Delve into the world of graphics and game development. Discuss creating stunning visuals, optimizing game mechanics, and share resources for game developers.

All subtopics
Posts under Graphics & Games topic

Post

Replies

Boosts

Views

Activity

GameKit Saved Games empty on iOS app running on Mac; iCloud reports app uninstalled
I’m investigating GameKit Saved Games synchronization between an iPhone and an iOS app running on an Apple-silicon Mac through TestFlight—not a native macOS or Mac Catalyst app. Environment: • iPhone 11, iOS 26.6.2 • Apple-silicon Mac, macOS 27.0 (26A428) • Built with Xcode 27 (27A266a), minimum iOS 16 • Same iCloud and Game Center accounts, confirmed manually • iCloud Drive enabled, including the app’s per-app syncing setting on Mac Expected: The Mac can discover and load the existing iPhone save. Observed: The iPhone’s fetchSavedGames returns two entries, one matching our expected save name, and loadData successfully reads that matching save. On the Mac, fetchSavedGames completes without an error but returns an empty array. This is the raw result before our filename filtering. Game Center authentication succeeds. Our app-level identity checks also resolve to the same game/environment/profile scope on both devices. We understand this does not independently prove the iCloud accounts match. The installed Mac app’s signed entitlements include: • Game Center • iCloud environment Production • CloudDocuments • The expected matching iCloud and ubiquity container identifiers The unusual system observation is that brctl status for this container reports: client:blocked-app-uninstalled SYNC DISABLED (app not installed) This persists while the app is installed and running. The container reports foreground when the app is open. Completed checks: Verified installed build and signed entitlements. Confirmed LaunchServices registration of the app wrapper and inner iOS bundle. Targeted re-registration, normal Mac restart, and reopening did not clear the state. In a separately triggered diagnostic inside the signed app, called FileManager.url(forUbiquityContainerIdentifier:) off the main thread with the exact entitled container identifier. That call returned a nonnil URL, but the immediate GameKit fetch and a subsequent delayed check still returned zero saves. The system block remained unchanged. The diagnostic did not write a save, restore progress, or directly read/write container files. We have not deleted app data, reset iCloud, or rewritten the source save to force synchronization. I found the authentication-timing discussion at: https://developer.apple.com/forums/thread/718541 However, our result persists across subsequent fetches and restarting the Mac, rather than only the first fetch near authentication. Questions: • Is there an additional supported configuration or initialization requirement for GameKit Saved Games in an iOS TestFlight app running on Apple silicon? • What targeted diagnostics would help explain why the container remains marked app-uninstalled despite the app being installed and running? • How should an app distinguish a genuinely empty Saved Games result from synchronization that is not yet available? We have not established that this is an Apple defect. We currently have an instrumented application, but not a separate minimal project reproducing the issue.
0
0
843
Sep ’26
macOS 27 beta: ProMotion refresh cadence is unstable, causing constant scroll judder
FB24091347 On macOS 27.0 beta (26A5388g), MacBook Pro M4 Pro, the built-in ProMotion display never settles on a stable refresh cadence. Scrolling in SwiftUI judders constantly. The same app binary was smooth on macOS 26, and is smooth on a 120 Hz ProMotion iPad. I captured two 60-second Instruments traces — same app, same scene, same scrolling, no external display — changing only the display's refresh-rate setting. On ProMotion the vsync interval standard deviation is 4.093 ms across six different cadences, mostly flip-flopping between 120 Hz and 60 Hz. Forced to a fixed 60 Hz it drops to 0.391 ms with a single cadence. The app presented an identical 59 fps median in both runs — frame production is perfectly steady, the display just holds each frame for an unpredictable length of time. That's what makes this nasty: it's invisible to every frame-rate metric, so it looks like the app got slow when nothing about the app changed. I spent most of a day profiling my own code before realising the app was never the problem. Workaround: force the built-in display to 60 Hz. Worth noting, because it complicates the picture: attaching a 60 Hz Studio Display makes the built-in smooth, but the Studio itself then judders — despite its own vsync cadence measuring perfectly stable. So refresh rate alone isn't the whole story, and there may be a second mechanism. The clean, reproducible, single-variable result is the ProMotion vs forced-60 Hz comparison on the built-in panel. If you can reproduce this on an M-series MacBook Pro on 27 beta, please file a duplicate referencing FB24091347.
6
2
2.8k
Sep ’26
RealityKit: How to read the current audio playback position, and sync audio across multiple entities?
Hi, in RealityKit, AudioPlaybackController exposes duration, gain, speed, play/pause/stop, and a completion handler, but I can't find a way to read the current playhead position while a resource is playing. I need this to trigger animations and other timed events in sync with the audio. Is there a supported way to read the current playback position on AudioPlaybackController? If not, is this planned any time soon? Also, is there a sample-accurate way to start/keep audio playback in sync across multiple entities in RealityKit? Appreciate any guidance, thanks.
1
0
1.2k
Sep ’26
MTL4FXTemporalDenoisedScaler initialization
I’m trying to use MTL4FXTemporalDenoisedScaler, and I’m seeing a crash during initialization even with a very simple sample app. I created a minimal sample here: https://github.com/tatsuya-ogawa/MetalFXInitExample The exception is: NSException: "-[AGXG16XFamilyHeap baseObject]: unrecognized selector sent to instance ..." What I found is: • This works: descriptor.makeTemporalDenoisedScaler(device: device) • This crashes: descriptor.makeTemporalDenoisedScaler(device: device, compiler: metal4Compiler) So the issue seems to happen only with the Metal4FX version. For testing, I’m using an iPhone 15 Pro. According to the Metal Feature Set Tables, MetalFX denoised upscaling should be supported on Apple9 and later, so I believe the device itself should meet the requirements. Reference: https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf Has anyone seen this before, or knows what might be causing it? I’d appreciate any advice. Thanks.
