CoopAllTheThings/tests/vk_capture_perf_test.cpp

// Reproducer + regression guard for the Vulkan capture performance collapse.
//
// Against a real 144 FPS game (Sphere Spectacle) the implicit-layer / inline-hook capture dropped
// the game to ~3 FPS. Measured cause: the read-back ran on the game's PRESENT THREAD and spent
// ~370 ms per 1080p frame reading the mapped staging buffer -- because the staging memory was a
// plain HOST_VISIBLE|HOST_COHERENT type (write-combined / uncached on a discrete GPU), where a
// scattered CPU read runs at PCIe latency.
//
// This test builds a known gradient image (in PRESENT_SRC layout) on a real device and measures the
// time the *present thread* spends per capture for two implementations:
//   * a synchronous reference (copy + WaitForFences + read the HOST_COHERENT mapping + swizzle, all
//     inline) -> reproduces the stall, and
//   * coop::hook::VkCapture (present thread only records+submits; a reaper thread does the
//     HOST_CACHED read-back + swizzle + upload off the critical path).
// It asserts VkCapture's present-thread cost is a small fraction of the synchronous cost, and that
// the captured image is byte-correct (BGRA->RGBA swizzle). `--sync` routes the measured path through
// the synchronous reference so the same assertion FAILS, demonstrating the test catches the regression.
//
// Needs a working Vulkan ICD (the dev box has one). With no vulkan-1.dll / no device it SKIPs.
#include <cstdint>
#include <cstdio>
#include <cstring>
#include <string>
#include <vector>

#include <windows.h>

#define VK_NO_PROTOTYPES
#include <vulkan/vulkan.h>

#include "vk_capture.hpp"

namespace {
double now_ms()
{
	LARGE_INTEGER f, c;
	QueryPerformanceFrequency(&f);
	QueryPerformanceCounter(&c);
	return 1000.0 * static_cast<double>(c.QuadPart) / static_cast<double>(f.QuadPart);
}

int g_failures = 0;
void check(bool ok, const char* what)
{
	std::printf("%s %s\n", ok ? "  ok:" : "FAIL:", what);
	if (!ok) {
		++g_failures;
	}
}

// Everything the test resolves from the device (superset of VkCapture::Fns + image-build helpers).
struct DevFns {
	PFN_vkGetDeviceProcAddr GetDeviceProcAddr;
	PFN_vkGetDeviceQueue GetDeviceQueue;
	PFN_vkCreateCommandPool CreateCommandPool;
	PFN_vkDestroyCommandPool DestroyCommandPool;
	PFN_vkAllocateCommandBuffers AllocateCommandBuffers;
	PFN_vkBeginCommandBuffer BeginCommandBuffer;
	PFN_vkEndCommandBuffer EndCommandBuffer;
	PFN_vkResetCommandBuffer ResetCommandBuffer;
	PFN_vkCmdPipelineBarrier CmdPipelineBarrier;
	PFN_vkCmdCopyImageToBuffer CmdCopyImageToBuffer;
	PFN_vkCmdCopyBufferToImage CmdCopyBufferToImage;
	PFN_vkQueueSubmit QueueSubmit;
	PFN_vkQueueWaitIdle QueueWaitIdle;
	PFN_vkCreateFence CreateFence;
	PFN_vkDestroyFence DestroyFence;
	PFN_vkWaitForFences WaitForFences;
	PFN_vkResetFences ResetFences;
	PFN_vkGetFenceStatus GetFenceStatus;
	PFN_vkCreateSemaphore CreateSemaphore;
	PFN_vkDestroySemaphore DestroySemaphore;
	PFN_vkCreateBuffer CreateBuffer;
	PFN_vkDestroyBuffer DestroyBuffer;
	PFN_vkGetBufferMemoryRequirements GetBufferMemoryRequirements;
	PFN_vkCreateImage CreateImage;
	PFN_vkDestroyImage DestroyImage;
	PFN_vkGetImageMemoryRequirements GetImageMemoryRequirements;
	PFN_vkAllocateMemory AllocateMemory;
	PFN_vkFreeMemory FreeMemory;
	PFN_vkBindBufferMemory BindBufferMemory;
	PFN_vkBindImageMemory BindImageMemory;
	PFN_vkMapMemory MapMemory;
	PFN_vkUnmapMemory UnmapMemory;
	PFN_vkInvalidateMappedMemoryRanges InvalidateMappedMemoryRanges;
	PFN_vkDeviceWaitIdle DeviceWaitIdle;
	PFN_vkGetPhysicalDeviceMemoryProperties GetPhysicalDeviceMemoryProperties;
};

