While in principle it would be possible (and desirable)
to control worker behaviour exclusively through the Work-Functor's return code,
in practice we must concede that Exceptions can always happen from situations
beyond our control. And while it is necessary for the WorkForce-dtor to
join and block (we can not just pull away the resources from running threads),
the same destructor (when called out of order) must somehow be able
at least to ask the running threads to terminate.
Especially for unit tests this becomes an obnoxious problem -- otherwise
each test failure would cause the test runner to hang.
Thus adding an emergency halt, and also improve setup for tests
with a convenience function to inject a work-function-λ
No new functionality, and implementation works as expected.
This test case covers an especially tricky setup, where a calculation
shall be triggered from an external event, while ensuring that the actual
processing can start only after also the regular time-bound scheduling
has taken place (this might be used to prevent an unexpectedly early
external signal to cause writing into an output buffer before the
defined window of data delivery)
...based on the new ability in the ActivityDetector, we can now assign
a custom λ, which deflects back the ctx.post() call into the ActivityLang
instance used for this test case.
While the previously seen behaviour was correct, it was not the call sequence
expected in the real implementation; with this change, on the main-chain
activation the post() now immediately dispatches the notification, which in turn
dispatches the rest of the chain, so that the JobFunctor is indeed
called in this second test case as expected
Up to now, the DiagnosticFun mock in ActivityDetector only
created an EventLog entry on invocation and was able to retunr
a canned result value. Yet for the job invocation scenario test,
it would be desirable to hook-in a λ with a fake implementation
into the ExecutionContext. As a further convenience, the
return value is now default initialised, instead of being
marked as uninitialised until invocation of "returning(val)"
...seems to work, but not really happy with the test setup,
since in real usage the post()-calls would dispatch, while here,
using the ActivityDetector, these calls just log invoation,
and thus the activation is not passed on
...regarding the kind of activity (the verb),
and also for some special case access of payload data;
deliberately asserting the correct verb, but no mandatory check,
since this whole Activity-Language is conceived as cohesive
and essentially sealed (not meant to be extended)
...to show in test that indeed the actual time is retrieved on each activation,
we can assign a λ -- which is rigged to increase the time on each access
It is not sufficient just to pass this "current time" as parameter
into the ActivityLang::dispatchChain(), since some Activities within
this chain will essentially be long-running (think rendering); thus
we need a real callback from within the chain. The obvious solution
is to make this part of the Execution Context, which is an abstraction
of the scheduler environment anyway
...turns out there is still a lot of leeway in the possible implementation,
and seemingly it is too early to decide which case to consider the default.
Thus I'll proceed with the drafted preliminary solution...
- on primary-chain, an inhibited Gate dispatches itself into future for re-check
- on Notification, activation happens if and only if this very notification opens the Gate
- provide a specifically wired requireDirectActivation() to allow enforcing a minimal start time
...assembled from parts already implemented
TODO
- need a way to access the »current scheduler time«
- need builder extension points to connect notifications
...this completes the basic setup
- Term builder mechanism working properly
- Memory allocator behaves sane
- the simple default wiring allows to invoke a Job
Solved by special treatment of a notification, which happens
to decrement the latch to zero: in this case, the chain is
dispatched, but also the Gate is locked permanently to block
any further activations scheduled or forwareded otherwise
TODO: while correct as implemented, the handling of the
notification seems questionable, since re-scheduling the chain immediately
may lead to multiple invocations of the chain, since it might have been "spinned"
and thus re-scheduled already, and we have no way to find out about that
...can not take a shortcut here, since the timing information
embedded into the POST-Activity must somehow be transported
to the Scheduler; key point to note is that the chain will
be performed in »management mode« (single threaded)
...attempt to get this intricate state machine sorted out
Notification turned out quite tricky, since it may emanate
from a concurrently executed phase and we try to avoid having
to protect the gate directly with a lock; rather we re-dispatch
the notification through the queue, which indirectly also ensures
that the worker de-queuing the NOTIFY-Activity operates in
management mode (single threaded, holding the GroomingToken)
Decision how to handle a failed Gate-check
- spin forward (re-scheduler) by some time amount
- this spin-offset parameter is retrieved from the Execution Context
- thus it will be some kind of engine parameter
With these determinations and the framework for the Execution Context
it is now possible to code up the logic for Gate check, which in turn
can then be verified by the watchGate diagnostics
due to technical limitations this requires to wire the adaptor
as replacement for the subject Activity, so that it can capture
and log the activation, and then pass it on to its watched subject
requires to supplement EventLog matching primitives
to pick and verify a specific positional argument.
Moreover, it is more or less arbitrary which job invocation parameters
are unpacked and exposed for verification; we'll have to see what is
actually required for writing tests...
doing so would contradict the fundamental architecture,
all kinds of failures and timeouts need to be handled within
Scheduler-Layer-2 rather.
Jobs are never aborted, nor do they need to know if and when they are invoked
Testcase (detect function invocation) passes now as expected
Some Library / Framework changes
- rename event-log-test.cpp
- allow the ExpectString also to work with concatenated expectation strings
Remark: there was a warning in the comment in event-log.hpp,
pointing out that negative assertions are shallow.
However, after the rework in 9/2018 (commit: d923138d1)
...this should no longer be true, since we perform proper backtracking,
leading to an exhaustive search.
ActivityMatch inherits privately from the EventMatch object,
and is thus able to delegate relevant matching queries, but
also to provide high-level special matchers.
This new design resolves the ambiguity regarding function arguments.
Moreover, we can now record the current sequence-Number as *attribute*
in the respective log record (this is the benefit of using structured
log entries instead of just a textual log), thereby avoiding the various
pitfalls with explicit bracketing sequence-number log entries
bottom line: this reworked design seems to be a better fit,
even while technically the implementation with the wrapped matcher
is somewhat ugly...
The EventLog seems to provide all the building blocks, but we need
some higher level special matchers (and maybe we also want to hide
some of the basic EventLog matchers). A soulution might be to wrap
the EventMatcher and delegate all follow-up builder calls.
This seems adequate, since the EventLog-Matcher is basically used as black box,
building up more elaborate matchers from the provided basic matchers...
Spent some time again to understand how EventLog matching works.
My feelings towards this piece of code are always the same: it is
somewhat too "tricky", but I am not aware of any other technique
to get this degree of elaborate chained matching on structured records,
short of building a dedicated matching engine from scratch.
The other alternative would be to use a flat textual log (instead of
the structured log records from EventLog), but then we'd have to
generate quite intricate regular expressions from the builder,
and I'm really doubtful it would be easier and clearer....
...turns out this is entirely generic and not tied to the context
within ActivityDetector, where it was first introduced to build a
mock functor to log all invocations.
Basically this meta-function generates a new instantiation of the
template X, using the variadic argument pack from template U<ARGS...>
...for coverage of the Activity-Language,
various invocations of unspecific functions must be verified,
with the additional twist that the implementation avoids indirections
and is thus hard to rig for tests.
Solution-Idea: provide a λ-mock to log any invocation into the
Event-Log helper, which was created some years ago to trace GUI communication...
Further extensive testing with parameter variations,
using the test setup in `BlockFlow_test::storageFlow()`
- Tweaks to improve convergence under extreme overload;
sudden load peaks are now accomodated typically < 5 sec
- Make the test definition parametric, to simplify variations
- Extract the generic microbenchmark helper function
- Documentation
There seems to be a ''sweet spot'' for somewhat larger Epoch sizes around 500 slots.
At least in the test setup used here, which works with a load of 200 Frames / sec,
which is significantly over the typical value of 50fps (video + audio) for simple playback.
The optimisation of averaged allocation times can not be much improved **below 30ns**.
Overall, this can be considered a good result,
since this allocation scheme does way more than just allocate memory,
it also provides a means to track dependencies and lifecycle.