5
2
1.7k
Sep ’26
iMac gpuRestart and then crash
Hi all, This has been bothering me for quite a while. Basically my new iMac (bought for a few months only) started to crash randomly. I went to the genius bar and they couldn't do anything to identify the issue, I tried reinstalling the OS and even reinstalling an older version of Mac OS as well, but still seeing this issue. Today it happened twice and below are the details: Mac specs: Model Name: iMac  Model Identifier: iMac20,1  Processor Name: 10-Core Intel Core i9  Processor Speed: 3.6 GHz  Number of Processors: 1  Total Number of Cores: 10  L2 Cache (per Core): 256 KB  L3 Cache: 20 MB  Hyper-Threading Technology: Enabled  Memory: 16 GB  Boot ROM Version: 1554.100.64.0.0 (iBridge: 18.16.14556.0.0,0)  Serial Number (system): xxx  Hardware UUID: xxx  Activation Lock Status: Enabled The DiagnosticReports around the time it crashed has a lot of files with .gpuRestart, e.g.: Kernel_2021-04-27-213412_Zhuzengs-iMac.gpuRestart and file WindowServer_2021-04-27-213319_Zhuzengs-iMac.userspace_watchdog_timeout.spin in between. The details of the the first gpuRestart file Tue Apr 27 21:32:13 2021 Event: GPU Reset Date/Time: Tue Apr 27 21:32:13 2021 Application: Path: Tailspin: /Library/Logs/DiagnosticReports/gpuRestart2021-04-27-213213.tailspin GPUSubmission Trace ID: 0 OS Version: Mac OS X Version 10.15.7 (Build 19H1030) Graphics Hardware: AMD Radeon Pro 5300 Signature: 2 Report Data: GPU Log Version: 1 Restart Channel: 18 VMPT --THE STATE OF THE DRIVER AMDRadeonX6000_AMDNavi14GraphicsAccelerator state: ENABLED PCIe Device: [3:0:0], DID=0x7340, RID=0x47, SSID=0x219 TotalVideoRAMBytes: 0x00000000ff000000 (4278190080) Uptime 21:50:05.077572 [00] Channel: GFX, last reset at 0:00:00.000000 CompletedTS = 0x005be078, SubmittedTS = 0x005be079 SentTS = 0x005be078, sent at 21:49:00.896511, ScheduledTS = 0x005be079, submitted at 21:50:03.672539 Wait for Channel 18, TS 0xef924 PendingEvent: YES NumberOfPendingCB = 1, FirstPendingTS = 0x005be079, LastPendingTS = 0x005be079 FirstPendingCB: ProcessID = 225, ProcessName = WindowServer, SubmitContext = Unknown (0) GPUAddress = 0x0000000431cef000, Size = 0x000001d3, VMID = 2 ContentValidation = PASS Buffer range 0x0 .. 0x100:c0012800 80000000 80000000 c0026900 00000081 80000000 40004000 c0026900 By searching online this seems to be happening to others as well but I failed to find a common fix for this. Any help would be hugely appreciated!!!
3
1
1.9k
Sep ’26
Data corruption when using MTLBlitCommandEncoder.copy from buffer to texture
When we use MTLBlitCommandEncoder.copy to copy from buffer to textures with specific size, we find that the texture content is corruppt. The specific rules we found are: On Apple GPUs (reproduced on Apple M5, macOS 26.3 and iPhone 13 / A15), for ASTC textures uploaded to private storage via the blit encoder, the sampler / texel-fetch mis-addresses the tail of a mip level when ALL of the following hold: that level's texel width is an exact multiple of the 16 KB page width (32 blocks) — 128 texels for ASTC 4x4, 192 for 6x6, 256 for 8x8; the texture width is not divisible by 2^level (a partial block column exists in the chain); that level is taller than one page (>32 block rows). The level's first 128 texel rows (first page row, 4x4) are always read correctly; everything beyond is mis-addressed. Reads landing on unmapped/invalid memory decode as opaque magenta (1,0,1); reads aliasing valid memory show wrong image content. This produced visible purple artifacts in a shipping game (impostor tree atlas, 514x1024 ASTC 4x4, mip2 bottom half). The minimum reproduce code is: ASTCMipTailBugRepro.swift The output is: 514x1024 blit(private) mip2: CORRUPTED (bottom of mip2 mis-addressed) 512x1024 blit(private) mip2: INTACT 514x1024 replaceRegion mip2: INTACT
1
0
1.6k
Sep ’26
Metal 4 and object lifetime
I have a metal kit view and drain the draw method of its delegate like shown below. Let's say I have one or more MTLBuffers with vertex resources bound via the argument table. When is it ok to drop these buffers? As far as I know one cannot schedule a completion handler In Metal 4 and I haven't been able to find any documentation about the lifetime requirements here. Any pointers/ideas appreciated. class RenderCoordinator: NSObject, MTKViewDelegate { public func draw(in view: MTKView) { let commandAllocator: any MTL4CommandAllocator = ... let commandBuffer: any MTL4CommandBuffer = ... let commandQueue: any MTL4CommandQueue = ... guard let drawable = view.currentDrawable else { return } commandBuffer.beginCommandBuffer(allocator: commandAllocator) let state: any MTLRenderPipelineState = ... let encoder: any MTL4RenderCommandEncoder = ... let argTable: any MTL4ArgumentTable = ... encoder.setRenderPipelineState(state) encoder.setArgumentTable(argTable, stages: .vertex) commandBuffer.endCommandBuffer() commandQueue.waitForDrawable(drawable) commandQueue.commit([commandBuffer]) commandQueue.signalDrawable(drawable) drawable.present() } }
2
0
1.7k
Sep ’26
Times New Roman superscript and Unicode fallback font
I was using Times New Roman and trying to use superscript numbers, but found out that Times New Roman only has superscript numbers 0, 1, 2, 3, ¹ (like so) and since I'm writing out transcriptions of pinyin, I also need superscript 4, 5, 6, 7. When I do that, the font changes to Lucida Grande automatically. I would like to change the default unicode font/alt font from Lucida Grande to EB Garamond (downloaded from Google) since it more closely resembles Times New Roman and unfortunately, I need to keep the font as close to Times New Roman as possible. Is there a way to change the default the computer chooses to when Times New Roman/alt font fails unicode? Alternatively, is there a font that looks exactly like Times New Roman that has the superscript numbers 0-9? 2022 laptop, M2, Tahoe 26.6.1 Move post at will if need be.
5
0
1.5k
Sep ’26
Core Animation Background Thread CATransaction
Hey everyone 👋 I'm trying to initialize a part of a CALayer hierarchy on a background thread and then attach the root of that hierarchy to a CALayer that backs a UIView. The motivation is to keep the main thread responsive when constructing a complex layer hierarchy. This isn't a case where I'm creating two or three layers and then switching back to the main thread. The hierarchy can potentially contain a large number of layers, with animations being created/configured for those layers as well. My first approach was to create and configure the layers entirely on a background thread. While the output might be the expected one (not always), CoreAnimation emits an assertion along the lines of: "Modifications to the layer tree from a background thread may not be committed" (or something like this). This makes sense to me if implicit CATransaction is thread-local. In that case, the implicit transaction opened by the layer modification on the background thread would not be part of the transaction that is already open on the main thread. Therefore, committing the main-thread transaction would not commit the changes made on the background thread. My second approach was to explicitly create and commit a CATransaction on the background thread. This appears to be accepted by Core Animation's threading model, but I'm seeing unreliable results. Sometimes parts of the hierarchy are missing, and in other cases the hierarchy is present but the animations don't appear to run at all. I do understand that this is private behavior of the framework, but I wanted to know if what I am trying to achieve is possible and, if so, what the solution would be (obviously if you can share this information). Besides this, I would also like to know what behavior CATransactions have when they are created on different threads. What I mean by this is that the transactions work as a stack, and the changes are committed when the stack is empty. Does this behavior still apply when having transactions on different threads? Any weird behaviours that might appear between transactions operated on main vs background threads? Thank you! Vlad.
