Upstream bc22df8ba skips waking waiters after a SUCCESSFUL conditional store
when the SPU sits at pc 0x11e4 with the SPURS control block reserved, unless
byte 0x73 of that block shows this thread going running->idle. It is a
throughput optimisation: SPURS kernels store to that block constantly and
waking every waiter each time is a thundering herd.
Both constants are assumptions about one specific SPURS kernel build. 0x11e4
is a guest code address and 0x73 an offset inside the guest's control block,
and SPURS ships in many versions across titles. On a kernel whose layout
differs, the running->idle test reads the wrong byte, answers no forever, and
the store succeeds while every waiter stays asleep -- reported by nothing.
Tales of Xillia (BLUS31006) hangs with all five graphics SPUs reserving that
block, each suppressing ~100,000 notifications, ~4.85M conditional stores at
a 16.5% failure rate. Its SPURS kernel then executes its own HALT at
pc=0x00f00: r16 & r17 = 0x40000000, two workload masks that must be disjoint
both claiming workload bit 30. The group never joins, no frame is ever
presented again, and every PPU thread parks.
Whether the missed wakeups cause that inconsistency or merely accompany it is
NOT established. What is established is that notifying is correct and
suppressing is the optimisation, so the optimisation goes. Expect a
throughput cost on SPURS-heavy titles; it is measurable and revertable.
The counter is kept, now recording how often the heuristic would have
suppressed, so that cost shows up in a log instead of being guessed at.
The stall report picked its detailed SPU by raddr == spurs_addr -- a kernel
waiting on its own control block, i.e. an IDLE one -- on the assumption that
every kernel parks at the same pc so any of them would do. Tales of Xillia
is the counter-example: one graphics kernel executed a guest HALT and sits
stopped at pc=0x00f00 while the other four idle normally at 0x011a8, so the
rule picked an idle SPU. The only thread in the process with anything to say
printed no registers at all.
A stopped, halted or exited SPU now wins outright, chosen in a pass before
the walk. Nothing else in a SPURS group stops on its own, so if one has, it
is why the group never joined and the rest are only waiting on it.
Also dump all 128 GPRs rather than the first 16. Xillia's assertion is a
validity check over r12/r16/r17/r19/r33/r34, so the sixteen that were printed
did not include a single operand of the test the log had just disassembled.
average_playtime_ratio divided m_average_playtime -- a rolling average of
get_enqueued_playtime(), in microseconds -- by audio_buffer_length, which is
a sample count (AUDIO_BUFFER_SAMPLES * channels, 512 for stereo). The result
was microseconds per sample: roughly 78 on a healthy stereo buffer and never
below 1, so the 'not as full as desired' branch it gates has not executed
once since it was written. The buffer target has therefore always been the
fixed desired_buffer_duration + half a block, with the adaptive widening
silently inert.
Wrong from introduction rather than drifted: m_average_playtime has been a
duration since before that line existed. audio_buffer_length had no other
reader anywhere in the tree, which is why nothing caught it; it is deleted
here so it cannot be picked up again by mistake.
The denominator is now desired_buffer_duration, which is what the comment
above it names and the only choice that makes the branch meaningful --
audio_block_period would leave the ratio near 7 and the branch just as dead.
Guarded against a zero denominator, and the existing max(ratio, 0.25) clamp
still bounds the widening at 4x.
This arms a path that has effectively never run, so it changes behaviour for
every game: a consistently under-filled buffer now raises the target instead
of being ignored, trading latency for fewer dropouts. That is what it was
written to do. The audio buffer report added alongside prints the ratio, so
the effect is visible in a log rather than inferred.
Progressive audio delay keeps being reported (#87: Guitar Hero titles, where
audio starts synchronised and falls further behind the notes the longer a
song runs, and pausing resets it). It reproduces on both Cubeb and Oboe, so
it is not the backend -- it is this ring, and how full this ring is IS the
delay the player hears. It was recorded nowhere, so every theory about it is
unfalsifiable from the logs we receive.
One line per 10s: queued depth, the target the algorithm is aiming for, and
the dynamic period as a percentage of nominal, which says how hard it is
correcting. Enough to see the curve across a song; few enough that the log
volume cannot become a stall in its own right.
