OpenSubdiv/documentation/sdc_overview.rst
jmooney c808f8df2d Added some minor grammar and spelling fixes
The changes are mostly limited to punctuation fixes, though some sentences
were restructured for clarity.

The files modified include:
	- documentation/porting.rst
	- documentation/sdc_overview.rst
2015-06-22 20:24:43 -07:00

352 lines
19 KiB
ReStructuredText

..
Copyright 2013 Pixar
Licensed under the Apache License, Version 2.0 (the "Apache License")
with the following modification; you may not use this file except in
compliance with the Apache License and the following modification to it:
Section 6. Trademarks. is deleted and replaced with:
6. Trademarks. This License does not grant permission to use the trade
names, trademarks, service marks, or product names of the Licensor
and its affiliates, except as required to comply with Section 4(c) of
the License and to reproduce the content of the NOTICE file.
You may obtain a copy of the Apache License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the Apache License with the above modification is
distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
KIND, either express or implied. See the Apache License for the specific
language governing permissions and limitations under the Apache License.
Sdc Overview
------------
.. contents::
:local:
:backlinks: none
Subdivision Core (Sdc)
======================
Sdc is the lowest level layer in OpenSubdiv. Its intent is to separate
the core subdivision details from any particular representation of a mesh
(it was previously bound to Hbr) to facilitate the generation of consistent
results with other mesh representations, both internal and external to
OpenSubdiv.
The functionality can be divided roughly into three sections:
* types, traits and options for the supported subdivision schemes
* computations required to support semi-sharp creasing
* computations for mask weights of subdivided vertices for all schemes
For most common usage, familiarity with only the first of these is necessary --
primarily the use of public types and constants for the choice of subdivision
scheme and its associated options. The latter two provide the basis for a
more comprehensive implementation of subdivision, which requires considerably
more understanding and effort.
Overall, the approach was to extract the functionality at the lowest level
possible. In some cases, the implementation is not far from being simple
global functions. The intent was to start at a low level and build any higher
level functionality as needed. What exists now is functional for ongoing
development and anticipated needs within OpenSubdiv for the near future.
The intent of Sdc is to provide the building blocks for OpenSubdiv and its
clients to efficiently process the specific set of supported subdivision
schemes. It is not intended to be a general framework for
defining customized subdivision schemes.
Types, Traits and Options
=========================
The most basic type is the enum *Sdc::SchemeType* that identifies the fixed set of
subdivision schemes supported by OpenSubdiv: *Bilinear*, *Catmark* and *Loop*.
With this alone, we intend to avoid all dynamic casting issues related to the
scheme by simply adding members to the associated subclasses for inspection.
In addition to the type enum itself, a class defining a fixed set of traits
associated with each scheme is provided. While these traits are available as
static methods in the interface of a class supporting more functionality for each
scheme (to be described shortly), the *SchemeTypeTraits* provide queries of the
traits for a variable of type *Sdc::SchemeType* -- enabling parameterization
of code by the value of a trait without templates or virtual inheritance (a
simple internal table of traits is constructed and trivially indexed).
The second contribution is the collection of all variations in one place that can
be applied to the subdivision schemes, i.e. the boundary interpolation rules,
creasing method, edge subdivision choices, etc. The fact that these are all
declared in one place alone should help clients see the full set of variations
that are possible.
A simple Options struct (a set of bitfields) aggregates all of these variations
into a single object (the equivalent of an integer in this case) that are passed
around to other Sdc classes and/or methods and are expected to be used at a higher
level both within OpenSubdiv and externally. By aggregating the options and
passing them around as a group, it allows us to extend the set easily in future
without the need to rewire a lot of interfaces to accommodate the new choice.
Clients can enable new choices at the highest level and be assured that they will
propagate to the lowest level where they are relevant.
Unlike other "options" structs used elsewhere to specify variations of a
particular method, *Sdc::Options* defines all options that affect the shape of
the underlying limit surface of a subdivision mesh. Other operations at higher
levels in the library may have options that approximate the shape and so create
a slightly different appearance, but *Sdc::Options* is a fundamental part of
the definition of the true limit surface.
Creasing support
================
Since the computations involved in the support of semi-sharp creasing are
independent of the subdivision scheme, the goal in Sdc was to encapsulate all
related creasing functionality in a similarly independent manner. Computations
involving sharpness values are also much less dependent on topology -- there
are vertices and edges with sharpness values, but knowledge of faces or boundary
edges is not required, -- so the complexity of topological neighborhoods required
for more scheme-specific functionality is arguably not necessary here.
Creasing computations have been provided as methods defined on a Crease class
that is constructed with a set of Options. Its methods typically take sharpness
values as inputs and compute a corresponding set of sharpness values
as a result. For the "Uniform" creasing method (previously known as *"Normal"*),
the computations may be so trivial as to question whether such an interface is
worth it, but for "Chaikin" or other schemes in the future that are non-trivial,
the benefits should be clear. Functionality is divided between both uniform and
non-uniform, so clients have some control over avoiding unnecessary overhead,
e.g. non-uniform computations typically require neighboring sharpness values
around a vertex, while uniform does not.
