vita-profiling

name	vita-profiling
description	Profile geometrizer on Vita3K — perf/DRC stat lines (opt-in triggers), gprof, scheduler tuning. Covers build configurations, deployment, and interpreting results.

Vita3K Profiling

Build Configurations

The Vita build profile is selected by the GEO_VITA_PROFILE CMake option, not CMAKE_BUILD_TYPE:

GEO_VITA_PROFILE	Flags	Use case
OFF (default)	-O3 -flto, -DGEO_VITA_RELEASE (logs/perf printf compiled out)	Shipping build; fastest, silent
ON	-g -pg -DGEO_PROFILE + vita-gprof + NOASLR, no LTO	gprof + runtime perf overlay

Release build (default — normal testing / shipping):

cmake -B build-vita -DVITA=ON \
  -DCMAKE_TOOLCHAIN_FILE=$VITASDK/share/vita.toolchain.cmake \
  && cmake --build build-vita -j

Profile build (gprof + perf overlay on Vita):

cmake -B build-vita-prof -DVITA=ON -DGEO_VITA_PROFILE=ON \
  -DCMAKE_TOOLCHAIN_FILE=$VITASDK/share/vita.toolchain.cmake \
  && cmake --build build-vita-prof -j

The default Vita build is now a true release: -O3 -flto, -pg/gprof stripped out, and all geo_log/perf-window sceClibPrintf output compiled out via -DGEO_VITA_RELEASE. You MUST pass -DGEO_VITA_PROFILE=ON to get gmon.out, the SELECT-button gprof hooks, or the on-TTY perf lines. LTO is force-disabled in the Profile build because it eliminates/reorders the mcount calls -pg injects. The release build uses GCC's plugin-aware gcc-ar/gcc-ranlib so LTO objects archive correctly. ARM libco backend (arm.c) is faster than sce (SceFiber) — add -DVITA_LIBCO_BACKEND=arm unless debugging fiber issues.

Deployment to Vita3K

Vita3K filesystem on macOS:

~/Library/Application Support/Vita3K/Vita3K/fs/
  ux0/app/GEOM00001/          ← deployed VPK contents
  ux0/data/geometrizer/       ← runtime data (writable)

Deploy script (one-liner):

FS="$HOME/Library/Application Support/Vita3K/Vita3K/fs"
APP="$FS/ux0/app/GEOM00001"
rm -rf "$APP" && mkdir -p "$APP"
unzip -o build-prof-armco/geometrizer.vpk -d "$APP" >/dev/null

Trigger Files (ux0:data/geometrizer/)

All runtime knobs are files in the data directory — no need to rebuild:

File	Purpose	Content
gpu_tilemaps	Enable GPU tilemap backend (empty file)	—
sched_max_slice	Override scheduler timeslice	Integer, e.g. 1000
drc_stats	Opt-in since 2026-07-10 (b8c1e4d): DRC/TGPDRC line every 30 frames + on-exit dumps	≠ 0
perf_fibers	Opt-in since 2026-07-10: per-fiber PERF line (clock read per co_switch — real overhead)	≠ 0
perf_log	Force the fps/cpu/video perf line on the release eboot	—
mix_log / audio_dump	Audio pacing counters / rendered-audio dump — see vita-hw-deploy §4b	—
nvram/vr.bin	NVRAM save (auto-created)	Binary

Create a trigger:

FS="$HOME/Library/Application Support/Vita3K/Vita3K/fs"
touch "$FS/ux0/data/geometrizer/drc_stats"
printf "1000\n" > "$FS/ux0/data/geometrizer/sched_max_slice"

All this instrumentation used to be ON by default and is now opt-in — a release eboot with no trigger files emits nothing and pays no counter cost.

Perf lines

Location: the perf/DRC lines go to sceClibPrintf (stderr), NOT to a file: Vita3K captures them in its own vita3k.log; on real HW you need catlog (see vita-hw-deploy §4). The old ux0:data/geometrizer/perf.log file sink was removed (per-line open/write/close was pure overhead).

