OpenSubdiv/documentation/far_overview.rst
David G Yu 1294e1b0c1 Updated far_overview to include PrimvarRefiner
This updates much of the initial overview documentation
to make a clearer distinction between topology refinement
and primvar refinement and describes the different classes
and methods that are used.
2015-06-19 16:36:10 -07:00

586 lines
26 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.
FAR Overview
------------
.. contents::
:local:
:backlinks: none
.. image:: images/api_layers_3_0.png
:width: 100px
:target: images/api_layers_3_0.png
Feature Adaptive Representation (Far)
=====================================
*Far* is the primary API layer for processing client-supplied mesh data
into subdivided surfaces.
The *Far* interface may be used directly and also may be used to prepare
mesh data for further processing by `Osd <osd_overview.html#Osd>`__.
The two main aspects of the subdivision process are *Topology Refinement*
and *Primvar Refinement*.
Topology Refinement
*******************
Topology refinement is the process of splitting the mesh topology
according to the specified subdivison rules to generate new topological
vertices, edges, and faces. This process is purely topological and does
not depend on the speciific values of any primvar data (point positions, etc).
Topology refinement can be either uniform or adaptive, where extraordinary
features are automatically isolated (see `feature adaptive subdivision <subdivision_surfaces.html#feature-adaptive-subdivision>`__).
The *Far* topology classes present a public interface for the refinement
functionality provided in `Vtr <vtr_overview.html#Vtr>`__,
The main classes in *Far* related to topology refinement are:
+-------------------------------+---------------------------------------------+
| TopologyRefiner | A class encapsulating mesh refinement. |
+-------------------------------+---------------------------------------------+
| TopologyLevel | A class representing one level of |
| | refinement within a TopologyRefiner. |
+-------------------------------+---------------------------------------------+
| TopologyRefinerFactory<MESH> | A factory class template specialized in |
| | terms of the application's mesh |
| | representation used to construct |
| | TopologyRefiner instances. |
+-------------------------------+---------------------------------------------+
Primvar Refinement
******************
Primvar refinement is the process of computing values for primvar data (points,
colors, normals, texture coordinates, etc) by applying weights determined by
the specified subdivision rules. There are many advantages gained by
distinguishing between topology refinement and primvar interpolation
including the ability to apply a single static topological refinement to
multiple primvar instances or to different animated primvar time samples.
*Far* supports methods to refine primvar data at the locations of topological
vertices and at arbitrary locations on the subdivision limit surface.
The main classes in *Far* related to primvar refinement are:
+-----------------------+--------------------------------------------------+
| PrimvarRefiner | A class implementing refinement of primvar data |
| | at the locations of topological vertices. |
+-----------------------+--------------------------------------------------+
| PatchTable | A representation of the refined surface topology |
| | that can be used for efficient evaluation of |
| | primvar data at arbitrary locations. |
+-----------------------+--------------------------------------------------+
| StencilTable | A representation of refinement weights suitable |
| | for efficient parallel processing of primvar |
| | refinement. |
+-----------------------+--------------------------------------------------+
| LimitStencilTable | A representation of refinement weights suitable |
| | for efficient parallel processing of primvar |
| | refinement at arbitrary limit surface locations. |
+-----------------------+--------------------------------------------------+
Far::TopologyRefiner
====================
TopologyRefiner is the building block for many other useful classes in
*Far*. It performs refinement of an arbitrary mesh and provides access to
the refined mesh topology. It can be used for primvar refinement directly
using PrimvarRefiner or indirectly by being used to create
`stencil table <#patch-table>`__, `patch table <#patch-table>`__, etc.
TopologyRefiner has public the refinement methods
*RefineUniform()* and *RefineAdapative()* which perform refinement
operations using Vtr and provides access to the refined topology via
TopologyLevel instances.
.. image:: images/topology_refiner.png
:align: center
Far::TopologyRefinerFactory
===========================
Consistent with other classes in Far, instances of TopologyRefiner are created
by a factory class -- in this case Far::TopologyRefinerFactory. This class
is an important entry point for clients its task is to map/convert data in a
client's mesh into the internal `Vtr <vtr_overview.html>`__ representation as
quickly as possible.
