Raw core overrides re-push after the curated settings, so a stale one silently
beats the UI with nothing on screen to explain it: the settings screen read SPU
Block Size = Safe for hours while config.yml read Mega.
Mega is the one that mattered. It produces very large compilation units, and those
are what fail AArch64 register allocation with "Cannot scavenge register without
an emergency spill slot" -- which is what put SPU threads on the interpreter
fallback at all. With it cleared, no block fails to compile and the fallback never
engages. Every "cannot be compiled" chased in these sessions traces back to it.
Cleared in every scope, because a title can pin a key the global also pins: Arkham
City carried Accurate SPU Reservations true as a raw per-title override against
false globally, so clearing one scope did nothing and the two readings looked
contradictory.
Per-title Accurate SPU Reservations values go too, except Web of Shadows, which is
the title it was measured on. Off is off-spec -- it forces the SPURS scheduler to
HLE and bypasses the reservation lock -- and a title left that way desyncs until
its SPU threads execute whatever they land on, which is how Arkham City ended up
dying with "Unknown STOP code: 0x0".
Adds CoreSettingOverrides.forgetEverywhere for the all-scopes case.
3072 was set to get the God of War 3 demo past an allocation failure, but that
failure was measured before the uninterruptible reclaim fix landed, and the cap is
not coordinated with the texture cache, which budgets itself up to 2560MB on
Android. Raising one without the other let the total grow with it: Batman: Arkham
City reached 5596MB resident against a 6246MB peak on a 7.2GB device and stalled
after a while, with no allocation failure to point at.
2048 is the value that shipped before, and it is where the sum of the two sat when
that game worked. Budgeting the cap and the cache together is the actual fix and
is not attempted here.
The RSX profiler was still recorded as a raw core override from the debugging
work, so config.yml read "RSX Profiler: true" while nothing in the UI said so --
the same divergence as the relaxed-ZCULL one, since overrides re-push at the tail
of applyTo. It writes a bucket report every 300 frames and keeps per-scope timers
on the RSX thread, which is not something to ship enabled. The first purge had
already marked itself done, so this takes a new key.
VRAM allocation limit is applied as VMA's pHeapSizeLimit, which makes it a hard
ceiling rather than an eviction threshold: once total allocations reach it VMA
returns OUT_OF_DEVICE_MEMORY however much the device has free. Lowering it does
not make the cache release earlier, it makes allocation fail earlier. The God of
War 3 demo was measured failing a routine 24MB request at 1024 while the process
held 1.6GB resident and 280MB in that pool, and failing at 2048 one screen later.
3072 leaves the caches room while keeping the bound that stops an unbounded quota
driving the process to 4.3GB and getting it killed.
The allocation failure now names the request size and the cap alongside it, since
"Out of video memory" alone cannot separate a full device from an artificial
ceiling, and those need opposite fixes.
Refusing outright skipped the allocator's own recovery. That path is "if
OUT_OF_DEVICE_MEMORY and vmm_handle_memory_pressure(...) succeeds, retry the
allocation", so returning false meant the retry never ran and the allocation died
having freed nothing: God of War 3 reached it with zero reclaim attempts and zero
recoveries logged.
Only the fatal branch needs the queue idle, which is what the flush inside it is
for. The rest is reachable while uninterruptible: the texture cache purges its
unreleased pool, and at severe it also drops unlocked sections. RPCS3 already runs
exactly that with no flush whenever pressure is non-fatal, so this is the existing
contract rather than a new risk. Severity is clamped below fatal so the
flush-dependent path stays unreachable.
Measured after: eviction runs and reports releasing resources, and the allocator
retries. God of War 3 still fails, but now for the honest reason -- the device is
out of memory, with 123MB free of 7.2GB and the emulator resident at 4.3GB -- and
not because nothing was ever given the chance to run.
on_vram_exhausted asserted that the renderer was interruptible. Eviction really
cannot run in that state, since it would touch resources the driver may still be
reading, but that is a reason to refuse rather than to kill the thread -- and
refusing is already the supported answer: the OOM path in VKDraw treats false as
using placeholder textures, which it notes can cause graphics glitches but
should not crash otherwise.
God of War 3 hit it by skipping the intro screens, which pushes a burst of surface
and texture allocation through a point where the renderer is uninterruptible. The
RSX thread died there, audio kept playing, and it presented as a hang. With the
refusal in place the same run reports the real problem instead:
VK_ERROR_OUT_OF_DEVICE_MEMORY from the allocator.
Which it genuinely is. VRAM allocation limit was also lowered from 2048 to 1024:
the first value was still above what the device could give us -- 5355MB resident,
99MB free of 7.2GB -- so the budget was never reached before the system ran dry,
which defeats its only purpose. It has to sit below what allocation can actually
satisfy, so eviction starts while there is still room to allocate.
Three things, all found by measurement after the interpreter fallback started
being used in anger.
Marking only the entry point made the interpreter release the thread after one
instruction, whereupon the recompiler tried the next address, failed the same way
and marked that too. 111 consecutive entries were recorded walking two blocks four
bytes at a time, each step paying a full failed LLVM compile. The failed set now
holds ranges, so a thread stays interpreted for the whole block and leaves when
execution genuinely moves past it: 111 markings became 1.
The range test then ran per interpreted instruction and took a reader lock each
time, which put shared_mutex::imp_lock_shared at 28% of the whole process against
23% for the interpreter itself. The extent is now cached on the thread when the
fallback engages, so the loop compares two integers.
The switch was also logged once per thread, but the flag is cleared on every exit,
so the guard fired on every re-entry: God of War 3 wrote thousands of lines a
second ping-ponging between two addresses. Removed; the block is still recorded
once when it is marked.
Separately, VRAM allocation limit was left at upstream's 65536 MB, which means no
limit and assumes a discrete card. Here the GPU shares system memory with the OS
and our own host allocations, so the texture cache is never asked to evict and
grows until allocation fails -- and failing is fatal: God of War 3 dies in
on_vram_exhausted on ensure(!vk::is_uninterruptible() && ...), because VRAM ran
out where the renderer cannot safely evict. Measured at the crash: 5355MB
resident, 99MB free of 7.2GB. 2048 leaves room for the guest's own memory, the
host caches and the OS.
The fallback flag was set once and never cleared, so a thread that met a single
block it could not compile interpreted everything it ran from then on. Correct,
but these are SPURS kernels doing real work, and Sonic Unleashed reached its
loading screen that way and then crawled through it.
The failed set holds entry points, so this keeps interpreting while pc sits on the
bad entry -- which is where a branch-to-self idle loop stays -- and releases the
thread as soon as execution moves past it. Only the block that cannot be compiled
is interpreted; the rest of the thread runs recompiled.
Leaving is safe at any instruction boundary, since all SPU state lives in
spu_thread, which is the assumption the JIT dispatch already makes. Re-entering
the bad block sets the flag again.
A block that fails to compile switches its thread to the interpreter. On ARM64
that fallback called spu_runtime::g_interpreter, which with a recompiler selected
is the LLVM-built interpreter, and calling it there executes nothing: measured a
million consecutive calls on Sonic Unleashed's stuck SPURS kernel without pc
moving once. The thread then spins in that loop forever at a fixed pc with no
flags set, which reads as a busy SPU and hangs the title with no diagnostic at
all. Any block that fails to compile landed there, so this was not one game.
old_interpreter is what the static decoder ultimately runs, through
tr_interpreter, and it is self-contained -- opcode table, thread, local store. Its
static-decoder-only check rejected exactly the case that needs it, so it now also
accepts a thread already marked for fallback.
Getting there also needed the give-up paths fixed: the TBL2/TBX2 retry could
return null with an empty error and fall through every branch unmarked and
unlogged, so nothing recorded that a block had been abandoned.
The stall dump now carries SPU event, MFC and interrupt state, which is what made
this findable: parked kernels showed pending=0 (no lost wakeup), intr_en was 0 on
healthy threads too (not interrupts), mfc_q was 0 everywhere (no stuck transfer),
and interp_fb=1 on the frozen thread pointed at the fallback itself.
Both waits in writer_lock are unbounded and silent. The acquire loop spins until
every range lock bit clears, and the range_lock path then spins until every
registered PPU thread reaches cpu_flag::wait. A thread that never gets there hangs
every other thread that takes a reservation, and leaves nothing behind: from
outside it reads as a clean guest deadlock with everything in a legitimate wait.
Both now log once, far past any plausible contention, naming the held range locks
or the PPU thread being waited on.
They paid for themselves immediately on Sonic Unleashed, which deadlocks at the
SEGA logo. Both stayed silent across several boots, which ruled out the VM lock
entirely -- worth having, since main_thread was pinned in cellSpursRemoveWorkload
carrying cpu_flag::memory without cpu_flag::wait, which looks exactly like this
bug and is not. The game hangs in a different state on different boots, so it is a
race elsewhere in SPURS.
Issue #16, both halves.
Installing a .pkg or .rap off a USB-OTG drive already went through the system
picker, and the descriptor it returns is handed to the native installer as a raw
fd, so a 40 GB package costs no copy. That holds only while the provider is
backed by real storage. The third-party USB-OTG and cloud apps people reach for
when the platform will not mount their drive return a PIPE, and every install
entry point seeks -- getFileType sniffs the magic and rewinds, package_reader
jumps around the archive -- so lseek failed with ESPIPE and a perfectly good
package was reported as unsupported or broken. Those descriptors are now
detected with the same lseek the core will make, and only those are copied to
real storage first, onto whichever of the emulator's own storage and the app
cache has more room. The copy is checked against the size the provider reported:
a short copy does not throw, it produces a truncated package that fails much
later as "broken", which reads as a bug report about the package.
Split releases picked through the system picker arrived in the order the user
tapped them, and installSplitPkg takes the order given as the part order, so
picking part 2 first extracted into a broken install rather than failing. Both
pick paths now sort the parts, digit runs numerically, since plain string order
puts part 10 between part 1 and part 2.
Installed titles were listed by title id alone -- NPUB90434, BLES01807 -- next
to an Uninstall button, which is where it hurt most: choosing which of two demos
to reclaim space from meant looking the ids up elsewhere. TITLE now comes out of
the install's own PARAM.SFO, read off disk rather than through the library cache
so a title the scanner has not seen yet is still named. The id stays on a second
line because patches, cheats and compatibility lists are keyed by it. Licence
files carry the same id inside their content id, so a .rap can name the game it
unlocks instead of being one of a row of indistinguishable hex strings.
The SFO field reader is the scanner's CATEGORY reader generalised rather than a
second copy of the 16-byte index-entry layout.
Accurate SPU Reservations off is worth a large amount in Spider-Man: Web of
Shadows and is not safe globally, so it goes in that title's own override rather
than the default.
Its SPURS reservation traffic serialises behind the global exclusive
vm::writer_lock that every reservation_op takes, which no amount of CPU can help:
all six SPU threads and several PPUs were measured yielding at the same rate with
18.8% of total CPU in sched_yield. Off, SPURS takes the lock-free path and
vm::writer_lock fell from 8.06% to 0.96%.
Kept per-title because it is off-spec -- upstream defaults it on, and Sonic
Unleashed fails EARLIER with it off, reaching neither the loading icon nor the
logo, which is consistent with the bypass being the SPURS area itself.
The seed writes only fields a title does not already carry, so a deliberate
change is never overwritten, and it runs once. Other regions of the same game
need their own entry.
Two things, both about the RSX waiting rather than working.
flush_command_queue ended by draining the present queue in case a queued frame
still held a ref to the command buffer just taken. It cannot: next() hands them
out from a 512 entry ring and the queued list is bounded at flip to
m_max_async_frames - 1, so the buffer being reused is hundreds of frames retired.
The guard was unreachable and the cost was not -- check_present_status pokes the
oldest queued frame's swap command buffer, and on Adreno vkGetFenceStatus blocks
until signalled rather than returning VK_NOT_READY, so a poll written to be cheap
became a full GPU sync. 1.32 times a frame at about 11ms: Fence poll 14.6ms ->
0.033ms, frame 44.5ms -> 36.6ms. Same fault as the two sites removed earlier;
this one sat inside flush_command_queue rather than on the present path. Ruled
out first: identical frame time at quarter resolution, and forcing the swapchain
pre-transform to match the surface left it unchanged.
The empty-ring yield is now charged to idle. It sits inside fifo_decode, which is
the enclosing scope of the whole run loop, so waiting on an empty ring was
reported as decode work -- Idle 0.003ms against FIFO decode 20.4ms, while a
native profile of the same thread put 34% of its cycles in sched_yield. The
bucket report and the profiler disagreed and the bucket report was wrong, which
has now produced two wrong conclusions in one session.
The 50us backoff in the FIFO_EMPTY path was added when the RSX thread was
measured spending 66% of its cycles in sched_yield on a machine starved for
cores: the affinity mask confined six SPU threads to four cores, and the
reservation path serialised everything behind a global lock, so a spinning RSX
took a core from threads that needed it.
Neither holds now, and the trade inverted with them. Measured after both were
fixed: 34% of eight cores busy, two to four threads runnable, five idle. Nothing
wants the core the sleep gives back, and the RSX sits on the frame's dependency
chain, so sleeping only delays it noticing the guest has produced work.
Also log the surface transform at swapchain creation. 30% of the frame is now in
check_present_status waiting on acquire_next_swapchain_image, which is not GPU
work -- a quarter-resolution run measured the same frame rate. Declaring IDENTITY
while the surface is rotated hands the rotation to the compositor, which can hold
images longer before releasing them for acquire. Logged rather than changed:
matching currentTransform means applying the rotation ourselves across the blit
and the overlay pass, and that is only worth doing if the two actually differ.
Dropped to 20 while the emulator was starved for cores, reasoning that a spinning
SPU steals a core from threads doing real work. Two things have changed
underneath that: the affinity mask no longer confines six SPU threads to four
cores, and the reservation path no longer serialises everything behind a global
lock. Measured after both, in game: 34% of eight cores busy, two to four threads
runnable, five idle, and no thread near saturation.
Tested at 100 and at 20 with no difference, which fits -- the wait is no longer on
the critical path, so how it waits does not matter. Upstream's value stands
rather than carrying a divergence that buys nothing.
The affinity migration turned the scheduler on so the big.LITTLE mask would
apply, keeping SPU and RSX off the A510s that run at roughly 27% of prime-core
capacity. That reasoning holds for one thread per core and breaks down at six.
Measured in game on a Snapdragon 8 Gen 2, reading the masks the threads actually
carry:
app cpuset (top-app) 0-7 Android grants every core
SPU[0..5] 3-6 six threads, four cores
rsx::thread 3-7
Six SPU threads sharing four cores get about two thirds of a core each, which is
worse than one thread owning an A510 outright, and it caps the emulator: the
device sat at 60% busy with cores 0-2 idle while frames were slow. Spider-Man:
Web of Shadows is visibly better at OS.
Worth being clear that the mask is ours and not Android's -- the app is in
top-app with all eight cores granted -- which is also why these devices are
reported to run better under native Linux, where no such policy is applied.
The other modes stay selectable for anyone whose device disagrees.
A DualShock 3 sends pressure bytes in every packet. The press setting governs how
much of the buffer the game is told is valid, not whether the pad produced the
values, so clearing the area diverges from the hardware: a game that reads a
pressure byte without having asked for press mode gets 0 where it would see a
press on a console.
Spider-Man: Web of Shadows does exactly that for R2. Tracing both entry points
showed it never calls cellPadInfoPressMode or cellPadSetPressMode, so the setting
stays at 0, yet it reads the R2 pressure byte to decide whether the trigger is
held. R2 did nothing in that game while every other button worked, from the
controller and from the touch overlay and after remapping to a different physical
button, because the digital bit was delivered correctly the whole time and was
never what the game looked at.
len is unchanged, so a game that honours it sees what it saw before.
The out-buffer check answers 'unlikely to be a loop', and that answer is not
free. It resets the spin count and leaves the busy-waiting switch at umax, so the
caller skips busy_wait, skips the sleep path, and returns immediately: the SPU
re-executes GETLLAR at full rate with no backoff. The spin count never reaches 4,
so the spin optimisation is never evaluated, and the 400ms fallback that would
force a sleep is never reached either. One SPU in that state holds a core flat
out, and the setting meant to control this has nothing to act on.
Spider-Man: Web of Shadows sits in that case. Its GETLLAR sites use an LSA in the
top 64K of local store, which is what the check looks for, and process_mfc_cmd
measured 55% of all CPU across the process while the game ran at 10-15fps.
Re-entering the same site with the same stack 32 times is itself the evidence
that it is a loop, whatever the LSA looks like. After that the verdict is dropped
and the normal spin detection decides between busy-waiting and sleeping. Any real
change of site or stack resets the count, so a genuine OUT buffer still gets the
original treatment.
SPU GETLLAR Busy Waiting Percentage defaults to 100 upstream, meaning always
busy-wait. That suits a desktop, where the SPU threads have cores of their own
and spinning costs nothing else. Here six of them share eight cores with the PPUs
and the RSX, so a spinning SPU takes a core from the threads doing the work.
Measured on Spider-Man: Web of Shadows: process_mfc_cmd accounted for 55% of all
CPU across the process, and making its inner loop cheaper did not move the frame
rate -- the loop just ran more iterations in the same wall clock. That is what
identified it as a spin rather than as work, after two rounds of optimising the
iteration itself.
