The new emerging NLE for GNU/Linux
State Draft
Date 2010-04-16
Proposed by Ichthyostega

Overview Engine Interface(s)

At the Engine Interfaces, Lumiera’s Backend and Session get connected and work together to produce rendered output. This design proposal intends to give an overview of the connection points and facilities involved, to define some terms and concepts and to provide a foundation for discussion and working out the APIs in detail.


Render Process

represents an ongoing calculation as a whole

Engine Model

encloses the details of the current engine configuration and wiring


translates a render process into the (planned) invocation of individual nodes


cares for calculations actually to happen, in the right order and just in time, if at all


abstraction of an processing unit, supports planning by the dispatcher, allows to pull data, thereby driving the actual calculation.

Render Process

The render process brackets an ongoing calculation as a whole. It is not to be confused with a operating system process or thread; rather it is a point of reference for the relevant entities in the Stage and Steam-Layer in need to connect to such a "rendering", and it holds the specific definitions for this calculation series. A render process corresponds to a single data stream to be rendered. Thus, when the play controller of some timeline in the model is in playing or paused state, typically multiple corresponding render processes exist.

  • there is an displayer- or output slot, which got allocated on creation of the process

  • the process disposes calculated data frames "into" this slot

  • the process can be paused/started and stopped (aborted, halted).

  • some processes allow for changing parameters dynamically (e.g. speed, direction)

  • each process has to ensure that the output/display slot gets closed or released finally

Process parameters

A process is linked to a single stream data format (a → stream implementation type).
It is configured with frame quantisation and timings, and a model port identifier and channel selector.


translates time values into frame numbers. (In the most general case this is a function, connected to the session)


a definition to translate global model time units in real clock time, including alignment to an external time grid.

model port

a point in the (high level) model where output can be produced.
This might be a global pipe in one of the model’s timelines, or it might be a probe point.


within the session and high level model, details of the stream implementation are abstracted. Typically, a global pipe (master bus or subgroup) corresponds to a multichannel stream, and each of these channels might be hooked up to an individual render process (we have to work out if that’s always the case or just under some circumstances)


While certainly the port and channel definition is fixed, unfortunately the quantisation and the timings are’nt. The timings may be changed in the middle of an ongoing render process, due to changed playback speed, shuffling or requirements forwarded from chase-and-lock synchronisation to an external source. We still need to discuss if Lumiera is going to support variable framerates (several media professionals I’ve talked to were rather positive we need to support that — personally I’m still in doubt we do). Variable framerates force us to determine the frame numbers by an integration over time from a start position up to the time position in question. The relevant data to be integrated is located in the session / high-level model; probably we’ll then create an excerpt of this data, but still the less quantisation will be a function of time. Anyway, it is the render processes job to translate all kinds of parameter changes into relevant internal API calls to reconfigure the calculation process to fit.

Engine Model (low-level Model)

The low level model is a network of interconnected render nodes. It is created by the build process to embody any configuration, setup and further parametrisation derived from the high-level description within the session. But the data structure of this node network is opaque and considered an implementation detail. It is not intended to be inspected and processed by outward entities (contrast this to the high-level model within the session, which provides an extensive discovery API and can be manipulated by model mutating commands). We just provide a set of query and information retrieval functions to suit the needs of the calculation process. The engine model is not persisted.

  • the engine model is partitioned by a segmentation of the time axis. Individual segments can be hot-swapped.

  • the engine has exit nodes, corresponding to the model ports mentioned above

  • each exit node provides a stream type definition plus quantisation and alignment constraints.

Thus, for any pair (port, time) it is possible to figure out a segment and an exit node to serve this position. The segmentation(s) for multiple ports might differ. To allow for effective dispatching, the model should provide convenience functions to translate these informations into frame number ranges. The mentioned quantisation and alignment constraints stem from the fact that the underlying media source(s) are typically themselves quantised and the timings might be manipulated within the processing chain. We might or might not be able to shift the underlying media source (it might be a live input or it might be tied to a fixed timecode)

Processing Node

In this context, a node is a conceptual entity: it is an elementary unit of processing. It might indeed be a single invocation of a processor (plugin or similar processing function), or it might be a chain of nodes, a complete subtree, it might represent a data source (file, external input or peer in case of distributed rendering), or it might stand for a pipeline implemented in hardware. The actual decision about these possibilities happened during the build process and can be configured by rules. Information about these decisions is retained only insofar it is required for the processing, most of the detailed type information is discarded after the wiring and configuration step. As mentioned above, each node serves two distinct purposes, namely to assist with the planning and dispatching, and to pull data by performing the calculations.

Nodes can be considered stateless — pulling a node has no effect outside the invocation context. While a node might actually be configured to drive a whole chain or subtree and propagate the pull request within this tree or chain internally, the node never propagates a pull request beyond its realm. The pull() call expects to be provided with all prerequisite data, intermediary and output buffers.

Dispatching Step

The dispatcher translates a render process (actually a calculation stream as part of a render process) into sequences of node invocations, which then can be analysed further (including planning the invocation of prerequisites) and scheduled. This mapping is assisted by the engine model API (to find the right exit node in the right segment), the render process (for quantisation) and the involved node’s invocation API (to find the prerequisites)

Node Invocation API

As nodes are stateless, they need to be embedded into an invocation context in order to be of any use. The node invocation has two distinct stages and thus the invocation API can be partitioned in two groups


During the planning phase, the dispatcher retrieves various informations necessary to schedule the following pull call. These informations include

  • reproducible invocation identifier, usable to label frames for caching

  • opaque source identifier (owned by the backed) in case this node represents a source

  • prerequisite nodes

  • index (channel) of the prerequisite’s output to be fed as input buffer(s)

  • number and size of the output buffers required

  • additional memory required

  • control data frame(s)

Node pull

  • the pull call expects to be provided with all the resources announced during the planning step

  • moreover, the pull call needs to know (or some way to figure out) the time coordinates

  • after retrieving automation, the control flow forwards to the actual processing function

  • there is an result/error code (assuming the scheduler prefers error codes over exceptions)


  • find out if we need to support variable framerate (→ yes, implementation deferred)

  • find out about the exact handling of multichannel data streams (✔ done)

  • design and prototypical implementation of frame quantisation (✔ done)

  • design a buffer descriptor (✔ done)

  • design a buffer designation scheme TODO

  • expand on the node identification scheme TODO

  • clarify how control data frames can be addressed TODO



Currently we’re focussing on how to implement this concept, not on evaluating alternatives. Especially the idea of scheduling individual frame jobs is a core concept of Lumiera. This RfC tries to bridge from the session model to an engine based on these concepts. It’s the attempt to link two concepts already defined and decided on….


  • allow for optimal resource use and avoid blocking of threads

  • shift away complexity from the engine into the builder, which is by far not so performance critical

  • allow to adjust the actual behaviour of the engine in a wide range, based on actual measurements

  • create a code structure able to support the foreseeable extensions (hardware and distributed rendering) without killing maintainability


State → Draft

Requirements and details of the design are sufficiently clear meanwhile. Ther seems to be not much room for alternative approaches, given our general planning for the application

Mi 11 Mai 2011 19:27:12 CEST Ichthyostega <>