special-case matrix invert for translate and scale

special-case matrix invert for translate and scale

update unittest to use diff scale-x and scale-y values, and tests for non-invertible scale matrices
Committed: https://code.google.com/p/skia/source/detail?r=7019

[reed1, 2013-01-04 15:49:16]

new

running bench [640 480] matrix_invert_maprect_typemask_nonpersp  8888: cmsecs = 
 8.89
running bench [640 480] matrix_invert_maprect_typemask_rectstaysrect  8888:
cmsecs =   4.36
running bench [640 480]  matrix_invert_maprect_persp  8888: cmsecs =   9.27
running bench [640 480] matrix_invert_maprect_nonpersp  8888: cmsecs =   7.57
running bench [640 480] matrix_invert_maprect_translate  8888: cmsecs =   2.87
running bench [640 480] matrix_invert_maprect_rectstaysrect  8888: cmsecs =  
3.18
running bench [640 480] matrix_invert_maprect_identity  8888: cmsecs =   2.99


old

running bench [640 480] matrix_invert_maprect_typemask_nonpersp  8888: cmsecs = 
 8.68
running bench [640 480] matrix_invert_maprect_typemask_rectstaysrect  8888:
cmsecs =   5.97
running bench [640 480]  matrix_invert_maprect_persp  8888: cmsecs =   9.01
running bench [640 480] matrix_invert_maprect_nonpersp  8888: cmsecs =   7.49
running bench [640 480] matrix_invert_maprect_translate  8888: cmsecs =   4.16
running bench [640 480] matrix_invert_maprect_rectstaysrect  8888: cmsecs =  
4.64
running bench [640 480] matrix_invert_maprect_identity  8888: cmsecs =   3.18

[robertphillips, 2013-01-04 15:58:33]

LGTM

https://codereview.appspot.com/7027055/diff/1/src/core/SkMatrix.cpp
File src/core/SkMatrix.cpp (right):

https://codereview.appspot.com/7027055/diff/1/src/core/SkMatrix.cpp#newcode856
src/core/SkMatrix.cpp:856: if (0 == (mask & ~(kScale_Mask | kTranslate_Mask))) {
NULL != inv

https://codereview.appspot.com/7027055/diff/1/src/core/SkMatrix.cpp#newcode859
src/core/SkMatrix.cpp:859: SkScalar invX = fMat[kMScaleX];
typo you already know about (kMScaleX -> kMScaleY)

[TomH, 2013-01-04 16:21:28]

The numbers are nice when it's a win, but there are a couple of cases that look
like regressions of a few %. Noise, or something to worry about?

[reed1, 2013-01-04 16:26:51]

From reading the profile, the slow-downs are a combination of noise (the invert
step is only a part of the bench), and just the few extra instructions /
branches to test the bits in the mask field.

My working assumption is that our *very* common case is that the matrix is
scale+translate when called by chrome or android.

Related to my first note: the speed-ups are actually much larger, if we were to
test invert() in isolation.

[JimVV, 2013-01-04 16:39:04]

If the common case is only scale+translate, is there any reason to not to pull
those out and apply them all the time? In the worst case you'd be multiplying by
1 and adding 0 twice. But it might save on branch stalls.

[robertphillips, 2013-01-07 13:38:02]

Without the SK_IGNORE_FAST_SCALEMATRIX_INVERT #define this causes the following
layout test failures (please see https://codereview.chromium.org/11797003/):

  fast/borders/border-image-massive-scale.html [ ImageOnlyFailure ]
  fast/borders/border-image-outset-in-shorthand.html [ ImageOnlyFailure ]
  fast/borders/border-image-outset.html [ ImageOnlyFailure ]
  fast/borders/border-image-scaled.html [ ImageOnlyFailure ]    - mac-only?
  fast/canvas/canvas-incremental-repaint.html [ ImageOnlyFailure ]
  fast/speech/input-appearance-searchandspeech.html [ ImageOnlyFailure ] -
linux-only?
 
platform/chromium/virtual/softwarecompositing/visibility/visibility-image-layers.html
[ ImageOnlyFailure ] - mac-only?
  svg/W3C-SVG-1.1-SE/pservers-grad-17-b.svg [ ImageOnlyFailure ]
  svg/W3C-SVG-1.1-SE/pservers-grad-20-b.svg [ ImageOnlyFailure ]
  svg/W3C-SVG-1.1/pservers-grad-17-b.svg [ ImageOnlyFailure ]
  svg/custom/non-scaling-stroke.svg [ ImageOnlyFailure ]
  tables/mozilla/bugs/bug137388-2.html [ ImageOnlyFailure ]
  tables/mozilla/bugs/bug23235.html [ ImageOnlyFailure ] - linux-only?
  tables/mozilla/bugs/bug4093.html [ ImageOnlyFailure ]