RadeonGL 0.15 -- 14.11.2026
WHAT IT IS
A minigl.library that draws with the 3D engine of an ATI Radeon from the
R200 family (8500, 9000, 9100, 9200, 9250) on a graphics card that
Picasso96 drives. No Warp3D: it talks to the Radeon's command processor
through Picasso96, while the Workbench keeps running.
It is a minigl.library 29.1 with dispatch ABI 5, so programs built with
the current MiniGL SDK open it like any other minigl.library 29.
SOURCE CODE
https://github.com/bubbob42/RadeonGL
The archive carries it too, in the drawer Source (no icon: Show All
Files, or a Shell): the library, the Installer, the tools, the demos, the
tests and the makefiles. Building needs m68k-amigaos-gcc, the Picasso96
developer files and, for the probes, Thomas Richter's openpci68k.lha
(Aminet driver/other); Source/README.md explains it.
FEATURES
* Every GL primitive, in immediate mode and from vertex arrays (8-, 16-
and 32-bit indices)
* Compiled vertex arrays as indexed draws; an upload cache keeps what a
program sends again unchanged in the card's memory
* Transformation and rasterising on the Radeon, the full float matrix
stack
* Two texture units; RGB565, ARGB4444, ARGB8888 (GL_RGB8, GL_RGBA8),
I8, AI88 in up to 16 MB of card memory; nearest, linear and mipmap
filters; repeat and clamp; MODULATE, REPLACE, DECAL, BLEND, ADD
* Lighting on the Radeon: eight lights, directional, local and spot,
with attenuation, two-sided; materials, colour material, normal
arrays
* Texture coordinate generation (object linear, eye linear, sphere
map), texture matrices, fog, wide lines, glReadPixels
* Blending, alpha test, depth test, polygon offset, culling, scissor,
colour mask, flat and smooth shading, point size
* The rest of MiniGL 29.1's calls: GLU (gluPerspective, gluLookAt,
gluBuild2DMipmaps, spheres, cylinders, disks), glArrayElement,
interleaved arrays, double-precision calls, uploads and glReadPixels
in every GL 1.1 layout and type, MiniGL's own main loop
* A Picasso96 screen of its own, 16 or 32 bit, up to 2048 x 2048, in
every mode of the card, switched while the program runs; vertical
sync, or three buffers without tearing
* Or a window on the Workbench (the program's or one of its own), or
straight into the program's bitmap, on 15-, 16- and 32-bit screens
* Several programs at once, each with its own context
* Every command list checked before the Radeon runs it (can be turned
off for speed); one Picasso96 lock a frame
* Installer, manuals in English and German, a card test, a card report
for bug reports, four demos
* RadeonStat: the card's engine load, command queue, card memory and
display, live in a window with a busy history (read-only)
NOT IMPLEMENTED YET
All 205 functions of minigl.library 29.1 are there. Not yet:
* Polygons as lines or points, stencil and accumulation buffers,
stipple, logic operations, reading the depth buffer
* Generic 32-bit textures are stored with 16 bits (GL_RGBA8 and
GL_RGB8 keep 8)
* Contexts on 24- and 8-bit screens; textures beyond 16 MB are not
swapped out
GLQUAKE BENCHMARKS
GLQuake for R200 68k (game/shoot/GLQuakeR200, the RadeonGL edition of
GLQuake for minigl.library, 1.2), its timedemos on the test machine
(A4000D, 68060 at 50 MHz on a CyberStorm MK II, Mediator, Radeon 9200,
AmigaOS 3.2, Picasso96 3.6), RadeonGL 0.8, command check off
(RadeonGLCheck=0), vertical sync off, frames a second:
Demo 320 x 240 640 x 480 1280 x 720
demo1 33.5 32.5 32.9
demo2 30.6 - 26.0
demo3 28.9 - 27.5
RadeonGL 0.11 runs them as fast as 0.9 (demo1 at 320 x 240 without
vsync, averaged over six load addresses: 29.44 seconds against 29.39).
