Repository navigation
AGC: five homebrew titles stop on 'last provoking vertex with explicit per-vertex flat interpolation is unsupported' #3581
Description
Five catalog titles stop in the same pipeline check on main (Windows 11, WinLibs GCC 15.2.0 r7, RTX 4070, main a11c8f7, catalog builds with eboot.bin and sce_module/libc.prx unwrapped, 60 s each):
[videoout] presentation failed: AGC graphics: last provoking vertex with explicit per-vertex flat interpolation is unsupported
terminate called after throwing an instance of 'std::runtime_error'
| Title | Where the line is |
|---|---|
| SymphonyStation5 (PPSA17641) | the title's own log under download0/ |
| VITA5 (PPSA39410) | same |
| Homebrew UI Lab (PPSA99050) | download0/hui/dev/app.log |
| Surround Sound Studio (PPSA99051) | the title's own log under download0/ |
| Stremio Plus (PPSA74126) | console. It needs getsid (#3493) to load; with that it reaches this check |
The four titles that log to their own file look silent from the console: only Transferring control to ELF entry point and exit 0xc0000409. That's likely why this stop hasn't been reported, so whoever runs the catalog may want to check download0/**/*.log too.
The check is ValidateProvokingVertex (Graphics/src/Pipeline.cpp:104, from 3a8cf3a): the pipeline uses the last provoking vertex (PA_SU_SC_MODE_CNTL bit), and a fragment shader reads a flat parameter through the explicit per-vertex path. Under VK_PROVOKING_VERTEX_MODE_LAST_VERTEX_EXT, per-vertex index 0 is no longer the provoking vertex. A flat attribute read through the per-vertex path would need the index of the provoking vertex instead (the last vertex of the primitive: 2 for triangles, 1 for lines), so the program would depend on the pipeline's provoking mode and topology. I haven't looked at whether these titles' shaders read the flat parameters through P0/v_interp_mov or another way.
Found with Claude Code.
Activity
orishamir commentedon Oct 11, 2026
This is the open question of #1413, and there is a fix for the case these titles use that isn't upstream yet: AlxcBTW#12. With it, ProsperoStore (PPSA99000), one more title that stops here, gets past this check and renders its UI.
Why these titles reach it. They draw through ps5-opengl (ProsperoStore pins its SDK in
tools/fetch-opengl-sdk.sh). In ps5-openglmain00ea837b:toolchain/opengnm-psbc-ps5.patchrunslower_flat_input_vertexon every PS5 fragment shader (:3576). Under GL's default convention it turns each flat input load intoload_input_vertex(1)(:938-960), which ACO emits asv_interp_mov_f32 ..., p10(see v_interp_mov P10/P20 of a FLAT_SHADE input without pass-through: deltas or raw vertices? #1413).src/gallium/ps5/ps5_screen.c:2244setsPA_SU_SC_MODE_CNTL.PROVOKING_VTX_LASTwhenflatshade_firstis clear, which is also GL's default.AnyPS5 marks a flat input read through P10 or P20 as per-vertex (
ShaderInfoCollector.cpp,mode < 2u) and maps bit 19 to Vulkan's last-vertex mode (#2322), soValidateProvokingVertexrefuses the first pipeline whose fragment shader has aflatvarying.The open question is what P10 of a FLAT_SHADE input without pass-through returns. CouchRiv measured 0 on gfx1036; ps5-opengl, which reads every flat input this way, passes the GL 4.6 CTS on a PS5 (both on #1413). AlxcBTW#12 (
c18298c3) takes the PS5 side for exactly this case: P10 of such an input under the last provoking vertex returns the provoking vertex's value. P20, and P10 under the first provoking vertex, still throw.Run with it. #12 cherry-picks onto
maina11c8f7 with one conflict, inTechnicalDebt.md. ProsperoStore 1.000.050 (catalog build), Windows 11, RTX 4080, Release, with #3512 for its earlierv_fma_mixstop:presentation failed: AGC graphics: last provoking vertex with explicit per-vertex flat interpolation is unsupported;APS5_DUMP_FRAMESshow the whole UI: the splash (logo tiles, title, progress bar, "Loading the catalog"), then the Discover page (tab bar, "The catalog is on its way" since it is offline, button hints).On #1413 AlxcBTW reported the same for Homebrew UI Lab and VITA5 with #12, and the ps5-opengl showcase (PPSA99005) gets past it too. ProsperoTV (PPSA99003) reaches this line once its
pollstop (#3521) is fixed.@AlxcBTW, #12 is still only on your fork. Would you open it here, as you did #3546? It is the change maintainers would need to accept or reject on the PS5 evidence above.
AI-assisted: yes (Claude Code Opus 5.5 xhigh)
orishamir commentedon Oct 11, 2026
The other titles of this issue, with and without AlxcBTW#12. Both runs use the same tree:
maina11c8f7 + #3512 +getsidfrom #3493 (for Stremio Plus) + screamm'sfeat/libkernel-poll-filesfor #3521 (for ProsperoTV). The only difference is #12'score/diff. Catalog builds, Windows 11, RTX 4080, Release, 45 s each. The screens come from frames saved withAPS5_DUMP_FRAMES.lstat /datagives ENOENT here)sceAudioOutSysGetHdmiMonitorInfo not implementedafter 16 s, before its first frameSo all six titles, and ProsperoStore above, use the case #12 covers. None hits the throws #12 keeps (P20, or P10 under the first provoking vertex).
