Skip to content

Commit a640dff

Browse files
winnerspiroswinnerspiros
authored
Merge pull request #293 from winnerspiros/copilot/fix-vulkan-rendering-issues
Bump osu-framework to 2026.503.4, Oboe 1× burst buffer, pre-GC collect on session entry
2 parents 38adcc2 + 6331e8e commit a640dff

6 files changed

Lines changed: 108 additions & 92 deletions

File tree

‎osu.Android.props‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -99,7 +99,7 @@
9999
</PropertyGroup>
100100

101101
<ItemGroup>
102-
<PackageReference Include="ppy.osu.Framework.Android" Version="2026.503.1" />
102+
<PackageReference Include="ppy.osu.Framework.Android" Version="2026.503.4" />
103103
<!-- `ppy.osu.Framework.NativeLibs` is a transitive dependency of `ppy.osu.Framework`
104104
that ships desktop-only natives (Linux/macOS/Windows) under `runtimes/<rid>/native/`
105105
— including a bare Linux `libbass.so`/`libbass_fx.so`/`libbassmix.so` for linux-arm64.

‎osu.Android/Native/oboe_bridge.cpp‎

Lines changed: 16 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -147,9 +147,22 @@ bool OboeBridge::open(int32_t sampleRate) {
147147
// This allows the system to prioritize our audio thread for stable low latency.
148148
stream_->setPerformanceHintEnabled(true);
149149

150-
// Set buffer size to 2x burst size for initial stability.
151-
// LatencyTuner will then attempt to shrink it to 1x burst if stable.
152-
stream_->setBufferSizeInFrames(stream_->getFramesPerBurst() * 2);
150+
// Start at the minimum possible buffer: exactly 1× burst.
151+
//
152+
// On devices with AAudio MMAP support (Pixel 3+, Snapdragon 8 Gen 1+, most
153+
// modern Android), the MMAP path writes directly to the hardware ring buffer.
154+
// Starting at 1× burst achieves the minimum possible end-to-end audio latency
155+
// immediately — no convergence period needed.
156+
//
157+
// Previously we started at 2× burst and relied on LatencyTuner to shrink it
158+
// over ~512ms (128 callbacks × 4ms/callback at 48kHz/192-frame burst).
159+
// That delay meant users experienced ~8ms extra audio latency for the first
160+
// half-second of every gameplay session.
161+
//
162+
// LatencyTuner is still active and will automatically increase the buffer
163+
// if underruns occur (backing off to 2× or more as needed), so stability
164+
// is not compromised on devices that cannot sustain 1× burst.
165+
stream_->setBufferSizeInFrames(stream_->getFramesPerBurst());
153166

154167
// Initialise LatencyTuner for dynamic buffer management.
155168
tuner_ = std::make_unique<oboe::LatencyTuner>(*stream_);