Noted while reading the algorithm, deliberately NOT changed here:
average_playtime_ratio divides m_average_playtime (microseconds) by
audio_buffer_length (samples times channels, 512 for stereo), so it reads
about 78 when healthy and its 'not as full as desired' branch has never once
executed. The intended denominator is a duration. Correcting it would arm a
widening path that has effectively never run, which raises latency in exactly
the under-buffered case -- the opposite of the symptom being chased. It wants
its own change, with the numbers this report will provide.
check_frame_stall() already detects a hang reliably and always-on: it
reports "No frame presented in 30s" and puts a message on screen. But the
one thing that says WHERE the guest is parked -- dump_guest_threads_stalled(),
which prints every PPU thread's registers, guest call stack and the
instructions around its cia, plus SPU state -- was only ever reached from the
RSX profiler's poll_stall(), and that returns false immediately unless the
profiler is switched on. Testers do not switch it on.
So every freeze report arrives with the detection line and nothing behind
it. Confirmed against a Kane & Lynch (BLUS30102) capture: guest execution
collapsed at 0:04:15 -- sys_mutex_lock fell from ~200,000 per 10s to ~300,
CPU from 43% to 5.5% -- and this fired at 0:04:45 having recorded none of
it. Three such reports across two different SoCs, none reproducible locally,
all equally undiagnosable.
Call it from the detector instead. Twice, ~15s apart, because one sample
cannot tell a thread spinning from one making very slow progress -- a cia
that has not moved between two samples is itself the finding. Twice and no
more, and re-armed only when a frame actually lands: the detail runs to
hundreds of lines per thread, and log volume alone is enough to stall the
emulator on Android. The profiler path is unchanged.
The backend asked for the guest's rate with sample rate conversion
disabled, then recorded the rate it ASKED for rather than the one the
stream actually opened at. A device that cannot run at the guest rate
opens at its own, and with conversion off nothing resamples -- so the
emulator feeds samples at one rate into a device consuming them at
another. That is a permanent rate mismatch, not a one-off pitch error:
the gap between what is heard and what is on screen grows for as long as
playback continues, and nothing downstream can see it, because every
later calculation (including the buffering algorithm's idea of how much
audio is queued) uses the rate we requested.
Check the opened rate and, on a mismatch, reopen letting Oboe convert.
Converting is a worse signal path than matching rates outright and a far
better one than not converting at all; Cubeb resamples for the same
reason. If the rate is still wrong after that, say so in the log rather
than drifting in silence.
Refs #87
load_iso() cannot report failure: it mounts the virtual device whether or
not the archive actually opened. An image that is unreadable, truncated,
or encrypted without a usable disc key therefore sailed past the mount and
died much further down as a generic 'invalid file or folder', with nothing
in the log distinguishing it from a dozen other causes. A bug report of
that failure carries no information at all.
Check the mount where the reason is still knowable. An image whose
filesystem did not parse now reports whether a disc key was missing or
rejected -- returning decryption_error, and naming both the key locations
searched -- or, failing that, says the image looks truncated. When the
archive is readable but has no EBOOT.BIN at the expected path, log what the
disc root did contain; install discs legitimately lack one, so that stays a
warning rather than a failure.
Also realign BootResult with game_boot_result. It was missing
firmware_version and database_config_missing, so every code from ordinal 10
down was reported as its neighbour -- still_running surfaced as
'AlreadyAdded' -- and the last two had no entry at all, making fromInt
throw rather than return. fromInt is now total.
Refs #88
present_generated_frame submitted its blit and presented immediately.
command_buffer::submit goes through queue_submit, which DEFERS to the offloader
thread when multithreaded RSX is on -- so the present waited on a semaphore whose
signal operation had not been submitted, which the spec forbids and which hangs
the present without returning the acquired image. The real present path flushes
first for exactly this reason; this now does the same.
Frame generation also refuses to run at all with multithreaded RSX enabled.
generate() submits inline on the RSX thread while the frame's own command buffer
is merely enqueued to the offloader, so the interpolation can reach the queue
before the frame it reads -- and two threads then submit to one VkQueue without
the external synchronisation vkQueueSubmit requires, which is what g_submit_mutex
exists for and what generate() bypasses by calling vkQueueSubmit directly.
Refusing is honest. Supporting it means routing through vk::queue_submit and
reasoning about deferred submission, which is not a change to make untested.
