Log	PowerBook Configuration Log xw9300 Opteron Configuration Log

2005 Page 1   Page 2

• Thu Jan 5 17:56:56 EST 2006

  To do:

  • Update Dr. Banks personal laptop to have latest OS patches/compilers/inventor.
  • Dress up Harvard's dataset (turn lineSets to tubes)
  • Pack up opteron.

• Wed Dec 14 14:28:02 EST 2005

  
  1000 Tracts .iv file renderLarge_fr.iv.gz (10 MB)

  Ward's histogram-based adjustment (brightness fix):

  
  Constant radius: 1k, 2k, 4k, 8k fibers

  
  Shrinking radius: 1k, 2k, 4k, 8k fibers

  
  Random seeding vs. grid seeding

  No brightness fix:

  
  Constant radius: 1k, 2k, 4k, 8k fibers

  
  Shrinking radius: 1k, 2k, 4k, 8k fibers

  
  Random seeding vs. grid seeding

  
  Still image of clustered tracts from Harvard LMI

  
  Vortex particle method
  Movie: gif avi

  
  1600x1200 close up of photons on tracts

• Mon Dec 5 02:44:53 EST 2005

  
  1000, 2000, 4000, 8000 fibers, radius constant

  
  1000, 2000, 4000, 8000 fibers, radius shrinking

  
  Grid vs. Random seeding, 1803 fibers both cases

  
  Visualization of 10000, 20000, 40000, 80000, 160000 photons on fibers

  
  Added displacement mapping to subdivision surfaces. Also added support for non-manifolds (meshes with boundaries). Left: original venus mesh, Right: subdivided into approximately 1 million triangles and displaced using Perlin turbulance.

  
  First stab at xmas tree fog. Fog is initialized with cone and perturbed with 12000 vortex particles (Selle et al.), on top of a fractally displaced heightfield.

• Wed Nov 30 19:39:16 EST 2005

  My vortices were askew. Also, having vorticities pointing counter-clockwise (when viewed from above, respective of center of floor, i.e. -i1) seems to produce the most similar results as Selle et al.

  Varying bouyancy again:
  
  

  ./fluid3D -j4 -n60x86x60 -p6000 -e0.01 -m -v0 -b.001  -i1 -f.06 -x3 -d.03 -T2
  ./fluid3D -j4 -n60x86x60 -p6000 -e0.01 -m -v0 -b.0005 -i1 -f.06 -x3 -d.03 -T2
  

  Constant, linear, quadratic, and cubic temperature ramps:
  
  

  ./fluid3D -j4 -n60x86x60 -p6000 -e0.01 -m -v0 -b.0005 -i1 -f.06 -x0 -d.03 -T1
  ./fluid3D -j4 -n60x86x60 -p6000 -e0.01 -m -v0 -b.0005 -i1 -f.06 -x1 -d.03 -T1
  ./fluid3D -j4 -n60x86x60 -p6000 -e0.01 -m -v0 -b.0005 -i1 -f.06 -x2 -d.03 -T1
  ./fluid3D -j4 -n60x86x60 -p6000 -e0.01 -m -v0 -b.0005 -i1 -f.06 -x3 -d.03 -T1
  

  First two show effect of increasing temperature bouyancy scalefactor, last shows increased divergence-driven expansion and no bouyancy:
  
  

  ./fluid3D -j4 -n36x52x36 -p6000 -e0.1 -m -v0 -b.0005 -T1.2 -i1 -f.04 -x0
  ./fluid3D -j4 -n36x52x36 -p6000 -e0.1 -m -v0 -b.001  -T1.2 -i1 -f.04 -x0
  ./fluid3D -j4 -n60x86x60 -p6000 -e0.1 -m -v0 -b0     -T1.8 -i1 -f.06 -x0 -d.03 
  

• Mon Nov 28 20:41:25 EST 2005

  Changed fluid sim to have non-equal dimensions, eg. 128x512x128. In the process I noticed a couple things. First, my viscosity is 0.001. Making it smaller (zero) gives more interesting results. Second, when adding forces, e.g. vorticity confinement, I multiply by dt twice instead of once.