7
0
1.2k
Sep ’26
Behaviour of a 0-value `accelerationStructureID`
When constructing a MTLIndirectAccelerationStructureInstanceDescriptor, one specifies an accelerationStructureID. In the equivalents in both Vulkan and DirectX12, one can set it to zero to be an "inactive" instance. However, on Metal, this field does not appear to have any documentation, and thus it is difficult to figure out if there is similar behavior (and no other Metal documentation seems to mention this). Does setting this field to 0 (i.e. null) disable the instance? If not, is there any other way to have an equivalent effect?
5
0
2.6k
Aug ’26
Stress-testing Metal compute pipelines using Autolykos workload characteristics
I have been using a small macOS research project to exercise Metal with a workload that differs from rendering and dense machine-learning kernels. Autolykos v2 is useful for this because it combines a large, height-dependent working set with pseudo-random reads, integer-heavy hashing, sustained execution, and periodic replacement of the dataset. The project happens to be a miner, but my question here is strictly about Metal compute behaviour. On the Apple M4 system used for these measurements, the full dataset contained 216,430,305 elements of 32 bytes each: 6.93 GB, or about 6.45 GiB, held in a .storageModePrivate buffer. For every nonce, the search kernel: computes an index seed, performs 32 pseudo-random 32-byte dataset reads, accumulates the eight 32-bit limbs into a 256-bit sum, and applies a final BLAKE2b compression and target comparison. The normal dispatch uses 128 threads per threadgroup. The wider pipeline also builds the next height's dataset in chunks on a separate command queue while search continues, and keeps two search command buffers in flight. I record command-buffer wall time, gpuStartTime/gpuEndTime, unions of overlapping intervals, and thermal state. To estimate the ceiling imposed by the random gathers, I added a non-consensus microbenchmark. It retains the normal seed calculation, index distribution, all 32 dataset reads, and the complete accumulation, but omits the final BLAKE2b compression. The accumulated result remains observable through a comparison, so the gather loop cannot simply disappear. I expected this stripped kernel to be at least slightly faster. Instead, an order-balanced campaign on an M4 produced: complete search kernel: 3.108 million nonces/s median active throughput gather-only kernel: 2.952 million nonces/s ratio: 105.3% All four same-round ratios were between 103.18% and 105.74%. Each measured run used the full dataset, a 30-second search interval, an excluded warm-up, and a start-temperature gate below 50 °C. Both compute pipeline states reported maxTotalThreadsPerThreadgroup == 1024. My conservative conclusion was not to pursue speculative register-pressure or manual memory-level-parallelism rewrites. The access pattern appears sufficiently dominant, while the supposedly simpler microbenchmark may have changed the compiled pipeline in a way that makes it a poor upper-bound model. My questions are: Can removing the trailing arithmetic legitimately make a memory-latency-heavy Metal kernel slower by changing register allocation, instruction scheduling, or the amount of useful latency hiding? Or would you first suspect a flaw in this kind of gather-only benchmark construction? Also, which Metal GPU counters are the most reliable way to distinguish memory-latency saturation from register-limited occupancy in a long-running compute kernel? I am looking at compute occupancy, buffer and ALU limiters, bandwidth, and cache behaviour, but maxTotalThreadsPerThreadgroup alone is clearly too coarse to explain the result. This is one hardware-specific observation rather than a general claim about Apple GPUs. If useful, I can reduce the workload to a smaller standalone reproducer. The source code, benchmark driver, and complete campaign report are available here: https://github.com/giffeler/ergometal The detailed measurements and validation procedure for this comparison are documented here: https://github.com/giffeler/ergometal/blob/main/Benchmarks/2026-08-15-search-gather-ceiling-ab.md
0
0
798
Aug ’26
Metal-cpp usability issue with MTL::Buffer and MTL::ResidencySet
I know this might be a peeve of mine, but looking into programming a simple Metal4 Compute Shader example, essentially updating the Performing Calculations on a GPU example code to work with CPP and Metal4. I found that MTL::Allocation and MTL::Buffer pointers can't be used interchangeably when you are trying to add allocations to a MTL::ResidencySet, this is forcing you to: Know by heart that they inherit from each other and that you can just cast them (this is a bit suspicious though, it did work for me). Forcefully either C-cast or reinterpret_cast the MTL::Buffer pointer to a MTL::Allocation pointer as the MTL::ResidencySet will only accept that type. I might as well just be plain wrong about how this is used, any tips on correct usage in that case? Is there any expectation to either provide a typecast operator or add inheritance to support the expected behaviour seen in Swift and ObjC, which is just passing the thing? Opened a report with # FB24534953 with some extra information. Bear in mind that the example code uses Premake5, but it can generate an Xcode solution easily.
0
0
777
Aug ’26
Xcode 27.0 b5, macOS 26.6.1, Metal build fails: symbols not found for air64_v28
I've just downloaded the Xcode 27.0 beta 5 on a macOS 26.6.1 machine and tried to build my app (which includes Metal CoreImage kernels). I'm met with a new (to me) error; /Users/…/Developer/…/air-lld:1:1 symbol(s) not found for target 'air64_v28-apple-macosx26.0.0' and from the build log; air-lld: warning: ignoring file '/Applications/Xcode-beta.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/System/Library/Frameworks/CoreImage.framework/Versions/A/CoreImage.metallib', file AIR version (2.9) is bigger than the one of the target being linked (2.8) air-lld: error: symbol(s) not found for target 'air64_v28-apple-macosx26.0.0' metal: error: air-lld command failed with exit code 1 (use -v to see invocation) I have no MTL_LANGUAGE_REVISION in my build settings. If I add one, with value Metal41 the app builds fine, but crashes at runtime as Metal 4.1 isn't supported on macOS 26. I imagine this is a beta Xcode and or macOS SDK bug, is there a workaround?
1
1
1.4k
Aug ’26
Can Materials not assigned to entities be retrieved from a .reality file?
like: let mat = try await ShaderGraphMaterial(named: "matname", from: "reality") currently I use an USD file with materials like so and it works: try await ShaderGraphMaterial(named: "/Root/matname", from: "file.usda", in: appBundle) when i try it with .reality i get "NameNotFound". so is it possible or do i have to have a bunch of dummy entities with my materials assigned so i can find the entity>components>material? or what's the best way to author materials in RCP3 for quick access in realitykit?
