Game candidates for the suite (research, 2026-10-16)

This is the survey that picked the benchmark suite, done on 2026-10-16 before any of these games ran on the PS3. Step 0 had rested on OpenArena alone, which misled us about VRAM and sound, and the SPU offload and memory decisions needed 5-10 titles. Every pick has run since, and so have most reserves: Yamagi Quake II, Extreme Tux Racer and Armagetron joined the suite, Teeworlds, SuperTux and Red Eclipse were smoke-tested, TORCS waits for the user, and Nexuiz runs on our own older DarkPlaces build. Current results are in games/README.md; where this survey and the measurements differ, the measurements win.

Where this lives now (2026-10-26). This survey moved from docs/research to gamedb/research. Every game it names has a bundle in gamedb/games/ (a draft when it is only a candidate, status and status_note saying where it stands) whose technical.md cites this file by section. The per-game tables and blocks here are kept by hand for now; they become generated blocks (tools/gamedb/blocks.py) in a later step, and until then a game’s current state is its bundle’s, not this file’s.

Sources: the debian-ports snapshot we pin (20261013T000000Z, ppc64 and all, contrib too), live debian-ports, buildd.debian.org logs, the Debian BTS, upstream trackers and the games’ source at the packaged versions. UNVERIFIED marks what no source or log confirmed. Sizes are installed sizes of what the image didn’t have yet (the dependency closure against build/rootfs/base.tar’s dpkg status). The same day, reviews against the source corrected several entries; the corrected facts are written in below and marked “(review)”.

The yardstick: what Mesa nv30 on NV47 offers

From nv30_screen.c and nv30_format.c (Mesa 26.2.4) and RSX.md:

• GL 2.1 compatibility, GLSL 1.20, GLES 2.0. No GL 3.x context of any profile: engines that ask for 3.2 core with no fallback fail at startup.
• Fragment programs: 4096 instructions, 224 vec4 constants, 32 temps, 8 varyings, 4 render targets. Vertex programs: 512 instructions, about 462 vec4 constants. No real loops (NIR unrolls).
• No vertex texture fetch (0 vertex samplers; nv40_verttex_validate is a stub).
• No float textures: 16- and 32-bit float formats are vertex formats only (no ARB_texture_float, no float FBOs). sRGB sampling yes, sRGB render targets no.
• NPOT, EXT_framebuffer_object, depth textures and ARB_shadow, occlusion queries, MSAA 2/4, S3TC. ARB_framebuffer_object is not exposed (mixed_framebuffer_sizes is false, the cap Mesa ties it to): engines that look only for the ARB name, or GL 3.0, treat FBOs as absent.
• No ARB_texture_rg (review): engines that upload R8 or RG textures fail.
• No texture arrays, instancing, geometry shaders, integer textures, OES_standard_derivatives. 3D textures sample fine (review; rendering into them is refused). TIMESTAMP and TIME_ELAPSED queries return a pixel count, not time (MESA-METHODS.md): any game’s GPU timings are garbage here.
• Known RSX traps: a graph exception wedges the whole RSX until a console reboot; vertex format type 0 and swizzled 1-pixel-wide render targets fault (the validator refuses them, and Mesa patch 0007 covers 1-pixel textures, review); single-attribute inline draws need Mesa 0008. Hardware mipmap generation was a new risk for Neverball, SuperTuxKart and Luanti (review).
• Mesa’s disk shader cache is on for nv30 (review).
• Texture compression at upload: generic GL_COMPRESSED_RGB becomes DXT1 and GL_COMPRESSED_RGBA DXT5 (st_format.c). libspeq’s encoder did DXT1 only then, so DXT5 fell back to Mesa’s PPE encoder at 2.6 Mpixel/s. (DXT5 on the SPEs came with libspeq 0.367 and Mesa 0032, 2026-10-18.)