VkPhysicalDeviceMemoryProperties g_memprops{};

bool find_mem(std::uint32_t type_bits, VkMemoryPropertyFlags want, std::uint32_t& out)
{
	for (std::uint32_t i = 0; i < g_memprops.memoryTypeCount; ++i) {
		if ((type_bits & (1u << i)) && (g_memprops.memoryTypes[i].propertyFlags & want) == want) {
			out = i;
			return true;
		}
	}
	return false;
}

coop::hook::VkCapture::Fns capture_fns(const DevFns& d)
{
	coop::hook::VkCapture::Fns f{};
	f.GetDeviceQueue = d.GetDeviceQueue;
	f.CreateCommandPool = d.CreateCommandPool;
	f.DestroyCommandPool = d.DestroyCommandPool;
	f.AllocateCommandBuffers = d.AllocateCommandBuffers;
	f.BeginCommandBuffer = d.BeginCommandBuffer;
	f.EndCommandBuffer = d.EndCommandBuffer;
	f.ResetCommandBuffer = d.ResetCommandBuffer;
	f.CmdPipelineBarrier = d.CmdPipelineBarrier;
	f.CmdCopyImageToBuffer = d.CmdCopyImageToBuffer;
	f.QueueSubmit = d.QueueSubmit;
	f.CreateFence = d.CreateFence;
	f.DestroyFence = d.DestroyFence;
	f.WaitForFences = d.WaitForFences;
	f.ResetFences = d.ResetFences;
	f.GetFenceStatus = d.GetFenceStatus;
	f.CreateSemaphore = d.CreateSemaphore;
	f.DestroySemaphore = d.DestroySemaphore;
	f.CreateBuffer = d.CreateBuffer;
	f.DestroyBuffer = d.DestroyBuffer;
	f.GetBufferMemoryRequirements = d.GetBufferMemoryRequirements;
	f.AllocateMemory = d.AllocateMemory;
	f.FreeMemory = d.FreeMemory;
	f.BindBufferMemory = d.BindBufferMemory;
	f.MapMemory = d.MapMemory;
	f.UnmapMemory = d.UnmapMemory;
	f.InvalidateMappedMemoryRanges = d.InvalidateMappedMemoryRanges;
	f.DeviceWaitIdle = d.DeviceWaitIdle;
	f.GetPhysicalDeviceMemoryProperties = d.GetPhysicalDeviceMemoryProperties;
	return f;
}
} // namespace

int main(int argc, char** argv)
{
	const bool sync_repro = argc > 1 && std::strcmp(argv[1], "--sync") == 0;
	const std::uint32_t W = 1920, H = 1080; // the resolution where the stall was measured

	HMODULE vk = LoadLibraryW(L"vulkan-1.dll");
	if (vk == nullptr) {
		std::printf("SKIP vk_capture_perf_test (no vulkan-1.dll)\n");
		return 0;
	}
	auto gipa = reinterpret_cast<PFN_vkGetInstanceProcAddr>(GetProcAddress(vk, "vkGetInstanceProcAddr"));
	if (gipa == nullptr) {
		std::printf("SKIP vk_capture_perf_test (no vkGetInstanceProcAddr)\n");
		return 0;
	}
#define IFN(name) reinterpret_cast<PFN_vk##name>(gipa(instance, "vk" #name))