‎osu.Android/OsuGameAndroid.cs‎

Lines changed: 63 additions & 85 deletions
Original file line numberDiff line numberDiff line change
@@ -399,35 +399,32 @@ protected override void LoadComplete()
399399
// out CPU pinning and the next iteration can target the next suspect with
400400
// the heartbeat data captured below.
401401
//
402-
// Vulkan-renderer override: when the user has selected Vulkan, ALSO skip
403-
// big-core affinity pinning of the Draw + Input threads, and ALSO skip
404-
// the LITTLE-core affinity pin of background workers further below
405-
// (only the renice-to-zero pass runs). Rationale: the Adreno / Mali /
406-
// Xclipse Vulkan driver spawns its own internal worker pool during
407-
// vkCreateInstance / vkCreateSwapchainKHR, and those workers are NOT
408-
// in our keep-alone list (we only know a subset of vendor-specific
409-
// comm names). Pinning the Draw thread to a fixed 5-core subset while
410-
// the driver workers are simultaneously demoted to the inverse 3-core
411-
// LITTLE subset reliably stalls vkQueuePresentKHR — the field-observed
412-
// failure mode is exactly "Update tick 1, Draw tick 0" with the Draw
413-
// thread blocked inside Veldrid's swapchain present queue and never
414-
// reaching the post-LoadComplete heartbeat lambda. On OpenGL/ANGLE the
415-
// same pinning is harmless (ANGLE is single-threaded on the GL driver
416-
// side) — hence the renderer-conditional override here rather than a
417-
// blanket policy change.
402+
// Vulkan background-worker affinity note: the LITTLE-core affinity pin
403+
// of background workers is still skipped for Vulkan (littleMask = 0
404+
// below). The Adreno / Mali / Xclipse driver spawns internal worker
405+
// threads whose comm names are not in our keep-alone list; if we push
406+
// all unknown workers to the LITTLE subset those driver threads end up
407+
// on slow cores and stall vkQueuePresentKHR. Renice-to-zero only is the
408+
// correct policy for background workers on Vulkan.
418409
//
419-
// The framework-side root cause (vkAcquireNextImageKHR with no timeout
420-
// → indefinite block when the swapchain is in a transient lost state)
421-
// lives in winnerspiros/veldrid `copilot/fix-vkacquirenextimage-deadlock`
422-
// and needs to be published as a new ppy.Veldrid NuGet for upstream
423-
// consumption. These overrides are the largest application-layer
424-
// mitigation we can apply until that lands.
410+
// Draw + Input threads CAN now be pinned to big cores, for two reasons:
411+
// 1. The workers are NOT pushed to little cores (littleMask = 0), so
412+
// Adreno driver threads remain free to run on any core. The original
413+
// stall was specifically the combination of Draw-on-big + workers-
414+
// on-little; with only the Draw pin active the driver workers are
415+
// unaffected.
416+
// 2. Veldrid now has a 100 ms bounded vkAcquireNextImageKHR timeout
417+
// (since ppy.osu.Framework 2026.503.1). Any residual contention
418+
// is capped to one 100 ms stall rather than an indefinite hang.
419+
// Pinning Draw to big cores significantly improves GPU command-recording
420+
// throughput and texture-upload burst performance — the primary cause of
421+
// the 35-40 fps observed in steady-state Vulkan gameplay.
425422
bool vulkanConfigured = false;
426423
try { vulkanConfigured = LogManagement.IsVulkanConfigured(); }
427424
catch (Exception e) { Debug.WriteLine($"[osu!] IsVulkanConfigured probe failed: {e.Message}"); }
428425

429426
if (vulkanConfigured)
430-
Logger.Log("[osu!] Vulkan renderer detected from framework.ini — backing off Draw/Input big-core pinning and background-worker LITTLE-core pinning to keep Adreno/Mali driver workers schedulable.", LoggingTarget.Performance);
427+
Logger.Log("[osu!] Vulkan renderer detected from framework.ini — pinning Draw/Input to big cores (worker LITTLE-core pin still skipped to keep Adreno/Mali driver workers schedulable).", LoggingTarget.Performance);
431428

432429
int affinityMask;
433430

@@ -477,11 +474,10 @@ protected override void LoadComplete()
477474
// elevation is what causes the inversion. Default SDL-set priorities are
478475
// sufficient and match upstream osu! / osu-framework behaviour.
479476

480-
// On Vulkan, do NOT pin the Draw or Input threads to big cores — see the
481-
// top-of-LoadComplete comment. The Update thread alone keeps its pin (it
482-
// doesn't directly contend with the GPU driver workers) so we still benefit
483-
// from kernel-clock stability on the game-loop tick.
484-
int mask = vulkanConfigured ? 0 : affinityMask;
477+
// Pin Draw + Input to big cores on all renderers.
478+
// For Vulkan, see the comment above: workers are NOT pushed to little
479+
// cores, so Adreno driver threads remain schedulable on any core.
480+
int mask = affinityMask;
485481

486482
if (mask != 0)
487483
{
@@ -594,21 +590,15 @@ protected override void LoadComplete()
594590
// Always select the highest refresh rate on startup, regardless of performance mode.
595591
// This ensures 120Hz+ displays are used at their native rate.
596592
//
597-
// Deferred by 5 s after LoadComplete so the initial display-mode change runs
598-
// AFTER the Vulkan swapchain has stabilised, the loader screen is up, and the
599-
// first burst of texture uploads (Toolbar et al.) has drained off the Draw
600-
// thread. On Samsung One UI / Adreno panels, writing PreferredDisplayModeId
601-
// and Surface.SetFrameRate during the cold-start swapchain bring-up can force
602-
// a non-seamless mode change that destroys the SurfaceView and stalls
603-
// vkAcquireNextImageKHR on the Draw thread; Update keeps ticking (so neither
604-
// the managed nor the native watchdog ever dumps), the screen never updates,
605-
// and ~10 s later Android raises a MotionEvent input-dispatch ANR — the
606-
// exact "cold-start black screen, no sound, no touch, ANR" pattern observed
607-
// in field reports. Deferring the initial call moves the mode change
608-
// out of the cold-start critical window; user-driven changes via the
609-
// SelectedDisplayRefreshRate dropdown and OnConfigurationChanged (DeX
610-
// connect/disconnect, rotation) remain immediate because they happen long
611-
// after the swapchain has settled.
593+
// Deferred by 5 s after LoadComplete so the initial Surface.setFrameRate call
594+
// runs AFTER the Vulkan swapchain has stabilised and the first burst of texture
595+
// uploads (Toolbar et al.) has drained off the Draw thread.
596+
//
597+
// Note: applyDisplayMode no longer writes window.Attributes.PreferredDisplayModeId
598+
// (see that method's comment). Previously that write was the main reason for the
599+
// cold-start ANR (non-seamless SurfaceView destruction mid-swapchain); the delay
600+
// is retained as a safety margin for Surface.setFrameRate even though its
601+
// ONLY_IF_SEAMLESS flag makes surface destruction unlikely.
612602
//
613603
// Under crash-loop safe-mode (previous launch died during startup) the delay
614604
// is extended to 15 s so a slow-loading device that needed >5 s to drain
@@ -1357,55 +1347,43 @@ private void applyRefreshRate(int targetHz)
13571347