20 still favours a short busy-wait, so a reservation that frees quickly is caught
without a scheduler round trip, and only a wait that history says is long goes to
sleep. A deliberate change in All Core Settings still wins, since core overrides
replay after this.
Gating the check on getllar_spin_count was not enough. That counter is reset from
several other paths, so it is frequently zero and the callstack was still rebuilt
constantly: measured 19.6% of all CPU inclusive in dump_callstack_list, the
largest single item after process_mfc_cmd itself.
Key on the values the answer actually depends on instead -- pc, the stack pointer
and the link register -- and recompute only when one of them moves. Only the
innermost frame is ever used, so that is all the memo keeps.
A stale answer across an unrelated LS write is acceptable here. This decides only
whether the address looks like a caller's OUT buffer, on a heuristic whose own
comment calls it 'unlikely to be a loop'.
The out-buffer check in the GETLLAR spin detector rebuilt the callstack on every
iteration of a busy-wait loop. dump_callstack_list walks the stack and calls
is_exec_code for each candidate, which allocates a vector<bool> and scans for
branch targets, so the cost is large next to what it decides.
On a whole-process profile of Spider-Man: Web of Shadows those three came to
about 14% of all CPU -- more than the RSX thread spent on the frame -- because
the game's SPU code spins on GETLLAR with an LSA in the top 64K of local store,
which is exactly the case the check looks at.
Once per sequence is enough. pc, ch_mfc_cmd.lsa, gpr[1] and addr are all compared
against the previous iteration a few lines above and any change resets the
sequence, so the callstack cannot move underneath a spin.
sched_yield does not idle a core. With every core already busy it returns almost
immediately and the RSX thread takes it again, running flat out producing
nothing. A native profile of Web of Shadows put 66% of this thread's cycles in
sched_yield and its kernel path against 9% in run_FIFO, which the bucket profiler
reports as a busy RSX because the yield happens inside the fifo_decode scope.
That is not free even with an empty ring. The RSX affinity mask covers the whole
fast cluster while the SPU mask is that cluster minus the prime core, so any of
this that lands off the prime core is taken from the SPU threads, and those are
what the frame is actually waiting on at 61% of all CPU.
The spin still runs 64 times before sleeping, so a producer that is merely slow
is met without a scheduler round trip. Only a ring that has genuinely gone quiet
reaches the 50us sleep, which is far below the frame times where this matters.
Android only.
A query that begins inside a render pass instance has to end inside that same
instance. Ending the pass underneath an open one leaves it permanently
unavailable, and on Turnip it takes the device with it, reported later against
poke_query because that is the first call that waits on a result.
Queries do begin inside render passes here. VKDraw only lifts them out when
use_strict_query_scopes() is set, and that is wired to Strict Rendering Mode, a
user performance setting rather than a driver quirk, so it is off for almost
everyone.
Twenty-one call sites end a render pass and only one, in VKDraw, ever paired
itself with a cleanup. change_image_layout alone ends 41 passes a frame in Web of
Shadows, and any of them can land while a query is open, which is why fixing the
two sites in the query pool moved the device loss from one minute to nearly three
rather than removing it. Holding the invariant in end_renderpass covers all of
them, including any added later.
The VKDraw site now cleans up before the pass ends rather than after, which is
the order the spec asks for; its own call becomes a no-op.
The migration that turned relaxed ZCULL on recorded it twice: once in the curated
store and once as a raw core override. The migration that turned it back off only
corrected the curated field, so the two stores disagreed, and the override is the
one that reaches the core -- overrides re-push at the tail of applyTo, after the
curated store has written the setting.
The toggle therefore read OFF while config.yml read 'Relaxed ZCULL Sync: true' on
every boot, with no way to change it from the UI. That is not cosmetic: relaxed
sync is what allows queries to be read while still pending, which is the path
behind the 'Dubious query data pushed to cond render' warnings, and it also
selects emulated predication in the VK backend.
Emulated conditional rendering exists for hardware that never had the extension.
Turning the extension off as a driver workaround enabled it by accident, because
both are selected by the same test, and the two halves do not fit together:
begin_conditional_rendering returns early without building m_cond_render_buffer,
while the vertex shader still reads that buffer at offset 0. It gets a zeroed
scratch buffer, predicates every draw away, and the game renders black with audio
and overlays still running. The all-ones word that disables predication sits at
offset 4 and is never reached, since the fallback leaves hw_cond_active set.
Off on these drivers means occlusion results stop culling draws, which is the
trade the workaround already documents.
The vendor comes off the GPU rather than from get_driver_vendor(), whose cached
value is not assigned until later in the same function and would still hold the
previous device's.
The existing gate covered only the proprietary driver and said Turnip was left
alone until there was evidence about it. There is now.
Web of Shadows loses the Vulkan device about a minute into gameplay on Turnip 26
/ Adreno 740. The assertion names poke_query, which is the first call that reads
a result rather than the one at fault. Conditional rendering is the only place we
record vkCmdCopyQueryPoolResults with VK_QUERY_RESULT_WAIT_BIT, and that form
makes the GPU block until the query resolves, so a query that never resolves
hangs the device instead of the caller and the watchdog ends the session. The
same run logged 169 'Dubious query data pushed to cond render' warnings, which is
this code being handed queries that are still pending.
It is also the churn: the aggregation barriers closed 42 of the 91 render passes
in a measured frame, and ending a pass on a tiler costs a tile store and reload.
Both drivers now fall back to thread::begin_conditional_rendering, the path
desktop already takes wherever the extension is absent.
Fixing the device loss by ending the render pass before vkCmdCopyQueryPoolResults
introduced a hang in its place. A query that begins inside a render pass instance
has to end inside that same instance; ending the pass underneath an open one
leaves it permanently unavailable, so get_query_result spins on poke_query with
no way out and the RSX thread stops.
Nothing reported it. The submit-time ensure() only checks that the query was
closed, and end_occlusion_query closes it a moment later, so the assert passes
while the result never arrives. The stall detector runs from do_local_task in the
FIFO loop, which the spin has already left, so the profiler charged the wait to
FIFO decode and the frame read as CPU-bound -- 93% in a bucket that was really
the thread sitting in sched_yield. Web of Shadows locked up this way after
reaching gameplay, audio and vblank still running.
Both sites that end a pass from the query path now close an open query first,
which keeps begin and end within one pass. The query is cut short, as it is
anywhere do_query_cleanup is used.
The wait itself is now bounded as well. It warns at one second and abandons at
three, using whatever the query holds: wrong culling for a frame is a better
failure than a thread that never returns, and the log names the cause.
The drain fix made every path that can idle or block publish GET immediately,
which is required for correctness: a producer waiting on ring space needs to see
the progress we made before we stopped consuming.
It publishes far more often than that requires. The empty and busy cases return
straight to the run loop, so a ring that has gone quiet re-enters them once per
iteration with GET unmoved. Web of Shadows measured 137000 loop iterations per
frame against 46000 method dispatches; the remaining 91000 were republishing a
value the guest already had.
GET shares a 64-byte line with put, which the guest PPU writes from another
cluster, so each of those is a coherence miss taken against the thread feeding
the ring. The cost lands on the producer rather than on the RSX, which is why it
presented as a freeze with sound still playing: the PPU stalls on the contended
line while threads that never touch it keep running.
GET is ours to write, so tracking the last published value and skipping an
unchanged store keeps the guarantee -- progress is still announced exactly once
after the last advance -- without the repeats.
vkCmdCopyQueryPoolResults must be recorded outside a render pass instance. This
recorded it inside one, with a comment saying we are technically supposed to stop
the pass first but that it does not matter on IMR hardware. It is not a
technicality -- inside a pass it is undefined behaviour, and a desktop GPU
tolerating it says nothing about a tiler.
This device lost the Vulkan device over it. The fault surfaced later, in
poke_query, because that is the first call that waits on a GPU result, so it read
as the query READ being at fault when the damage was done at record time. Only
the RSX thread died, so the process kept running with audio and vblank alive and
it presented as a hard freeze rather than a crash. Verified gone: zero device
losses on a run that previously died within a minute.
The pass is ended only when one is actually open, on a path that already stalls
for a GPU result, so the flush the upstream comment worried about is paid where
we were blocking anyway -- and disabling occlusion queries is not the
alternative, measured here at 80ms frames with broken visuals.
The two by-pass tables were joined on counters that reset at different points.
tick_frame runs from on_frame_end, before flip; the GPU timer rotates its slot at
the top of flip and then drops every non-frame region on the fresh slot, which is
flip's own overlay and calibration passes -- and those still incremented the CPU
counter. So the CPU ordinal ran ahead by the number of present-path passes and
the two tables described different passes. A whole anomaly came out of that: a
pass whose GPU cost was joined to a neighbour's workload read as 36x the per-draw
cost of its peers. The comment claiming both reset on the same boundary was
wrong. Reset where the GPU slot actually rotates instead.
Also adds a Storage Access Framework route to the package installer. The in-app
browser walks java.io.File, which only reaches storage this process can open by
path, so a .pkg on a USB-OTG drive or an SD card was unreachable and had to be
copied to internal storage first. Packages are handed over as the descriptor SAF
already returned -- the native side takes a raw fd, so nothing is copied and a
4 GB package costs no extra space; licences are 16 bytes and their installer
wants a real file, so those alone are staged.
The low-memory serialisation added three days ago holds a std::mutex across the
LLVM compile itself. A worker that hits LLVM's fatal handler leaves through
pthread_exit, and bionic unwinds nothing on that path, so the mutex stays locked
by a thread that no longer exists and every remaining worker waits on it for the
rest of the session. Memory only falls as modules accumulate, so the tight-memory
branch is likeliest late in a run -- which is why it reads as the PPU cache
getting stuck at the very end, and why dropping to the interpreter avoids it.
Reported as Saint Seiya: The Sanctuary never finishing its module cache. A claim
taken by compare-and-swap and waited on with a timeout costs a stranded claim a
wait rather than the session; the memory back-pressure either side of it is
unchanged. Third time this fork has been bitten by an unbounded wait around a
thread bionic can kill without unwinding.
overlay_audio.cpp already accounts for a platform with no video source; the same
ensure() was left in overlay_video.cpp. Android's make_video_source returns
nullptr, and overlay_save_dialog builds a video_view for EVERY entry on all three
of its paths, so opening a save list aborted as soon as there was one save to
draw. It presents as the save menu never opening -- reported against Ratchet &
Clank: Tools of Destruction and Devil May Cry 4, and against Web of Shadows,
which stalls only once a save exists to be listed. Bundling the overlay icons
was necessary but not sufficient: the dialog still could not survive drawing.
The still image is what an entry needs; the animated ICON1.PAM is the part no
backend here can supply. Also dumps SPU thread pc and block hash alongside the
PPU dump when frames stop, which is what named the SPURS kernels as idle rather
than spinning in guest code.
overlay_controls.cpp loads a fixed set of PNGs -- button glyphs, save.png,
new.png, spinner -- through fs::get_config_dir() + Icons/ui/. Desktop ships them
beside the binary; nothing put them on Android, so every load failed and the log
said so on each boot. The visible cost was cellSaveData's list: it is a native
overlay that draws its rows with save.png/new.png, so the load-save menu a game
opens never appeared. Reported against Ratchet and Clank: Tools of Destruction
and Devil May Cry 4, both fine on emulators that ship the icons.
Bundled from bin/Icons/ui and staged into config/Icons/ui once, revision-guarded,
before the core can draw its first overlay.
The nineteen translation files were ARMSX2-era: about nine hundred of their keys
still existed and showed the old PS2 wording -- worse than the English fallback,
which is at least right -- and roughly eight hundred current keys had no
translation at all. Regenerated all nineteen from the 1041-key map, batch plus
per-line retry, with every %s/%d checked against the source so no broken format
string ships. A string that would not translate is omitted and falls back to
English rather than shipping wrong.
The memory-card and PNACH patch screens were PS2 concepts with no PS3 counterpart
and no caller left -- the drawer had already been cleaned, so they were dead code
holding dead strings. The PNACH downloader went with its only consumer, and the
PCSX2-Android.ini seed could never exist under this package. The session log now
announces ARMSX3_INIT instead of PCSX2_INIT, which had every bug report opening
with the name of a different emulator.
The English map drops 46 PS2 strings and 619 orphans nothing references (1705 ->
1041 keys), rewords the four live strings that still said memory card, and renames
about.pcsx2.* to about.rpcs3.* to match what they already said.
A guest that stops progressing presents nothing further, so whatever was drawn
last stays on screen for good. When that frame held the boot progress bar it read
as stuck compiling at 1s remaining, and it looked identical across five unrelated
faults -- it sent every report of them to the wrong place, including this week's.
Nothing contradicted it either, since the emulator has not crashed and logs no
error.
Reports once, to the log and to the screen, after thirty seconds with no frame
and nothing claiming to be in progress. The progress text is what separates
working quietly from stopped: a shader or PPU compile presents no frames for
minutes and holds a dialog saying so. Drawing it needs the native UI flip,
because the guest is not flipping -- which is the point.
GET goes out on a bounded lag, every eighth packet, to keep a cross-cluster
coherence miss off the per-packet path. That is only sound while more packets are
coming to flush it, and the paths that can block were given a forced publish for
exactly that reason -- but the one where the ring runs dry was not, and it is the
one where nothing further will ever flush it.
The guest reads GET to see how far the RSX has consumed. Draining the ring left
it up to seven packets behind with no more packets to publish, so the guest waited
on progress that had already been made and never announced.
It presents as a boot or a load that hangs with the RSX perfectly healthy and idle,
every guest thread in a legitimate wait, and no error anywhere: bisected to this
across six rounds after five wrong theories, because nothing is broken at the point
it stops. Only bites when the packet count is not a multiple of eight as the ring
drains, which is why it was game- and timing-dependent and why the same title could
boot yesterday and hang today.
Publishes on both paths that leave the consume loop. The lag stays.
The RSX-side stall report says what the RSX is doing, which on every hang chased
so far has been idling while the guest waits -- and nothing said which guest
thread or what it was in. The syscall stats name the syscall without the caller,
and a thread that has not started reads from /proc exactly like one that is
blocked.
One line per PPU thread with its name, state, PC and current function, on the
same condition and cadence as the RSX report. Reads the id map unlocked on
purpose: this runs on the RSX thread, and taking that lock to diagnose a hang
would add the kind of dependency being diagnosed.
The profiler arms and reports only from on_frame_end, and dumps once 300 frames
have accumulated, so a boot that hangs before presenting left it switched off and
silent however the setting was set -- the one case where what the RSX thread is
looping in is the whole question. Armed and polled from do_local_task as well,
which the FIFO loop reaches whether or not frames advance.
The vblank thread is the only source of the interrupt gcm waits on, and said
nothing about being alive, blocked or gone. A heartbeat and an exit reason
separate those, which are three different faults that look identical from
outside: on a Demon's Souls boot it delivered about 120 vblanks and then parked
in the send path with the queue undrained.
Waiting on the claim was untimed, so a waiter that missed the owner's transition
waited for the rest of the session, and the duplicate waiter could only leave on
the failure state -- an owner that published state 2 without publishing a
function left it waiting on something that was never coming.
SPURS brings all of its kernels to the same block at once, so this is four
threads at a time, and the PPU then blocks on SPUs that never answer. Measured
during one: the PPU thread took no CPU at all across eleven minutes while the
SPU threads churned two-to-one system time.
Both waits are bounded now and the duplicate leaves when the owner has finished
and published nothing.
Auto-save on exit, auto-load on boot and the interval auto-save were ARMSX2 shims
returning false that were never ported, so all three toggles persisted and read
back while doing nothing -- the interval job woke on schedule to call a function
that always failed. They now use a reserved slot above the ten the picker shows,
reusing the numbered-slot path rather than growing a second mechanism.
Import treated getGamePathSlot as a file path, but it answers with the title id:
File(id).exists() was false for every slot, so the first OCCUPIED slot read as
free and the destination resolved against the process working directory. It
copied the file nowhere useful and reported the slot it had not written.
Occupancy now comes from the core and the destination from the real path.
Also says how large a state is before the storage bill arrives, and stops the
interval description promising a pause when a PS3 save is a stop and a reload.
The reader rooted the path at systemDirPosix(), which is null unless a data
folder was explicitly picked, so on a default install it returned before it ever
looked and every slot drew as a blank tile with the thumbnail sitting on disk
beside the state it belongs to. Falls back to getExternalFilesDir, which is what
fs::get_config_dir() resolves to on that install and where the writer put it --
the same fallback inputProfilesDir() makes for the same reason.
CallFromMainThread without a wake_up is a post: upstream hands the callback to
the GUI thread and returns. This ran it inline instead, under whatever locks the
caller held.
lv2_obj::sleep_unlocked posts one while holding lv2_obj::g_mutex, which is what
the comment on that call site is about. The callback is FinalizeRunRequest, the
wake for a restored savestate, so it took g_mutex against itself and every thread
stopped there: the log reaches Final Thread and goes quiet with the SPUs spinning
and the progress overlay frozen on its last figure. It took out loading a state
and saving one alike, a save being a stop and a restore.
Callers passing wake_up are waiting on completion and still run inline.
