EVALUATION

The initial implementation of the pipeline has been completed.
Currently it's disabled by default pending resolution of
some quality issues on some video boards.
To enabled, supply -Dsun.java2d.d3d=true property to the vm.

At this point the following primitives are accelerated 
by the D3D pipeline:
   - managed image to VolatileImage blits/transforms/scaling,
     w/ and w/o alpha compositing, nearest neighbor and bilinear
     filtering
   - D3DRenderer:
      - draw[Line, Rect, Poly*], FillRect - directly via
        DrawPrimitive for thin stroke and single color
      - the rest is implemented via FillSpans
   - MaskFill (AA fills)
   - Text rendering (AA/no AA): done via texture mapping, with a glyph
     cache
   - hardware clipping via stencil buffer

Note:
  - Xor paint mode is not accelerated (our GDI pipeline is used
    as a fallback).
  - no gradient paint acceleration

At the time of the initial putback, the following performance
improvement were registered:
 Performance gains vary with benchmarks (data for Radeon 9800 Pro on WinXP):
  NinJo (mostly lines, polygon fills/draws), in milliseconds:
     nod3d (d3d disabled): 352021 ms
     d3d   (d3d enabled ):  85933 ms
     ogl   (ogl enabled ):  84176 ms

  FireStarter: (TR-transformed, AC-AlphaComposited, tr - translucent
  source image, op - opaque, bm - bitmask), number of rocks:
               TR+AC:    AC:       TR:       no AC/TR:
    old dd/d3d      1   4300         1       4600(tr)    1300(op/bm)
    new d3d      3000   4500      3300       4500(tr/bm) 5300(opaque)

  J2DBench results for most accelerated primitives:
    http://javaweb.sfbay/~tdv/bugs/d3d_pipeline/benchmarks/2005-03-15_20h47m/j2dbench/Summary_Report.html
  In general, it's a win for most primitives, huge in some cases,
  especially when compositing or transformations are involved.

  Swingmark is currently about 15% slower with the new pipeline for
  ocean look and feel.
  Here are some swingmark data logs for those interested.
  (scroll to the bottom, look at the 'Mean'. 'target0' is the new
  pipeline, 'base' is jdk1.6.0):
    http://javaweb.sfbay/~tdv/bugs/d3d_pipeline/benchmarks/2005-03-15_20h47m/swingmark/swingmark.run.log
  (full logs can be found in the same directory)

Note that some followup work considerably improved performance of
Swing applications and the SwingMark scores. For more details see bug:
  6252147: D3D: minimize device state changes in the pipeline

###@###.### 2005-04-09 00:33:01 GMT