The TopologyRefinerFactory class is a class template parameterized by and
specialized for the client's mesh class, i.e. TopologyRefinerFactory<MESH>.
Since a client' mesh representation knows best how to identify the topological
neighborhoods required, no generic implementation would provide the most
direct means of conversion possible, and so we rely on specialization. For
situations where mesh data is not defined in a boundary representation, a
simple container for raw mesh data is provided (TopologyDescriptor) along
with a Factory specialized to construct TopologyRefiners from it.
So there are two ways to create TopologyRefiners:
* use the existing TopologyRefinerFactory<TopologyDescriptor> with a
populated instance of TopologyDescriptor
* specialize TopologyRefinerFactory<class MESH> for more efficient
conversion
TopologyDescriptor is a simple struct with pointers to raw mesh data in a
form common to mesh constructors. Topologically, the minimal requirement
consists of:
* the number of vertices and faces of the mesh
* an array containing the number of vertices per face
* an array containing the vertices assigned to each face
These last two define one of the six topological relations that are needed
internally by Vtr, but this one relation is sufficient to construct the rest.
Additional members are available to assign sharpness values per edge and/or
vertex, hole tags to faces, or to define multiple sets (channels) of
face-varying data.
Specialization of TopologyRefinerFactory<class MESH> should be done with care
as the goal here is to maximize the performance of the conversion and so
minimize overhead due to runtime validation. The template provides the
high-level construction of the required topology vectors of the underlying
Vtr. It requires the specification/specialization of two methods with the
following purpose:
* specify the sizes of topological data so that vectors can be pre-allocated
* assign the topological data to the newly allocated vectors
As noted above, the assumption here is that the client's boundary-rep knows best
how to retrieve the data that we require most efficiently. After the factory class
gathers sizing information and allocates appropriate memory, the factory provides
the client with locations of the appropriate tables to be populated (using the
same `Array <vtr_overview.html#arry-type>`__ classes and interface used to access
the tables). The client is expected to load a complete topological description
along with additional optional data, i.e.:
* the six topological relations required by Vtr, oriented when manifold
* sharpness values for edges and/or vertices (optional)
* additional tags related to the components, e.g. holes (optional)
* values-per-face for face-varying channels (optional)
While there is plenty of opportunity for user error here, that is no different
from any other conversion process. Given that Far controls the construction
process through the Factory class, we do have ample opportunity to insert
runtime validation, and to vary that level of validation at any time on an
instance of the Factory.
A common base class has been created for the factory class, i.e.:
.. code:: c++
template <class MESH>
class TopologyRefinerFactory : public TopologyRefinerFactoryBase
both to provide common code independent of <MESH> and also potentially to
protect core code from unwanted specialization.
Far::PrimvarRefiner
===================
PrimvarRefiner supports refinement of arbitrary primvar data at the locations
of topological vertices. A PrimvarRefiner accesses topology data directly
from a TopologyRefiner.
Different methods are provided to support three different classes of primvar
interpolation. These methods may be used to refine primvar data to a specified
refinement level.
+-------------------------------+-------------------------------------------+
| Interpolate(...) | Interpolate using vertex weights |
+-------------------------------+-------------------------------------------+
| InterpolateVarying(...) | Interpolate using linear weights |
+-------------------------------+-------------------------------------------+
| InterpolateFaceVarying(...) | Interpolate using face-varying weights |
+-------------------------------+-------------------------------------------+
Additional methods allow primvar data to be interpolated to the final limit
surface including the calculation of first derivative tangents.
