/* Copyright (c) 2017-2026 Hans-Kristian Arntzen
 *
 * Permission is hereby granted, free of charge, to any person obtaining
 * a copy of this software and associated documentation files (the
 * "Software"), to deal in the Software without restriction, including
 * without limitation the rights to use, copy, modify, merge, publish,
 * distribute, sublicense, and/or sell copies of the Software, and to
 * permit persons to whom the Software is furnished to do so, subject to
 * the following conditions:
 *
 * The above copyright notice and this permission notice shall be
 * included in all copies or substantial portions of the Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
 * IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY
 * CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
 */

#define NOMINMAX
#include "wsi.hpp"
#include "environment.hpp"
#include <algorithm>

#ifdef _WIN32
#define WIN32_LEAN_AND_MEAN
#include <windows.h>
#endif

#if defined(ANDROID) && defined(HAVE_SWAPPY)
#include "swappy/swappyVk.h"
#endif

static constexpr uint32_t PresentTimingQueueSize = 16;

namespace Vulkan
{
WSI::WSI()
{
	// With frame latency of 1, we get the ideal latency where
	// we present, and then wait for the previous present to complete.
	// Once this unblocks, it means that the present we just queued up is scheduled to complete next vblank,
	// and the next frame to be recorded will have to be ready in 2 frames.
	// This is ideal, since worst case for full performance, we will have a pipeline of CPU -> GPU,
	// where CPU can spend 1 frame's worth of time, and GPU can spend one frame's worth of time.
	// For mobile, opt for 2 frames of latency, since TBDR likes deeper pipelines and we can absorb more
	// surfaceflinger jank.
#ifdef ANDROID
	present_frame_latency = 2;
#else
	present_frame_latency = 1;
#endif

	present_frame_latency = Util::get_environment_uint("GRANITE_VULKAN_PRESENT_WAIT_LATENCY", present_frame_latency);
	LOGI("Targeting VK_KHR_present_wait latency to %u frames.\n", present_frame_latency);

	// Primaries are ST.2020 with D65 whitepoint as specified.
	hdr_metadata.displayPrimaryRed = { 0.708f, 0.292f };
	hdr_metadata.displayPrimaryGreen = { 0.170f, 0.797f };
	hdr_metadata.displayPrimaryBlue = { 0.131f, 0.046f };
	hdr_metadata.whitePoint = { 0.3127f, 0.3290f };

	// HDR10 range? Just arbitrary values, user can override later.
	hdr_metadata.minLuminance = 0.01f;
	hdr_metadata.maxLuminance = 1000.0f;
	hdr_metadata.maxContentLightLevel = 1000.0f;
	hdr_metadata.maxFrameAverageLightLevel = 200.0f;
}

void WSI::set_present_wait_latency(uint32_t latency)
{
	present_frame_latency = latency;
}

void WSI::set_hdr_metadata(const VkHdrMetadataEXT &hdr)
{
	hdr_metadata = hdr;
	valid_hdr_metadata = true;

	if (swapchain && swapchain_surface_format.colorSpace == VK_COLOR_SPACE_HDR10_ST2084_EXT &&
	    device->get_device_features().supports_hdr_metadata)
	{
		table->vkSetHdrMetadataEXT(device->get_device(), 1, &swapchain, &hdr_metadata);
	}
}

void WSIPlatform::set_window_title(const std::string &)
{
}

void WSIPlatform::destroy_surface(VkInstance instance, VkSurfaceKHR surface)
{
	vkDestroySurfaceKHR(instance, surface, nullptr);
}

uintptr_t WSIPlatform::get_fullscreen_monitor()
{
	return 0;
}

uintptr_t WSIPlatform::get_native_window()
{
	return 0;
}

const VkApplicationInfo *WSIPlatform::get_application_info()
{
	return nullptr;
}

void WSI::set_window_title(const std::string &title)
{
	if (platform)
		platform->set_window_title(title);
}

double WSI::get_smooth_elapsed_time() const
{
	return smooth_elapsed_time;
}

double WSI::get_smooth_frame_time() const
{
	return smooth_frame_time;
}

bool WSI::init_from_existing_context(ContextHandle existing_context)
{
	VK_ASSERT(platform);
	if (platform && device)
		platform->event_device_destroyed();
	device.reset();
	context = std::move(existing_context);
	table = &context->get_device_table();
	return true;
}

bool WSI::init_external_swapchain(std::vector<ImageHandle> swapchain_images_)
{
	VK_ASSERT(context);
	VK_ASSERT(device);
	swapchain_width = platform->get_surface_width();
	swapchain_height = platform->get_surface_height();
	swapchain_aspect_ratio = platform->get_aspect_ratio();

	external_swapchain_images = std::move(swapchain_images_);

	swapchain_width = external_swapchain_images.front()->get_width();
	swapchain_height = external_swapchain_images.front()->get_height();
	swapchain_surface_format = { external_swapchain_images.front()->get_format(), VK_COLOR_SPACE_SRGB_NONLINEAR_KHR };

	LOGI("Created swapchain %u x %u (fmt: %u).\n",
	     swapchain_width, swapchain_height, static_cast<unsigned>(swapchain_surface_format.format));

	platform->event_swapchain_destroyed();
	platform->event_swapchain_created(device.get(), VK_NULL_HANDLE, swapchain_width, swapchain_height,
	                                  swapchain_aspect_ratio,
	                                  external_swapchain_images.size(),
	                                  swapchain_surface_format.format, swapchain_surface_format.colorSpace,
	                                  swapchain_current_prerotate);

	device->init_external_swapchain(this->external_swapchain_images);
	platform->get_frame_timer().reset();
	external_acquire.reset();
	external_release.reset();
	return true;
}

void WSI::set_platform(WSIPlatform *platform_)
{
	platform = platform_;
}

bool WSI::init_device()
{
	VK_ASSERT(context);
	VK_ASSERT(!device);
	device = Util::make_handle<Device>();
	device->set_context(*context);
	platform->event_device_created(device.get());