The hash it changed is one of several computed over the same function data, and
only this one moved, so the cache rebuild it forced ran through a path whose
other sites disagreed. Loading a state then wedged in Building SPU cache with
nothing compiling.
Reuse of blocks across a change of the setting is still wrong, but it is an
upstream behaviour that predates this and is better addressed by invalidating
the object cache once when the mode changes than by moving one hash out from
under the others.
already_added reports that the title was already in games.yml, which is the
normal case for anything booted once before, and RPCS3's own front-end passes it
through for that reason. The bridge failed every result that was not NoErrors,
so the boot was abandoned and the user was returned to the library with "Game
failed to start: AlreadyAdded".
The bridge translates each section it knows and returns false for the rest, and
there was no Savestate case, so the write was dropped on the floor. The setting
could not be turned on at all -- not by the default, not from the settings row --
and savestates failed to lock the SPUs while telling the user to enable exactly
the option that was being discarded.
The setting changes the code generated for blocking channel reads -- the GPRs
are stored rather than the thread being marked unsavable -- but the cache key
was a hash of the guest code alone, and the compiled object is cached under it.
A block built in one mode was therefore reused unchanged in the other, so
turning the setting on left the old unsavable blocks in place and savestates
went on failing to lock the SPUs with a message telling the user to enable a
setting they had already enabled.
Mixed in only when set, so caches built in the default mode stay valid.
A save that fails with "missing SPU setting" reads as broken rather than as a
setting waiting to be found, and the setting is not one a player would think to
look for. Upstream defaults it off to protect SPU performance; here the feature
not working at all is the worse trade.
Costs are unchanged and still stated on the switch: it slows the SPUs while it
is on, and a PS3 state runs 500MB to 3GB. Turning it off restores upstream
behaviour and gives the performance back, and that choice is now respected on
every boot rather than overwritten.
Release 0.4.1.
Save states could not be taken at all. Saving has to stop every SPU somewhere it
can be serialised from, which is what Compatible Savestate Mode does, and this
port wrote that setting to false on every boot -- so the save failed with
"missing SPU setting" no matter what the user did, and nothing on screen
connected the two.
The reasoning was sound: the mode costs SPU performance, and a PS3 state runs
500MB to 3GB, so a few saves fill a phone. Both of those are costs to disclose,
not reasons to decide for someone. It is now a setting, still off by default, so
nobody pays for a feature they did not ask for and installs from the window when
it was pushed as true are corrected by the same write.
Placed with the SPU rows rather than under a savestate heading, because that is
where the cost lands, and the description says what both costs are before the
switch is touched.
run_recoverable_llvm runs code generation on a disposable thread and terminates
it through pthread_exit when LLVM invokes its fatal error handler. bionic does
not force-unwind C++ frames on pthread_exit, so the lock MCJIT holds over the
execution engine is never released and stays held by a thread that no longer
exists.
Every entry point into the engine takes that lock, so the next compile hangs and
so does teardown. One recovered error, and the emulator is finished until it is
killed -- and the error is recovered, which is the point: it is meant to be
survivable.
Their Kotlin log-channel screen is left out; this port has its own.
From MaxsTechReview in PS3Native.
Two places compute a size from values a file supplies and use it without
checking it is possible.
A SELF or SCE header gives the metadata offset and the header size, and the
buffer between them is sized by subtracting one from the other. Both are
unsigned, so a truncated or malformed dump that puts the offset past the header
end underflows into a near-SIZE_MAX allocation, which fails as an out-of-memory
rather than as the bad file it is. Reject the layout instead.
Texture uploads take the mip levels the guest describes and write them into an
image built from the destination's own dimensions, which can hold fewer. Drop
the levels that do not fit rather than writing past what was allocated.
From MaxsTechReview in PS3Native.
get_cycles passes the pthread handle to pthread_getcpuclockid, which glibc
answers with an error for a thread that has already exited -- the else branch
below returns the last known value for exactly that case. bionic instead looks
the handle up in its list of live threads and aborts the process when it is not
there.
m_thread is never cleared when a thread ends and the performance overlay samples
every PPU, SPU and RSX thread on a timer, so one finished thread is enough to
take the emulator down with it. Latent here rather than absent: it needs the
overlay on and a thread to have gone.
Record the kernel tid at initialize and build the per-thread clock id from it
the way bionic does once its own lookup succeeds, so clock_gettime simply fails
for a dead thread, which is what the surrounding code already expects. Cleared
at finalize so a thread stops being sampled before it goes away. Other platforms
keep the original path.
Found and fixed by Zulux91 in PS3Native.
A licence installed from a locked game's own menu reported success and left the
game locked. The install writes the file into exdata under its own name, because
a RAP's name is conventionally the content id it unlocks, and that convention is
the whole lookup: the core opens exdata/<content id>.rap and nothing else. A file
saved as "license(1).rap", renamed, or tidied up on the way over therefore lands
where nothing will look for it, and the only feedback is the game asking again.
Where the game is known, ask it. The native path decrypts the EBOOT's
supplemental header to read the content id, which works on a locked game because
that header is not what the licence protects, and names the file correctly
whatever the user's copy is called. Falls back to the name when there is no game
to ask or the header cannot be read.
The package screen keeps the name-based path: a licence installed on its own has
no game to resolve against.
Those two are 7.4 ms a frame in Arkham City, a fifth of it, and the GPU side of
the same work is 0.76 ms, so it is host work. Each handler does several unrelated
things and nothing separates them: a read barrier that can force a readback, the
memory copy the transfer exists to perform, and in the blit engine a software
scale through ffmpeg for the cases the GPU path does not take.
Scope the three. The scale is scoped inside convert_scale_image rather than at
its four call sites in the blit engine, and only the RSX thread is ever reported,
so calls from elsewhere cost nothing to cover.
Also let a scope be closed early, so a region ending part-way through a function
does not need a block introduced purely to place a brace.
The name table is keyed by register index, and both method reports passed the
byte offset. A lookup therefore matched whichever unrelated method happened to
have that value as its enum, so the costliest entry in Arkham City came out as
NV4097_SET_CONTEXT_DMA_VERTEX_B, which has no handler and cannot cost anything.
It was NV406E_SEMAPHORE_ACQUIRE: the RSX waiting for the guest to signal, which
is the one entry in that list that is supposed to block and the one that should
not be optimised.
A wrong name is worse than none here. The hex fallback was right the whole time
and is left as the byte offset, which is what a reader looks up.
The handler bodies are 60% of the RSX thread in Arkham City and the dispatch
machinery around them is 3.5%, so the question is which handlers. The method
histogram cannot answer it: it counts calls, and the busiest method may be a
register write while a rare one does the work.
Keep the interval the dispatch site already measures. It brackets the call with
two counter reads to fill the method_call bucket and then throws the difference
away; billing it to the method's slot as well costs one add.
Inclusive of whatever the handler calls into, including scopes that charge
themselves elsewhere. For ranking handlers that is the useful reading, and the
per-bucket totals stay exclusive as they were.
Arkham City spends 38.5 ms a frame in FIFO decode, 58% of the RSX thread, at
164 ns a dispatch. Sonic manages 45 ns on the same loop, the same decode and the
same counters, so the difference is in what the handlers do rather than in the
dispatch. Nothing separates the two: fifo_decode encloses the whole loop, so it
holds every handler body as well as the machinery around them.
The two handlers already scoped, transform program and transform constant,
measure 0.02 and 0.06 ms here, which rules them out and leaves the rest of the
mix unaccounted for.
Wrap the handler call. This is the only per-dispatch scope in the profiler and an
earlier attempt at one measured mostly itself; it is affordable here because it
brackets a call rather than a loop iteration, and handlers carrying their own
scope still attribute inward. It costs a few percent of the bucket it splits, so
the split is the number to read, not the total.
Booting a second game without restarting the app builds a new RSX thread.
set_enabled is the only thing that binds the profiler to a thread, and it
early-returns when the setting has not changed, so it stayed bound to the
previous game's thread. Every scope then failed its owner check, nothing
switched buckets, and the whole window was charged to whichever bucket happened
to be current.
That prints as "FIFO decode 100.0%", which is indistinguishable from a genuine
finding about a command-bound title, and was briefly read as one.
Notice the change per frame and re-bind, dropping the accumulated window and the
per-pass counters: they belong to a thread that is gone, and keeping them would
blend two games into one report.
The include list came over from ARMSX2 and names sstates, memcards, gamesettings,
cheats and snaps. None of those exist here, so a backup collected a few kilobytes
of controller profiles, reported success, and left every save behind. Nothing
warned: skipping an absent folder silently is right for an optional one and wrong
for a list aimed at a different emulator.
Name RPCS3's paths instead. The part that matters is config/dev_hdd0/home, which
holds save data, trophies and licences, and is the only thing in here that cannot
be rebuilt or re-downloaded. Save states, input configs and patches come along.
Installed titles are left out on purpose, along with firmware and dev_hdd1: a PKG
reinstalls and a PUP reinstalls, a save does not, and that is the line this list
is drawn on. Including them would have taken the archive from tens of megabytes
to nearly five hundred on the device this was sized against.
The description on the screen said memory cards and artwork too, so the one place
a user could have noticed agreed with the bug.
Everything measured so far describes what a draw contains: vertices, pixels,
shader length, subdraws, barriers. By all of them pass six should be the
cheapest of the expensive passes, and it is the dearest by a factor of seven.
An occlusion query is none of those things. On a tiler it makes the visibility
stream resolve, it costs the same whatever the framebuffer size, and no counter
here would show it. That matches every property this pass has: indifferent to a
sixteen fold cut in pixels, indifferent to tiling being switched off, no
barriers, one subdraw per draw, shorter shaders than the passes it dwarfs.
ZCULL is active, and emit_geometry opens a query whenever the command buffer
carries the occlusion flag. Count them where they open.
Installing a .rap never worked at all. The package screen routed licences to
installKey, whose RAP branch works out the content id by decrypting the game's
EBOOT, so it needs a game path, and the only caller passed an empty one. Every
attempt died at "Failed to fetch NPDRM of SELF". A RAP's filename is the content
id it unlocks, which is why RPCS3 desktop's InstallFileInExData simply copies the
file into exdata. That is what this does now, lower case extension included,
because unself.cpp searches for it that way.
Picking a game together with its licence could not work either. The installer
routed on file count rather than file kind, so any multi file selection went to
installSplitPkg, whose first act is to reject anything that is not a .pkg part.
Multiple selection has been allowed since split packages landed, so the obvious
thing to do was the one thing guaranteed to fail. The selection is split by kind
now, packages first, since a licence unlocks content the package has to have
written already.
Both failures showed the same generic "Install failed. The file may be encrypted,
incomplete or not a PS3 package", which reads as a bad file rather than a bug in
the app. The reason the native side already reported now reaches the screen.
A licence-locked title also looked like any other until it refused to boot. The
core works that flag out by attempting decrypt_self on the EBOOT, but the library
never asked it. The scan asks now, and a locked game gets a badge on its cover, an
Install licence entry in its context menu, and a prompt instead of a doomed boot
from every launch path: the library cards, the context menu, the controller, and
the settings screen's Play button.
Boot failures were silent besides. Rpcs3Bridge.boot threw away BootGame's return
code and MainActivityRuntime dropped runVMThread's result, so a failed boot was
indistinguishable from a game that started and exited immediately. Both are
reported now, which is how I found the licence problem in the first place.
External intents and launcher shortcuts are not covered, because externalGameInfo
builds a fresh GameInfo where locked defaults to false. Those still fall back to
the boot failure message.
Uninstalling only ever removed dev_hdd0/game/<TITLEID>, so the title's compiled
code and shader cache stayed on disk forever. On my device that was between 7 and
58 MB per title, and one of those caches belonged to a game I had already removed.
I made it a checkbox on the existing confirmation rather than doing it silently,
defaulted on, which is how RPCS3 desktop's own remove dialog treats caches. The
row is hidden when there is no cache, and it shows the measured size so you can
see what you are freeing. The size is measured off the main thread because a cache
directory holds hundreds of files and this runs while the dialog is opening.
The cache goes only after the native uninstall reports success, since dropping the
cache for a title that is still installed would just cost a recompile. The title
id is validated before the recursive delete: it comes from a directory listing,
but a path separator or a dot dot in it would resolve outside the per title
folder, so anything that is not a single plain segment is refused.
Save data, trophies and licences are deliberately left alone. Those belong to the
user rather than to the install, and desktop does not offer to remove them either.
A session ended in a fatal VK_ERROR_OUT_OF_DEVICE_MEMORY, the first in any log
here. On this GPU that is system memory, and the device had two gigabytes free
of seven with the emulator holding most of the rest.
Two frames in flight is what makes the CPU and the GPU overlap, and it is also a
second frame's worth of resources alive before anything retires them. That trade
is worth making at rest and not worth making into a crash on a handheld sharing
memory with everything else.
Fall back to the single frame this used to run with when the memory load is
above low. Slower, and slower is recoverable.
Shader length settled that pass six is not the game's workload: it has the
shortest shaders of the expensive passes, a quarter of the vertices of a pass
that costs a seventh as much, no barriers, and no reaction to resolution or to
tiling being switched off. Every quantity measured so far says it should be
cheap, and it takes nine milliseconds.
The draw count is the one that has been lying. It counts clauses, and a clause
is expanded over its subranges, so a single entry can become thousands of draws.
Batching them through VK_EXT_multi_draw, which this device does support, saves
our command overhead and changes nothing about how many the GPU processes.
Count them at every submission site. Thousands of tiny draws at a fixed cost
each is the last shape that fits, and nothing else measured would reveal it.
Pass six costs about 26 times what pass eight does per vertex: 123 draws and 68
thousand vertices for 9.15 ms against 532 draws and 253 thousand vertices for
1.26 ms. It has no barriers, does not care about resolution, and does not change
when TU_DEBUG=sysmem takes tiling and binning out of the picture entirely. The
only thing left that behaves that way is the shader.
Record vertex and fragment ucode length per pass. This decides whether there is
a bug here at all, which nothing measured so far can: shaders genuinely that
much longer are the game's own workload and there is nothing to fix, while
comparable ones mean something is happening to those draws that should not be.
The previous commit read driver_env.txt before the log file was opened, so the
one thing worth knowing -- whether the option was applied -- was written into a
listener that did not exist yet and then thrown away when the log rotated.
Move the read to just after the log file is created and report each option by
reading it back rather than echoing what was meant to be set. There is no other
honest confirmation available: /proc/<pid>/environ is the snapshot taken at exec
and never reflects a runtime setenv, and Mesa's own logging goes to stderr,
which Android discards. Still long before any Vulkan instance exists, which is
the only ordering Mesa cares about.
This device needs Turnip; the stock Adreno driver does not render the game at
all. Turnip is steered by environment variables such as TU_DEBUG, and the usual
way to set one on Android, the wrap.<package> property, is ignored on a user
build. It can be set and read back while never reaching the process, which makes
a flag that never applied look exactly like a flag that made no difference. That
is how the first attempt at this measured stock Turnip twice and called it a
result.
Read NAME=VALUE lines from <root>/driver_env.txt during initialize, before any
Vulkan instance exists, since Mesa caches each option the first time it is read.
A missing file does nothing, which is the normal case.
Picking a scale and launching a game still rendered at native. The previous
attempt read the launch-time write from ps3.resolutionScale, which turns out to
be the wrong end of it: that field has no writer anywhere in the UI, so it holds
its default of 100 permanently.
applyTo pushed that default onto Video@@Resolution Scale, the same node the
upscale multiplier writes, and applyTo runs after the launch path, so the orphan
won every time. Changing the scale in game appeared to work only because nothing
calls applyTo again afterwards.
Emit the node from upscaleFloat instead, which is what the preset grid, the
custom percentage slider and the in-game overlay all write, using the same
conversion and clamp as the other writer so the two cannot disagree. Restores
the launch-time call to the multiplier it always used.
Picking a resolution scale and then launching a game ran at native. The UI kept
showing the chosen value, the config held the default, and changing it in game
worked, which made it look like the setting was not saving.
Both settings write the same native node. applyTo writes the PS3 percentage to
Video@@Resolution Scale, and renderUpscalemultiplier writes the ARMSX2-lineage
multiplier times a hundred to the same place, from the launch path, after
applyTo. So the last writer won and it was the one carrying a default of 1.0.
Changing the value in game appeared to work only because nothing writes the node
again afterwards.
Drive the launch-time write from the PS3 setting so the two agree. Same node,
one owner.
Pass six spends 9.8 ms on 44 draws and 33 thousand vertices at ordinary
resolution, which is 300 ns a vertex. That is not vertex work, and the 2048
square shadow map next to it costs under a quarter of a millisecond with twice
the geometry, so it is not target size either. What is left is the GPU being
serialised inside the pass.
texture_barrier keeps the pass open on Android and issues a by-region
self dependency instead, which was the right trade against a tile store and
reload. But that barrier still makes a tiler resolve the tile and fetch it back,
and one per draw would cost about what pass six is costing. Nothing counts them.
Count barriers issued while a pass is open, per pass, and how many came from a
cyclic reference. If pass six shows one per draw the mechanism is named; if it
shows none, the serialisation is somewhere else and this rules out the obvious
candidate cheaply.
