perf(indium): rebase lazy wakeups and batch tuning onto current polling

Keep the merged per-device completion budget, bounded parsing, idle blocking and timeout accounting. Scope the registration/blocking handshake to each device, restrict live batch windows to the timer resolution, and test registration interleavings, spin-disabled polling, shutdown, opt-out and independent devices with actual patched methods.

Co-authored-by: SypherrDev <[email protected]>

docs/SETUP.md

@@ -135,6 +135,9 @@ Indium also contains experimental hand-off changes. They are **off by default**:
 | `/tmp/metal-gpu` | diagnostic: selects the Vulkan device used for Metal by index or case-insensitive name substring instead of the loader's first device. The OpenGL context that presents does not follow it; see [hybrid laptops](TROUBLESHOOTING.md#hybrid-graphics-laptops-amdintel-igpu--nvidia). `METAL_INDIUM_GPU` is the environment equivalent; the file wins |
 | `/tmp/metal-all-devices` | a `VkDevice` is created for every supported Vulkan device, as before 2026-10-16. Without it only the selected device gets one; on a machine with lavapipe installed this also drops lavapipe's idle rasterizer threads |
 | `/tmp/metal-no-drawable-vram` | drawables of a layer with `framebufferOnly` off (League sets it off) go back to host-visible system RAM, as before 2026-10-16. Without the file they are allocated in GPU memory like `framebufferOnly` drawables; nothing maps drawable memory from the CPU. On the RX 6600 this gave about 20 FPS more in a practice-tool fight and the same FPS when idle, not yet measured in an ARAM. `/tmp/metal-device.log` shows `drawable-vram=on` and `storage=vram` |
+| `/tmp/metal-no-lazy-wakeup` | every completion registration signals the event loop's wake-up semaphore from the CPU again, as before 2026-10-18. The handshake is per device. Without the file a registration signals only while the poller has announced a blocking wait (a spinning poller re-reads the registrations on its next pass, and its spin ends when one arrives). On the NVIDIA proprietary driver each `vkSignalSemaphore` is an RM control ioctl that waits on a driver-wide lock, about 0.3 ms each and ~25 per League frame, and it was the largest kernel cost of the submit worker. Re-read every 256 registrations; the wait log counts `wakeupSignalled` and `wakeupSkipped` |
+| `/tmp/metal-submit-batch-us` | the submit batch window in microseconds (500–100000, default 500); re-read every 256 checks, so it can be changed during a match. A presenting commit and every CPU wait still flush the batch at once. On the NVIDIA proprietary driver every `vkQueueSubmit2` is an RM control ioctl that waits on a driver-wide lock (about 0.4 ms each in a League match); with the default window a frame took about 8.5 submits, with `5000` about 2.5 (RTX 4090: 89 → 113 FPS, same match) |
+| `/tmp/metal-submit-batch-max` | command buffers per batch (1–256, default 16), re-read like the window; `64` lets one batch hold a whole frame |
 | `/tmp/metal-debug-waits` | diagnostic: counts how often and how long threads block at the hand-off waits and in `nextDrawable`, every 5 s in `/tmp/metal-trace.log`; the first check writes a `waits enabled` line |
 
 Create these files from inside the guest, for example `darling shell touch /tmp/metal-debug-waits`, or while Darling is shut down. The guest `/tmp` is `~/.darling/private/tmp` seen through Darling's overlay mount. A file created on the host in that directory while Darling runs may stay invisible to guest processes that already looked for it. The same applies to removing a file. The older `/tmp/metal-no-present-dispatch-wait`, `/tmp/metal-no-deferred-blit` and `/tmp/metal-no-skip-cpu-devices` files are ignored.

docs/TROUBLESHOOTING.md

@@ -202,6 +202,11 @@ Compare matches on the same map: Bilgewater runs a few percent below Hexgate.
 
 On Hyprland, also check that direct scanout is off (`hyprctl getoption render:direct_scanout` prints `int: 0`). The first match with it off ran about 16 FPS faster than the previous one on the same build; see the next section.
 