+-------------------------------+-------------------------------------------+
| Limit(dst) | Interpolate to the limit surface using |
| | vertex weights |
+-------------------------------+-------------------------------------------+
| Limit(dst, dstTan1, dstTan2) | Interpolate including first derivatives |
| | to the limit surface using vertex weights |
+-------------------------------+-------------------------------------------+
| LimitFaceVarying(...) | Interpolate to the limit surface using |
| | face-varying weights |
+-------------------------------+-------------------------------------------+
PrimarRefiner provides a straightforward interface for refining primvar data,
but depending on the application use case, it can be more efficient to create
and use a `StencilTable <#patch-table>`__, or `PatchTable <#patch-table>`__,
to refine primvar data.
Far::PatchTable
================
The patch table is a serialized topology representation. This container is
generated using *Far::PatchTableFactory* from an instance
*Far::TopologyRefiner* after a refinement has been applied. The
FarPatchTableFactory traverses the data-structures of the TopologyRefiner and
serializes the sub-faces into collections of bi-linear and bi-cubic patches as
dictated by the refinement mode (uniform or adaptive). The patches are then
sorted into arrays based on their types.
.. container:: notebox
**Release Notes (3.0.0)**
The organization and API of Far::PatchTable is likely to change
in the 3.1 release to accommodate additional functionality including:
smooth face-varying interpolation on patches, and dynamic feature
adaptive isolation (DFAS), and patch evaluation of Loop subdivision
surfaces.
Patch Arrays
************
The patch table is a collection of control vertex indices. Meshes are decomposed
into a collection of patches, which can be of different types. Each type
has different requirements for the internal organization of its
control-vertices. A PatchArray contains a sequence of multiple patches that
share a common set of attributes.
While all patches in a PatchArray will have the same type, each patch in the
array is associated with a distinct *PatchParam* which specifies additional
information about the individual patch.
Each PatchArray contains a patch *Descriptor* that provides the fundamental
description of the patches in the array.
The PatchArray *ArrayRange* provides the indices necessary to track the records
of individual patches in the table.
.. image:: images/far_patchtables.png
:align: center
:target: images/far_patchtables.png
Patch Types
***********
The following are the different patch types that can be represented in the
PatchTable:
+---------------------+------+---------------------------------------------+
| Patch Type | #CVs | Description |
+=====================+======+=============================================+
| NON_PATCH | n/a | *"Undefined"* patch type |
+---------------------+------+---------------------------------------------+
| POINTS | 1 | Points : useful for cage drawing |
+---------------------+------+---------------------------------------------+
| LINES | 2 | Lines : useful for cage drawing |
+---------------------+------+---------------------------------------------+
| QUADS | 4 | Bi-linear quads-only patches |
+---------------------+------+---------------------------------------------+
| TRIANGLES | 3 | Bi-linear triangles-only mesh |
+---------------------+------+---------------------------------------------+
| LOOP | n/a | Loop patch (currently unsupported) |
+---------------------+------+---------------------------------------------+
| REGULAR | 16 | B-spline Basis patches |
+---------------------+------+---------------------------------------------+
| GREGORY | 4 | Legacy Gregory patches |
+---------------------+------+---------------------------------------------+
| GREGORY_BOUNDARY | 4 | Legacy Gregory Boundary patches |
+---------------------+------+---------------------------------------------+
| GREGORY_BASIS | 20 | Gregory Basis patches |
+---------------------+------+---------------------------------------------+
The type of a patch dictates the number of control vertices expected in the
table as well as the method used to evaluate values.
Patch Parameterization
**********************
Each patch represents a specific portion of the parametric space of the
coarse topological face identified by the PatchParam FaceId. As topological
refinement progresses through successive levels, each resulting patch
corresponds to a smaller and smaller subdomain of the face.
The PatchParam UV origin describes the mapping from the uv domain of the
patch to the uv subdomain of the topological face. We encode this uv
origin using log2 integer values for compactness and efficiency.
It is important to note that this uv parameterization is the intrinsic
parameterization within a given patch or coarse face and is distinct
from any client specified face-varying channel data.
Patches which result from irregular coarse faces (non-quad faces in the
Catmark scheme, or non-trianglular faces in the Loop scheme) are offset
by the one additional level needed to "quadrangulate" or "triangulate"
the irregular face.
