What evidence exists that it's safe to call nextDrawable() on CAMetalLayer off the main thread? I have seen developers claiming that it's OK, but the official docs are silent on the topic. Attempting to do so with Strict Concurrency Checking set to Complete complains that CAMetalLayer is not @Sendable. I want to call it off the main thread since there doesn't seem to be any way to prevent it from blocking the UI for up to a second. I have read hints and allegations that this won't happen if you avoid asking for too many drawables, but that doesn't seem to be true 100% of the time in my experience. Supposing it is allowed, I wonder how races are handled such as when the layer's size is changed on the main thread, or if the layer is removed from the layer hierarchy.

I am trying to use the SVGF denoiser to denoise my ray traced shadows (and also other textures later). I do get a smoothed image, but with wonky denoising. I need the depth-normal textures and motion textures for the SVGF and assume that these are badly filled in my case. However, neither in the above linked documentation nor in the WWDC19 video I find how they should be defined. I am looking to answers to: Is depth in red or alpha channel for the depth-normal texture? Are the normals in screen space? Is depth linear? Is it distance or z coordinate in view space? Or even logarithmically scaled or something else? Are the motion vectors supposed to be in pixels per frame? What is the orientation of the axis? Is y up or down? Are there are other restrictions on the formats? Also the linked code did not help me (I have not found any SVGF so far; also all the code is in Objective-C++, not Swift, but that's a different topic). So how should I fill these textures. Can someone point me to the documentation where these kinds of questions are answered?

Hello! I need to "draw" a set of particles into the texture. It would be trivial in render encoder of course. However, I would like to implement the task in compute kernel. Every particle draw operation is expected to set 5 texels - "center" one and left/right/upper/lower. Particles can and will overlap, so concurrent draws are to be expected. I tried using texture atomics - atomic_store() to be more precise. This worked, albeit pretty slowly - too slow for my purpose. Just to test what would happen, I tried using normal texture write(). I was expecting to see some kind of visual artefacts, but to my surprise, it worked very well (and much faster). My question: is it safe? I understand that calling write() doesn't guarantee any ordering of the operations, so if multiple threads write to the same texel, the final value may come from any of those threads. But suppose all the threads were to write the very same color? Can I assume that the texel in question will have said color after the compute kernel finishes? I am using M2 Pro MacBook, but ideally I would love to get the answer for the all Apple Silicon devices. My texture format is R32Int (so as to be able to use atomics), but I could do with any single-channel format, the purpose of the texture is to be binary mask of sorts. Thanks!

Hello! I'm currently porting a videogame console emulator to iOS and I'm trying to make the renderer (tested on MacOS) work on iOS as well. The emulator core is written in C++ and uses metal-cpp for rendering, whereas the iOS frontend is written in Swift with SwiftUI. I have an Objective-C++ bridging header for bridging the Swift and C++ sides. On the Swift side, I create an MTKView. Inside the MTKView delegate, I run the emulator for 1 video frame and pass it the view's backing layer for it to render the final output image with. The emulator runs and returns, but when it returns I get a crash in Swift land (callstack attached below), inside objc_release, which indicates I'm doing something wrong with memory management. My bridging interface (ios_driver.h): #pragma once #include <Foundation/Foundation.h> #include <QuartzCore/QuartzCore.h> void iosCreateEmulator(); void iosRunFrame(CAMetalLayer* layer); Bridge implementation (ios_driver.mm): #import <Foundation/Foundation.h> extern "C" { #include "ios_driver.h" } <...> #define IOS_EXPORT extern "C" __attribute__((visibility("default"))) std::unique_ptr<Emulator> emulator = nullptr; IOS_EXPORT void iosCreateEmulator() { ... } // Runs 1 video frame of the emulator and IOS_EXPORT void iosRunFrame(CAMetalLayer* layer) { void* layerBridged = (__bridge void*)layer; // Pass the CAMetalLayer to the emulator emulator->getRenderer()->setMTKLayer(layerBridged); // Runs the emulator for 1 frame and renders the output image using our layer emulator->runFrame(); } My MTKView delegate: class Renderer: NSObject, MTKViewDelegate { var parent: ContentView var device: MTLDevice! init(_ parent: ContentView) { self.parent = parent if let device = MTLCreateSystemDefaultDevice() { self.device = device } super.init() } func mtkView(_ view: MTKView, drawableSizeWillChange size: CGSize) {} func draw(in view: MTKView) { var metalLayer = view.layer as! CAMetalLayer // Run the emulator for 1 frame & display the output image iosRunFrame(metalLayer) } } Finally, the emulator's render function that interacts with the layer: void RendererMTL::setMTKLayer(void* layer) { metalLayer = (CA::MetalLayer*)layer; } void RendererMTL::display() { CA::MetalDrawable* drawable = metalLayer->nextDrawable(); if (!drawable) { return; } MTL::Texture* texture = drawable->texture(); <rest of rendering follows here using the drawable & its texture> } This is the Swift callstack at the time of the crash: To my understanding, I shouldn't be violating ARC rules as my bridging header uses CAMetalLayer* instead of void* and Swift will automatically account for ARC when passing CoreFoundation objects to Objective-C. However I don't have any other idea as to what might be causing this. I've been trying to debug this code for a couple of days without much success. If you need more info, the emulator code is also on Github Metal renderer: https://github.com/wheremyfoodat/Panda3DS/blob/ios/src/core/renderer_mtl/renderer_mtl.cpp#L58-L68 Bridge implementation: https://github.com/wheremyfoodat/Panda3DS/blob/ios/src/ios_driver.mm Bridging header: https://github.com/wheremyfoodat/Panda3DS/blob/ios/include/ios_driver.h Any help is more than appreciated. Thank you for your time in advance.