+On the NVIDIA proprietary driver, check the submit worker's kernel time: a thread of the game process spending most of its time in the kernel (`/proc/<pid>/task/<tid>/stat` system time, or `perf trace -s -p <pid>`) points at driver calls. Each `vkQueueSubmit2` and each CPU-side `vkSignalSemaphore` is an RM control ioctl (command `0xc020462a`) that queues on a driver-wide lock. Before 2026-10-18 every command buffer signalled the event loop's wake-up semaphore, ~25 times per frame; with lazy wake-ups (default) and a 5 ms batch window (`/tmp/metal-submit-batch-us` = `5000`, `/tmp/metal-submit-batch-max` = `64`) an RTX 4090 ran at 206–552 FPS against 91–147 FPS with `/tmp/metal-no-lazy-wakeup`, in alternating 20 s phases of one match (the later phases with wake-ups on were capped at 240 FPS) (see [setup](SETUP.md#metal-and-indium-overlays)).
+
+The NVIDIA performance figures above were reported on an earlier runtime base;
+they have not been remeasured after integration with current `main`.
+
 ## Hybrid graphics laptops (AMD/Intel iGPU + NVIDIA)
 
 Metal renders with Vulkan, and each finished frame is handed to an OpenGL (EGL) context that presents it. The hand-off uses opaque file descriptors, which only work when Vulkan and OpenGL run on the same GPU. Indium takes the first Vulkan device the loader lists, and on a hybrid laptop both that device and the EGL context are normally the integrated GPU, so the discrete GPU stays idle. A report in [issue #3](https://github.com/lucannez64/darling-lol/issues/3) (Ryzen 9 5900HS, RTX 3060 Laptop) measured 60-70 FPS in ARAM on the integrated GPU; on the NVIDIA GPU the game showed a black screen. **Running on an NVIDIA GPU is not validated.**

patches/metal--deps--indium.patch

@@ -781,7 +781,7 @@ index 8546daa..7e4993c 100644
  		void updateBindings();
  
 diff --git a/private-include/indium/device.private.hpp b/private-include/indium/device.private.hpp
-index 58f1794..1f4bd5a 100644
+index 58f1794..3e7ee47 100644
 --- a/private-include/indium/device.private.hpp
 +++ b/private-include/indium/device.private.hpp
 @@ -13,6 +13,8 @@
@@ -793,7 +793,7 @@ index 58f1794..1f4bd5a 100644
  
  namespace Indium {
  	class PrivateDevice;
-@@ -25,8 +27,11 @@ namespace Indium {
+@@ -25,8 +27,14 @@ namespace Indium {
  	class PrivateDevice: public Device, public std::enable_shared_from_this<PrivateDevice> {
  	private:
  
@@ -802,10 +802,13 @@ index 58f1794..1f4bd5a 100644
  		std::mutex _eventLoopMutex;
  		std::mutex _pollingMutex;
 +		uint64_t _completionSpinNanoseconds = 0;
++		// Sequentially consistent handshake: either re-snapshot or signal.
++		std::atomic<uint64_t> _eventLoopRegistrations {0};
++		std::atomic<unsigned> _eventLoopBlockedPollers {0};
  
  		// these are stored separately (rather than in a single structure)
  		// so that we can easily copy them individually in the event loop polling function.
-@@ -59,6 +64,8 @@ namespace Indium {
+@@ -59,6 +67,8 @@ namespace Indium {
  		}
  
  		virtual std::string name() const override;
@@ -814,7 +817,7 @@ index 58f1794..1f4bd5a 100644
  		virtual std::shared_ptr<CommandQueue> newCommandQueue() override;
  		virtual std::shared_ptr<RenderPipelineState> newRenderPipelineState(const RenderPipelineDescriptor& descriptor) override;
  		virtual std::shared_ptr<ComputePipelineState> newComputePipelineState(const ComputePipelineDescriptor& descriptor, PipelineOption options, std::shared_ptr<ComputePipelineReflection> reflection) override;
-@@ -73,6 +80,12 @@ namespace Indium {
+@@ -73,6 +83,12 @@ namespace Indium {
  		virtual void pollEvents(uint64_t timeoutNanoseconds) override;
  		virtual void wakeupEventLoop() override;
  
@@ -827,7 +830,7 @@ index 58f1794..1f4bd5a 100644
  		void waitForSemaphore(VkSemaphore semaphore, uint64_t targetValue, std::function<void()> callback);
  
  		/**
-@@ -86,6 +99,11 @@ namespace Indium {
+@@ -86,6 +102,11 @@ namespace Indium {
  		 * Similar to getTimelineSemaphore(), but wraps it in a shared pointer
  		 * that automatically releases the semaphore when all references to it die out.
  		 */
@@ -839,7 +842,7 @@ index 58f1794..1f4bd5a 100644
  		std::shared_ptr<TimelineSemaphore> getWrappedTimelineSemaphore();
  
  		BinarySemaphore getBinarySemaphore(bool exportable = false);
-@@ -111,5 +129,34 @@ namespace Indium {
+@@ -111,5 +132,34 @@ namespace Indium {
  