.. image:: images/far_patchUV.png
:align: center
:target: images/far_patchUV.png
A patch along an interpolated boundary edge is supported by an incomplete
sets of control vertices. For consistency, patches in the PatchTable always
have a full set of control vertex indices and the PatchParam Boundary bitmask
identifies which control vertices are incomplete (the incomplete control
vertex indices are assigned values which duplicate the first valid index).
Each bit in the boundary bitmask corresponds to one edge of the patch
starting from the edge from the first vertex and continuing around the
patch. With feature adaptive refinement, regular B-spline basis patches
along interpolated boundaries will fall into one of the eight cases
(four boundary and four corner) illustrated below:
.. image:: images/far_patchBoundary.png
:align: center
:target: images/far_patchBoundary.png
Transition edges occur during feature adaptive refinement where a patch
at one level of refinement is adjacent to pairs of patches at the next
level of refinement. These T-junctions do not pose a problem when evaluating
primvar data on patches, but they must be taken into consideration when
tessellating patches (e.g. while drawing) in order to avoid cracks.
The PatchParam Transition bitmask identifies the transition edges of
a patch. Each bit in the bitmask corresponds to one edge of the patch
just like the encoding of boundary edges.
After refining an arbitrary mesh, any of the 16 possible transition edge
configurations might occur. The method of handling transition edges is
delegated to patch drawing code.
.. image:: images/far_patchTransition.png
:align: center
:target: images/far_patchTransition.png
Single-Crease Patches
**************************
Using single-crease patches allows a mesh with creases to be represented
with many fewer patches than would be needed otherwise. A single-crease
patch is a variation of a regular BSpline patch with one additional crease
sharpness parameter.
.. container:: notebox
**Release Notes (3.0.0)**
Currently, the crease sharpness parameter is encoded as a separate
PatchArray within the PatchTable. This parameter may be combined
with the other PatchParam values in future releases. Also, evaluation
of single-crease patches is currently only implemented for OSD patch
drawing, but we expect to implement support in all of the evaluation
code paths for future releases.
Local Points
************
The control vertices represented by a PatchTable are primarily refined points,
i.e. points which result from applying the subdivision scheme uniformly or
adaptively to the points of the coarse mesh. However, the final patches
generated from irregular faces, e.g. patches incident on an extraordinary
vertex might have a representation which requires additional local points.
.. container:: notebox
**Release Notes (3.0.0)**
Currently, representations which require local points also require
the use of a StencilTable to compute the values of local points.
This requirement, as well as the rest of the API related to local
points may change in future releases.
Legacy Gregory Patches
**********************
Using Gregory patches to approximate the surface at the final patches
generated from irregular faces is an alternative representation which does
not require any additional local points to be computed. Instead, when
Legacy Gregory patches are used, the PatchTable must also have an alternative
representation of the mesh topology encoded as a vertex valence table
and a quad offsets table.
.. container:: notebox
**Release Notes (3.0.0)**
The encoding and support for Legacy Gregory patches may change
in future releases. The current encoding of the vertex valence
and quad offsets tables may be prohibitively expensive for some
use cases.
Far::StencilTable
==================
The base container for stencil data is the StencilTable class. As with most
other Far entities, it has an associated StencilTableFactory that requires a
TopologyRefiner:
Advantages
**********
Stencils are used to factorize the interpolation calculations that subdivision
schema apply to vertices of smooth surfaces. If the topology being subdivided
remains constant, factorizing the subdivision weights into stencils during a
pre-compute pass yields substantial amortizations at run-time when re-posing
the control cage.
Factorizing the subdivision weights also allows to express each subdivided
vertex as a weighted sum of vertices from the control cage. This step effectively
removes any data inter-dependency between subdivided vertices : the computations
of subdivision interpolation can be applied to each vertex in parallel without
any barriers or constraint. The Osd::Compute module leverages these properties
on massively parallel GPU architectures to great effect.