Two passes hold 69% of GPU time and one of them, pass six, costs 76us a draw
against 2.4us in pass eight while holding 7% of the frame's draws. Rendering at
quarter resolution changed nothing, so it is not fragment work, and the ordinal
on its own says nothing about what the pass is for.
Record the render target size and the vertex count per pass alongside the draw
count. Size names the pass in the game's terms, since a shadow map, a reflection
and the main scene do not share dimensions. Vertices per draw separates a lot of
geometry from a lot of cost per vertex, which is the question the timing cannot
answer and which decides what a fix would even look like.
Rendering at a quarter resolution changed the GPU time not at all, which rules
out fill rate, fragment shading and tile traffic in one measurement, since all
three scale with pixels. What is left inside the passes is geometry, binning and
per-draw cost. It also retires the tile bandwidth theory the previous two
attempts were built on: that traffic would have fallen sixteen fold.
So the draw total needs splitting, and the timer already measures each pass
individually and only reports the sum. Report the distribution instead, keyed by
the pass ordinal within the frame: the frame structure is stable, so pass N is
the same logical pass each time, which is what makes it something to act on.
Count draws per pass alongside it, on the same ordinal. A pass that is expensive
holding few draws is expensive per draw; one holding most of the frame's draws
is carrying the geometry. Same milliseconds, opposite fixes.
Reporting only. No new timestamps and nothing recorded that was not already
being measured.
vkCmdClearAttachments needs the pass open, and the pass opens with LOAD_OP_LOAD,
so clearing a target reads the whole framebuffer into tile memory and then
throws it away. LOAD_OP_CLEAR skips the read. On a tiler that read is the whole
attachment every time, and this title runs about thirty passes a frame at 720p
with colour and depth.
Taken only when the clear covers the entire render area and no pass is already
open. A partial clear is not a load op, and ending an open pass to change its
load ops would store the framebuffer in order to discard it, which costs more
than it saves. Colour is all attachments or none, since a load op applies to the
attachment as a whole. Depth and stencil get separate bits because clearing one
and keeping the other is common.
Whether an open instance can serve a request now compares the key with the clear
bits masked off rather than the pass pointer. Load ops do not affect render pass
compatibility, so the two variants are interchangeable for an open instance and
for the pipelines inside it; comparing pointers would have ended the instance to
begin an equivalent one, paying the store and reload this is meant to avoid and
discarding the clear on the way. Callers that pass no key keep the old pointer
comparison.
The collector had gathered eight frames in five thousand flips and its ring was
parked on slot zero with every slot unreset and empty. All of that follows from
one thing: the frame region was opened at device init and in flush_command_queue
only, and this title takes that path roughly never, so the region opened once at
boot, closed on the first submit and was never opened again.
Everything else depends on it. The slot's query range is reset when the frame
region opens, the ring only advances past a slot once something in it has
completed, and collection refuses a slot that was never reset. So a timer that
initialised cleanly and logged its tick period produced no report for an entire
session, which reads the same as a GPU with nothing to do.
Open it where the primary command buffer is actually begun for the next frame.
The GPU timer initialises, reports its tick period, and then never produces a
report: eighteen RSX profiles came and went in one session against zero GPU
profiles. Collection has several preconditions and the report only prints once
three hundred frames have been gathered, so a collector stuck on any of them
prints nothing at all, which reads exactly like a GPU that is idle.
Log the collector's state periodically while it has nothing, with the slot
flags, the open regions and the drop count, so the precondition that is not
being met can be read instead of guessed at.
Also stop recording anything but the frame region into a slot that still needs
its reset. Writing a timestamp into a range that has not been reset is invalid,
and the reset only happens when the frame region opens, so a render pass that
begins first -- the ones flip() runs after next_frame has already rotated the
slot -- was writing into stale queries.
The GPU timer measures the whole frame, readbacks, blits and uploads, and the
one region it names but never records is draw. So the split it exists to provide
has been missing exactly where it matters: with the RSX thread no longer waiting
on a fence, the Adreno sits at 99% busy at its top clock and nothing says how
much of that is drawing the game.
Bracket the render pass at the only place one actually starts, not at the
wrapper, which early-outs when the same pass and framebuffer are already bound.
Roughly thirty passes a frame, comfortably inside the per-frame event cap.
Both timestamps sit outside the pass rather than inside it. On a tiler the load
at the start and the store at the end are the expensive part, and timing from
within would exclude the cost worth knowing about.
Take the command buffer by const reference, which is what the render pass
helpers hold and what the conversion operator already permits.
Ending the open pass to change an image layout costs a tile store and a reload
on a tiler, and it happens about twenty times a frame out of twenty nine passes.
The counter on it says how many and never which: change_image_layout is reached
from seventy five call sites, tagging them by hand would be tedious and would
still miss the next one added.
Record the return address instead. Two levels, because image::change_layout
funnels most callers and one level would name that function for nearly
everything. Only recorded when a pass is actually open, so the count is
teardowns caused rather than layout changes attempted, and only while profiling
is armed, on a path taken twenty times a frame.
Reported as a symbol where the dynamic table has one and as a module offset
otherwise, which llvm-symbolizer resolves against the unstripped core.
FIFO decode is 56.7% of the RSX thread now that it is no longer waiting on a
fence, and it is the enclosing scope of the dispatch loop, so it holds every
method handler body as well as the loop itself. 201 ns a dispatch is far too
much for reading a word and calling a handler, so the cost is in a handler or in
the per-dispatch machinery, and nothing in the report separates those.
Scope the two batching handlers and the FIFO cache refill. All three run a few
thousand times a frame at most rather than per dispatch, so unlike the earlier
attempt at a per-command scope none of them measures mostly itself.
Count calls, not methods. Both handlers consume a run and skip the rest, so
their share of the method histogram counts what they swallowed rather than how
often they ran, and dividing by it would price a batch as a single method.
Neither vkGetFenceStatus nor vkWaitForFences with a zero timeout returns without
waiting on this driver: both measured 17-28 ms a call and neither returned
not-ready once in 300 frames. So the poll cannot be made honest at the call
site, and the previous commit's zero timeout changed nothing.
What can move is where the wait happens. Draining the present queue at the first
draw of a frame meant the CPU started recording only once the GPU had finished,
and frame time became GPU plus CPU rather than the larger of the two: 42 ms made
of 27.7 GPU and 14.3 CPU, which is the two of them end to end. Stop draining
there and let the throttle at flip bound the pipeline, which is the same wait
placed after the frame's recording rather than in front of it, so recording runs
while the GPU is still busy.
The rotated context should already be retired by then. If it is not, borrow the
aux context as before, and if that is busy too, wait for this one specifically
rather than trip the ensure behind it.
vkGetFenceStatus measured 19.7 ms per call on Adreno and returned VK_NOT_READY
zero times in 300 frames. Every caller of poke() wants "is it done, do not
wait", so the one call per frame turned an intended poll into a full GPU sync
and ran the CPU and the GPU in series: frame time was CPU plus GPU rather than
the larger of the two, with the GPU only 66% busy at less than its top clock.
Use vkWaitForFences with a zero timeout, which is specified to answer without
waiting.
That poll was also the only thing bounding the pipeline, because a queue whose
oldest entry is always retired before the next is added never holds more than
one. With an honest answer the frames accumulate, so bound them on purpose, one
below the frame context count: the queue and the context rotation advance
together, so retiring the front is what frees the context about to be handed
out. Allowing the full count would route every frame through the single aux
context borrow and hit the ensure behind it.
The remaining wait is real frame pacing and is charged to swap_wait, where it
can be read.
The fallback CPU table was missing cores found in recent handheld SoCs
(Cortex-A510, A715, X3, A520, A720, X4). Because get_cpu_name() bails out
when any detected MIDR is unknown, a single missing core sent the whole
lookup to the cortex-a78 fallback whenever LLVM host detection returned
"generic".
Display names were also reused as LLVM -mcpu values, which happens to work
for the Cortex names but not for Qualcomm Oryon: the display name lowercased
to "x-elite", which is not an LLVM processor, so the JIT silently lost
per-CPU scheduling. Entries now carry an explicit canonical LLVM name
alongside the human-readable one; get_cpu_brand() keeps using the latter.
The Qualcomm entry is named "Oryon" rather than "X-Elite" because MIDR
0x51/0x001 only identifies an Oryon core, not the SoC it sits in.
MIDRs cannot identify the SoC at all, which made bug reports ambiguous.
Android's own SOC_MANUFACTURER/SOC_MODEL are now passed to the core and
logged as a separate "SoC:" line, so SoC identity, core topology and the
resolved LLVM target are three distinct values. The LLVM target reported by
system info now comes from the same resolution path the JIT uses, rather
than from the fallback alone, so it no longer disagrees with the target
actually compiled for.
The Vulkan renderer logs one verdict for the adapter it selected, recording
whether BC1-BC3 support keeps DXT textures compressed or whether they are
decoded on the CPU. It sits in render_device::create rather than where the
flag is resolved, because physical_device::create runs for every GPU of
every instance, and not in TextureUtils, whose fallback branches run per
texture and per mip level.
SoC information travels through a new optional _rpcsx_setSocInfo export
instead of an added _rpcsx_initialize parameter. The core is dlopen()ed and
can be updated independently of the JNI glue, so changing an existing
export's signature would make older glue call it with a garbage argument.
Older glue simply never calls the setter, and newer glue null-checks the
symbol against older cores.
No JIT feature policy and no texture decoding behaviour changed.
Verified on an AYN Odin 3 (ayn CQ8725S, 8x Oryon, Adreno 830) running
Turnip 26.2.99: SoC line reads "ayn CQ8725S (Snapdragon 8 Elite-class)",
the brand line reports Oryon rather than X-Elite, the JIT resolves to
oryon-1, and a single BC verdict reports the GPU path. That BC result
applies to the Turnip driver tested; stock-driver behaviour is unmeasured.
Resource destruction was the stated suspect and measured 0.056 ms, so the time
is in vkGetFenceStatus, which is over half the RSX thread. Every call site that
reaches it appears to run once or twice a frame, and a status query that blocks
for milliseconds would be a driver problem while one called a hundred thousand
times would be ours. Nothing in the report distinguishes those.
Count the calls and how many come back not ready. This is the same denominator
the FIFO buckets needed twice already, once for packets against commands and
once for draws against setup.
The present check kept its 18.9 ms after the fence wait and the reclaim both
measured zero, which leaves the poke, and the only thing in a poke that can
sleep is the event completion callback. Without multithreaded RSX that callback
runs inline, and popping an event scope runs the destructor for every GPU object
that event retired, so the RSX thread frees a frame's worth of images and memory
through the kernel driver before it can record the next draw.
Scope the destruction and the fence status query separately. If neither holds
the time, what is left is the lock at the top of the poke, and that is a
different bug again.
The mid-draw present check turned out to be two thirds of the RSX thread, which
the previous commit could only say as one number. It covers three things that
mean opposite things: a fence wait, a poke that takes a shared lock, and the
per-frame resource reclaim. A wait says the GPU is the bottleneck and every CPU
change aimed at this path was aimed at nothing; the reclaim says the opposite.
Scope the fence wait and the reclaim separately, so whatever is left in the
present check bucket is the poke and its lock. Count entries and cleanups too:
the block is written as a rare async flip fixup, so how often it runs is the
first thing worth knowing about it.
Draw setup was 66.7% of the RSX thread, but the bucket only ever held whatever
VKGSRender::begin and end did not charge to a nested scope. Everything with a
body of its own now carries one: the surface write barriers, the render target
on_write pass, the temporary texture release, the mid-draw present check, and
rsx::thread's own prologue and epilogue, which were unscoped on both backends.
Count draws too. A large per-draw bucket is a lot of draws at a fair price or a
few at an unfair one, and those want opposite fixes; the FIFO buckets already
learned that lesson the hard way when a per-packet figure was read per command.
Draw setup keeps the leftovers, which is now the draw clause loop and nothing
that can hide 23 ms.
The two-point probe read the incoming words as be_t<u64>, an eight byte swap,
and compared that against a destination written by copy_data_swap_u32, which
swaps each word on its own. The wide swap also exchanges the two words, so the
comparison was (w0,w1) against (w1,w0) and could only match when w0 equalled w1.
It never reported a match.
Every upload therefore set vertex_program_ucode_dirty. That forces a full vertex
program re-analysis per draw clause, drops the program cache hint, nulls the
bound program so load_program runs again, and re-uploads the transform constants
unconditionally. Sonic '06 issues 8088 of these a frame against 3429 draws, and
a corrected profile puts 24.7ms of a 36.4ms frame in draw setup, which is what
all of that lands in.
Rotating the source back by 32 bits puts both sides in the same word order. The
change can only remove spurious invalidations: a clean verdict from the probe is
still confirmed word for word by the full compare below it, so a false clean is
not reachable.
Upstream inherited, introduced in ae39c5b8cb.
fifo_decode is the enclosing scope of the whole RSX loop, and the profiler is
exclusive, so it holds whatever no nested scope claimed. VKGSRender::begin()
carries a scope and end() did not, so essentially every per-draw cost landed
there: load_texture_env's texture cache search and sampler lookup, the vertex
and fragment ucode analysis, and the write barriers. shader_translate and
barrier had no instrumentation sites anywhere in the tree.
That is why the bucket read as 100% of a 29ms frame while the decode loop itself
only accounts for a couple of milliseconds: 36728 packets and 48737 dispatches
cannot cost 29ms when the loop body is an inlined exchange, a table load and an
indirect call.
Scopes added to end(), load_texture_env and analyse_current_rsx_pipeline, so the
next capture shows where the frame actually goes.
Three changes to the same loop, all measured against 36728 packets and 48737
dispatches per frame in Sonic '06.
GET is published on a bounded lag rather than every packet. It is a release
store into guest DMA memory, and get shares a 64-byte line with put which the
guest PPU writes from another CPU cluster, so each publish was a cross-cluster
coherence miss. The guest reads GET to size its free ring space and is far ahead
of us here, FIFO stalls measuring 0.1 a frame, so lag is invisible to it.
Anything that can idle or block publishes immediately: the put wait in inc_get,
the NOP path, and set_get.
The FIFO accuracy setting is snapshotted once per packet instead of being read
per argument. Reading it goes through a seq_cst atomic load, which on ARM64 is
an ldar the compiler cannot hoist out of the loop.
The again poll is relaxed. That flag is only ever set by this thread, by the
handler invoked immediately before, so sequential consistency buys nothing and
cost another ldar per dispatch.
Three separate fixes to the same hot path.
The profiler's FIFO figure counted the wrong thing. g_fifo_commands is
incremented once per run_FIFO entry, and one of those drains a whole packet, so
dividing by it priced a packet rather than a method. Sonic '06 averages about 17
methods per packet, so the reported 612 ns per command was really 612 ns per
packet and the per-method cost was closer to 36 ns. Cross-checks: 413.6 FIFO
refills a frame at 4096 bytes is 1.69 MB, about 423000 words, which 24248
packets can only consume at roughly 17 words each. Dispatches are now counted
where they are dispatched, both figures are reported, and the per-method
histogram divides by the right one. The line is labelled packets, and notes that
fifo_decode is a catch-all holding every handler body too, since no handler
carries its own scope.
rsx_state::decode was a cross-TU call per dispatched method. The body is one
exchange, but the definition lived in rsx_methods.cpp and LTO is disabled
project-wide, so it never inlined, and being opaque it also forced the caller to
reload its context pointer afterwards. Moved to the header.
set_transform_constant and set_transform_program read ctrl->put through a
seq_cst load. That is an ldar on ARM64, on a cache line shared with ctrl->get
which the guest PPU writes from another cluster, in the two hottest handlers in
this title. Relaxed: a stale value only shrinks the batch, and the remainder is
picked up on the next call.
copy_data_swap_u32 and its compare variant are assembled by asmjit under
ARCH_X64 only. Every ARM64 build fell through to the scalar per-word loop,
reached through a function pointer so it could not be inlined, and LTO is
disabled project-wide so nothing recovered it afterwards.
It is not a cold path: transform constants, transform programs and vertex data
all upload through it, and a Sonic '06 profile on a Snapdragon 8 Gen 2 put 82.5%
of the RSX thread in FIFO decode with thousands of these blocks per frame.
vrev32q_u8 reverses bytes within each 32-bit lane, which is the same swap the
scalar path does per element. The compare variant accumulates differences and
reduces once at the end rather than branching per element. Checked against the
scalar version over 20000 randomised trials at counts 0 to 39, covering every
tail remainder, for both variants: identical output and identical return value.
transform_constant_load_modifier_barrier decoded its argument into
NV4097_SET_TRANSFORM_PROGRAM_LOAD. The barrier is pushed by
nv4097::set_transform_constant_load, so it should target
NV4097_SET_TRANSFORM_CONSTANT_LOAD.
A title that moves its constant load pointer mid-draw therefore had the move
dropped, leaving every constant after it written at the old offset, and had its
vertex program upload position overwritten with a constant index at the same
time.
An unbounded wait polled vkGetFenceStatus in a tight loop with nothing but a
pause hint between calls. command_buffer::flush() takes that path for the submit
fence, so it is what a frame does while it waits on the GPU: a core pinned at
100% for the whole wait, hammering a driver entry point while the driver is
trying to do the work being waited on.
Cheap on a desktop with cores to spare. Not here, where it competes with the SPU
and PPU threads for a handful of cores. Arkham City measured 24ms of a 53ms
frame in this function with the GPU only 71-77% busy, which is what a stall
looks like when the waiter is too busy spinning to prepare the next submission.