  		INDIUM_PROPERTY(VkPhysicalDeviceMemoryProperties, m, M,emoryProperties);
  		INDIUM_PROPERTY_READONLY(Feature, f, F,eatures);
@@ -1934,7 +1937,7 @@ index 23fd2d3..1a48b81 100644
 +	return _gpuAddress;
  };
 diff --git a/src/indium/command-buffer.cpp b/src/indium/command-buffer.cpp
-index 43f4a84..aaa1acb 100644
+index 43f4a84..f7640c5 100644
 --- a/src/indium/command-buffer.cpp
 +++ b/src/indium/command-buffer.cpp
 @@ -1,3 +1,8 @@
@@ -2271,7 +2274,7 @@ index 43f4a84..aaa1acb 100644
  	auto encoder = std::make_shared<Indium::PrivateComputeCommandEncoder>(shared_from_this(), descriptor);
  	{
  		std::scoped_lock lock(_mutex);
-@@ -91,10 +354,239 @@ std::shared_ptr<Indium::ComputeCommandEncoder> Indium::PrivateCommandBuffer::com
+@@ -91,10 +354,269 @@ std::shared_ptr<Indium::ComputeCommandEncoder> Indium::PrivateCommandBuffer::com
  	return computeCommandEncoder(ComputePassDescriptor { {}, dispatchType });
  };
  
@@ -2393,6 +2396,35 @@ index 43f4a84..aaa1acb 100644
 +	}
 +}
 +
++// darling-lol: tunable batch window and size. On the NVIDIA proprietary driver every vkQueueSubmit2 is an RM control
++// ioctl that queues on a driver-wide lock (~0.4 ms each in a League match, ~8.5 submits per frame with the 500 us
++// window), so fewer, larger batches matter there. Guest files, re-read every 256 checks: /tmp/metal-submit-batch-us
++// (window in microseconds, 500-100000, default 500) and /tmp/metal-submit-batch-max (command buffers per batch,
++// 1-256, default 16). A presenting commit and every CPU wait still flush the batch at once.
++long readTuning(const char* path, long fallback, long low, long high) {
++	FILE* file = fopen(path, "r");
++	if (!file) return fallback;
++	long value = fallback;
++	if (fscanf(file, "%ld", &value) != 1) value = fallback;
++	fclose(file);
++	return std::clamp(value, low, high);
++}
++
++struct SubmitBatchTuning {
++	std::atomic<long> windowMicroseconds {500};
++	std::atomic<long> maxItems {16};
++	std::atomic<unsigned> calls {0};
++};
++
++SubmitBatchTuning& submitBatchTuning() {
++	static auto* tuning = new SubmitBatchTuning;
++	if ((tuning->calls.fetch_add(1, std::memory_order_relaxed) & 255) == 0) {
++		tuning->windowMicroseconds.store(readTuning("/tmp/metal-submit-batch-us", 500, 500, 100000), std::memory_order_relaxed);
++		tuning->maxItems.store(readTuning("/tmp/metal-submit-batch-max", 16, 1, 256), std::memory_order_relaxed);
++	}
++	return *tuning;
++}
++
 +void startSubmitBatchTimer(const std::shared_ptr<Indium::PrivateDevice>& device) {
 +	auto& state = submitBatchState();
 +	if (state.timerStarted) return;
@@ -2407,8 +2439,9 @@ index 43f4a84..aaa1acb 100644
 +				 * merging with latency; the old 2ms added visible input lag. */
 +				submitBatchWake().wait_for(lock, std::chrono::microseconds(500));
 +			}
++			const auto window = std::chrono::microseconds(submitBatchTuning().windowMicroseconds.load(std::memory_order_relaxed));
 +			std::unique_lock<std::mutex> queueLock(device->graphicsQueueMutex());
-+			if (!state.items.empty() && std::chrono::steady_clock::now() - state.since >= std::chrono::microseconds(500)) flushSubmitBatchLocked();
++			if (!state.items.empty() && std::chrono::steady_clock::now() - state.since >= window) flushSubmitBatchLocked();
 +		}
 +	}).detach();
 +}
@@ -2511,7 +2544,7 @@ index 43f4a84..aaa1acb 100644
  
  	for (const auto& encoder: _commandEncoders) {
  		if (auto renderEncoder = std::dynamic_pointer_cast<PrivateRenderCommandEncoder>(encoder)) {
-@@ -110,16 +602,28 @@ void Indium::PrivateCommandBuffer::commit() {
+@@ -110,16 +632,28 @@ void Indium::PrivateCommandBuffer::commit() {
  			}
  		}
  
@@ -2541,7 +2574,7 @@ index 43f4a84..aaa1acb 100644
  
  	// for the event loop
  	auto timelineSemaphore = _privateDevice->getWrappedTimelineSemaphore();
-@@ -138,14 +642,49 @@ void Indium::PrivateCommandBuffer::commit() {
+@@ -138,14 +672,49 @@ void Indium::PrivateCommandBuffer::commit() {
  			readWriteTextures.insert(readWriteTextures.end(), readWrite.begin(), readWrite.end());
  		}
  
