PS5_RetroArch/docs/ACTIVE.md

Active work

Volatile by design. Rewrite this file in place and keep it under about 120 lines. Specifications belong in docs/PLAN.md, and run results belong in docs/PHASE_LOG.md. This is the only file that carries a date: the line below is the last thing read, so changing it costs nothing.

Updated: 2026-09-19

Now

Frames run. An unattended tools/run-title.sh --no-build --watch 25 records 644 frames and 643 menu draws and is still alive when the script closes it: "VERDICT: it ran for 25s and this script closed it" (klog/run-PPSA99169-022103.log, klog/trace-r3o.txt). The probe-free shipping image behaves the same way on its own run (klog/run-shipping-r3.log, klog/run-PPSA99169-023553.log: "it ran for 20s and this script closed it", no fatal signal, grep -c 'fatal signal' = 0). video_vulkan is selected, the device and the 3840x2160 khr_display swapchain are created, the stock shaders compile through the driver's own compiler, and vulkan_frame -> chain draw -> vkQueueSubmit repeats for the whole run.

The driver's own messages were what made this ordinary. The messenger patch (0019) put every refusal in the trace, and each one named its reason:

Refusal	What it was	Fix
set 0 binding 1: a combined image sampler write names no sampler	all four display samplers were refused at creation (borderColor opaque white; the driver creates only transparent black), so the menu quad wrote a NULL sampler	0024 (the colour the driver names), 0025 (a refused sampler falls back to the nearest one)
set 0 binding 2 is not bound or holds no write	the display layout's second sampler binding: the driver requires a write for every binding its stage metadata names	0026 (write the same image at binding 2, as the HDR path already does)
samples a view whose component mapping is not the identity	the menu texture used B4G4R4A4 with a B/R view swizzle	0027 (32-bit, swizzle-free path), 0028 (the CPU conversion's channels in R8G8B8A8 order)

What is still refused, all at init and none of them fatal: seven triangle-strip clears and one storage-image compute upload of the frontend's blank texture, plus the filter chain's non-clamp sampler modes (the chain's sampler table is repaired by 0023). They cost the frontend work, not the frame.

Nothing proves pixels yet, and the capture path is now half-built. The port has no screenshot: --max-frames/--max-frames-ss were compiled out (#ifdef HAVE_SCREENSHOTS), and tools/build-retroarch.sh now defines it, so the frontend builds and links with the options present. They still do not fire: runloop.h's RUNLOOP_TIME_TO_EXIT compares max_frames against the core's frame_count, and this title loads no content, so the counter never advances - three runs with --max-frames=200 ran the full watch window and wrote no shot.png. src/main.cpp reads /app0/args.txt (one argument per line) so a run can be given those options without the launcher, verified by the mark argv extras from /app0/args.txt = 3.

Next

1. Run the capture - it is built and waiting. The port now reads the frame with the driver's own vulkan_read_viewport and writes a PPM itself (patches 0031/0032, committed as 949fdb9; the frontend builds 276 of 276 and the path is in the image). It has not been run: the console stopped answering ("No route to host") before the run, and the run that was attempted failed at its deploy step for that reason. Arm /app0/args.txt with --ps5-capture=90 and --ps5-capture-path=/app0/shot.ppm, run tools/run-title.sh --no-build --watch 18, fetch /app0/shot.ppm and look at it. The trace line says view viewport WxH, read_viewport -> ok|failed.
2. Then the init refusals: triangle strips at init (the topology patch 0016 applies to the chain's quad) and the blank texture's compute upload (its staging texture could match its destination, as 0027 does for the menu texture).
3. /app0/retroarch.log is still the stale 1200 bytes from an earlier round: the frontend's file logger is set up (--log-file=/app0/retroarch.log) but never flushes, and a run that exits by itself would flush it - which is the same exit the screenshot step needs.
4. Config loading is still parked (parked/config-path.patch.py, step 0006): the command line passes -c /app0/retroarch.cfg and the port does not yet prove the file is read.
5. Then the first frame, and with it the objective: frames in the trace, the menu on screen, and tools/verify.sh green.
6. Then the config file. Content loading still discards the title's -c; the fix (0006) is parked in parked/config-path.patch.py.

Working notes

• bash tools/run-title.sh is the only way to run the title. It owns the whole sequence and prints the title's own trace. Do not hand-deploy.
• python3 tools/symbolize-crash.py <capture> reads a console backtrace, but only approximately: the addresses are the converted image's and build/title.map is the linked one, so a frame can land one function off. The trace's assertion failed: line is exact - prefer it.
• Assertions only speak because stderr is unbuffered. If that line vanishes from the trace, check src/main.cpp's setvbuf before anything else.
• A probe is not gone until the objects are. Reverting the probe set leaves the strings in build/ra/obj; strings build/llvm-pie.elf | grep 'probe ' is the check and rm -rf build/ra/obj is the fix. Two stale probe strings survived a revert this round and needed a clean rebuild.
• An insert must not repeat its anchor. The probe tool writes insert + anchor, so an insert ending with its anchor duplicates that line - which broke the build three times in one round. tests/test_frontend.py's ProbeSet now checks every probe for that, for a marker in its own text, for a note and for a trailing newline.
• Do not hand-edit the configured tree. Slicing a probe out of build/ra-conf by text removed a declaration line and a brace's partner this round, and the file only went back together by diffing it against vendor/retroarch. Revert probes with the tool; if the tree is damaged, diff it against its vendor copy (it carries no port patches beyond the ones the patcher names).
• vendor/ is not in the diff. Anything a build needs from it must be regenerable or live in a tracked directory.
• /app0/... is not an FTP path. The title's files are at /data/homebrew/PPSA99169/.
• src/ is linked before the archives, so a definition there wins over the same symbol in libretroarch.a or the driver's archives - that is how src/locale_shims.c supplies the _l locale functions, and why a libc symbol cannot be overridden (the native converter refuses to publish an application export).
• The console's address and credentials come from the ignored .env; console captures stay in the ignored klog/ tree.

Last verified

Check	Result
bash tools/run-title.sh (probed build)	Partial: selected="vulkan", khr_display, 4 textures, vkCreateGraphicsPipelines -> 0, vulkan_init live, pad opened, vulkan_frame -> vkQueueSubmit, then assertion failed: cmd_buffer->render_pass == NULL
bash tools/verify.sh (all gates)	PASS (format unit build integration evidence)
python3 -m unittest discover -s tests	PASS: 20 tests, including the probe-set checks and the patch count
strings build/llvm-pie.elf | grep 'probe '	Only the driver's own messages; no probe text (after a clean rebuild)
Config path (0006)	PARKED: reading the config still crashes in command_event; see parked/config-path.patch.py

Open findings

• The first frame leaves a render pass open before the backbuffer pass; the marks narrow it to the offscreen-passes call and the menu upload, and the next measurement splits those two.
• The frontend's own log never reaches disk. --log-file=/app0/retroarch.log is in argv and the file exists, but it is a stale 1200 bytes: RetroArch never closes the logger on this path, so its buffer is lost. The trace is the instrument that works.
• src/locale_shims.c exists because glslang and SPIRV-Cross are written against FreeBSD's xlocale interface and the console SDK exports only the unsuffixed functions. If the SDK ever ships the _l set, that file goes.
• The console's FTP will not replace sce_module/libc.prx; the title runs against the console's copy. A release has to solve this.
• Pad input cannot be verified unattended yet: nothing in the pipeline presses a button.