dist/docs/viewperf.html

<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01 Transitional//EN" "http://www.w3.org/TR/html4/loose.dtd">
<html lang="en">
<head>
  <meta http-equiv="content-type" content="text/html; charset=utf-8">
  <title>Viewperf Issues</title>
  <link rel="stylesheet" type="text/css" href="mesa.css">
</head>
<body>

<div class="header">
  <h1>The Mesa 3D Graphics Library</h1>
</div>

<iframe src="contents.html"></iframe>
<div class="content">

<h1>Viewperf Issues</h1>

<p>
This page lists known issues with
<a href="http://www.spec.org/gwpg/gpc.static/vp11info.html" target="_main">SPEC Viewperf 11</a>
when running on Mesa-based drivers.
</p>

<p>
The Viewperf data sets are basically GL API traces that are recorded from
CAD applications, then replayed in the Viewperf framework.
</p>

<p>
The primary problem with these traces is they blindly use features and
OpenGL extensions that were supported by the OpenGL driver when the trace
was recorded,
but there's no checks to see if those features are supported by the driver
when playing back the traces with Viewperf.
</p>

<p>
These issues have been reported to the SPEC organization in the hope that
they'll be fixed in the future.
</p>

<p>
Some of the Viewperf tests use a lot of memory.
At least 2GB of RAM is recommended.
</p>


<h2>Catia-03 test 2</h2>

<p>
This test creates over 38000 vertex buffer objects.  On some systems
this can exceed the maximum number of buffer allocations.  Mesa
generates GL_OUT_OF_MEMORY errors in this situation, but Viewperf
does no error checking and continues.  When this happens, some drawing
commands become no-ops.  This can also eventually lead to a segfault
either in Viewperf or the Mesa driver.
</p>


<h2>Catia-03 tests 3, 4, 8</h2>

<p>
These tests use features of the
<a href="http://www.opengl.org/registry/specs/NV/fragment_program2.txt"
target="_main">
GL_NV_fragment_program2</a> and
<a href="http://www.opengl.org/registry/specs/NV/vertex_program3.txt"
target="_main">
GL_NV_vertex_program3</a> extensions without checking if the driver supports
them.
</p>
<p>
When Mesa tries to compile the vertex/fragment programs it generates errors
(which Viewperf ignores).
Subsequent drawing calls become no-ops and the rendering is incorrect.
</p>


<h2>sw-02 tests 1, 2, 4, 6</h2>

<p>
These tests depend on the
<a href="http://www.opengl.org/registry/specs/NV/primitive_restart.txt"
target="_main">GL_NV_primitive_restart</a> extension.
</p>

<p>
If the Mesa driver doesn't support this extension the rendering will
be incorrect and the test will fail.
</p>

<p>
Also, the color of the line drawings in test 2 seem to appear in a random
color.  This is probably due to some uninitialized state somewhere.
</p>


<h2>sw-02 test 6</h2>

<p>
The lines drawn in this test appear in a random color.
That's because texture mapping is enabled when the lines are drawn, but no
texture image is defined (glTexImage2D() is called with pixels=NULL).
Since GL says the contents of the texture image are undefined in that
situation, we get a random color.
</p>


<h2>Lightwave-01 test 3</h2>

<p>
This test uses a number of mipmapped textures, but the textures are
incomplete because the last/smallest mipmap level (1 x 1 pixel) is
never specified.
</p>

<p>
A trace captured with
<a href="https://github.com/apitrace/apitrace" target="_main">API trace</a>
shows this sequences of calls like this:

<pre>
2504 glBindTexture(target = GL_TEXTURE_2D, texture = 55)
2505 glTexImage2D(target = GL_TEXTURE_2D, level = 0, internalformat = GL_RGBA, width = 512, height = 512, border = 0, format = GL_RGB, type = GL_UNSIGNED_SHORT, pixels = blob(1572864))
2506 glTexImage2D(target = GL_TEXTURE_2D, level = 1, internalformat = GL_RGBA, width = 256, height = 256, border = 0, format = GL_RGB, type = GL_UNSIGNED_SHORT, pixels = blob(393216))
2507 glTexImage2D(target = GL_TEXTURE_2D, level = 2, internalformat = GL_RGBA, width = 128, height = 128, border = 0, format = GL_RGB, type = GL_UNSIGNED_SHORT, pixels = blob(98304))
[...]
2512 glTexImage2D(target = GL_TEXTURE_2D, level = 7, internalformat = GL_RGBA, width = 4, height = 4, border = 0, format = GL_RGB, type = GL_UNSIGNED_SHORT, pixels = blob(96))
2513 glTexImage2D(target = GL_TEXTURE_2D, level = 8, internalformat = GL_RGBA, width = 2, height = 2, border = 0, format = GL_RGB, type = GL_UNSIGNED_SHORT, pixels = blob(24))
2514 glTexParameteri(target = GL_TEXTURE_2D, pname = GL_TEXTURE_MIN_FILTER, param = GL_LINEAR_MIPMAP_LINEAR)
2515 glTexParameteri(target = GL_TEXTURE_2D, pname = GL_TEXTURE_WRAP_S, param = GL_REPEAT)
2516 glTexParameteri(target = GL_TEXTURE_2D, pname = GL_TEXTURE_WRAP_T, param = GL_REPEAT)
2517 glTexParameteri(target = GL_TEXTURE_2D, pname = GL_TEXTURE_MAG_FILTER, param = GL_NEAREST)
</pre>