#ifdef HAVE_WSI_DXGI_INTEROP
	dxgi.reset(new DXGIInteropSwapchain);
	if (!dxgi->init_interop_device(*device))
		dxgi.reset();
	else
		platform->get_frame_timer().reset();
#endif
	return true;
}

bool WSI::init_device(DeviceHandle device_handle)
{
	VK_ASSERT(context);
	device = std::move(device_handle);
	platform->event_device_created(device.get());

#ifdef HAVE_WSI_DXGI_INTEROP
	dxgi.reset(new DXGIInteropSwapchain);
	if (!dxgi->init_interop_device(*device))
		dxgi.reset();
	else
		platform->get_frame_timer().reset();
#endif
	return true;
}

#ifdef HAVE_WSI_DXGI_INTEROP
bool WSI::init_surface_swapchain_dxgi(unsigned width, unsigned height)
{
	if (!dxgi)
		return false;

	// Anything fancy like compute present cannot use DXGI.
	if (current_extra_usage)
		return false;

	HWND hwnd = reinterpret_cast<HWND>(platform->get_native_window());
	if (!hwnd)
		return false;

	VkSurfaceFormatKHR format = {};
	switch (current_backbuffer_format)
	{
	case BackbufferFormat::UNORM:
		format = { VK_FORMAT_A2B10G10R10_UNORM_PACK32, VK_COLOR_SPACE_SRGB_NONLINEAR_KHR };
		break;

	case BackbufferFormat::sRGB:
		format = { VK_FORMAT_B8G8R8A8_SRGB, VK_COLOR_SPACE_SRGB_NONLINEAR_KHR };
		break;

	case BackbufferFormat::HDR10:
		format = { VK_FORMAT_A2B10G10R10_UNORM_PACK32, VK_COLOR_SPACE_HDR10_ST2084_EXT };
		break;

	case BackbufferFormat::scRGB:
		format = { VK_FORMAT_R16G16B16A16_SFLOAT, VK_COLOR_SPACE_EXTENDED_SRGB_LINEAR_EXT };
		break;

	default:
		return false;
	}

	constexpr unsigned num_images = 3;

	if (!dxgi->init_swapchain(hwnd, format, width, height, num_images))
		return false;

	LOGI("Initialized DXGI interop swapchain!\n");

	swapchain_width = width;
	swapchain_height = height;
	swapchain_aspect_ratio = platform->get_aspect_ratio();
	swapchain_current_prerotate = VK_SURFACE_TRANSFORM_IDENTITY_BIT_KHR;
	swapchain_surface_format = dxgi->get_current_surface_format();
	has_acquired_swapchain_index = false;

	const uint32_t queue_present_support = 1u << context->get_queue_info().family_indices[QUEUE_INDEX_GRAPHICS];
	device->set_swapchain_queue_family_support(queue_present_support);

	swapchain_images = { dxgi->get_vulkan_image() };
	device->init_swapchain(swapchain_images, swapchain_width, swapchain_height,
	                       swapchain_surface_format.format,
	                       swapchain_current_prerotate,
	                       VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT | VK_IMAGE_USAGE_TRANSFER_SRC_BIT);

	platform->event_swapchain_destroyed();
	platform->event_swapchain_created(device.get(), swapchain, swapchain_width, swapchain_height,
	                                  swapchain_aspect_ratio, num_images,
	                                  swapchain_surface_format.format,
	                                  swapchain_surface_format.colorSpace,
	                                  swapchain_current_prerotate);

	return true;
}
#endif

bool WSI::init_surface_swapchain()
{
	VK_ASSERT(surface == VK_NULL_HANDLE);
	VK_ASSERT(context);
	VK_ASSERT(device);

	unsigned width = platform->get_surface_width();
	unsigned height = platform->get_surface_height();

#ifdef HAVE_WSI_DXGI_INTEROP
	if (init_surface_swapchain_dxgi(width, height))
		return true;
	else
		dxgi.reset();
#endif

	surface = platform->create_surface(context->get_instance(), context->get_gpu());
	if (surface == VK_NULL_HANDLE)
	{
		LOGE("Failed to create VkSurfaceKHR.\n");
		return false;
	}

	swapchain_aspect_ratio = platform->get_aspect_ratio();

	VkBool32 supported = VK_FALSE;
	uint32_t queue_present_support = 0;

	// TODO: Ideally we need to create surface earlier and negotiate physical device based on that support.
	for (auto &index : context->get_queue_info().family_indices)
	{
		if (index != VK_QUEUE_FAMILY_IGNORED)
		{
			if (vkGetPhysicalDeviceSurfaceSupportKHR(context->get_gpu(), index, surface, &supported) ==
			    VK_SUCCESS && supported)
			{
				queue_present_support |= 1u << index;
			}
		}
	}

	if ((queue_present_support & (1u << context->get_queue_info().family_indices[QUEUE_INDEX_GRAPHICS])) == 0)
	{
		LOGE("No presentation queue found for GPU. Is it connected to a display?\n");
		return false;
	}

	device->set_swapchain_queue_family_support(queue_present_support);

	if (!blocking_init_swapchain(width, height))
	{
		LOGE("Failed to create swapchain.\n");
		return false;
	}

	device->init_swapchain(swapchain_images, swapchain_width, swapchain_height, swapchain_surface_format.format,
	                       swapchain_current_prerotate,
	                       current_extra_usage | VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT);
	platform->get_frame_timer().reset();
	return true;
}

bool WSI::init_simple(unsigned num_thread_indices, const Context::SystemHandles &system_handles)
{
	if (!init_context_from_platform(num_thread_indices, system_handles))
		return false;
	if (!init_device())
		return false;
	if (!init_surface_swapchain())
		return false;
	return true;
}

bool WSI::init_context_from_platform(unsigned num_thread_indices, const Context::SystemHandles &system_handles)
{
	VK_ASSERT(platform);
	auto instance_ext = platform->get_instance_extensions();
	auto device_ext = platform->get_device_extensions();
	auto new_context = Util::make_handle<Context>();