__For context__:
- we should strive at processing one frame in ~ 10ms
- for 10 Activity records per Frame, we currently use < 0.5 µs for
memory and dependency management in the scheduler
- this leaves enough room for the further administrative efforts
(priority queue, job planning, buffer management)
BUT -> +50% runtime in -O3 (+20ns)
Investigation seems to indicate
- that the increased (+1 Epochs, 10 -> 11) moving average
caused the Algo to perform worse (strong effect)
- that the Optimiser has problems with boost::rational, which however
yields only a minute effect (+5ns), and only on the critical path
The access via Meyers Singleton has no adverse effect,
rather the new setup gives a tiny benefit (46ns -> 37ns).
Surprisingly, the increased pre-allocation has no observable effect.
On the long run, there will be a central Render Engine parametrisation;
some parameters can even be expected to be dynamic; thus prepare the
BlockFlow allocator to fit in with this expectation
For comparison: use individual managment by refcount.
This supports the conclusion that BlockFlow is more than just a
custom allocator; it also supports a non-trivial lifetime management,
and this comes at a cost.
Playing around with various load patterns uncovers further weak spots
in the regulation mechanism. As a remedy, introduce a stronger feed-back
and especially set the target load factor from 100% -> 90%
to add some headroom to absorb intermittent load peaks
Presumably ''much more observation and fine-tuning'' will be necessary
under real-world load conditions (⟹ Ticket #1318 for later)
- BUG: must prevent the Epoch size to become excessive low
- Problem: feedback signal should not be overly aggressive
Fine-Tuning:
- Dose for Overflow-compensation is delicate
- Moving average and Overflow should be balanced
- ideally the compensatory actions should be one order of magnitude
slower than the characteristic regulation time
Improvement: perform Moving-Average calculations in doubles
...leading to PATHETICALLY bad timing comparison
...it seems clear that the Epoch-Step went to zero
(which was neither anticipated, nor protected against)
However, even individual heap allocations fare surprisingly well
under full optimisation; just they don't solve our problem with
tracking dependencies; the most simplest solution that would
also fulfil this requirement would be using shared_ptr
..as a heuristic to regulate optimal Epoch duration;
when Epochs are discarded, the effective fill factor can be used
to guess an Epoch duration time, which would (in hindsight)
lead to perfect usage of storage space
..using a simplistic implementation for now: scale down the
Epoch-stepping by 0.9 to increase capacity accordingly.
This is done on each separate overflow event, and will be
counterbalanced by the observation of Epoch fill ratio
performed later on clean-up of completed Epochs
further implementation makes clear that the AllocationHandle,
which is the primary usage front-end, has to rely both on
services of the underlying ExtentFamily allocator, as well
as on the BlockFlow itself for managing the Epoch spacing.
- fix a bug in IterExplorer: when iterating a »state core« directly,
the helper CoreYield passed the detected type through ValueTypeBindings.
This is logically wrong, because we never want to pick up some typedefs,
rather we always want to use the type directly returned from CORE::yield()
Here the iterator returns an Epoch&, which itself is again iterable
(it inherits from std::array<Activity, N>). However, it is clear
that we must not descent into such a "flatMap" style recursive expansion
- draft a simple scheme how to regulate Epoch lengths dynamically
- add diagnostics to pinpoint a given Activity and find out into which
Epoch it has been allocated; used to cover the allocator behaviour
- add preliminary deadline-check (directly instead of using the Activity)
- with this shortcut, now able to implement discarding obsoleted Epochs
- Iteration and use of the underlying `ExtentFamily` is also settled by now
💡 ''Implementation concept for the allocation scheme complete and validated''
...with the preceding IterableDecorator refactoring,
the navigation and access to the storage extents can now be
organised into a clear progression
Allocator::iterator -> EpochIter -> Epoch&
Convenience management and support functions can then be
pushed down into Epoch, while iteration control can be done
high-level in BlockFlow, based on the helpers in Epoch
..this is the most simple case, where no Epochs are opened yet
..add diagnostics to inspect alloc count and deadlines
..add accessors for the first/last underlying Extent
...continue to proceed test-driven
...scheduler internals turn out to be intricate and cohesive,
and thus the only hope is to adhere to strict testing discipline
Library: add "obvious" utility to the IterExplorer, allowing to
materialise all contents of the Pipeline into a container
...use this to take a snapshot of all currently active Extent addresses
- use a checksum to prove that ctor / dtor of "content" is not invoked
- let the usage of active extents "wrap around" so that the mem block is re-used
- verify that the same data is still there
The low-level allocator is basically implemented now,
but we still need to check thoroughly that the tricky
wrap-around and expansion logic behaves sane...
(see #1311)
Iteration should just yield an Reference to an Extent,
thereby hiding all details of the actual raw storage (char[]).
This can be achieved by usind a wrapper type around a pointer
into the managing vector; from this pointer we may convert
into a vector::iterator with the trick described here
https://stackoverflow.com/a/37101607/444796
Furthermore, continued planning of the Activity-Language,
basically clarified the complete usage scenario for now;
seems all implementable right away without further difficulties
- the idea is to use slot-0 in each extent for administrative metadata
- to that end, a specialised GATE-Activity is placed into slot-0
- decision to use the next-pointer for managing the next free slot
- thus we need the help of the underlying ExtentFamily for navigating Extents
Decision to refrain from any attempt to "fix" excessive memory usage,
caused by Epochs still blocked by pending IO operations. Rather, we
assume the engine uses sane parametrisation (possibly with dynamic adjustment)
Yet still there will be some safety limit, but when exceeding this limit,
the allocator will just throw, thereby killing the playback/render process
- decision to favour small memory footprint
- rather use several Activity records to express invocation
- design Activity record as »POD with constructor«
- conceptually, Activity is polymorphic, but on implementation
level, this is "folded down" into union-based data storage,
layering accessor functions on top
- decision how to handle the Extent storage (by forced-cast)
- decision to place the administrative record directly into the Extent
TODO not clear yet how to handle the implicit limitation for future deadlines
using a simple yet performant data structure.
Not clear yet if this approach is sustainable
- assuming that no value initialisation happens for POD payload
- performance trade-off growth when in wrapped-state vs using a list
The second design from 2017, based on a pipeline builder,
is now renamed `TreeExplorer` ⟼ `IterExplorer` and uses
the memorable entrance point `lib::explore(<seq>)`
✔
after completing the recent clean-up and refactoring work,
the monad based framework for recursive tree expansion
can be abandoned and retracted.
This approach from functional programming leads to code,
which is ''cool to write'' yet ''hard to understand.''
A second design attempt was based on the pipeline and decorator pattern
and integrates the monadic expansion as a special case, used here to
discover the prerequisites for a render job. This turned out to be
more effective and prolific and became standard for several exploring
and backtracking algorithms in Lumiera.
An extended series of refactoring and partial rewrites resulted
in a new definition of the `Dispatcher` interface and completes
the buildup of a Job-Planning pipeline, including the ability
to discover prerequisites and compute scheduling deadlines.
At this point, I am about to ''switch to the topic'' of the `Scheduler`,
''postponing'' the completion of the `RenderDrive` until the related
questions regarding memory management and Scheduler interface are settled.
- allow to configure the expected job runtime in the test spec
- remove link to EngineConfig and hard-wire the engine latency for now
... extended integration testing reveals two further bugs ;-)
... document deadline calculation
This finishes the last series of refactorings; the basic concept
remains the same, but in the initial version we arranged the expander
function in the pipeline to maintain a Tuple (parent, child) for the
JobTickets. Unfortunately this turned out to be insufficient, since
JobTicket is effectively const and responsible for a complete Sement,
so there is no room to memorise a Deadline for the parent dependency.
