Bug 42435

Created attachment 52954 [details]
Test case and stuff

core did not fit: >3000K size

Worked around the issue by moving all texture sampling to fragment shader.
(even works on 965)

Also, just moving sampling outside conditional blocks did not change anything.

The workaround did not help r600g. Same black screen. In addition, the non-llvm swrast draws point sprites 60 pixels in size instead of requested 16 (my original shaders get texture sub-image size via varyings too), plus it draws solid-black-textured point sprites as single pixels. r600c (as of Oct 27 git, just before being axed) works though.

Not knowing whether any given black screen is a bug, or legitimate undefined behaviour that just happens to work on some other renderer is driving me mad.

I managed to cut down the test case to bare minimum.

Usage:

./testbed <shaderset> <mesadriver>

shaderset is either "working" or "failing" - loads similarly named shaders.

"working" shader set does texture sampling in fragment shader
"failing" shader set does texture sampling in vertex shader and passes
the color as a varying vec4.

mesadriver is:
hw - default, LIBGL_ALWAYS_SOFTWARE and LIBGL_DRIVERS_PATH unset.
hw-alt - use .so from dri-alternates (x64-86 hardcoded, sorry).
sw - LIBGL_ALWAYS_SOFTWARE=y
sw-alt - as above, but use .so from dri-alternates.

For the Oct 27 git these map to r600g, r600c, swrast, llvmpipe on my box.

Of those:

"failing" llvmpipe - works as intended, vertical colored bars (composed of 16x16 point sprites), one sprite is dark green - left unpainted intentionally.
"working" llvmpipe - same (ok)
"failing" DRI - ok
"working" DRI - ok
"failing" r600g - all sprites black 
"working" r600g - all sprites black 
"failing" old-swrast  - single pixels instead of 16x16, but looks like color is right
"working" old-swrast  - same as "failing"

The hardware is rv770 9440.

I'd say there's a problem with r600g and a different one with old swrast.

There's also intermittent display corruption mainly on areas in xterm redrawn (but not those that get scrolled up) immediately after launching the testcase.

Created attachment 53025 [details]
minimized test case

Created attachment 53028 [details]
xterm corruption screenshot

xterm corruption screenshot, got it right after I first run those shaders:

#version 120
#line 2 0

uniform float posz;
uniform float width;
attribute vec2 posn;

varying float idx;

void main() { 
    vec2 posn2 = posz*posn;
    idx = posn2.x/width;
    gl_Position = gl_ModelViewProjectionMatrix*vec4(posn2.x, posn2.y, 0.0, 1.0);
    gl_PointSize = posz;
}
#version 120
#line 2 0
uniform sampler2D ansi;

varying float idx;

void main() {
    vec2 pc = gl_PointCoord;
    gl_FragColor = texture2D(ansi, vec2(0.5, 0));
}

 - got grey screen as expected, and then reset fragment shader back to 

#version 120
#line 2 0
uniform sampler2D ansi;

varying float idx;

void main() {
    vec2 pc = gl_PointCoord;
    gl_FragColor = texture2D(ansi, vec2(idx, 0));
}

Further reducing shaders I got to this point:

#version 120
#line 2 0

varying float idx;

void main() { 
    vec2 posn2 = posz*posn;
    idx = 0.625;
    gl_Position = gl_ModelViewProjectionMatrix*vec4(128, 128, 0.0, 1.0);
    gl_PointSize = 128;
}

#version 120
#line 2 0
varying float idx;

void main() {
    gl_FragColor = vec4(idx, 0, 0, 1);
}

This is in fact very close to glsl-novertexdata from the piglit.

Which still does not work on r600g and swrast while working ok on r600c and llvmpipe.

Created attachment 53029 [details]
piglit test case for the problem

Created attachment 53030 [details]
piglit test case for the problem as a patch

Created attachment 53031 [details] [review]
piglit test case for the problem as a patch

Problem manifests itself only when gl_PointSize is set in the shader. Corrected patch attached.

Created attachment 53033 [details] [review]
[PATCH] r600g: take into account PSIZE output for spi mapping

Does this patch help?

Yes, thank you.

I experience a similar situation with Ironlake and Mesa 8.0.1, only with far more complex shaders.

Apparently, in these shaders, all varyings get bogus data on the fragment shader.

The only workaround I can find is passing that data through the built-in varying in gl_TexCoord, although that's quite suboptimal for other drivers so I would prefer not to.

r600g fixed in:
http://cgit.freedesktop.org/mesa/mesa/commit/?id=16c0a3a9720c03c3d5cbfd66e2db63014dfc4f19