	VkInstance instance = VK_NULL_HANDLE;
	{
		auto create = reinterpret_cast<PFN_vkCreateInstance>(gipa(nullptr, "vkCreateInstance"));
		VkApplicationInfo app{VK_STRUCTURE_TYPE_APPLICATION_INFO};
		app.apiVersion = VK_API_VERSION_1_1;
		VkInstanceCreateInfo ci{VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO};
		ci.pApplicationInfo = &app;
		if (create == nullptr || create(&ci, nullptr, &instance) != VK_SUCCESS) {
			std::printf("SKIP vk_capture_perf_test (vkCreateInstance failed)\n");
			return 0;
		}
	}

	auto EnumeratePhysicalDevices = IFN(EnumeratePhysicalDevices);
	auto GetPhysicalDeviceQueueFamilyProperties = IFN(GetPhysicalDeviceQueueFamilyProperties);
	auto CreateDevice = IFN(CreateDevice);
	auto DestroyDevice = IFN(DestroyDevice);
	auto DestroyInstance = IFN(DestroyInstance);
	auto gdpa = IFN(GetDeviceProcAddr);

	std::uint32_t n = 0;
	EnumeratePhysicalDevices(instance, &n, nullptr);
	if (n == 0) {
		std::printf("SKIP vk_capture_perf_test (no physical devices)\n");
		DestroyInstance(instance, nullptr);
		return 0;
	}
	std::vector<VkPhysicalDevice> phys(n);
	EnumeratePhysicalDevices(instance, &n, phys.data());
	VkPhysicalDevice gpu = phys[0];

	std::uint32_t qn = 0;
	GetPhysicalDeviceQueueFamilyProperties(gpu, &qn, nullptr);
	std::vector<VkQueueFamilyProperties> qf(qn);
	GetPhysicalDeviceQueueFamilyProperties(gpu, &qn, qf.data());
	std::uint32_t qfam = UINT32_MAX;
	for (std::uint32_t i = 0; i < qn; ++i) {
		if (qf[i].queueFlags & VK_QUEUE_GRAPHICS_BIT) {
			qfam = i;
			break;
		}
	}
	if (qfam == UINT32_MAX) {
		std::printf("SKIP vk_capture_perf_test (no graphics queue)\n");
		DestroyInstance(instance, nullptr);
		return 0;
	}

	const float prio = 1.0f;
	VkDeviceQueueCreateInfo qci{VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO};
	qci.queueFamilyIndex = qfam;
	qci.queueCount = 1;
	qci.pQueuePriorities = &prio;
	VkDeviceCreateInfo dci{VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO};
	dci.queueCreateInfoCount = 1;
	dci.pQueueCreateInfos = &qci;
	VkDevice device = VK_NULL_HANDLE;
	if (CreateDevice(gpu, &dci, nullptr, &device) != VK_SUCCESS) {
		std::printf("SKIP vk_capture_perf_test (vkCreateDevice failed)\n");
		DestroyInstance(instance, nullptr);
		return 0;
	}