#ifdef HAVE_FFMPEG_VULKAN
	constexpr ContextCreationFlags video_context_flags =
			CONTEXT_CREATION_ENABLE_VIDEO_DECODE_BIT |
			CONTEXT_CREATION_ENABLE_VIDEO_ENCODE_BIT |
			CONTEXT_CREATION_ENABLE_VIDEO_H264_BIT |
			CONTEXT_CREATION_ENABLE_VIDEO_H265_BIT;
#else
	constexpr ContextCreationFlags video_context_flags = 0;
#endif

	new_context->set_application_info(platform->get_application_info());
	new_context->set_num_thread_indices(num_thread_indices);
	new_context->set_system_handles(system_handles);

	constexpr ContextCreationFlags context_flags =
			CONTEXT_CREATION_ENABLE_ADVANCED_WSI_BIT |
			CONTEXT_CREATION_ENABLE_PUSH_DESCRIPTOR_BIT |
			CONTEXT_CREATION_ENABLE_DESCRIPTOR_BUFFER_BIT |
			CONTEXT_CREATION_ENABLE_DESCRIPTOR_HEAP_BIT |
#ifdef GRANITE_VULKAN_SYSTEM_HANDLES
			//CONTEXT_CREATION_ENABLE_PIPELINE_BINARY_BIT |
#endif
			video_context_flags;

	if (!new_context->init_instance(
			instance_ext.data(), instance_ext.size(),
			context_flags))
	{
		LOGE("Failed to create Vulkan instance.\n");
		return false;
	}

	VkSurfaceKHR tmp_surface = platform->create_surface(new_context->get_instance(), VK_NULL_HANDLE);

	bool ret = new_context->init_device(
			VK_NULL_HANDLE, tmp_surface,
			device_ext.data(), device_ext.size(),
			context_flags);

	if (tmp_surface)
		platform->destroy_surface(new_context->get_instance(), tmp_surface);

	if (!ret)
	{
		LOGE("Failed to create Vulkan device.\n");
		return false;
	}

	return init_from_existing_context(std::move(new_context));
}

void WSI::reinit_surface_and_swapchain(VkSurfaceKHR new_surface)
{
	LOGI("init_surface_and_swapchain()\n");
	if (new_surface != VK_NULL_HANDLE)
	{
		VK_ASSERT(surface == VK_NULL_HANDLE);
		surface = new_surface;
	}

	swapchain_width = platform->get_surface_width();
	swapchain_height = platform->get_surface_height();
	update_framebuffer(swapchain_width, swapchain_height);
}

VkResult WSI::wait_for_present(uint64_t id, uint64_t timeout)
{
	if (!swapchain)
		return VK_SUCCESS;

	if (id > present_last_id)
		return VK_NOT_READY;

	bool timeout_is_fault = false;
	if (device->get_device_features().supports_post_mortem && timeout == UINT64_MAX)
	{
		timeout = PostMortemTimeout;
		timeout_is_fault = true;
	}

	VkResult vr;

	if (supports_present_wait2 && device->get_device_features().present_wait2_features.presentWait2)
	{
		VkPresentWait2InfoKHR wait_info = { VK_STRUCTURE_TYPE_PRESENT_WAIT_2_INFO_KHR };
		wait_info.presentId = id;
		wait_info.timeout = timeout;
		vr = table->vkWaitForPresent2KHR(context->get_device(), swapchain, &wait_info);
	}
	else if (device->get_device_features().present_wait_features.presentWait)
		vr = table->vkWaitForPresentKHR(context->get_device(), swapchain, id, timeout);
	else
		return VK_NOT_READY;

	if (timeout_is_fault && (vr == VK_TIMEOUT || vr == VK_ERROR_DEVICE_LOST))
		device->managers.breadcrumbs.notify_device_hung();
	return vr;
}

void WSI::nonblock_delete_swapchain_resources()
{
	// If we can help it, don't try to destroy swapchains until we know the new swapchain has presented at least one frame on screen.
	if (swapchain != VK_NULL_HANDLE && wait_for_present(1, 0) != VK_SUCCESS)
		return;

	Util::SmallVector<DeferredDeletionSwapchain> keep;
	size_t pending = deferred_swapchains.size();
	for (auto &swap : deferred_swapchains)
	{
		if (!swap.fence || swap.fence->wait_timeout(0))
		{
			platform->destroy_swapchain_resources(swap.swapchain);
			table->vkDestroySwapchainKHR(device->get_device(), swap.swapchain, nullptr);
		}
		else if (pending >= 2)
		{
			swap.fence->wait();
			platform->destroy_swapchain_resources(swap.swapchain);
			table->vkDestroySwapchainKHR(device->get_device(), swap.swapchain, nullptr);
		}
		else
			keep.push_back(std::move(swap));

		pending--;
	}

	deferred_swapchains = std::move(keep);

	auto itr = std::remove_if(deferred_semaphore.begin(), deferred_semaphore.end(), [](DeferredDeletionSemaphore &sem) {
		return !sem.fence || sem.fence->wait_timeout(0);
	});
	deferred_semaphore.erase(itr, deferred_semaphore.end());
}

void WSI::drain_swapchain(bool in_tear_down)
{
	release_semaphores.clear();
	device->set_acquire_semaphore(0, Semaphore{});
	device->consume_release_semaphore();