I'm trying to use MTLBinaryArchive. I collected a BinaryArchive from one device and used metal-tt to translate it for all supported iPhone devices, ranging from iPhone 7 Plus to iPhone 16. However, this BinaryArchive is quite large, around 1.5GB uncompressed, and about 500MB compressed in the IPA. I'm wondering how to address the size issue. I watched the WWDC 2022 video, which mentioned that the operating system or app installation process would handle compatibility. Does this compatibility support different GPU chips? I tried installing an IPA with a BinaryArchive collected only from an iPhone 12 on an iPhone 13, but the BinaryArchive didn't take effect. I also saw that Apple supports App Thinning. However, it seems that resources in the Asset Catalog cannot be accessed via URL, and creating an MTLBinaryArchive requires a URL. Is it possible for MTLBinaryArchive to be distributed through App Thinning? The WWDC 2022 video also mentioned using the -Os optimization flag to reduce size. Can this give an estimate of how much compression it would achieve? Are there any methods to solve the BinaryArchive size issue without impacting performance?

Hello! I have a question about how thread groups work with tile shading. When running "traditional" compute, I get to choose both thread group size and the grid size. However, when using tile shading kernel I only have dispatchThreadsPerTile method - this controls how many threads will be ran in each tile. So far so good, but what about thread groups? The examples in video "Tile Shading on A11" seem to suggest that there will be only one thread group per tile. In the video, [[thread_index_in_threadgroup]] is called "local_id" and it is used to access the image block. I assume this is the default configuration. So when one does the following: Creates MTLRenderPassDescriptor with tileWidth set to W and tileHeight set to H Fires up the tile shading kernel using dispatchThreadsPerTile with MTLSize size = { W, H, 1 } I understand that the result is 1-to-1 mapping between the tile "pixels" and kernel threads. Now, what I would like to do is to have more than one thread group there. I want this for performance reasons: I have a certain compute kernel which I know executes very well with small thread group size. In fact, { 32, 1, 1 } seems to be the fastest. My understanding is that even if I set tile size to 16x16, and so I am executing 256 threads there, there will only be one SIMD group active in a thread group. Meaning that this SIMD group has to execute 8 times over the tile. Is it possible somehow? Or perhaps the limitations of the API are pointing at the limitations of hardware itself, and if I want to execute with SIMD group sized thread groups I have to use "traditional" compute encoder? Will be grateful for help. Michał

Hi, I am working with a large project. We are compiling each material to its own .metallib. They all include many common files full of inline functions. Finally we link it all together at the end with a single big pathtrace kernel. Everything works as expected, however the compile times have gotten completely out of hand and it takes multiple minutes to compile at runtime (to native code). I have gathered that I can do this offline by using metal-tt however if I am wondering if there is a way to reduce the compile times in such a scenario, and how to investigate what the root cause of the problem is. I suspect it could have to do with the fact that every materials metallib contains duplications of all the inline functions. Any ideas on how to profile and debug this? Thanks, Rasmus

The code is pretty simple kernel void naive( constant RunParams *param [[ buffer(0) ]], const device float *A [[ buffer(1) ]], // [N, K] device float *output [[ buffer(2) ]], uint2 gid [[ thread_position_in_grid ]]) { uint a_ptr = gid.x * param->K; for (uint i = 0; i < param->K; i++, a_ptr++) { val += A[b_ptr]; } output[ptr] = val; } when uint a_ptr = gid.x * param->K, the code got 150 GFLops when uint a_ptr = gid.y * param->K, the code got 860 GFLops param->K = 256; thread per group: [16, 16] I'd like to understand why the performance is so different, and how can I profile/diagnose this to help with further optimization.

