ioquake3

Debian’s ioquake3 source with our patches: its package.

Technical notes

Our build of Debian’s ioquake3 (1.36+u20260719.58839361+dfsg-1+ps3.N, packages/ioquake3) replaces Debian’s and carries 30 patches (there is no 0017). OpenArena runs on it (its notes), and Quake III Arena with the player’s own data (its notes). Every change sits behind a cvar, so a recipe variant can turn it back to upstream’s behaviour. The PS3 runs ps3.959~dev (0031) since 2026-10-28; the release candidate is ps3.396 (the archived release plan).

renderergl2 memory (0001-0010)

These are the patches that made the GLSL renderer fit in 217 MB of RAM (the investigation).

Patch	What	Cvar back to upstream
0001	GLSL shader objects deleted once their program links (Mesa keeps no source or IR)	`r_glslKeepShaders 1`
0002	no GPU bone-animation programs (27 programs index a 105-matrix array, which nv30 turns into compare trees)	`r_glslMaxBones -1`
0003	none of the 15 shadow and post-processing programs that need framebuffer objects	`r_glslAllPrograms 1`
0004	compiled programs kept across renderer restarts (no second compile burst at the map load)	`r_glslKeepPrograms 0`
0005	opt-in: each program built at its first use (`r_glslLazy 1`); it adds stalls when a variant is first drawn
0006	the VAO cache allocated only with `r_vaoCache` (21 MB never drawn from)
0007	no permanent CPU copy of MD3 vertex frames (about 87 MB of hunk for the bench demo’s models)	`r_keepModelVerts 1`
0008	no tangents in MD3 VBOs without normal mapping
0009	`r_md3Lods`, and no LODs that `r_lodbias` never draws
0010	`r_gpuVertexAnimation 0` animates MD3 on the CPU as opengl1 does

Engine and opengl1 CPU (0011-0018)

On in every build since ps3.362: no rescan of the shader text after a hash miss (r_shaderTextScan 1 for the old way), no crc32 of pk3 entries (fs_zipCrc 1), the AltiVec MD3 interpolation and lighting built and on (com_altivec 0), no memsets per decal clip plane (r_markClear 1), a player’s sounds loaded at their first play (s_alLazy 0), pk3 entries inflated by libdeflate (fs_zipFast 0), the message Huffman trees built from tables (msg_huffSlow 1, a command-line cvar). They were never A/B’d one by one on the PS3; the recipe’s up variant turns all of them back.

opengl1 batching (0019-0024)

The lightmap atlas r_mergeLightmaps (0019, 0023: values 1 and 2, r_mergeLightmapsSize), a GL state cache r_glStateCache, no glLockArraysEXT around batches (r_lockArrays 0) and the sky box from vertex arrays r_skyArrays (islanddm’s sky went from 1453 calls to 18). On llvmpipe at GL 2.1 the bench demo went from 3022 GL calls and 198 draws a frame to 377 and 79, pixels within 1-4 levels and identical with every switch off. 0024 makes r_mergeLightmaps 0 the default, because any atlas page cost about 10 s of hitches a run on the RSX (the investigation).

Loads during play (0025-0026)

r_keepImages 1 keeps the textures of image files across a reload of the same map; cl_preloadLate 1 registers again, at the end of the next load of the same map, what the cgame loaded during the last one’s play (bots joining). Both are on by default. cl_deferPlayers 1 (opt-in) ignores loaddefered, so bots keep a substitute model, which changes the picture (the investigation).

renderergl2 vertex formats (0027-0029)

Patch	Cvar (default)	What
0027	`r_vaoSnormColor 1`	vertex colours as GL_SHORT normalized, which nv30 fetches, instead of 16-bit unsigned that u_vbuf converted at every draw (and that made nouveau keep malloc copies of every buffer it read: about 62 MB for the demo’s models)
0028	`r_mipMapPowf 0`	the gamma-correct mipmap bytes from a threshold table instead of a `powf` per texel channel (7.8% of a profile’s samples); the same bytes
0029	`r_md3PackedFrames 0` (opt-in, needs `r_normalMapping 0`)	animated MD3 frames in 12 bytes a vertex instead of 20 (16-bit positions, 8-bit normals): the demo’s model VBOs 62.2 to 37.5 MB of VRAM

The investigation has the measurements.

Quake III’s menus (0030)

Patch	Cvar (default)	What
0030	`cl_cdkeyCheck 1`	0: the UI’s CD key check answers valid, so Quake III’s and Team Arena’s menus don’t open on the CD KEY screen at every start (`ui_cdkeychecked` is `CVAR_ROM` in both UIs: a value set from outside is forced back to 0). Quake III’s autoexec.cfg sets 0; no key is written or read

Signals (0031)

Patch	Cvar (default)	What
0031	`sys_signalQuit 1`	SIGTERM, SIGINT, SIGHUP and SIGQUIT only set a flag, and the main loop quits before its next frame as the `quit` command does (exit status 1, `Client Shutdown (Received signal 15)` as before, the PID file removed); a second one exits at once, and `alarm(10)` ends a quit that doesn’t come. Upstream ran the shutdown inside the handler, which hung 4 of 6 SIGTERMs on the PS3 (libxcb’s lock, held by the interrupted buffer swap, wanted again by the shutdown’s `XGrabPointer`); now 12 of 12 exit in 1-4 s. 0: upstream’s handler

The investigation has the stacks.

Sound dropouts

OpenAL Soft’s [ALSOFT] (EE) mmap commit error: Broken pipe (and available update failed: Broken pipe) is an audible dropout, 80-440 ms each: OpenAL’s mixer thread starved past its 32 ms buffer and recovered. It runs at normal priority (no RT limit in the session, no RTKit). 150 s with 5 bots: Quake III 1-2 dropouts a run, OpenArena 1-17; with LimitRTPRIO=10 (OpenAL takes SCHED_RR by itself) Quake III had none in 3 runs and OpenArena 3. The suite measurement the same day (docs/SPU-AUDIO.md “Real-time mixer”) found no audible dropout in 20 live runs with bots and 18 timedemo runs, with the limit or without, and a 0.4-2% fps cost, so the RT limit is not shipped.

Log noise

ioquake3 probes for extensions nv30 doesn’t have (GL_ARB_framebuffer_object, texture_float, seamless_cube_map, texture_compression_bptc, a core context) and carries on; there is no IPv6 in the kernel, so it prints socket errors for it. Builds before 0031 sometimes printed DOUBLE SIGNAL FAULT: Received signal 11 right after Client Shutdown at the harness’s SIGTERM (the shutdown crashing inside the signal handler), harmless to the numbers. The engine’s analysis.toml holds the log rules for these.