Polls briefly first, since most waits are for a fence about to signal and
blocking would cost a syscall and a wake-up for nothing, then hands the wait to
vkWaitForFences so the driver can sleep the thread. The blocking call was
already there, three lines up, used only when a finite timeout was supplied.
Same shape as wait_for_event below, which already had this treatment.
Installing a licence is a copy into exdata, a directory nothing on the screen
read, so a success looked exactly like a failure: a licence belongs to no title,
never appears under Installed titles, and left nothing visible anywhere in the
app. Reported as the .rap doing nothing, when the file had in fact been written
correctly.
Listed now beside the installed titles, refreshed on the same events.
Every directory under dev_hdd0/game was listed with an Uninstall button beside
it. RPCS3 keeps its own lock directory in there, get_hdd0_locks_dir() being
get_hdd0_game_dir() + "$locks/", so the screen offered to delete the emulator's
lock state, and any folder a failed install left behind was offered as a title.
A PARAM.SFO is the test now. Game data installs keep theirs and stay listed on
purpose: a 1.1GB BLUS30464_INSTALL is the kind of thing someone opens this
screen to reclaim, bootable or not.
sys_fs_readdir fires once per directory ENTRY, unlike opendir and closedir
either side of it which fire once per operation. A game scanning its own USRDIR
emits a line per file: one such scan measured 346 lines in 22ms, during boot,
for no diagnostic gain.
Moved to trace, which is where the other per-datum calls already sit
(sys_fs_read, sys_fs_write). opendir and closedir stay at warning, so a scan is
still visible in the log without being enumerated.
The instance asked for 1.2 unconditionally. A loader that predates it may answer
VK_ERROR_INCOMPATIBLE_DRIVER to a higher request, and the spec tells applications
to check the version first for that reason, so on those devices the renderer
never started at all.
Queried through the global procedure address, since vkEnumerateInstanceVersion is
itself a 1.1 entry point and its absence means 1.0, then clamped. A no-op wherever
1.2 or better is available, and logged when it is not so a device report says so.
Every JNI object handed to native code is a local reference, reclaimed only when
the frame that created it returns to Java. The frames this runs on do not return:
the main thread processor and the compilation queue are infinite loops inside a
single JNI call.
Progress took one reference per instance from FindClass and one per report() from
NewStringUTF, and released neither. There is no DeleteLocalRef, PushLocalFrame or
NewGlobalRef anywhere in the native tree. The progress dialog server pushes
several updates per tick and a firmware precompile emits thousands of ticks, so
ART's local reference table filled and the runtime aborted.
Time-proportional, which is why it showed as a crash during firmware install on
slower devices and not on faster ones.
Copy construction is deleted along with it: the class owns a reference now, and a
copy would have had its destructor release one the original still used.
The Android port keeps one global config.yml: settingsSet persists through
SaveSettings(g_cfg.to_string(), "") and an empty title id is the global path.
apply() returned early for any title without an entry, writing nothing, so
Uncharted 3's Stub PPU Traps = 1 stayed set once it had been booted and every
game launched afterwards ran with a PPU that silently skips an instruction on
any trap rather than stopping.
Nothing said so on screen and nothing else writes that node: it is not in the
curated push, and CoreSettingOverrides only replays paths the user recorded
themselves.
Every managed path is now written on every boot, this title's value where it has
one and the upstream stock value where it does not. Anything added to BY_SERIAL
has to gain its default in STOCK.
The recursive scan descends into anything that is not itself a game folder and
then accepts any file whose extension is in gameExtensions. "img" is one of
them and a title's own data is full of them, so once a package unpacked over
dev_hdd0/game the library filled with GTA IV's archives: manhat01, props_ab,
vehicles, script, weapons.
dev_hdd0/game is the emulator's own install root and holds one directory per
title, so it is now read that way. A direct child is a title or it is not
listed. Folders a user pointed us at keep the recursive scan, because games
legitimately sit at any depth there.
The extractor no longer unpacks into this directory either, but the library
should not have depended on that, and existing installs still have the debris.
A package's install directory is taken straight from its own metadata and was
never checked. Both sources can produce nothing: read_metadata sizes the string
to 9 and reads the title ID over it without testing the result, so a short read
leaves nine NUL bytes, and the DLC path takes c_str() + 8, which is empty
whenever byte 8 is a NUL.
Appending either left the destination as dev_hdd0/game itself, so the package
unpacked its contents over the games root. Users reported a library full of
asset directories, storage consumed with nothing listed as installed, and
folders that outlived uninstalling the title, because uninstall only removes
dev_hdd0/game/<TITLEID>. It looked random because it depends on the individual
package's metadata.
Checked against c_str() so the nine-NUL case reads as empty. Separators and dot
entries are refused as well: this is one path component chosen by the package
and it has no business pointing anywhere else.
While chasing an unrelated SPURS hang I noticed the JIT publishes
freshly written code on ARM64 with no instruction-cache maintenance at
all. A grep for clear_cache or flushInstructionCache over the JIT layer
comes back empty. The branch-rewrite sites only issue ISB; DSB ISH,
which performs no D-cache clean or I-cache invalidation and is ordered
backwards for self-modifying code besides. On ARMv8 a correct
publication needs the DC CVAU / IC IVAU broadcast sequence; x86 has a
coherent instruction cache, so none of this was ever visible there.
All sites use the bundled asmjit::VirtMem::flushInstructionCache(),
which emits that sequence portably across toolchains.
This covers every publication path I could find:
- MemoryManager1::finalizeMemory() and MemoryManager2::finalizeMemory()
were both no-ops. RuntimeDyld calls finalizeMemory() after writing
code and relies on it for cache maintenance, so LLVM emitted PPU and
SPU code was never flushed. MemoryManager1 serves the primary PPU
JIT, MemoryManager2 the SPU JIT and auxiliary engines. Both managers
now record code section allocations and flush them on finalize. I
confirmed at runtime that the MemoryManager2 path executes (about
12800 calls per cold boot).
- jit_runtime_base::_add() copies asmjit output into executable memory
with no flush.
- jit_runtime::finalize() restores an executable code snapshot in place
during emulator restart with only the ISB/DSB pair.
- spu_runtime::rebuild_ubertrampoline() publishes a hand-written
trampoline via CAS with no flush; the flush now happens before the
publication.
- spu_runtime::make_branch_patchpoint() writes a patchpoint byte by
byte and returns it with only the ISB/DSB pair.
- Both 16-byte branch-site rewrites (dispatch and branch) atomically
overwrite live code and only issued the ISB/DSB pair.
The ISB/DSB pairs adjacent to the new flushes are removed along with
their misleading "flush all cache lines" comments: the flush helper
already issues the trailing barriers, and the pairs never performed
any cache maintenance in the first place.
I want to be upfront that this was not the cause of the hang I was
debugging (a same-item compilation race, fixed separately), and I have
not observed a failure that this change alone fixes. It is a latent
correctness issue on any ARM64 host: nothing prevents another core
from fetching stale instruction bytes for freshly published code.
Five SPURS kernel threads executing the same uncached code at the same
address all reached spu_llvm_recompiler::compile() for one spu_item.
add_empty() returns the existing item for an identical program without
telling the caller it did not insert, and the entry-point dedup only
catches equivalent code at a different address. Each thread then
compiled the program with its own LLVM instance, racing the compiled
pointer publication, the ubertrampoline rebuild, and the waiter
notification.
On my 8-core ARM64 device this wedged SPURS bring-up on every single
cold-cache boot of Virtua Tennis 4 (BLUS30529). The kernels ended up
parked polling zeroed workload state, the PPU main thread blocked
forever in sys_event_queue_receive on a queue no SPU would ever signal,
and the title never reached the menu. Warm boots never hit it because
cached programs are compiled before SPU execution starts, one
presentation per program. When I instrumented the compile path I saw up
to five concurrent compilations of a single item, around 790 collision
events per boot, with duplicates accounting for roughly two thirds of
all cold compilation work.
This change gives spu_item an explicit LLVM compilation state
(unclaimed, compiling, complete, failed). The first compiler claims the
item; later arrivals wait and take the published result, mirroring the
existing dedup-wait path. I made the claim a state on the item rather
than an inserted-flag from add_empty() so that an item pre-inserted by
spu_fast is still claimed by the first LLVM worker, which preserves the
asynchronous optimized replacement on x86-64. A scope guard marks the
item failed on any early exit so waiters cannot be stranded.
The pre-existing wait for relocated duplicates (same program, different
entry point) is also covered: it still waits on the compiled pointer,
because that result can be published by spu_fast from the asmjit path
which never touches the LLVM state, but it now observes the failure
state on each wakeup with a bounded timeout, and the failure guard
wakes those waiters too. Without this an owner that bailed out early
would have stranded them forever.
I verified this on device: 11 out of 11 cold boots stalled before the
change, 7 out of 7 pass after it, plus 2 out of 2 warm controls, and
Mirror's Edge now reaches gameplay past its previous SPURS stall.
Cold-boot SPU compilation dropped from about 12900 blocks to 3900.
Every .rap failed. A .rap is 16 raw bytes of key and carries nothing that says
which content it unlocks; that lives in the filename, as the content id. Upstream
copies the file into dev_hdd0/home/<usr>/exdata/ under its own name and is done,
so the name is the whole mechanism.
This called the native installKey with an empty game path instead. That path
decrypts the GAME's EBOOT to read an NPDRM header out of it, so with no game path
there was no EBOOT, no header, and the install could never succeed. Handing it a
bare file descriptor had already thrown the name away regardless.
Licences are now handled before any descriptor is opened. Reported for Resident
Evil 4 HD and for DLC licences generally.
SONIC THE HEDGEHOG (2006) renders only its HUD and skybox and flickers
everything else in and out of existence. That is upstream RPCS3 issue
#4122, open since 2018: a PPU/SPU race where an SNR is overwritten while
still non-empty, so the SPU jobs feeding geometry lose their signal.
Upstream never fixed it in code. The fix is elad335's canary patch, which
hooks RPCS3_HLE_LIBRARY:WaitForSPUsToEmptySNRs -- a function that already
exists in the core and has no other user. The patch is not in the official
feed: neither our stored database nor a fresh pull from rpcs3.net carries
a single BLUS30008 entry, jumpf op, or RPCS3_HLE_LIBRARY reference. Desktop
users add it by hand through the Patch Manager's import button, which has
no equivalent here, so on Android the game was simply broken.
Ship it as an asset and merge it into patch.yml at boot, before the core
reads that file and once the config directory is known to exist. Enable it
rather than only listing it: a user who has to find and tick a box before
the game renders has already decided the emulator is broken. Gated on a
stored revision, not run every boot, so turning it off sticks.
Everything ARM64 needs is already in this tree -- the calloc code-cave
registry, the is_faux_function guard that keeps the tail-call return trap
from firing on a patch-point, and the PPU LLVM filter for patched
functions -- so the patch applies here rather than crashing as it once did
on ARM.
Keyed by PPU hash, so it covers only the dump it was built for; other
regions and revisions need their own entry.
get_per_core_usage() fills the per-core vector with zeros and then wraps
the whole Linux /proc/stat body in #ifndef ANDROID, because an app cannot
read the per-cpu lines. The exclusion left the zeros in place, so the
monitor did not go quiet -- it reported an idle machine. On device it
logged "CPU Usage: Total: 0.0%, Cores: 0.0%, 0.0%, ..." while three
emulator threads were pegged at 100%, which hides exactly the class of
problem the monitor exists to surface.
Report process usage from times(), which is POSIX and readable by our own
process, and leave the per-core vector empty so perf_monitor prints no
"Cores:" list rather than a fabricated one.
When the transfer queue was drained the offload thread called
std::this_thread::yield() in a tight loop, which is a sched_yield()
spin rather than a wait. Measured on an Adreno 740 handheld it burned
1142s of CPU over a 1145s session, 85% of it system time, for no work
at all -- a whole core taken from the RSX and SPU threads that need it,
plus the battery and thermal budget that goes with it.
Wait on the queue instead. lf_queue::push notifies on the empty to
non-empty transition so a real job still wakes the thread immediately
and transfer latency is unchanged, and if a push lands between the
emptiness check and the wait the atomic is already non-zero so there is
no lost wakeup. The timeout is only there so the loop can re-check
thread_ctrl::state(): aborting pushes nothing, so an untimed wait would
never return and shutdown would hang instead of spin.
dma_manager::sync() is unaffected -- it only waits while work is
outstanding, which is exactly when the thread is not idle.
Ours was false against upstream's true, so the curated push wrote false over the
node on every boot and every install ran with an accuracy fixup disabled that
RPCS3 ships on. Games that depend on it get NaNs through vector maths, which
shows up as geometry behaving impossibly rather than as any kind of error.
The disagreement was not free either. The node feeds ppu_settings::fixup_vnan,
which is part of the compiled PPU object filename, so flipping it renames every
object and the next boot recompiles all of them. A device here carries both
variants, 76 modules under one key and 58 under the other, from one flip.
Audited the other two settings in that key: Accurate Cache Line Stores and Use
Accurate DFMA both already match upstream. This was the only one adrift.
All three now say that changing them rebuilds compiled PPU code, which is the
part that had no way of being known from the screen.
The unresponsive-stop callback reported that a stop was hanging but not what was
holding it, and that is the one thing that cannot be recovered afterwards: from
outside, a wedged shutdown looks like a sleeping thread with no CPU time and no
reason attached.
Enumerate the SPU threads still registered and log their cpu_flag state. Whether
exit was ever delivered, whether the thread is parked in a wait, or whether it
never left its initial stopped state are three different faults with three
different fixes, and the flags separate them.
This is what exposed the failed savestate: the thread that would not stop was
debris from a save that had already been abandoned, not a shutdown bug.
It was turned on so savestates could save at all, and savestates are no longer a
feature: a PS3 state runs 500MB to 3GB, which fills a phone in a handful of
saves. What is left is the cost, and it is a real one, which is why upstream
defaults it off.
Written as false rather than dropped from the push. It shipped as true for a
while, so installs from that window have it persisted in config.yml and would go
on paying SPU performance for something nothing uses.
Also drops the thumbnail read logging, which did its job: it proved the read
path was never called, because the in-game menu has its own numbered slot list
and only the touch overlay's picker asks for a preview.
Savestates could not work without it. Saving locks every SPU thread into a state
it can be serialised from, and with this off that lock fails on any title with
SPU work running. The save is abandoned with "failed to lock SPU threads
execution", and the SPU it gave up on then never answers the stop request, so
the join thread spins and the app hangs on the next save or load.
The deadlock chased through the SPU shutdown path was this setting all along.
The state on disk looked valid because a previous attempt had written one.
Upstream defaults it off for SPU performance, which suits a desktop where
savestates are optional. Here they are on the in-game menu and the touch
overlay, and a save that wedges the emulator costs more than slightly slower
SPU emulation.
RPCS3's stop watchdog calls on_emulation_stop_no_response once shutting the VM
down has taken about ten seconds. That callback was a no-op here, so the one
thing upstream does about a deadlocked stop, say so, did not happen: the join
thread went on spinning, the overlay sat at its last figure, and the app looked
frozen with nothing in the log to explain it.
Reported rather than aborted. The desktop build offers to terminate but asks
first, and slow is not the same as stuck: a savestate write was measured here
taking 71 seconds legitimately, so killing the process on a timer would trade a
hang for a corrupted save.
The core already renders the finished frame and hands it to take_screenshot
whenever a screenshot is asked for. That override was empty, so the picture was
built and thrown away and every slot tile had nothing to draw.
Cached rather than captured on demand: a save kills the VM, and the point where
the state is known to be on disk is after_kill_callback, by which time there is
no renderer left to ask. So the save requests a frame on its way in, the
override catches whatever the renderer produces, and the capture step writes it
out once the state has landed.
Stored downscaled to a ~320px edge as AX3T + width + height + RGBA, and
re-encoded to PNG on the Kotlin side, which is where a Bitmap can be built from
raw pixels directly. That keeps a compressor out of the core for something only
ever drawn as a tile. A slot with no picture still loads.
Loading a slot closed the game and dropped to the library. BootGame failed in
the same millisecond it was called, without ever reading the file, and the
Emulation Join Thread warnings then ran for seconds after: the previous VM was
still coming down while BootGame was already being asked to bring the next one
up, so it bailed.
boot_current_game_savestate gets away with that sequence because it runs from a
shortcut handler rather than from inside the emulation callback queue. Kill
first and boot from after_kill_callback, which is the shape the save path uses
and the reason that one works.
The boot result is logged too. A bare failure could not tell a corrupt state
from a version mismatch from a path the loader never accepted.
Every slot tile read as empty and Load was disabled on all ten, whatever was on
disk. Occupancy comes from getGamePathSlot, which was a stub returning the empty
string, so saving had been working with no way to see that it had.
Answer it from hasState, which was written for this and had no callers. The
subtitle is the title id rather than the state's path: the picker strips to the
last segment and drops the extension, which would render a real path as
slot0.SAVESTAT.
RetroArch shaders shifted the colours, pink over anything green. The output
image handed to librashader is the swapchain image but the format reported with
it was the source framebuffer's. librashader builds its output view and render
pass from that, so with a BGRA swapchain and an RGBA source it wrote through a
mismatched view and swapped red with blue instead of converting.