3
0
2.0k
Aug ’26
Is there a working example for Fog post effect for non AR game?
i can read the depth and mix it with the rendered image and get a linear fog effect but I can't get a true radial fog. I've been going back and forth with gemini and chatgpt and neither can do it. gemini got me a radial gradient but it's jsut projected flat across all 3d objects. it's been all day trial after trial this is what gives me the flat radial gradient using namespace metal; struct DepthFogEffectConstants { float4 fogColor; float density; // We no longer need the inverse projection matrix here! }; kernel void depthFogKernel( texture2d<half, access::read> inColor [[texture(0)]], texture2d<float, access::read> inDepth [[texture(1)]], texture2d<half, access::write> outColor [[texture(2)]], constant DepthFogEffectConstants& uniforms [[buffer(0)]], uint2 gid [[thread_position_in_grid]]) { float w = outColor.get_width(); float h = outColor.get_height(); if (gid.x >= w || gid.y >= h) { return; } half4 originalColor = inColor.read(gid); float rawDepth = inDepth.read(gid).r; // 1. Guard check for empty backgrounds/skyboxes if (rawDepth <= 0.00001f || rawDepth >= 0.9999f) { outColor.write(originalColor, gid); return; } // 2. Map screen pixels from the center of the lens (-1.0 to 1.0) float2 screenPos = float2( ((float(gid.x) / w) * 2.0f) - 1.0f, 1.0f - ((float(gid.y) / h) * 2.0f) ); // 3. Since rawDepth is already acting as a view-space Z proxy, // we use it to calculate the true spherical ray distance from the lens center. // The hypotenuse of screen offset (X, Y) and depth (Z) gives the radial distance. float radialDistance = sqrt(screenPos.x * screenPos.x + screenPos.y * screenPos.y + rawDepth * rawDepth); // 4. Calculate exponential fog matching your visual test float fogFactor32 = exp(-radialDistance * uniforms.density); half fogFactor = half(clamp(fogFactor32, 0.0f, 1.0f)); half4 fogColor = half4(uniforms.fogColor); // Mix and write colors out cleanly half4 finalColor = mix(fogColor, originalColor, fogFactor); outColor.write(radialDistance, gid); } in this code i'm just outputing the radial distance to see that calculation and it's just wrong. i don't know what to do anymore
1
1
1.6k
Aug ’26
D3DMetal: crash on Cmd+Tab out of a fullscreen D3D11 app — FindClosestMatchingMode reads modes[count-1] (FB24422691)
Filed as FB24422691. Posting here as well because the crash leaves no usable stack in the application's own crash handler, and searching for "0x1BFFFFFFE4" or "FindClosestMatchingMode" returns nothing anywhere, so this may save someone else the debugging. SYMPTOM A Direct3D 11 Windows application (Unity 2021.3) running fullscreen under Game Porting Toolkit dies when you Cmd+Tab out of it and back in. Exclusive and borderless fullscreen both crash; windowed never does. It is intermittent: on my machine it takes 5 to 9 switches. Always the same signature: EXCEPTION_ACCESS_VIOLATION (0xC0000005), reading address 0x1BFFFFFFE4, faulting RIP at D3DMetal+0xFA8F. Identical in all 16 crash reports I collected, which is what makes it a deterministic underflow rather than heap corruption. CAUSE D3DMetal's DXGIOutput::FindClosestMatchingMode calls GetDisplayModeList and then reads the last element of the list without checking the count or the buffer pointer: call DXGIOutput::GetDisplayModeList(...) ; count returned in [rsp+0x3c] mov eax, [rsp+0x3c] ; eax = count dec eax ; 0xFFFFFFFF when count == 0 lea rcx, [rax + 8rax] ; rcx = 28 * eax (28 = sizeof DXGI_MODE_DESC) lea rcx, [rcx + 2rcx] add rcx, rax mov rax, [r14 + rcx] ; reads modes[-1] 28 * 0xFFFFFFFF = 0x1BFFFFFFE4, and r14 (the mode buffer) is NULL on that path, so the read lands exactly on the address seen in the crash reports. WHY THE LIST IS EMPTY winemac.drv rebuilds the display device list on every application activation. With WINEDEBUG=+display, four Cmd+Tab activations produce exactly four full rebuilds (macdrv_UpdateDisplayDevices: GPU count, adapter, monitor). An application that queries the closest matching mode mid-rebuild gets zero modes back. Fullscreen makes that query on focus changes; windowed does not, which is exactly why windowed never crashes. The empty-list condition is not unique to this race: GetDisplayModeList has also been reported returning 0 modes for DXGI_FORMAT_R16G16B16A16_FLOAT on D3DMetal 2.1 (github.com/vec715/enfusion-dxgi-fix). DXVK and DXMT both return DXGI_ERROR_NOT_FOUND for an empty list instead of dereferencing it. MEASUREMENTS Alternating applications automatically and verifying every focus change: borderless fullscreen: crash after 5, 6, 6 and 8 switches (4 of 4 runs) exclusive fullscreen: crash after 9 switches windowed: 105+ switches, no crash As a control I patched a local copy of D3DMetal so the read is skipped when count == 0 or the buffer is NULL, keeping the requested mode. Same machine, same setup: 140 switches in fullscreen with no crash, and the unpatched binary crashed again after 8 switches immediately afterwards. VERSIONS The unchecked read is present in both D3DMetal builds I have: 2.0 (built for macOS 13.3) at 0x1453A, and 3.0 (built for macOS 15.4) at 0xFA82. Environment: macOS 26.5.2 (25F84), Apple M5, D3DMetal 3.0 inside Game Porting Toolkit, Wine 7.7, 64-bit prefix, fullscreen at 2560x1664 with winemac.drv Retina mode on. WORKAROUND UNTIL IT IS FIXED Run the application windowed, or interpose a dxgi proxy DLL that returns DXGI_ERROR_NOT_FOUND from FindClosestMatchingMode when the mode count is 0. Suggested fix: return DXGI_ERROR_NOT_FOUND from FindClosestMatchingMode / FindClosestMatchingMode1 when the count is 0 or the buffer is NULL, instead of indexing modes[count-1].