@@ -2594,7 +2627,7 @@ index 43f4a84..aaa1acb 100644
  	}
  
  	std::vector<VkSemaphoreSubmitInfo> signalInfos;
-@@ -202,11 +741,13 @@ void Indium::PrivateCommandBuffer::commit() {
+@@ -202,11 +771,13 @@ void Indium::PrivateCommandBuffer::commit() {
  		auto privateTexture = std::dynamic_pointer_cast<PrivateTexture>(texture);
  		const auto& presentSema = presentationSemaphores[i];
  
@@ -2613,7 +2646,7 @@ index 43f4a84..aaa1acb 100644
  
  		handleTextureSemaphores(privateTexture);
  	}
-@@ -215,7 +756,15 @@ void Indium::PrivateCommandBuffer::commit() {
+@@ -215,7 +786,15 @@ void Indium::PrivateCommandBuffer::commit() {
  	//        but it also leaves the resources for this command buffer tied up, essentially becoming a memory leak.
  	//        i've observed this in the cube example, and it happens more than once (because the example display semaphore is exhausted and never signaled).
  	// UPDATE: upon further testing, it seems that this only occurs when the view is off-screen/hidden. weird.
@@ -2630,7 +2663,7 @@ index 43f4a84..aaa1acb 100644
  		{
  			std::unique_lock lock(self->_mutex);
  			self->_completed = true;
-@@ -234,8 +783,47 @@ void Indium::PrivateCommandBuffer::commit() {
+@@ -234,8 +813,47 @@ void Indium::PrivateCommandBuffer::commit() {
  		// the handlers may have captured a reference to us (from their surrounding scope), so clear out handlers so those references go away
  		self->_scheduledHandlers.clear();
  		self->_completedHandlers.clear();
@@ -2668,7 +2701,7 @@ index 43f4a84..aaa1acb 100644
 +			}
 +			batch.items.push_back(std::move(item));
 +			startSubmitBatchTimer(_privateDevice);
-+			if (batch.items.size() >= 16) flushSubmitBatchLocked();
++			if (batch.items.size() >= (size_t) submitBatchTuning().maxItems.load(std::memory_order_relaxed)) flushSubmitBatchLocked();
 +			return;
 +		}
 +	}
@@ -2678,7 +2711,7 @@ index 43f4a84..aaa1acb 100644
  	commandBufferInfo.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_SUBMIT_INFO;
  	commandBufferInfo.commandBuffer = _commandBuffer;
  
-@@ -252,23 +840,41 @@ void Indium::PrivateCommandBuffer::commit() {
+@@ -252,23 +870,41 @@ void Indium::PrivateCommandBuffer::commit() {
  	//        the Device constructor tries to choose command queues that support as many operations as possible, but it's possible
  	//        that a particular device only supports certain operations on certain queues (e.g. maybe it only supports transfer operations
  	//        on an exclusive queue that doesn't support graphics or compute).
@@ -2725,7 +2758,7 @@ index 43f4a84..aaa1acb 100644
  };
  
  void Indium::PrivateCommandBuffer::presentDrawable(std::shared_ptr<Drawable> drawable) {
-@@ -307,9 +913,12 @@ void Indium::PrivateCommandBuffer::addCompletedHandler(std::function<void(std::s
+@@ -307,9 +943,12 @@ void Indium::PrivateCommandBuffer::addCompletedHandler(std::function<void(std::s
  };
  
