United StatesChange Country, Oracle Worldwide Web Sites Communities I am a... I want to...
Bug ID: JDK-6695820 D3D: performance degradation for AA primitives on older hardware
JDK-6695820 : D3D: performance degradation for AA primitives on older hardware

Details
Type:
Bug
Submit Date:
2008-04-29
Status:
Resolved
Updated Date:
2010-10-14
Project Name:
JDK
Resolved Date:
2008-05-13
Component:
client-libs
OS:
windows_xp
Sub-Component:
2d
CPU:
x86
Priority:
P3
Resolution:
Fixed
Affected Versions:
6u10
Fixed Versions:
6u10 (b24)

Related Reports
Relates:
Relates:

Sub Tasks

Description
The fix for 6690659 introduced some performance optimizations for
rendering antialised and/or transformed primitives.

On some older hardware this optimization actually results in performance
degradation.

                                    

Comments
EVALUATION

The pixel shader which does antialiasing proves to be too much for
older pre-PS3.0 level hardware. Even though it is compiled to target
PS2.0a pixel shader profile it seems that it runs completely in
software on older boards like FX5600 or Nvidia Quadro NV280.

NV280, drawLine: AA
graphics.render.tests.drawLine,graphics.opts.sizes=20,graphics.opts.transform=ident,graphics.render.opts.antialias=true,graphics.render.opts.paint=single:
old: 958.0026809 (var=0.34%) (100.0%)
new: 365.6308515 (var=0.4%) (38.17%)
-- 
graphics.render.tests.drawLine,graphics.opts.sizes=20,graphics.opts.transform=rot15,graphics.render.opts.antialias=true,graphics.render.opts.paint=single:
old: 883.5955056 (var=0.26%) (100.0%)
new: 404.6854082 (var=0.47%) (45.8%)
-- 
graphics.render.tests.drawLine,graphics.opts.sizes=250,graphics.opts.transform=ident,graphics.render.opts.antialias=true,graphics.render.opts.paint=single:
old: 599.5514147 (var=6.41%) (100.0%)
new: 46.89521345 (var=109.44%) (7.82%)
-- 
graphics.render.tests.drawLine,graphics.opts.sizes=250,graphics.opts.transform=rot15,graphics.render.opts.antialias=true,graphics.render.opts.paint=single:
old: 833.9368005 (var=1.26%) (100.0%)
new: 69.48881789 (var=75.58%) (8.33%)

  NV280, drawLine NO AA: the only improvement is 2x for TX:
 graphics.render.tests.drawLine,graphics.opts.sizes=20,graphics.opts.transform=ident,graphics.render.opts.antialias=false,graphics.render.opts.paint=single:
old: 90419.85205 (var=2.11%) (100.0%)
new: 87247.61840 (var=3.16%) (96.49%)
-- 
graphics.render.tests.drawLine,graphics.opts.sizes=20,graphics.opts.transform=rot15,graphics.render.opts.antialias=false,graphics.render.opts.paint=single:
old: 17255.51455 (var=0.6%) (100.0%)
new: 40464.60200 (var=0.13%) (234.5%)
-- 
graphics.render.tests.drawLine,graphics.opts.sizes=250,graphics.opts.transform=ident,graphics.render.opts.antialias=false,graphics.render.opts.paint=single:
old: 76994.82098 (var=1.69%) (100.0%)
new: 74372.50664 (var=0.13%) (96.59%)
-- 
graphics.render.tests.drawLine,graphics.opts.sizes=250,graphics.opts.transform=rot15,graphics.render.opts.antialias=false,graphics.render.opts.paint=single:
old: 63822.01875 (var=1.15%) (100.0%)
new: 61676.05633 (var=0.2%) (96.64%)