MTRSX defaults off, so this affects only users who enabled it -- and it is the
profile that produces intermittent DEVICE_LOST under load.
capture_presented_frame() was handed target_layout -- COLOR_ATTACHMENT_OPTIMAL or
TRANSFER_DST_OPTIMAL -- and stored it for generate() to use. But generate()
records into its own command buffer, submitted after the frame's, and the frame's
own transition to present_layout runs in between. So by the time the barrier
executed, its oldLayout named a layout the image had already left, and the
matching barrier on the way out restored that same wrong layout -- leaving the
image outside PRESENT_SRC_KHR when it was presented.
Both are violations: oldLayout must be the current layout or UNDEFINED
(VUID-VkImageMemoryBarrier-oldLayout-01197), and a presented image must be in
PRESENT_SRC_KHR (VUID-VkPresentInfoKHR-pImageIndices-01430). It also desynced the
renderer's own layout bookkeeping for that swapchain image on every subsequent
frame.
Passing present_layout instead describes what the image will actually be in, so
the barrier is a no-op transition that reads it where it lies and leaves it there.
Found by a Vulkan lifetime and layout audit. Not verified on device.
Two findings from the Vulkan audit, both of which let the interpolation read an
image it had no valid dependency on.
The barrier before Process used TOP_OF_PIPE with MEMORY_READ. TOP_OF_PIPE in
srcStageMask specifies no stage of execution, so the first synchronization scope
was empty and no availability operation happened for the colour-attachment and
transfer writes the frame made into that image in the PREVIOUS submission.
Submission order on a queue orders execution; it is not a memory dependency. So
the interpolation could read a partially-flushed frame -- which presents as
smearing during motion and does not respond to any interpolation setting, since
the interpolation is not what is wrong. Source scope is now
COLOR_ATTACHMENT_OUTPUT|TRANSFER with the matching write access, and the mirror
barrier's empty BOTTOM_OF_PIPE second scope becomes ALL_COMMANDS.
Separately, the swapchain was created COLOR_ATTACHMENT|TRANSFER_DST, but
FrameGen::Process copies FROM the presented image as TRANSFER_SRC_OPTIMAL. Both
that layout and vkCmdCopyImage require TRANSFER_SRC usage. Nothing else in the
renderer reads a WSI image, so it had never been needed. Requested only where the
surface reports it, with a warning otherwise, since it is not guaranteed on
Android.
Frame generation's state is file-scope global and had no teardown path.
shutdown() only finalised the old dlopen'd library, and release_shared_images()
was never called at all, so at emulation stop the fences, command pool, command
buffers, allocator, queue and the whole pass chain stayed live as children of a
destroyed VkDevice.
The reachable consequence is worse than the leak. On the NEXT boot in the same
process g_native.valid() was still true, so build_native_stack() was skipped and
generate() waited on a fence from the dead device, reset freed command buffers
and submitted to its queue. g_shared_w/h still matched too, so the output images
were not recreated either.
That is boot -> stop -> boot, and it is deterministic rather than intermittent.
It is the likeliest explanation for frame generation working in one game and then
failing immediately in the next one launched without restarting the app, which is
exactly how it was reported.
release_device_resources() destroys the stack and the images and resets the state
that gates rebuilding, including g_disabled -- a new device may not fail where the
last one did. Called from ~VKGSRender after vkDeviceWaitIdle, while the device is
still alive.
Found by a Vulkan lifetime audit. Not verified on device; the hardware was
unavailable.
shader_count() read the ENTIRE SPIR-V cache off disk and rebuilt every module,
then threw the result away -- and it ran twice per frame on the RSX thread, from
capture_presented_frame() and generated_frame_count(). At ~322KB that is roughly
644KB of file IO, four stats, two opens and hundreds of allocations per frame, on
the thread driving the whole present path.
That is CPU cost, which is why halving the GPU shader cost with fp16 changed
nothing, and why ~3ms of measured interpolation was costing ~20ms of frame time.
It arrived with the fix for shaders not loading at all: making the probe ask the
cache was right, doing it per frame was not. The verdict can only change on
import, so it is memoised and cleared there, as ARMSX2 does for the same reason
(GSLsfg.cpp:53-58, whose comment names this exact hazard).
Also: GenerateInto writes through a storage image view, which requires
VK_IMAGE_LAYOUT_GENERAL, and the blit that follows declares TRANSFER_SRC. Nothing
transitioned those images either way -- a storage write outside GENERAL is
undefined, and on Adreno a mismatched layout on a UBWC image forces a
conservative full-surface decompress. Both transitions added, with the layout
tracked across frames.