Only the shader path was affected because the bilinear and nearest passes reach
the swapchain through vkCmdBlitImage, which converts formats itself. Carry the
swapchain format through to the pass, refreshed per frame so a swapchain rebuilt
at a new format cannot leave a stale value behind.
RPCS3 keeps savestates as a rolling history addressed by age, 1 being the most
recent, and the app offers ten numbered slots. The slot was dropped on save and
read as an age on load, so saving to slot 3 pushed a new newest state and
loading slot 3 fetched the fourth-newest. States appeared to wander between
slots on their own.
The history stays RPCS3's. A slot is now a copy parked under
savestates/<title>/armsx3_slots/, in its own directory because get_savestate_file
derives the next auto id by listing the title's directory and must not be fed
names it never generated. Copy rather than move: the restart after a save boots
from the file the core just wrote.
At 0 the four direction keys meet in the middle, which only reads correctly on
the flat built-in drawables. The bundled skins draw four separate keys and they
came out as one clumped blob.
Migrated once for existing installs: the key is written by the bulk save, so
anyone who had ever opened a touch setting already had 0 stored and would never
have picked up a new default. Only a stored 0 is rewritten.
The button was always there. The LazyColumn above it had no weight, so in a
Column it took the whole remaining height and left nothing for the row below,
putting Back off the bottom of the screen. With a controller you could still
leave; on a touch-only device this was the one screen with no visible way out.
Weight the list so it shares the space instead of consuming it.
No preview rendered for any downloadable skin. The published manifest carries a
trailing comma before two of its closing braces, which org.json rejects, so
parseManifest returned null and fetch fell through to the git tree, which only
ever reported filenames and set previewPath to null for everything. One comma in
a file this app does not own cost the whole index.
Retry the parse with trailing commas dropped, string aware so a name containing
a comma survives, and pair Previews images to skins by name in the tree fallback
so a broken manifest costs detail rather than every thumbnail.
Changing DEFAULT_SKIN_ID moved nobody. resolveRaw only falls back to the default
when nothing was ever stored, and every install from before the change has a
stored value already, so existing users kept the old pad and reported the new
default as not applying.
Rewrite the stored global skin once, and only when it is absent or the explicit
no-skin sentinel. An install pointing at a real skin chose it on purpose. The
flag is set either way so picking the ARMSX2 row afterwards sticks.
The second row read 'Built-in (default)', which was accurate when the built-in
drawables were the default and wrong the moment a bundled pack took over: it
still claimed to be the default while ARMSX3 Textured actually was. Name it
'ARMSX2' for what it is and tag whichever row matches DEFAULT_SKIN_ID, so the
label follows the constant instead of restating it.
Adds two packs from bagas as app assets, ARMSX3 Textured and ARMSX1, and makes
ARMSX3 Textured what a fresh install gets. The order offered is ARMSX3 Textured,
ARMSX2, ARMSX1, NetherSX2, NetherSX2 Old. Their files arrive named plainly
(L1.png, cross.png), which the loader already handles: it strips ic_controller_
and a trailing _button, so they are renamed to the canonical form on the way in
and both packs land on exactly the key set nethersx2 uses.
The default needed untangling first. Null meant the built-in ARMSX2 drawables
AND was what an install with no stored choice got, so the two were the same
thing and nothing had to tell them apart. They are different now. Resolution
moved into one function both callers use, because ensureLoaded and
applyForSerial each resolved separately and disagreeing would have changed the
pad the first time a game started:
nothing stored the default, ARMSX3 Textured
NONE the ARMSX2 row was chosen, so the drawables
an id that skin, or the default if it names one no longer installed
setActive stores NONE at the global tier rather than removing the key. Removing
it used to mean the drawables; with a bundled default it would now mean "never
chose", so picking ARMSX2 would have silently handed back ARMSX3 Textured on the
next launch. Nobody's existing choice moves: only the absence of a stored value
resolves to the default, and that is not written back, so the default can change
again later without pinning every install to today's answer.
Every built-in now shows a preview strip beside its name, matching what the
downloader already does for remote skins. Baked in as preview.png per pack
rather than composed at runtime: the list draws several at once and compositing
ten bitmaps per row per recomposition is not worth it for a picture that never
changes. keyForFilename ignores the file, so it is not counted as a button. The
ARMSX2 look has no asset folder, so its strip is a drawable.
ARMSX3 Dark was offered and is deliberately not included. It ships 17 images
against the other packs' 18, with no start button.
Verified in the APK rather than assumed: four asset packs present, 18 images
each in the two new ones, four asset previews plus the drawable.
It was reachable only from the library drawer, which is a global screen. That
put the moment a core node is actually worth touching, a game that needs one
while that game is loaded, in the one place it could not be set without setting
it for every other game too.
Add it to the Options pane beside All Settings, opening over the paused game
like the other manager screens. Scope and serial come from the overlay's own
scope state, resolved for the running title when the menu opened, so an edit
made mid-session is remembered for that title alone.
The core settings screen recorded every edit into a single global store, so a
node set to get one title running was then set for every title that booted
afterwards, with nothing on screen to say so.
Give that store the two tiers ConfigStore already gives curated settings: a
global set, and a per-game set keyed on the same settingsKey, so both stores
agree on which title is being configured. Replay pushes global first and the
running title's set on top, so a per-game value wins where the two disagree and
every node a title never touched still follows global. With no game loaded the
per-game tier is skipped rather than guessed at, so the last title played cannot
leak into a BIOS boot.
The screen now takes its scope and serial from the caller instead of resolving
them itself: the drawer route is global and would otherwise inherit whatever
scope the last per-game settings visit left behind. It also states which tier
the next edit lands in, since the screen looks identical either way.
Ports ARMSX2's GitHub-release updater. The UI hooks and all seventeen update.*
strings were already here from the UI port; only the implementation was missing,
so both hooks were sitting behind IN_APP_UPDATER doing nothing.
A "Check for updates" panel at the top of the App tab queries the ARMSX3
releases API, semver-compares the tag against the installed build, downloads the
.apk asset with a progress bar into externalCacheDir/updates/, and hands it to
the system package installer through a FileProvider. The user confirms the
install; nothing happens silently. Two opt-in toggles come with it, both default
off: check on launch, which pops a prompt only when something newer exists, and
include nightly builds.
Points at ARMSX2/ARMSX3 releases, not ARMSX2's. Download goes to the app cache
directory, which the OS can evict, and each download clears the folder first, so
it never accumulates.
ONE DELIBERATE DIFFERENCE FROM ARMSX2, and it needs to be understood before any
Play build exists. ARMSX2 keeps this out of its bundle with a src/github versus
src/play flavor split plus a build script that fails closed if
REQUEST_INSTALL_PACKAGES appears in the bundle manifest. ARMSX3 has neither a
Play build nor flavors, so the code, the permission and the provider live in
src/main behind the runtime flag. That is fine today and NOT fine the moment a
Play target appears: Play rejects the permission in the bundle, and a runtime
flag does not remove a permission. Adding a Play build means doing the flavor
split first. The requirement is written at the manifest, at the buildConfigField
and in the file header, because one comment is easy to miss.
Verified in the produced APK rather than assumed: the permission is present, the
provider is registered as com.armsx3.updateprovider, and UpdaterEntry is in
classes18.dex.
Demon's Souls precompiles for 27 minutes and was killed partway through. The
trajectory says why serialising alone could not save it: the process sat between
4.3GB and 5.8GB for the whole run on a 7GB device, so this was sustained
footprint rather than a transient overlap the existing mutex could flatten. It
died when the system wanted memory back, and Zygote logged signal 9 for four
processes at once, so the pressure was not ours alone.
Adds a second tier. Below 2GB free, workers already compile one at a time; below
1GB free, the worker holding that lock now waits for the system to recover before
starting the next module, up to ten seconds, rechecking every 100ms.
The wait happens AFTER taking the serialisation lock deliberately. Waiting first
would have the other worker still allocating, so the wait would be watching
memory it is not allowed to influence.
Safe to wait here because precompilation writes each object to disk and links
none of them: the modules are not mapped into the VM, so is_being_used_in_emulation
is false and no JIT instance is created. Pausing costs time and nothing else.
Bounded rather than indefinite, because failing to compile is worse than
compiling under pressure, and it gives up cleanly if the emulator is stopping.
The readback in VKPresent was already gated on Video@@RSX Profiler, but
gpu_timer::begin and end were not. A release build with the profiler off, which
is the default and what everyone runs, still paid a vkCmdResetQueryPool and two
vkCmdWriteTimestamp calls per region per frame for results nothing would ever
read. Timestamp writes are not free on a tiled GPU; they are pipeline sync
points, which is the opposite of what a diagnostic should cost when disabled.
Both sides are gated now. end() deliberately clears m_open before returning
rather than bailing first: disarming between a region's begin and its end would
otherwise leave the flag set, and every later begin for that region would drop
itself as unbalanced, silently killing that region's timing for the rest of the
session. The timestamp is skipped, the bookkeeping is not.
The pool is still created at device init so arming mid-session works without a
restart. That is one small allocation, not a per-frame cost.
Arkham City runs about 83 render passes a frame for roughly 339 draws, and on a
tiled GPU every pass is a tile store plus a reload of the attachment. Attributing
every end site showed where they come from:
ImgHelper:43 37-48/frame change_image_layout
Barrier:inout 20-23/frame insert_texture_barrier
Draw:1093 13-18/frame subpass mismatch
Barrier:img 3-5/frame
insert_texture_barrier handles the feedback case, an attachment sampled while it
is still bound. It ended the pass because it had no choice: Vulkan forbids
vkCmdPipelineBarrier inside a render pass unless the subpass declares a
dependency on itself, and this render pass cache declared no dependencies at all.
The function already took a preserve_renderpass flag; there was simply no way for
a caller to use it legally.
So the pass now declares a by-region self-dependency, and the feedback barrier
opts in. Three parts that have to agree:
VKRenderPass.cpp declares the self-dependency, framebuffer-local stages
barriers.cpp drops the vertex stage when preserving, since
VK_DEPENDENCY_BY_REGION_BIT permits framebuffer-space
stages only and naming the vertex stage would make the
barrier invalid
VKRenderTargets.cpp passes preserve_renderpass at the cyclic-reference site
Correct for the use: the feedback case is a fragment shader sampling the
attachment its own fragments write. Anything needing vertex-stage visibility
leaves the flag false and still gets the pass ended.
Android only at the call site. The self-dependency itself is declared everywhere,
which is harmless where nothing issues an in-pass barrier.
Does not touch ImgHelper:43, the larger site. Layout transitions of
non-attachment images are illegal inside a pass whatever dependencies exist, so
that one needs resource preparation hoisted ahead of the pass instead.
Arkham City runs about 90 render passes a frame for roughly 339 draws, under four
draws per pass, on a tiled GPU where every pass costs a tile load and store. That
is worth attacking, since Fence wait is 24ms of a 53ms frame and disabling
culling to add GPU work pushed it to 43.8ms, which is what being GPU bound looks
like.
But only about 19 of those 90 ends were attributed: Draw:1093 at 16.1 and
Texture:921 at 2.9. The other 71 happened at sites with no counter, so the
report pointed at the wrong two.
Instruments the rest: the four barrier sites, both texture cache sites, the image
helper, and the occlusion query one in VKGSRender. The barrier sites are the ones
to watch, since every barrier that cannot preserve the pass ends it, and that is
the same mechanism as the Adreno leak fixed earlier: this driver does real work
on every vkCmdEndRenderPass.
Counters only. Which site dominates decides whether the fix is barrier batching,
texture cache scheduling, or something else, and guessing between those has a
poor record here.
A simpleperf profile of the RSX thread during Arkham City gameplay, 83261 samples:
24.91% vk::query_pool_manager::get_query_result
11.63% [kernel]
10.16% rsx::nv406e::semaphore_acquire
9.37% rsx::FIFO::FIFO_control::fetch_u32_refill
3.96% memcpy_opt
3.26% VKGSRender::do_local_task
get_query_result is a quarter of the thread on its own, and what it does is spin:
pause(), re-poke the query, repeat, until the GPU has the result. That is a
defensible trade on a desktop, where the answer lands in microseconds and there
are cores going spare. On a tiled mobile GPU the result is not available until
the tile pass resolves, so the wait is much longer, and this device runs eleven
hot emulator threads across five usable cores. The spin does not make the result
arrive sooner; it just denies the core to an SPU thread that had work.
Keeps a short 64-iteration spin so a nearly-ready result still returns without a
scheduler round trip, then yields. Android only.
Also scopes the wait as fence_wait, which is what it is. This was completely
invisible before: the bucket report attributed 0.072ms/frame to ZCULL and showed
nothing here, because the ZCULL scope covers zcull_ctrl->update and this is
reached by another path. A bucket reading zero means no scope reached it, not
that the work is free.
The per-command figure came back at 1017ns across 43870 commands per frame. That
is not a fair price for reading a word and calling a handler, so the cost is
concentrated in particular handlers rather than spread across the dispatch, and
the useful question is which.
Counts commands per method register and reports the top eight with their share,
named through gcm_printing so they read the same as the log's own FIFO traces.
A handful of methods dominating means a fast path is worth writing for them. An
even spread means the dispatch itself is the problem and this was the wrong tree.
Either way it is the last thing hidden inside fifo_decode, which now holds 44.6ms
of a 59.2ms frame with every sub-unit around it measured and small: ZCULL 0.111ms,
page protect 0.133ms, local tasks 1.357ms, the whole draw path under 6ms.
One increment behind the enabled() branch, no counter-timer read, 64KB of
counters touched only while armed.
Every sub-unit inside the RSX dispatch loop is now measured and every one is
small. ZCULL is 0.083ms/frame, page protection 0.094ms, local tasks 1.691ms,
and the entire draw path under 5ms of a 53.4ms frame. Fence wait takes 9.2ms.
That leaves 34.9ms in fifo_decode with nothing left to attribute it to except
the dispatch itself, and FIFO stalls at 1.0/frame say it is not waiting to be
fed either.
Which leaves one question worth asking: is that a lot of commands at a fair
cost each, or few commands at an unfair one. Those want opposite work. A lot of
commands means the volume is the problem and the answer is upstream of the
loop; an unfair per-command cost means the dispatch is the problem and can be
attacked directly.
So the loop counts its iterations, and the report divides fifo_decode by them.
An increment behind the enabled() branch with no counter-timer read: the
earlier attempt at a per-command SCOPE read cntvct_el0 twice per command and
took run_FIFO from 3% of samples to 35%, which is the mistake this avoids.
Page protection turned out to be 0.090ms/frame, 0.2%, so the 38ms sitting in
fifo_decode is not mprotect and not the drawing either: the entire draw path
comes to under 5ms of a 54ms frame, and fence wait accounts for 8.4ms more.
Roughly 38ms had no owner.
fifo_decode is scoped around the whole RSX loop rather than around run_FIFO,
deliberately, because a scope inside run_FIFO reads the counter-timer twice per
FIFO command and previously turned a 3% bucket into 35% of samples. So anything
in that loop without a scope of its own accumulates there, and the loop's
per-64-cycle sub-unit updates are the largest unmeasured things left in it.
Both now have buckets. They run once per 64 commands, so the counter read is
amortised and the earlier distortion does not apply.
ZCULL is the specific suspect: this title logs "Reports area at location
CELL_GCM_LOCATION_MAIN was accessed. ZCULL optimizations will be disabled" and
then runs the unoptimised path for the whole session, with Accurate ZCULL stats
on, while the game polls occlusion reports.
Measurement only, no behaviour change.
Arkham City under load spends 36.9ms of a 46.5ms frame in fifo_decode, which is the
bucket the FIFO loop leaves active and therefore holds everything without a scope of
its own. The drawing is not the cost: draw_setup, vertex, pipeline, descriptors,
texture upload, RT prep, blit and submit together come to under 5ms. Fence wait falls
to 5.4% under load, so it is not waiting on the GPU either.
The one number in the report large enough to explain the hole is page protection:
16.8 mprotect calls covering 40MB every frame. That is on the order of ten thousand
pages of kernel page-table work plus TLB shootdowns across eight cores, and it is
reached from RSX state handling, so every microsecond of it lands in fifo_decode.
Counting it was not enough to know whether it is most of that 36.9ms or almost none
of it, and the two point at completely different work. So it gets a bucket and a
scope at the syscall itself, charged only on the RSX thread.
No behaviour change; the scope compiles to a branch on the profiler flag, which is
off by default.
Arkham City dies of SIGSEGV about 140ms after the game writes PS3Progress_Frame_1,
on both the Qualcomm driver and Turnip, with the database on or off, Multithreaded
RSX on or off, and the RSX profiler on or off. A table of ten runs showed no setting
correlates with it.
The reason it took so long to even establish that it WAS a segfault: nothing records
it. No tombstone (/data/tombstones is root-only, so its emptiness proves nothing), no
logcat crash-buffer entry, and no line from our own handler. The single witness
anywhere is Zygote:
I Zygote : Process <pid> exited due to signal 11 (Segmentation fault)
That combination is what a stack overflow looks like here. The handler is installed
with SA_SIGINFO alone, so it runs on the faulting thread's own stack; if that stack is
what overflowed there is nowhere to run, it faults again immediately, and the kernel
applies the default action having written nothing.