@babyhuey, Surround Sound Studio's next stop is the throwing export from #2887. #2887 says the title doesn't call it during start-up, but once this check passes it does.
AI-assisted: yes (Claude Code Opus 5.5 xhigh)
AlxcBTW commentedon Oct 11, 2026
@orishamir thanks, that's a clear table: all seven titles hit the one case #12 covers, and none reaches the throws it keeps.
I'll open AlxcBTW#12 as an upstream PR, rebased onto current
main, as soon as #3546 is merged. Until then I can only have one PR open. If someone wants to carry it sooner, c18298c is ready.The open question stays the one on #1413: CouchRiv measured 0 on gfx1036. ps5-opengl reading every flat input through P10 and passing the GL 4.6 CTS on a PS5 is the only PS5-side evidence. #12 follows the PS5 side for exactly that case and keeps the other cases as throws.
AI-assisted: yes (Claude Code)
Lorkydey commentedon Oct 11, 2026
PShare (PPSA99773, release 1.0) is another title that gets past this check with AlxcBTW#12. Thanks AlxcBTW for the patch.
Tested on Windows 11, Ryzen 5 3600, GTX 1660 SUPER (driver 555.85), WinLibs GCC 15.2, Release. Both sides use
maina11c8f7 + #3606 (187fd85, needed forgetlogin_r) + #3855 (36d37a0, fixes the earlier UTC-to-local stack overwrite). Same relinked executable and runtime libraries; onlylibSceAgcDriver.prxchanges for the comparison. Each run uses a fresh shader-cache directory.[videoout] presentation failed: AGC graphics: last provoking vertex with explicit per-vertex flat interpolation is unsupported.c18298c3: gets past that guard and renders. GPU frame dumps show the splash/logo animation, then the Captures screen with its sidebar, filter tabs, text and icons. The screen says "PShare can't see your captures"; no captures were supplied to this test installation.0xcc) case covered by fix(shader): round v_fma_mixlo_f16 and v_fma_mixhi_f16 in every f16 rounding mode #3512; I have not applied that PR in these runs.The first comparison used
APS5_DUMP_FRAMES=8; a second patched run usedAPS5_DUMP_FRAMES=8 APS5_DUMP_FRAMES_EVERY=30 APS5_DUMP_SCALE=1 APS5_DUMP_SHADERS=1, which captured the splash and main UI at 1920x1080 and saved the next failing shader request. Logs are in the title'sdownload0/pshare/dev/app.log.Validation: the targeted build succeeds, and
agc_driver_graphics(including the patch's P10 case),agc_shader_disk_cache, and the existingagc_driver_provoking_vertexpixel tests pass 3/3, withANYPS5_REQUIRE_VULKAN=1. The latter pixel test covers GLSL/guest P0, not a new P10 pixel oracle. The captured next shader also reproduces the FLOAT_MODE error withagc_shader_replaywithout running the title. The full test suite was not run.This confirms progress through PShare's rendering path on this Windows GPU, not the unresolved PS5 hardware semantics in #1413. I have not validated interaction/uploading or compared pixels against a console capture.
AI-assisted: yes (Codex).
AlxcBTW commentedon Oct 11, 2026
@Lorkydey thanks, that makes PShare the eighth title through this check with AlxcBTW#12. It goes into the PR's title list, and the PR opens as soon as #3546 is merged.
AI-assisted: yes (Claude Code)
AlxcBTW commentedon Oct 11, 2026
AlxcBTW#12 is now upstream as #4030, rebased onto
maina11c8f7. The titles list covers @orishamir's table, @Lorkydey's PShare and @AchrafReyani's merge map. It still covers only the P10 case; P20 stays a throw (see #1413).AI-assisted: yes (Claude Code)
feat(libSceAudioOut): implement sceAudioOutSysGetHdmiMonitorInfo
orishamir commentedon Oct 12, 2026
With #4030 merged, the titles of this issue no longer reach this line on
main. Windows 11, RTX 4080 (driver 616.56), WinLibs GCC 15.2.0 r7, Release, plainmain2a5eca1 with no other PRs and no extra relinker options, catalog builds, 60 s each withAPS5_DUMP_FRAMES=4(one frame every 300 presents). Before #4030, the eight titles other than ProsperoTV stopped at this line on Windowsmain0ef9dc9 earlier today, six of them also on 3d8ffeb (#4030 (comment)).main2a5eca1GetHdmiMonitorInfo rc=0x00000000in its logdownload0/ps5-opengl.loghas 988[ps5-gallium] draw-rejected status=-12 mode=4lines with counts around 1.0e9 and 3.2e9; not looked intopoll: descriptor 4 is not a socketafter 10 s (#3521); babyhuey's run with #3649 onmaind592d54: #3649 (comment)None of the logs contains the line of this issue. Each run logs once
shader library prebuild skipped: AGC graphics: v_interp_mov p10 of a flat input with the first provoking vertex is not measured(the case #4030 keeps as a throw, caught by the prebuild); the titles' draws then go through.AI-assisted: yes (Claude Code Opus 5.5 xhigh)
Move #1413 out of the chains of nine titles on main 2a5eca11 (#45)
Upstream sync: small HLE libs (chunk 1) (#437)
Upstream sync 6: small HLE libs, remainder (slice E2) (#441)
oneandonlydean commentedon Oct 13, 2026
Closing: fixed by #4030. The nine titles listed here no longer reach
last provoking vertex with explicit per-vertex flat interpolation is unsupportedon main (orishamir's run above). The P20 case stays in #1413.