• Mon Nov 28 17:31:10 EST 2005 To-do list:
  • Create figures
  • Make and print out holiday cards
  • Visit Boston (install inventor on solaris)
  • Contact Selle and Fedkiw
  • Organize thesis work, pane
  • Prepare for paper talk
  • Make pane tutorial videos
  • Make fluid simulation web page

• Wed Nov 23 19:09:01 EST 2005

  
  Got blackbody emission to work. Also, I found several constants defined in "Physically Based Modeling and Animation of Fire" by Nguyen at al., including bouyancy, smoke density, etc. Left to right: Smoke simulation with new constants, flame (5000K), flame (50000K), flame together with smoke (flame is scaled by a constant, a hack).

  
  Implemented Perlin noise ("Hypertextures" by Ken Perlin), in both Matlab and C++. Shown are 3D and 2D turbulence (resp.) using Amira and Matlab (resp.).

• Tue Nov 22 00:13:54 EST 2005

  
  Ahhh, seeding all 6000 particles in the initial cubic smoke source area during the divergence driven expansion produces much finer detail, as shown in this animated .gif.

  
  I created a flame area in the plume by adding an orange emissive component proportional to the square of the temperature. It doesn't look that great, but I know how to make it better.

• Wed Nov 16 18:16:57 EST 2005

  
  Figures for abstract.

• Tue Nov 15 14:06:26 EST 2005

  
  
  It wasn't the filtering, it was the number of samples. I doubled the number for the 4D rendering and now they look about the same. Rendering time increased to 2.4 minutes.

  
  3D texture comparison of levelset-extraction rendering (left, 34 minutes) and 4D heightfield rendering (right, 1.1 minutes), using only direct illumination. Most of the difference in the appearance is due to slightly different filtering.

  
  Full GI tracts movie, with motion blur. Some frames pop because of architecture implementation issues (e.g., the Itanium2 computes sqrtf() slightly differently, resulting in a different photon map). Here's an individual frame.

• Sun Nov 13 14:58:31 EST 2005

  
  Results of pane benchmark on hp xw9300, Dell Precision 650, and SGI Prism. In this test the xw9300 is 190% as fast as photon, while the prism is about 88% as fast. The times to render a scene using 4 threads are:

  Machine	Time (s)
  xw9300	917.42
  Dell	1748.08
  SGI	1979.09

• Fri Nov 11 08:56:49 EST 2005

  
  Implemented camera motion blur using QMC sampling from Keller's Strictly Deterministic Sampling sec. 4.2. a) Cornell box b) brain tracts movie, approximate, 1000 frames.

  
  Revamped anti-aliasing code. It's *much* improved. I implemented Keller's QMC anti-aliasing from his course notes, Strictly Deterministic Sampling. This produced about the same results as the tent-filter warped (QMC) point sets from RRT 2nd Ed. that I was using before. What really made a drastic difference is that I now properly filter sub-pixel samples onto all neighboring pixels (using formula 7.3, pg. 351 of PBRT), whereas before sub-pixel values were box filtered only onto the pixel they were inside of. This is a much better reconstruction method and reduces postaliasing (aliasing arising from reconstruction). Instead of a tent filter I'm using a Mitchell-Netravali cubic filter. a) Comparison of warped QMC, box filtered vs. stratified (jittered) pointset, mitchell filtered vs. generalized Hammersly, mitchell filtered, for 16 samples/pixel of a radial sin() function of increasing frequency. b) Comparison of brain tracts image at 4 samples/pixel (enlarged 300%). c) New brain tracts image, 900x900, 4 samples/pixel (compare against previous version).

• Tue Nov 8 15:46:00 EST 2005

  
  1) Test rendering, and 2) full GI movie. 400 frames, 512x512, 8.7 MB. Rendered on heterogenous cluster of about 50 dual Xserves, a dual Opteron, 5 dual Xeons, 6 Pentium4's, and a quad Itanium2, in about three hours. 3) Screenshot of rendered .iv file. .iv files are stored as both a single and as separate .iv files.

  
  New cutaway image. Movie of colored tracts soon.

• Sun Nov 6 03:08:29 EST 2005

  
  New colors for Harvard collaboration. I'm rendering an .iv file that has the same colors stored per-vertex in a triangle mesh, including separate individual .iv mesh files for each color component.