0
0
283
Aug ’26
MPS bf16 softmax produces NaN on M5 Max (regression from M4) — breaks all on-device diffusion inference
Metal Performance Shaders produces from bf16/fp16 softmax for large attention tensors on , forcing the entire local generative-AI ecosystem to fall back to fp32. This is a . Environment Minimal reproduction Decomposed softmax on MPS — diffs = x - maxes produces NaN even though every element is identical (result should be all zeros): Real-world impact Running ComfyUI (the dominant local generative-AI UI) on M5 Max: Root cause (per PyTorch MPS maintainers) PyTorch maintainers (@drisspg, @albanD) have traced this to in the MPS/MPSGraph fused kernels. The NaN originates in the x - maxes subtraction inside softmax, then propagates through the attention block and the entire network. PyTorch cannot fix this — it is in the Metal/MPS kernel layer. Not isolated The M4-era fix ("fixed on macOS 15.1") did not survive onto M5, indicating the MPS fused-kernel precision fix was either reverted or not ported to the M5 GPU architecture. Request
1
0
1.1k
Aug ’26
GameKit Saved Games empty on iOS app running on Mac; iCloud reports app uninstalled
I’m investigating GameKit Saved Games synchronization between an iPhone and an iOS app running on an Apple-silicon Mac through TestFlight—not a native macOS or Mac Catalyst app. Environment: • iPhone 11, iOS 26.6.2 • Apple-silicon Mac, macOS 27.0 (26A428) • Built with Xcode 27 (27A266a), minimum iOS 16 • Same iCloud and Game Center accounts, confirmed manually • iCloud Drive enabled, including the app’s per-app syncing setting on Mac Expected: The Mac can discover and load the existing iPhone save. Observed: The iPhone’s fetchSavedGames returns two entries, one matching our expected save name, and loadData successfully reads that matching save. On the Mac, fetchSavedGames completes without an error but returns an empty array. This is the raw result before our filename filtering. Game Center authentication succeeds. Our app-level identity checks also resolve to the same game/environment/profile scope on both devices. We understand this does not independently prove the iCloud accounts match. The installed Mac app’s signed entitlements include: • Game Center • iCloud environment Production • CloudDocuments • The expected matching iCloud and ubiquity container identifiers The unusual system observation is that brctl status for this container reports: client:blocked-app-uninstalled SYNC DISABLED (app not installed) This persists while the app is installed and running. The container reports foreground when the app is open. Completed checks: Verified installed build and signed entitlements. Confirmed LaunchServices registration of the app wrapper and inner iOS bundle. Targeted re-registration, normal Mac restart, and reopening did not clear the state. In a separately triggered diagnostic inside the signed app, called FileManager.url(forUbiquityContainerIdentifier:) off the main thread with the exact entitled container identifier. That call returned a nonnil URL, but the immediate GameKit fetch and a subsequent delayed check still returned zero saves. The system block remained unchanged. The diagnostic did not write a save, restore progress, or directly read/write container files. We have not deleted app data, reset iCloud, or rewritten the source save to force synchronization. I found the authentication-timing discussion at: https://developer.apple.com/forums/thread/718541 However, our result persists across subsequent fetches and restarting the Mac, rather than only the first fetch near authentication. Questions: • Is there an additional supported configuration or initialization requirement for GameKit Saved Games in an iOS TestFlight app running on Apple silicon? • What targeted diagnostics would help explain why the container remains marked app-uninstalled despite the app being installed and running? • How should an app distinguish a genuinely empty Saved Games result from synchronization that is not yet available? We have not established that this is an Apple defect. We currently have an instrumented application, but not a separate minimal project reproducing the issue.
Replies
0
Boosts
0
Views
843
Activity
Sep ’26
macOS 27 beta: ProMotion refresh cadence is unstable, causing constant scroll judder
FB24091347 On macOS 27.0 beta (26A5388g), MacBook Pro M4 Pro, the built-in ProMotion display never settles on a stable refresh cadence. Scrolling in SwiftUI judders constantly. The same app binary was smooth on macOS 26, and is smooth on a 120 Hz ProMotion iPad. I captured two 60-second Instruments traces — same app, same scene, same scrolling, no external display — changing only the display's refresh-rate setting. On ProMotion the vsync interval standard deviation is 4.093 ms across six different cadences, mostly flip-flopping between 120 Hz and 60 Hz. Forced to a fixed 60 Hz it drops to 0.391 ms with a single cadence. The app presented an identical 59 fps median in both runs — frame production is perfectly steady, the display just holds each frame for an unpredictable length of time. That's what makes this nasty: it's invisible to every frame-rate metric, so it looks like the app got slow when nothing about the app changed. I spent most of a day profiling my own code before realising the app was never the problem. Workaround: force the built-in display to 60 Hz. Worth noting, because it complicates the picture: attaching a 60 Hz Studio Display makes the built-in smooth, but the Studio itself then judders — despite its own vsync cadence measuring perfectly stable. So refresh rate alone isn't the whole story, and there may be a second mechanism. The clean, reproducible, single-variable result is the ProMotion vs forced-60 Hz comparison on the built-in panel. If you can reproduce this on an M-series MacBook Pro on 27 beta, please file a duplicate referencing FB24091347.
Replies
6
Boosts
2
Views
2.8k
Activity
Sep ’26
Game
In duo because of this design, most games like Call of Duty or Minecraft didn’t support this
Replies
0
Boosts
0
Views
712
Activity
Sep ’26
RealityKit: How to read the current audio playback position, and sync audio across multiple entities?
Hi, in RealityKit, AudioPlaybackController exposes duration, gain, speed, play/pause/stop, and a completion handler, but I can't find a way to read the current playhead position while a resource is playing. I need this to trigger animations and other timed events in sync with the audio. Is there a supported way to read the current playback position on AudioPlaybackController? If not, is this planned any time soon? Also, is there a sample-accurate way to start/keep audio playback in sync across multiple entities in RealityKit? Appreciate any guidance, thanks.
Replies
1
Boosts
0
Views
1.2k
Activity
Sep ’26
MTL4FXTemporalDenoisedScaler initialization
I’m trying to use MTL4FXTemporalDenoisedScaler, and I’m seeing a crash during initialization even with a very simple sample app. I created a minimal sample here: https://github.com/tatsuya-ogawa/MetalFXInitExample The exception is: NSException: "-[AGXG16XFamilyHeap baseObject]: unrecognized selector sent to instance ..." What I found is: • This works: descriptor.makeTemporalDenoisedScaler(device: device) • This crashes: descriptor.makeTemporalDenoisedScaler(device: device, compiler: metal4Compiler) So the issue seems to happen only with the Metal4FX version. For testing, I’m using an iPhone 15 Pro. According to the Metal Feature Set Tables, MetalFX denoised upscaling should be supported on Apple9 and later, so I believe the device itself should meet the requirements. Reference: https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf Has anyone seen this before, or knows what might be causing it? I’d appreciate any advice. Thanks.
