Repository navigation
Tags
Tags
Present: a broken completion fence falls back to glFinish instead of … …ending the run Two tester reports (79fed084 ROCKNIX, d37c9b06 AisRock; both RK3562 with libmali g24p0) ended at the title screen with "Presentation completion fence failed EGL=0x3000 GL=0x0": glClientWaitSync on the presentation thread's post-blit fence returned GL_WAIT_FAILED with no GL error although the frames were fine (60 fps for ten seconds). The same SoC on dArkOS is unaffected, so it is that driver build. The first GL_WAIT_FAILED now marks the fence as broken, logs once and finishes that frame and every later one with glFinish; drivers that honour the fence are unchanged (matrix match on RG351P r13p0 clean, branch never taken). Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
PortMaster release notes: the defect sweep, the Speed and Controls pa… …ge and the doldecomp sync Verification of this build: host playthrough of every mode (about 1,270 matches at 8x), all 11 matrix cases on the RG35XX SP and the RG351P, 20 minutes of every mode on a Pi 5. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
PortMaster release notes: the upstream sync and where the dynamic lib… …stdc++ was checked The release carries the doldecomp merge (35 commits) as well as the dynamic libstdc++ link, so the notes name it and say how it was checked. The libstdc++ entry lists the systems the released configuration ran on (Raspberry Pi OS, AmberELEC, stock Anbernic, and symbol resolution against glibc 2.30 with libstdc++ 6.0.28) and points at the previous build. Co-Authored-By: Claude Code <noreply@anthropic.com>
portmaster: Aurora merged with encounter/aurora main; a damaged pipel… …ine cache no longer crashes start-up Pins zalo/aurora-arm gles-direct-submission at dbf7f91: the port's branch merged with 25 upstream commits (input rewrite, layout-keyed pipeline references, CARD fixes, SDK header updates), the pipeline-cache fix for a damaged database (sqlite3_clear_bindings on a finalized statement), and the port's east = B / west = X layout kept for Xbox-type pads. Upstream now clears only the copy rectangle on an EFB copy clear. Melee's copies cover what the frame goes on to use, and pass fusion and the small-copy interval rely on whole-target load-op clears, so the game sets AuroraConfig::wholeTargetCopyClear; MELEE_FLIP_COPY_CLEAR_RECT=1 selects upstream's behaviour for comparison. [flip-config] logs the setting. Co-Authored-By: Claude Code <noreply@anthropic.com>
portmaster: compile the game's text literals as Shift-JIS; renew the … …index streams for libmali r20p0 Two fixes, released together. Blank name plates at character select, a name-entry keyboard with the same letter on every key, no "Lv." numbers in the Event Match list: the game's text routines read Shift-JIS byte pairs, the decomp sources spell those strings as UTF-8 and the GameCube build converts each file with sjiswrap before compiling. The native build compiled the UTF-8 bytes, so every glyph lookup missed. native/tools/sjis_literals.py rewrites the string and character literals of the eleven affected files as CP932 octal escapes (behind a #line marker, so diagnostics still name the original), and CMake compiles those copies. Compiler-independent on purpose: the host build is clang, the cross build gcc. Tests cover the converter and check the packaged binary holds the CP932 bytes and not the UTF-8 ones. Memory growth for the length of a session on the Allwinner H700 handhelds (RG35XX, RG34XX; soak reports 6cdec219 and 4b5e0568): libmali r20p0 keeps a table per index buffer inside glDrawRangeElements and doubles it for as long as the same persistently mapped buffer is drawn from. Aurora 40071d0 replaces each mapped index stream every 256 frames (AURORA_GLES_INDEX_STREAM_FRAMES, 0 = never). RG35XX SP, one hour of the soak playlist: 246 MB anonymous memory and flat, against 431-451 MB before, same frame rate. Hour-long runs on the Miyoo Flip and RG351P and a half hour on a Raspberry Pi 5 show no change. The combined build ran a match, name entry and fifteen minutes of the playlist on the RG35XX SP, Miyoo Flip and RG351P (character select captured on the last two) and fifteen minutes of the playlist on the Pi 5, without an error. Co-Authored-By: Claude Code <noreply@anthropic.com>