13581348
private void applyDisplayMode(global::Android.Views.Display display, global::Android.Views.Display.Mode mode)
13591349
{
1360-
var window = gameActivity.Window;
1361-
1362-
if (window == null)
1363-
return;
1364-
13651350
gameActivity.RunOnUiThread(() =>
13661351
{
13671352
try
13681353
{
1369-
if (window.Attributes is WindowManagerLayoutParams layoutParams)
1370-
{
1371-
layoutParams.PreferredDisplayModeId = mode.ModeId;
1372-
window.Attributes = layoutParams;
1373-
currentRefreshRate = (int)mode.RefreshRate;
1374-
1375-
// Set frame rate at the surface level for better compositor scheduling.
1376-
// FRAME_RATE_COMPATIBILITY_FIXED_SOURCE tells Android we render at a
1377-
// fixed rate; CHANGE_FRAME_RATE_ONLY_IF_SEAMLESS restricts the request
1378-
// to mode changes the platform can perform without blanking the display
1379-
// and recreating the SurfaceView's backing buffers.
1380-
//
1381-
// We previously passed CHANGE_FRAME_RATE_ALWAYS, which permits the
1382-
// compositor to perform a non-seamless transition. On Samsung One UI /
1383-
// Adreno panels that path momentarily destroys the SurfaceView and
1384-
// invalidates the active VkSurfaceKHR; if it lands while the Draw
1385-
// thread is mid-swapchain (e.g. during the cold-start texture-upload
1386-
// burst), vkAcquireNextImageKHR can stall the present queue
1387-
// indefinitely. Update keeps ticking (heartbeats fire, neither the
1388-
// managed nor the native watchdog ever dumps), the screen never
1389-
// updates, and ~10 s later Android raises a MotionEvent input-dispatch
1390-
// ANR — the "cold-start black screen, no sound, no touch, ANR" pattern
1391-
// observed in field reports. The seamless-only restriction keeps the
1392-
// 120 Hz request honoured when the panel can do it without a surface
1393-
// tear, and silently no-ops otherwise; either outcome is visually
1394-
// unchanged but the swapchain stays alive.
1395-
try
1396-
{
1397-
var surface = gameActivity.GetSurface()?.Holder?.Surface;
1354+
currentRefreshRate = (int)mode.RefreshRate;
13981355

1399-
if (surface != null && surface.IsValid)
1400-
surface.SetFrameRate(mode.RefreshRate, FRAME_RATE_COMPATIBILITY_FIXED_SOURCE, CHANGE_FRAME_RATE_ONLY_IF_SEAMLESS);
1401-
}
1402-
catch
1403-
{
1404-
// Surface.SetFrameRate may not be available on all binding versions.
1405-
}
1356+
// Request the refresh rate via Surface.setFrameRate() ONLY.
1357+
//
1358+
// We deliberately do NOT touch window.Attributes.PreferredDisplayModeId.
1359+
// Setting PreferredDisplayModeId asks the compositor to switch the display
1360+
// to a specific hardware mode. On Samsung One UI / Adreno devices this
1361+
// triggers a non-seamless transition even when CHANGE_FRAME_RATE_ONLY_IF_SEAMLESS
1362+
// is passed to SetFrameRate — it momentarily destroys the SurfaceView and
1363+
// invalidates the active VkSurfaceKHR. When that surface loss lands while
1364+
// the Draw thread is mid-render it causes the Veldrid swapchain to enter
1365+
// its surface-lost recovery path, producing visual corruption (multiple
1366+
// overlaid layers, missing textures, tiled frames) and a sustained FPS
1367+
// drop until the swapchain is fully rebuilt.
1368+
//
1369+
// Surface.setFrameRate(FIXED_SOURCE, ONLY_IF_SEAMLESS) is the correct
1370+
// API on Android 11+ (minSdkVersion=33) for requesting a refresh-rate
1371+
// change: the platform honours it without a surface tear when possible
1372+
// and silently no-ops when a seamless switch isn't available — the
1373+
// swapchain is never touched either way.
1374+
try
1375+
{
1376+
var surface = gameActivity.GetSurface()?.Holder?.Surface;
14061377