<p>
Note that one would expect call 2514 to be glTexImage(level=9, width=1,
height=1) but it's not there.
</p>

<p>
The minification filter is GL_LINEAR_MIPMAP_LINEAR and the texture's
GL_TEXTURE_MAX_LEVEL is 1000 (the default) so a full mipmap is expected.
</p>

<p>
Later, these incomplete textures are bound before drawing calls.
According to the GL specification, if a fragment program or fragment shader
is being used, the sampler should return (0,0,0,1) ("black") when sampling
from an incomplete texture.
This is what Mesa does and the resulting rendering is darker than it should
be.
</p>

<p>
It appears that NVIDIA's driver (and possibly AMD's driver) detects this case
and returns (1,1,1,1) (white) which causes the rendering to appear brighter
and match the reference image (however, AMD's rendering is <em>much</em>
brighter than NVIDIA's).
</p>

<p>
If the fallback texture created in _mesa_get_fallback_texture() is
initialized to be full white instead of full black the rendering appears
correct.
However, we have no plans to implement this work-around in Mesa.
</p>


<h2>Maya-03 test 2</h2>

<p>
This test makes some unusual calls to glRotate.  For example:
</p>
<pre>
glRotate(50, 50, 50, 1);
glRotate(100, 100, 100, 1);
glRotate(52, 52, 52, 1);
</pre>
<p>
These unusual values lead to invalid modelview matrices.
For example, the last glRotate command above produces this matrix with Mesa:
<pre>
1.08536e+24 2.55321e-23 -0.000160389 0
5.96937e-25 1.08536e+24 103408 0
103408 -0.000160389 1.74755e+09 0
0 0 0 nan
</pre>
and with NVIDIA's OpenGL:
<pre>
1.4013e-45 0 -nan 0
0 1.4013e-45 1.4013e-45 0
1.4013e-45 -nan 1.4013e-45 0
0 0 0 1.4013e-45
</pre>
<p>
This causes the object in question to be drawn in a strange orientation
and with a semi-random color (between white and black) since GL_FOG is enabled.
</p>


<h2>Proe-05 test 1</h2>

<p>
This uses depth testing but there's two problems:
<ol>
<li>The glXChooseFBConfig() call doesn't request a depth buffer
<li>The test never calls glClear(GL_DEPTH_BUFFER_BIT) to initialize the depth buffer
</ol>
<p>
If the chosen visual does not have a depth buffer, you'll see the wireframe
car model but it won't be rendered correctly.
</p>
If (by luck) the chosen visual has a depth buffer, its initial contents
will be undefined so you may or may not see parts of the model.
<p>
Interestingly, with NVIDIA's driver most visuals happen to have a depth buffer
and apparently the contents are initialized to 1.0 by default so this test
just happens to work with their drivers.
</p>

<p>
Finally, even if a depth buffer was requested and the glClear(GL_COLOR_BUFFER_BIT)
calls were changed to glClear(GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT)
the problem still wouldn't be fixed because GL_DEPTH_WRITEMASK=GL_FALSE when
glClear is called so clearing the depth buffer would be a no-op anyway.
</p>


<h2>Proe-05 test 6</h2>

<p>
This test draws an engine model with a two-pass algorithm.
The first pass is drawn with polygon stipple enabled.
The second pass is drawn without polygon stipple but with blending
and GL_DEPTH_FUNC=GL_LEQUAL.
If either of the two passes happen to use a software fallback of some
sort, the Z values of fragments may be different between the two passes.
This leads to incorrect rendering.
</p>

<p>
For example, the VMware SVGA gallium driver uses a special semi-fallback path
for drawing with polygon stipple.
Since the two passes are rendered with different vertex transformation
implementations, the rendering doesn't appear as expected.
Setting the SVGA_FORCE_SWTNL environment variable to 1 will force the
driver to use the software vertex path all the time and clears up this issue.
</p>

<p>
According to the OpenGL invariance rules, there's no guarantee that
the pixels produced by these two rendering states will match.
To achieve invariance, both passes should enable polygon stipple and
blending with appropriate patterns/modes to ensure the same fragments
are produced in both passes.
</p>


</div>
</body>
</html>