+-------------+---------------------+
TYPE PREFIXES

Those have to be prefixes in order to be honored by DOLLSYM:
[$1norm] should expand to [fnorm], [lfnorm], [#norm], etc.

Those prefixes are necessary in order to achieve polymorphism through
abstraction arguments.

CURRENT:
  f   float
  ~   signal

FUTURE (from PureData):
  s   symbol
  p   gpointer
  a   anything
  l   list (of whatever)
  lf  list of floats
  ls  list of symbols
  lp  list of pointers

FUTURE (from DesireData):
  t   stringpointer
  L   listpointer
  v   varpointer (instance symbol)

FUTURE (from GridFlow):
  #   grid (of whatever)
  #b  grid of bytes (uint8)
  #s  grid of shorts (int16)
  #i  grid of ints (int32)
  #l  grid of longs (int64)
  #f  grid of floats (float32)
  #d  grid of doubles (float64)
  #r  grid of rubies (VALUE*)

for a type prefix to be considered implemented, it has to
have the following class set:

  metaabstraction   for floats   for signals   for grids
  [$1.inlet]         [inlet]      [inlet~]      [inlet]
  [$1.outlet]       [outlet]     [outlet~]     [outlet]
  [$1.do $2 $3]     [$2 $3]      [$2~ $3]      [# $2 $3]
  [$1.taa]          [t a a]      noop          [t a a]
  [$1.swap]         [swap]       noop          TODO
  [$1.norm]         [abs]        [env~]        [# sq]->[#ravel]->[#fold +]->[#export]->[sqrt]
  [$1.packunpack3]  pack,unpack  noop          TODO

The first two cannot be implemented as abstractions and instead must be
defined as aliases in pureunity.c.

extra metaabstractions:
  [$1.rand]         [f.rand]     [~.rand]TODO  [#.rand]TODO

+-+-+--+---+-----+--------+-------------+---------------------+
OTHER PROTOCOLS

Those four classes are operators that give verify algebraic properties
of other operators. The more their outputs are close to zero, the more
those other operators are faithful to an algebraic property.

(here, supported $types are f and ~)

[commutator  $type $class]          (2 inlets)     ab-ba
[associator  $type $class]          (2 inlets)  (ab)c-a(bc)
[distributor $type $class1 $class2] (3 inlets) a&(b^c)-(a&b^a&c)
[invertor    $type $class1 $class2] (2 inlets)   ab/b-a

+-+-+--+---+-----+--------+-------------+---------------------+
TESTS AND RULES

For each class, a test file's name is the class name followed by "-test.pd",
and a rule file's name is the class name followed by "-rule.pd",
in the same way as it is for help files.

for a class called $foo, the protocol (aka interface aka rule) $foo is the
set of behaviours expected from the $foo class; the class called $foo-rule
must repect the $foo protocol as well, plus it should test that the inputs
are valid, and if they are, it should test for one or several results and
report any errors.

(((To report errors and inaccuracies, output them through the properties outlet at the right. If there is no
properties outlet in $foo (curently almost nothing in Pd has one),
then $foo-rule must have one more outlet than $foo.)))

(((Float messages coming out of the properties outlet of $foo-rule report
accuracy. Named error messages come out with selector "error" followed by
an error-symbol and then its arguments.)))

(((In the case of true/false logic, a value of 0 means that a test has passed
and a 1 means that a test has failed. Those values represent failure and not
success. The reason is so that it matches with accuracy levels, where 0 is
perfectly accurate, but any inaccuracy shows up as a relative error fraction.
Any finite nonnegative value is allowed for accuracy, because it is expected
to be the result of a norm))).

(((In standard math, "Discrete Metric" is when there are only two possible
distances between objects: together=0 and apart=1)))

+-+-+--+---+-----+--------+-------------+---------------------+
RANDOMIZERS (?)

+-+-+--+---+-----+--------+-------------+---------------------+
ETC


(write me!)




If +-test.pd tests [+], it can test for hotness, coldness, it can test
that only one result is produced per hot message, that all results are
float, that a few example additions work, and that with random inputs it
respects commutativity, associativity, invertibility, within appropriate
relative-error bounds, etc.

However +-test.pd can't test that errormessages aren't printed during the
testing. This may be something that we want to check for, and currently
the best way to handle it is to search the console for error messages, and
if there are any, restart the tests in verbose mode and see where the
error happens exactly.

[...]

Floating-point is the scientific notation for numbers that we all
learned on paper in school. Rounding and inaccuracy are two sides
of the same coin. They are required when it is stupid to have perfect
results, that is, when it would mean too many computations for little
gain.

However sometimes we want to make sure that our math is accurate enough.
Many algorithms are data-recursive: each computation uses previous
results. Many of those algorithms have chaotic and/or unstable
behaviours, which means that the inaccuracies may skyrocket instead of
fading out.

+-+-+--+---+-----+--------+-------------+---------------------+

Date: Fri, 13 Jan 2006 04:07:59 +0900
From: Mathieu Bouchard <matju@artengine.ca>
Reply-To: ruby-core@ruby-lang.org
To: ruby-core@ruby-lang.org
Subject: Re: Design contracts and refactoring (was Re: mathn: ugly warnings)

On Fri, 13 Jan 2006, mathew wrote:

> *Dean Wampler *<deanwampler gmail.com> writes:
> > Let me suggest an XP-style alternative; make thorough unit tests
> > required and make sure they "document" - and test! - the design
> > "contract".
> Unit tests are not an alternative. They are an additional requirement.

