Bug 25068

Interesting, I was expecting a boost since the shader is simpler and we're transferring fewer bytes.

Can you report the scale of the regression?

Since centre-point sampling in combination with 1x1R textures is buggy, I can't simply revert this change (and the centre-point sampling is required to prevent off-by-one rendering errors, i.e the occasional black rectangle around images).

On my i945, prior to this commit I get 378k/s, and afterwards 370k/s.

Alternates that I have tried so far:
  using per-vertex colors: 361k/s
  using shader constants instead of defaults: 359k/s.

On 945GM ia32, it drops 10% from 602000.0 to  536000.0 with rgb10text, and drops 11% from 812000.0 to  716000.0 with aa10text. 
On 945GME ia32, its rgb10text drops 19% from 330000.0 to  266000.0 but its aa10text only drops a little from 347000.0 to 338000.0. 

Dropping priority as we seem to be hitting a gpu bottleneck on a path that I believe is required for correct rendering elsewhere (with similar setup).

Ok, I think I've found the cause of the damage here.

i915:
Baseline:
3600000 trep @   0.0071 msec (142000.0/sec)
3200000 trep @   0.0082 msec (122000.0/sec)

Adjusting libXft to use SolidFills:
4000000 trep @   0.0066 msec (150000.0/sec)
3600000 trep @   0.0076 msec (132000.0/sec)

Improving the driver to avoid reading back (from system memory) a pixel to determine color for SolidFills:
4000000 trep @   0.0067 msec (149000.0/sec)
3600000 trep @   0.0068 msec (147000.0/sec)

And for good measure reverting the change to libXft i.e. back to using a solid pixmap (this should be close to the original code & performance):
8000000 trep @   0.0062 msec (162000.0/sec)
4000000 trep @   0.0069 msec (145000.0/sec)

PineView:
Baseline:
8000000 trep @   0.0040 msec (248000.0/sec)
8000000 trep @   0.0043 msec (232000.0/sec)
Updating libXft:
8000000 trep @   0.0040 msec (249000.0/sec)
8000000 trep @   0.0043 msec (231000.0/sec)
Improved driver:
8000000 trep @   0.0038 msec (262000.0/sec)
8000000 trep @   0.0040 msec (251000.0/sec)
And reverting the change to libXft...:
8000000 trep @   0.0039 msec (259000.0/sec)
8000000 trep @   0.0041 msec (244000.0/sec)


So, the cause would appear to be the readback of the single pixel. The remaining question is whether to use a pixmap or a diffuse color.

(Still a mystery 3x performance hit, but the improvement is consistent and the code should be pretty close to the original, so...)

commit 21c1c3c7f6eb2b5070d2153b15a8fb1afe938bbb
Author: Chris Wilson <chris@chris-wilson.co.uk>
Date:   Mon May 10 10:19:28 2010 +0100

    i915: Use 1x1R pixmap for solid drawables
    
      x11perf has a regression
      https://bugs.freedesktop.org/show_bug.cgi?id=25068
    
    caused by
    
      commit e581ceb7381e29ecc1a172597d258824f6a1d2d3
      i915: Use the color channels to pass along solid sources and masks.
    
    Do not convert 1x1R pixmaps into a solid color as the readback from the
    bo negates all the performances advantages of using a smaller vertex
    buffer and fewer samplers.
    
    Before (PineView):
      aa=66800 glyph/s, rgb=28800 glyphs/s
    
    Now:
      aa=96800 glyphs/s, rgb=48500 glyphs/s
    
    Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

Rechecked on my i945GME:

Before:
12000000 trep @   0.0025 msec (404000.0/sec): Char in 80-char aa line (Charter 10)
12000000 trep @   0.0026 msec (380000.0/sec): Char in 80-char rgb line (Charter 10)

After:
12000000 trep @   0.0024 msec (417000.0/sec): Char in 80-char aa line (Charter 10)
12000000 trep @   0.0025 msec (399000.0/sec): Char in 80-char rgb line (Charter 10)

which seems consistent with the original regression.

x11perf improves a lot on pineview(i915) recently, so verified.