	if (device->get_device_features().swapchain_maintenance1_features.swapchainMaintenance1)
	{
		// If we're just resizing, there's no need to block, defer deletions for later.
		if (in_tear_down)
		{
			if (last_present_fence)
			{
				last_present_fence->wait();
				last_present_fence.reset();
			}

			for (auto &old_swap : deferred_swapchains)
			{
				if (old_swap.fence)
					old_swap.fence->wait();
				platform->destroy_swapchain_resources(old_swap.swapchain);
				table->vkDestroySwapchainKHR(context->get_device(), old_swap.swapchain, nullptr);
			}

			deferred_swapchains.clear();
			deferred_semaphore.clear();
		}
	}
	else if (swapchain != VK_NULL_HANDLE)
	{
		wait_for_present(present_last_id);

		device->external_queue_lock();
		table->vkDeviceWaitIdle(device->get_device());
		device->external_queue_unlock();
	}
}

void WSI::tear_down_swapchain()
{
#ifdef HAVE_WSI_DXGI_INTEROP
	// We only do explicit teardown on exit.
	dxgi.reset();
#endif

	drain_swapchain(true);
	platform->event_swapchain_destroyed();
	platform->destroy_swapchain_resources(swapchain);
	table->vkDestroySwapchainKHR(context->get_device(), swapchain, nullptr);
	swapchain = VK_NULL_HANDLE;
	has_acquired_swapchain_index = false;
	next_present_id = 1;
	present_last_id = 0;
	device->set_present_id(VK_NULL_HANDLE, 0);
}

void WSI::deinit_surface_and_swapchain()
{
	LOGI("deinit_surface_and_swapchain()\n");

	tear_down_swapchain();

	if (surface != VK_NULL_HANDLE)
	{
		platform->destroy_surface(context->get_instance(), surface);
		surface = VK_NULL_HANDLE;
	}
}

void WSI::set_external_frame(unsigned index, Semaphore acquire_semaphore, double frame_time)
{
	external_frame_index = index;
	external_acquire = std::move(acquire_semaphore);
	frame_is_external = true;
	external_frame_time = frame_time;
}

bool WSI::begin_frame_external()
{
	device->next_frame_context();

	// Need to handle this stuff from outside.
	if (has_acquired_swapchain_index)
		return false;

	auto frame_time = platform->get_frame_timer().frame(external_frame_time);
	auto elapsed_time = platform->get_frame_timer().get_elapsed();

	// Assume we have been given a smooth frame pacing.
	smooth_frame_time = frame_time;
	smooth_elapsed_time = elapsed_time;

	// Poll after acquire as well for optimal latency.
	platform->poll_input();

	swapchain_index = external_frame_index;
	platform->event_frame_tick(frame_time, elapsed_time);

	platform->event_swapchain_index(device.get(), swapchain_index);
	device->set_acquire_semaphore(swapchain_index, external_acquire);
	external_acquire.reset();
	return true;
}

Semaphore WSI::consume_external_release_semaphore()
{
	Semaphore sem;
	std::swap(external_release, sem);
	return sem;
}

//#define VULKAN_WSI_TIMING_DEBUG

void WSI::wait_swapchain_latency()
{
	unsigned effective_latency = low_latency_mode_enable_present ? 0 : present_frame_latency;

	if (device->get_device_features().supports_low_latency2_nv && swapchain && low_latency_mode_enable_gpu_submit)
	{
		if (!low_latency_semaphore)
			low_latency_semaphore = device->request_semaphore(VK_SEMAPHORE_TYPE_TIMELINE);

		auto wait_ts = device->write_calibrated_timestamp();
		VkLatencySleepInfoNV sleep_info = { VK_STRUCTURE_TYPE_LATENCY_SLEEP_INFO_NV };
		sleep_info.signalSemaphore = low_latency_semaphore->get_semaphore();
		sleep_info.value = ++low_latency_semaphore_value;
		if (device->get_device_table().vkLatencySleepNV(device->get_device(), swapchain, &sleep_info) == VK_SUCCESS)
			low_latency_semaphore->wait_timeline(low_latency_semaphore_value);
		else
			LOGE("Failed to call vkLatencySleepNV.\n");
		device->register_time_interval("WSI", std::move(wait_ts), device->write_calibrated_timestamp(), "low_latency_sleep");

		VkSetLatencyMarkerInfoNV latency_marker_info = { VK_STRUCTURE_TYPE_SET_LATENCY_MARKER_INFO_NV };
		latency_marker_info.marker = VK_LATENCY_MARKER_INPUT_SAMPLE_NV;
		latency_marker_info.presentID = next_present_id;
		device->get_device_table().vkSetLatencyMarkerNV(device->get_device(), swapchain, &latency_marker_info);

		latency_marker_info.marker = VK_LATENCY_MARKER_SIMULATION_START_NV;
		device->get_device_table().vkSetLatencyMarkerNV(device->get_device(), swapchain, &latency_marker_info);

		// Avoid conflicting wait cycles when doing reflex style latency limiting.
		effective_latency = std::max<uint32_t>(effective_latency, 2);
	}
	else if (device->get_device_features().anti_lag_features.antiLag)
	{
		auto wait_ts = device->write_calibrated_timestamp();