FX5600, drawLine: AA - losing across the board:

graphics.render.tests.drawLine,graphics.opts.sizes=20,graphics.opts.transform=ident,graphics.render.opts.antialias=true,graphics.render.opts.paint=single:
old: 588.2263134 (var=0.88%) (100.0%)
new: 475.1216122 (var=0.77%) (80.77%)
-- 
graphics.render.tests.drawLine,graphics.opts.sizes=20,graphics.opts.transform=rot15,graphics.render.opts.antialias=true,graphics.render.opts.paint=single:
old: 564.7548690 (var=1.29%) (100.0%)
new: 523.8857972 (var=1.24%) (92.76%)
-- 
graphics.render.tests.drawLine,graphics.opts.sizes=250,graphics.opts.transform=ident,graphics.render.opts.antialias=true,graphics.render.opts.paint=single:
old: 415.3290529 (var=0.62%) (100.0%)
new: 65.86402266 (var=25.73%) (15.86%)
-- 
graphics.render.tests.drawLine,graphics.opts.sizes=250,graphics.opts.transform=rot15,graphics.render.opts.antialias=true,graphics.render.opts.paint=single:
old: 557.3379579 (var=0.65%) (100.0%)
new: 81.57756123 (var=109.07%) (14.64%)

FX5600, drawLine: NO AA: 2x improvement for TX:

graphics.render.tests.drawLine,graphics.opts.sizes=20,graphics.opts.transform=ident,graphics.render.opts.antialias=false,graphics.render.opts.paint=single:
old: 42154.74207 (var=17.04%) (100.0%)
new: 43618.63665 (var=3.87%) (103.47%)
-- 
graphics.render.tests.drawLine,graphics.opts.sizes=20,graphics.opts.transform=rot15,graphics.render.opts.antialias=false,graphics.render.opts.paint=single:
old: 9276.336499 (var=0.07%) (100.0%)
new: 19064.87641 (var=0.27%) (205.52%)
-- 
graphics.render.tests.drawLine,graphics.opts.sizes=250,graphics.opts.transform=ident,graphics.render.opts.antialias=false,graphics.render.opts.paint=single:
old: 305632.56148 (var=0.4%) (100.0%)
new: 303968.88913 (var=0.54%) (99.46%)
-- 
graphics.render.tests.drawLine,graphics.opts.sizes=250,graphics.opts.transform=rot15,graphics.render.opts.antialias=false,graphics.render.opts.paint=single:
old: 118543.53123 (var=0.47%) (100.0%)
new: 241079.35076 (var=0.47%) (203.37%)


Similar story with drawRect, fillRect: very bad with AA,
some improvement (up to 8x) for NO AA with TX:

FX5600:
graphics.render.tests.fillRect,graphics.opts.sizes=20,graphics.opts.transform=rot15,graphics.render.opts.antialias=false,graphics.render.opts.paint=single:
old: 33762.81271 (var=0.54%) (100.0%)
new: 288962.08509 (var=0.33%) (855.86%)
-- 
graphics.render.tests.fillRect,graphics.opts.sizes=250,graphics.opts.transform=rot15,graphics.render.opts.antialias=false,graphics.render.opts.paint=single:
old: 636079.10480 (var=1.94%) (100.0%)
new: 1271898.62160 (var=0.87%) (199.96%)

So it looks like as a short term solution we should only enable the AA shader 
for PS3.0 level of hardware.
                                     
2008-04-29
EVALUATION

Another data point: with the fix for 6690659 we have put in a better
mechanism for flushing for the Direct3D pipeline.

The mechanism is the same we used for the old pre-6u10 pipeline:
create a RT surface and lock/unlock it on sync request.

This resulted in better flushing behavior: the J2DBench
benchmark shows much smaller score variations (< 1%) while
the old mechanism sometimes went into double digits.

The new flushing mechamism degrades the benchmark scores by 
around 2-3% though.
                                     
2008-04-30
SUGGESTED FIX

http://sa.sfbay.sun.com/projects/java2d_data/6u10/6695820.1
                                     
2008-05-02



Hardware and Software, Engineered to Work Together