This leads to the better idea to link the JobPlanning aggregators
themselves by parent-child references, which is possible since the
whole dependency chain actually sits in the stack embedded into the
Expander (in the pipeline)
...in the hope that the Optimiser is able to elide those references entirely,
when (as is here the case) they point into another field of a larger object compound
...as a preparation for solving a logical problem with the Planning-Pipeline;
it can not quite work as intended just by passing down the pair of
current ticket and dependent ticket, since we have to calculate a chained
calculation of job deadlines, leading up to the root ticket for a frame.
My solution idea is to create the JobPlanning earlier in the pipeline,
already *before* the expansion of prerequisites, and rather to integrate
the representation of the dependency relation direcly into JobPlanning
...using hard coded values instead of observation of actual runtimes,
but at least the calculation scheme (now relocated from TimeAnchor to JobPlanning)
should be a reasonable starting point.
TODO: test fails...
The initial implementation effort for Player and Job-Planning
has been reviewed and largely reworked, and some parts are now
obsoleted by the reworked alternative and can be disabled.
The basic idea will be retained though: JobPlanning is a
data aggregator and performs the final step of creating a Job
- had to fix a logical inconsistency in the underlying Expander implementation
in TreeExplorer: the source-pipeline was pulled in advance on expansion,
in order to "consume" the expanded element immediately; now we retain
this element (actually inaccessible) until all of the immediate
children are consumed; thus the (visible) state of the PipeFrameTick
stays at the frame number corresponding to the top-level frame Job,
while possibly expanding a complete tree of flexible prerequisites
This test now gives a nice visualisation of the interconnected states
in the Job-Planning pipeline. This can be quite complex, yet I still think
that this semi-functional approach with a stateful pipeline and expand functors
is the cleanest way to handle this while encapsulating all details
- fix a bug in the MockDispatcher, when duplicating the ExitNodes.
A vector-ctor with curly braces will be interpreted as std::initializer_list
- add visualisation of the contents appearing at the end of the pipeline
*** something still broken here, increments don't happen as expected
`steam/engine/mock-dispatcher.hpp |cpp` now integrates this
''complete mock setup for render jobs and frame dispatching.''
The exising `DummyJob` has been slightly adapted and renamed
to `MockJob` and is tightly integrated with the other mocks.
The implementation of a `MockDispatcher` necessitated to change
the use of `MockJobTicket`. The initial attempts used a complete
mock implementation, but this approach turned out not to be viable.
Instead — based on the ideas developed for the mock setup —
now the prospective real implementation of `JobTicket` is available
and will be used by the mock setup too. Instead of a synthetic spec,
now a setup of recursively connected `ExitNode`(s) is used; the latter
seems to develop into some kind of Facade for the render node network.
Based on this mock setup, we can now demonstrate the (mostly) complete
Job-Planning pipeline, starting from a segmentation up to render jobs,
and verify proper connectivity and job invocation.
✔
- has to be prepared / supported by the RenderEnvironmentClosure
- actual translation happens when building the Dispatcher-Pipeline
- implementation delegate through
virtual size_t Dispatcher::resolveModelPort (ModelPort)
...ouch this was insidious: the STL implementation for list does not
return a pointer to the element just allocated, but rather retrieves
and dereferences the back() / front() iterator after returning from emplace_back|front()
...which in case of re-entrant allocations is something wildly different
than the initial allocation. Thus a *cheap* and dirty placeholder implementation
just using a STL container is not possible, and we need at least
to code up likewise cheesy placeholder implementation by hand.
- separate allocation and ctor all
- use an inline buffer in the STL container
- explicitly handle ctor failures to discard allocation
- NOT THREADSAFE and likely WASTFUL in terms of performance
==> MockSupport_test now back to GREEN after complete refactoring
The existing implementation of the Player from 2012~2015 inclduded
an additional differentiation by media channel (for multichannel media)
and would build a separate CalcStream for each channel.
The in-depth analysis conducted for the ongoing »Vertical Slice« effort
revealed that this differentiation is besides the point and would never
be materialised: Since -- by definition -- all media processing has
to be done by the engine, also the generation of the final output format
including any channel multiplexing will happen in render nodes.
The only exception would be when only a single channel of multichannel
media is extracted -- yet this case would then translate into a
dedicated ModelPort.
Based on this reasoning, a lot of complexity (and some contradictions)
within the JobTicket implementation can be removed -- together with
some further leftovers of the fist attempt to build JobTickets always
from a Mock specification (we now use construction by the Segment,
based on an ExitNode, which is the expected actual implementation
for production setup)
...by defining a new scheme for access to custom allocators
...and then passing a reference to such an accessor into the
JobTicket ctor, thereby allowing the ticket istelf recursively
to place further JobTicket instances into the allocation space
--> success, test passes (finally)
Up to now, a draft/mock implementation was used, relying on a »spec tuple«,
which was fabricated by MockJobTicket. But with the introduction of
NodeGraphAttachment, the MockSequence now generates a nested ExitNode structure,
and thus the JobTicket will be created through the "real" ctor, and
no longer via MockJobTicket.
Thus it is possible to skip this whole interspersed »spec tuple«,
since ExitNode *is* already this aggregated / abstracted Spec
PROBLEM: can not implement Spec-generation, since
- we must use a λ for internal allocation of JobTickets
- but recursive type inference is not possible
Will thus need to abandon the Spec-Tuple and relocate this
traversal-and-generation code into JobTicket itself
Use another unit-test (FixtureSegment_test) to guide and cover
the transition from the existing fake-implementation to the
actual implementation, where the JobTicket will be generated
on-demand, from a NodeGraphAttachment
It turns out that the real (not mocked) implementation of JobTicket creation
is already required now for this planned (mock)Dispatcher setup;
moreover, this real implementation turns out to be almost identical
to the mock implementation written recently -- just nested structure
of prerequiste JobTickets need to be changed into a similar structur
of ExitNodes
-- as an aside: rearrange various tests to be more in-line
with the envisioned architecture of playback and engine
...this opens up yet another difficult question and a host of new problems
- how are prerequisites detected or arranged by the Builder
- how are prerequisites represented?
- what is an ExitNode in terms of implementation? A subclass of ProcNode?
- how will the actual implementation of JobTicket creation (on-demand) work?
- how to adapt the Mock implementation, while retaining the Specification
for Segments and prerequisites?
...it turns out that we actually do not need to wrap TreeExplorer
on the builder types, because basically there is only a single active
builder type, and the complete processing pipeline can be assembled
in a single terminal function.
The type rebinding problem can thus be solved just by a simple
marker struct, which inherits from a template parameter
...hard to tackle...
The idea is to wrap the TreeExplorer builder, so that our specific
builder functions can delegated to the (inherited) generic builder functions
and would just need to supply some cleverly bound lambdas. However,
resulting types are recursive, which does not play nice with type inference,
and working around that problem leads to capturing a self reference,
which at time of invocation is already invalidated (due to moving the
whole pipeline into the final storage)
...which leads to the next daunting problems:
- we need some mocked ModelPort and DataSink placeholders
- we need a way how to inherit from a partial TreeExplorer pipeline
...introduced in preparation for building the Dispatcher pipeline,
which at its core means to iterate over a sequence of frame positions;
thus we need a way to stop rendering at a predetermined point...
several years ago, it seemed like a good idea to incorporate
the link between nominal time and wall-clock time into a dedicated
anchor point, which also regulates the continued frame planning.
But it turned out that such a design mixes up several concepts
and introduces confusion regarding the meaning of "real time"
- latency can not be reasonably defined for a whole planning chunk
- skipping or sliding due to missed deadlines can not reasonably handled
within such an abstract entity; it must be handled rather at the
level of a playback process
- linking the frame grid generation directly to a planning chunk
undercuts the possible abstraction of a planning pipeline
...which is build a »Job planning pipeline« step by step
in a test setup, and then factor that out as RenderDrive,
to supersede the existing CalcPlanContinuation and get
rid of the Monads this way...
