Skip to content

README.md

Latest commit

 

History

History
103 lines (76 loc) · 6.19 KB

README.md

File metadata and controls

103 lines (76 loc) · 6.19 KB
Raw

Runtime tests

These tests drive parts of the runtime directly on the host, with no game and no SDL window, and check them against the firmware's behaviour. Run each one from this folder. They all share the runtime's include path:

INC="-I ../core -I ../gfx -I ../audio -I ../media -I ../input -I ../net -I ../ui -I ../platform -I ../debug -I ../../recompiler -I ../.."

The kernel: kernel_test.c

Checks runtime/core/kernel.c, the thread manager, against the firmware's semantics and error codes. It covers:

  • Synchronisation: threads, semaphores, mutexes, lightweight mutexes and conditions, kernel conditions, event flags, simple events, SceUlt objects, reader-writer locks and message pipes.
  • Waking rules:
    • several waiters on one event flag, each with its own mask and clear mode;
    • a condition wait releasing and re-taking its mutex with the same recursion count;
    • Signal, SignalAll and SignalTo;
    • semaphore queue order and overflow;
    • timeouts returning SCE_KERNEL_ERROR_WAIT_TIMEOUT with the remaining time written back;
    • deletion waking waiters with WAIT_DELETE.
  • Threads: info, priority, exit status and thread event handlers.
  • Reader-writer locks: recursive reads, writer preference, and the order an unlock grants in.
  • Message pipes: ASAP and FULL transfers, sender order, and WAIT_CANCEL from Cancel.
  • Callbacks: one notified while its owner sleeps in WaitSemaCB, DelayThreadCB or WaitThreadEndCB runs on the owner mid-wait with its arguments, and the wait carries on. Also CheckCallback, and a callback deleting itself by returning non-zero.
  • Fibers: running, returning to and being resumed by their thread with arguments passed both ways, switching between fibers, and a fiber without a context restarting each run.

A final section switches on the Vita scheduler (RECOMP_SCHED=vita) and runs real guest threads:

  • on one core, a higher-priority thread preempts its starter at the starter's next system call, and the starter resumes only once the new thread blocks;
  • three equal-priority spinners all make progress by round robin, never two at once;
  • a blocked high-priority thread lets a low-priority one run, and a signal from it hands the core back at its next system call;
  • ChangeThreadPriority reorders waiters and preempts at once;
  • nothing deadlocks when every thread sleeps;
  • the timer evicts a spinner that makes no system calls;
  • on two cores, a new thread preempts the lowest-priority holder, and affinity masks pin threads to cores.

A hang in that section fails after 30 seconds with every thread's scheduler state. The test includes kernel.c and supplies the pieces it needs from main.c and display.c: a 4 GiB guest memory block, the allocator and the function lookup.

cc -O1 -g -Wall -Wno-unused-function $INC kernel_test.c -o /tmp/kernel_test -lpthread && /tmp/kernel_test

It prints scheduler (RECOMP_SCHED=vita): N checks, 0 failed, then N checks, 0 failed, and exits 0. A failed check names its line. Waiter tests use short sleeps to order threads, so a run takes about three seconds.

Touch gestures: gesture_test.c

Checks runtime/input/gesture.c, the Vita's gesture library (SceSystemGesture): its touch tracker and its tap, drag, hold, pinch and rotation recognisers.

  1. Scripted taps, as a headless run makes them: a front-panel touch held over a range of frames, fed once a frame and also as 64 identical copies a frame, the way Killzone reads them. Each tap must give exactly one TAP, and each touch one DRAG BEGIN and one END. If two touches share an ID, the second drag must not begin; that's the firmware's rule, and why display.c gives each contact its own ID.
  2. The console's own scenario (gesture_scenario.h) runs through our handlers, and its trace is compared record by record with one the firmware produced on a real Vita. That firmware.trace stays on your machine, since no firmware output belongs in the repository. Without it, only the scripted taps run.
cc -O1 -g -Wall -Wno-unused-function $INC gesture_test.c -o /tmp/gesture_test -lm -lpthread \
  && /tmp/gesture_test [path/to/firmware.trace]

It prints scripted taps: N checks, 0 failed and console trace: 684 records, 0 differ.

JPEG: jpeg_test.c

Checks runtime/media/jpeg.c (SceJpeg and SceJpegEnc) the way Uncharted's photo minigame uses it: a 960x544 picture encoded, then decoded twice, once as YCbCr at quarter size (the 240x136 thumbnail) and once as RGBA, checking the pitches and the colours that come back. There's no console reference yet; the layout follows Vita3K's MJPEG module.

cc -O1 -g -Wall -Wno-unused-function $INC jpeg_test.c -o /tmp/jpeg_test && /tmp/jpeg_test

It prints all passed.

The camera: camera_test.c

Checks runtime/media/camera.c (SceCamera) the way Uncharted's hold-it-to-the-light parchment puzzles use it: the rear camera opened with YUV420 planes at 640x480, and frames read with the lamp off and on (F7), checking the planes come out dark and then bright, along with the frame index and status. Then buffers given per read, and a 32-bit ABGR format. The behaviour follows Vita3K's SceCamera module; there's no console reference.

cc -O1 -g -Wall -Wno-unused-function $INC $(pkg-config --cflags --libs sdl3) camera_test.c -o /tmp/camera_test && /tmp/camera_test

It prints all passed.

Metal's scratch folders: mtl_archive_test.m (macOS)

Checks the sweep in runtime/gfx/mtl_archive.m. Metal gives every binary archive a pipeline is added to a scratch folder (com.apple.gpuarchiver/gpuarchiver-<hex> in the user's cache folder) and never removes it; the archive keeps one per shader library, so a week of test runs once left 43 GB of them. The sweep removes the ones untouched for 20 minutes. The test checks, in a temporary folder, that only idle gpuarchiver-* folders go. Then, with Metal itself, it checks the two facts the sweep relies on: the folder is where the sweep looks, and a written archive (and its file) still finds its pipelines once the folder is gone. A macOS update that changes either fails here.

clang -O1 -g -Wall -I ../gfx mtl_archive_test.m -framework Metal -framework Foundation -o /tmp/mtl_archive_test && /tmp/mtl_archive_test

It prints all passed.