1407-
Logger.Log($"[osu!] Display mode applied: {mode.RefreshRate}Hz (mode {mode.ModeId}, {mode.PhysicalWidth}x{mode.PhysicalHeight})", LoggingTarget.Performance);
1378+
if (surface != null && surface.IsValid)
1379+
surface.SetFrameRate(mode.RefreshRate, FRAME_RATE_COMPATIBILITY_FIXED_SOURCE, CHANGE_FRAME_RATE_ONLY_IF_SEAMLESS);
14081380
}
1381+
catch
1382+
{
1383+
// Surface.SetFrameRate may not be available on all binding versions.
1384+
}
1385+
1386+
Logger.Log($"[osu!] Display mode applied: {mode.RefreshRate}Hz (mode {mode.ModeId}, {mode.PhysicalWidth}x{mode.PhysicalHeight})", LoggingTarget.Performance);
14091387
}
14101388
catch (Exception e)
14111389
{

‎osu.Android/Performance/AndroidHighPerformanceSessionManager.cs‎

Lines changed: 25 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -48,6 +48,31 @@ private void enterSession()
4848
if (!gcLatencyModeSupported)
4949
return;
5050

51+
// Pre-drain accumulated garbage before entering the low-latency window.
52+
// SustainedLowLatency suppresses Gen2 (major) GC, so any garbage already
53+
// on the heap will persist for the entire session. A non-blocking hint here
54+
// asks the runtime to schedule a collection immediately — the call returns
55+
// in microseconds and the GC runs in background. On .NET runtimes that
56+
// support it, this eliminates the most common source of a multi-frame GC
57+
// stall right at the start of gameplay (the "first-note hitbox miss"
58+
// symptom observed across multiple field sessions).
59+
//
60+
// GCCollectionMode.Optimized + blocking:false requires .NET Core 3.0+ / .NET 5+.
61+
// On Mono (older .NET for Android runtimes) it throws NotSupportedException,
62+
// and on some niche OEM runtimes it may throw PlatformNotSupportedException.
63+
// The catch-all deliberately swallows these: the call is a best-effort hint
64+
// and the cost of it failing is exactly zero (the code path below proceeds
65+
// identically).
66+
try
67+
{
68+
GC.Collect(GC.MaxGeneration, GCCollectionMode.Optimized, blocking: false);
69+
}
70+
catch
71+
{
72+
// Non-critical hint; intentionally swallows NotSupportedException /
73+
// PlatformNotSupportedException on older or non-.NET-Core runtimes.
74+
}
75+
5176
try
5277
{
5378
originalGCMode = GCSettings.LatencyMode;

‎osu.Game/osu.Game.csproj‎

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -38,10 +38,10 @@
3838
<IncludeAssets>runtime; build; native; contentfiles; analyzers; buildtransitive</IncludeAssets>
3939
</PackageReference>
4040
<PackageReference Include="Realm" Version="20.1.0" />
41-
<PackageReference Include="ppy.osu.Framework" Version="2026.503.1" />
41+
<PackageReference Include="ppy.osu.Framework" Version="2026.503.4" />
4242
<!--
4343
Explicitly pin `ppy.Veldrid.SPIRV` to the winnerspiros fork build that
44-
`ppy.osu.Framework 2026.503.1` was compiled against.This version is the only
44+
`ppy.osu.Framework 2026.503.4` was compiled against. This version is the only
4545
one whose `runtimes/android-arm64/native/libveldrid-spirv.so` is aligned to 16 KB
4646
pages (required by Android 16+). It lives only as a release asset on
4747
<https://github.com/winnerspiros/veldrid-spirv/releases/tag/1.0> and is vendored

‎osu.iOS.props‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -33,6 +33,6 @@
3333
<SuppressTrimAnalysisWarnings>true</SuppressTrimAnalysisWarnings>
3434
</PropertyGroup>
3535
<ItemGroup>
36-
<PackageReference Include="ppy.osu.Framework.iOS" Version="2026.503.1" />
36+
<PackageReference Include="ppy.osu.Framework.iOS" Version="2026.503.4" />
3737
</ItemGroup>
3838
</Project>

0 commit comments

Comments
 (0)