So: an alternate signal stack per thread in thread_base::initialize, and SA_ONSTACK on
the handler. Thread-local rather than shared, because two threads can fault at once and
a shared stack would corrupt whichever report lost the race.
This does not fix the fault. It makes the fault reportable, which is the thing that has
been missing all along: the next occurrence should log where it came from instead of
vanishing. Android only.
Arkham City's first RSX profile put 83.9% of a 33.45ms frame in FIFO decode and
essentially zero everywhere else. That was not a finding, it was a gap: fifo_decode
is the bucket the FIFO loop leaves active, so everything without a scope of its own
accumulates there, and draw_setup, vertex, texture_upload, shader_translate,
shader_compile, barrier and cmdbuf had no scope sites at all. They could only ever
read zero.
Adds the three that account for the draw path:
begin() -> draw_setup per-draw setup, with the existing pipeline,
descriptors and texcache_lookup scopes nesting
inside so each is charged to itself
emit_geometry() -> vertex vertex and index upload plus the draw
bind_texture_env() -> texture_upload sampler setup and any upload the bind forces
load_program already carried a pipeline scope and reported 0.006ms/frame, so shader
and pipeline work is genuinely negligible here rather than unmeasured, and is left
alone.
What the first profile did establish stands: Present wait and Fence wait at ~0 mean
the thread is not waiting on the GPU, Idle at 15.7% means it is not starved, and
FIFO stalls at 0/frame mean it is not waiting on the guest. The 28ms is real CPU
work in the draw path. This says which part.
The worker count is decided once, from a reading taken before the emulator has
mapped the PS3 address space or the game has loaded anything. On a cold cache
that reading goes stale almost immediately, and by the time it matters the
count is fixed and cannot respond.
Measured on Arkham City: a first boot peaked at 5228MB where the same session
with the modules already cached sits at 2636MB, and sampling RssAnon against
RssFile and RssShmem put the growth in anon, so it is the compilers holding
LLVM contexts rather than the GPU caches. The process was killed partway
through; on a warm cache the identical settings run fine.
So the number of workers is not really the problem, their overlap is. Below
2GB free, a worker now takes a mutex around a single module's compilation,
which makes it one at a time exactly when that is the difference between
finishing and being killed. Checked per module, at the point of use, because
rechecking is the entire point. With headroom the check fails and nothing is
serialised, so there is no cost in the common case.
The lock is scoped to the compile alone, so a worker waiting on it is never
holding a context while it waits.
Android only.
Follow-up to the previous commit, which lowered modules-per-JIT to a flat 25
on Android. That fixed the kill but paid for it everywhere, including on
devices with memory to spare and on titles that were never at risk. The two
things the constant governs are not the same concern: branch reachability is
about translated game code, while the symbol resolver is a one-shot boot-time
initialiser that fills the jumptable and never runs again. Only the second is
a memory problem, so only the second should be allowed to force a split.
The group is now sized from MemAvailable, the same way the compile worker
count already is. A quarter of what is free, with the ceiling set at what a
full group of 100 costs, since budgeting past the point where nothing would
be split buys nothing.
~2GB free -> 26 per JIT, resolver peak ~500MB
~4.4GB -> 57 per JIT, resolver peak ~1.1GB (this device)
8GB+ -> 100 per JIT, i.e. upstream, untouched
Two separate reasons this stays free in the common case. A device with room
lands on 100 and is not split at all. And a title whose parts fit in one group
gets a single JIT instance at any limit at or above its part count, so every
title below the threshold is unaffected regardless: same instances, same
codegen. At roughly 4000 functions per part that covers everything under about
100k analysed functions, which is most of the library. Arkham City, at ~457k,
is not, and that is the point.
The 5KB per function and 4000 functions per part are measured off the run that
died: 2.3GB across ~457k functions, with parts holding 2800 to 4900 each.
Arkham City was killed partway through compiling its PPU modules. Memory sat
level around 4GB for four minutes of ordinary compile-and-free, then went
4010MB -> 6323MB in seventeen seconds and the log stops mid-compile, no
tombstone: the kernel OOM killer on a 7GB device.
The module part that carries jit_bounds also builds the symbol resolver, and
GetSymbolResolver spans every function across its entire JIT instance rather
than its own part: an LLVM Function declaration and a constant-array entry per
function, then a relocation each through MCJIT. Arkham City analyses to about
457k functions, and at 100 modules per instance the first resolver covered all
of them at once. The log records it plainly, 457209 functions generated in one
module where every other module in the run reports between 2800 and 4900.
No new mechanism was needed. ppu_initialize already splits modules across JIT
instances and jit_mod.symbol_resolvers is already a vector with one entry per
instance, executed in a loop. The group size was simply tuned for a desktop.
Upstream's own comment on the constant names this exact trade: lowering it
lowers continuous memory requirements, at the cost of more branches unable to
reach with a direct B. The resolver's cost is linear in the functions it spans,
so a quarter of the group size is a quarter of the peak.
Android only. Desktop keeps 100.
Enable Time Stretching read true on the device, though upstream defaults it
false, Settings writes ps3.audioTimeStretch which is false, and no core
override touches it. Something later was setting it back.
Six PCSX2 keys reach RPCS3 nodes that a PS3/ key already owns:
Enable Time Stretching <- SPU2/Output/SyncMode
Desired Audio Buffer Duration <- SPU2/Output/OutputLatencyMS
Enable Buffering <- SPU2/Output/BufferMS
Anisotropic Filter Override <- EmuCore/GS/MaxAnisotropy
Clocks scale <- Framerate/NominalScalar
Frame limit <- EmuCore/GS/SyncToHostRefreshRate
They date from before the PS3/ pseudo-sections existed, which were added for
exactly this reason. Being unambiguous in the bridge was not enough: put()
calls setSetting immediately rather than collecting into a map, so applyTo's
source order is the call order, and every PS3 write lands first (L939-994)
with the PCSX2 one later (L1044-1765). The leftover won every time.
The field each UI row is bound to is the PS3 one in all six cases; the PCSX2
counterparts appear only in the reset-fields list or nowhere. So the Audio
tab's Time Stretching toggle did nothing at all: it wrote false, and SyncMode
wrote true ninety lines later.
Two were independently wrong. BufferMS turned a buffer size in milliseconds
into the boolean "buffering enabled". SyncToHostRefreshRate mapped to Frame
limit "Display", which resolves to the host panel's refresh, so on a 120Hz
handheld it asked for a 120fps cap, the same mistake just fixed in
setDisplayRefreshRate; and its `if (asBool(value))` guard could only ever set
the mode, never clear it, so turning the setting off left the cap in place.
All six now fall through to the unhandled path. Verified no RPCS3 setter is
left with more than one writer.
The frame rate still ran past 60. The live config on the device explains it:
Vblank Rate 120, Frame limit Auto, and Auto resolves to the vblank rate.
setDisplayRefreshRate was writing the HOST PANEL's refresh into
Video@@Vblank Rate. Those are not the same quantity. Vblank Rate is the
frequency of the emulated console's vblank, and a PS3 runs 60Hz whatever
display is attached, so a 120Hz handheld was asking the emulator for 120 frames
a second: twice the RSX command volume, twice the GPU work, for frames no PS3
game was ever written to produce.
It also could not be corrected from settings. EmulationSurface reports the panel
rate from surfaceChanged, which runs after ApplySettings on boot and again on
every rotation and resume, so it overwrote the pushed 60 every single time. That
is why recording 60 as a core override did not hold.
Dropped rather than redirected. RPCS3 reads the host rate itself through
get_display_refresh_rate() for Frame limit Display, and never wanted to be told.
Frame limit now goes through the curated push next to Vblank Rate, so the cap
does not depend on the vblank path holding. A deliberate choice on the core
screen still wins, since the overrides replay after.
Also drops the two core overrides the profiling work left behind. RSX Profiler
belongs at its false default in a build meant for playing, and Eager Surface
Readback names a node this build no longer has. Removed by path, because
clearing the store would take the user's real edits with it.
Rotating the device left the picture corrupt in both orientations.
Two independent faults, and both have to go.
m_swapchain_dims held the size passed to swapchain::init, but the WSI backend
replaces that with the surface's currentExtent whenever the platform reports
one, so after any window reshape the two disagreed. That value is not
bookkeeping: it sizes the framebuffer the swapchain image is attached to, and
the present blit region, so a frame was drawn at one size into images of
another. Nor did it recover, because flip() decides whether to rebuild by
comparing that same stale number, so it kept confirming itself. Both init sites
now adopt the size the swapchain actually got.
The other half is that nothing told the renderer the window had changed shape.
The activity handles orientation itself, so rotation keeps the same Surface and
the same ANativeWindow, leaving no new handle to notice; and once a swapchain is
connected, ANativeWindow_getWidth answers about buffers rather than about the
window. SurfaceHolder.Callback::surfaceChanged is the one authoritative source
and it was being discarded, so it now reaches the core and drives client_width().
Three smaller things on the same path. ANativeWindow_fromSurface returns an
already acquired reference and the old code acquired again on top of it, leaking
a window on every surfaceChanged. The event code was computed after
currentSurface had been overwritten, so SURFACE_CREATED could never be sent and
the resume paired with the surface loss pause was dead code. That resume now
undoes only a pause the surface loss itself caused, so it cannot fight the
overlay's.
The Qualcomm driver does not expose VK_EXT_conditional_rendering, which is why
the vendor gate added earlier never fired: the feature was already off, so
turning it off changed nothing.
But VKGSRender::begin_conditional_rendering does its work regardless of support.
It allocates the predicate buffer, copies query results into it and barriers it,
then falls through to the base implementation. Only vkCmdBeginConditionalRendering
reads that buffer, so without the extension all of it is discarded.
On Adreno the waste is not merely wasted. insert_buffer_memory_barrier ends the
open render pass, and that driver allocates on every vkCmdEndRenderPass and does
not release it. A heap profile of a Skate 3 session put its largest allocation
stacks, 157MB, 152MB, 150MB and more, on exactly this path through
qglinternal::vkCmdEndRenderPass into calloc, with the process killed at 4.3GB of
anonymous memory after about 2.4GB arrived in eight seconds.
Returns to the base implementation immediately when the extension is missing.
Behaviour is unchanged, since nothing was being predicated anyway; what goes is
the buffer traffic, the barriers and the render pass ends.
A heap profile of a Skate 3 session, sampled through the crash, put every one of
the top allocation stacks on the same path:
run_FIFO -> VKGSRender::begin -> begin_conditional_rendering
-> insert_buffer_memory_barrier -> end_renderpass
-> qglinternal::vkCmdEndRenderPass -> calloc
157MB, 152MB, 150MB and more from that single stack. begin_conditional_rendering
inserts a buffer memory barrier, which ends the render pass, and every
vkCmdEndRenderPass makes the Qualcomm driver allocate memory it never returns.
The process reached 4.3GB of anonymous memory, took the device to 54MB free with
3GB in swap, and was killed. About 2.4GB of it arrived in eight seconds.
Anonymous, so unreclaimable, and not ours to free: it belongs to the driver.
Without the extension RSX falls back to thread::begin_conditional_rendering,
which performs the draws instead of predicating them. Occlusion results stop
culling, which costs some GPU work, in exchange for sessions that do not end in
an OOM kill.
Scoped to the proprietary driver, which is what was measured. Turnip is a
separate implementation and is left alone until there is evidence about it.
Three earlier guesses at this, the texture cache quota, the SPU JIT and the
eager readback, were all wrong. This one came from allocation stacks.
Two separate fixes.
Memory: the texture and surface caches size their quotas from
device_local_total_bytes. On a discrete GPU that is right, a 3GB texture cache
out of 8GB of dedicated VRAM costs system RAM nothing. A phone has one pool for
both, so that figure is system RAM and the caches budget memory the OS also
needs. Here it reported 7446MB, which resolved the texture cache quota to
2978MB.
Watched a Skate 3 session die: steady around 1.6GB for two minutes, then 2GB
allocated in fourteen seconds, available memory down to 54MB, 3GB pushed into
swap, killed at a 4.3GB peak with no tombstone. That is the quota being honoured
on a device that cannot pay it.
Adds get_budgetable_device_memory, which is the device local heap everywhere
except Android, where it is what is actually free less room for the emulator,
clamped between 1GB and 2.5GB. Both caches now evict against that.
Rotation: getNativeWindow only re-read the window size when the window pointer
changed, and rotating keeps the same ANativeWindow while changing its
dimensions, so the cached size stayed at whatever the first orientation was and
the swapchain was rebuilt at portrait extent inside a landscape window. Size is
re-read on every query now.
Skate 3 still aborted in llvm::report_bad_alloc_error during PPU compilation with
4.6GB reported available, where the previous figures allowed three workers.
Both numbers were too optimistic. A single large PPU module can take well over a
gigabyte through MCJIT and relocation processing, so 1GB per worker does not
cover a big title, and the reading is taken before the emulator maps the PS3
address space, so some of what it counts is already spoken for.
Reserves 2GB for the emulator and budgets 1.5GB per worker. On a 7GB phone with
4.6GB free that is a single worker: slower to compile, but it finishes. Three was
faster right up to the point it killed the process, and a game that aborts during
compilation cannot be played at all.
Minecraft exposed this class of bug after a cache clear and Skate 3 exposed that
the first fix was not enough; neither is new, the original cap sized against
installed rather than available memory.
RSX and every SPU thread shared one affinity mask, so on a 4+3+1 phone that is
six hot threads over five cores. RSX is the thread the frame waits on, and it
was measured spending about 10ms per frame inside its own loop without running:
not blocked on the GPU, not faulting, just waiting for a core.
Carves the single fastest core out of the SPU mask and leaves it in the RSX one.
RSX keeps the whole fast cluster and only loses the contention for the best
core; the SPUs lose one core out of several. Skipped when it would leave the
SPUs with a single core, which would be worse than the problem.
Only takes effect under the alternative scheduler. On Operating System mode
Android places threads itself and this code does not run.
A PS3 runs a 60Hz vblank and every game was written against it. The stored value
was 120, which asks for twice the frames the hardware ever produced: twice the
RSX command volume, twice the vertex upload, twice the GPU work, on a handheld
chasing a panel refresh rate the games predate.
Recording it as a core override did not hold. The override is stored correctly
and the two beside it apply, but they only persist because toggling them in the
UI writes config.yml directly, and CoreSettingOverrides.replay never ran at boot
in any log taken today. So this goes through the curated push instead, which
runs on every apply.
A deliberate change in All Core Settings still wins, since the override replay
happens immediately after.
set_native_priority used pthread_setschedparam with sched_priority. Android
threads run under SCHED_OTHER, where sched_priority must be zero and
sched_get_priority_max returns zero, so the call succeeded and changed nothing.
The RSX thread asks for a boost when it starts and was still measured at nice 0,
taking about 5300 involuntary preemptions a second, roughly 130 per frame, from
the PPU, SPU and audio threads sharing its cores. It is the thread everything
else waits on.
Under SCHED_OTHER the scheduler weights by nice, which setpriority does set, and
Android gives apps enough RLIMIT_NICE headroom to go negative for their own
threads. Applies -8 for a raise and +8 for a drop, and warns rather than fails
if the headroom is not there.
Modest on purpose: a hint to be scheduled ahead of the other emulator threads,
not a bid to starve them.
Timing each render pass overran the GPU timer's per-frame event cap by two
orders of magnitude: 554600 events dropped over 300 frames, so no draw region
ever completed and the line was missing from the report entirely.
That failure is the finding. Roughly 1800 render pass begins per frame, for a
game that should need a handful. On a tiled GPU every pass boundary stores the
tile buffer to memory and reloads it, which is the most expensive thing the
architecture does, and it would account for the 20ms of GPU time on its own.
Counts them plainly instead, since a counter cannot be overrun. The blit,
upload and readback regions stay timed and are all tiny: 0.012, 0.463 and 0.144
ms per frame, so none of them is where the GPU time goes.
Clearing the shader cache forces every PPU module to recompile at once, and the
process aborted partway through: scudo internal map failure, NO MEMORY, in
RuntimeDyldELF relocation processing on a PPU worker thread.
The existing cap allowed one worker per 1.5GB of total RAM, so four on this
device. Total RAM is the wrong number. The same device reported 7.3GB installed
while sitting at 76MB actually free, because it is also holding everything else
the user is running. Four LLVM workers on top of the emulator's own couple of
gigabytes had nowhere to go.
Adds utils::get_avail_memory, reading MemAvailable from /proc/meminfo, which is
the kernel's own estimate of what can be handed out without swapping. Workers are
then budgeted against that with a 1.5GB floor reserved for the emulator, falling
back to a more conservative slice of total where it cannot be read.
At 4.9GB available that allows three workers instead of four, and near zero it
correctly allows one.
The texture cache now reports zero misses and zero hard faults, so the eager
readback is doing its job and the guest no longer faults on surfaces. Yet
submissions per frame are still around six, and GPU work plus fence wait still
sum to the whole frame, which is the serialisation keeping this at 40fps.
So the drains are not readback faults, which is what the last several changes
assumed. Counts the three remaining callers instead: GCM label release with
texture loads outstanding, the FIFO sync hint that declares a hard sync coming,
and flushes requested by another thread.