Replies
5
Boosts
2
Views
1.7k
Activity
Sep ’26
iMac gpuRestart and then crash
Hi all, This has been bothering me for quite a while. Basically my new iMac (bought for a few months only) started to crash randomly. I went to the genius bar and they couldn't do anything to identify the issue, I tried reinstalling the OS and even reinstalling an older version of Mac OS as well, but still seeing this issue. Today it happened twice and below are the details: Mac specs: Model Name: iMac  Model Identifier: iMac20,1  Processor Name: 10-Core Intel Core i9  Processor Speed: 3.6 GHz  Number of Processors: 1  Total Number of Cores: 10  L2 Cache (per Core): 256 KB  L3 Cache: 20 MB  Hyper-Threading Technology: Enabled  Memory: 16 GB  Boot ROM Version: 1554.100.64.0.0 (iBridge: 18.16.14556.0.0,0)  Serial Number (system): xxx  Hardware UUID: xxx  Activation Lock Status: Enabled The DiagnosticReports around the time it crashed has a lot of files with .gpuRestart, e.g.: Kernel_2021-04-27-213412_Zhuzengs-iMac.gpuRestart and file WindowServer_2021-04-27-213319_Zhuzengs-iMac.userspace_watchdog_timeout.spin in between. The details of the the first gpuRestart file Tue Apr 27 21:32:13 2021 Event: GPU Reset Date/Time: Tue Apr 27 21:32:13 2021 Application: Path: Tailspin: /Library/Logs/DiagnosticReports/gpuRestart2021-04-27-213213.tailspin GPUSubmission Trace ID: 0 OS Version: Mac OS X Version 10.15.7 (Build 19H1030) Graphics Hardware: AMD Radeon Pro 5300 Signature: 2 Report Data: GPU Log Version: 1 Restart Channel: 18 VMPT --THE STATE OF THE DRIVER AMDRadeonX6000_AMDNavi14GraphicsAccelerator state: ENABLED PCIe Device: [3:0:0], DID=0x7340, RID=0x47, SSID=0x219 TotalVideoRAMBytes: 0x00000000ff000000 (4278190080) Uptime 21:50:05.077572 [00] Channel: GFX, last reset at 0:00:00.000000 CompletedTS = 0x005be078, SubmittedTS = 0x005be079 SentTS = 0x005be078, sent at 21:49:00.896511, ScheduledTS = 0x005be079, submitted at 21:50:03.672539 Wait for Channel 18, TS 0xef924 PendingEvent: YES NumberOfPendingCB = 1, FirstPendingTS = 0x005be079, LastPendingTS = 0x005be079 FirstPendingCB: ProcessID = 225, ProcessName = WindowServer, SubmitContext = Unknown (0) GPUAddress = 0x0000000431cef000, Size = 0x000001d3, VMID = 2 ContentValidation = PASS Buffer range 0x0 .. 0x100:c0012800 80000000 80000000 c0026900 00000081 80000000 40004000 c0026900 By searching online this seems to be happening to others as well but I failed to find a common fix for this. Any help would be hugely appreciated!!!
Replies
3
Boosts
1
Views
1.9k
Activity
Sep ’26
Data corruption when using MTLBlitCommandEncoder.copy from buffer to texture
When we use MTLBlitCommandEncoder.copy to copy from buffer to textures with specific size, we find that the texture content is corruppt. The specific rules we found are: On Apple GPUs (reproduced on Apple M5, macOS 26.3 and iPhone 13 / A15), for ASTC textures uploaded to private storage via the blit encoder, the sampler / texel-fetch mis-addresses the tail of a mip level when ALL of the following hold: that level's texel width is an exact multiple of the 16 KB page width (32 blocks) — 128 texels for ASTC 4x4, 192 for 6x6, 256 for 8x8; the texture width is not divisible by 2^level (a partial block column exists in the chain); that level is taller than one page (>32 block rows). The level's first 128 texel rows (first page row, 4x4) are always read correctly; everything beyond is mis-addressed. Reads landing on unmapped/invalid memory decode as opaque magenta (1,0,1); reads aliasing valid memory show wrong image content. This produced visible purple artifacts in a shipping game (impostor tree atlas, 514x1024 ASTC 4x4, mip2 bottom half). The minimum reproduce code is: ASTCMipTailBugRepro.swift The output is: 514x1024 blit(private) mip2: CORRUPTED (bottom of mip2 mis-addressed) 512x1024 blit(private) mip2: INTACT 514x1024 replaceRegion mip2: INTACT
Replies
1
Boosts
0
Views
1.6k
Activity
Sep ’26
Metal 4 and object lifetime
I have a metal kit view and drain the draw method of its delegate like shown below. Let's say I have one or more MTLBuffers with vertex resources bound via the argument table. When is it ok to drop these buffers? As far as I know one cannot schedule a completion handler In Metal 4 and I haven't been able to find any documentation about the lifetime requirements here. Any pointers/ideas appreciated. class RenderCoordinator: NSObject, MTKViewDelegate { public func draw(in view: MTKView) { let commandAllocator: any MTL4CommandAllocator = ... let commandBuffer: any MTL4CommandBuffer = ... let commandQueue: any MTL4CommandQueue = ... guard let drawable = view.currentDrawable else { return } commandBuffer.beginCommandBuffer(allocator: commandAllocator) let state: any MTLRenderPipelineState = ... let encoder: any MTL4RenderCommandEncoder = ... let argTable: any MTL4ArgumentTable = ... encoder.setRenderPipelineState(state) encoder.setArgumentTable(argTable, stages: .vertex) commandBuffer.endCommandBuffer() commandQueue.waitForDrawable(drawable) commandQueue.commit([commandBuffer]) commandQueue.signalDrawable(drawable) drawable.present() } }
Replies
2
Boosts
0
Views
1.7k
Activity
Sep ’26
Times New Roman superscript and Unicode fallback font
I was using Times New Roman and trying to use superscript numbers, but found out that Times New Roman only has superscript numbers 0, 1, 2, 3, ¹ (like so) and since I'm writing out transcriptions of pinyin, I also need superscript 4, 5, 6, 7. When I do that, the font changes to Lucida Grande automatically. I would like to change the default unicode font/alt font from Lucida Grande to EB Garamond (downloaded from Google) since it more closely resembles Times New Roman and unfortunately, I need to keep the font as close to Times New Roman as possible. Is there a way to change the default the computer chooses to when Times New Roman/alt font fails unicode? Alternatively, is there a font that looks exactly like Times New Roman that has the superscript numbers 0-9? 2022 laptop, M2, Tahoe 26.6.1 Move post at will if need be.
Replies
5
Boosts
0
Views
1.5k
Activity
Sep ’26
Efootball mobile
Tengo un iphone 13, estoy con ios 27 ultima beta reciente y juego efootball mobile pero anda horrible. Anda muy pesado.
Replies
0
Boosts
0
Views
1.1k
Activity
Sep ’26
Core Animation Background Thread CATransaction
Hey everyone 👋 I'm trying to initialize a part of a CALayer hierarchy on a background thread and then attach the root of that hierarchy to a CALayer that backs a UIView. The motivation is to keep the main thread responsive when constructing a complex layer hierarchy. This isn't a case where I'm creating two or three layers and then switching back to the main thread. The hierarchy can potentially contain a large number of layers, with animations being created/configured for those layers as well. My first approach was to create and configure the layers entirely on a background thread. While the output might be the expected one (not always), CoreAnimation emits an assertion along the lines of: "Modifications to the layer tree from a background thread may not be committed" (or something like this). This makes sense to me if implicit CATransaction is thread-local. In that case, the implicit transaction opened by the layer modification on the background thread would not be part of the transaction that is already open on the main thread. Therefore, committing the main-thread transaction would not commit the changes made on the background thread. My second approach was to explicitly create and commit a CATransaction on the background thread. This appears to be accepted by Core Animation's threading model, but I'm seeing unreliable results. Sometimes parts of the hierarchy are missing, and in other cases the hierarchy is present but the animations don't appear to run at all. I do understand that this is private behavior of the framework, but I wanted to know if what I am trying to achieve is possible and, if so, what the solution would be (obviously if you can share this information). Besides this, I would also like to know what behavior CATransactions have when they are created on different threads. What I mean by this is that the transactions work as a stack, and the changes are committed when the stack is empty. Does this behavior still apply when having transactions on different threads? Any weird behaviours that might appear between transactions operated on main vs background threads? Thank you! Vlad.