Challenges
- there is a inconsistency with channel usage
- need to establish a way how to transport the output-Sink into the JobFunctor
- need a way to propagate the current frame number to the next planning chunk
The prototypical setup of data structures and test support components
is largely complete by now — with the exception of the `MockDispatcher`,
which will be completed while moving to the next steps pertaining the
setup of a frame dispatch pipeline.
* the existing `DummyJob` was augmented to allow verification of
association between Job and `JobTicket`
* the existing implementation of `JobTicket` was verified and augmented
to allow coverage of the whole usage cycle
* a `MockJobTicket` was implemented on top, which can be generated
from a symbolical test specification (rather than from the real
Fixture data structure)
* a complete `MockSegmentation` was developed, allowing to establish
all the aforementioned data structures without an actual backing
Render Engine. Moreover, `MockSegmentation` can be generated
from the aforementioned symbolic test specification.
* as part of this work, an algorithm to split an existing Segmentation
and to splice in new segments was developed and verified
Last testcase: add deeply nested Prerequisites.
Turns out that the allocator must be able to handle
re-entrant allocations, which std::deque can not fulfil.
Thus using std::list here for the Mock implementation.
In the end, the real allocations will be done by our custom
allocator (AllocationCluster), which can be arranged easily
to support re-entrant allocation calls (since the whole point
is to just place those objects into a pre-allocated large block
and only de-allocate them later in one sway. Thus the allocator
does not need to wait for the object constructor to finish, which
trivially allows for re-entrant calls)
...which uncovers further deeply nested problems,
especially when referring to non-copyable types.
Thus need to construct a common type that can be used
both to refer to the source elements and the expanded elements,
and use this common type as result type and also attempt to
produce better diagnostic messages on type mismatch....
...the improved const correctness on STL iterators uncovered another
latent problem with out diagnositc format helper, which provide
consistently rounded float and double output, but failed to take
CV-qualifiaction into account
This is a subtle and far reaching fix, which hopefully removes
a roadblock regarding a Dispatcher pipeline: Our type rebinding
template used to pick up nested type definitions, especially
'value_type' and 'reference' from iterators and containers,
took an overly simplistic approach, which was then fixed
at various places driven by individual problems.
Now:
- value_type is conceptually the "thing" exposed by the iterator
- and pointers are treated as simple values, and no longer linked
to their pointee type; rather we handle the twist regarding
STL const_iterator direcly (it defines a non const value_type,
which is sensible from the STL point of view, but breaks our
generic iterator wrapping mechanism)
To complete the mock setup, the next step would be to extend the GenNode-based spec langage
to allow defining prerequisite Mock-JobTickets. Setting this up seems rather straight forward --
however, defining a simple testcase to cover this extension runs into surprisingly tricky problems..
- for one, the singleValIterator from Itertools has serious difficulties handling references
- but even more surprising, it seems impossible to make the "prerequisites iterator"
fit into the Tree-Explorer framework (which I intend to use as replacement
for the monadic approach)
after some extended analysis of generic types and template instances,
it seems that not TreeExplorer as such is the primary problem, but rather
there is a conceptual mismatch somewhere deep down in Itertools or Iter-Adapter
By reasoning and analysis I conclude that the differentiation into
multiple channels is likely misplaced in JobTicket; it belongs ratther
into the Segment and should provide a suitable JobTicket for each ModelPort
Handling of prerequisites also needs to be reshaped entirely after
switching to a pipeline builder for the Job-planning pipeline; as
preliminary access point, just add an iterator over the immediate
prerequisites, thereby shifting the exploration mechanism entirely
out of the JobTicket implementation
Testcase: A simple Sementation with a single and bounded Segment
As aside, figured out how to unpack an iterator such as to
tie a fixed number of references through a structural binding:
auto const& [s1,s2,s3] = seqTuple<3> (mockSegs.eachSeg());
...now able to build a mock segmentation which issues dummy jobs,
and is wired such as to verify the right job is invoked for each segment.
And this allows to build and verify the Dispatcher,
without being able to invoke actual render jobs yet.
- only the parts actually touched by the algo will be re-allocated
- when a segment is split, the clone copies carry on all data
Library: add function to check for a bare address (without type info)
...this is something I should have done since YEARS, really...
Whenever working with symbolically represented data, tests
typically involve checking *hundreds* of expected results,
and thus it can be really hard to find out where the
failure actually happens; it is better for readability
to have the expected result string immediately in the
test code; now this expected result can be marked
with a user-defined literal, and then on mismatch
the expected and the real value will be printed.
There are 12 distinct cases regarding the orientation of two intervals;
The Segmentation::splitSplice() operation shall insert a new Segment
and adjust / truncate / expand / split / delete existing segments
such as to retain the *Invariant* (seamless segmentation covering
the complete time axis)
- how to pass-in a specification given as GenNode
- now this might be translated into a MockJobTicket allocated in the MockSegmentation
Unimplemented: actually build the Segment with suitable start/end time
right now we're lacking a complete working implementation of render node invocation,
and thus the Dispatcher implementation can only be verified with the help
of mocked jobs. However, at least a preliminary implementation of tagging the
invocation instance is available, and thus we're able to verify that
a given job instance indeed belongs to and is "backed" by a specific JobTicket.
This is prerequisite for building up a (likewise mocked) Fixture datastructure,
and this in turn was meant to form the basis for attacking an actual Scheduler
implementation, followed by a real render node invocation.
- can now create a Job from JobTicket::NIL
- on invocation this Job will to nothing
Only when the first real output backend is implemented,
we can decide if this simplistic implementation is enough,
or if an empty output must be explicitly generated...
* using a simplified preliminary implementation of hash chaining (see #1293)
* simplistic implementation of hashing for time values (half-rotation)
* for now just hashing the time into the upper part of the LUID
Maybe we can even live with that implementation for some time,
depending on how important uniform distribution of hash values is
for proper usage of the frame cache.
Needless to say, various further fine points need more consideration,
especially questions of portability (32bit anyone?). Moreover, since
frame times are typically quantised, the search space for the hashed
time values is drastically reduced; conceivably we should rather
research and implement a good hash function for 128bit and then combine
all information into a single hash key....
...using the MockJobTicket setup as point of reference,
since the actual invocation of render nodes will only be drafted
later in this "Vertical Slice" integration effort...
- introduce a JobTicket::NOP (null-object pattern)
- assuming that the function splitSplice() will retain complete coverage allways
Remark:
`Fixture::getPlaylistForRender()` is a leftover from the very early implementation drafts.
This function was more or less based on the way Cinelerra works; it is clear by now
that Lumiera can not possibly work this way, given that we'll build a low-level model
and dispatch precompiled render jobs....
The Fixture and the low-level model backbone deserve a distinct namespace on their own.
Since it's built by the Builder from the Session contents, and also used by the frame dispatch,
we can expect dependence on some types from Steam-Layer, and thus this namespace
needs to reside in Steam-Layer rather, while the actual low-level Model
might become part of Vault-Layer, creating a hierarchy of data structures.
(Remark: likely also the session related namespaces will need a reorganisation)
The idea is to escape a "design deadlock" by using a test-driven prototype
implementation of the data structure to back a further development
of the Dispatcher and Scheduler implementation, which then can be used
to gradually elaborate and switch over to an actual implementation
data structure
...requires a first attempt towards defining a `JobTiket`.
This turns out quite tricky, due to using those `LinkedElements`
(intrusive single linked list), which requires all added records
actually to live elsewhere. Since we want to use a custom allocator
later (the `AllocationCluster`), this boils down to allocating those
records only when about to construct the `JobTicket` itself.
What makes matters even worse: at the moment we use a separate spec
per Media channel (maybe these specs can be collapsed later non).
And thus we need to pass a collection -- or better an iterator
with raw specs, which in turn must reveal yet another nested
sequence for the prerequisite `JobTickets`.
