Skip to content

Latest commit

 

History

History
130 lines (94 loc) · 4.45 KB

README.md

File metadata and controls

130 lines (94 loc) · 4.45 KB

Partio - A library for particle IO and manipulation

This is the initial source code release of partio a tool we used for particle reading/writing. It started out as an abstraction for the commonalities in particle models (i.e. accessing many attributes associated with an index or entity).

Super impatient building guide

    # Install Location ~ adjust accordingly
    prefix=$HOME/local
    git clone https://github.com/wdas/partio.git
    cd partio
    make -j prefix=$prefix install

Getting Started

CMake is used to build the project, but we provide a top-level Makefile for convenience that takes care of all the steps.

See the Makefile for the user-tweakable variables and corresponding cmake options.

The typical usage for an installation into /usr/local with a temporary staging directory of /tmp/stage is:

make DESTDIR=/tmp/stage prefix=/usr/local install

Source code overview

    src/
       lib/      Library code (public API in root)
       lib/core  Core library (KDtree traversal, data representations)
       lib/io    Input/Output (Different file formats)
       py/       SWIG based python bindings
       doc/      Doxygen documentation and (the start of) a manual
       tests/    Start of regression tests (I need more)
       tools/    Useful tools
                 partconvert <input format> <output format>
                 partinfo <particle file>
                 partview <particle file>

Class Model

The goal of the library is to abstract the particle interface from the data representation. That is why Partio represents particles using three classes that inherit and provide more functionality

ParticlesInfo - Information about # of particles and attributes ParticlesData - Read only access to all particle data ParticlesDataMutable - Read/write access to all particle data

The functions used to get particle access are these:

    readHeaders()
       returns ParticlesInfo
       reads only the minimum data necessary to get number of particles and
       attributes

    readCached()
       returns ParticlesData
       For multiple users in different threads using the same particle file
       ParticlesData

    create() and read()
       returns ParticlesDataMutable
       allows read/write access

Behind the scenes you could implement these classes however you like. Headers only representation is called core/ParticleHeader.{h,cpp}. Simple non-interleaved attributes is core/ParticleSimple.{h,cpp}.

Attribute Data Model

All particles have the same data attributes. They have the model that they are of three basic types with a count of how many scalar values they have.

    VECTOR[3]
    FLOAT[d]
    INT[d]

    VECTOR[3] and FLOAT[3] have the same data representations.
    VECTOR[4] is invalid however FLOAT[4] is valid as is FLOAT[1...infinity]

This seems to encompass the most common file formats for particles

Iterating

There are multiple ways to access data in the API. Here are some tips

  • Use SIMD functions when possible prefer dataAsFloat(),data(arrayOfIndices) as opposed to data(int singleIndex) which accesses multiple pieces of data at once

  • Cache ParticleAttributes for quick access instead of calling attributeInfo() over a loop of particles

  • Use iterators to do linear operations over all particles They are much more optimized than both data() and the dataAsFloat or

Backends

Behind the scenes there are SimpleParticles, ParticleHeaders, and SimpleParticlesInterleaved. In the future I would like to write a disk-based cached back end that can dynamically only load the data that is necessary. create(), read() and readCached could be augmented to create different structures in these cases.

Readers/Writers

New readers and writers can be added in the io/ directory. You simply need to implement the interface ParticlesInfo, ParticlesData and ParticlesDataMutable (or as many as you need). Editing the io/readers.h to add prototypes and io/ParticleIO.cpp to add file extension bindings should be easy.

Building the python Package for PyPi

To the partio for python and publish it to we have to build it using docker and upload it to PyPi.

# build the docker
  docker build -t partio:latest .
  # run the build
  docker run --rm -v $(pwd):/io partio:latest
  # use twine to upload to pypi
  twine upload dist/*
  • Andrew Selle, Walt Disney Animation Studios