Replies
7
Boosts
0
Views
1.2k
Activity
Sep ’26
Behaviour of a 0-value `accelerationStructureID`
When constructing a MTLIndirectAccelerationStructureInstanceDescriptor, one specifies an accelerationStructureID. In the equivalents in both Vulkan and DirectX12, one can set it to zero to be an "inactive" instance. However, on Metal, this field does not appear to have any documentation, and thus it is difficult to figure out if there is similar behavior (and no other Metal documentation seems to mention this). Does setting this field to 0 (i.e. null) disable the instance? If not, is there any other way to have an equivalent effect?
Replies
5
Boosts
0
Views
2.6k
Activity
Aug ’26
Stress-testing Metal compute pipelines using Autolykos workload characteristics
I have been using a small macOS research project to exercise Metal with a workload that differs from rendering and dense machine-learning kernels. Autolykos v2 is useful for this because it combines a large, height-dependent working set with pseudo-random reads, integer-heavy hashing, sustained execution, and periodic replacement of the dataset. The project happens to be a miner, but my question here is strictly about Metal compute behaviour. On the Apple M4 system used for these measurements, the full dataset contained 216,430,305 elements of 32 bytes each: 6.93 GB, or about 6.45 GiB, held in a .storageModePrivate buffer. For every nonce, the search kernel: computes an index seed, performs 32 pseudo-random 32-byte dataset reads, accumulates the eight 32-bit limbs into a 256-bit sum, and applies a final BLAKE2b compression and target comparison. The normal dispatch uses 128 threads per threadgroup. The wider pipeline also builds the next height's dataset in chunks on a separate command queue while search continues, and keeps two search command buffers in flight. I record command-buffer wall time, gpuStartTime/gpuEndTime, unions of overlapping intervals, and thermal state. To estimate the ceiling imposed by the random gathers, I added a non-consensus microbenchmark. It retains the normal seed calculation, index distribution, all 32 dataset reads, and the complete accumulation, but omits the final BLAKE2b compression. The accumulated result remains observable through a comparison, so the gather loop cannot simply disappear. I expected this stripped kernel to be at least slightly faster. Instead, an order-balanced campaign on an M4 produced: complete search kernel: 3.108 million nonces/s median active throughput gather-only kernel: 2.952 million nonces/s ratio: 105.3% All four same-round ratios were between 103.18% and 105.74%. Each measured run used the full dataset, a 30-second search interval, an excluded warm-up, and a start-temperature gate below 50 °C. Both compute pipeline states reported maxTotalThreadsPerThreadgroup == 1024. My conservative conclusion was not to pursue speculative register-pressure or manual memory-level-parallelism rewrites. The access pattern appears sufficiently dominant, while the supposedly simpler microbenchmark may have changed the compiled pipeline in a way that makes it a poor upper-bound model. My questions are: Can removing the trailing arithmetic legitimately make a memory-latency-heavy Metal kernel slower by changing register allocation, instruction scheduling, or the amount of useful latency hiding? Or would you first suspect a flaw in this kind of gather-only benchmark construction? Also, which Metal GPU counters are the most reliable way to distinguish memory-latency saturation from register-limited occupancy in a long-running compute kernel? I am looking at compute occupancy, buffer and ALU limiters, bandwidth, and cache behaviour, but maxTotalThreadsPerThreadgroup alone is clearly too coarse to explain the result. This is one hardware-specific observation rather than a general claim about Apple GPUs. If useful, I can reduce the workload to a smaller standalone reproducer. The source code, benchmark driver, and complete campaign report are available here: https://github.com/giffeler/ergometal The detailed measurements and validation procedure for this comparison are documented here: https://github.com/giffeler/ergometal/blob/main/Benchmarks/2026-08-15-search-gather-ceiling-ab.md
Replies
0
Boosts
0
Views
798
Activity
Aug ’26
Metal-cpp usability issue with MTL::Buffer and MTL::ResidencySet
I know this might be a peeve of mine, but looking into programming a simple Metal4 Compute Shader example, essentially updating the Performing Calculations on a GPU example code to work with CPP and Metal4. I found that MTL::Allocation and MTL::Buffer pointers can't be used interchangeably when you are trying to add allocations to a MTL::ResidencySet, this is forcing you to: Know by heart that they inherit from each other and that you can just cast them (this is a bit suspicious though, it did work for me). Forcefully either C-cast or reinterpret_cast the MTL::Buffer pointer to a MTL::Allocation pointer as the MTL::ResidencySet will only accept that type. I might as well just be plain wrong about how this is used, any tips on correct usage in that case? Is there any expectation to either provide a typecast operator or add inheritance to support the expected behaviour seen in Swift and ObjC, which is just passing the thing? Opened a report with # FB24534953 with some extra information. Bear in mind that the example code uses Premake5, but it can generate an Xcode solution easily.
Replies
0
Boosts
0
Views
777
Activity
Aug ’26
Xcode 27.0 b5, macOS 26.6.1, Metal build fails: symbols not found for air64_v28
I've just downloaded the Xcode 27.0 beta 5 on a macOS 26.6.1 machine and tried to build my app (which includes Metal CoreImage kernels). I'm met with a new (to me) error; /Users/…/Developer/…/air-lld:1:1 symbol(s) not found for target 'air64_v28-apple-macosx26.0.0' and from the build log; air-lld: warning: ignoring file '/Applications/Xcode-beta.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/System/Library/Frameworks/CoreImage.framework/Versions/A/CoreImage.metallib', file AIR version (2.9) is bigger than the one of the target being linked (2.8) air-lld: error: symbol(s) not found for target 'air64_v28-apple-macosx26.0.0' metal: error: air-lld command failed with exit code 1 (use -v to see invocation) I have no MTL_LANGUAGE_REVISION in my build settings. If I add one, with value Metal41 the app builds fine, but crashes at runtime as Metal 4.1 isn't supported on macOS 26. I imagine this is a beta Xcode and or macOS SDK bug, is there a workaround?
Replies
1
Boosts
1
Views
1.4k
Activity
Aug ’26
Can Materials not assigned to entities be retrieved from a .reality file?
like: let mat = try await ShaderGraphMaterial(named: "matname", from: "reality") currently I use an USD file with materials like so and it works: try await ShaderGraphMaterial(named: "/Root/matname", from: "file.usda", in: appBundle) when i try it with .reality i get "NameNotFound". so is it possible or do i have to have a bunch of dummy entities with my materials assigned so i can find the entity>components>material? or what's the best way to author materials in RCP3 for quick access in realitykit?