  Also on the way is a new cutaway image. I'm also rendering an emissive streamlines image but I don't think it's going to finish within a week so I'll probably cancel it since I need the machine (hp) for benchmarking.

  
  
  Same isovalue, four surfaces, new colors and slightly different lighting. Note that, despite the file names, the prominent filopodium time frame is chronologically last.

  
  Emissive fMRI surface. Horrible artifacts. Rendering thousands of tubes with complicated BRDF is really slow. Also, it doesn't help to have several hundred lights. Further optimization and careful BRDF picking are needed. Having the tubes be emissive is impractical (it's still rendering) as that creates several thousand lights, and portions of the direct lighting code are linear in the number of lights.

  
  Same isovalue (roughly, based on best guess), different timesteps.

• Sat Nov 5 21:23:18 EST 2005

  
  Overlapping neuron isosurfaces. Same isovalue, with GI using uffizi IBL. Same, with cornell box and less saturated colors. New version coming.

  
  Fiber tracts images from collaboration with Harvard. Third image shows all tracts, with 10/99 having non-blue color, using GI in cornell box.

• Sat Nov 5 01:21:38 EST 2005

  
  (Sigh) This is about the 50th simulation parameter test and the smoke still doesn't billow the same as in the vortex-particles paper. Getting the smoke to behave in a particular way is a nightmare. Rise, or no rise? It's like the dirt particle in LAPR. Second image links to a compilation movie of 56 smoke test-run simulations. 5.6 MB, 4:47 s.

  
  Got Yagi's code to produce an .iv file. Was kind of tricky since his program wasn't reporting some command line option errors (-iosfile instead of -isofile, and -writeIv instead of -writeIV). As a result it was loading the wrong isofile and taking an eternity to run. Now I just need to consolidate the material nodes so I can edit them easily and render some scenes.

• Fri Nov 4 05:44:51 EST 2005

  
  Full global illumination rendering of nested isosurfaces, which is not very good. Spaghetti cutaway with 1000 tubes. The global illumination in provided by a rendered .iv file. That explains why there is no shadow on the floor, since it is two triangles with interpolated vertex shading.

  
  It was difficult to wrangle amira to display the cutaway surfaces properly. The key is to use the pruning features of the oblique slices, and blend transparency mode (I'm not positive latter is crucial).

  
  
  Nested isosurface and approximate global illumination solution using Uffizi light probe. I predict the finished result is not going to look very good and will require code modification. This is because the direct lighting will be computed using caustic photons.

• Wed Nov 2 03:15:40 EST 2005

  
  Optimized the fluid simulation a little bit, including parallelizing the pressure solver with pthreads (only ~10% speedup), using trilinear (instead of cubic) interplation for everything except velocity advection, and explicitly setting more constants as single precision (e.g., 1.0f). Tried three-level bricking from Parker et al.'s Ray Tracing Volume Data but found this to cause a slow down. Found two bugs in my vortex particle method: I was only updating 2 of 3 vortex components (a remnant of the 2D solver) and was not including the gaussian filter constant as a weight. Shown are three tests of the vortex particle method, using eps=10, 100, 1000, with 4000 particles. Resolution is 100x100x100 and step time was about 4.7 seconds on a hp xw9300 dual dual-core opteron system.

• Sun Oct 30 17:49:57 EST 2005

  
  Vortex-particle method test animation. Maybe there is missing fine scale structure? Plus the smoke seems really wispy and thin. It's strange the smoke doesn't ever really sink. Probably some constant. Rendering is with single-scattering only. For some strange reason, g++ still takes 30 seconds to compile, or doesn't work at all.

• Sat Oct 29 01:11:15 EDT 2005

  
  First attempt at 3D fluid simulation. This is just a straight-forward extension of my 2D solver to 3D. This image shows a rising plume, using vorticity confinement. The simulation may be buggy, I'm not sure. Each time-step took about 7 seconds using one processor on an HP opteron xw9300. Resolution is 100x100x100. Rendering time was 6.5 minutes on a Dual Xeon 3.0 GHz. I'm running another simulation using Fedkiw's vortex-particle method for comparison. I don't like the plume's shape, and there seems to be missing smoke at the bottom, these are things I can easily fix, though. The Bresenham line integration artifacts are plainly visible. Maybe jittering the values a little bit will help. (This image was made despite site wide, persistant file system response delays of 30 seconds.)

