1. Polishing X with Glamor
Keith Packard HP

2. Glamor, KMS and X Delivering on the promise from 2004:
Getting X Off the Hardware (OLS, 2004)
2D Drawing with GL
Hardware-independent X drivers
Developing X with Xephyr

3. Glamor: GL-based X acceleration
Started by Eric Anholt in 2008
Original target was Mesa from that era
GL support in Mesa is far more modern today

4. GL in one slide Vertex shaders Compute destination and texture coords
Hardware
Interpolate coords across primitive
Fragment shader
Fetch texture and other data
Compute destination pixel value operand
Composite/raster op to destination

5. X Core Rendering Geometry Spans, Rectangles, Text, Thin Lines
Arcs, Wide Lines, Filled Arcs Polygons
converted to spans
Fill Style
Solid, tiled, stippled
Raster Op
Bit-wise operations. Irrelevant today.
Plane Mask
Used for colormap tricks. Irrelevant today.
Dash patterns
For lines and arcs

6. Glamor last year A few fixed GLSL programs
Coordinates computed with CPU
Tile/stipples repeated on CPU
Quite a few fallback cases
Not great performance

7. My Glamor Plan Move computation to GPU Deliver raw X coordinates
No fallbacks
Template GLSL construction

8. Constructing GLSL Programs
'facet's
declarations
code
uniforms
Vertex and Fragment templates to glue facets together
Geometry primitive facet
Fill style facet
Built on-the-fly as needed

9. Rectangle Facet
static const glamor_facet glamor_facet_polyfillrect_130 = {
.name = "poly_fill_rect",
.version = 130,
.vs_vars = "attribute vec4 primitive;\n",
.vs_exec = (" vec2 pos = primitive.zw * vec2(gl_VertexID&1,”
“(gl_VertexID&2)>>1);\n"
GLAMOR_POS(gl_Position, (primitive.xy + pos))), };

10. Solid Fill Facet const glamor_facet glamor_fill_solid = {
.name = "solid",
.fs_exec = " gl_FragColor = fg;\n",
.locations = glamor_program_location_fg,
.use = use_solid, };

11. Solid Rectangle Vertex Shader
#version 130
attribute vec4 primitive;
uniform vec4 v_matrix;
void main() {
vec2 pos = primitive.zw *vec2(gl_VertexID&1,
(gl_VertexID&2)>>1);
gl_Position.xy = (primitive.xy + pos).xy * v_matrix.xz v_matrix.yw;
gl_Position.zw = vec2(0.0,1.0); }

12. Solid Rectangle Fragment Shader
#version 130
#ifdef GL_ES
precision mediump float;
#endif
uniform vec4 fg;
void main() {
gl_FragColor = fg; }

13. Core X Text Indexed by character ASCII, ISO-Latin 1, Shift-JIS
8 or 16 bits
16 bits are either “2d” or “1d”
Bitmap per glyph
Glamor stores entire font in one texture
8 bits per pixel to save space
Fragment shader fetches bits with shift/mask

14. Line Dashing GL supports only a few fixed dash pattern
X supports huge dash patterns
Unlimited list of on/off lengths
Glamor stores dash list in texture
0xff for 'on' pixel, 0x00 for 'off' pixel
Vertex shader computes correct interpolation
Checks for X major vs Y major lines
Interpolates dash coordinate along major axis
Dash offset computed for each segment by CPU
Fragment shader fetches dash value

15. Glamor Core Accelerated Operations
Filled Rectangles
Filled Spans
Copy Area
Copy Plane (n to n)
Points
Thin lines (solid and dashed)
Text

16. Interesting Results Solid/tiled/stippled all run at the same speed
Glamor accelerates “weird” operations trivially, like stippled text (which was used in Motif years ago)
GLSL is way nicer than GPU Assembly

18. Lessons from x11perf Large batched operations are awesome rectangles
lines
text
Small operations are dominated by GL library
window configuration
polygons

19. Potential performance ideas
Batch operations across X requests
Leave operations pending until flush
Check previous queued operation
Fewer uniforms, more vertex attributes
Solid colors

20. Render Extension Render added to X in 2000
Anti-aliased text was the motivator
Conflicting goals:
Satisfy 2D application requirements
Perform well on existing hardware
Exposed best to apps with cairo