Anyhow, now we're able at least to create an empty `JobTicket`,
backed by a dummy `JobFunctor`....
Looks like we'll actually retain and use this low-level solution
in cases where we just can not afford heap allocations but need
to keep polymorphic objects close to one another in memory.
Since single linked lists are filled by prepending, it is rather
common to need the reversed order of elements for traversal,
which can be achieved in linear time.
And while we're here, we can modernise the templated emplacement functions
- build the reworked Job-planning pipeline more or less from scratch
- back that with mocked `Dispatcher` and `JobTicket`
- then transfer this into a `RenderDrive`, which can be tested as well
- could continue then to a `CalcStream` integration test....
- decision: the Monad-style iteration framework will be abandoned
- the job-planning will be recast in terms of the iter-tree-explorer
- job-planning and frame dispatch will be disentangled
- the Scheduler will deliberately offer a high-level interface
- on this high-level, Scheduler will support dependency management
- the low-level implementation of the Scheduler will be based on Activity verbs
This finishes a long lasting effort to rework the top-level of the Lumiera GTK UI,
to adapt to GTK-3 and the new asynchronous message based architecture.
Special credits and thanks to
* Joel Holdsworth
* Stefan Kangas
Without their relentless foundational work, the Lumiera UI could
never be where it is now. Even if some code was rewritten and several
parts of the old GTK-2 implementation are now obsolete, numerous ideas
solutions and inspirations were drawn from those early contributions
and live on as part of the reworked GUI.
Note: changing behaviour of TimeSpan to possibly flip start and end,
and also to use Offset as Offset and then re-orient,
since this seems the least surprising behaviour.
These changes carry over into changed default and limiting
on ZoomWindow constructor and various mutators, and most
notably shifting the time span always into allowed domain.
...the implementation was way too naive; in some cases we could go
into an infinite loop. In the end, using Newton approximation was not
necessary (and thus there is no loop anymore), but it helped me get
at a much better solution with very small error margin on average case.
All these corner cases are obviously "academic" to some degree,
but it turns out there is no clear-cut point where you'd be able
just so set a limit and be sure that fractional integer arithmetic
works flawless in all cases.
Thus the choice is
- give up (fractional) integers and work with floats and have to
deal with error accumulation
- or do something as chosen here, namely add a boundary zone, where
fractional integer arithmetic can be kept under control, while admitting
small errors, and in turn get the absolutely precise integers in all
everyday standard cases
The value used previously was too conservative, and prevented ZommWindow
from zooming out to the complete Time domain. This was due to missing the
Time::SCALE denominator, which increaded the limit by factor 1e6
In fact the code is able to handle even this extremely reduced limit,
but doing so seems over the top, since now detox() kicks in on several
calculations, leading to rather coarse grained errors.
Thus I decided to use a compromise: lower the limit only by factor 1000;
with typical screen pixel widths, we can reach the full time domain,
while most scaling and zoom calculations can be performed precisely,
without detox() kicking in. Obviously this change requires adjusting
a lot of the test case expectations, since we can now zoom out maximally.
As it turns out, the calculation path initially choosen for the mutateScale(Rat)
was needlessly indirect, and also duplicated several of the safeguards,
meanwhile implemented way better in conformWindowToMetric(Rat)
Thus, instead of relatively re-scaling the window, now we just
limit the given zoomFactor and pass it to conformWindowToMetric()
There is a built-in limitation, which now is even
lowered to 100000 pixels horizontally.
With the techniques introduced in this changeset, it seems possible
to support more -- yet this would be a case of unnecessary genricity;
handling such large numbers will drive more computations into the
danger zone, and doing so incurs cost in terms of testing and debugging.
Placing that into context, contemporary displays are not even 4K on
average, and it does not look likely even for cinema display to go
way beyond 8k -- so yes, I want display hardware with 100000 pixels!!
The key takeaway of this changeset:
- can calculate px = trunc(zoomFactor * duration) step wise,
even when the direct calculation would lead to wrap-around
- can safely adjust and fix the zoomFactor using Newton approximation
...even zooming out to span the complete time domain (~19000 years).
But only under the condition that the display window is sufficiently
large in terms of pixels, so we can handle the computation without
glitches.
This should not be a relevant limitation in practice, since a window
size of some 100 pixels is enough to handle Duration::MAX. Needless to add
that it's hard to imagine a media timeline of such tremendous size...
building on these Library changes, plus the safe-add function
developed some days ago, it is now possible to mark a large displacement
as `time::Offset`, and apply this to yield any valid time position,
even extreme negative values
...building on these Library changes, plus the safe-add function
developed some days ago, it is now possible to mark a large displacement
as `time::Offset`, and apply this to yield any valid time position,
even extreme negative values
The APIs for time quantisation were drafted in an early stage of the project
and then never followed-up. Especially Grid::gridAlign has no
real-world usage yet, and is only massaged in some tests.
When looking at QuantiserBasics_test, I was puzzled and led astray,
since this function suggests to materialise a continuous time into
a quantised time -- which it doesn't (there is another dedicated
function Quantiser::materialise() to that end); so, without engaging
into the discussion if this function is of any use, I'll hereby
choose a name better reflecting what it does.
This is a deep refactoring to allow to represent the distance
between all valid time points as a time::Offset or time::Duration.
By design this is possible, since Time::MAX was defined as 1/30 of
the maximum value technically representable as int64_t. However,
introducing a different limiter for offsets and durations turns
out difficult, due to the inconsistencies in the exiting hierarchy
of temporal entities. Which in turn seems to stem from the unfortunate
decision to make time entities immutable, see #1261
Since the limiter is hard wired into the `time::TimeValue` constructor,
we are forced to create a "backdoor" of sorts, to pass up values
with different limiting from child classes. This would not be so
much of a problem if calculations weren't forced to go through `TimeVar`,
which does not distinguish between time points and time durations.
This solution rearranges all checks to be performed now by time::Offset,
while time::Duration will only take the absolute value at construction,
based on the fact that there is no valid construction path to yield
a duration which does not go through an offset first.
Later, when we're ready to sort out the implementation base of time values
(see #1258), this design issue should be revisited
- either we'll allow derived classes explicitly to invoke the limiter functions
- or we may be able to have an automatic conversion path from clearly
marked base implementation types, in which case we wouldn't use the
buildRaw_() and _raw() "backdoor" functions any more...
While the calculation was already basically under control, I just was not content
with the achieved numeric precision -- and the fact that the test case in fact
misses the bar, making it difficult do demonstrate that the calculation
is not derailed. I just had the gut feeling that it must be somehow possible
to achieve an absolute error level, not just a relative error level of 1e-6
Thus I reworked this part into a generic helper function, see #1262
The end result is:
* partial failure. we can only ''guarantee'' the relative error margin of 1e-6
* but in most cases not out to the most extreme numbers, the sophisticated
solution achieves much better results way below the absolute error level of 1µ-Tick
Thus with using rational numbers, we have now a solution that is absolutely precise
in the regular case, and gradually introduces errors at the domain bound
but with an guaranteed relative error margin of 1e-6 (== Time::SCALE)
...in a similar vein as done for the product calculation.
In this case, we need to check the dimensions carefully and pick
the best calculation path, but as long as the overall result can
be represented, it should be possible to carry out the calculation
with fractional values, albeit introducing a small error.
As a follow-up, I have now also refactored the re-quantisation
functions, to be usable for general requantisation to another grid,
and I used these to replace the *naive* implementation of the
conversion FSecs -> µ-Grid, which caused a lot of integer-wrap-around
However, while the test now works basically without glitch or wrap,
the window position is still numerically of by 1e-6, which becomes
quite noticeably here due to the large overall span used for the test.
...using a requantisation trick to cancel out some factors in the
product of two rational numbers, allowing to calculate the product
without actual multiplication of (dangerously large) numbers.
with these additional safeguards, the anchorWindowAtPosition()
succeeds without Integer-wrap, but the result is not fully correct
(some further calculation error hidden somewhere??)