		VkAntiLagDataAMD anti_lag = { VK_STRUCTURE_TYPE_ANTI_LAG_DATA_AMD };
		VkAntiLagPresentationInfoAMD present_info = { VK_STRUCTURE_TYPE_ANTI_LAG_PRESENTATION_INFO_AMD };
		anti_lag.pPresentationInfo = &present_info;
		present_info.stage = VK_ANTI_LAG_STAGE_INPUT_AMD;
		present_info.frameIndex = ++low_latency_semaphore_value;
		anti_lag.mode = low_latency_mode_enable_gpu_submit ? VK_ANTI_LAG_MODE_ON_AMD : VK_ANTI_LAG_MODE_OFF_AMD;
		device->get_device_table().vkAntiLagUpdateAMD(device->get_device(), &anti_lag);
		low_latency_anti_lag_present_valid = low_latency_mode_enable_gpu_submit;
		device->register_time_interval("WSI", std::move(wait_ts), device->write_calibrated_timestamp(),
		                               "low_latency_sleep");

		if (low_latency_mode_enable_gpu_submit)
		{
			// Avoid conflicting wait cycles when doing reflex style latency limiting.
			effective_latency = std::max<uint32_t>(effective_latency, 2);
		}
	}

	// If we're using duped frames, make sure we're waiting for the previous "real" frame,
	// instead of a duped one.
	// E.g. when doing frame dupes:
	// 0,      1,    2,   3,    4,   5 ...
	// real, dup, real, dup, real, dup ...
	// With present frame latency of 1 (default), after presenting 2
	// we will wait for 0 to be done rather than 1.
	// Similarly, after presenting 3 we'll still wait for 0 to be done, so we can get on submitting work
	// for the next real frame, 4 before the GPU drains of work.
	effective_latency += last_duplicated_frames;

	if ((device->get_device_features().present_wait_features.presentWait || supports_present_wait2) &&
	    present_last_id > effective_latency &&
	    current_present_mode == PresentMode::SyncToVBlank)
	{
		// The effective latency is more like present_frame_latency + 1.
		// If 0, we wait for vblank, and we must do CPU work and GPU work in one frame
		// to hit next vblank.
		uint64_t target = present_last_id - effective_latency;

#ifdef VULKAN_WSI_TIMING_DEBUG
		auto begin_wait = Util::get_current_time_nsecs();
#endif
		auto wait_ts = device->write_calibrated_timestamp();
		auto wait_result = wait_for_present(target);

		device->register_time_interval("WSI", std::move(wait_ts),
		                               device->write_calibrated_timestamp(), "wait_frame_latency");
		if (wait_result < 0)
			LOGE("vkWaitForPresentKHR failed, vr %d.\n", wait_result);
#ifdef VULKAN_WSI_TIMING_DEBUG
		auto end_wait = Util::get_current_time_nsecs();
				LOGI("WaitForPresentKHR took %.3f ms.\n", 1e-6 * double(end_wait - begin_wait));
#endif
	}
}

void WSI::emit_end_of_frame_markers()
{
	if (device->get_device_features().supports_low_latency2_nv && swapchain &&
	    low_latency_mode_enable_gpu_submit)
	{
		VkSetLatencyMarkerInfoNV latency_marker_info = { VK_STRUCTURE_TYPE_SET_LATENCY_MARKER_INFO_NV };
		latency_marker_info.marker = VK_LATENCY_MARKER_SIMULATION_END_NV;
		latency_marker_info.presentID = next_present_id;
		device->get_device_table().vkSetLatencyMarkerNV(device->get_device(), swapchain, &latency_marker_info);

		latency_marker_info.marker = VK_LATENCY_MARKER_RENDERSUBMIT_END_NV;
		device->get_device_table().vkSetLatencyMarkerNV(device->get_device(), swapchain, &latency_marker_info);
	}
}

void WSI::emit_marker_pre_present()
{
	if (device->get_device_features().supports_low_latency2_nv && swapchain &&
	    low_latency_mode_enable_gpu_submit)
	{
		VkSetLatencyMarkerInfoNV latency_marker_info = { VK_STRUCTURE_TYPE_SET_LATENCY_MARKER_INFO_NV };
		latency_marker_info.marker = VK_LATENCY_MARKER_PRESENT_START_NV;
		latency_marker_info.presentID = next_present_id;
		device->get_device_table().vkSetLatencyMarkerNV(device->get_device(), swapchain, &latency_marker_info);
	}
	else if (device->get_device_features().anti_lag_features.antiLag && low_latency_anti_lag_present_valid)
	{
		VkAntiLagDataAMD anti_lag = { VK_STRUCTURE_TYPE_ANTI_LAG_DATA_AMD };
		VkAntiLagPresentationInfoAMD present_info = { VK_STRUCTURE_TYPE_ANTI_LAG_PRESENTATION_INFO_AMD };
		anti_lag.pPresentationInfo = &present_info;
		present_info.stage = VK_ANTI_LAG_STAGE_PRESENT_AMD;
		present_info.frameIndex = low_latency_semaphore_value;
		anti_lag.mode = low_latency_mode_enable_gpu_submit ? VK_ANTI_LAG_MODE_ON_AMD : VK_ANTI_LAG_MODE_OFF_AMD;
		device->get_device_table().vkAntiLagUpdateAMD(device->get_device(), &anti_lag);
		low_latency_anti_lag_present_valid = false;
	}
}

void WSI::emit_marker_post_present()
{
	if (device->get_device_features().supports_low_latency2_nv && swapchain &&
	    low_latency_mode_enable_gpu_submit)
	{
		VkSetLatencyMarkerInfoNV latency_marker_info = { VK_STRUCTURE_TYPE_SET_LATENCY_MARKER_INFO_NV };
		latency_marker_info.marker = VK_LATENCY_MARKER_PRESENT_END_NV;
		latency_marker_info.presentID = next_present_id;
		device->get_device_table().vkSetLatencyMarkerNV(device->get_device(), swapchain, &latency_marker_info);
	}
}

void WSI::set_present_low_latency_mode(bool enable)
{
	low_latency_mode_enable_present = enable;
}