And the generated-frame blit is source-size to destination-size, so LINEAR
filtering bought nothing and cost a filtered sampling pass instead of the fast
copy path. NEAREST.
Found by a comparative audit against the ARMSX2 driver. Not yet verified on
device -- the hardware was unavailable when this was written.
The fence wait at the head of generate() blocked for up to a second on the
previous slot's work. When it actually waited, the long frame interval it
produced went straight into the pacer -- whose clock is sampled after this point
-- and any interval over 100ms trips MINIMUM_BASE_RATE and stops generation for a
full second (FrameGenPacer.cpp:29/40/97-100).
So a stall here made the pacer stand down, and it was measuring a rate we had
slowed ourselves. Measured against that: the pacer asked for a generated frame on
only ~40% of real frames, which matches one >100ms interval every couple of
seconds costing a second of generation each time.
A zero timeout asks whether the previous frame's work is done and skips this
frame if it is not. Skipping costs one generated frame; stalling cost a second of
them. With three rotating slots the work has three frames to finish, and it
measures ~3ms, so skips should be rare.
VK_TIMEOUT is now a normal outcome rather than an error, so only real failures --
VK_ERROR_DEVICE_LOST being the one actually suspected here -- reach the counter
that disables the feature.
Every non-success result from vkWaitForFences was reported as a timeout, which
hid the one detail that mattered: 60 of them arrived in 3 seconds against a
1-second timeout, so the wait was failing immediately rather than expiring.
VK_ERROR_DEVICE_LOST is the obvious candidate and means something entirely
different from a slow frame -- the device is gone and frame generation is the
messenger, not the cause.
The failure is intermittent, so this logs the result code once per stall rather
than trying to fix a cause that has not been identified.
The single-submit path broke two games -- one reported "failed", another stopped
progressing past its logo -- and reverting it restores both. It was chasing ~8 ms
of measured per-submit overhead, and the idea may still be right, but it changed
swapchain image ownership incrementally and got it wrong four separate ways:
acquiring before the count was known (leaving semaphores signalled and every
later acquire failing), sharing semaphores across in-flight frames, leaking
acquired images that were never presented, and writing storage images outside
GENERAL. That is a design that needs settling before it is written, not patched
into place against a device.
Kept, all measured:
- The acquire semaphore. Generated frames were acquired with VK_NULL_HANDLE and
written immediately -- a swapchain image used before the presentation engine
had finished with it. That was the judder: invariant to every interpolation
setting because the interpolation was never at fault, present at any frame
rate because a race has no cadence, and smearing rather than stutter because a
torn image is parts of two frames at once. Gone entirely at x2.
- The guest extent. The passes size their optical flow from the ratio between
what the game rendered and what is presented; passing the presented size for
both made that ratio 1.0 and ran the flow at full 1080p for a title that
rendered 720p. 4.8 ms -> 3.0 ms of GPU time.
- fp16 shaders, where the device reports and enables them.
- Fence timeouts are transient. disable() is permanent for the session, so one
stall during a load spike left frame generation reading "failed" until the
game was restarted with nothing wrong.
Frame generation costs ~3 ms of GPU time per frame. On a title already at the
refresh rate that is enough to cross a vblank boundary, so it trades real frames
for interpolated ones and gains nothing; it is worth having where a game has
headroom below the panel rate, not where it is already at the cliff edge.
Measured, not guessed: a GPU timer around the passes says they cost 4.1 ms per
frame, steady, while frame time went from ~17 ms to ~30 ms with generation on.
Roughly 8 ms per frame was therefore not interpolation -- it was the cost of
submitting it.
ARMSX2 records Process, every GenerateInto, and every copy into the acquired
swapchain images into ONE command buffer and issues ONE vkQueueSubmit that waits
on each acquire and signals each present. ARMSX3 submitted framegen's work and
then a separate command buffer and submit per generated frame, each its own GPU
batch on the same queue with its own scheduling gap.
generate_into_targets() takes the acquired images and does the whole frame in one
buffer and one submit; the present path acquires them up front and issues the
presents afterwards. present_generated_frame is left in place for the pipelined
path, which is still disabled, and no longer runs on this one.
