xenocara/dist/Mesa/docs/shading.html
2007-11-24 17:25:28 +00:00

313 lines
8.4 KiB
HTML

<HTML>
<TITLE>Shading Language Support</TITLE>
<link rel="stylesheet" type="text/css" href="mesa.css"></head>
<BODY>
<H1>Shading Language Support</H1>
<p>
This page describes the features and status of Mesa's support for the
<a href="http://opengl.org/documentation/glsl/" target="_parent">
OpenGL Shading Language</a>.
</p>
<p>
Last updated on 28 March 2007.
</p>
<p>
Contents
</p>
<ul>
<li><a href="#unsup">Unsupported Features</a>
<li><a href="#notes">Implementation Notes</a>
<li><a href="#hints">Programming Hints</a>
<li><a href="#standalone">Stand-alone Compiler</a>
<li><a href="#implementation">Compiler Implementation</a>
<li><a href="#validation">Compiler Validation</a>
</ul>
<a name="unsup">
<h2>Unsupported Features</h2>
<p>
The following features of the shading language are not yet supported
in Mesa:
</p>
<ul>
<li>Dereferencing arrays with non-constant indexes
<li>Comparison of user-defined structs
<li>Linking of multiple shaders is not supported
<li>gl_ClipVertex
<li>The derivative functions such as dFdx() are not implemented
<li>The inverse trig functions asin(), acos(), and atan() are not implemented
<li>The gl_Color and gl_SecondaryColor varying vars are interpolated
without perspective correction
</ul>
<p>
All other major features of the shading language should function.
</p>
<a name="notes">
<h2>Implementation Notes</h2>
<ul>
<li>Shading language programs are compiled into low-level programs
very similar to those of GL_ARB_vertex/fragment_program.
<li>All vector types (vec2, vec3, vec4, bvec2, etc) currently occupy full
float[4] registers.
<li>Float constants and variables are packed so that up to four floats
can occupy one program parameter/register.
<li>All function calls are inlined.
<li>Shaders which use too many registers will not compile.
<li>The quality of generated code is pretty good, register usage is fair.
<li>Shader error detection and reporting of errors (InfoLog) is not
very good yet.
<li>The ftransform() function doesn't necessarily match the results of
fixed-function transformation.
</ul>
<p>
These issues will be addressed/resolved in the future.
</p>
<a name="hints">
<h2>Programming Hints</h2>
<ul>
<li>Declare <em>in</em> function parameters as <em>const</em> whenever possible.
This improves the efficiency of function inlining.
</li>
<br>
<li>To reduce register usage, declare variables within smaller scopes.
For example, the following code:
<pre>
void main()
{
vec4 a1, a2, b1, b2;
gl_Position = expression using a1, a2.
gl_Color = expression using b1, b2;
}
</pre>
Can be rewritten as follows to use half as many registers:
<pre>
void main()
{
{
vec4 a1, a2;
gl_Position = expression using a1, a2.
}
{
vec4 b1, b2;
gl_Color = expression using b1, b2;
}
}
</pre>
Alternately, rather than using several float variables, use
a vec4 instead. Use swizzling and writemasks to access the
components of the vec4 as floats.
</li>
<br>
<li>Use the built-in library functions whenever possible.
For example, instead of writing this:
<pre>
float x = 1.0 / sqrt(y);
</pre>
Write this:
<pre>
float x = inversesqrt(y);
</pre>
<li>
Use ++i when possible as it's more efficient than i++
</li>
</ul>
<a name="standalone">
<h2>Stand-alone Compiler</h2>
<p>
A unique stand-alone GLSL compiler driver has been added to Mesa.
<p>
<p>
The stand-alone compiler (like a conventional command-line compiler)
is a tool that accepts Shading Language programs and emits low-level
GPU programs.
