Archive for the ‘code’ Category

Using glfw background window in matlab mex thread

Monday, April 3rd, 2017

A while ago, I tried to get glfw playing nicely with matlab’s mex files. I didn’t quite succeed.

My current project requires “GP-GPU” programming. I’m currently using glfw’s “background window” feature to create an OpenGL context. My first tests show that matlab’s mex will play nicely if glfwTerminate() is never called:

#include <igl/opengl/glfw/background_window.h>
#include <mex.h>
void mexFunction(int nlhs, mxArray *plhs[], int nrhs, const mxArray *prhs[])
  GLFWwindow * window;

You can compile this (on mac os x) with:

mex( ...
  '-largeArrayDims','-DMEX','CXXFLAGS=$CXXFLAGS -std=c++11', ...
  'LDFLAGS=\$LDFLAGS -framework Foundation -framework AppKit -framework Accelerate -framework OpenGL -framework AppKit -framework Carbon -framework QuartzCore -framework IOKit ', ...
  '-I/usr/local/igl/libigl//external/nanogui/ext/eigen/', ...
  '-I/usr/local/igl/libigl/include', ...
  '-I/usr/local/igl/libigl/external/nanogui/ext/glfw/include/', ...
  '-L/usr/local/igl/libigl/lib/','-lglfw3', ...

I don’t know why including glfwTerminate() causes matlab to sometimes crash. The error report is impossible to read and seems to have something to do with Quartz.

Offset surface of triangle mesh in matlab

Wednesday, March 22nd, 2017

Here’s a little demonstration of how to use gptoolbox and MATLAB to generate an offset surfaces of a triangle mesh. This takes a mesh in V,F and creates a mesh SV,SF of the isosurface at signed distance iso:

% Extract offset at minus 3% of bounind box diagonal length
iso = -0.03;
% Resolution grid → resolution of extracted offset surface
side = 60;
% Amount of smoothing to apply to distance field
sigma = 1.4;
bbd = norm(max(V)-min(V));
% Pad generously for positive iso values 
[BC,side,r] = voxel_grid([V;max(V)+iso*1;min(V)-iso*1],side);
D = signed_distance(BC,V,F);
D = reshape(D,side([2 1 3]));
% Smooth signed distance field
D = imfilter(D,fspecial('gaussian',9,sigma),'replicate');
BC3 = reshape(BC,[side([2 1 3]) 3]);
% Use matlab's built-in marching cubes iso-surface mesher (works most of the time)
surf = isosurface(BC3(:,:,:,1),BC3(:,:,:,2),BC3(:,:,:,3),D,iso*bbd);
SV = surf.vertices;
SF = surf.faces;

Here’s a blue bunny with a positive offset surface, an orange “cage”:

bunny with positive offset surface

Here’s a blue bunny with a negative offset surface. This is useful for hollowing out objects to 3d print:

bunny with positive offset surface

Because the iso-surface extraction will over tesselate low curvature patches of the output surface, it would make a lot of sense to remesh/decimate this mesh.

(to create these fancy renderings:)

hold on;
t  = tsurf(F,V,'EdgeColor','none',fsoft,  'FaceVertexCData',repmat(blue,size(V,1),1),'FaceAlpha',1+(iso<0)*(0.35-1),fphong);
ts = tsurf(SF,SV,'EdgeAlpha',0.2+(iso<0)*(0-0.2),fsoft,'FaceVertexCData',repmat(orange,size(SV,1),1),fphong,'FaceAlpha',1+(iso>0)*(0.2-1));
hold off;
axis equal;
t.SpecularStrength = 0.04;
l = light('Position',[5 -5 10],'Style','local');
add_shadow(t,l,'Color',0.8*[1 1 1],'Fade','local','Ground',[0 0 -1 min([V(:,3);SV(:,3)])]);
set(gca,'pos',[0 0 1 1])

Quad meshing in matlab using gptoolbox

Friday, February 17th, 2017

This morning I played around with a very simple way of quad meshing a triangle mesh. This is more or less the “ideal pipeline” for quad meshing via parameterization. It will likely only work for very simple meshes. I’m pleasantly surprised that it works at all.