Format (one line per 30-frame window):

fps=48.4  cpu=19.07ms  video=1.43ms

• fps — measured over the 30-frame window (wall clock)
• cpu — average time for model1_run_frame + model1_prepare_render + model1_end_frame (V60, TGP, M68K, scheduler)
• video — average time for model1_video_update + model1_video_present (software 3D, tilemaps, upload)

DRC stats (when the drc_stats trigger is present):

DRC new_blocks=0 total_exec=374760 jit_blocks=374760 ir_blocks=0 jit_cov=100.0% jit_cycle_cov=100.0%

• new_blocks — new V60 IR blocks translated this window (0 = fully cached)
• total_exec — total V60 blocks executed (JIT + IR)
• jit_blocks / ir_blocks — blocks run via JIT vs IR interpreter
• jit_cov — percentage of blocks compiled to JIT
• jit_cycle_cov — percentage of cycles executed via JIT

Interpreting the perf lines

Pattern	Diagnosis
cpu high, video low	CPU bottleneck (V60, TGP, M68K, scheduler)
cpu low, video high	Video bottleneck (software 3D, tilemaps, upload)
new_blocks > 0 for many windows	DRC cache thrashing (new code paths)
jit_cov < 100%	JIT rejections (check [V60DRC] stderr log on exit)
ir_blocks > 0	Blocks falling back to IR interpreter

Typical VR results (Vita3K, ARM libco, GPU tilemaps)

Scene	FPS	CPU	Video
Stable (menu/idle)	~37	~25ms	~1.6ms
Post-menu (heavy 3D)	~16-21	~42-55ms	~5.6-7.3ms

Instrumentation overhead is real: with the (old default-on) stats the stable FPS read ~37 vs ~48 without. They are opt-in now — only enable what you need.

gprof (Profile builds only)

Requirements: configure with -DGEO_VITA_PROFILE=ON (adds -pg -DGEO_PROFILE + vita-gprof + NOASLR)

Output: ux0:data/geometrizer/gmon.out (written on clean exit)

Note: gmon.out is ARM (PS Vita) — you CANNOT run gprof with the macOS binary. The file is collected for future analysis with an ARM gprof tool or for comparison.

# After run, gmon.out is at:
FS="$HOME/Library/Application Support/Vita3K/Vita3K/fs"
ls -la "$FS/ux0/data/geometrizer/gmon.out"

Vita3K Launch (macOS)

# Clean start
pkill -f Vita3K 2>/dev/null
perl -e "select(undef,undef,undef,1)"
open -a /Applications/Vita3K.app --args -l 1 -r GEOM00001 >/dev/null 2>&1

# Run for N seconds then kill
perl -e "select(undef,undef,undef,$N)"
pkill -f Vita3K 2>/dev/null

# Read the perf/DRC lines (Vita3K captures stderr in its own log)
grep -E "PERF|DRC " "$HOME/Library/Application Support/Vita3K/Vita3K/vita3k.log" | tail

Full Iteration Script

FS="$HOME/Library/Application Support/Vita3K/Vita3K/fs"
APP="$FS/ux0/app/GEOM00001"
mkdir -p "$FS/ux0/data/geometrizer"
printf "1000\n" > "$FS/ux0/data/geometrizer/sched_max_slice"
touch "$FS/ux0/data/geometrizer/gpu_tilemaps"
touch "$FS/ux0/data/geometrizer/drc_stats"   # optional (opt-in since b8c1e4d)
touch "$FS/ux0/data/geometrizer/perf_fibers" # optional PERF line
rm -rf "$APP" && mkdir -p "$APP"
unzip -o build-prof-armco/geometrizer.vpk -d "$APP" >/dev/null
pkill -f Vita3K 2>/dev/null
perl -e "select(undef,undef,undef,1)"
open -a /Applications/Vita3K.app --args -l 1 -r GEOM00001 >/dev/null 2>&1
perl -e "select(undef,undef,undef,35)"
pkill -f Vita3K 2>/dev/null
perl -e "select(undef,undef,undef,1)"
grep -E "PERF|DRC " "$HOME/Library/Application Support/Vita3K/Vita3K/vita3k.log" | tail

Common Gotchas

• stderr on Vita3K goes to ~/Library/Application Support/Vita3K/Vita3K/vita3k.log — NOT to a separate stderr.log. Search with grep "V60DRC" vita3k.log.
• Instrumentation is opt-in (2026-07-10, b8c1e4d): no trigger files → no PERF/DRC lines and no per-co_switch clock reads. The old per-line perf.log file sink is gone; lines go to sceClibPrintf only.
• gmon.out is ARM — cannot be analyzed with macOS gprof. Use an ARM gprof or wait for Vita3K to expose profiling data.
• Vita3K window close vs clean exit — killing the window (⌘Q) may lose gmon.out (and anything stdio-buffered the app had open).