Extreme Tux Racer

Extreme Tux Racer 0.8.4 runs with the right picture and sound at about 59 fps in the benchmark, not a vsync cap (59 uncapped too), after our patches cut a frame from 19,400 GL calls to 850. Its played session (2026-10-24, with the controller patch) had a 58 fps median, the controller fully usable, and 608 sound xruns the player didn’t hear, which the new SFML audio period is meant to remove.

Packages and patches

extremetuxracer, extremetuxracer-data (0.8.4-2+ps3.N, packages/etr), replacing Debian’s. SFML 3.1 already has its big-endian Ogg fix.

• 0001-0006 send fewer, larger GL submissions, each behind a switch (ETR_GL=0 turns all off): a bind cache and the HUD numbers (ETR_GL_BIND, ETR_GL_HUD), the terrain from buffer objects in one quadtree walk (ETR_GL_TERRAIN), Tux’s spheres from cached vertex arrays and the shadow in one draw (ETR_GL_TUX, ETR_GL_SHADOW), track marks as arrays of what the camera sees (ETR_GL_TRACKS), trees and items in one draw per texture (ETR_GL_TREES, ETR_GL_TREES_GROUP), snow spray, falling snow and curtains in one draw (ETR_GL_SNOW, ETR_GL_PARTICLES). A frame went from 19,400 GL calls and 697 draws to 850 and 72 (the investigation).
• 0007 and 0008, game controllers: SFML 3 reads /dev/input/js*, so a button is the joydev index (DualShock 3 through hid-sony: 0 Cross, 1 Circle, 2 Triangle, 3 Square, 4 L1, 5 R1, 6 L2, 7 R2, 8 Select, 9 Start, 13-16 the d-pad; L2/R2 are buttons, not axes). In menus the d-pad and left stick are the arrow keys, Cross and Start Enter, Circle Escape; in a race the left stick (analog) or d-pad steer, R2/R1 paddle, L2/L1 brake, Cross jumps (hold to charge), Square tricks, Triangle changes the view, Select pauses, Start opens the race’s menu. The keys keep working and nothing is written to the options file; the scheme is on the Help screen in a second column.
• libsfml (packages/patched/audio/libsfml/, 3.1.0+dfsg-4+ps3.N): SFML 3.1 (miniaudio) opened the ALSA device with a 64-frame period and 3 periods, a 4 ms buffer, which ETR can’t change; ours opens it with a 1024-frame period (the investigation).

How it performs

• Bench (batch 2026-10-18-batch17): 59 steady, 2.9 s of stalls (11 frames over 100 ms), 1% low 27.3, longest 483 ms, 29 MB free, 58 MB of VRAM, no underruns; the window holds the game-over screen and a reload. Picture and sound right.
• Its CPU outside Mesa is mostly freetype and harfbuzz (29% of the samples) and SFML audio.
• Its memory grows (MEMORY.md, the MEM=1 batch of 2026-10-18 on Mesa ps3.299): over a 117 s window malloc’s used memory went 17.5 to 18.5 to 19.9 MB (+2.4 MB) and its GART 8.8 to 12.8 to 16.8 MB (161 to 163 handles), two new buffers of about 4 MB, one a minute. The candidate is the track marks in the snow, a growing list drawn through our stream buffers; only a longer run tells a cap from a leak (the game’s todo.toml; the long-session soak).

The benchmark (recipe): ETR has no demo mode, so the user’s race was recorded as keyboard events (evemu-record, variant record) and is replayed into a virtual keyboard from when ETR’s window appears. assets/etr-race.evemu: Practice, one course, 69 s, the menus up to 24.7 s; the race is raced again after the game-over screen from ETR_LOOP_FROM up to ETR_LOOPS times. fps from the HUD window (WINDOW_STOP=1). ETR’s physics take the frame time as their step and the inputs replay at wall-clock times, so at another frame rate Tux’s line can drift from the recorded one, more in later races.

Investigations

• 2026-10-17 game-side batching: 52 to 59 fps, the CPU 83.6 to 53 s.
• 2026-10-24 the 4 ms sound buffer.

Research

What the surveys and plans said about it, newest first; each line names its source, the section and the date.

• 2026-10-15, the user’s first wish list (backlog.txt, commit 7d58e3d), “2. GPU driver and game support”: a strong candidate: GL 1.x and simple, a nice controller game.