% Load a mesh with boundary and no topological (donut) holes
[V,F] = load_mesh('~/Dropbox/models/');
V = V*axisangle2matrix([1 0 0],-pi/2);
% Construct the LSCM matrix
[~,Q] = lscm(V,F,[],[]);
M = massmatrix(V,F);
% Use Fiedler vector as parameterization
[EV,ED] = eigs(Q,repdiag(M,2),5,'sm');
U = reshape(EV(:,end-3),[],2);
% try to find "canonical" rotation
[sU,sS,sV] = svd(U'*U);
U = U*sU;
% Lay down a grid in 2D
h = ((max(U(:,1))-min(U(:,1)))/(160-2));
[X,Y] = meshgrid(min(U(:,1))-h:h:max(U(:,1))+h,min(U(:,2))-h:h:max(U(:,2))+h);
% Find all quads that are strictly inside the parameterized mesh
[QQ,QU] = surf2patch(X,Y,0*Y);QU = QU(:,1:2);
[sqrD,I,C] = signed_distance([QU QU(:,1)*0],[U U(:,1)*0],F);
QQ = QQ(all(abs(sqrD(QQ))<1e-10,2),:);
[QU,I] = remove_unreferenced(QU,QQ);
QQ = fliplr(I(QQ));
% Snap boundary vertices of quad mesh to boundary of mesh, and
% solve little Dirichlet problem to diffuse the displacements smoothly
QF = [QQ(:,[1 2 3]);QQ(:,[1 3 4])];
L = cotmatrix(QU,QF);
b = unique(outline(QQ));
[~,~,QU(b,:)] = point_mesh_squared_distance(QU(b,:),U,outline(F));
QU = min_quad_with_fixed(-L,[],b,QU(b,:));
% Locate each quad mesh vertex in the parameterized mesh
[sqrD,I,C] = signed_distance([QU QU(:,1)*0],[U U(:,1)*0],F);
B = barycentric_coordinates(C(:,1:2),U(F(I,1),:),U(F(I,2),:),U(F(I,3),:));
% Interpolate 3D positions at quad mesh vertex locations
QV = V(F(I,1),:).*B(:,1) + V(F(I,2),:).*B(:,2) + V(F(I,3),:).*B(:,3);

To create the image above use:

hold on;
tq = tsurf(QQ,QV,'FaceVertexCData',repmat(blue,size(QV,1),1),'EdgeColor','none',fsoft,fphong);
tt = tsurf(F,V-[0.8 0 0],'FaceVertexCData',repmat(1-0.8*(1-blue),size(V,1),1),'EdgeAlpha',0.5,fsoft,fphong);
axis equal;
plot_edges(QV,[QQ(:,2:3);QQ(:,[4 1])],'Color',orange);
plot_edges(QV,[QQ(:,1:2);QQ(:,3:4)],'w');hold off;
l = light('Position',[-10 0 10],'Style','infinite');
sq = add_shadow(tq,l,'Fade','infinite','Color',[0.8 0.8 0.8]);
st = add_shadow(tt,l,'Fade','infinite','Color',[0.8 0.8 0.8]);

There’s a lot to improve about the quad mesh: the boundaries are not aligned with the meshing direction, the meshing directions are not always aligned with principle curvature, etc. Not bad for 30 mins of coding using tools that are just lying around.

Using git and github for course assignments

Tuesday, January 10th, 2017

Here’s my tested plan for using git and github to manage student assignments for my upcoming course.

Initial setup as the instructor:

We’re going to create a student repo and an instructor fork for each assignment. The motivation for separating assignments is to ensure that the diffs that students are submitting for their solutions are tidy. In my case, there’s not much shared code between assignments

Create public* repo that students will see: test-gp-hw0-student.

If you try to fork this on Github, it will tell you that you already have a fork. So instead trick github into forking it via the github importer, import test-gp-hw0-student as a private test-gp-hw0-instructor. Now you have two repos. We’ll set up a branch on test-gp-hw0-instructor to track test-gp-hw0-student for easy merging:

git clone$INSTRUCTOR_REPO.git
git remote add student$STUDENT_REPO.git
git fetch student
git checkout -b student-tracker --track student/master
git checkout master
cd ..