I have run into an issue where I am trying to use atomic_float in a swift package but I cannot get things to compile because it appears that the Swift Package Manager doesn't support Metal 3 (atomic_float is Metal 3 functionality). Is there any way around this? I am using // swift-tools-version: 6.1 and my Metal code includes: #include <metal_stdlib> #include <metal_geometric> #include <metal_math> #include <metal_atomic> using namespace metal; kernel void test(device atomic_float* imageBuffer [[buffer(1)]], uint id [[ thread_position_in_grid ]]) { } But I get an error on the definition of atomic_float . Any help, one more importantly, where I could have found this information about this limitation, would be helpful. -RadBobby

Hello! I'm a developer working on a plugin for the Elgato Stream Deck, called GPU Metrics. The plugin currently only works on Windows but I'd like to bring it to macOS. However, based on forum posts I've read (and StackOverflow) there isn't a very clear path to query GPU metrics like usage, temperature, used GPU memory, and power consumption. There are some tools out there that do similar things, but I wanted to see what would be the recommendation from Apple's engineering team to get this data via a public API. Requirements: Access GPU utilization, temperature, memory usage, power usage C/C++ based API for querying the metrics so I can expose the data to JavaScript via Node Addon No need to compatibile with Intel-based Macs, as Apple silicon will be fine for now Plugin GitHub Thank you! Noah

hi When analyzing our game using Instruments, I've always been confused about the two items "Drawable Present" and "Drawable Presented" in the GPU column. The timing of Drawable Present seems to be when the CPU layer calls commandbuffer:present, rather than when the actual encoding is completed on the GPU. Also, what does drawable presented specifically mean? In our case, when a CPU stall occurs, it appears that the vsync interval changes in the next frame, and a surface that has already been calculated is not displayed. Why is this happening?

Context I’m deploying large language models on iPhone using llama.cpp. A new iPhone Air (12 GB RAM) reports a Metal MTLDevice.recommendedMaxWorkingSetSize of 8,192 MB, and my attempt to load Llama-2-13B Q4_K (~7.32 GB weights) fails during model initialization. Environment Device: iPhone Air (12 GB RAM) iOS: 26 Xcode: 26.0.1 Build: Metal backend enabled llama.cpp App runs on device (not Simulator) What I’m seeing MTLCreateSystemDefaultDevice().recommendedMaxWorkingSetSize == 8192 MiB Loading Llama-2-13B Q4_K (7.32 GB) fails to complete. Logs indicate memory pressure / allocation issues consistent with the 8 GB working-set guidance. Smaller models (e.g., 7B/8B with similar quantization) load and run (8B Q4_K provide around 9 tokens/second decoding speed). Questions Is 8,192 MB an expected recommendedMaxWorkingSetSize on a 12 GB iPhone? What values should I expect on other 2025 devices including iPhone 17 (8 GB RAM) and iPhone 17 Pro (12 GB RAM) Is it strictly enforced by Metal allocations (heaps/buffers), or advisory for best performance/eviction behavior? Can a process practically exceed this for long-lived buffers without immediate Jetsam risk? Any guidance for LLM scenarios near the limit?

The following minimal snippet SEGFAULTS with SDK 26.0 and 26.1. Won't crash if I remove async from the enclosing function signature - but it's impractical in a real project. import Metal import MetalPerformanceShaders let SEED = UInt64(0x0) typealias T = Float16 /* Why ran in async context? Because global GPU object, and async makeMTLFunction, and async makeMTLComputePipelineState. Nevertheless, can trigger the bug without using global @MainActor let myGPU = MyGPU() */ @main struct CMDLine { static func main() async { let ptr = UnsafeMutablePointer<T>.allocate(capacity: 0) async let future: Void = randomFillOnGPU(ptr, count: 0) print("Main thread is playing around") await future print("Successfully reached the end.") } static func randomFillOnGPU(_ buf: UnsafeMutablePointer<T>, count destbufcount: Int) async { // let (device, queue) = await (myGPU.device, myGPU.commandqueue) let myGPU = MyGPU() let (device, queue) = (myGPU.device, myGPU.commandqueue) // Init MTLBuffer, async let makeFunction, makeComputePipelineState, etc. let tempDataType = MPSDataType.uInt32 let randfiller = MPSMatrixRandomMTGP32(device: device, destinationDataType: tempDataType, seed: Int(bitPattern:UInt(SEED))) print("randomFillOnGPU: successfully created MPSMatrixRandom.") // try await computePipelineState // ^ Crashes before this could return // Or in this minimal case, after randomFillOnGPU() returns // make encoder, set pso, dispatch, commit... } } actor MyGPU { let device : MTLDevice let commandqueue : MTLCommandQueue init() { guard let dev: MTLDevice = MPSGetPreferredDevice(.skipRemovable), let cq = dev.makeCommandQueue(), dev.supportsFamily(.apple6) || dev.supportsFamily(.mac2) else { print("Unable to get Metal Device! Exiting"); exit(EX_UNAVAILABLE) } print("Selected device: \(String(format: "%llX", dev.registryID))") self.device = dev self.commandqueue = cq print("myGPU: initialization complete.") } } See FB20916929. Apparently objc autorelease pool is releasing the wrong address during context switch (across suspension points). I wonder why such obvious case has not been caught before.