#define DFN(name) reinterpret_cast<PFN_vk##name>(gdpa(device, "vk" #name))
	DevFns d{};
	d.GetDeviceProcAddr = gdpa;
	d.GetDeviceQueue = DFN(GetDeviceQueue);
	d.CreateCommandPool = DFN(CreateCommandPool);
	d.DestroyCommandPool = DFN(DestroyCommandPool);
	d.AllocateCommandBuffers = DFN(AllocateCommandBuffers);
	d.BeginCommandBuffer = DFN(BeginCommandBuffer);
	d.EndCommandBuffer = DFN(EndCommandBuffer);
	d.ResetCommandBuffer = DFN(ResetCommandBuffer);
	d.CmdPipelineBarrier = DFN(CmdPipelineBarrier);
	d.CmdCopyImageToBuffer = DFN(CmdCopyImageToBuffer);
	d.CmdCopyBufferToImage = DFN(CmdCopyBufferToImage);
	d.QueueSubmit = DFN(QueueSubmit);
	d.QueueWaitIdle = DFN(QueueWaitIdle);
	d.CreateFence = DFN(CreateFence);
	d.DestroyFence = DFN(DestroyFence);
	d.WaitForFences = DFN(WaitForFences);
	d.ResetFences = DFN(ResetFences);
	d.GetFenceStatus = DFN(GetFenceStatus);
	d.CreateSemaphore = DFN(CreateSemaphore);
	d.DestroySemaphore = DFN(DestroySemaphore);
	d.CreateBuffer = DFN(CreateBuffer);
	d.DestroyBuffer = DFN(DestroyBuffer);
	d.GetBufferMemoryRequirements = DFN(GetBufferMemoryRequirements);
	d.CreateImage = DFN(CreateImage);
	d.DestroyImage = DFN(DestroyImage);
	d.GetImageMemoryRequirements = DFN(GetImageMemoryRequirements);
	d.AllocateMemory = DFN(AllocateMemory);
	d.FreeMemory = DFN(FreeMemory);
	d.BindBufferMemory = DFN(BindBufferMemory);
	d.BindImageMemory = DFN(BindImageMemory);
	d.MapMemory = DFN(MapMemory);
	d.UnmapMemory = DFN(UnmapMemory);
	d.InvalidateMappedMemoryRanges = DFN(InvalidateMappedMemoryRanges);
	d.DeviceWaitIdle = DFN(DeviceWaitIdle);
	d.GetPhysicalDeviceMemoryProperties = reinterpret_cast<PFN_vkGetPhysicalDeviceMemoryProperties>(
	    gipa(instance, "vkGetPhysicalDeviceMemoryProperties"));

	d.GetPhysicalDeviceMemoryProperties(gpu, &g_memprops);

	VkQueue queue = VK_NULL_HANDLE;
	d.GetDeviceQueue(device, qfam, 0, &queue);

	// A small command pool + fence the test uses for setup and for the synchronous reference.
	VkCommandPool pool = VK_NULL_HANDLE;
	VkCommandPoolCreateInfo pci{VK_STRUCTURE_TYPE_COMMAND_POOL_CREATE_INFO};
	pci.flags = VK_COMMAND_POOL_CREATE_RESET_COMMAND_BUFFER_BIT;
	pci.queueFamilyIndex = qfam;
	d.CreateCommandPool(device, &pci, nullptr, &pool);
	VkCommandBuffer cb = VK_NULL_HANDLE;
	VkCommandBufferAllocateInfo cbai{VK_STRUCTURE_TYPE_COMMAND_BUFFER_ALLOCATE_INFO};
	cbai.commandPool = pool;
	cbai.level = VK_COMMAND_BUFFER_LEVEL_PRIMARY;
	cbai.commandBufferCount = 1;
	d.AllocateCommandBuffers(device, &cbai, &cb);
	VkFence fence = VK_NULL_HANDLE;
	VkFenceCreateInfo fci{VK_STRUCTURE_TYPE_FENCE_CREATE_INFO};
	d.CreateFence(device, &fci, nullptr, &fence);

	auto submit_wait = [&](VkCommandBuffer c) {
		VkSubmitInfo si{VK_STRUCTURE_TYPE_SUBMIT_INFO};
		si.commandBufferCount = 1;
		si.pCommandBuffers = &c;
		d.ResetFences(device, 1, &fence);
		d.QueueSubmit(queue, 1, &si, fence);
		d.WaitForFences(device, 1, &fence, VK_TRUE, UINT64_MAX);
	};