- detailed documentation of known problematic behaviour
when working with rational fractions
- demonstrate the heuristic predicate to detect dangerous numbers
- add extensive coverage and microbenchmarks for the integer-logarithm
implementation, based on an example on Stackoverflow. Surprising result:
The std::ilog(double) function is of comparable speed, at least for
GCC-8 on Debian-Buster.
Especially rational numbers with large denominator can be insidious,
since they might cause numeric overflow on seemingly harmless operations,
like adding a small number.
A solution might be to *requantise* the number into a different,
way smaller denominator. Obviously this is a lossy operation;
yet a small and controlled numeric error is always better than
an uncontrolled numeric wrap-around.
- protection against negative numbers seems adequate
- a possible concern are handling of very large time spans
- definitively have to guard against "poisonous" fractions
(e.g. n / INT_MAX)
- some test definitions were simply numerically wrong
- changed some aspects of the specified behaviour, to be more consistent
+ scrolling is more liberal and always allows to extend canvas
+ setting window to a given duration expands around anchor point
Rearrange the internal mutator functions to follow a common scheme,
so that most of the setters can be implemented by simple forwarding.
Move the change-listener triggering up into the actual setters.
This makes further test cases pass
- verify_setup
- verify_calibration
...implying that the pixel width is now retained
and basic behaviour matches expectations
Extensive tests with corner cases soon highlighted this problem
inherent to integer calculations with fractional numbers: it is
possible to derail the calculation by numeric overflow with values
not excessively large, but using large numbers as denominator.
This problem is typically triggered by addition and subtraction,
where you'd naively not expect any problems.
Thus changed the approach in the normalisation function, relying
on an explicitly coded test rather, and performing the adjustment
only after conversion back to simple integral micro-tick scale.
Getting all those requirements translated into code turns out to be a challenging task;
and the usual ascent to handle such a situation is to define **Invariants**
in conjunction with a normalisation scheme; each manipulation will then be
translated into invocation of one of the three fundamental mutators,
and these in turn always lead into the common normalisation sequence.
__Invariants__
- oriented and non-empty windows
- never alter given pxWidth
- zoom metric factor < max zoom
- visibleWindow ⊂ Canvas
Writing this specification unveiled a limitation of our internal
time base implementation, which is a 64bit microsecond grid.
As it turns out, any grid based time representation will always
be not precise enough to handle some relevant time specifications,
which are defined by a divisor. Most notably this affects the precise
display of frame duration in the GUI, and even more relevant,
the sample accurate editing of sound in the timeline.
Thus I decided to perform the internal computation in ZoomWindow
as rational numbers, based on boost::rational
Note: implementation stubbed only, test fails
This ZoomWindow_test highlights again the question about the intended usage
of the Lumiera time entities. In which way do we want to perform time calculations,
and under which circumstances is it adequate to perform arithmetic on
raw time values?
These questions made me think about rather far reaching concerns regarding
subsidiarity and implicit or explicit usage context. Basically I could
reconfirm the design choices taken some years ago -- while I must admit
that the project is headed towards a way larger scale and more loose
coupling of the parts, than I could imagine several years ago, at the
time when the design started...
As a side note: we can not avoid that some knowledge about the time implementation
leaks out from the support lib; time codes themselves are tightly coupled
to the usage scenario within the session and can not be used as means
for implementing UI concerns. And the more generic time frameworks,
like std::chrono (as much as it is desirable to have some integration here)
will not be of any help for most of our specific usage patterns.
The reason is, for film editing we do not have a global time scale,
rather the truth is when the film starts....
implement the first test case: nudge the zoom factor
⟹ scale factor doubled
⟹ visible window reduced to half size
⟹ visible window placed in the middle of the overall range
The header "format-cout.hpp" offers a convenience function
to print pretty much any object or data in human readable form.
However, the formatter for pointers used within this framework
switched std::cout into hexadecimal display of numbers and failed
to clean-up this state.
Since the "stickyness" of IOS stream manipulators is generally a problem,
we now provide a RAII helper to capture the previous stream state and
automatically restore it when leaving the scope.
Complete the investigation and turn the solution into a generic
mix-in-template, which can be used in flexible ways to support
this qualifier notation.
Moreover, recapitulate requirements for the ElementBoxWidget
- move construct into the buffer
- directly invoke the payload constructor through PlantingHandle
- reconsider type signature and size constraint
- extend the unit test
- document a corner case of c++ "perfect forwarding",
which caused me some grief here
...this extension was spurred by the previeous refactoring.
Since 'emplace' now clearly denotes an operation to move-embed an existing object,
we could as well offer a separate 'create' API, which would take forwarding
arguments as usual and just delegates to the placement-new operation 'create'
already available in the InPlaceBuffer class.
Such would be a convenience shortcut and is not strictly necessary,
since move-construction is typically optimised away; yet it would also
allow to support strictly non-copyable payload types.
This refactoring also highlights a fuzziness in the existing design,
where we just passed the interface type, while being sloppy about the
DEFAULT type. In fact this *is* relevant, since any kind of construction
might fail, necessitating to default-construct a placeholder, since
InPlaceBuffer was intended for zero-overhead usage and thus has in itself
no means to know about the state of its buffer's contents. Thus the
only sane contract is that there is always a valid object emplaced
into the buffer, which in turn forces us to provide a loophole for
class hierarchies with an abstract base class -- in such a case the
user has to provide a fallback type explicitly.
...for the operation on a PlantingHandle, which allows
to implant a sub type instance into the opaque buffer.
* "create" should be used for a constructor invocation
* "emplace" takes an existing object and move-constructs
this allows to avoid multi-step indirection
when translating mouse dragging pixel coordinates
into a time offset for the dragged clip widget.
Moreover this also improves the design,
since the handling of canvas metric is pretty much
a self contained, separate concern
some bugfixes,
but also a notable change: detect the completion of the gesture
directly when the button is released; this is necessary, because
seemingly we do not get motion_events when no button is pressed,
at least not in this test setup based on a Gtk::Button widget.
GTK doesn't expose a first-class API for this,
since -- by design -- the extension of a widget is negotiated.
Thus I'm looking for some kind of workaround for our specific use-case,
where a clip widget must be rendered with a well defined horizontal size,
corresponding to its length.
Thus far, we're only able to increase the size of the Button widget
used as placeholder, but we can not forcibly shrink that button,
probably because the embedded Gtk::Lable requires additional extension.
Partially as a leftover from the way more ambitious initial design,
we ended up with CanvasHook as an elaboration/specialisation of the
ViewHook abstraction. However, as it stands, this design is tilted,
since CanvasHook is not just an elaboration, but rather a variation
of the same basic idea.
And this is now more like a building pattern and less of a generic
framework, it seems adequate to separate these two variations completely,
even if this incurs a small amount of code duplication.
Actually this refactoring is necessary to resolve a bug, where
we ended up with the same Clip widgets attached two times to the
same Canvas control, one time through the ViewHook baseclass,
and a second time by the ctor of the "derived" CanvasHook
...in an attempt to clarify why numerous cross links are not generated.
In the end, this attempt was not very successful, yet I could find some breadcrumbs...
- file comments generally seem to have a problem with auto link generation;
only fully qualified names seem to work reliably
- cross links to entities within a namespace do not work,
if the corresponding namespace is not documented in Doxygen
- documentation for entities within anonymous namespaces
must be explicitly enabled. Of course this makes only sense
for detailed documentation (but we do generate detailed
documentation here, including implementation notes)
- and the notorious problem: each file needs a valid @file comment
- the hierarchy of Markdown headings must be consistent within each
documentation section. This entails also to individual documented
entities. Basically, there must be a level-one heading (prefix "#"),
otherwise all headings will just disappear...