.. image:: images/far_stencil5.png
:align: center
Principles
**********
Iterative subdivision algorithms converge towards the limit surface by
successively refining the vertices of the coarse control cage. Each successive
iteration interpolates the new vertices by applying polynomial weights to a
*basis of supporting vertices*.
The interpolation calculations for any given vertex can be broken down into
sequences of multiply-add operations applied to the supporting vertices.
Stencil table encodes a factorization of these weighted sums : each stencils is
created by combining the list of control vertices from the 1-ring.
With iterative subdivision, each refinement step is dependent upon the previous
subdivision step being completed, and a substantial number of steps may be
required in order approximate the limit : each subdivision step incurs an
O(4\ :superscript:`n`) growing amount of computations.
Instead, once the weights of the contributing coarse control vertices for a
given refined vertex have been factorized, it is possible to apply the stencil
and directly obtain the interpolated vertex data without having to process the
data for the intermediate refinement levels.
.. image:: images/far_stencil7.png
:align: center
Cascading Stencils
******************
Client-code can control the amount of factorization of the stencils : the tables can be
generated with contributions all the way from a basis of coarse vertices, or
reduced only to contributions from vertices from the previous level of
refinement.
The latter mode allows client-code to access and insert modifications to the
vertex data at set refinement levels (see `hierarchical vertex edits
<subdivision_surfaces.html#hierarchical-edits>`_). Once the edits have been
applied by the client-code, another set of stencils can be used to smoothe the
vertex data to a higher level of refinement.
.. image:: images/far_stencil8.png
:align: center
See implementation details, see the Far cascading stencil `tutorial
<tutorials.html>`_
Limit Stencils
**************
Stencil tables can be trivially extended from discrete subdivided vertices to
arbitrary locations on the limit surface. Aside from extraordinary points, every
location on the limit surface can be expressed as a closed-form weighted
average of a set of coarse control vertices from the 1-ring surrounding the
face.
The weight accumulation process is similar : the control cage is adaptively
subdivided around extraordinary locations. A stencil is then generated for each
limit location simply by factorizing the bi-cubic Bspline patch weights over
those of the contributing basis of control-vertices.
The use of bi-cubic patches also allows the accumulation of analytical
derivatives, so limit stencils carry a set of weights for tangent vectors.
.. image:: images/far_stencil0.png
:align: center
Once the stencil table has been generated, limit stencils are the most direct
and efficient method of evaluation of specific locations on the limit of a
subdivision surface, starting from the coarse vertices of the control cage.
Also: just as discrete stencils, limit stencils that are factorized from coarse
control vertices do not have inter-dependencies and can be evaluated in
parallel.
For implementation details, see the `glStencilViewer <glStencilViewer.html>`_
code example.
Sample Location On Extraordinary Faces
**************************************
Each stencil is associated with a singular parametric location on the coarse
mesh. The parametric location is defined as face location and local [0.0 - 1.0]
(u,v) triplet:
In the case of face that are not quads, a parametric sub-face quadrant needs to
be identified. This can be done either explicitly or implicitly by using the
unique ptex face indices for instance.
.. image:: images/far_stencil6.png
:align: center
Code example
************
When the control vertices (controlPoints) move in space, the limit locations can
be very efficiently recomputed simply by applying the blending weights to the
series of coarse control vertices:
.. code:: c++
class StencilType {
public:
void Clear() {
memset( &x, 0, sizeof(StencilType));
}
void AddWithWeight( StencilType const & cv, float weight ) {
x += cv.x * weight;
y += cv.y * weight;
z += cv.z * weight;
}
float x,y,z;
};
std::vector<StencilType> controlPoints,
points,
utan,
vtan;
// Update points by applying stencils
controlStencils.UpdateValues<StencilType>( reinterpret_cast<StencilType const *>(
&controlPoints[0]), &points[0] );
// Update tangents by applying derivative stencils
controlStencils.UpdateDerivs<StencilType>( reinterpret_cast<StencilType const *>(
&controlPoints[0]), &utan[0], &vtan[0] );