Also included as part of the Crease class is the Rule enum -- this indicates if
a vertex is Smooth, Crease, Dart or Corner (referred to as the "mask" in Hbr)
and is a function of the sharpness values at and around a vertex. Knowing the
Rule for a vertex can accelerate mask queries, and the Rule can often be
inferred based on the origin of a vertex (e.g. it originated from the middle of
a face, was the child of a Smooth vertex, etc.).
Methods are defined for the Crease class to:
* subdivide edge and vertex sharpness values
* determine the Rule for a vertex based on incident sharpness values
* determine the transitional weight between two sets of sharpness values
Being all low-level and working directly on sharpness values, it is a client's
responsibility to coordinate the application of any hierarchical crease edits
with their computations.
Similarly, in keeping with this as a low-level interface, values are passed as
primitive arrays. This follows the trend in OpenSubdiv of dealing with data of
various kinds (e.g. weights, component indices, now sharpness values, etc.) in
small contiguous sets of values. In most internal cases we can refer to a set
of values or gather what will typically be a small number of values on the stack
for temporary use.
Scheme-specific support
=======================
While the SchemeTypeTraits class provides traits for each subdivision scheme
supported by OpenSubdiv (i.e. *Bilinear*, *Catmark* and *Loop*), the Scheme class
provides these more directly, Additionally, the Scheme class provides methods
for computing the various sets of weights used to compute new vertices resulting
from subdivision. The collection of weights used to compute
a single vertex at a new subdivision level is typically referred to as a
*"mask"*. The primary purpose of the Scheme class is to provide such masks in a
manner both general and efficient.
Each subdivision scheme has its own values for its masks, and each are provided
as specializations of the template class *Scheme<SchemeType TYPE>*. The intent is to
minimize the amount of code specific to each scheme.
The computation of mask weights for subdivided vertices is the most significant
contribution of Sdc. The use of semi-sharp creasing with each
non-linear subdivision scheme complicates what are otherwise simple
masks determined solely by the topology, and packaging that functionality to
achieve both the generality and efficiency desired has been a challenge.
Mask queries are defined in the *Scheme* class template, which has
specializations for each of the supported subdivision schemes. Mask queries
are defined in terms of interfaces for two template parameters: the first
defining the topological neighborhood of a vertex, and a second defining a
container in which to gather the individual weights:
.. code:: c++
template <typename FACE, typename MASK>
void ComputeFaceVertexMask(FACE const& faceNeighborhood, MASK& faceVertexMask, ...) const;
Each mask query is expected to call methods defined for the **FACE**, **EDGE** or
**VERTEX** classes to obtain the information they require ; typically these
methods are simple queries about the topology and associated sharpness values.
Clients are free to use their own mesh representations to gather the requested
information as quickly as possible, or to cache some subset as member variables
for immediate inline retrieval.
In general, the set of weights for a subdivided vertex is dependent on the following:
* the topology around the parent component from which the vertex originates
* the type of subdivision *Rule* applicable to the parent component
* the type of subdivision *Rule* applicable to the new child vertex
* a transitional weight blending the effect between differing parent and child rules
This seems fairly straight-forward, until we look at some of the dependencies involved:
* the parent *Rule* requires the sharpness values at and around the parent component
* the child *Rule* requires the subdivided sharpness values at and around the new
child vertex (though it can sometimes be trivially inferred from the parent)
* the transitional weight between differing rules requires all parent and child
sharpness values
Clearly the sharpness values are inspected multiple times and so it pays to have
them available for retrieval. Computing them on an as-needed basis may be simple
for uniform creasing, but a non-uniform creasing method requires traversing
topological neighborhoods, and that in addition to the computation itself can be
costly.
The point here is that it is potentially unreasonable to expect to evaluate the
mask weights completely independent of any other consideration. Expecting and
encouraging the client to have subdivided sharpness values first, for use in more
than one place, is therefore recommended.
The complexity of the general case above is also unnecessary for most vertices.
Any client using Sdc typically has more information about the nature of the vertex
being subdivided and much of this can be avoided -- particularly for the smooth
interior case that often dominates. More on that in the details of the Scheme classes.
Given that most of the complexity has been moved into the template parameters for
the mask queries, the Scheme class remains fairly simple. Like the Crease class,
it is instantiated with a set of Options to avoid them cluttering the interface.
It is currently little more than a few methods for the limit and refinement masks
for each vertex type, plus the few fixed traits of the scheme as static methods.
The mask queries have been written in a way that greatly simplifies the
specializations required for each scheme. The generic implementation for both
the edge-vertex and vertex-vertex masks take care of all of the creasing logic,
requiring only a small set of specific masks to be assigned for each Scheme:
smooth and crease masks for an edge-vertex, and smooth, crease and corner masks
for a vertex-vertex. Other than the *Bilinear* case, which will specialize the
mask queries to trivialize them for linear interpolation, the specializations
for each *Scheme* should only require defining this set of masks -- and with
two of them common (edge-vertex crease and vertex-vertex corner) the Catmark
scheme only needs to define three.
