pathlength thoughts 

Issue descriptionCool math for exactly quadratic arclength http://segfaultlabs.com/docs/quadraticbeziercurvelength Webkit code for pathlength http://trac.webkit.org/browser/trunk/Source/WebCore/platform/graphics/PathTraversalState.cpp Skia has this in SkPathMeasure. 1. Is SkPathMeasure as fast as it should be? 2. Can we make it more accurate w/o hurting speed (much)? Dashing and textOnPath are the big users of our pathmeasure.
Sadly the cool math page is now a sexy lady smoking a cigarette and the Webkit code does nothing more interesting than subdivide until the curve is flat. Marking Invalid  feel free to reopen if you come across more cool math in the future. Copying the implementation for future generations float blen(v* p0, v* p1, v* p2) { v a,b; a.x = p0>x  2*p1>x + p2>x; a.y = p0>y  2*p1>y + p2>y; b.x = 2*p1>x  2*p0>x; b.y = 2*p1>y  2*p0>y; float A = 4*(a.x*a.x + a.y*a.y); float B = 4*(a.x*b.x + a.y*b.y); float C = b.x*b.x + b.y*b.y; float Sabc = 2*sqrt(A+B+C); float A_2 = sqrt(A); float A_32 = 2*A*A_2; float C_2 = 2*sqrt(C); float BA = B/A_2; return ( A_32*Sabc + A_2*B*(SabcC_2) + (4*C*AB*B)*log( (2*A_2+BA+Sabc)/(BA+C_2) ) )/(4*A_32); };
The following revision refers to this bug: https://skia.googlesource.com/skia.git/+/17bc0851d34afe8c89d7455c061a1d419f76af8a commit 17bc0851d34afe8c89d7455c061a1d419f76af8a Author: caryclark <caryclark@google.com> Date: Mon Dec 21 13:32:53 2015 check in direct quad length measure Add code so that it at minimum won't bitrot. Next: add tests to see if it works. R=reed@google.com BUG=skia:1036 GOLD_TRYBOT_URL= https://gold.skia.org/search2?unt=true&query=source_type%3Dgm&master=false&issue=1541523002 Review URL: https://codereview.chromium.org/1541523002 [modify] http://crrev.com/17bc0851d34afe8c89d7455c061a1d419f76af8a/src/core/SkPathMeasure.cpp
The code fails for trivial (arguably degenerate) quads. For instance: (0, 0) (0, 1) (0, 2) computes a = (0, 0) A = 0 A_2 = 0 BA = NaN This implies that quads close to straight lines will suffer a similar numerical fate. This isn't worth pursing for shippable code.
Possibly, though we have to detect nearly straight quads even in our quadrootfinder, so perhaps this is no different:  first see if its nearly straight  if so, take a simpler case (length is trivial for this case)  else do the tricky part, which hopefully won't compute degenerate values I'd like us to spend a little time on this and investigate further As Mike suggested, with some additional work, the code can handle the error cases. By computing the current piecewise approach and the direct computation approach described above for a quad of the form (0,0) (A,B) (C,D) where A/B/C/D are random values +/1000. , is {0, 0}, {780.426f, 746.948f}, {433.807f, 417.938f}} The (old) piecewise length is 1445.86 The (new) direct quad length is 1447.13 Drawing the quad, it nearly folds back on itself. The extrema as drawn is at (305.341, 291.826). The length found by adding the length of two lines from end point to the extrema, the empirical length is 1447.12 Next up is handling subdividing the quad when its length is nonlinear with respect to t.


►
Sign in to add a comment 
Comment 1 by hcm@google.com, Sep 10 2014