Surprises

• Four Debian games don’t build on ppc64 (latest buildd logs):
  • wesnoth-1.18: links fail on boost::context (Debian #1146645: Boost doesn’t set BOOST_USE_UCONTEXT on ppc64). Not a Wesnoth bug; it built on powerpc and s390x.
  • hedgewars 1.0.3: builds, then 2 of 7 engine Lua tests fail (EInvalidOp in stuckcake.lua, wrong positions in gravity_get_set.lua). Looked like FPC floating point on PPC, not byte order (UNVERIFIED then; later reading of the build log: two different failures, see Hedgewars’ investigation).
  • ddnet 20.1.1: the test run hangs in Net.Ipv4AndIpv6Work and the buildd kills it.
  • warzone2100 4.7.0: a real big-endian bug, a static assert in the bundled GameNetworkingSockets (minbase_endian.h:26, LittleWord() on a uint64). Moot anyway: it needs GL 3.0.
• Luanti has no ppc64 binary: debian/control lists only LuaJIT architectures. Big-endian itself works: s390x builds 5.15.2 and passes --run-unittests. A 2025 ppc64 attempt (5.10.0, ports LuaJIT) aborted in TestSAO. LuaJIT exists for ppc64 and Debian strips the bundled Lua, so our build keeps LuaJIT (review; building without it means restoring lib/lua and -DREQUIRE_LUAJIT=FALSE).
• contrib is arch:all only on debian-ports: nothing arch-specific there is built, so Yamagi Quake II (source package yquake2, review) doesn’t exist for ppc64 although it builds on s390x and ppc64el.
• Debian’s darkplaces is GL 3.2 core only (since 0~20180908), so Nexuiz and Xonotic are out with Debian’s engine.
• BananaBread (Debian’s cube2) and Red Eclipse turn texture compression off on Mesa: if(!mesa) usetexcompress = 2 in rendergl.cpp; only pre-made DDS files go in compressed.
• QuakeSpasm 0.96.3 has a big-endian bug: wrong colours and dlight wraparound from the packed-pixels path (quakespasm issue #114, a G5 on Debian ppc64); -nopackedpixels avoids it, upstream master disables the path on big-endian.
• Freedoom 0.13’s built-in demos desync (freedoom issue #1532).
• Hardly any game runs tests on the buildds (most dh_auto_test steps do nothing), so “builds on ppc64” is the only big-endian evidence for most. Exceptions: OpenTTD (97/97 ctest pass on ppc64), Frozen Bubble (tests pass on ppc64), Luanti (s390x).
• The OpenArena GLSL renderer (cl_renderer opengl2) was tried once: commit 82b6b5f (2026-10-15) went back to opengl1 because “the GLSL one ran out of RAM compiling its shaders”, no log kept. That was before the VRAM swap and game mode, so it deserved another try.

Candidates

Tier is driver readiness: A likely runs as is, B likely needs driver work, C unknown until tried, X can’t run on NV47. MB = installed (engine + data + new libraries).

Every game here takes a resolution on the command line or in its config and none was expected to need a window manager for fullscreen, except possibly SuperTuxKart (IRR_DISABLE_NETWM=1 forces its legacy fullscreen; UNVERIFIED). In practice SDL’s fullscreen windows needed one: the runner starts Openbox.

The recommended suite and why

Ten games, eleven entries (OpenArena counts twice: its two renderers are different workloads).

PrBoom+ GL (on the user’s list) is covered by dsda-doom, its successor with the same GL renderer, packaged for ppc64. Sound stacks: OpenAL Soft (1-3, 6, 7), SDL audio with the game’s own mixer (4, 8, 10), SDL_mixer (5, 9, 11); SFML audio only in Extreme Tux Racer. Compressed textures: OpenArena (both renderers), SuperTuxKart (a setting, effect unverified), BananaBread (with a one-line patch). For a DXT5 case without patching, Teeworlds and TORCS were the reserves.

Reserves, in order: Yamagi Quake II gl1 (an OpenAL shooter with a timedemo; needs our build), Extreme Tux Racer (SFML audio; no benchmark), Armagetron (a long-running stability soak), Teeworlds (DXT5 via gfx_texture_compression, 3D textures), TORCS (driver-side DXT1/DXT5), SuperTux.

The suite that ran has 13 games: these ten plus Yamagi Quake II, Extreme Tux Racer and Armagetron.

Per-game notes from the research

What the survey expected for each pick, before the runs (the commands that run now are in games/SUITE.md):

