contents
Some nasty problems are recurring time and again. Maybe a trick could be found somewhere in the net, and a library function was created to settle this damn topic once and for all. Maybe even a nice test and demo is provided. And then the whole story will be forgotten.
Sounds familiar? ⇒ ☹☻☺👻 … then please consider to leave some traces here …
some basic descriptive statistics computations are defined in lib/stat/statistic.hpp
the simple case for linear regression is also implemented there
Gnuplot provides also common statistics functions, which may come in handy when the goal is anyway to create a visualisation (→ see below)
Reformulate the research and the findings into a test, which can be read top-down like a novel. Start with documenting the basics, package helpers into a tool class, or package setup into a workbench-style class, with individual tool modules. Provide a short version of this test with the basic demo, which should be able to run with the regular test suite. Extended long-running tests can be started conditionally with commandline-arguments. See scheduler-stress-test.cpp
Use code generation to visualise data structures or functions and observation statistics. Typically these generation statements can be packaged into an invocation helper and included into a relevant test, but commented-out there.
generate Graphviz diagrams: lib/dot-gen.hpp provides a simple DSL. See test-chain-load-test.cpp
generate Gnuplot scripts: use the Text-Template engine to fill in data, possibly from a generated data table
lib/gnuplot-gen.hpp provides some pre-canned scripts for statistics plots
used by the stress-test-rig.cpp, in combination with data.hpp to collect measurement results
build a diagnostic output which shows the nesting
use configuration by template parameters and use simple numbers or strings as part components
render the output and compare it with ""_expect (the ExpectStr → see lib/test/test-helper.hpp)
break the expect-string into parts with indentation, by exploiting the C string gaps
However, this works only up to a certain degree of complexity.
A completely different approach is to transform the structured result-data into an ETD (GenNode tree)
and then directly compare it to an ETD that is created in the test fixture with the DSL notation (MakeRec())
either use approximate comparison
almostEqual() → see lib/test/test-helper.hpp
util-quant.hpp has also an almostEqual(a,b, ulp)
TODO 2024 should be sorted out → #1360
or render the floating point with the diagnostic-output functions, which deliberately employ a built-in rounding to some sensible amount of places (which in most cases helps to weed-out the “number dust”)
simply match it against an ExpectStr — which implicitly converts to string, thereby
also applying the aforementioned implicit rounding
→ See also about formatting below; in a nutshell, #include 'lib/format-output.hpp'
A first attempt should be made to bring them into some regular band. This can be achieved by
automatically calibrating the measurement function (e.g. do a timing calibration beforehand).
Then the actual value can be matched by the isLimited(l, val, u) notation (see lib/uitl.hpp)
Control the seed! Invoke the seedRand() function once in each test instance.
This draws an actually random number as seed and re-seeds the lib::defaultGen. The seed is written to the log.
But the seed can be set to a fixed value with the --seed parameter of the test runner.
This is how you reproduce a broken test case.
Moreover, be sure either to draw all random values from the defaultGen or to build a well organised
tree of PRNG instances, which seed from each other. This is especially valuable when the test starts
several threads; each thread should use its own generator then (yet this can be handled sloppy if
the quality of the random number actually does not matter!)
It can be very helpful to use policy based design for some complex evaluation or algorithm, even in cases where it is clear that the actual usage will only ever involve one specific data type and configuration. Not only does such a structured approach help with clarifying the design itself, since the core of the evaluation logic can be abstracted from data representation — yet also the testing and verification of correctness is simplified, as the test can be written to use simple, human-readable data (strings, symbols), and, furthermore, many kinds of instrumentation can be included into the test data to gain detailed insight what is happening in the test subject.
Use the lib::EventLog:
within the test, notable points are marked into the log, possibly with arguments
after the test, the Event Log can be verified with a search-sequence
this way it is possible to ensure that some events did happen, and happened in a very specific order, or that some event did not happen after another event
the test::Tracker defined in src/test/tracking-dummy.hpp will mark all constructor,
destructor and copy events into an embedded Event Log instance. By inheriting or mix-in
from that type, the allocation and instance behaviour of some service can be scrutinised.