After watching WWDC 2025 session "Combine Metal 4 machine learning and graphics", I have decided to give it a shot to integrate the latest MTL4MachineLearningCommandEncoder to my existing render pipeline. After a lot of trial and errors, I managed to set up the pipeline and have the app compiled. However, I am now stuck on creating a MTLLibrary with .mtlpackage. Here is the code I have to create a MTLLibrary according the WWDC session https://developer.apple.com/videos/play/wwdc2025/262/?time=550: let coreMLFilePath = bundle.path(forResource: "my_model", ofType: "mtlpackage")! let coreMLURL = URL(string: coreMLFilePath)! do { metalDevice.makeLibrary(URL: coreMLURL) } catch { print("error: \(error)") } With the above code, I am getting error: Error Domain=MTLLibraryErrorDomain Code=1 "Invalid metal package" UserInfo={NSLocalizedDescription=Invalid metal package} What is the correct way to create a MTLLibrary with .mtlpackage? Do I see this error because the .mtlpackage I am using is incorrect? How should I go with debugging this? I'd really appreciate if I could get some help on this as I have been stuck with it for some time now. Thanks in advance!

Hi, I’m testing Unity’s Spaceship HDRP demo on iPhone 17 Pro Max and iPad Pro M4 (iOS 26.1). Everything renders correctly, and my custom MetalFX Spatial plugin initializes successfully — it briefly reports active scaling (e.g. 1434×660 → 2868×1320 at 50% scaling), then reverts to native rendering a few frames later. Setup: Xcode 16.1 (targeting iOS 18) Unity 2022.3.62f3 (HDRP) Metal backend Dynamic Resolution enabled in HDRP assets and cameras Relevant Xcode console excerpt: [MetalFXPlugin] MetalFX_Enable(True) called. [SpaceshipOptions] MetalFX enabled with HDRP dynamic resolution integration. [SpaceshipOptions] Disabled TAA for MetalFX Spatial. [SpaceshipOptions] Created runtime RenderTexture: 1434x660 [MetalFX] Spatial scaler created (1434x660 → 2868x1320). [MetalFX] Processed frame with scaler. [MetalFXPlugin] Sent RenderTexture (1434x660) to MetalFX. Output target 2868x1320. [SpaceshipOptions] MetalFX target set: 1434x660 [SpaceshipOptions] Camera targetTexture cleared after MetalFX handoff. It looks like HDRP clears the camera’s target texture right after MetalFX submits the frame, which causes it to revert to native rendering. Is there a recommended way to persist or rebind the MetalFX output texture when using HDRP on iOS? Unity doesn’t appear to support MetalFX in the Editor either: Thanks!

