Warzone 2100

Warzone 2100 4.3.5 (our ppc64 build, gamedb/games/warzone2100/packages/warzone2100/: warzone2100 4.3.5-1+ps3.N, plus warzone2100-data and warzone2100-music, both arch all; Debian has 4.5/4.7, which need GL 3): smoke test, not benchmarked. 4.3.5 is the last release with an OpenGL 2.1 / GLES 2 renderer (GLSL 1.20 shaders: 19 of them, no instancing, UBOs or texture arrays; docs/research/gl3-census/C-rts.md). The game asks SDL for a 3.1 core context first (Mesa's nv30 refuses it, the log says "Failed to create an OpenGL Core context") and then for a 2.1 compatibility one; gfxbackend stays opengl. What base does: an AI-vs-AI skirmish on Sk-HighGround (2 players, the smallest 2-player map: 48x96 tiles; SemperFi both, Easy and Medium; base level 2 = bases with defences, no scavengers), straight from the command line, no input needed: `--autogame --skirmish=ps3bench.json`. The local player is an AI too, so nobody is needed at the keyboard; the camera sits on that player's base (the player with the HUD, power bar and minimap) and the other AI is played against it at normal speed. Upstream's --autogame runs the game at 500x speed (for automated tests: 324 s of game time in 13 s of real time on the PC); our patch 0002 reads WZ_AUTOGAME_SPEED, which this recipe sets to 1. The match is deterministic: it ends (the game quits) when player 0 is defeated, at 318002 ms of game time (5.3 min), on the PC and in the ppc64 build run headless under qemu alike. A 5.3 min match at 1x is shorter than the default window (load + 60 s warm-up + 120 s): the run just ends with the game. Variant lobby: the same settings but as a human: `--skirmish=ps3bench.json` opens the skirmish lobby with the map and the AI already set; click "Start Hosting Game" (bottom of the options column) and then the READY box of the Player row (top right), and the game starts with you as the player against SemperFi (mouse: edge scrolling, left-drag selects, right click orders; Esc is the menu). Variant menu: the main menu. Variants nomusic, nosound: audio off. Settings: a private config directory, $HOME/.local/share/warzone2100-ps3bench (--configdir; the game's own ~/.local/share/warzone2100-4.3 stays untouched), with config written by setup() (BACKUP restores it; the game rewrites it at exit) and tests/ps3bench.json (the match). The lowest settings that keep the picture right: fullscreen 1920x1080, texturesize 512 (the default 2048 costs +66 MB of texture RAM on the PC; 128 and 512 measured the same), shadows off, fog off, vsync off, no anti-aliasing, UPnP off, FPS counter on (bottom right), the music at its default volume (the CPU cost of Opus on the PPE is unknown: variant nomusic). 4.3.5's textures are PNG (the Basis Universal encoder is off in our build): the game compresses RGBA pages to DXT5 itself when it uploads (texturecompression=1, the default; 485 of 1348 textures), the rest is RGBA8 or LUMINANCE (on big-endian the game's own DXT encoder, etcpak, made wrong blocks: black terrain and missing models until patch 0004; GCC 16 at -O3 -mcpu=cell broke the mipmaps, made by stb_image_resize, until patch 0007: games/smoke-tested.md). First start only: the game writes its stats.db and caches in the config directory. Videos: the campaign videos (sequences.wz, the warzone2100-video package) aren't installed; fmvmode is irrelevant to that, the game shows a "Campaign videos are missing" notification at the menu and over the match until it is dismissed with "do not show again", so setup() writes that choice into notifications.json. Risks: RAM. Measured on the PC (amd64 build of the same source, Debian sid, llvmpipe with MESA_GL_VERSION_OVERRIDE=2.1, Xvfb 1920x1080, these settings, 1x speed AI-vs-AI): VmHWM 366-379 MB (97 MB of it libLLVM and Mesa's libraries); heaptrack: peak heap 259 MB = llvmpipe 118 (incl. the textures, ~77 MB live) + its JIT 49 + QuickJS 72 (the AI scripts, rules and the stats as JS objects: the game's biggest own consumer) + 20 other. Human against one AI: 370 MB (the second AI costs only ~5). Without any GL (--headless, 500x autogame, PC): VmHWM 164 MB, which is the game itself with its libraries. PS3 estimate: 180-205 MB (that 164 plus ~15 for Mesa's nv30 userland, or the census formula: RSS - 100 llvmpipe baseline - texture MB - buffers + 10); the textures are VRAM here. Knobs measured (PC RSS): texturesize 2048 +66 MB, texturecompression=0 +72 MB, shadows on +2, fog on +1. Watch dmesg.log, MemAvailable and swap-ins. Other things to look at: the first load (the game reads base.wz, 180 MB, and mp.wz through physfs and builds the stats: 10 s on the PC, ~10x that on the PPE?), the JS AIs' CPU (semperfi runs every game tick, 10 per second), the GL call load (171 draws and ~2900 GL calls per frame, 429 glTexParameteri per frame: PPE-bound), nouveau "out of temps" on the terrain or water fragment shaders (grep stderr and the game log), and the update check the game starts at launch (a curl to wz2100.net; it fails quietly without network). Benchmark idea (not wired up): the base command again as a SUITE=1 recipe with the HUD window: a fixed match (deterministic, see above, so the scene is the same each run), WARMUP 30 s after the load, MEASURE 120 s (the match ends at 318 s of game time, so the window must fit before that); fps from the Gallium HUD (the game's own "FPS:" counter, bottom right, is in screenshots). The AI-only cost with no drawing: `--autogame --skirmish=... --headless` (WZ_AUTOGAME_SPEED=1 for real time, 500 for the whole match): 198 s at 500x under qemu on the PC, a PPE timing wanted.