I find unit-tests to be often decomposable like this. Start with something
like this:

  raise if Blah.new(666) != Blah.new(666)
  raise if Blah.new(747) != Blah.new(747)
  raise if Blah.new(242) != Blah.new(242)
  raise if Blah.new(69)  != Blah.new(69)
  raise if Blah.new(37)  != Blah.new(37)

then generalize it ("equality is defined based on the arg of .new"):

  for x in [666,747,242,69,37] do
    raise if Blah.new(x) != Blah.new(x)
  end

then extract a contract from it:

  class CheckedBlah < Blah
    def self.new(x)
      r = super(x)
      raise if r != super(x)
      r
    end
  end

so now all Blah object creation may be checked throughout actual uses of a
program and not just unit tests. The unit test now reduces to:

  for x in [666,747,242,69,37] do Blah.new(x) end

so for many unit tests, all you have to do is just do things and discard
the results, and the contract will do the job of checking.

 _ _ __ ___ _____ ________ _____________ _____________________ ...
| Mathieu Bouchard - tél:+1.514.383.3801 - http://artengine.ca/matju
| Freelance Digital Arts Engineer, Montréal QC Canada

+-+-+--+---+-----+--------+-------------+---------------------+

Date: Fri, 13 Jan 2006 05:05:19 +0900
From: Mathieu Bouchard <matju@artengine.ca>
Reply-To: ruby-core@ruby-lang.org
To: ruby-core@ruby-lang.org
Subject: Re: Design contracts and refactoring (was Re: mathn: ugly warnings)

On Fri, 13 Jan 2006, mathew wrote:

> For example, consider a simple vector addition routine in a 3D library.
> The unit tests might test its behavior with Float and Integer vectors,
> since that's why it was written.

Here's another way to factor unit-tests that I haven't mentioned in the
last mail.

suppose you test for + using:

  class IntegerTest
    def test; 2+2==4 or raise; end
  end
  class FloatTest
    def test; 2.0+2.0==4.0 or raise; end
  end
  class RationalTest
    def test; Rational(2,1)+Rational(2,1)==Rational(4,1) or raise; end
  end

you can refactor those tests like this:

  class NumericTest
    def initialize nt; @nt; end
    def make x; raise "abstract class" end
    def test; make(2)+make(2)==make(4) or raise; end
  end
  class  IntegerTest; def make x;  Integer(x)   end end
  class    FloatTest; def make x;    Float(x)   end end
  class RationalTest; def make x; Rational(x,1) end end

> However, to do that you need to know whether the feature of supporting
> (say) Complex vectors or BigDecimal vectors is intended or not. The unit
> tests won't tell you this.

I once called unit-tests "test-by-example" and contracts "test-by-rule". I
think that those names are preferable to the more common names. I also had
listed "test-by-use" which is to use the software in practice: this may
include testing a component A using the unit tests for B because B uses A.
The last I had listed was "test-by-proof", which is rarer and is the only
one that requires analysing the implementation.

> > One limitation of documentation is that it has no enforcement power,
> > so you have to write tests anyway to test conformance.
> Unit tests have no enforcement power either, because you can just change the
> test. Indeed, I've already had to do this once when it turned out that the
> unit test was wrong. (In net/ftp.)

That was a pretty bad case of strawman argument. Dean was assuming that
your documentation was not executable when you had quite clearly stated
that it was the contracts that acted as documentation!

I've thought of a triad:

  A. "the real thing"
  B. what it's documented as
  C. a way to verify that (1) and (2) agree

and another one:

  changing A to match B+C:
    programming

  changing B to match A+C:
    the scientific method (aka reverse engineering)

  changing C to match A+B:
    unit-tests and contracts and scientific experiments

 _ _ __ ___ _____ ________ _____________ _____________________ ...
| Mathieu Bouchard - tél:+1.514.383.3801 - http://artengine.ca/matju
| Freelance Digital Arts Engineer, Montréal QC Canada

+-+-+--+---+-----+--------+-------------+---------------------+

Date: Fri, 13 Jan 2006 07:36:36 +0900
From: Mathieu Bouchard <matju@artengine.ca>
Reply-To: ruby-core@ruby-lang.org
To: ruby-core@ruby-lang.org
Subject: Re: Design contracts and refactoring (was Re: mathn: ugly warnings)

On Fri, 13 Jan 2006, mathew wrote:

> >  The XP view is
> > that you should eliminate the redundancy.
> Except it's not redundancy.
> Unit tests define a set of functionality that is required. Documentation tells
> you the functionality that is supported, which is generally a superset of the
> functionality required by the unit tests.

Let's follow the argument of both of you to the end.

1. Unit-tests often match inputs with outputs on a case-by-case basis.

2. Redundancy should be eliminated.

(1) suggests that there is a shorter way to express the unit-tests.
Suppose you are able to find a formula for generating output-validators
from inputs. Then that formula is a postcondition of a contract, and the
explicit output-validators of the unit-tests are redundant.

(2) because part of the unit-tests are redundant, part of the unit-tests
should be eliminated. This causes the postconditions to become an
essential part of unit-testing.

Unit-tests vs contracts is a false debate.

 _ _ __ ___ _____ ________ _____________ _____________________ ...
| Mathieu Bouchard - tél:+1.514.383.3801 - http://artengine.ca/matju
| Freelance Digital Arts Engineer, Montréal QC Canada


+-+-+--+---+-----+--------+-------------+---------------------+