21. Render Overview

22. Render Objects Sources Images. Regular X pixmaps.
Gradients. Smooth gradations between colors.
Solid colors. Solid colors.
Masks
Any source, but use only the alpha value
Geometry (trapezoids, triangles, rectangles)
Glyphs (special storage for image masks)

23. Render Operations Composite Paint a single rectangle Glyphs
Paint multiple rectangles as described by a sequence of glyph masks
Trapezoid/triangle
Construct a mask from a set of polygons
Paint a single bounding rectangle using that mask and a single source

24. Render Glyphs Client constructs “glyph sets”
Contain a set of alpha masks, one per glyph
Each glyph also has spacing information
Generally one glyph set per face
Arbitrary indices; no semantic value
Draw a list of glyphs from one or more sets

25. Glamor Glyph Storage Glyph sets are often larger than max texture size
Applications often draw glyphs from multiple glyph sets in one operation
Glamor uses glyph-set independent “caches”
Append-only on miss
Flush drawing operation and allocate new cache when full
Avoids complex 2D allocation
Better cache space utilization
Avoids stalling GPU on miss
Never replacing existing glyph with new glyph

26. Drawing Glyphs For each glyph Ensure glyph is in cache
Append glyph info to vertex list
Flush vertex list when cache is full or done drawing.

27. Future Render Work Got side-tracked by GSoC failure
Figure out Render source management
Fragment shader gradient computation
Image sources, including repeat modes
Add composite code
Simple, once source operands works.
Hardware trapezoids?

28. Measuring X Performance
x11perf is not representative of real applications
We used to have cairo traces
Those are fairly stale, and draw only off-screen now
We have gtkperf. It shows only old gtk+ widget-based performance and core graphics.
Few (no?) “real” applications provide any benchmarking mode
For GL, many games provide useful benchmark data internally

29. GL vs X. Sometimes it's not great
Texture formats vs X pixmaps and pictures
GLES limitations
Older hardware

30. GL textures vs X pixmaps
X pixmaps define only depth
Color data defined only in relation to other objects
Core X drawing has no color meaning
Purely pixel values
Render pictures are one usage
Picture format defines pixel format
Can create multiple Pictures per Pixmap
gtk+ actually does this
GL Textures format defined by data
Store texture data and you have a format

31. Glamor fails at this today
GL format approximated by remembering last Render format
Kludges for multiple pictures using different formats in one operation (as Gtk+ does)
Reformat texture when switching formats
This is expensive and wrong

32. Need ARB_texture_view extension
Requires ARB_texture_storage
Allows multiple textures from one bag of pixels
Separation looks much like picture/pixmap
Supported on most hardware
Fallback by fetching/storing data to temporary or new pixmaps.

33. GLES2 limitations GLES2 is popular on embedded GPUs Like VC4
Missing lots of stuff Glamor uses today
GL_QUADS
32-bit integers
instanced drawing
texture pack/unpack formats
logic ops

34. Older Hardware and Glamor
To be specific: i915 and i945
Hardware lacks vertex shaders
executed on CPU
Mesa version is interpreted and slow
Glamor performance is pretty bad here
Options:
Fix Mesa. LLVM vertex shaders on CPU? Special case glamor shaders in Mesa with custom code?
Use existing xfree86 video drivers?
Construct new drawing-only driver interface?

35. One driver to rule them all.
KMS + GBM + Glamor = X Modesetting driver
Merged into the X server code base for 1.17
Track API/ABI changes easily
Uses Glamor where GL is available
Uses software rendering otherwise

36. Modesetting feature support
DRI2
Mesa, VaAPI
DRI3/Present
Missing page flipping in 1.17 Xv
Only textured video for 1.17
RandR
Including shadow buffers for rotation

37. Modesetting saves a bit of code
intel
ati
nouveau
Total
modesetting 5032

38. Modesetting Plans Add page flipping Code should be ready shortly
Fix shadow support
Conflict between “regular” shadowing and RandR shadows
Just force RandR shadowing on when “regular” shadows are requested.

39. Xephyr: Single Machine X Development
Nested X server
Used to be software only
Now includes Glamor support
Allows lots of Glamor debugging without restarting main X server
Vastly simplified GDB experience

40. X 1.17 and 1.18 1.17 contains: First integrated modesetting driver
Improved core X glamor code
For 1.18, we'll add
Page flipping and better shadows for modesetting
At least new glyphs for glamor
GLES support could be fixed with some effort