Aside from using some generic tracing utility, sometimes it is much easier to setup a specifically tailored instrumentation. This can be achieved with incidence-count.hpp
for instrumentation, the markEnter() and markLeave() functions need to be invoked
after the test run, various statistics can be computed and evaluated, including cumulated time and concurrency for the "cases" marked with enter / leave.
use the lib::stat::DataTable (data.hpp) with CSV rendering → see data-csv-test.cpp
represent it as Lumiera ETD, that is as a tree of GenNode elements.
be sure to understand the essential idea: the receiver should act based on knowledge about the structure — not by introspection and case-switching
however — for the exceptional case that you absolutely must discover the structure, then use the visitor feature. This has the benefit of concentrating the logic at one place.
can represent the notion of a nested scope, with iteration
provides a convenient DSL notation, especially for testing, which helps to explain the expected structure visually .
also useful as output format, both for debugging and because it can be matched against an expected structure, which is generated with the DSL notation.
(planned) will be able to map this from/to JSON easily.
unique-ownership collection that holds a sequence of objects in a preallocated fixed
storage block. Notably even non-copyable objects can be managed (which is not possible
with a std::vector), and type-erased instances of subclasses can be mixed. Furthermore
there is a constructor variant that allows to populate a ScopedCollection in RAII-style.
This can e.g. be used to fire-off a set of Threads or open a set of wrapped connections.
They are designed for convenient usage with low boilerplate (even while this means wasting some CPU cycles or memory). They are deliberately much simpler than STL iterators, can be iterated only once, can be bool checked for iteration end, and can be used both in a for-each construct and in while-loops.
The function lib::explore(IT) builds on top of these features and is meant to basically
iterate anything that is iterable — this can be used to level and abstract away the details.
can be filtered and transformed
can be reduced or collected into a vector
can be used to build complex layered search- and evaluation schemes
A set of convenience functions like eachElm(CON) — available in iter-adapter-stl.hpp.
Also allows to iterate each key or value, and to take snapshots
This is a special Lumiera iterator implementation which delegates through a virtual (OO) interface. Thus the source of the data can be abstracted away and switched transparently at runtime. Especially relevant for APIs, to produce a data sequence once and without coupling to details; even the production-state can be hooked up into the IterSource instance (with a smart-ptr). This allows e.g. to send a Diff to the UI through the UI-Bus, while the actual generation and application both happen demand-driven or lazy…
When some algorithm involves a statefull evaluation, it might be helpful to package that into
a State Core container. This is a class that is expected to expose three operations, which
allow to build an iteration on top. The header iter-adapter.hpp provides several adapters
and wrappers, notable lib::IterStateWrapper wraps any state core and provides a
Lumiera Iterator interface; likewise, IterExplorer heavily relies on that concept and
is able to accept and iterate any state core in a pipeline.
Some additional adaptors are provided as building block so that specific evaluations can be made to conform with the »Lumiera Iterator« concept
several given iterators can be combined into a iteration over tuples with lib/iter-zip.hpp
there is also a variation lib::izip that adds a running index into the zipping operation
IterExplorer provides a flatten() operation, that can be used to join an iterator-of-iterators
into a single seamless iteration lazily.
when a STL container also supports backwards iteration, then a lumiera iterator can be created that can switch direction dynamically, with the help of iter-cursor.hpp
a STL »deque« container can be used both as a stack and a queue; the header iter-stack.hpp provides an iteration adapter for both usage schemes, so that either the stack or the queue can be "discharged" as a Lumiera Iterator (possibly with an IterExplorer pipeline)
implement a conversion-to-string operator.
include the C++ IOStreams via lib/format-cout.hpp → this magically uses the util::toString()
for testing, temporarily include lib/test/diagnostic-output and use the SHOW_EXPR macro.
use util::join (lib/format-util.hpp) to join arbitrary elements with util::toString() conversion
use printf-style formatters from Boost-format. We provide a light-weight front-end via lib/format-string.hpp
the heavyweight boost-format include is only required once, for lib/format-string.cpp.
the templated front-end passes-through most basic types and types with string-conversion
all invocations are strictly error safe (never throw) and can thus be used from catch-handlers
use the Text-Template engine. See text-template-test.cpp.
Can be used with simple map bindings, or even a definition string "x=42, y=why-not?", but can also
draw data o from a lib::GenNode tree, or even from a custom defined DataSource template.
Supports placeholders, conditionals and simple loops (and that’s it — because there are way more capable solutions out there ☺)
Generally speaking, we prefer constructor wiring whenever possible, to avoid creating unnecessary statefulness (as would be the case when connecting dependencies through setters). This design also favours decomposition into self-contained components with a single purpose, and simplifies error handling with the help of the RAII pattern.