	// --- Build the known source image (BGRA gradient) in PRESENT_SRC layout ------------------------
	const VkDeviceSize bytes = static_cast<VkDeviceSize>(W) * H * 4;
	std::vector<unsigned char> gradient(bytes);
	for (std::uint32_t y = 0; y < H; ++y) {
		for (std::uint32_t x = 0; x < W; ++x) {
			unsigned char* p = &gradient[(static_cast<size_t>(y) * W + x) * 4];
			p[0] = static_cast<unsigned char>(x & 0xFF);       // B
			p[1] = static_cast<unsigned char>(y & 0xFF);       // G
			p[2] = static_cast<unsigned char>((x + y) & 0xFF); // R
			p[3] = 255;
		}
	}
	// Upload buffer (host-visible).
	VkBuffer upbuf = VK_NULL_HANDLE;
	VkDeviceMemory upmem = VK_NULL_HANDLE;
	{
		VkBufferCreateInfo bci{VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO};
		bci.size = bytes;
		bci.usage = VK_BUFFER_USAGE_TRANSFER_SRC_BIT;
		d.CreateBuffer(device, &bci, nullptr, &upbuf);
		VkMemoryRequirements mr{};
		d.GetBufferMemoryRequirements(device, upbuf, &mr);
		std::uint32_t mt = 0;
		find_mem(mr.memoryTypeBits, VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT | VK_MEMORY_PROPERTY_HOST_COHERENT_BIT, mt);
		VkMemoryAllocateInfo mai{VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO};
		mai.allocationSize = mr.size;
		mai.memoryTypeIndex = mt;
		d.AllocateMemory(device, &mai, nullptr, &upmem);
		d.BindBufferMemory(device, upbuf, upmem, 0);
		void* mp = nullptr;
		d.MapMemory(device, upmem, 0, VK_WHOLE_SIZE, 0, &mp);
		std::memcpy(mp, gradient.data(), bytes);
		d.UnmapMemory(device, upmem);
	}
	// Device-local source image.
	VkImage img = VK_NULL_HANDLE;
	VkDeviceMemory imgmem = VK_NULL_HANDLE;
	{
		VkImageCreateInfo ici{VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO};
		ici.imageType = VK_IMAGE_TYPE_2D;
		ici.format = VK_FORMAT_B8G8R8A8_UNORM;
		ici.extent = {W, H, 1};
		ici.mipLevels = 1;
		ici.arrayLayers = 1;
		ici.samples = VK_SAMPLE_COUNT_1_BIT;
		ici.tiling = VK_IMAGE_TILING_OPTIMAL;
		ici.usage = VK_IMAGE_USAGE_TRANSFER_DST_BIT | VK_IMAGE_USAGE_TRANSFER_SRC_BIT;
		ici.initialLayout = VK_IMAGE_LAYOUT_UNDEFINED;
		if (d.CreateImage(device, &ici, nullptr, &img) != VK_SUCCESS) {
			std::printf("SKIP vk_capture_perf_test (CreateImage failed)\n");
			return 0;
		}
		VkMemoryRequirements mr{};
		d.GetImageMemoryRequirements(device, img, &mr);
		std::uint32_t mt = 0;
		find_mem(mr.memoryTypeBits, VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT, mt);
		VkMemoryAllocateInfo mai{VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO};
		mai.allocationSize = mr.size;
		mai.memoryTypeIndex = mt;
		d.AllocateMemory(device, &mai, nullptr, &imgmem);
		d.BindImageMemory(device, img, imgmem, 0);
	}
	auto barrier = [&](VkCommandBuffer c, VkImageLayout from, VkImageLayout to, VkAccessFlags src, VkAccessFlags dst) {
		VkImageMemoryBarrier b{VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER};
		b.srcAccessMask = src;
		b.dstAccessMask = dst;
		b.oldLayout = from;
		b.newLayout = to;
		b.srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
		b.dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
		b.image = img;
		b.subresourceRange = {VK_IMAGE_ASPECT_COLOR_BIT, 0, 1, 0, 1};
		d.CmdPipelineBarrier(c, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT, 0, 0, nullptr,
		                     0, nullptr, 1, &b);
	};
	{
		VkCommandBufferBeginInfo bi{VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO};
		bi.flags = VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT;
		d.BeginCommandBuffer(cb, &bi);
		barrier(cb, VK_IMAGE_LAYOUT_UNDEFINED, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, 0, VK_ACCESS_TRANSFER_WRITE_BIT);
		VkBufferImageCopy r{};
		r.imageSubresource = {VK_IMAGE_ASPECT_COLOR_BIT, 0, 0, 1};
		r.imageExtent = {W, H, 1};
		d.CmdCopyBufferToImage(cb, upbuf, img, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, 1, &r);
		barrier(cb, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, VK_IMAGE_LAYOUT_PRESENT_SRC_KHR, VK_ACCESS_TRANSFER_WRITE_BIT,
		        VK_ACCESS_MEMORY_READ_BIT);
		d.EndCommandBuffer(cb);
		submit_wait(cb);
	}