Replies
3
Boosts
0
Views
2.0k
Activity
Aug ’26
Is there a working example for Fog post effect for non AR game?
i can read the depth and mix it with the rendered image and get a linear fog effect but I can't get a true radial fog. I've been going back and forth with gemini and chatgpt and neither can do it. gemini got me a radial gradient but it's jsut projected flat across all 3d objects. it's been all day trial after trial this is what gives me the flat radial gradient using namespace metal; struct DepthFogEffectConstants { float4 fogColor; float density; // We no longer need the inverse projection matrix here! }; kernel void depthFogKernel( texture2d<half, access::read> inColor [[texture(0)]], texture2d<float, access::read> inDepth [[texture(1)]], texture2d<half, access::write> outColor [[texture(2)]], constant DepthFogEffectConstants& uniforms [[buffer(0)]], uint2 gid [[thread_position_in_grid]]) { float w = outColor.get_width(); float h = outColor.get_height(); if (gid.x >= w || gid.y >= h) { return; } half4 originalColor = inColor.read(gid); float rawDepth = inDepth.read(gid).r; // 1. Guard check for empty backgrounds/skyboxes if (rawDepth <= 0.00001f || rawDepth >= 0.9999f) { outColor.write(originalColor, gid); return; } // 2. Map screen pixels from the center of the lens (-1.0 to 1.0) float2 screenPos = float2( ((float(gid.x) / w) * 2.0f) - 1.0f, 1.0f - ((float(gid.y) / h) * 2.0f) ); // 3. Since rawDepth is already acting as a view-space Z proxy, // we use it to calculate the true spherical ray distance from the lens center. // The hypotenuse of screen offset (X, Y) and depth (Z) gives the radial distance. float radialDistance = sqrt(screenPos.x * screenPos.x + screenPos.y * screenPos.y + rawDepth * rawDepth); // 4. Calculate exponential fog matching your visual test float fogFactor32 = exp(-radialDistance * uniforms.density); half fogFactor = half(clamp(fogFactor32, 0.0f, 1.0f)); half4 fogColor = half4(uniforms.fogColor); // Mix and write colors out cleanly half4 finalColor = mix(fogColor, originalColor, fogFactor); outColor.write(radialDistance, gid); } in this code i'm just outputing the radial distance to see that calculation and it's just wrong. i don't know what to do anymore
Replies
1
Boosts
1
Views
1.6k
Activity
Aug ’26
D3DMetal: crash on Cmd+Tab out of a fullscreen D3D11 app — FindClosestMatchingMode reads modes[count-1] (FB24422691)
Filed as FB24422691. Posting here as well because the crash leaves no usable stack in the application's own crash handler, and searching for "0x1BFFFFFFE4" or "FindClosestMatchingMode" returns nothing anywhere, so this may save someone else the debugging. SYMPTOM A Direct3D 11 Windows application (Unity 2021.3) running fullscreen under Game Porting Toolkit dies when you Cmd+Tab out of it and back in. Exclusive and borderless fullscreen both crash; windowed never does. It is intermittent: on my machine it takes 5 to 9 switches. Always the same signature: EXCEPTION_ACCESS_VIOLATION (0xC0000005), reading address 0x1BFFFFFFE4, faulting RIP at D3DMetal+0xFA8F. Identical in all 16 crash reports I collected, which is what makes it a deterministic underflow rather than heap corruption. CAUSE D3DMetal's DXGIOutput::FindClosestMatchingMode calls GetDisplayModeList and then reads the last element of the list without checking the count or the buffer pointer: call DXGIOutput::GetDisplayModeList(...) ; count returned in [rsp+0x3c] mov eax, [rsp+0x3c] ; eax = count dec eax ; 0xFFFFFFFF when count == 0 lea rcx, [rax + 8rax] ; rcx = 28 * eax (28 = sizeof DXGI_MODE_DESC) lea rcx, [rcx + 2rcx] add rcx, rax mov rax, [r14 + rcx] ; reads modes[-1] 28 * 0xFFFFFFFF = 0x1BFFFFFFE4, and r14 (the mode buffer) is NULL on that path, so the read lands exactly on the address seen in the crash reports. WHY THE LIST IS EMPTY winemac.drv rebuilds the display device list on every application activation. With WINEDEBUG=+display, four Cmd+Tab activations produce exactly four full rebuilds (macdrv_UpdateDisplayDevices: GPU count, adapter, monitor). An application that queries the closest matching mode mid-rebuild gets zero modes back. Fullscreen makes that query on focus changes; windowed does not, which is exactly why windowed never crashes. The empty-list condition is not unique to this race: GetDisplayModeList has also been reported returning 0 modes for DXGI_FORMAT_R16G16B16A16_FLOAT on D3DMetal 2.1 (github.com/vec715/enfusion-dxgi-fix). DXVK and DXMT both return DXGI_ERROR_NOT_FOUND for an empty list instead of dereferencing it. MEASUREMENTS Alternating applications automatically and verifying every focus change: borderless fullscreen: crash after 5, 6, 6 and 8 switches (4 of 4 runs) exclusive fullscreen: crash after 9 switches windowed: 105+ switches, no crash As a control I patched a local copy of D3DMetal so the read is skipped when count == 0 or the buffer is NULL, keeping the requested mode. Same machine, same setup: 140 switches in fullscreen with no crash, and the unpatched binary crashed again after 8 switches immediately afterwards. VERSIONS The unchecked read is present in both D3DMetal builds I have: 2.0 (built for macOS 13.3) at 0x1453A, and 3.0 (built for macOS 15.4) at 0xFA82. Environment: macOS 26.5.2 (25F84), Apple M5, D3DMetal 3.0 inside Game Porting Toolkit, Wine 7.7, 64-bit prefix, fullscreen at 2560x1664 with winemac.drv Retina mode on. WORKAROUND UNTIL IT IS FIXED Run the application windowed, or interpose a dxgi proxy DLL that returns DXGI_ERROR_NOT_FOUND from FindClosestMatchingMode when the mode count is 0. Suggested fix: return DXGI_ERROR_NOT_FOUND from FindClosestMatchingMode / FindClosestMatchingMode1 when the count is 0 or the buffer is NULL, instead of indexing modes[count-1].
Replies
0
Boosts
0
Views
283
Activity
Aug ’26
MPS bf16 softmax produces NaN on M5 Max (regression from M4) — breaks all on-device diffusion inference
Metal Performance Shaders produces from bf16/fp16 softmax for large attention tensors on , forcing the entire local generative-AI ecosystem to fall back to fp32. This is a . Environment Minimal reproduction Decomposed softmax on MPS — diffs = x - maxes produces NaN even though every element is identical (result should be all zeros): Real-world impact Running ComfyUI (the dominant local generative-AI UI) on M5 Max: Root cause (per PyTorch MPS maintainers) PyTorch maintainers (@drisspg, @albanD) have traced this to in the MPS/MPSGraph fused kernels. The NaN originates in the x - maxes subtraction inside softmax, then propagates through the attention block and the entire network. PyTorch cannot fix this — it is in the Metal/MPS kernel layer. Not isolated The M4-era fix ("fixed on macOS 15.1") did not survive onto M5, indicating the MPS fused-kernel precision fix was either reverted or not ported to the M5 GPU architecture. Request
Replies
1
Boosts
0
Views
1.1k
Activity
Aug ’26