- sometimes the doc/devel/doxygen-warnings.txt gives further clues
...by relying on the newly implemented automatic standard binding
Looks like a significant improvement for me, now the actual bindings
only details aspects, which are related to the target, and no longer
such technicalitis like how to place a Child-Mutator into a buffer handle
After this long break during the "Covid Year 2020",
I pick this clean-up task as a means to fresh up my knowledge about the code base
The point to note is, when looking at all the existing diff bindings,
seemingly there is a lot of redundancy on some technical details,
which do not cary much meaining or relevance at the usage site:
- the most prominent case is binding to a collection of DiffMutables hold by smart-ptr
- all these objects expose an object identity (getID() function), which can be used as »Matcher«
- and all these objects can just delegate to the child's buildMutator() function
for entering a recursive mutation.
...and the result was very much worth the effort,
leading to more focused and cleaner code.
- all the concerns of moving widgets and translating coordinates
are now confined to the second abstraction layer (CanvasHook)
- while the ViewHook now deals exclusively with attachment, detachment
and reordering of attachment sequence
As it turned out, it is rather easy to extend the existing listener
for structural changes to detect also value assignments. Actually
it seems we'd need both flavours, so be it.
Yeah, C++17, finally!
...not totally sure if we want to go that route.
However, the noise reduction in terms of code size at call site looks compelling
this draft commit reshifts the (meanwhile broken) test code from:
03c358fe86
Now the marker Buttons are injected again, but without any detailed
positioning code at call site. This demonstrates the viability of the
Structure-Change / ViewHook refactoring.
To make this change viable, it was necessary to remove the ViewHooked<>
marker template from the rehook() callback. As it turns out, this was
added rather for logical reasons, and is in fact not necessary in
any of the existing ViewHook implementations (and I don't expect any
other implementations to come)
BUT the actual positioning coordinates are still wrong (which seems
to re related to other conceptual problems in coordinate offset handling)
the reason for the failure, as it turned out,
is that 'noexcept' is part of the function signature since C++17
And, since typically a STL container has const and non-const variants
of the begin() and end() function, the match to a member function pointer
became ambuguous, when probing with a signature without 'noexcept'
However, we deliberately want to support "any STL container like" types,
and this IMHO should include types with a possibly throwing iterator.
The rationale is, sometimes we want to expose some element *generator*
behind a container-like interface.
At this point I did an investigation if we can emulate something
in the way of a Concept -- i.e. rather than checking for the presence
of some functions on the interface, better try to cover the necessary
behaviour, like in a type class.
Unfortunately, while doable, this turns out to become quite technical;
and this highlights why the C++20 concepts are such an important addition
to the language.
So for the time being, we'll amend the existing solution
and look ahead to C++20
as it turns out, "almost" the whole codebase compiles in C++17 mode.
with the exception of two metaprogramming-related problems:
- our "duck detector" for STL containers does not trigger anymore
- the Metafunction to dissect Function sigantures (meta::_Fun) flounders
"%broken" is not broken anymore, but renders a boolean,
and we configured the formatter not to complain on missing values.
Fortunately "%madness" is still broken ;-)
When drafting the time handling framework some years ago,
I foresaw the possible danger of mixing up numbers relating
to fractional seconds, with other plain numbers intended as
frame counts or as micro ticks. Thus I deliberately picked
an incompatible integer type for FSecs = boost::rational<long>
However, using long is problematic in itself, since its actual
bit length is not fixed, and especially on 32bit platforms long
is quite surprisingly defined to be the same as int.
However, meanwhile, using the new C++ features, I have blocked
pretty much any possible implicit conversion path, requiring
explicit conversions in the relevant ctor invocations. So,
after weighting in the alternatives, FSecs is now defined
as boost::rational<int64_t>.
basically the solution was a bit too naive and assumed everything is similar to a vector.
It is not, and this leads to some insidious problems with std::map, which hereby
are resolved by introducing ContainerTraits
All of the existing "simple" tests for the »Diff Framework« are way to much low-level;
they might indeed be elementary, but not introductory and simple to grasp.
We need a very simplistic example to show off the idea of mutation by diff,
and this simple example can then be used to build further usage test cases.
My actual goal for #1206 to have such a very basic usage demonstration and then
to attach a listener to this setup, and verify it is actually triggered.
PS: the name "GenNodeBasic_test" is somewhat pathetic, this test covers a lot
of ground and is anything but "basic". GenNode in fact became a widely used
fundamental data structure within Lumiera, and -- admittedly -- the existing
implementation might be somewhat simplistic, while the whole concept as such
is demanding, and we should accept that as the state of affairs
now the lifecycle of widget and hook are tightly interwoven.
Indeed the test uncovered a situation where a call into the
already destroyed Canvas might halt the application.
...basically it occurred to me that in practice we will never have to deal
with isolated ViewHooks, rather with widgets-combinded-with-a-hook.
So the idea is to combine both into a template ViewHooked<W>
basically this attempts to work around an "impedance mismatch" caused by relying on Lumiera's Diff framework.
Applying a diff might alter the structural order of components, without those componets
being aware of the change. If especially those components are attached into some
UI layout, or otherwise delegate to display widgets, we need a dedicated mechanism
to reestablish those display elements in proper order after applying the change.
The typical examples is a sequence of sub-Tracks, which might have been reordert due
to applying rules down in the Steam Layer. The resulting diff will propagate the
new order of sub-Tracks up into the UI, yet now all of the elaborate layout and
space allocation done in the presentation code needs to be adjusted or even
recomputed to accomodate the change.
...which serves to solve the problem with Canvas access.
Basically we do not want each and every Clip widget to be aware of the concrete canvas implementation widget;
and in addition, automated removal of widgets from the Canvas seems desirable
For context: The »Advice System« was coined a long time ago, in 2010,
based on the vague impression that it might be useful for that kind of application
we are about to build here. And, as can be expected, none of the usage situations
envisioned at that time was brought to bear. Non the less, the facility came in
handy at times, precisely because it is cross-cutting and allows to pass
information without imposing any systematic relationship between the
communication partners.
And now we've got again such a situation.
The global style manager in the UI has to build a virtual CSS path,
which is needed by drawing code somewhere deep down, and we absolutely
do not want to pass a reference to the style manager over 20 recursive calls.
The alternatives would be
(1) to turn the style manager into a public service
(2) to have a static access function somewhere
(3) to use a global variable.
For rationale, (1) would be overblown, because we do not actually request
a service to do work for us, rather we need some global piece of information.
(2) would be equivalent to (1), just more confusing. And (3) is basically
what the Advice system does, with the added benefit of a clear-cut service
access point and a well defined lifecycle.
This changeset adds the ability to check if actual Advice has been published,
which allows us to invoke the (possibly expensive) GTK path building and
style context building code only once.
Mostly, std::regexp can be used as a drop-in replacement.
Note: unfortunately ECMA regexps do not support lookbehind assertions.
This lookbehind is necesary here because we want to allow parsing values
from strings with additional content, which means we need explicitly to
exclude mismatches due to invalid syntax.
We can work around that issue like "either line start, or *not* one of these characters.
Alternatively we could consider to make the match more rigid,
i.e we would require the string to conain *only* the timecode spec to be parsed.
The existing implementation created a Buffer-Type based on various traits,
including the constructor and destructor functions for the buffer content.
However, this necessitates calculating the hash_value of a std::function,
which (see #294) is generally not possible to implement.
So with this changeset we now store an additional identity hash value
right into the TypeHandler, based on the target type placed into the buffer
This was prompted by a test failing under Boost-1.65 (--> see #294)
When reviewed now, the whole idea of testing Steam-Layer Commands for
equivalence feels a bit sketchy.