• Thu Oct 27 16:28:22 EDT 2005

  
  15000 streamlines following eigenvectors of a diffusion-tensor MRI scan of a human brain, layered isosurfaces of a manatee MRI scan, layered isosurfaces of a mouse dendrite filopodium.

  
  Sphere and bumpy sphere exhibiting caustics in a full global illumination simulation of participating media, using photon mapping and ray marching.

• Sat Oct 22 20:32:21 EDT 2005

  
  
  Implemented a 2D fluid solver and visualization based on Jos Stam's FFT solver. I've added texture and particle advection, velocity magnitude visualization, as well as smoke density and temperature advection (with gravity), vorticity confinement (not shown), and the vortex particle method (not shown).

  I'm working on switching to Gauss-Seidel relaxation as used in Stam's "Real-Time Fluid Dynamics for Games" and promoting to 3D.

  
  Implemented single scattering for volumetric grid data for point lights. Non-primary rays are integrated along a line from a 3D Bresenham line-drawing algorithm, which produces mild artifacts such as those in the blow-up of the shadow region shown in the 2nd pic (however it is much faster as there is no interpolation).

  
  Added ability to rotate volume data, added indirect illumination using photon mapping, and added area light support. Render time for the last image is 12.7 minutes on dual 3.0 GHz Xeon.

• Tue Oct 11 20:35:02 EDT 2005

  My log for trying to get the hp xw9300 demo unit up and running

  Installing Open Inventor on x86_64:

  • Edit "libFL/ang/flfreetype.h" to add at the top:

      #include <ft2build.h>
      #include FT_FREETYPE_H
      

  • Use the command `g++ -E -dM empty.c` to determine that "|| __x86_64__" needs to be added to the two lines in inventor/lib/database/include/machine.h "#if __i386__ || __ia64__".
  • `up2date libjpeg-devel`
  • Change "inventor/make/ivcommondefs" to say "X11LIBDIR = $(X11DIR)/lib64" in the usingLinux section.
  • `cd /usr/X11R6/lib64; ln -s libX11.so.6 libX11.so`
  • `ln -s libGLU.so.1 libGLU.so`
  • `up2date openmotif-devel`

• Thu Sep 22 17:45:18 EDT 2005

  
  Rendered Yagi's brain image with 5000 fibers on the hp xw9300 w/8 GB RAM.

• Thu Sep 8 21:17:14 EDT 2005

  Implemented Catmull-Clark subdivision surfaces
  

  Also, I think I can implement advanced surfaces in inventor by having a state (stack of advanced materials) that push/popped using Pre/Post call backs, and inherited by any following SoMaterial nodes. Not having an advanced material results in them inheriting a default initial advanced material state.

• Sat Jul 30 10:38:18 EDT 2005

  I've graduated (M.S.), and I'm going to SIGGRAPH 2005. I'll be in L.A. from August 1-13.
  I'm looking for a job. If you can help, please call my cell: (850)545-5977. Thanks!

• Thu Jul 28 14:56:23 EDT 2005
  • Spellcheck final document
  • Make final changes (Gaussian -> Gaussian-like)
  • Picture for plaque - line art, less arrows?
  • Thesis page - pdf, power point pres, link to pane, log, content, CODE
  • Pane 3D, Pane 4D (both versions), brToRGBA, compareSurf, compareTexture, HDR help (mention radiance in help), surface, modSurface, viewIso (give credit), amira network, sample dat sets, should pointed
  • Make it so that someone from another university can download and compile and run

• Sun Jul 17 20:47:10 EDT 2005

  
  Power loader (by Derrin Proctor) from Aliens in a driveway.

  
  Renderings of the global illumination test scenes.

  
  First stab at the Sponza atrium.

  
  Here's the final poster image, it's 2172x1118 resolution with 72 samples/pixel. It took 7.9 hours on flowspace (Dual 3.0 GHz Xeon w/ 4 GB RAM).

  
  All the environment maps, at 1086x559 resolution.

• Sat Jul 9 02:33:01 EDT 2005

  
  
  Images for my poster

2005 Page 1   Page 2

Kevin Beason /

Home