	// --- Synchronous reference (HOST_COHERENT staging, inline read-back on the present thread) ------
	VkBuffer ref_buf = VK_NULL_HANDLE;
	VkDeviceMemory ref_mem = VK_NULL_HANDLE;
	void* ref_mapped = nullptr;
	{
		VkBufferCreateInfo bci{VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO};
		bci.size = bytes;
		bci.usage = VK_BUFFER_USAGE_TRANSFER_DST_BIT;
		d.CreateBuffer(device, &bci, nullptr, &ref_buf);
		VkMemoryRequirements mr{};
		d.GetBufferMemoryRequirements(device, ref_buf, &mr);
		std::uint32_t mt = 0; // OLD selection: first HOST_VISIBLE|HOST_COHERENT (often write-combined)
		find_mem(mr.memoryTypeBits, VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT | VK_MEMORY_PROPERTY_HOST_COHERENT_BIT, mt);
		VkMemoryAllocateInfo mai{VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO};
		mai.allocationSize = mr.size;
		mai.memoryTypeIndex = mt;
		d.AllocateMemory(device, &mai, nullptr, &ref_mem);
		d.BindBufferMemory(device, ref_buf, ref_mem, 0);
		d.MapMemory(device, ref_mem, 0, VK_WHOLE_SIZE, 0, &ref_mapped);
	}
	std::vector<unsigned char> ref_rgba(bytes);
	auto sync_capture = [&]() {
		VkCommandBufferBeginInfo bi{VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO};
		bi.flags = VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT;
		d.ResetCommandBuffer(cb, 0);
		d.BeginCommandBuffer(cb, &bi);
		barrier(cb, VK_IMAGE_LAYOUT_PRESENT_SRC_KHR, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL, VK_ACCESS_MEMORY_READ_BIT,
		        VK_ACCESS_TRANSFER_READ_BIT);
		VkBufferImageCopy r{};
		r.imageSubresource = {VK_IMAGE_ASPECT_COLOR_BIT, 0, 0, 1};
		r.imageExtent = {W, H, 1};
		d.CmdCopyImageToBuffer(cb, img, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL, ref_buf, 1, &r);
		barrier(cb, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL, VK_IMAGE_LAYOUT_PRESENT_SRC_KHR, VK_ACCESS_TRANSFER_READ_BIT,
		        VK_ACCESS_MEMORY_READ_BIT);
		d.EndCommandBuffer(cb);
		submit_wait(cb);
		const auto* src = static_cast<const unsigned char*>(ref_mapped);
		const size_t row = static_cast<size_t>(W) * 4;
		for (std::uint32_t y = 0; y < H; ++y) {
			const unsigned char* in = src + static_cast<size_t>(y) * row;
			unsigned char* o = ref_rgba.data() + static_cast<size_t>(y) * row;
			for (std::uint32_t x = 0; x < W; ++x) {
				o[x * 4 + 0] = in[x * 4 + 2];
				o[x * 4 + 1] = in[x * 4 + 1];
				o[x * 4 + 2] = in[x * 4 + 0];
				o[x * 4 + 3] = 255;
			}
		}
	};

	// Time the synchronous reference (a few iterations; this is the per-present cost an inline
	// read-back puts on the game's present thread).
	sync_capture(); // warm
	double sync_ms = 0;
	const int iters = 8;
	for (int i = 0; i < iters; ++i) {
		const double t0 = now_ms();
		sync_capture();
		sync_ms += now_ms() - t0;
	}
	sync_ms /= iters;
	std::printf("synchronous reference (old design): %.2f ms per present-thread capture (%ux%u)\n", sync_ms, W, H);

	double sut_ms = sync_ms; // in --sync repro mode the measured path IS the synchronous one
	bool image_ok = true;

	if (!sync_repro) {
		// --- The fix under test: VkCapture (async reaper) -----------------------------------------
		coop::hook::VkCapture cap;
		cap.init(gpu, device, qfam, capture_fns(d), GetCurrentProcessId(), nullptr);