Hi all, I'm encountering an issue with Metal raytracing on my M5 MacBook Pro regarding Instance Acceleration Structure (IAS). Intersection tests suddenly stop working after a certain point in the sampling loop. Situation I implemented an offline GPU path tracer that runs the same kernel multiple times per pixel (sampleCount) using metal::raytracing. Intersection tests are performed using an IAS. Since this is an offline path tracer, geometries inside the IAS never changes across samples (no transforms or updates). As sampleCount increases, there comes a point where the number of intersections drops to zero, and remains zero for all subsequent samples. Here's a code sketch: let sampleCount: UInt16 = 1024 for sampleIndex: UInt16 in 0..<sampleCount { // ... do { let commandBuffer = commandQueue.makeCommandBuffer() // Dispatch the intersection kernel. await commandBuffer.completed() } do { let commandBuffer = commandQueue.makeCommandBuffer() // Use the intersection test results from the previous command buffer. await commandBuffer.completed() } // ... } kernel void intersectAlongRay( const metal::uint32_t threadIndex [[thread_position_in_grid]], // ... const metal::raytracing::instance_acceleration_structure accelerationStructure [[buffer(2)]], // ... ) { // ... const auto result = intersector.intersect(ray, accelerationStructure); switch (result.type) { case metal::raytracing::intersection_type::triangle: { // Write intersection result to device buffers. break; } default: break; } Observations Encoding both the intersection kernel and the subsequent result usage in the same command buffer does not resolve the problem. Switching from IAS to Primitive Acceleration Structure (PAS) fixes the problem. Rebuilding the IAS for each sample also resolves the issue. Intersections produce inconsistent results even though the IAS and rays are identical — Image 1 shows a hit, while Image 2 shows a miss. Questions Am I misusing IAS in some way ? Could this be a Metal bug ? Any guidance or confirmation would be greatly appreciated.

My app has a number of heterogeneous GPU workloads that all run concurrently. Some of these should be executed with the highest priority because the app’s responsiveness depends on them, while others are triggered by file imports and the like which should have a low priority. If this was running on the CPU I’d assign the former User Interactive QoS and the latter Utility QoS. Is there an equivalent to this for GPU work?

Hello, I am currently working on a research project under ENINCA Consulting, focused on advanced diagnostic tools for pseudorandom number generators (structural metrics, multi-seed stability, cross-architecture reproducibility, and complementary indicators to TestU01). To validate this diagnostic framework, I prototyped a small non-linear 64-bit PRNG (not as a goal in itself, but simply as a vehicle to test the methodology). During these evaluations, I observed something interesting on Apple Silicon (Mac M1): • bit-exact reproducibility between M1 ARM CPU and M1 Metal GPU, • full BigCrush pass on both CPU and Metal backends, • excellent p-values, • stable behaviour across multiple seeds and runs. This was not the intended objective, the goal was mainly to validate the diagnostic concepts, but these results raised some questions about deterministic compute behaviour in Metal. My question: Is there any official guidance on achieving (or expecting) deterministic RNG or compute behaviour across CPU ↔ Metal GPU on Apple Silicon? More specifically: • Are deterministic compute kernels expected or guaranteed on Metal for scientific workloads? • Are there recommended patterns or best practices to ensure reproducibility across GPU generations (M1 → M2 → M3 → M4)? • Are there known Metal features that can introduce non-determinism? I am not sharing the internal recurrence (this work is proprietary), but I can discuss the high-level diagnostic observations if helpful. Thank you for any insight, very interested in how the Metal engineering team views deterministic compute patterns on Apple Silicon. Pascal ENINCA Consulting

Hello, I am currently working on a research project under ENINCA Consulting, focused on advanced diagnostic tools for pseudorandom number generators (structural metrics, multi-seed stability, cross-architecture reproducibility, and complementary indicators to TestU01). To validate this diagnostic framework, I prototyped a small non-linear 64-bit PRNG (not as a goal in itself, but simply as a vehicle to test the methodology). During these evaluations, I observed something interesting on Apple Silicon (Mac M1): • bit-exact reproducibility between M1 ARM CPU and M1 Metal GPU, • full BigCrush pass on both CPU and Metal backends, • excellent p-values, • stable behaviour across multiple seeds and runs. This was not the intended objective, the goal was mainly to validate the diagnostic concepts, but these results raised some questions about deterministic compute behaviour in Metal. My question: Is there any official guidance on achieving (or expecting) deterministic RNG or compute behaviour across CPU ↔ Metal GPU on Apple Silicon? More specifically: • Are deterministic compute kernels expected or guaranteed on Metal for scientific workloads? • Are there recommended patterns or best practices to ensure reproducibility across GPU generations (M1 → M2 → M3 → M4)? • Are there known Metal features that can introduce non-determinism? I am not sharing the internal recurrence (this work is proprietary), but I can discuss the high-level diagnostic observations if helpful. Thank you for any insight, very interested in how the Metal engineering team views deterministic compute patterns on Apple Silicon. Pascal ENINCA Consulting

I am puzzled by the setAddress(_:attributeStride:index:) of MTL4ArgumentTable. Can anyone please explain what the attributeStride parameter is for? The doc says that it is "The stride between attributes in the buffer." but why? Who uses this for what? On the C++ side in the shaders the stride is determined by the C++ type, as far as I know. What am I missing here? Thanks!