To test this out, let’s create a file visible to everyone (students and instructors) in the student repo:

git clone$STUDENT_REPO.git
echo "visible to everyone" >>
git add
git commit -m "add file visible to everyone"
git push
cd ../

Now, let’s merge change into the instructor repo

git checkout master
git pull student master
cd ..

Just to be thorough, let’s try the reverse: creating a file just visible to instructors:

echo "visible to just instructors" >>
git add
git commit -m "add file visible to just instructors"
git push
cd ..

Getting assignment and handing in assignment as the (dummy) student

Open an incognito/private window and create a new github account for a fake student (pro tip: gmail lets you put dots anywhere in your email address so that redirects to, but most websites don’t recognize that these are the same). My fake students name is stewartdent.

Clone the repository and try to access the instructor repo (assuming the url was leaked):

git clone$STUDENT_REPO.git stewartdent
cd stewartdent
git remote add instructor$INSTRUCTOR_REPO.git
git fetch instructor

You should see an error:

remote: Repository not found.
fatal: repository '' not found

Now, let’s act as the student to submit the homework

# cd stewartdent
echo "stewartdent was here" >>
git add
git commit -m "Stewart Dent's Assignment 01 submission"
git push
cd ..

The student has pushed their submission onto their own repo. They can continue to work on it using git in all its glory. When they’re down they “hand in” their homework via a pull request onto the public test-gp-hw0-student repo. The student does this by visiting the Github page for their own fork of test-gp-hw0-student and clicking “New Pull Request”.

Grading as the instructor

The instructor can now grade this. I’m imagining that the test-gp-hw0-instructor repo has a solution that can be compared (in some way or another) to the students work. To pull the students work into the instructor copy (but never merge):

git checkout -b stewartdent-master master
git pull master
# Record grade.
git checkout master
git branch -D stewartdent-master

Then the instructor can close the pull request on test-gp-hw0-student.

Other perks

Questions or discussions related to the homework can be conducted on the “issues” page of test-gp-hw0-student

* The fact that this repo is public should immediately indicate that student’s work submitted via pull request will also be public and importantly will be visible to other students. Unlike the most common form of cheating, this means that Student A can copy the work of Student B without the explicit consent of Student B. Student B’s only possible defense against this would be to submit their work just before the deadline. As far as I know, I can’t hide pull requests from the public (that would be a nice solution) and I prefer not to use any method that relies on student’s having to create a private fork (since they cost money; although, other class require very over-priced textbooks, so a 6-month github subscription is not so crazy, but it does go against the open-source spirit; there’s also the Big Tobacco style get’em while their young

How does matlab’s hypot function work?

Saturday, January 7th, 2017

The hypot function in matlab purports to be more numerically stable at computing the hypotenuse of a (right-)triangle in 2D. The example the help hypot gives is:

a = 3*[1e300 1e-300];
b = 4*[1e300 1e-300];
c1 = sqrt(a.^2 + b.^2)
c2 = hypot(a,b)

where you see as output:

c1 =

   Inf     0


c2 =

       5e+300       5e-300

this is a compiled built-in function so you can’t just open hypot to find out what its doing. It might just be pre-dividing by the maximum absolute value. There’s probably a better reference for this, but I found it in: “Vector length and normalization difficulties” by Mike Day.

Continuing this example:

m = max(abs([a;b]))
c3 = m.*sqrt((a./m).^2 + (b./m).^2)


c3 =

       5e+300       5e-300

While matlab’s hypot only accepts two inputs, pre-dividing by the maximum obviously extends to any dimension.

Triangle area in nD

Saturday, January 7th, 2017

My typical go-to formula for the area of a triangle with vertices in nD is Heron’s formula. Actually this is a formula for the area of a triangle given its side lengths. Presumably it’s easy to compute side lengths in any dimension: e.g., a = sqrt((B-C).(B-C)).

Kahan showed that the vanilla Heron’s formula is numerically poor if your (valid) triangle side lengths are given as floating point numbers. He improved this formula by sorting the side-lengths and carefully rearranging the terms in the formula to reduce roundoff.