  void Indium::PrivateCommandBuffer::waitUntilCompleted() {
@@ -2838,7 +2871,7 @@ index 5f62ad0..5957bec 100644
  	DynamicVK::vkCmdBindDescriptorSets(buf->commandBuffer(), VK_PIPELINE_BIND_POINT_COMPUTE, _pso->layout(), 0, descriptorSets.size(), descriptorSets.data(), 0, nullptr);
  };
 diff --git a/src/indium/device.cpp b/src/indium/device.cpp
-index 322e186..4da420a 100644
+index 322e186..f712c7a 100644
 --- a/src/indium/device.cpp
 +++ b/src/indium/device.cpp
 @@ -1,4 +1,9 @@
@@ -3161,11 +3194,11 @@ index 322e186..4da420a 100644
 +			}
 +			creationOrder.push_back(i);
 +		}
- 	}
++	}
 +
 +	for (size_t index: creationOrder) {
 +		globalDeviceList.push_back(std::make_shared<PrivateDevice>(candidates[index].device));
-+	}
+ 	}
 +
 +	appendDeviceLog("created VkDevice for " + std::to_string(globalDeviceList.size()) + " of "
 +		+ std::to_string(candidates.size()) + " supported device(s)"
@@ -3291,7 +3324,32 @@ index 322e186..4da420a 100644
  		funcInfo.bindings.insert(funcInfo.bindings.end(), info.bindings.begin(), info.bindings.end());
  		funcInfo.embeddedSamplers.insert(funcInfo.embeddedSamplers.end(), info.embeddedSamplers.begin(), info.embeddedSamplers.end());
  	}
-@@ -396,6 +722,26 @@ void Indium::PrivateDevice::pollEvents(uint64_t timeoutNanoseconds) {
+@@ -382,6 +708,19 @@ std::shared_ptr<Indium::Device> Indium::createSystemDefaultDevice() {
+ 	return globalDeviceList.empty() ? nullptr : globalDeviceList.front();
+ };
+ 
++namespace {
++bool lazyWakeupEnabled() {
++    static std::atomic<bool> enabled {true};
++    static std::atomic<unsigned> calls {0};
++    if ((calls.fetch_add(1, std::memory_order_relaxed) & 255) == 0)
++        enabled.store(access("/tmp/metal-no-lazy-wakeup", F_OK) != 0, std::memory_order_relaxed);
++    return enabled.load(std::memory_order_relaxed);
++}
++void noteWakeup(const char* site) {
++    if (Indium::Waits::enabled()) Indium::Waits::record(site, 0);
++}
++}
++
+ void Indium::PrivateDevice::pollEvents(uint64_t timeoutNanoseconds) {
+ 	// this is held for the entire duration of the poll so we can
+ 	// ensure we're the only one polling, which allows us to avoid
+@@ -393,9 +732,35 @@ void Indium::PrivateDevice::pollEvents(uint64_t timeoutNanoseconds) {
+ 	const std::vector<VkSemaphore> semaphores = _eventLoopSemaphores;
+ 	const std::vector<uint64_t> values = _eventLoopWaitValues;
+ 	const std::vector<std::function<void()>> callbacks = _eventLoopCallbacks;
++	// All registrations covered by this snapshot, under the same mutex.
++	const uint64_t snapshotRegistrations = _eventLoopRegistrations.load(std::memory_order_seq_cst);
  
  	lock.unlock();
  
@@ -3308,6 +3366,10 @@ index 322e186..4da420a 100644
 +                anyReady = DynamicVK::vkGetSemaphoreCounterValue(_device, semaphores[i], &count) == VK_SUCCESS && count >= values[i];
 +            }
 +            if (anyReady) break;
++            if (_eventLoopRegistrations.load(std::memory_order_relaxed) != snapshotRegistrations) {
++                anyReady = true; // Re-snapshot registrations instead of waiting on stale handles.
++                break;
++            }
 +            __builtin_ia32_pause();
 +        } while (std::chrono::steady_clock::now() < deadline);
 +        const uint64_t elapsed = std::chrono::duration_cast<std::chrono::nanoseconds>(
@@ -3318,8 +3380,43 @@ index 322e186..4da420a 100644
  	VkSemaphoreWaitInfo info {};
  	info.sType = VK_STRUCTURE_TYPE_SEMAPHORE_WAIT_INFO;
  	info.semaphoreCount = semaphores.size();
-@@ -480,9 +826,18 @@ void Indium::PrivateDevice::waitForSemaphore(VkSemaphore semaphore, uint64_t tar
- 	wakeupEventLoop();
+@@ -403,7 +768,19 @@ void Indium::PrivateDevice::pollEvents(uint64_t timeoutNanoseconds) {
+ 	info.pValues = values.data();
+ 	info.flags = VK_SEMAPHORE_WAIT_ANY_BIT;
+ 
+-	auto result = DynamicVK::vkWaitSemaphores(_device, &info, timeoutNanoseconds);
++    // A registrant increments its counter before inspecting blocked pollers.
++    // The poller announces its wait before rechecking registrations. With both
++    // operations seq_cst, either this wait becomes nonblocking or the registrant
++    // signals the wake-up semaphore. The counters belong to this device only.
++    const bool announcedBlock = timeoutNanoseconds > 0;
++    if (announcedBlock) {
++        _eventLoopBlockedPollers.fetch_add(1, std::memory_order_seq_cst);
++        if (_eventLoopRegistrations.load(std::memory_order_seq_cst) != snapshotRegistrations)
++            timeoutNanoseconds = 0;
++    }
++    auto result = DynamicVK::vkWaitSemaphores(_device, &info, timeoutNanoseconds);
++    if (announcedBlock)
++        _eventLoopBlockedPollers.fetch_sub(1, std::memory_order_seq_cst);
+ 
+ 	if (result == VK_TIMEOUT) {
+ 		// timed out with no semaphores ready
+@@ -474,15 +851,30 @@ void Indium::PrivateDevice::waitForSemaphore(VkSemaphore semaphore, uint64_t tar
+ 		_eventLoopSemaphores.push_back(semaphore);
+ 		_eventLoopWaitValues.push_back(targetValue);
+ 		_eventLoopCallbacks.push_back(callback);
++		_eventLoopRegistrations.fetch_add(1, std::memory_order_seq_cst);
+ 	}
+ 
+-	// now wakeup the event loop so it can start waiting on this new semaphore
+-	wakeupEventLoop();
++	// A spinning poller re-snapshots registrations without a driver signal.
++    if (!lazyWakeupEnabled() || _eventLoopBlockedPollers.load(std::memory_order_seq_cst) > 0) {
++        noteWakeup("wakeupSignalled");
++        wakeupEventLoop();
++    } else {
++        noteWakeup("wakeupSkipped");
++    }
  };
  
