Optimization & vectorization UU Crowd Simulation Software - - PowerPoint PPT Presentation

▶
optimization vectorization
SMART_READER_LITE
LIVE PREVIEW

Optimization & vectorization UU Crowd Simulation Software - - PowerPoint PPT Presentation

Aug 08, 2023 273 likes •622 views

Optimization & vectorization UU Crowd Simulation Software Roland Geraerts October 14, 2019 Unity3D plugin However Global framework Agents are simulated in parallel using OpenMP Real-time performance UUCS: simulates 15K

slide-1
SLIDE 1

Optimization & vectorization

UU Crowd Simulation Software

Roland Geraerts October 14, 2019

slide-2
SLIDE 2

Unity3D plugin

slide-3
SLIDE 3

However…

  • Global framework

– Agents are simulated in parallel using OpenMP

  • Real-time performance

– UUCS: simulates 15K agents – Unity: animates and visualizes 1.5K agents

slide-4
SLIDE 4

30K in real-time on a fast laptop

slide-5
SLIDE 5

What changed?

  • UUCS

– Made some remaining code run in parallel

  • Unity

– From objected oriented to data driven implementation

  • Entity component system & Job system

– From main thread to separate threads – From CPU animations to GPU shader-based animations

slide-6
SLIDE 6

However…

  • Current optimizations in UUCS

– Theoretical running times: O(…) – Parallel code using OpenMP

  • So much is still possible…

– …but we first need to understand the framework

slide-7
SLIDE 7

can you simulate a human crowd interactively?

How

slide-8
SLIDE 8

Crowd simulation framework

Simulation step Animation Local movement Route following Global route planning High-level planning start/goal positions indicative route preferred velocity velocity Level 1 Level 2 Level 3 Level 4 Level 5
  • Representation environment
  • Level 5

– Plans actions

  • Level 4

– Creates indicative routes

  • Level 3

– Traverses the routes – Yields speed/direction pairs

  • Level 2

– Adapts routes – E.g. to avoid collisions

  • Level 1

– Moves the agents

Representation of the environment
slide-9
SLIDE 9

Crowd simulation framework

Simulation step Animation Local movement Route following Global route planning High-level planning start/goal positions indicative route preferred velocity velocity Level 1 Level 2 Level 3 Level 4 Level 5
  • Representation environment
  • Level 5

– Plans actions

  • Level 4

– Creates indicative routes

  • Level 3

– Traverses the routes – Yields speed/direction pairs

  • Level 2

– Adapts routes – E.g. to avoid collisions

  • Level 1

– Moves the agents

Representation of the environment
slide-10
SLIDE 10

Representation of the environment

  • Computation of walkable areas and navigation mesh

Van Toll et al, 2018: The Medial Axis of a Multi-Layered Environment and its Application as a Navigation Mesh

slide-11
SLIDE 11

Crowd simulation framework

Simulation step Animation Local movement Route following Global route planning High-level planning start/goal positions indicative route preferred velocity velocity Level 1 Level 2 Level 3 Level 4 Level 5
  • Representation environment
  • Level 5

– Plans actions

  • Level 4

– Creates indicative routes

  • Level 3

– Traverses the routes – Yields speed/direction pairs

  • Level 2

– Adapts routes – E.g. to avoid collisions

  • Level 1

– Moves the agents

Representation of the environment
slide-12
SLIDE 12

Level 4: Indicative routes

  • We use the Explicit Corridor Map (ECM)

– Compact navigation mesh – Supports any agent radius – Multi-layered environments – Dynamic updates

slide-13
SLIDE 13

Crowd simulation framework

Simulation step Animation Local movement Route following Global route planning High-level planning start/goal positions indicative route preferred velocity velocity Level 1 Level 2 Level 3 Level 4 Level 5
  • Representation environment
  • Level 5

– Plans actions

  • Level 4

– Creates indicative routes

  • Level 3

– Traverses the routes – Yields speed/direction pairs

  • Level 2

– Adapts routes – E.g. to avoid collisions

  • Level 1

– Moves the agents

Representation of the environment
slide-14
SLIDE 14

Level 3: Path following

  • Smoothly follow a desired path

– Input: indicative route, non-smooth indication of the path – In each simulation step, compute an attraction point – Leads to a preferred velocity for the next level

  • Indicative Route Method (IRM, 2009)
  • MIRAN: improvement

by Jaklin et al. (2013)

– Supports weighted regions – Better smoothness/ shortcut control

slide-15
SLIDE 15

Crowd simulation framework

Simulation step Animation Local movement Route following Global route planning High-level planning start/goal positions indicative route preferred velocity velocity Level 1 Level 2 Level 3 Level 4 Level 5
  • Representation environment
  • Level 5

– Plans actions

  • Level 4

– Creates indicative routes

  • Level 3

– Traverses the routes – Yields speed/direction pairs

  • Level 2

– Adapts routes – E.g. to avoid collisions

  • Level 1

– Moves the agents

Representation of the environment
slide-16
SLIDE 16

Level 2: Local movement

  • Roughly move in the preferred direction, while...