Follows the acquire-semaphore fix in the previous commit -- that was the judder
(a swapchain image written before the presentation engine had finished with it),
this is the cost.
present_generated_frame acquired its swapchain image with VK_NULL_HANDLE -- no
semaphore, no fence -- and then blitted into it and presented immediately.
Vulkan requires waiting on the acquire before touching the image, so this was
writing into one the presentation engine could still be scanning out.
That accounts for what nothing else did. The artifact was invariant to flow
scale, shader family, fp16 and multiplier, because all of those change how a
pair is interpolated and the interpolation was never at fault. It appeared at
every frame rate, because a race has no cadence. It looked like smearing rather
than stutter, because a torn image is parts of two frames at once, and it only
showed during motion, because that is when consecutive frames differ.
The real frame has always had an acquire/present semaphore pair in
frame_context_t. Only the generated path went without, which is why nothing
inside frame generation could reveal it.
ARMSX2 acquires with a fence and blocks on it before submitting, and says so.
This waits on the GPU instead: the blit waits on the acquire, the present waits
on the blit, one pair per swapchain image so none is reused while in flight.
Also restores the fp16 probe consistency in shader_count(), which a git checkout
during this session reverted -- three callers ask which shader variant to use and
any two disagreeing makes a freshly imported cache read as "no shaders".
Frame generation was running the fp32 shader family on hardware that supports
fp16. Measured on an Adreno 740: real frame rate fell from ~60 to ~25 with
generation on, so the output was worse than not using it at all.
The cause was copying a hardcode without its reason. ARMSX2 sets allow_fp16 and
prefer_fp16 false and says why: PCSX2's Vulkan backend never asks for
shaderFloat16 when it creates its logical device, so a module declaring the
Float16 capability would be invalid usage there -- and its own comment ends
"restore Eden's two lines if PCSX2 ever enables the feature". Eden does gate on
the device, and Eden is the one that performs. RPCS3 enables shaderFloat16 where
the driver is trusted, and clears its own flag on the Adreno drivers whose
compiler rejects native float16, so asking the device is both possible and safe.
The device in question logs "GPU/driver supports float16 data types natively".
The compat Device answers a third query now, as Eden's does. BuildShaderCache
takes the flags rather than choosing its own, because the cache header stores
them and rejects an entry whose flags differ -- the same mismatch that made a
freshly written cache read as "no shaders".
Found by measuring instead of reasoning: the CPU side of the passes costs
0.000-0.061 ms/frame, which said the cost was GPU-side and not in how the work
was sequenced, after four attempts at the sequencing.
One command buffer and one fence meant every frame waited for the immediately
previous interpolation to finish before it could reset the buffer. The GPU
pipeline drained once per frame and the CPU could never get ahead, which is a
stall no amount of moving the work around removes -- it is the depth that is
wrong, not the placement.
ARMSX2 keeps one slot per swapchain image, each with its own command buffer and
fence, and waits on the fence belonging to the slot it is about to reuse. That
fence is several frames old, so in the normal case it is already signalled and
the wait costs nothing. Three slots here, rotated per frame.
This is the last of the differences found by reading the ARMSX2 driver rather
than reasoning about my own: the copy, then the pacer's clock, then the frame's
own command buffer, and now the depth.
Discord: the emulator logo now appears beside the cover art, as it does in
ARMSX2. By URL rather than asset key -- a key that is not in the portal makes
Discord reject the entire presence update, which is what left the card blank
before. It is the launcher icon out of this repository, so it cannot drift from
the app. With no cover the logo takes the large slot instead of sending no assets
at all.
The previous commit removed the capture copy by recording Process() into the
frame's command buffer. That traded one problem for a worse one: the present path
waits on that buffer, so the interpolation's GPU cost came straight off the real
frame rate. Measured on device, 1920x1080 content ran at 24-28 real fps with
generation on, while the CPU side of the same call cost 0.013-0.061 ms -- the
work was not expensive to record, it was expensive to be waited on.
ARMSX2 runs Process in framegen's OWN command buffer, after the frame, against
the same presented image. That costs neither the copy nor the frame, and it is
what this now does: the capture side only remembers the image and lets the pacer
plan, and generate() transitions it, runs Process, runs GenerateInto per planned
frame, and submits -- one buffer, one submission, off the critical path.
Defaults were checked against ARMSX2 first and are identical (flow scale 100,
performance shaders on, target rate 0, multiplier 2), so this was never a settings
difference.