 -// TODO: create a semaphore pool to avoid constantly creating and destroying semaphores
@@ -3338,7 +3435,7 @@ index 322e186..4da420a 100644
  	VkSemaphoreTypeCreateInfo typeInfo {};
  	typeInfo.sType = VK_STRUCTURE_TYPE_SEMAPHORE_TYPE_CREATE_INFO;
  	typeInfo.semaphoreType = VK_SEMAPHORE_TYPE_TIMELINE;
-@@ -505,7 +860,14 @@ Indium::TimelineSemaphore Indium::PrivateDevice::getTimelineSemaphore() {
+@@ -505,7 +897,14 @@ Indium::TimelineSemaphore Indium::PrivateDevice::getTimelineSemaphore() {
  };
  
  void Indium::PrivateDevice::putTimelineSemaphore(const TimelineSemaphore& semaphore) {
@@ -3354,7 +3451,7 @@ index 322e186..4da420a 100644
  };
  
  std::shared_ptr<Indium::TimelineSemaphore> Indium::PrivateDevice::getWrappedTimelineSemaphore() {
-@@ -551,3 +913,160 @@ std::shared_ptr<Indium::BinarySemaphore> Indium::PrivateDevice::getWrappedBinary
+@@ -551,3 +950,160 @@ std::shared_ptr<Indium::BinarySemaphore> Indium::PrivateDevice::getWrappedBinary
  		delete ptr;
  	});
  };