– ...responding to collisions with other characters – ...avoiding future collisions – ...adapting to the surrounding streams of people – ...maintaining social group behavior – etc.

slide-17
SLIDE 17

Crowd simulation framework

  • Representation environment
  • Level 5

– Plans actions

  • Level 4

– Creates indicative routes

  • Level 3

– Traverses the routes – Yields speed/direction pairs

  • Level 2

– Adapts routes – E.g. to avoid collisions

  • Level 1

– Moves the agents

Simulation step Animation Local movement Route following Global route planning High-level planning start/goal positions indicative route preferred velocity velocity Level 1 Level 2 Level 3 Level 4 Level 5 Representation of the environment
slide-18
SLIDE 18

DEMOLITION TIME

The UUCS engine in action

slide-19
SLIDE 19

IMPLEMENTATION DETAILS

Crowd simulation in the UUCS framework

slide-20
SLIDE 20

Simulation step

  • performStep(Δt)

– For each agent: path following

  • Update pointers along indicative route
  • Update attraction point, preferred velocity

– For each agent: collision avoidance

  • Compute new velocity vNew
  • Smoothen vNew (optional)
  • Compute collision forces F (optional)

– For each agent

  • Update velocity: v := vNew + Δt · F/mass
  • Update position: p := p + Δt · v

– Update nearest-neighbor data structure

(There are actually many more substeps) Why separate loops? → Order of agents does not matter → Agents are independent, each loop can be parallellized

slide-21
SLIDE 21

Performance (without visualization)

  • 1 thread (2015)
slide-22
SLIDE 22

Performance (without visualization)

  • 8 threads: 4 cores (2015)
slide-23
SLIDE 23

Assignments

slide-24
SLIDE 24

Collision avoidance 1 / 2

  • Algorithm

– Focus on Optimal Reciprocal Collision Avoidance (ORCA) – Appears to be the most expense part of UUCS

  • Code

– src/Simulation/CollisionAvoidance/CollisionAvoidance_RVO.cpp – 444 lines of code – Code includes solving a linear program

  • Literature

– Paper: http://gamma.cs.unc.edu/RVO/icra2008.pdf – GPU tips: https://arxiv.org/abs/1908.10107 – LP GPU implementation: https://rgb-lp-docs.readthedocs.io/en/latest/

slide-25
SLIDE 25

Collision avoidance 2 / 2

  • Goal

– Optimize CPU code, or – Convert to GPU implementation

  • Performance criterion

– Relative difference in total running time of collision avoidance during 60s (600 frames) in city environment with 25K agents

slide-26
SLIDE 26

KD-tree 1 / 2

  • Algorithm

– NanoFlann – Computes and queries a nearest neighbors KD-tree

  • Code

– src/external/nanoflann/nanoflann.hpp – 1946 lines – Optimized templated C++ code

  • Literature

– https://github.com/jlblancoc/nanoflann

slide-27
SLIDE 27

KD-tree 2 / 2

  • Goal

– Optimize C++ code

  • Performance criterion

– Relative difference in total running time of building the KD-tree and all nearest neighbor queries during 60s (600 frames) in city environment with 25K agents

slide-28
SLIDE 28

UUCS 1 / 2

  • Algorithm

– UUCS codebase

  • Code

– Mainly src/Simulation/* – 10K lines?

  • Literature

– Framework:

https://www.staff.science.uu.nl/~gerae101/UU_crowd_simulation_p ublications_framework.html

– PhD thesis:

https://www.staff.science.uu.nl/~gerae101/pdf/PhD_Thesis_Wouter _van_Toll_Navigation_for_characters_and_crowds_in_complex_virtua l_environments.pdf

slide-29
SLIDE 29

UUCS 2 / 2

  • Goal

– Optimize C++ code

  • Performance criterion

– Relative difference in total running time of the whole simulation during 60s (600 frames) in city environment with 25K agents

October 14, 2019 INFOMCRWS: UU Crowd Simulation Software 29
slide-30
SLIDE 30

Prizes

  • 1. Arduino starter kit
  • 2. Arduino starter kit
  • 3. Arduino starter kit
slide-31
SLIDE 31

Getting started

  • Sign EULA

– Improvements may be integrated in UUCS – IP goes to University so that education and research is secured – ucrowds.com/eula

  • After signing, you will get access to

– UUCS library and demo projects

  • https://git.science.uu.nl/UUCS/explicit-corridor-map-framework
  • To compile the project

– Follow the instructions listed in README.md – You can get some help

slide-32
SLIDE 32

Technical support

  • Compilation

– Geert-Jan Giezeman – g.j.giezeman@uu.nl – BBG 5.77 – Please send him an e-mail first

  • Weekly visit hour

– Monday 10.00 - 11.00 – UtrechtInc, Padualaan 8, Office W125 – Contact Yiran Zhao

  • yiran@ucrowds.com
slide-33
SLIDE 33

Questions

Roland Geraerts R.J.Geraerts@uu.nl uu.nl/staff/RJGeraerts BBG 4.07 06 28 80 49 01