Two defects, both from grafting the passes onto the existing shape instead of
adopting the one they were written for.
A copy per frame that ARMSX2 never makes. Process() copies the source into its
chain itself (FrameGen.cpp, CopyPresentedFrame), and ARMSX2 hands it the real
swapchain image. This handed it a capture copy made by a separate full-frame blit,
so every frame paid for two 1920x1080 copies instead of one. The capture existed
only because generation happened later, from another submission, by which time the
swapchain image was gone -- so the copy was paying for a problem the ordering
created.
And the pacer planned against a clock that jittered. Plan() derives the base frame
rate from the interval between its own calls; it sat inside generate(), behind
three early returns and a conditional call site, so it measured an interval the
game never had and planned against it. That is judder by construction, and no
amount of tuning the multiplier reaches it.
Process now runs in the frame's OWN command buffer, against the presented image,
where the frame is finished and the image is still ours -- which removes the copy
and the reason for it together. The pacer plans there too, once per frame,
unconditionally. generate() keeps only the generation half: record GenerateInto
per planned frame, submit, and let queue order carry it into the blits.
The input capture images are gone entirely.
Reported as judder and lower performance against ARMSX2 and Eden running the same
passes, which was the right comparison to make.
The cache was written with one flag set and read with another. BuildShaderCache
asked for allow_fp16 = true while LsfgShaders, the only consumer, asks for
(false, false), and the header stores those flags and is rejected when they
differ -- so the cache written at import could never satisfy the read after it.
On a desktop that is invisible: the source DLL is still there, LoadShaderModules
falls through to re-parsing the executable, and the cache is dead weight rewritten
every launch. Android has nothing to fall through to, because the picked file is
a copy in app cache that the system may clear, so it surfaced as "no shaders"
with a valid cache sitting next to it. Worth telling Camille: Eden and ARMSX2 are
both re-parsing Lossless.dll every launch rather than using their cache.
The cache also survives a missing source now. It is validated against the source's
size and mtime, which is right where an install stays put and wrong here; passing
0 skips those checks, matching what source_hash and variant already do. A replaced
Lossless Scaling is no longer noticed automatically, which is the trade -- against
losing the cache to a routine cache sweep, and re-importing is explicit.
Performance: the fence wait moved from after the submit to before the next one.
Waiting at the end put the whole interpolation on the critical path -- the thread
sat idle until the GPU finished, every frame -- when nothing required it: frame
generation and the present blits share one queue, so submission order already
orders them. Waiting at the START only blocks when the previous frame's passes
have not finished in time. The queue is now taken from the device rather than a
second vkGetDeviceQueue, so that sharing is explicit rather than incidental.
Settings: motion detail is a slider rather than three stops, and performance
shaders is a switch rather than a pair of buttons.
Three defects from the switchover, all mine.
vmaCreateAllocator crashed at pc 0 the first time frame generation was switched
on. The allocator was created with physicalDevice and device but no instance and
no vulkanApiVersion; Android builds with VK_NO_PROTOTYPES, so VMA resolves its
table dynamically through vkGetInstanceProcAddr, which answers only for global
functions when the instance is null. It stored nulls and called one. memory.cpp
documents this exact failure at its own vmaCreateAllocator -- the two getters were
supplied and the instance beside them was not.
shader_count() then answered for the wrong cache, twice. First it still counted
g_shaders, which import_shaders stopped filling when it moved to the ported
extractor: generated_frame_count() multiplies by it, so generate() returned on its
first line and frame generation sat on "starting" with nothing logged, because it
never got far enough to fail. Replacing that with GetInstalledLosslessStatus() was
also wrong -- that validates the DLL PATH, and the UI copies the picked file into
app cache, which Android may purge. A re-import wrote a good 322KB shader cache and
it still said "no shaders". It now asks whether the cache LOADS, which is the
actual question: the shaders are extracted once and the source is not needed again.
The dead initialize() call in generate() is gone with it. It built the dlopen'd
library's context, nothing uses it, and a failure there could still disable the
feature.
Settings: the in-game menu offered the multiplier alone. Someone opening it
mid-game is there to answer "it is generating but the picture is unsteady" or
"it is generating but it is too expensive", and neither is the multiplier. Target
rate, flow scale and performance mode now sit beside it. Target rate is new --
0 keeps the fixed multiplier, non-zero is Camille's adaptive pacing.