A refill pays a fixed cost regardless of size: read_put, the iotable lookup, the
reservation lock, and a reservation_acquire per line. Measured at about 10us per
refill against roughly 1us of actual copying, so nine tenths of it was that fixed
cost, and a heavy scene ran over 1200 refills per frame to move 1.29MB of FIFO,
which is around 324,000 command words.
Fetching 4KB at a time pays that cost a quarter as often for the same bytes.
The line mask was a u8 pinned to 8 lines, so it widens to u32 with the full-mask
case special-cased, since 1u << 32 is undefined. The traversal order was checked
against the original at both widths before changing it.
Also adds counters for refill stalls, which ruled out the alternative
explanation: the retry spin is 0.006 ms/frame, so the refill is doing real work
rather than waiting on the guest.
The expansion indices for primitives the host cannot draw natively depend only
on the index, never on the draw's data, so the buffer for N primitives is
exactly a prefix of the buffer for any larger N. They were regenerated on every
draw call anyway, one u16 at a time, written straight into mapped GPU memory.
Builds each table once and copies the prefix. Tens of thousands of dependent
scalar stores become a single bulk copy, which also suits write-combined memory
far better than a scatter of small writes.
Minecraft draws quads throughout, so this ran on essentially every draw and
measured 5.7% of RSX thread samples.
Patterns were diffed against the original loops across edge cases before
replacing them; u16 indices bound both tables at 65536 vertices and anything
larger keeps the old path.
fetch_u32 is called once per FIFO command word and the guest pushes about 52,000
of them per frame, while the cache refills only around 208 times. Better than
99% of calls are a compare and a load, but the function lived out of line in
RSXFIFO.cpp for the sake of the rare refill, so every one of those 52,000 was a
real call into another translation unit.
It measured 15.6% of RSX thread samples, which no amount of refill work
explains: 208 refills of a 1KB cache is a few tens of microseconds of copying.
The cost was call overhead, not work.
Splits the refill into fetch_u32_refill and leaves the hit path inline.
Two measurement faults, both mine, both of which sent this investigation at the
wrong targets.
The scope timing the swapchain acquire was declared at function scope, so it
lived until flip() returned and charged the whole present path against
present_wait. That bucket read 10.12 ms and looked like the largest cost in the
frame. It is now braced around the acquire alone.
The accounting was in thread locals, and under the generic TLS model every scope
enter and exit went through the linker's tlsdesc resolver: 21% of RSX thread
samples against 0.07% uninstrumented, landing hardest on the FIFO bucket, which
is why that number would not move when the FIFO fetch path was bypassed.
initial-exec removed the cost but stopped every game booting.
Since only the RSX thread is ever reported, there is now one copy of the
accounting and a thread-pointer check at the door instead of a copy per thread.
On ARM64 that is a single register read with no relocation, cheaper than one
tlsdesc access where a scope previously did six.
CPU profiling has taken this as far as it goes. With the readback wait backed
off, the RSX thread drops to ~22% host CPU while the Adreno holds 81-88% busy at
its maximum clock and nothing throttles, so the GPU is the limiter and every
remaining question is on its side.
A whole-frame GPU timestamp would not have been worth writing: the kernel
already exposes gpu_busy_percentage and it says the same thing. What no counter
exposes is the split, specifically how much GPU time goes to copying render
targets back for the guest to read rather than to drawing. So the regions are
labelled, and readback is bracketed where the copy is actually recorded.
Readback of the queries is deferred: results come from ring slots written frames
earlier, a slot that is not ready is retried later, and no path waits on the GPU.
Reading a query in the frame that wrote it would stall on exactly the thing being
measured. vkDeviceWaitIdle is never called.
Frames are retired from flip rather than from the frame region closing, because
the primary command buffer is submitted once per flush_command_queue and several
times per frame, which would divide every per-frame figure by the wrong number.
Untimed events past the per-frame cap are reported rather than dropped silently,
so a truncated frame cannot read as a cheap one.
vkCmdWriteTimestamp had to be added to the Android Vulkan loader, which resolves
an explicit list of entry points and did not include it.
vk::wait_for_event spun on vkGetEventStatus with nothing but an isb between
calls. Profiling the RSX thread put 44 to 53 percent of its samples in that
loop, on a build with no instrumentation in it, while the GPU sat at 88 percent
busy pinned at its maximum clock and nothing was thermally throttled.
A readback issued mid frame queues behind everything already submitted, so the
wait drains the whole pipeline and runs into milliseconds. Spinning through that
does not make the event arrive sooner. It holds a core at peak clock and, worse
on a tiled mobile part, keeps reading memory the GPU is writing, taking
bandwidth from the device we are blocked on.
Polls hot for a bounded window so short waits behave as before, then backs off
to 50us. Against a wait measured in milliseconds that granularity is noise.
Also pins the profiler's thread locals to initial-exec. Reading them through the
generic model routed every scope through the linker's tlsdesc resolver and cost
21 percent of RSX thread samples against 0.07 percent uninstrumented, which
inflated the FIFO bucket it was meant to measure.
The overlay's RSX percentage measures what the thread is not doing: get_load()
counts everything outside four idle sites, so a thread decoding commands and a
thread spinning on a Vulkan fence both read as fully loaded. That is exactly the
distinction that decides what is worth optimising, and it could not be read off
any existing counter.
Splits RSX thread time into exclusive buckets instead. Entering a scope charges
elapsed time to whatever was active and switches, so nesting attributes to the
innermost scope and the totals sum to wall clock rather than double counting a
caller with its callee. Reports to the log every 300 frames, including an
explicit unscoped remainder so the percentages cannot read as complete coverage
when they are not.
Behind the "RSX Profiler" video setting, dynamic, so it can be armed once a
slowdown has already started. Off by default, costing one predictable branch per
scope. Accounting is per thread because some of these paths run on whichever
guest thread faulted rather than on RSX; only the RSX thread's copy is reported.
No optimisation here, only measurement.
Adds an apply toggle and a remove action next to the download row, so a title can
be run with and without its database entry to compare. Disabling renames the split
files aside rather than deleting them, so flipping back does not mean re-fetching
two thousand titles; remove deletes both copies and returns to stock.
Also drops Multithreaded RSX from the deny list. It was listed on the belief that
it froze Minecraft, which was wrong: the freeze was Accurate SPU DMA plus Accurate
Cache Line Stores, and it reproduced with Multithreaded RSX off and an empty
database. It is a real upstream feature backed by the RSXOffload thread, so there
was never evidence against it, and the toggle is the honest way to settle whether
a database entry helps or hurts.
Accurate SPU DMA and Accurate Cache Line Stores were both on. Together they
route every 128-byte SPU DMA store through do_cell_atomic_128_store, turning
bulk DMA into one atomic reservation store per cache line. Reservation
contention then outruns the rate it can drain, an SPU spins in do_putllc
forever, and the PPU stalls behind it on a semaphore the SPU never signals
while the RSX idles.
Both default off upstream. Cache Line Stores was wrong in our defaults;
SPU DMA already defaulted off but was stored on, so a default change alone
would not reach anyone who had already run the app, hence the migration.
Seen on Minecraft, which froze loading world chunks, that being bulk SPU DMA
and little else. Load dependent, so it presented as an intermittent freeze.
It defers nothing on this port: the Android call_from_main_thread callback runs
the function inline on the caller's thread. All it added was RPCS3's state-guard
wrapper, a std::function allocation and a log call, on a path applyTo hits eleven
times per settings change.
The guard that skips the reset entirely while stopped is what actually prevents
the crash it was meant to fix, and that stays.
The recommended per-title settings the desktop build offers under "Download
Config Database". The core already knows how to use them: Emulator::Load asks
through the get_database_config callback and passes the result to BootGame as
db_config, which sits under the user's own settings.
That callback was returning a PATH, which was simply the wrong thing. The return
value is taken as the config CONTENT, so what it got back was a filename that
failed to parse and was discarded, and the feature has never done anything here.
It reads the config for the title now.
The download and the JSON parsing were the only parts missing, and they lived in
rpcs3qt, which this build does not compile. Those are done in the app instead:
api.rpcs3.net returns every title in one object, and it is split into one YAML
per title so the callback, which runs on the boot path, does not have to parse a
database of every PS3 game to find one entry.
Manual rather than automatic on launch, since it is a network call to a third
party. 2164 titles at the time of writing.
Emu.Kill() spawns an Emulation Join Thread that joins every emulator thread, and
nothing stopped two of those existing at once. They end up joining each other and
neither finishes, so the emulator never reaches stopped. On device that showed as
four live join threads and six leaked AudioTrack threads after a few close then
boot cycles, with the join thread logging that it was waiting on itself.
Boot took the lifecycle lock only after calling Kill, so a Close and the boot that
followed it each started a teardown and deadlocked. The lock is now taken first
and held across the whole boot, which is also what the disc probe and shutdown
already use, so probing, booting and killing can no longer overlap at all.
This is what was behind the reports of being kicked back to the library when
opening a game, of crashes when switching games quickly, and of the save screen
never appearing: all of them were the emulator failing to reach a stopped state
rather than anything wrong with booting or with saves.
Disc probing was the cause of the game switching crashes. probeDisc mounts the
image into the emulator's GLOBAL vfs to read its PARAM.SFO, vfs::mount begins
with g_fxo->need<vfs_manager>(), and the scan runs on a background thread. So a
scan overlapping a boot or a teardown aborted the process inside vfs::mount.
Closing a game does both at once: Emu.Kill() resets g_fxo and returning to the
library starts a rescan. Waiting ten seconds only worked because the scan had
finished by then.
Three parts. The probe cache is now seeded from the last scan, so a disc that
has been seen before is never mounted again, which removes almost all of the
window and stops rescans re-reading multi gigabyte images. Probing, booting and
killing take a shared lock, so what is left cannot overlap. Booting waits for the
previous VM to actually stop, since BootGame's failure path asserts IsStopped and
aborts otherwise, which is what killed PES 2018 on a failed boot.
The overlay reset that settingsSet performs is posted through CallFromMainThread
rather than run on the JNI caller's thread. It walks the overlay manager into
perf_metrics_overlay::update(), which is RSX-owned, and doing that inline is what
crashed when a setting was changed while a game was starting or closing.
All Core Settings edits are recorded and replayed at the tail of applyTo. The
curated store pushes about 165 nodes on every settings change and on every boot,
so anything set on that screen which overlapped one of them was reverted moments
later. Per-title required settings land just before them, so a game that needs a
workaround gets it while an explicit user choice still wins. Uncharted 3 needs
Stub PPU Traps to get past its own crash handler.
Also adds .rap and .edat licence installing, which routes to installKey rather
than install, and migrates the stored Scaling Mode value so PR 10's change of
meaning does not silently move everyone from Bilinear to Nearest.
SkipDuplicateFrames in PCSX2 means "do not present a frame identical to the
last one". It is harmless there and on by default, which is why Settings
defaults it to true. RPCS3 has no equivalent, and the bridge mapped it onto
Enable Frame Skip, which means something completely different: drop one frame
in every two, unconditionally.
So every game presented at half the rate the guest asked for, on a fresh
install, with nothing touched. It also made the frameskip row in the in-game
menu useless, because applyToInner pushes the explicit frameskip first and this
key afterwards, so this one always won no matter what you picked.
Measured on Mirror's Edge, which asks for 30 flips a second. SurfaceFlinger was
presenting 15.0 fps, one frame every 66.65 ms, rock steady. The RSX thread
reported 0-5% load with six of eight cores idle, which is what sent me looking
for a performance problem that was never there. With the mapping gone it sits
at 30.0 fps, 33.32 ms.
Dropped rather than remapped. Frameskip belongs to the explicit control, which
is the one the user actually set.
Four keys were writing Output Scaling Mode and the two with nothing behind them
were winning. IntegerScaling and linear_present_mode are PCSX2 keys that no
screen in this app exposes, both wrote the node unconditionally, and
IntegerScaling was emitted last, so it overwrote whatever the visible Scaling
Mode row had asked for.
On top of that the row itself was misread. It offers Nearest, Bilinear and FSR
and stores the index in casMode, but the bridge treated casMode as PCSX2's CAS
mode, where anything above zero means "CAS on". So picking Bilinear asked for
FSR, picking Nearest asked for nothing, and none of it survived anyway.
The two invisible keys no longer touch the node, and CASMode maps its own
indices. CAS is emitted before ShaderChainEnabled now so the chain still gets
the last word when both are on, which is what the bridge already documented.
casMode defaults to Bilinear rather than Nearest. Bilinear is what the core has
been using all along, and what RPCS3 itself defaults to, so nobody's picture
changes. Leaving the default at 0 would have quietly switched every user to
nearest neighbour the moment the mapping started working.
Checked on an Odin 3, reading Output Scaling Mode back out of config.yml:
picking FSR now gives FidelityFX Super Resolution where it used to give
Bilinear, and picking Bilinear gives Bilinear for the right reason.
Closes#3.
setAspectRatio was written against ARMSX2's PS2 aspect enum, where index 1 was
4:3, and its javadoc still described that enum. ARMSX3's picker in RendererTab
is a different one: 0 stretch, 1 Auto, 2 is 4:3 and 3 is 16:9. So the check for
index 1 was matching Auto, and Auto is the default.
Every game therefore came up in 4:3 on a fresh install, and RPCS3's own default
for that node is 16:9. Setting the picker to 16:9 by hand worked, which is what
the issue reports, because index 3 fell through to the 16:9 branch.
Compare against index 2 instead, so only an explicit 4:3 selects 4:3 and
everything else lands on 16:9. That also lines it up with the other writer of
this node, writeGsToNative, which maps the same indices by name and already
treated anything other than "4:3" as widescreen.
Verified on an Odin 3: config.yml went from "Aspect ratio: 4:3" to
"Aspect ratio: 16:9" with the picker left on Auto.
Two things end up writing the same config node. RPCS3 has its own Performance
Overlay switch, in OverlayTab and the in-game menu, both writing
ps3.overlayEnabled. ARMSX2 has its twelve per-stat osdShow* flags, and
osdApplyFlags derives Enabled purely from those, all of which default to off.
So it pushed Enabled=false right over the switch.
Boot order decided the winner. MainActivityRuntime calls applyTo(), which
pushes the switch, and then applyStoredOsdMode() one line later, which pushes
the flags. Turning the overlay on in settings did nothing at all.
Only that one key was affected, because osdApplyFlags returns as soon as it
sees nothing enabled and never reaches the graph, font size and opacity
settings. That is why config.yml held the user's values for those while
Enabled sat at false.
Re-assert the switch after the flags have been applied. Scoped to the Custom
path, since that is the mode that reads saved settings and the one that runs
at boot. Full and Min pass explicit flags, and Off is meant to be off.
A game data install looks exactly like an installed HDD game on disk, a
PARAM.SFO next to USRDIR, so isPs3GameFolder accepted it. Every title you
install data for therefore got a second tile that cannot boot, since game data
holds no EBOOT. Skate 3 ships a 1.1 GB BLUS30464_INSTALL and showed up twice.
CATEGORY is what tells them apart. GD is data, HG and DG are games. The native
scanner already rejects these, because fetchGameInfo requires BOOTABLE and game
data does not set it, but the folder path never opened the SFO at all.
Reads just that one field rather than adding a general SFO parser. Nothing else
needs it, and an unreadable or malformed file falls back to listing the folder,
so the worst case is a stray tile and never a hidden game.
ScanSchemaVersion goes up as the comment on cacheKey asks, since the cache
stores the scan result and an existing one would keep serving the old entry.
The precompilation queue calls vm::init() so that ppu_register_range has
somewhere to register, but nothing ever unmapped it again. vm::close() is only
reached from Emu.Stop(), and precompilation deliberately never boots anything.
So the blocks stayed mapped past "Finalization". The next vm::init(), either
the next workload or Emulator::Load() booting a game, assigns over g_locations
and destroys them, which trips ensure(!is_valid()) in ~block_t() and takes the
whole process down.
Easy to reproduce: install firmware and then boot a game without restarting the
app. The firmware pass leaves five live blocks behind and the boot dies inside
vm::ps3_::init() before it reads a byte of the disc.
Two sources, both writing real PNGs to shared storage where Android's media
scanner indexes them and they turn up in the camera roll.
The data root holds firmware assets: trophy icons under dev_hdd0/home, the whole
dev_flash VSH resource set, and an ICON0.PNG per installed game. That was 216
images. A .nomedia now goes in before the core initialises, since that is what
unpacks the firmware and creates most of them.
The ROM folder is the one people actually notice, and only since folder format
games started working. An .iso is a single opaque file so the scanner sees
nothing inside it, but a disc in folder form lays its ICON0.PNG, PIC1.PNG and
every DLC image out in the open. One Minecraft folder accounted for 245 images,
which was every image in the whole ROM tree.
The ROM marker goes at the configured directory root rather than inside a game
folder, so it covers current and future folder games with one file and never
leaves a stray file inside content that gets mounted as a disc. Writing it also
triggers a rescan, because the marker alone does not drop what MediaStore has
already indexed. Zero bytes and reversible either way.
2026-08-07 00:25:47 -04:00
240 changed files with 30398 additions and 21330 deletions
Some files were not shown because too many files have changed in this diff
Show More
Reference in New Issue
Block a user
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
jpolo1224
Zulux91
Zulux91
