lumiera_documentation/doc/devel/rfc/RenderOptimizer.txt

Design Process : Optimiser
==========================


[options="autowidth"]
|====================================
|*State*       | _Parked_
|*Date*        | _2007-06-07_
|*Proposed by* | ct
|====================================

Render Optimizer
----------------
Render only parts of a frame which are necessary for the Output;
optimize render pipeline for efficiency


Description
~~~~~~~~~~~
This Idea is just stored here for later reference/implementation.

Effects give some information on which data their output depends (like
transitions, temporal dependencies, color/alpha etc) and what the operation
costs. Based on this information we optimize the render pipeline, for example
if the output is a zoom, then we only need to calculate the parts of a frame
which will be viewable in the output (plus some more dependencies, like blur
has radius and so on). Further in some cases it might be favorable to reorder
some effects for the actual render process, as long it would produce the same
output as the original sequence of effects.





Comments
--------

Possible classification for video filters:

 1. The filter only changes the color of each pixel in the same way
 2. The filter deforms the image but leaves the color
 3. The filter makes complex things. The only additional hint it can export is
 the number of referenced past frames, if such a limit exists (sometimes it
 doesn't).

Filters of type 1 and type 2 never use any previous frames, and are strictly
one frame in -- one frame out. Filters of type 1 can always be swapped with
filters of type 2, the output is the same. All other filters cannot be swapped
in general.

The good news is, that:

 1. All commonly used filters are either type 1 or type 2
    (type 3 are more the fun effects)
 2. Filters of type 2 are colormodel agnostic
 3. If a filter of type 1 makes only linear transformations of the color vectors +
    (i.e.: new_color ≔ matrix · old_color),
 then the matrix can be transformed from e.g. RGB to YUV, so these filters can
 always work in both colorspaces directly



.State -> Parked
Generally this is accepted but needs some more polishing when we go over it.

Christian Thaeter::  'Thu 14 Apr 2011 03:06:42 CEST'


.Reassessment : remains valid
[underline]#Relevance confirmed 2025.#
This RfC is written narrow enough to just define the goal of an optimised
render pipeline -- which _was_ and _is_ a central goal of the Lumiera project.

Approaching this topic however _requires much more work_ and notably a detailed
analysis of the possibilities, based on *facts and measurements*. There is the
real danger to get attached to some implementation idea deemed enticing,
leading to very specific tweaks which limit the ability to adapt to
other relevant requirements.

Furthermore, since Lumiera aims at being library / framework agnostic,
a significant difficulty to overcome first is the build-up of an abstract
framework to describe and engage viable optimisations, which can then be
applied as transformations within the Builder -- quite similar to the
proceedings of an optimising compiler. The description of possible
optimisations likely involves elements of knowledge representation,
combined with heuristical aspects, possibly augmented by
machine learing.

To summarise: this is a *long-term vision*.

Ichthyostega:: '2025-09-20'


''''
Back to link:/x/DesignProcess.html[Lumiera Design Process overview]