</p>
<p>
This tool is useful for:
<p>
<ul>
<li>Inspecting GPU code to gain insight into compilation
<li>Generating initial GPU code for subsequent hand-tuning
<li>Debugging the GLSL compiler itself
</ul>
<p>
To build the glslcompiler program (this will be improved someday):
</p>
<pre>
cd src/mesa
make libmesa.a
cd drivers/glslcompiler
make
</pre>
<p>
Here's an example of using the compiler to compile a vertex shader and
emit GL_ARB_vertex_program-style instructions:
</p>
<pre>
glslcompiler --arb --linenumbers --vs vertshader.txt
</pre>
<p>
The output may look similar to this:
</p>
<pre>
!!ARBvp1.0
0: MOV result.texcoord[0], vertex.texcoord[0];
1: DP4 temp0.x, state.matrix.mvp.row[0], vertex.position;
2: DP4 temp0.y, state.matrix.mvp.row[1], vertex.position;
3: DP4 temp0.z, state.matrix.mvp.row[2], vertex.position;
4: DP4 temp0.w, state.matrix.mvp.row[3], vertex.position;
5: MOV result.position, temp0;
6: END
</pre>
<p>
Note that some shading language constructs (such as uniform and varying
variables) aren't expressible in ARB or NV-style programs.
Therefore, the resulting output is not always legal by definition of
those program languages.
</p>
<p>
Also note that this compiler driver is still under development.
Over time, the correctness of the GPU programs, with respect to the ARB
and NV languagues, should improve.
</p>
<a name="implementation">
<h2>Compiler Implementation</h2>
<p>
The source code for Mesa's shading language compiler is in the
<code>src/mesa/shader/slang/</code> directory.
</p>
<p>
The compiler follows a fairly standard design and basically works as follows:
</p>
<ul>
<li>The input string is tokenized (see grammar.c) and parsed
(see slang_compiler_*.c) to produce an Abstract Syntax Tree (AST).
The nodes in this tree are slang_operation structures
(see slang_compile_operation.h).
The nodes are decorated with symbol table, scoping and datatype information.
<li>The AST is converted into an Intermediate representation (IR) tree
(see the slang_codegen.c file).
The IR nodes represent basic GPU instructions, like add, dot product,
move, etc.
The IR tree is mostly a binary tree, but a few nodes have three or four
children.
In principle, the IR tree could be executed by doing an in-order traversal.
<li>The IR tree is traversed in-order to emit code (see slang_emit.c).
This is also when registers are allocated to store variables and temps.
<li>In the future, a pattern-matching code generator-generator may be
used for code generation.
Programs such as L-BURG (Bottom-Up Rewrite Generator) and Twig look for
patterns in IR trees, compute weights for subtrees and use the weights
to select the best instructions to represent the sub-tree.
<li>The emitted GPU instructions (see prog_instruction.h) are stored in a
gl_program object (see mtypes.h).
<li>When a fragment shader and vertex shader are linked (see slang_link.c)
the varying vars are matched up, uniforms are merged, and vertex
attributes are resolved (rewriting instructions as needed).
</ul>
<p>
The final vertex and fragment programs may be interpreted in software
(see prog_execute.c) or translated into a specific hardware architecture
(see drivers/dri/i915/i915_fragprog.c for example).
</p>
<h3>Code Generation Options</h3>
<p>
Internally, there are several options that control the compiler's code
generation and instruction selection.
These options are seen in the gl_shader_state struct and may be set
by the device driver to indicate its preferences:
<pre>
struct gl_shader_state
{
...
/** Driver-selectable options: */
GLboolean EmitHighLevelInstructions;
GLboolean EmitCondCodes;
GLboolean EmitComments;
};
</pre>
<ul>
<li>EmitHighLevelInstructions
<br>
This option controls instruction selection for loops and conditionals.
If the option is set high-level IF/ELSE/ENDIF, LOOP/ENDLOOP, CONT/BRK
instructions will be emitted.
Otherwise, those constructs will be implemented with BRA instructions.
</li>
<li>EmitCondCodes
<br>
If set, condition codes (ala GL_NV_fragment_program) will be used for
branching and looping.
Otherwise, ordinary registers will be used (the IF instruction will
examine the first operand's X component and do the if-part if non-zero).
This option is only relevant if EmitHighLevelInstructions is set.
</li>
<li>EmitComments
<br>
If set, instructions will be annoted with comments to help with debugging.
Extra NOP instructions will also be inserted.
</br>
</ul>
<a name="validation">
<h2>Compiler Validation</h2>
<p>
A new <a href="http://glean.sf.net" target="_parent">Glean</a> test has
been create to exercise the GLSL compiler.
</p>
<p>
The <em>glsl1</em> test runs over 150 sub-tests to check that the language
features and built-in functions work properly.
This test should be run frequently while working on the compiler to catch
regressions.
</p>
<p>
The test coverage is reasonably broad and complete but additional tests
should be added.
</p>
</BODY>
</HTML>