jadnohra / old_tech_experience

Table of Contents

• 0. Currently
• 1. At Havok
  • a. Nintendo WiiU math library
  • b. Physics contact solver perfromance
  • c. Network physics R&D
  • d. Job queue combinatorial delay mode
  • e. Geometry numerical issues
  • f. TOI Argument
  • g. High fidelity solver
  • h. Misc.
• 2. At Microsoft
• 3. At Guerrila Games
• 4. Before that
• 5. Problem Solving Approach
• 6. Miscellaneous

---

0. Currently

• CV
  • https://github.com/jadnohra/daisy, a traffic simulator
  • https://github.com/jadnohra/PaCoS, a concurrency determinism simulator
  • https://github.com/jadnohra/hinges_py, a rigid-body-dynamics simulator
  • https://github.com/jadnohra/trace, a real-time data plotting tool
  • https://github.com/jadnohra/tag_export
  • https://github.com/jadnohra/TheNotes
• Scenario language
• Map search
• Scenario augmenter design
• Test validation concept

1. At Havok

a. Nintendo WiiU math library

1. Instruction-level ARM (handicapped two-float SIMD)
2. Buggy compiler in terms of performance optimization
   • 

b. Physics-2014 contact solver perfromance

1. Mental models of CPUs at race with CPU improvements

   • Benchmarking ignoring real data becomes dangerous

2. Research (https://www.agner.org/optimize/#manuals) -> branch prediction as main actor in this context

3. Proof using counter example :

4. New (intrusive) optimization : coupling geometry processing to solver using data rearrangement

5. Also, wrote whole new contact solver, critical component in all games shipped since ~2014

   • To get an idea Bullet solver
   • Havok's focus: performance
   • No bugs ever reported
   • Defensive code (tunable defensiveness -> math library only has INFTY and NAN warnings)
   • Found bugs in other pieces of code, including hard to find bugs due to slow accumulation of wrong calculation
   • tunable per-platform using a very carefully configured set of macros (> 10 options -> code path combinations)

6. Deterministic scaling of cores (Intel many-core CPU R&D)

   • 

c. Network physics R&D

1. Research report and proposed designs
   • 
2. Prototype
   • 
   • 
3. Including low-level framework
   • 

d. Job queue combinatorial delay mode

• 1/100000 repro -> found in an hour

e. Geometry numerical issues

• Jira epsilon
  • 
• Ray sphere: improvement by intuitive solution
  • 
• Ray triangle
  • Improvement by analytical solution
    • Books like 'Floating-Point Computation (Strebenz)', resources like '???'
  • Telescoping debugger
  • 
  • 
  • 
  • 
• Led to interest about numerical analysis (Studying, etc.)
  • Sampler prototype (RRT)
    • 
  • Interval math library with pessimistic approximate but fast bounds (2x slower than normal arithmetic)
  • Also first steps into numerical analysis for dynamics : Residuals
    • 

f. TOI Argument

• Open problem ever since the 'TOI' engine was abandoned
• Multiple attempts that replaced artifacts with other artifacts
• Proved that the problem is not solvable: TOI argument
  • 
• Effect: stop searching for the solution, focus on which artifacts to trade and when

g. HKDS

• Overview : HKDS overview
  • 
• A high level view of the approaches
  • 
• More detail
  • 
• Fast forward: videos of results
  • Titan arm video
    • 
  • Digger video
    • 
• Prerequisites that allowed to even start tackling this
  • Rotations and Basics
    • 
    • 
  • Work on stabilizing certain constraints
  • Notebook notes on 'universal constraint'
  • Substepping algorithm Report
    • A stumbling block for others at multiple occasions
    • Local problem:
    • Approximations:
    • Algorithm:
• MLCP: Notebook 'MLCP'
• Some code
  • Prototyping using Julia : Gitlab
    • 
  • MLCP->LCP (error in paper)
  • Co-simulation
• Dynamics debugging tool:
  • Terminal python trace.py -test
  • Trace user manual

h. Misc.

• 
• 

2. At Microsoft

• 
• 
• 

3. At Guerrila Games

1. AI environment-aware physical behavior: Video
   • 
2. Automatic region generation
   1. Porting of recast 'ugly and cryptic' code
      • Improvements of corner cases that other companies presented as challenges during GDC
      • 
   2. Research sekeltonization algorithms
   3. Choose and implement
      • 
   4. Modify to fit our application (stray edges)
      • 
3. Vehicle AI : Video
   • 
   • Prototyping VehicleTest on gitlab
   • 
   • 
4. SPU + PS3 multithreading/DMA tricky bug chasing

4. Before that

1. World of Football
   • On github
     • 
   • Youtube videos
     • 
   • Youtube videos
2. Mocap loader, skelton semantics, footplant detection, UI framework
   • Bigfoot video
     • 
   • Bigeye on gitlab
     • 
   • example: smart pointers 10 years ago
     • 

5. Problem Solving Approach

1. I have too many topics that I care about
2. In the past, I had the luxury to focus one at a time, since some years, I cannot afford this anymore
3. Stack keeps growing bigger
4. Tried many techniques to mitigate this
   1. Physical notebooks
      • Show GK notes and paper
        1. https://gitlab.com/jadnohra/project_gk/blob/master/simp-hing-analyt/simp-hing-analyt.pdf
           • public arxiv.org version: 'Uniqueness of Minima of a Certain Least Squares Problem'
           • 
           • 
        2. https://gitlab.com/jadnohra/project_gk/blob/master/Code/hinge_nlsq_simul.py
        3. https://gitlab.com/jadnohra/project_gk/blob/master/Code/plot_tool.py
        4. https://gitlab.com/jadnohra/project_gk/blob/master/Code/plot_tool_repo.txt
        5. https://gitlab.com/jadnohra/jad-pre-2015-dabblings/blob/master/Lab2015/asinus_salutem/scripts/gk_notes_p45.txt
        6. https://gitlab.com/jadnohra/jad-pre-2015-dabblings/blob/master/Lab2015/asinus_salutem/scripts/test_df.txt
        7. Terminal: python hinges_py2.py -scene chain -help -fancy -wobble_body -fwd_euler -control
   2. Large single-topic latex documents
      1. show PL/QL printed book
   3. Currently: combination of physical notes, latex summaries, and gitlab project management
      1. https://gitlab.com/jadnohra/study/issues/17
      2. https://gitlab.com/jadnohra/study/issues/7
      3. https://gitlab.com/jadnohra/study/issues/13 -> https://gitlab.com/jadnohra/study/issues/15
      4. I also use this for developing all the high-level knowledge I need about SIM & L4+
      5. The biggest challenge is still memorization, I do not do it actively anymore, except when I really have to
         • I try to work more as a 'processing machine', I open my notes, switch context for a while, and process for results
         • The older the topic, the longer the lead time, can take from a couple of hours to a couple of weeks

6. Miscellaneous

• Existing attempts
  • http://chrishecker.com/The_mixed_linear_complementarity_problem
  • https://github.com/thomasmarsh/ODE/blob/df82c09d967d486822ea1715a09291593ba471a6/ode/src/lcp.cpp
  • Overview: http://box2d.org/files/GDC2014/ErwinCoumans_ExploringMLCPSolversAndFeatherstone.pdf
• More notes
  • https://www.dropbox.com/home/LibraryLive/TheNotes2
  • https://www.dropbox.com/home/LibraryLive/TheNotes2?preview=pointillistic.pdf
  • https://www.dropbox.com/home/LibraryLive/TheNotes