Reduce calls to SkMatrix::reset from GrDrawState

Next in the continuing series of setup cost reduction CLs. This CL removes the 3.8% spent in SkMatrix::reset when running:
 "bench --match drawpointsdash_5_5_aa"

Performance numbers will be forthcoming.

[bsalomon, 2013-04-25 02:55:12]

lgtm

[robertphillips, 2013-04-25 13:51:05]

This appears to be a win. Here are the results for the GPU benchs (old is left;
new is right) with everything <= .5ms or <= 3% removed (using the geometry-base
rect drawing method):

                gradient_create_opaque GPU c   15.69   16.30  -0.61   -3.9%

                       zeroradroundrect GPU c   24.61   23.81  +0.80   +3.3%
              picture_playback_drawText GPU c   32.87   31.65  +1.22   +3.7%
          picture_playback_drawPosTextH GPU c   38.38   36.78  +1.60   +4.2%
            picture_record_dictionaries GPU c   13.53   12.96  +0.57   +4.2%
                    blurrect_gaussian_2 GPU c   24.08   23.02  +1.06   +4.4%
 bitmap_8888_update_scale_rotate_filter GPU c   74.63   71.27  +3.36   +4.5%
        bitmap_8888_update_scale_filter GPU c   75.84   72.18  +3.66   +4.8%
                     bitmap_8888_update GPU c   75.60   71.52  +4.08   +5.4%
                  drawpointsdash_5_5_bw GPU c   27.28   25.75  +1.53   +5.6%
                  drawpointsdash_5_5_aa GPU c   28.08   26.22  +1.86   +6.6%
                  drawpointsdash_1_1_bw GPU c  116.86  108.61  +8.25   +7.1%
                  drawpointsdash_3_1_bw GPU c   43.82   40.72  +3.10   +7.1%
                  drawpointsdash_1_1_aa GPU c   90.04   83.11  +6.93   +7.7%
                  drawpointsdash_3_1_aa GPU c   40.89   37.60  +3.29   +8.0%
                                aarects GPU c   50.01   45.96  +4.05   +8.1%
                             tablebench GPU c   41.54   35.11  +6.43  +15.5%

As expected the benchs that draw a lot of primitives/rects benefit the most.

[robertphillips, 2013-04-25 15:00:22]

Here are the corresponding results for the shader-based rect drawing method:

           circles_fill GPU c   15.92  17.21   -1.29  -8.1%
         circles_stroke GPU c   22.44  23.52   -1.08  -4.8%
           arbroundrect GPU c   48.89  51.05   -2.16  -4.4%
                readpix GPU c   79.51  81.97   -2.46  -3.1%

      dashline_2_square GPU c   53.95  52.29   +1.66  +3.1%
      dashline_0_circle GPU c   16.67  16.09   +0.58  +3.5%
  drawpointsdash_5_5_bw GPU c   27.13  26.17   +0.96  +3.5%
  drawpointsdash_3_1_bw GPU c   42.04  40.27   +1.77  +4.2%
bitmap_8888_update_volatile_scale_rotate_filter GPU c  14.02 13.22 +0.80 +5.7%
  drawpointsdash_5_5_aa GPU c   31.22  28.90   +2.32  +7.4%
  drawpointsdash_3_1_aa GPU c   45.16  41.73   +3.43  +7.6%
  drawpointsdash_1_1_aa GPU c  104.06  92.93  +11.13 +10.7%
             tablebench GPU c   40.08  34.53   +5.55 +13.8%
                aarects GPU c   62.22  52.28   +9.94 +16.0%

[Tom Hudson, 2013-04-25 15:31:20]

Any sense of why there's such a loss on e.g. circles_fill and circles_stroke?

[robertphillips, 2013-04-25 17:15:48]

Despite running on a Mac with repeat count set to 200 we are still subject to
noise. Here is a re-run of the test:

        aaclip_build_path_BW GPU c    3.00    3.65   -0.65   -21.7%
        aaclip_build_path_AA GPU c    6.40    7.50   -1.10   -17.2%

       drawpointsdash_5_5_bw GPU c   26.99   25.73   +1.26    +4.7%
       drawpointsdash_1_1_bw GPU c  113.20  107.57   +5.63    +5.0%
 picture_record_dictionaries GPU c   13.82   13.05   +0.77    +5.6%
       drawpointsdash_3_1_bw GPU c   42.25   39.74   +2.51    +5.9%
       drawpointsdash_5_5_aa GPU c   31.00   28.75   +2.25    +7.3%
       drawpointsdash_3_1_aa GPU c   44.98   41.54   +3.44    +7.6%
       drawpointsdash_1_1_aa GPU c  103.22   93.63   +9.59    +9.3%
                  tablebench GPU c   40.21   34.87   +5.34   +13.3%
                     aarects GPU c   62.36   52.56   +9.80   +15.7%

I only believe the drawpointsdash*, tablebench and aarects numbers are
real(ish).

[robertphillips, 2013-04-25 17:24:03]

committed as r8861