void WSI::set_gpu_submit_low_latency_mode(bool enable)
{
	if (device && device->get_device_features().supports_low_latency2_nv && swapchain &&
	    low_latency_mode_enable_gpu_submit != enable)
	{
		VkLatencySleepModeInfoNV sleep_mode_info = { VK_STRUCTURE_TYPE_LATENCY_SLEEP_MODE_INFO_NV };
		sleep_mode_info.lowLatencyBoost = enable;
		sleep_mode_info.lowLatencyMode = enable;
		if (table->vkSetLatencySleepModeNV(context->get_device(), swapchain, &sleep_mode_info) != VK_SUCCESS)
			LOGE("Failed to set low latency sleep mode.\n");
	}

	low_latency_mode_enable_gpu_submit = enable;
}

#ifdef HAVE_WSI_DXGI_INTEROP
bool WSI::begin_frame_dxgi()
{
	Semaphore acquire;

	while (!acquire)
	{
		if (!dxgi->acquire(acquire))
			return false;

		swapchain_index = 0;
		acquire->signal_external();
		has_acquired_swapchain_index = true;

		// Poll after acquire as well for optimal latency.
		platform->poll_input();

		// Polling input may trigger a resize event. Trying to present in that situation without ResizeBuffers
		// cause wonky issues on DXGI.
		if (platform->should_resize())
			update_framebuffer(platform->get_surface_width(), platform->get_surface_height());

		// If update_framebuffer caused a resize, we won't have an acquire index anymore, reacquire.
		if (!has_acquired_swapchain_index)
			acquire.reset();
	}

	auto wait_ts = device->write_calibrated_timestamp();
	if (!dxgi->wait_latency(present_frame_latency))
	{
		LOGE("Failed to wait on latency handle.\n");
		return false;
	}
	device->register_time_interval("WSI", std::move(wait_ts), device->write_calibrated_timestamp(),
	                               "DXGI wait latency");

	auto frame_time = platform->get_frame_timer().frame();
	auto elapsed_time = platform->get_frame_timer().get_elapsed();

	smooth_frame_time = frame_time;
	smooth_elapsed_time = elapsed_time;

	platform->event_frame_tick(frame_time, elapsed_time);
	platform->event_swapchain_index(device.get(), swapchain_index);
	device->set_acquire_semaphore(swapchain_index, std::move(acquire));

	return true;
}
#endif

void WSI::update_present_timing_properties()
{
	VkSwapchainTimingPropertiesEXT props = { VK_STRUCTURE_TYPE_SWAPCHAIN_TIMING_PROPERTIES_EXT };
	uint64_t counter = 0;
	if (table->vkGetSwapchainTimingPropertiesEXT(context->get_device(), swapchain, &props, &counter) != VK_SUCCESS)
		return;

	if (counter == present_timing.refresh_counter && present_timing.has_refresh_feedback)
		return;

	present_timing.refresh_counter = counter;
	present_timing.refresh_duration = props.refreshDuration;
	present_timing.refresh_interval = props.refreshInterval;

	const char *refresh_mode = "Unknown";

	if (props.refreshInterval == 0)
	{
		// Wayland issue: It cannot figure out VRR vs FRR. Seems like we will need presentation-timing v3.
		present_timing.refresh_mode = RefreshMode::Unknown;
	}
	else if (props.refreshInterval == UINT64_MAX)
	{
		present_timing.refresh_mode = RefreshMode::VRR;
		refresh_mode = "VRR";
	}
	else
	{
		present_timing.refresh_mode = RefreshMode::FRR;
		refresh_mode = "FRR";
	}

	if (present_timing.refresh_duration)
	{
		(void)refresh_mode;
#ifdef VULKAN_DEBUG
		LOGI("Present timing (count %llu): detected refresh duration of %llu nsec (%.6f Hz), mode %s.\n",
			 static_cast<unsigned long long>(counter),
		     static_cast<unsigned long long>(present_timing.refresh_duration),
		     1e9 / double(present_timing.refresh_duration), refresh_mode);
#endif
	}

	present_timing.has_refresh_feedback = true;
}

void WSI::recalibrate_present_timing_domains()
{
	if (!context->get_enabled_device_features().supports_calibrated_timestamps)
		return;

	present_timing.calibration.clear();

	uint32_t count;
	vkGetPhysicalDeviceCalibrateableTimeDomainsKHR(context->get_gpu(), &count, nullptr);
	Util::SmallVector<VkTimeDomainKHR> domains(count);
	vkGetPhysicalDeviceCalibrateableTimeDomainsKHR(context->get_gpu(), &count, domains.data());
	domains.resize(count);

#ifdef _WIN32
	constexpr auto host_domain = VK_TIME_DOMAIN_QUERY_PERFORMANCE_COUNTER_KHR;
#else
	constexpr auto host_domain = VK_TIME_DOMAIN_CLOCK_MONOTONIC_KHR;
#endif

	if (std::find(domains.begin(), domains.end(), host_domain) == domains.end())
	{
		LOGW("Cannot calibrate timestamp domain %u\n", host_domain);
		return;
	}

	present_timing.calibration.resize(present_timing.time_domains.size());

	for (size_t i = 0, n = present_timing.time_domains.size(); i < n; i++)
	{
		if (present_timing.time_domains[i] == host_domain)
		{
			present_timing.calibration[i] = { 1, { 1, 1, 1, 1 } };
			continue;
		}

		if (std::find(domains.begin(), domains.end(), present_timing.time_domains[i]) == domains.end())
		{
			LOGW("Cannot calibrate timestamp domain %u\n", host_domain);
			return;
		}