Wires up the previous commit. Frame generation now runs on OUR VkDevice, in our
own present path, instead of inside a dlopen'd library with a VkDevice of its own.
What that removes: every frame used to cross between devices as an
AHardwareBuffer -- allocated, imported as a VkImage on our device so the renderer
could blit into it, handed over as a raw buffer, imported again on theirs. The
capture images are plain device-local VkImages now, because the passes read the
very images the renderer already wrote. The AHardwareBuffer external-memory
extension was also the narrowest gate on the whole feature and is no longer
required.
Kept deliberately:
- The capture itself. The swapchain image is presented and reused, so the passes
still need a stable copy; only its backing changed.
- commit_capture(), now near-vestigial. It existed because a second VkDevice had
no semaphore joining it to ours. One device and one queue means submission
order already says this, but the call site is where the capture becomes
readable and that is worth keeping named.
- The fence wait in generate(). The present path blits the generated images
immediately after, so the old contract -- return only when they are ready --
still holds. A semaphore would do it without stalling the thread and is the
obvious next step, on a path that has broken three times from partial fixes.
Frame generation records into its OWN command buffer, as the ARMSX2 driver does:
the passes have to run after the captured frame is complete and before the
generated images are blitted, and the renderer's buffer is already closed by then.
import_shaders now drives the ported extractor rather than the library. Both kept
their own copy of the shaders, and importing through the library would have looked
like it worked -- a count comes back, the settings screen agrees -- while the cache
the passes actually read stayed empty. Its failures are now phrased for the person
who chose the file.
VKPresent.cpp is untouched: generate(), generated_frame_count(), generated_image(),
capture_presented_frame() and commit_capture() all kept their signatures and their
meaning, which is the whole reason the switchover is one commit and not five.
Builds clean. NOT yet run on hardware -- neither this nor the device-feature change
under it has drawn a frame.
Compiles; not yet wired. VKFrameGen still drives the old dlopen path and nothing
calls these passes -- that is the next commit. Landing it here keeps the port and
the switchover separable, because the present path is the part that has broken
before.
The implementation is Camille LaVey's, from eden-emu PR #4263, by way of ARMSX2
2.6.6.8. She licensed it GPL-2.0-or-later for use here: ARMSX3 derives from
RPCS3, which is GPL-2.0-ONLY with no "or later" clause, so the original
GPL-3.0-or-later terms could not be carried across. Every file records that.
Eden is unaffected -- "or later" leaves its own use exactly as it was.
Why it is worth having: the current implementation runs behind a dlopen'd .so
with its OWN VkDevice, so every frame crosses devices as an AHardwareBuffer --
allocated, imported on our device, handed over, imported again on theirs. These
passes run on our device in our own present path and that round-trip disappears.
It also brings adaptive pacing: instead of a fixed x2, generate as many frames as
it takes to hold a target refresh rate, which is the answer to a game oscillating
between 60 and 30 on a 60Hz panel.
Porting notes, all of it concentrated in three seams rather than spread:
- LsfgVkCompat's Device adapter wrapped GSDeviceVK and now wraps vk::render_device.
28 of the 31 files never mention either and were not touched.
- GSConfig lives in FrameGenConfig.h, mapping RPCS3's mode enum and cfg fields to
the names the passes read, so a later merge from Eden touches passes not wiring.
- LosslessDll's platform calls (FileSystem/Path/GSXXH) became fs:: and an inline
FNV-1a; the hash only has to notice the DLL changed since the cache was written.
Two changes outside the tree were prerequisites, not cleanups:
- vkCreateDevice now ENABLES vulkanMemoryModel and robustness2's nullDescriptor
when the device reports them. Both were already probed and never enabled, and a
shader that declares the memory model on a device where it was not enabled is
invalid usage rather than a soft fallback -- desktop drivers shrug it off while
Adreno takes the device down mid-frame, so it would have looked fine until it
was not. Only nullDescriptor is turned on, not robustness2's expensive half.
- The Android Vulkan loader gained vkFreeDescriptorSets, which it never loaded.
New settings: Frame Generation Target Rate (0 = fixed multiplier) and Frame
Generation Lossless Path. Nothing ships the shaders; they are read from the
user's own legitimately purchased copy, exactly as before.
jpolo1224