tests/fixtures/indium-wakeup.cpp

@@ -0,0 +1,146 @@
+#include <algorithm>
+#include <array>
+#include <atomic>
+#include <cassert>
+#include <chrono>
+#include <condition_variable>
+#include <cstdint>
+#include <cstdlib>
+#include <functional>
+#include <mutex>
+#include <thread>
+#include <vector>
+using VkSemaphore = int;
+struct VkSemaphoreWaitInfo { int sType, flags; size_t semaphoreCount; const VkSemaphore* pSemaphores; const uint64_t* pValues; };
+struct VkSemaphoreSignalInfo { int sType; VkSemaphore semaphore; uint64_t value; };
+constexpr int VK_STRUCTURE_TYPE_SEMAPHORE_WAIT_INFO=1, VK_SEMAPHORE_WAIT_ANY_BIT=1;
+constexpr int VK_STRUCTURE_TYPE_SEMAPHORE_SIGNAL_INFO=2, VK_SUCCESS=0, VK_TIMEOUT=1;
+static std::array<std::atomic<uint64_t>,512> counters {};
+static std::atomic<int> signalCalls {0};
+static std::atomic<uint64_t> lastTimeout {0};
+static std::mutex kernelMutex;
+static std::condition_variable kernelWake;
+static std::function<void()> onWait, onCounter;
+static bool lazy = true;
+static bool lazyWakeupEnabled() { return lazy; }
+static void noteWakeup(const char*) {}
+struct TestMutex {
+    std::mutex mutex;
+    std::function<void()> afterUnlock;
+    void lock() { mutex.lock(); }
+    void unlock() {
+        auto hook=std::move(afterUnlock);
+        afterUnlock=nullptr;
+        mutex.unlock();
+        if(hook) hook();
+    }
+};
+namespace DynamicVK {
+int vkGetSemaphoreCounterValue(int, int semaphore, uint64_t* value) {
+    auto hook=std::move(onCounter); onCounter=nullptr;
+    if(hook) hook();
+    *value=counters[semaphore].load(); return VK_SUCCESS;
+}
+int vkSignalSemaphore(int, const VkSemaphoreSignalInfo* info) {
+    { std::lock_guard lock(kernelMutex); counters[info->semaphore]=info->value; ++signalCalls; }
+    kernelWake.notify_all(); return VK_SUCCESS;
+}
+int vkWaitSemaphores(int, const VkSemaphoreWaitInfo* info, uint64_t timeout) {
+    lastTimeout=timeout;
+    auto hook=std::move(onWait); onWait=nullptr;
+    if(hook) hook();
+    auto ready=[&]{
+        for(size_t i=0;i<info->semaphoreCount;++i)
+            if(counters[info->pSemaphores[i]].load()>=info->pValues[i]) return true;
+        return false;
+    };
+    if(ready()) return VK_SUCCESS;
+    if(!timeout) return VK_TIMEOUT;
+    std::unique_lock lock(kernelMutex);
+    // A missed wakeup fails in a bounded interval instead of hanging the suite.
+    assert(kernelWake.wait_for(lock,std::chrono::seconds(2),ready));
+    return VK_SUCCESS;
+}
+}
+namespace Indium {
+class PrivateDevice {
+public:
+    std::mutex _pollingMutex;
+    TestMutex _eventLoopMutex;
+    int _device=0;
+    uint64_t _completionSpinNanoseconds=0;
+    std::atomic<uint64_t> _eventLoopRegistrations {0};
+    std::atomic<unsigned> _eventLoopBlockedPollers {0};
+    std::vector<VkSemaphore> _eventLoopSemaphores={1};
+    std::vector<uint64_t> _eventLoopWaitValues={1};
+    std::vector<std::function<void()>> _eventLoopCallbacks={nullptr};
+    void pollEvents(uint64_t);
+    void waitForSemaphore(VkSemaphore,uint64_t,std::function<void()>);
+    void wakeupEventLoop();
+};
+}
+/* ACTUAL_PATCHED_METHODS */
+static void reset() {
+    for(auto& value:counters) value=0;
+    signalCalls=0; lastTimeout=0; lazy=true; onWait=nullptr; onCounter=nullptr;
+}
+int main() {
+    // Registration before snapshot is included without a host-side wakeup.
+    reset(); {
+        Indium::PrivateDevice d; int done=0; counters[2]=1;
+        d.waitForSemaphore(2,1,[&]{++done;}); d.pollEvents(UINT64_MAX);
+        assert(done==1 && signalCalls==0 && d._eventLoopBlockedPollers==0);
+    }
+    // Registration after snapshot but before announcing a blocking wait.
+    reset(); {
+        Indium::PrivateDevice d; int done=0;
+        d._eventLoopMutex.afterUnlock=[&]{ counters[2]=1; d.waitForSemaphore(2,1,[&]{++done;}); };
+        d.pollEvents(UINT64_MAX);
+        assert(lastTimeout==0 && signalCalls==0 && done==0 && d._eventLoopBlockedPollers==0);
+        d.pollEvents(UINT64_MAX); assert(done==1);
+    }
+    // Registration after announcing a blocking wait must signal that wait.
+    reset(); {
+        Indium::PrivateDevice d; int done=0;
+        onWait=[&]{ assert(d._eventLoopBlockedPollers==1); counters[2]=1; d.waitForSemaphore(2,1,[&]{++done;}); };
+        d.pollEvents(UINT64_MAX);
+        assert(signalCalls==1 && d._eventLoopBlockedPollers==0 && done==0);
+        d.pollEvents(UINT64_MAX); assert(done==1);
+    }
+    // A registration during the spin ends it and causes a fresh snapshot.
+    reset(); {
+        Indium::PrivateDevice d; int done=0; d._completionSpinNanoseconds=1000000;
+        d.waitForSemaphore(2,1,nullptr);
+        onCounter=[&]{ counters[3]=1; d.waitForSemaphore(3,1,[&]{++done;}); };
+        d.pollEvents(UINT64_MAX); assert(lastTimeout==0 && signalCalls==0);
+        d.pollEvents(UINT64_MAX); assert(done==1);
+    }
+    // Zero-timeout polls never announce a blocking wait.
+    reset(); {
+        Indium::PrivateDevice d;
+        onWait=[&]{ assert(d._eventLoopBlockedPollers==0); d.waitForSemaphore(2,1,nullptr); };
+        d.pollEvents(0); assert(lastTimeout==0 && signalCalls==0);
+    }
+    // Opt-out restores unconditional registration wakeups; shutdown always wakes.
+    reset(); {
+        Indium::PrivateDevice d; lazy=false; d.waitForSemaphore(2,1,nullptr);
+        assert(signalCalls==1); lazy=true; d.wakeupEventLoop(); assert(signalCalls==2);
+    }
+    // A blocked poller on another device cannot change this device's policy.
+    reset(); {
+        Indium::PrivateDevice a,b; b._eventLoopBlockedPollers=1;
+        a.waitForSemaphore(2,1,nullptr); assert(signalCalls==0 && b._eventLoopRegistrations==0);
+    }
+    // Exercise real producer/poller concurrency with spinning enabled and disabled.
+    for(uint64_t spin:{0ULL,1000000ULL}) {
+        reset(); Indium::PrivateDevice d; d._completionSpinNanoseconds=spin;
+        std::atomic<int> done {0}; constexpr int count=200;
+        std::thread poller([&]{ while(done<count) d.pollEvents(UINT64_MAX); });
+        for(int i=2;i<count+2;++i) {
+            counters[i]=1;
+            d.waitForSemaphore(i,1,[&]{ ++done; });
+            if(i%3==0) std::this_thread::yield();
+        }
+        poller.join(); assert(done==count && d._eventLoopBlockedPollers==0);
+    }
+}