However, sometimes a piece of code depends on some remote and even global service; in these
cases, the lib::Depend framework can be used to access that service via type name. The
actual service can either be created on-demand as a singleton, or it can be installed
explicitly as a service instance. Furthermore, it is possible to "push aside" the
current service instance temporarily and replace it by a test mock.
One point to note is that lib::Depend is a monostate, and not a singleton.
The actual service should be written as a free-standing class, that can be instantiated
directly for the purpose of testing.
use a templated constructor, possibly even with varargs
use a deduction guide to pick the right ctor and arguments → see
ThreadJoinable in thread.hpp, 698
DataSource<string> specialisation only when argument can be converted to string,
in text-template.hpp, 745
prevent shadowing of automatically generated copy operations.
See #963. Based on the “disable if” SFINAE technique.
A ready-made templated typedef lib::meta::disable_if_self can be found in lib/meta/util.hpp
The key trick is to define an index sequence template, which can then be matched against a processing template for a single argument; and the latter can have partial specialisations
see variadic-argument-picker-test.cpp
but sometimes it is easier to use the tried and true technique of the Loki-Typelists, which
can be programmed recursively, similar to LISP. The »bridge« is to unpack the variadic argument pack
into the lib::meta::Types<ARGS...>
When you need to rebind and manipulate a variadic sequence, it helps to transform the sequence
into one of our meta sequence representations (lib::meta::Types or the Loki typelists).
In variadic-helper.hpp, we define a convenient rebinding template lib::meta::Elms<>,
which can work transparently on any type sequence or any tuple-like entity
to get at the variadics in a sequence representation
to get a matching index sequence
to rebind into another variadic template, using the same variadic sequence
to apply a meta-function on each of the variadic types
to compute a conjunction or disjunction of meta-predicates over the sequence
This is a concept to match on any type in compliance with the »tuple protocol«
such a type must inject a specialisation of std::tuple_size<TY>
and it must likewise support std::tuple_element_t<N, TY>
and, based on that, expose a constexpr getter function
together this also implies, that such a type can be used in structured bindings (but note, structured bindings also work on plain arrays and on simple structs, which are not considered tuple-like by themselves)
Unfortunately, the standard has a glaring deficiency here, insofar it defines an exposition only
concept, which is then hard mapped to support only some fixed types from the STL (tuples, pairs,
std::array and some range stuff). This was one of the main reasons to define our own concept tuple_like
but unfortunately, std::apply was fixed with C++20 to allow only the above mentioned fixed set of
types from the STL, while in theory there is no reason why not to allow any tuple-like entity
this forces us to define our own lib::meta::apply as a drop-in replacement for std::apply
in addition to that, we also define a lib::meta::getElm<N, TY>, which is an universal getter
to work on any tuple-like, either with a get member function or a free-ADL function get
note that starting with C++20 it is forbidden to inject function overloads into namespace std,
and thus a free get function must be injected as friend via ADL and used appropriately, i.e.
unqualified.
A common trick is to use apply in combination with a fold-expression
provided as lib::meta::forEach in 'lib/meta/tuple-helper.hpp
The design of the DataTable with CSV-Formatting is based on this technique, see lib/stat/data.hpp
lib/iter-zip.hpp uses this to construct a tuple-of-iterators
Under controlled conditions this is possible (even while it seems like time travel from the runtime into the compile-time domain). The number of results to extract from the iterator must be known at compile time, and the possible result types must be limited, so that a visitor can be used for double-dispatch.
see tuple-record-init-test.cpp
used in command-simple-closure.hpp to receive parameter values sent via UI-Bus and package them into a tuple for invocation of a Steam-Layer command.
This is another example of bridging the compile-time / runtime barrier, using a trampoline. Obviously such a solution requires that it is somehow possible to define a generic functor / λ that can accept any of the tuple elements. And, in order to build a runtime index access, we additionally need to convert each of the tuple elements into a common reconciled type.
If such a type exists, can be figured out with the help of the meta::CommonReturn metafunction
lib::meta::TupleIdxAdaptor provides the necessary machinery to define a subscript operator[]
see TupleIdxAdaptor_test for some usage examples, including a runtime iteration…