However, if you’re vertices don’t actually live in R^n but rather F^n where F is the fixed precision floating-point grid, then computing side lengths will in general incur some roundoff error. And it may so happen (and it does so happen) that this roundoff error invalidates the triangle side lengths. Here, invalid means that the side lengths don’t obey the triangle inequality. This might happen for nearly degenerate (zero area) triangles: one (floating point) side length ends up longer than the sum of the others. Kahan provides an assertion to check for this, but there’s no proposed remedy that I could find. One could just declare that these edge-lengths must of come from a zero-area triangle. But returning zero might let the user happily work with very nonsensical triangle side lengths in other contexts not coming from an embedded triangle. You could have two versions: one for “known embeddings” (with assertions off and returning 0) and one for “intrinsic/metric only” (with assertions on and returning NaN). Or you could try to fudge in an epsilon a nd hope you choose it above the roundoff error threshold but below the tolerance for perceived “nonsense”. These are messy solutions.

An open question (open in the personal sense of “I don’t know the answer”) is what is the most robust way to compute triangle area in nD with floating point vertex positions. A possible answer might be: compute the darn floating point edge lengths, use Kahan’s algorithm and replace NaNs with zeros. That seems unlikely to me because (as far as I can tell) there are 4 sqrt calls, major sources of floating point error.

Re-reading Shewchuk’s robustness notes, I saw his formula for triangle area for points A,B,C in 3D:

Area3D(A,B,C) = sqrt( Area2D(Axy,Bxy,Cxy)^2 + Area2D(Ayz,Byz,Cyz)^2 + Area2D(Azx,Bzx,Czx)^2 ) / 2

where Area2D(A,B,C) is computed as the determinant of [[A;B;C] 1]. This formula reduces the problem of computing 3D triangle area into computing 2D triangle area on the “shadows” (projections) of the triangle on each axis-aligned plane. This lead me to think of a natural generalization to nD:

AreaND(A,B,C) = sqrt( ∑_{i<j} ( Area2D(Aij,Bij,Cij)^2 ) / 2

This formula computes the area of the shadow on all (N choose 2) axis-aligned planes. Since the sqrt receives a sum of squares as in argument there’s no risk of getting a NaN. There’s a clear drawback that this formula is O(N^2) vs Heron’s/Kahan’s O(N).

Friday, January 6th, 2017

Like clockwork, my matlab updates have gotten out of sync with Xcode updates. It seems like fixing this SDK error always requires a different hack each year. This year I got the error:

No supported compiler or SDK was found. For options, visit 

To fix it, I replaced all occurrences of 10.9 with 10.11 in /Applications/{++,}_maci64.xml

I’m still getting linker warnings:

ld: warning: object file was built for newer OSX version (10.11) than being linked (10.9)

For now, I’m assuming that I can ignore them. We’ll see how far that gets me.

Plot a bunch of edges in matlab with per-edge color data

Monday, December 19th, 2016

Suppose you have a list of vertices V (#V by dim) and a list of edges E (#E by 2) indexing V, then one hacky to draw all of these edges with per-vertex color data DV (#V by 1) is to use trisurf:

trisurf(E(:,[1 1 2]),V(:,1),V(:,2),V(:,3),'CData',DV,'EdgeColor','interp');

However, if you try to use this trick to draw per-edge color data DE (#E by 1), you’ll find that 'EdgeColor','flat' does not work. Instead you can explode you compact “mesh” of edges into individual edges and repeat you edge color data for each edge point:


Command line program to view 3d meshes from files and piped input

Thursday, December 15th, 2016

Here’s a little C++ program to directly render meshes in 3D viewer from the command line.