Just the comparison for the command ''identity'' alone seems sufficient,
i.e. the test if a command-ID is associated with the same backend-handle
and thus the same functor binding.
the template lib::PolymorphicValue seemingly picked the wrong
implementation strategy for "virtual copy support": In fact it is possible
to use the optimal strategy here, since our interface inherits from CloneSupport,
yet the metaprogramming logic picked the mix-in-adapter (which requires one additional "slot"
of storage plus a dynamic_cast at runtime).
The reason for this malfunction was the fact that we used META_DETECT_FUNCTION
to detect the presence of a clone-support-function. This is not correct, since
it can only detect a function in the *same* class, not an inherited function.
Thus, switching to META_DETECT_FUNCTION_NAME solves this problem
Well, this solution has some downsides, but since I intend to rewrite the
whole virtual copy support (#1197) anyway, I'll deem this acceptable for now
TODO / WIP: still some diagnostics code to clean up, plus a better solution for the EmptyBase
...which, in the end, can even be considered the more logical design choice,
since the "verb visitor" is a more elaborated and sophisiticated Verb-Token,
adding the special twist of embedded storage for variable function arguments
...but bad news on the main issue:
the workaround consumes the tuple and thus is not tenable!
And what is even worse: the textbook implementation of std::apply is
equivalent to our workaround and also consumes the argument tuple
A simple yet weird workaround (and basically equivalent to our helper function)
is to wrap the argument tuple itself into std::forward<Args> -- which has the
effect of exposing RValue references to the forwarding function, thus silencing
the compiler.
I am not happy with this result, since it contradicts the notion of perfect forwarding.
As an asside, the ressearch has sorted out some secondary suspicions..
- it is *not* the Varargs argument pack as such
- it is *not* the VerbToken type as such
The problem clearly is related to exposing tuple elements to a forwarding function.
this is a generalisation of what we use in the diff framework;
typically you'd package the VerbToken into some kind of container,
together with the concrete invocation argument.
However, the specific twist here is that we want *variable arguments*,
depending on the actual operation called on the interpreter interface.
...which leads to a specific twist here; while in the simple version
we still could hope to get away with a simple uniform uint argument,
the situation has changed altogether now. The canvas has turned into
some generic component, since it is instantiated two times, onece for
the time ruler and once for the actual body content. Thus all of the
specifics of the drawing code need to be pushed into a new, dedicated
renderer component. And this more or less forces us to pass all the
actual presentation variations through the invocation arguments of
the visitor.
So we're now off again for a digression, we need a more generalised visitor
- we got occasional hangups when waiting for disabled state
- the builder was not triggered properly, sometimes redundant, sometimes without timeout
As it turned out, the loop control logic is more like a state machine,
and the state variables need to be separated from the external influenced variables.
As a consequence, the inChange_ variable was not calculated properly when disabled in a race,
and then the loop went into infinite wait state, without propagating this to
the externally waiting client, which caused the deadlock
effectively we rely in the micro tick timescale promoted by libGAVL,
but it seems indicated to introduce our own constant definition.
And also clarify some comments and tests.
(this changeset does not change any values or functionality)
basically we can pick just any convention here, and so we should pick the convention in a way
that makes most sense informally, for a *human reader*. But what we previously did, was to pick
the condition such as to make it simple in some situations for the programmer....
With the predictable result: even with the disappointingly small number of usages we have up to now,
we got that condition backwards several times.
OK, so from now on!!!
Time::NEVER == Time::MAX, because "never" is as far as possible into the future
- most notably the NOBUG logging flags have been renamed now
- but for the configuration, I'll stick to "GUI" for now,
since "Stage" would be bewildering for an occasional user
- in a similar vein, most documentation continues to refer to the GUI
...it should have been explicit from start, since there is no point
in converting an EntryID into a plain flat string without further notice
this became evident, when the compiler picked the string overload on
MakeRec().genNode(specialID)
...which is in compliance to the rules, since string is a direct match,
while BareEntryID would be an (slicing) upcast. However, obviously we
want the BareEntryID here, and not an implicit string conversion,
thereby discarding the special hash value hidden within the ID
...rather extend the "object builder" DSL notation to allow passing in a given EntryID literally.
Rationale is, we should handle the problem of unique IDs on the level of the domain model.
If we attempt to "fix" this within GenNode, the price would be to make the ETD creation stateful
seems to work surprisingly well...
the diff application poceeds in the GUI up to the point
where the TrackPresenter need to be inserted into a two-fold display context
As it turns out, using the functional-notation form conversion
with *parentheses* will fall back on a C-style (wild, re-interpret) cast
when the target type is *not* a class. As in the case in question here, where
it is a const& to a class. To the contrary, using *curly braces* will always
attempt to go through a constructor, and thus fail as expected, when there is
no conversion path available.
I wasn't aware of that pitfall. I noticed it since the recently introduced
class TimelineGui lacked a conversion operator to BareEntryID const& and just
happily used the TimelineGui object itself and did a reinterpret_cast into BareEntryID
...these magical strings are already spreading dangerously throughout the code base
PS: also fixup for c6b8811af0 (broken whitespace in test definition)
this initially (on 1.9.18) triggered this extended digression;
The initial naive implementation (without backtracking) did not allow
to express such a simple thing like "function XXX" not invoked (again) after "function XXX"
...seems basically sane now.
Just we still need to wrap it one more time into IterableDecorator;
which means the overall scheme how to build and package the whole pipeline
is not correct yet.
Maybe it is not possible to get it packaged all into one single class?
on closer investigation it turned out that the logic of the
first design attempt was broken altogether. It did not properly
support backtracking (which was the reason to start this whole
exercise) and it caused dangling references within the lambda
closure once the produced iterator pipeline was moved out
into the target location.
Reasoning from first principles then indicated that the only sane
way to build such a search evaluation component is to use *two*
closely collaborating layers. The actual filter configuration
and evaluation logic can not reside and work from within the
expander. Rather, it must sit in a layer on top and work in
a conventional, imperative way (with a while loop).
Sometimes, functional programming is *not* the natural way
of doing things, and we should then stop attempting to force
matters against their nature.
...and TADAA ... there we get an insidious bug:
we capture *this by reference into the expansion functor,
and then we move *this away, out from the builder into the target....
Up to now, we had a very simplistic configuration option just
to search for a match, and we had the complete full-blown reconfiguration
builder option, which accepts a functor to work on and reconfigure the
embedded Filter chain.
It occurred to me that in many cases you'd rather want some intermediary
level of flexibility: you want to replace the filter predicate entirely
by some explicitly given functor, yet you don't need the full ability
to re-shape the Filter chain as a whole. In fact the intended use case
for IterChainSearch (which is the EventLog I am about to augment with
backtracking capabilities) will only ever need that intermediate level.
Thus wer're adding this intermediary level of configurability now.
The only twist is that doing so requires us to pass an "arbitrary function like thing"
(captured by universal reference) through a "layer of lambdas". Which means,
we have to capture an "arbitrary thingie" by value.
Fortunately, as I just found out today, C++14 allows something which comes
close to that requirement: the value capture of a lambda is allowe to have
an intialiser. Which means, we can std::forward into the value captured
by the intermediary lambda. I just hope I never need to know or understand
the actual type this captured "value" takes on.... :-)
with the augmented TreeExplorer, we're now able to get rid of the
spurious base layer, and we're able to discard the filter and
continue with the unfiltered sequence starting from current position.
build a special feature into the Explorer component of TreeExplorer,
causing it to "lock into" the current child sequence and discard
all previous sequences from the stack of child explorations
So we have now a reworked version of the internals of TreeExplorer in place.
It should be easier to debug template instantation traces now, since most
of the redundancy on the type parameters could be remove. Moreover, existing
pipelines can now be re-assigned with similarily built pipelines in many cases,
since the concrete type of the functor is now erased.
The price tag for this refactoring is that we have now to perform a call
through a function pointer on each functor invocation (due to the type erasure).
And seemingly the bloat in the debugging information has been increased slightly
(this overhead is removed by stripping the binary)