The Aminet source built for the 68060, on RadeonGL 0.6: demo1 24.4 /
23.7 / 23.3, demo2 21.7 / - / 17.9, demo3 21.4 / - / 19.9. With the
command check on (as installed) slower: 0.4 took demo1 from 24.4 to
19.1. The 68060 sets the pace, not the Radeon.
A tester's A4000D (BFG9060, 68060 at 75 MHz, Prometheus Firestorm PCI
bridge, RadeonGL 0.6, check off, demo1 at 320 x 240 without sound):
43.0 fps. RadeonGL works with both the Mediator and the Prometheus.
INSTALLATION
Double-click Install. It keeps a minigl.library that is not RadeonGL's
as LIBS:minigl.library.old. RadeonGL.guide explains going back.
MADE WITH AI
RadeonGL was made with the help of AI. Claude, an AI model by Anthropic,
wrote most of its code, its tests and its manuals, working under the
direction of Marcus Gerards, who made the design decisions. Every release
was tested on real hardware.
CREDITS
* Thomas Richter: Picasso96 3, its Radeon driver, openpci.library, and
his advice on using Picasso96 with the Radeon
* Alexander Kneer and Tobias Abt: Picasso96
* Hyperion Entertainment (Hans-Joerg and Thomas Frieden): MiniGL
* Steffen "MagicSN" Haeuser: minigl.library 29.1 and its SDK
* Frank Wille, Steffen Haeuser and Dennis Boon: GLQuake for
minigl.library
* AMD: the R200 microcode
* ATI, VA Linux Systems, Precision Insight, The Weather Channel and
Tungsten Graphics, Red Hat and the authors of the Linux radeon and
Mesa r200 drivers: the public record of the R200
* amiga-gcc (Bebbo), amitools (Christian Vogelgsang), FS-UAE (Frode
Solheim), libnix, SAS/C, the AmigaOS 3.2 NDK: the tools
LICENCE
minigl.library contains MiniGL and is therefore under the Hyperion MiniGL
Open Source License: AmigaOS only, and the source code of every change is
published: see SOURCE CODE above. RadeonGL's own code is also under the
MIT licence. The R200 microcode is AMD's (MIT). The licence texts are in
the package (the installed drawer's Licences, and Source/LICENSE).
HISTORY
0.15 (14.11.2026)
- fixed: panning tore the picture (0.14 was wrong to call it tear-free)
- the upload cache starts at 4 MB and grows with the scene
- fixed: a stale upload-cache block under the static arrays hint
- GL_RGL_bitmap_view: board bitmaps as textures without a copy
0.14 (13.11.2026)
- panning (three buffers) 21-40% faster, its task now below the
program: faster than vsync (but it tore, fixed in 0.15);
RadeonGLFlip=csbnowait for tests; faster 32-bit texture uploads,
GL_BGRA 0xAARRGGBB words
0.13 (13.11.2026)
- public port "RadeonGL" (is anything drawing?), GL_ARB_multitexture
named on its own, GL_EXT_texture_env_add, a ChangeScreenBuffer fix
0.12 (12.11.2026)
- triple buffering (panning or ChangeScreenBuffer), vsync without
tearing, wireframe and points, faster lit vertices, contexts open
in 26 ms
0.11 (8.11.2026)
- all 205 functions of minigl.library 29.1: display lists, GLUT
0.10 (8.11.2026)
- 177 of minigl.library 29.1's 205 functions: GLU, MiniGL's main loop,
every pixel layout, copy textures, blend equations, palettes; Trio
0.9 (8.11.2026)
- contexts in a window and in a bitmap, several programs at once,
lighting, texture coordinate generation, fog, wide lines,
glReadPixels, 8-bit-per-component textures
0.8 (7.11.2026)
- every screen mode, live mode switches, vertical sync, no stale
columns at widths not a multiple of 64
0.7 (6.11.2026)
- static arrays and index lists (own extension), 2 MB upload cache,
pointer hidden, card tests on 24-bit Workbenches
0.6 (5.11.2026)
- up to 18% faster: the Radeon draws while the CPU builds the next frame
- texture memory independent of the Workbench's screen mode
0.5 (5.11.2026)
- 16 MB of textures on 8- and 16-bit Workbenches too
0.4 (3.11.2026)
- first release
|