		VkCalibratedTimestampInfoKHR infos[2] = {};
		infos[0].sType = VK_STRUCTURE_TYPE_CALIBRATED_TIMESTAMP_INFO_KHR;
		infos[1].sType = VK_STRUCTURE_TYPE_CALIBRATED_TIMESTAMP_INFO_KHR;
		infos[0].timeDomain = host_domain;
		infos[1].timeDomain = present_timing.time_domains[i];
		uint64_t timestamps[2] = {};
		uint64_t max_deviation;

		VkSwapchainCalibratedTimestampInfoEXT swapchain_info =
				{ VK_STRUCTURE_TYPE_SWAPCHAIN_CALIBRATED_TIMESTAMP_INFO_EXT };

		if (present_timing.time_domains[i] == VK_TIME_DOMAIN_PRESENT_STAGE_LOCAL_EXT)
		{
			// Every present stage has a unique time domain, query each stage one-by-one.
			infos[1].pNext = &swapchain_info;
			swapchain_info.timeDomainId = present_timing.time_domain_ids[i];
			swapchain_info.swapchain = swapchain;

			Util::for_each_bit(supports_present_timing.feedback, [&](unsigned bit) {
				swapchain_info.presentStage = 1u << bit;

				if (table->vkGetCalibratedTimestampsKHR(context->get_device(), 2, infos,
				                                        timestamps, &max_deviation) != VK_SUCCESS)
				{
					LOGE("Failed to get calibrated timestamps.\n");
					return;
				}

				present_timing.calibration[i].host_time = timestamps[0];
				present_timing.calibration[i].stage_times[bit] = timestamps[1];
			});
		}
		else if (present_timing.time_domains[i] == VK_TIME_DOMAIN_SWAPCHAIN_LOCAL_EXT)
		{
			infos[1].pNext = &swapchain_info;
			swapchain_info.timeDomainId = present_timing.time_domain_ids[i];
			swapchain_info.swapchain = swapchain;

			if (table->vkGetCalibratedTimestampsKHR(context->get_device(), 2, infos,
			                                        timestamps, &max_deviation) != VK_SUCCESS)
			{
				LOGE("Failed to get calibrated timestamps.\n");
				return;
			}

			present_timing.calibration[i].host_time = timestamps[0];
			Util::for_each_bit(supports_present_timing.feedback, [&](unsigned bit) {
				present_timing.calibration[i].stage_times[bit] = timestamps[1];
			});
		}
		else
		{
			// This probably needs spec clarification.
			// We assume it's normal nanoseconds for now.
			if (infos[1].timeDomain == VK_TIME_DOMAIN_DEVICE_KHR &&
			    (context->get_gpu_props().limits.timestampPeriod != 1.0f ||
			     context->get_queue_info().timestamp_valid_bits != 64))
			{
				LOGW("Implementation reports DEVICE domain timestamps, "
					 "but it's unclear how to deal with non-trivial periods and valid bits.\n");
			}

			if (table->vkGetCalibratedTimestampsKHR(context->get_device(), 2, infos,
													timestamps, &max_deviation) != VK_SUCCESS)
			{
				LOGE("Failed to get calibrated timestamps.\n");
				return;
			}

			present_timing.calibration[i].host_time = timestamps[0];
			Util::for_each_bit(supports_present_timing.feedback, [&](unsigned bit) {
				present_timing.calibration[i].stage_times[bit] = timestamps[1];
			});
		}
	}

#ifdef _WIN32
	LARGE_INTEGER freq;
	QueryPerformanceFrequency(&freq);
	for (auto &calibration : present_timing.calibration)
		calibration.host_time = uint64_t(1e9 * calibration.host_time / double(freq.QuadPart));
#endif

	if (present_timing.present_stage == 0)
	{
		// Make sure that our requests use a supported time domain, otherwise NV driver gets confused.
		for (size_t i = 0, n = present_timing.calibration.size(); i < n; i++)
		{
			auto &cal = present_timing.calibration[i];
			if (cal.host_time &&
			    std::any_of(std::begin(cal.stage_times), std::end(cal.stage_times), [](uint64_t v) { return v != 0; }))
			{
				present_timing.time_domain = present_timing.time_domains[i];
				present_timing.time_domain_id = present_timing.time_domain_ids[i];
				break;
			}
		}
	}
}

void WSI::update_time_domain_properties()
{
	VkSwapchainTimeDomainPropertiesEXT time_domain_properties = { VK_STRUCTURE_TYPE_SWAPCHAIN_TIME_DOMAIN_PROPERTIES_EXT };
	time_domain_properties.timeDomainCount = 1; // Workaround VVL.
	if (table->vkGetSwapchainTimeDomainPropertiesEXT(context->get_device(), swapchain, &time_domain_properties, nullptr) != VK_SUCCESS)
	{
		LOGE("Failed to query time domain properties.\n");
		return;
	}

	present_timing.time_domains.resize(time_domain_properties.timeDomainCount);
	present_timing.time_domain_ids.resize(time_domain_properties.timeDomainCount);
	time_domain_properties.pTimeDomains = present_timing.time_domains.data();
	time_domain_properties.pTimeDomainIds = present_timing.time_domain_ids.data();

	if (table->vkGetSwapchainTimeDomainPropertiesEXT(context->get_device(), swapchain,
	                                                 &time_domain_properties,
	                                                 &present_timing.time_domain_counter) != VK_SUCCESS)
	{
		LOGE("Failed to query time domain properties.\n");
		return;
	}

#ifdef VULKAN_DEBUG
	for (uint32_t i = 0; i < time_domain_properties.timeDomainCount; i++)
	{
		LOGI("Got time domain %u, ID %llu\n",
		     present_timing.time_domains[i],
		     static_cast<unsigned long long>(present_timing.time_domain_ids[i]));
	}
#endif

	present_timing.has_time_domain_props = true;
	present_timing.need_recalibration = true;
}

void WSI::poll_present_timing_feedback()
{
	VkPastPresentationTimingEXT timings[PresentTimingQueueSize] = {};
	VkPresentStageTimeEXT stage_time[PresentTimingQueueSize][4] = {};

	for (uint32_t i = 0; i < PresentTimingQueueSize; i++)
	{
		auto &t = timings[i];
		t.sType = VK_STRUCTURE_TYPE_PAST_PRESENTATION_TIMING_EXT;
		t.pPresentStages = stage_time[i];
		t.presentStageCount = 4;