This let’s you write little test programs without worrying about linking to a 3D viewer. You just need to output a mesh in a standard format. For example, here’s a tiny program that outputs a cube in an .off format:

#include <igl/read_triangle_mesh.h>
#include <Eigen/Core>
#include <iostream>

int main(int argc, char * argv[])
  using namespace Eigen;
  MatrixXd V(8,3);
  MatrixXi Q(6,4);
    "OFF\n"<<V.rows()<<" "<<Q.rows()<<" 0\n"<<
    V.format(IOFormat(FullPrecision,DontAlignCols," ","\n","","","","\n"))<<
    (Q.array()).format(IOFormat(FullPrecision,DontAlignCols," ","\n","4 ","","","\n"));
  return EXIT_SUCCESS;

Compile this into cube_off then issue:

./cube_off | view mesh

**Update: ** And here’s a funny, little one-liner you can call from matlab to display a mesh via the .obj format:

system(sprintf('echo \"%s%s\" | /usr/local/bin/viewmesh',sprintf('v %0.17f %0.17f %0.17f\n',V'),sprintf('f %d %d %d\n',F')))

Background computation threads with igl::viewer::Viewer

Sunday, December 4th, 2016

Here’s a minimal example showing how to launch background computation threads for each mesh in a list of meshes. Meanwhile the main thread runs a mesh viewer with all meshes concatenated into one huge multi-component mesh. Whenever a computation thread signals that an update to the mesh needs to be made, the main thread will re-concatenate the meshes and update the viewer. In this example, the “computation” is determining how much to move a clock “hand” (random mesh).

Here’s the program running on the cow, cheburashka and knight models:

// Tiny example to demonstrate spawning a background computation thread for
// each mesh in a list and update the viewer when computation results in a
// change to (one of) the meshes.
// In this example, three meshes are read in and interpreted as "hands" of a
// clock. The "computation" is just a busy-wait until the hand should move
// (after one second, one minute, one hour). This is, of course, a terrible way
// to implement a clock.
// ./test libigl/tutorial/shared/{,,}
#include <igl/read_triangle_mesh.h>
#include <igl/point_mesh_squared_distance.h>
#include <igl/combine.h>
#include <igl/viewer/Viewer.h>
#include <Eigen/Geometry>
#include <thread>
int main(int argc, char * argv[])
  using namespace std;
  using namespace Eigen;
  // Read in k meshes (<=3 will be used)
  std::vector<Eigen::MatrixXd> VV(argc-1);
  std::vector<Eigen::MatrixXi> FF(argc-1);
  for(int i = 1;i<argc;i++)
    VV[i-1].col(0).array() -= VV[i-1].col(0).mean();
    VV[i-1].col(1).array() -= VV[i-1].col(1).minCoeff();
  bool continue_computing = true;
  // Flag to redraw and mutex to guard it
  bool redraw = false;
  std::mutex m;
  // After waiting `tic` seconds, rotate `V` by `deg` degrees
  const auto rot = [&continue_computing,&redraw,&m](
    const double tic, const double deg, Eigen::MatrixXd& V)
      // Let's watch this at 500x: Real clocks are slow.
      V *= Eigen::AngleAxisd(deg/180.*igl::PI,
        std::lock_guard<std::mutex> lock(m);
        redraw = true;
  // Launch background "computation" threads for each "hand" of the clock
  // std::ref is important, otherwise std::bind passes by copy
  std::vector<std::thread> threads;
    case 3:
    case 2:
    case 1:
    default: break;
  igl::viewer::Viewer v;;
  // Helper function to view k meshes
  const auto & set_meshes = [](
    const std::vector<Eigen::MatrixXd> & VV,
    const std::vector<Eigen::MatrixXi> & FF,
    igl::viewer::Viewer & v)
    Eigen::MatrixXd V;
    Eigen::MatrixXi F;
  // Continuous draw loop. TODO: trigger draws using glfwPostEmptyEvent
  v.core.is_animating = true;
  // Before every draw check if the meshes have changed. 
  v.callback_pre_draw = 
    [&redraw,&m,&VV,&FF,&set_meshes](igl::viewer::Viewer & v)->bool
          std::lock_guard<std::mutex> lock(m);
          redraw = false;
      return false;
  // Tell computation threads to stop
  continue_computing = false;
  // Join with computation threads: return to serial main thread.
  for(auto & t : threads) if(t.joinable()) t.join();

mesh clock igl viewer threads

This is pretty hacky, but it will allow me to use the standard libigl viewer for making comparisons of mesh-editing methods whose performances are very different.