The <FACE>, <EDGE> and <VERTEX> interfaces
******************************************
Mask queries require an interface to a topological neighborhood, currently
labeled **FACE**, **EDGE** and **VERTEX**. This naming potentially implies more
generality than intended, as such classes are only expected to provide the
methods required of the mask queries to compute its associated weights. While
all methods must be defined, some may rarely be invoked, and the client has
considerable flexibility in the implementation of these: they can defer some
evaluations lazily until required, or be pro-active and cache information in
member variables for immediate access.
An approach discussed in the past has alluded to iterator classes that clients
would write to traverse their meshes. The mask queries would then be parameterized
in terms of a more general and generic mesh component that would make use of more
general traversal iterators. The advantage here is the iterators are written once,
then traversal is left to the query and only what is necessary is gathered. The
disadvantages are that clients are forced to write these to do anything, getting
them correct and efficient may not be trivial (or possible in some cases), and that
the same data (e.g. subdivided sharpness) may be gathered or computed multiple
times for different purposes.
The other extreme was to gather everything possible required at once, but that is
objectionable. The approach taken here provides a reasonable compromise between
the two. The mask queries ask for exactly what they want, and the provided classes
are expected to deliver it as efficiently as possible. In some cases the client
may already be storing it in a more accessible form and general topological
iteration can be avoided.
The information requested of these classes in the three mask queries is as follows:
For **FACE**:
* the number of incident vertices
For **EDGE**:
* the number of incident faces
* the sharpness value of the parent edge
* the sharpness values of the two child edges
* the number of vertices per incident face
For **VERTEX**:
* the number of incident faces
* the number of incident edges
* the sharpness value of the parent vertex
* the sharpness values for each incident parent edge
* the sharpness value of the child vertex
* the sharpness values for each incident child edge
The latter should not be surprising given the dependencies noted above. There
are also a few more to consider for future use, e.g. whether the **EDGE** or
**VERTEX** is manifold or not. In most cases, additional information can be
provided to the mask queries (i.e. pre-determined Rules), and most of the child
sharpness values are not necessary. The most demanding situation is a
fractional crease that decays to zero -- in which case all parent and child
sharpness values in the neighborhood are required to determine the proper
transitional weight.
The <MASK> interface
********************
Methods dealing with the collections of weights defining a mask are typically
parameterized by a *MASK* template parameter that contains the weights. The set of
mask weights is currently divided into vertex-weights, edge-weights and
face-weights -- consistent with previous usage in OpenSubdiv and providing some
useful correlation between the full set of weights and topology. The
vertex-weights refer to parent vertices incident the parent component from which a
vertex originated, the edge-weights the vertices opposite incident edges of the
parent, and the face-weights the center of incident parent faces. Note the latter
is **NOT** in terms of vertices of the parent but potentially vertices in the child
originating from faces of the parent. This has been done historically in
OpenSubdiv but is finding less use -- particularly when it comes to providing
greater support for the Loop scheme -- and is a point needing attention.
So the mask queries require the following capabilities:
* assign the number of vertex, edge and/or face weights
* retrieve the number of vertex, edge and/or face weights
* assign individual vertex, edge and/or face weights by index
* retrieve individual vertex, edge and/or face weights by index
through a set of methods required of all *MASK* classes. Since the maximum
number of weights is typically known based on the topology, usage within Vtr,
*Far* or *Hbr* is expected to simply define buffers on the stack. Another
option is to utilize pre-allocated tables, partitioned into the three sets
of weights on construction of a *MASK*, and populated by the mask queries.
A potentially useful side-effect of this is that the client can define their
weights to be stored in either single or double-precision. With that
possibility in mind, care was taken within the mask queries to make use of a
declared type in the *MASK* interface (*MASK::Weight*) for intermediate
calculations. Having support for double-precision masks in *Sdc* does enable it
at higher levels in OpenSubdiv if later desired, and that support is made
almost trivial with *MASK* being generic.
It is important to remember here that these masks are being defined consistent
with existing usage within OpenSubdiv: both *Hbr* and the subdivision tables
generated by *Far*. As noted above, the "face weights" correspond to the
centers of incident faces, i.e. vertices on the same level as the vertex for
which the mask is being computed, and not relative to vertices in the parent
level as with the other sets of weights. It is true that the weights can be
translated into a set in terms solely of parent vertices, but in the general
case (i.e. *Catmark* subdivision with non-quads in the base mesh) this requires
additional topological association. In general we would need N-3 weights for
the N-3 vertices between the two incident edges, where N is the number of
vertices of each face (typically 4 even at level 0). Perhaps such a
translation method could be provided on the mask class, with an optional
indication of the incident face topology for the irregular cases. The *Loop*
scheme does not have *"face weights"*, for a vertex-vertex mask, but for an
edge-vertex mask it does require weights associated with the faces incident the
edge -- either the vertex opposite the edge for each triangle, or its center
(which has no other use for Loop).