• OpenArena opengl2: renderergl2 (read at upstream main) needs GL 2.0 and #version 120 shaders, 30 source files, about 66 KB, the biggest lightall_fp 13.8 KB; about 100-115 programs compiled at every start, no binary cache, any compile or link error fatal. FBO features (HDR, tone mapping, sun and player shadows with 1024 and 512 depth maps, SSAO with an R32F target) need ARB_framebuffer_object or GL 3.0, which nv30 exposes neither of, so they should be off by themselves. Cubemaps are forced off below GL 3.0; no vertex texture fetch. The likely RAM cause: bone variants index a 105-entry uniform array, which nv30 turns into four 105-leaf if/else trees per program, compiled at every start though OpenArena’s models never use them; about 200 shader objects stay attached all session. Plan: min settings first (normal, specular, deluxe mapping, sun, dlights, depth prepass, HDR, post-processing, tone mapping, auto exposure, SSAO, cube maps and FBOs off; lightall takes its USE_FAST_LIGHT path, about 18 fewer programs without the sun), then Debian’s defaults; peak RSS during R_Init.
• ClassiCube: no timedemo; a fixed spawn and the HUD, a flight-path plugin later.
• QuakeSpasm: -noglslgamma (a full-screen GLSL pass at 1080p) and -noglsl as variants; load time from the out.log stamp of the first demo line.
• BananaBread: no demo playback in the client was assumed, so a fixed spawn for a fixed time; texcompress (textures of 1024 and up) only works when usetexcompress is set, which the engine skips on any GL_VERSION containing “Mesa”: a one-line patch (engine/rendergl.cpp:536).
• SuperTuxKart: --benchmark with force_legacy_device replays benchmark_black_forest.replay and writes a CSV; the GPU CSV uses timer queries, wrong on the RSX. Lower RAM through max_texture_size, HD textures off, geometry_level, animated_characters off.
• Luanti: viewing_range 60, client_mapblock_limit 1000, mesh_generation_threads 1, fps_max 1000; a ps3bench mod to fly a fixed path and call core.request_shutdown().
• Neverball: fps only on screen and it stops at a menu, so either the HUD with a timeout or a 10-line patch in st_demo_play (that patch is what we did).
• dsda-doom: record our own demo with the same version and IWAD (-record bench -warp 1 -skill 4); sound on by default.
• OpenTTD: simulation -v null:ticks=N timed by the runner minus a ticks=1 run for loading; rendering for a fixed time with HUD fps. Music is MIDI through fluidsynth with a soundfont (TimGM6mb comes with the dependencies).
• X-Moto: replays are little-endian on every host; paced in real time, about 144 s, so the timeout must be above 180 s.

Packages proposed then: a ps3-games metapackage next to ps3-apps (Depends on the light picks, Recommends the heavy ones); the Luanti engine as our package with its unit tests in the build; the Quake shareware deb from game-data-packager (non-free, LAN repository only, installed with --no-install-recommends because game-data-packager-runtime pulls GTK 4; LibreQuake, a free replacement, only in Debian experimental); bench assets of our own in its benchmark (gamedb/games/<id>/benchmark/assets/).

The planned first session

One launch per game, the first failure of each recorded, a driver bug list ordered by how many games each blocks; through ps3-bench.sh with the VRAM swap off and a 180 s timeout. Because a graph exception ends the session, the order went from safest to riskiest:

Out of scope: anything with tier X (darkplaces from Debian, Tesseract, Warzone 2100).

Sources

• Package data: snapshot.debian.org/archive/debian-ports/20261013T000000Z (main ppc64 and all, contrib ppc64), deb.debian.org/debian-ports (live), deb.debian.org/debian Contents-all.
• Build status and logs: buildd.debian.org/status/package.php?p=<src>&suite=sid and logs.php?pkg=<src>&arch=ppc64; Boost: bugs.debian.org/1146645.
• Upstream: github.com/ioquake/ioq3 (code/renderergl2), github.com/sezero/quakespasm/issues/114, github.com/yquake2/yquake2/issues/138, github.com/freedoom/freedoom/issues/1532, supertuxkart.net/FAQ and stk-code 1.5 (main.cpp, central_settings.cpp, profiler.cpp), luanti 5.15.2 (doc/os-compatibility.md, builtin/settingtypes.txt, cmake/Modules/FindLua.cmake), OpenTTD 15.1 (src/video/opengl.cpp, null_v.cpp), warzone2100 4.7.0 (lib/ivis_opengl/gfx_api_gl.cpp), cube2 and redeclipse src/engine/rendergl.cpp, tesseract rendergl.cpp, xmoto 0.6.3 XMArgs.cpp, StateReplaying.cpp, dsda-doom prboom2/src, game-data-packager 89 data/quake.yaml, data/quake2.yaml.
• Mesa 26.2.4: src/gallium/drivers/nouveau/nv30/nv30_screen.c, nv30_format.c, src/mesa/state_tracker/st_extensions.c, st_format.c.

Older versions and the dated history: archive/GAME-SUITE.md.