tests/test_indium_wakeup.py

@@ -0,0 +1,63 @@
+"""Compile the actual event-loop methods against a deterministic Vulkan wait fixture."""
+from pathlib import Path
+import shutil
+import subprocess
+import tempfile
+import unittest
+import test_runtime_compatibility as sources
+
+ROOT = Path(__file__).resolve().parents[1]
+
+
+@unittest.skipUnless(shutil.which('c++'), 'C++ compiler required')
+class WakeupChecks(unittest.TestCase):
+    patched = sources.CompatibilityChecks.patched
+
+    def test_registration_interleavings_opt_out_shutdown_and_two_devices(self):
+        source = self.patched('metal--deps--indium', 'src/indium/device.cpp')
+        methods = '\n'.join(sources.function(source, 'Indium::PrivateDevice::' + name)
+                            for name in ('pollEvents', 'waitForSemaphore', 'wakeupEventLoop'))
+        fixture = (ROOT / 'tests/fixtures/indium-wakeup.cpp').read_text()
+        self.run_cpp(fixture.replace('/* ACTUAL_PATCHED_METHODS */', methods))
+
+    def test_live_batch_controls_respect_timer_resolution_and_bounds(self):
+        source = self.patched('metal--deps--indium', 'src/indium/command-buffer.cpp')
+        tuning = source[source.index('long readTuning('):source.index('void startSubmitBatchTimer(')]
+        fixture = r'''
+#include <algorithm>
+#include <atomic>
+#include <cassert>
+#include <cstdio>
+static const char* fixturePath;
+static FILE* test_fopen(const char*, const char* mode) { return ::fopen(fixturePath,mode); }
+#define fopen test_fopen
+''' + tuning + r'''
+#undef fopen
+static void write(const char* text) {
+    FILE* f=fopen(fixturePath,"w"); assert(f); fputs(text,f); fclose(f);
+}
+int main(int argc,char** argv) {
+    assert(argc==2); fixturePath=argv[1];
+    write("100");
+    auto& initial=submitBatchTuning();
+    assert(initial.windowMicroseconds==500 && initial.maxItems==100);
+    write("200000");
+    for(int i=0;i<256;++i) submitBatchTuning();
+    assert(initial.windowMicroseconds==100000 && initial.maxItems==256);
+    write("invalid");
+    for(int i=0;i<256;++i) submitBatchTuning();
+    assert(initial.windowMicroseconds==500 && initial.maxItems==16);
+}
+'''
+        self.run_cpp(fixture, config=True)
+
+    def run_cpp(self, text, config=False):
+        with tempfile.TemporaryDirectory() as folder:
+            path = Path(folder)
+            (path / 'check.cpp').write_text(text)
+            compile = subprocess.run(['c++', '-std=c++17', '-pthread', str(path/'check.cpp'),
+                                      '-o', str(path/'check')], capture_output=True, text=True)
+            self.assertEqual(compile.returncode, 0, compile.stderr)
+            result = subprocess.run([str(path/'check'), *([str(path/'tuning')] if config else [])],
+                                    capture_output=True, text=True, timeout=15)
+            self.assertEqual(result.returncode, 0, result.stderr)

tests/test_nvidia_runtime.py

@@ -126,6 +126,7 @@ int main() {
         poll = sources.function(source, 'Indium::PrivateDevice::pollEvents')
         harness = '''
 #include <algorithm>
+#include <atomic>
 #include <chrono>
 #include <cassert>
 #include <cstdint>
@@ -159,6 +160,8 @@ public:
     std::mutex _pollingMutex, _eventLoopMutex;
     int _device=0;
     uint64_t _completionSpinNanoseconds=1000000;
+    std::atomic<uint64_t> _eventLoopRegistrations {0};
+    std::atomic<unsigned> _eventLoopBlockedPollers {0};
     std::vector<VkSemaphore> _eventLoopSemaphores={1};
     std::vector<uint64_t> _eventLoopWaitValues={1};
     std::vector<std::function<void()>> _eventLoopCallbacks={nullptr};