portmaster: start on EGL drivers with missing entry points (TrimUI Sm… …art Pro), say so when no driver works Report 5c135d7e (TrimUI Smart Pro, stock OS): "Couldn't get proc eglCreateImageKHR", Null backend, abort. The PowerVR driver lists EGL_KHR_image_base without its entry points, and on Knulli its eglGetProcAddress returns null for core functions. The dlsym fallback added for the Knulli case on 2026-09-27 patched DisplayEGL::InitializeWithDynamicLoading, which this port never runs: Aurora hands Dawn MeleeFlipEGLProc and the display, so Dawn takes CreateFromProcAndDisplay. - display.cpp: MeleeFlipEGLProc falls back to dlsym when eglGetProcAddress has no answer, and the log gets one "[flip-display] EGL <version> (<vendor>): <extensions>" line. - dawn-egl-optional-ext-procs.patch: an EGL extension whose entry points are missing is dropped with a warning instead of failing the backend (reusable sync goes with a missing sync family below EGL 1.5). - runtime_main.cpp, Melee.sh: when Aurora falls through to the Null backend the game exits with "No usable graphics driver" and the launcher says so, instead of aborting a moment later. Checked with an LD_PRELOAD stand-in for the driver under qemu (both reports reproduce on the released build and run with this one); regression runs on the Flip, RG351P, RG35XX SP and Pi 5. Co-Authored-By: Claude Code <noreply@anthropic.com>
portmaster: Mesa 25.3-26.1.3 dropped the shaders' immediates; stop th… …e uniform array trimming there Same linker bug as the cached-program crash: once the explicit-location tint_immediates array is trimmed, its location keeps the "inactive explicit location" marker and glUniform on it is ignored without an error, so the shaders never get their immediates (fighters drawn as huge spikes, textures missing). Found by the tester with the Pi 5 on Batocera 43 (Mesa 25.3.6) once 60a6819 no longer crashed there, together with the driver option that avoids it. The restart for affected Mesa releases now also sets disable_uniform_array_resize=true (a driconf option Mesa reads from the environment at screen creation). The shader cache stays off. Reproduced under qemu with Mesa 25.3.3 (llvmpipe): 60a6819 draws garbage in a match, this build's frame of the same frozen match state matches Mesa 22.3.6 apart from particle phase; warm relaunches and the Mesa 22.3.6 control are unchanged. Co-Authored-By: Claude Code <noreply@anthropic.com>
portmaster: work around the Mesa 25.3-26.1.3 cached-program crash at …
…start-up
Tester report 777957ff (Pi 5, Batocera 43, Mesa 25.3.6): SIGSEGV at
0xffffffffffffffff in libgallium on the first pipeline of every launch after
the first. Those Mesa releases keep the "inactive explicit location" marker in
the uniform remap table of a program whose explicit-location uniform array had
trailing elements trimmed (Tint's tint_immediates), and dereference it when the
program is read back from a cache: glProgramBinary, and their own on-disk
shader cache under glLinkProgram. Fixed upstream in 26.1.4 and 26.2.0
("util/u_range_remap: allow insert to truncate range").
On an affected GL_VERSION the game now re-executes itself once with
MESA_SHADER_CACHE_DISABLE=true, and reports no program binary formats to Dawn,
since Mesa goes on offering binaries with its disk cache off.
MELEE_TEST_MESA_CACHE_QUIRK forces the path on any driver.
Reproduced and verified under qemu with Mesa 25.3.3 (llvmpipe); the restart
path also checked on a Pi 5 (Mesa 26.2.2, X11) and an RG351P (libmali, KMSDRM).
Co-Authored-By: Claude Code <noreply@anthropic.com>
portmaster: on-screen frame rate option, permanent latest-download li… …nk in the docs Port Settings > Show Frame Rate (settings.cfg show_fps, MELEE_SHOW_FPS for one launch) draws the presented frames a second in the top right corner, refreshed every half second, sized relative to a 480-line screen like the online lobby. When the game's own update rate over the last two seconds is under 95 % of the retrace rate, the speed is shown beside it. The [perf] log lines are unchanged. The release job already publishes every release as the latest one under fixed asset names, which the soak updater relies on. The release notes, README and testing thread now give https://github.com/zalo/melee/releases/latest/download/melee.zip, and WorkflowTests holds the workflow and soak/update-url.txt to it. Co-Authored-By: Claude Code <noreply@anthropic.com>
PreviousNext