		// Consume the present semaphore VkCapture hands back, exactly like a real vkQueuePresentKHR.
		auto consume = [&](VkSemaphore sem) {
			VkPipelineStageFlags stage = VK_PIPELINE_STAGE_ALL_COMMANDS_BIT;
			VkSubmitInfo si{VK_STRUCTURE_TYPE_SUBMIT_INFO};
			si.waitSemaphoreCount = 1;
			si.pWaitSemaphores = &sem;
			si.pWaitDstStageMask = &stage;
			d.QueueSubmit(queue, 1, &si, VK_NULL_HANDLE);
		};

		// Warm up (first calls allocate staging / create the D3D texture on the reaper).
		for (int i = 0; i < 16; ++i) {
			VkSemaphore sem = VK_NULL_HANDLE;
			if (cap.present(img, VK_FORMAT_B8G8R8A8_UNORM, W, H, nullptr, 0, sem)) {
				consume(sem);
			}
			Sleep(2);
		}
		// Steady-state present-thread cost.
		int queued = 0;
		double t = 0;
		const int loop = 240;
		for (int i = 0; i < loop; ++i) {
			VkSemaphore sem = VK_NULL_HANDLE;
			const double t0 = now_ms();
			const bool did = cap.present(img, VK_FORMAT_B8G8R8A8_UNORM, W, H, nullptr, 0, sem);
			t += now_ms() - t0;
			if (did) {
				consume(sem);
				++queued;
			}
			Sleep(1); // ~1 kHz present loop; lets the reaper drain
		}
		sut_ms = t / loop;
		std::printf("VkCapture (fix): %.3f ms per present-thread call  (queued %d/%d, published %llu)\n", sut_ms,
		            queued, loop, static_cast<unsigned long long>(cap.frames_published()));

		// Let the reaper finish, then verify the captured image is byte-correct (BGRA->RGBA swizzle).
		Sleep(50);
		std::vector<unsigned char> got;
		std::uint32_t gw = 0, gh = 0;
		if (cap.last_frame(got, gw, gh) && gw == W && gh == H) {
			image_ok = std::memcmp(got.data(), ref_rgba.data(), bytes) == 0;
		} else {
			image_ok = false;
		}
		check(cap.frames_published() > 0, "VkCapture published frames");
		check(image_ok, "VkCapture image matches the source gradient (BGRA->RGBA swizzle correct)");

		d.DeviceWaitIdle(device);
		cap.shutdown();
	}

	// The core assertion: the measured present-thread cost must be a small fraction of the synchronous
	// read-back cost (VkCapture moves the read-back off the present thread). In --sync repro mode the
	// measured path IS the synchronous one, so this fails -- demonstrating the test catches the bug.
	std::printf("present-thread cost: measured %.3f ms vs synchronous %.2f ms (ratio %.3f)\n", sut_ms, sync_ms,
	            sut_ms / sync_ms);
	check(sut_ms * 4.0 < sync_ms, "capture stays off the present thread (measured << synchronous)");

	// cleanup
	d.DeviceWaitIdle(device);
	d.UnmapMemory(device, ref_mem);
	d.DestroyBuffer(device, ref_buf, nullptr);
	d.FreeMemory(device, ref_mem, nullptr);
	d.DestroyImage(device, img, nullptr);
	d.FreeMemory(device, imgmem, nullptr);
	d.DestroyBuffer(device, upbuf, nullptr);
	d.FreeMemory(device, upmem, nullptr);
	d.DestroyFence(device, fence, nullptr);
	d.DestroyCommandPool(device, pool, nullptr);
	DestroyDevice(device, nullptr);
	DestroyInstance(instance, nullptr);

	std::printf(g_failures == 0 ? "PASS vk_capture_perf_test\n" : "FAILED vk_capture_perf_test (%d)\n", g_failures);
	return g_failures == 0 ? 1 - 1 : 1; // 0 on pass
}