007 First Light (PPSA11386)

[SleglJan]

Existing thread

• I searched this category for the title ID and no thread exists for this title on any platform.

Title

007 First Light (PPSA11386)

Content version tested

01.018.000

AnyPS5 commit

main 34355ee (2026-10-14) plus my integration branch share/007-v5 at 65c5b68 (159 commits):

• open PRs: feat(agc): implement image and sampler tables indexed at runtime offsets #757 and feat(agc): implement image tables loaded through 64-bit pointers #758 (older heads), fix(libSceAgcDriver): retire a depth surface once another writer covers its memory #454 (ported onto main's depth code), fix(agc): advance stored layers when a write-back keeps CPU-written blocks #4663, fix(agc): store small pending storage results at each flip #4841, perf(agc): bind a sparse descriptor heap's committed run in place from its import #5249 (older head), a port of draft perf(agc): pipeline queue-0 draw commits behind the packet worker #1201 (opt-in) and the GPU-side part of draft perf(agc): pipeline indirect draws behind the packet worker (APS5_PIPELINE_INDIRECT) #1896 (off here);
• from closed PRs: fix(agc): preserve storage image dependencies during ahead dispatch capture #4724 (closed by a maintainer) and one commit of feat(shader): support bindless image table stores #2500 (closed by the intake pause);
• fix(shader): apply cube output modifiers whenever IEEE mode is off SleglJan/AnyPS5#21 (my fork);
• two commits from Scramblrrrr's fork without a PR (snapshot clearing, a joined-import test);
• the rest are local commits without a PR: fixes, tests, diagnostics, performance changes and input helpers.

System tested

CachyOS (Linux 7.2.9), GCC 16.2.1, Release with SPIRV-Tools off; RTX 3080 10 GB (NVIDIA 615.71.09); Ryzen 9 7950X3D, 64 GB RAM.

Furthest point

In game

What is shown

The "(L) Move" tutorial, then Bond in third person on the beach rocks ("R Look"), at 0.2–0.35 FPS (per-run medians 0.22–0.30).

• Keyboard and a DualSense both control Bond. Audio plays.
• Picture: white blowouts and colour washes (white, green, magenta) in many frames; the clearer ones show a purple fog and stray coloured specks and streaks. In the first gameplay run 75 of 142 screenshots were washed or blown out. Causes not identified.
• The first run from boot to gameplay (2026-10-14, about 1 h of play, no crash) was on the previous rebase of this stack (main 3e0e01f). On the branch above, at least 5 runs from the checkpoint saved there reached the tutorial again.

Current blocker

On main c13dc7e (2026-10-13), 007 stops about 2 s after start, before any draw, with VOP3 source modifiers are not implemented: row 8 of the table in #2523. Newer main was not run alone.

Still needed beyond main, in the order the title reached them (order from earlier bases, not bisected on 34355ee):

1. Row 8, cube instructions with an output modifier: fix(shader): apply cube output modifiers whenever IEEE mode is off SleglJan/AnyPS5#21. Upstream fix(shader): accept cube output modifiers #2463 was closed by the intake pause; no upstream PR is open.
2. Row 19 at the first flip, bindless image table has 18960 entries (0 materials), limit 16: feat(agc): implement image and sampler tables indexed at runtime offsets #757 and feat(agc): implement image tables loaded through 64-bit pointers #758, plus local follow-ups that narrow each table to the keys its shader reads.
3. Local, no PR: a pixel shader stops at its first draw with shader sampler pairs exceed runtime heap capacity (21 samplers instead of 7); the fix folds again after masked-select elimination in every program, not only with a live packed ancillary input.
4. Local, no PR: vertex programs with all eight guest clip distances fail (guest clip distances plus the all-zero clip position distance exceed 8 clip distances); the fix adds the all-zero distance only when a slot is free.
5. Level load, local: registered program refers to an incompatible shader binary type. 007 registers its hull programs as one merged LS+HS binary (type 3); the stack runs the whole program in the hull stage, one invocation per output control point.
6. Level load, local: indexed indirect draws whose declared index range reaches unbacked pages; the unbacked part is bound as zeros.
7. First cinematic, local:
   • image-table stores (the feat(shader): support bindless image table stores #2500 commit and a store path for tables selected at run time); without them frames come out pink and white;
   • workgroup count … exceeds device limits on an indirect dispatch: storing pending storage-image results waits for newer recorded writers of the same memory;
   • out of device memory on 10 GB: texture allocations release unreferenced cache entries and retry, device buffers fall back to host memory.

Also carried, not shown to be needed: #454, #4663, #4841, #4724. Later stops now pass with merged PRs: rows 9, 12, 18 (#2471, #3379, #3500), depth storage views (#425, #4500, #2904), #423, #3523, #4364, #1491, #4031.

Open problems:

• Intermittent guest SIGSEGV in the title's terrain code during the level load after Resume: 5 of 12 runs from the in-game save on this branch, 3 of 15 on earlier stacks (007 First Light (PPSA11386) on Linux: start-up stops in order and the PRs that pass them #2523). Cause not identified; the branch's guest-fault dump now prints the write tracker's view of the faulting memory, and the first fault with it is being read.
• Performance: queue 0's packet worker is about 99% busy in gameplay. The largest measured item is the flush wait (pending storage-image results stored behind recorded writers): 0.6–1.1 s per 10 s, 6–11% of the worker.

How it was run

1. relinker --registry (no --to-intel); param.json, icon0.png and playgo-chunk.dat from the package in sce_sys.
2. bubblewrap sandbox: no network, tmpfs home (NVIDIA's own shader cache is dropped), X11 through XWayland.
3. Warm AnyPS5 shader and Vulkan pipeline caches from the previous run; a cold pipeline cache adds about 11 minutes of compute pipeline creation at boot.
4. A diagnostics configuration (traces, profiling). Switches that matter: APS5_HOST_IMPORT_MIB=12288, APS5_RESOURCE_CACHE_ENTRIES=64 (not re-checked since fix(agc): evict resource-cache images with storage entries #3474), APS5_NO_IDLE_STORE=1 (stack switch, turns off fix(agc): store pending storage-image results when queues idle #3777's idle store), and for speed APS5_NO_DRAW_SRT_ENTRIES=1, APS5_PIPELINED_DRAWS=8, APS5_AHEAD_PIPELINES=2, APS5_LOCAL_LARGE_CLASSES=1.
5. Title, Cross, "Disconnected" dialog (continue offline), main menu, Square (Resume Game) from a save of an earlier run; the runs on this branch resume from the checkpoint at the tutorial.
6. The underwater cutscene's "L" prompt with a dial needs the left stick rotated: up, right, down, left (W, D, S, A, or 4 s steps through APS5_PAD_INPUT_FILE). Holding one direction does not pass it.

AI-assisted: yes (Claude Code)

[SleglJan]

Update on the open problem "intermittent guest SIGSEGV in the title's terrain code after Resume": cause and evidence are in #6172. The title reads a GPU feedback buffer after a bounded wait (2 s, or skipped once the job is two frames old); our level entry stops presenting for 26-360 s, so the buffer is read before the GPU wrote it. A local title patch that widens both limits removed the crash in 8 of 8 runs.

AI-assisted: yes (Claude Code)