		// VVL workaround
		for (auto &s : stage_time[i])
			s.stage = VK_PRESENT_STAGE_QUEUE_OPERATIONS_END_BIT_EXT;
	}

	VkPastPresentationTimingInfoEXT info = { VK_STRUCTURE_TYPE_PAST_PRESENTATION_TIMING_INFO_EXT };
	VkPastPresentationTimingPropertiesEXT props = { VK_STRUCTURE_TYPE_PAST_PRESENTATION_TIMING_PROPERTIES_EXT };
	props.presentationTimingCount = PresentTimingQueueSize;
	props.pPresentationTimings = timings;
	info.swapchain = swapchain;

	if (table->vkGetPastPresentationTimingEXT(context->get_device(), &info, &props) != VK_SUCCESS)
		return;

	if (props.presentationTimingCount == 0)
		return;

	if (props.timingPropertiesCounter != present_timing.refresh_counter)
		update_present_timing_properties();

	// NV bug on X11: props.timingPropertiesCounter is always 0, even if TimingPropertiesEXT returns 1.
	if (props.timingPropertiesCounter != 0 && props.timingPropertiesCounter != present_timing.refresh_counter)
	{
		LOGW("Got presentation timing counter (%llu) which does not map to current state of swapchain (%llu).\n",
		     static_cast<unsigned long long>(props.timingPropertiesCounter),
		     static_cast<unsigned long long>(present_timing.refresh_counter));
	}

	// NV bug on X11: timeDomainsCount remains 0.
	if ((props.timeDomainsCounter != 0 && props.timeDomainsCounter != present_timing.time_domain_counter) ||
	    !present_timing.has_time_domain_props)
	{
		update_time_domain_properties();
	}

	auto current_time = Util::get_current_time_nsecs();
	if (present_timing.last_recalibration_time + 1000000000 < current_time)
	{
		present_timing.need_recalibration = true;
		present_timing.last_recalibration_time = current_time;
	}

	if (present_timing.need_recalibration)
	{
		recalibrate_present_timing_domains();
		present_timing.need_recalibration = false;
	}

	for (uint32_t i = 0; i < props.presentationTimingCount; i++)
	{
		// By default, these reports must be sorted based on QueuePresent().

		auto &timing = props.pPresentationTimings[i];

		if (!timing.reportComplete)
		{
			// This should never happen since we don't request partial timestamps.
			LOGE("Implementation does not report complete timestamps?\n");
			continue;
		}

#ifdef VULKAN_DEBUG
		LOGI("Timing for presentID %llu, time domain %u, time domain ID %llu:\n",
		     static_cast<unsigned long long>(timing.presentId),
		     timing.timeDomain, static_cast<unsigned long long>(timing.timeDomainId));
#endif

		present_timing.present_stage = 0;
		present_timing.reference_time = 0;
		present_timing.present_id = timing.presentId;
		present_timing.time_domain = timing.timeDomain;
		present_timing.time_domain_id = timing.timeDomainId;

		// Try to calibrate the timestamp so we can report them in host time domain.
		const CalibratedTimestamp *calibrated = nullptr;
		for (size_t j = 0, n = present_timing.time_domain_ids.size(); j < n; j++)
		{
			if (present_timing.time_domain_ids[j] == timing.timeDomainId)
			{
				if (j < present_timing.calibration.size())
					calibrated = &present_timing.calibration[j];
				break;
			}
		}

		if (calibrated && calibrated->host_time == 0)
			calibrated = nullptr;

		std::sort(timing.pPresentStages, timing.pPresentStages + timing.presentStageCount,
		          [](const VkPresentStageTimeEXT &a, const VkPresentStageTimeEXT &b) { return a.stage < b.stage; });

		present_timing.present_done_host_time = 0;
		present_timing.gpu_done_host_time = 0;

		for (uint32_t stage_index = 0; stage_index < timing.presentStageCount; stage_index++)
		{
			auto &stage = timing.pPresentStages[stage_index];

			// Safety. Shouldn't happen since we don't request it.
			// Time of 0 can happen according to spec and means timing information is not available.
			// Can happen for e.g. discarded, occluded surfaces on WL. Just skip the update.
			if (stage.stage > VK_PRESENT_STAGE_IMAGE_FIRST_PIXEL_VISIBLE_BIT_EXT || stage.stage == 0 || stage.time == 0)
				continue;

			uint64_t calibrated_stage_time = calibrated ? calibrated->stage_times[Util::trailing_zeroes(stage.stage)] : 0;
#ifdef VULKAN_DEBUG
			static const char *stage_tags[] = { "QueueOperations", "Dequeued", "FirstPixelOut", "FirstPixelVisible" };
			const char *stage_tag = stage_tags[Util::trailing_zeroes(stage.stage)];
#endif

			if (calibrated)
			{
				uint64_t calibrated_ts = calibrated->host_time + (stage.time - calibrated_stage_time);
				if (stage.stage == VK_PRESENT_STAGE_QUEUE_OPERATIONS_END_BIT_EXT)
					present_timing.gpu_done_host_time = calibrated_ts;
				else
					present_timing.present_done_host_time = calibrated_ts;
#ifdef VULKAN_DEBUG
				LOGI("  %s: %.3f s (calibrated)\n", stage_tag, calibrated_ts * 1e-9);
#endif
			}
#ifdef VULKAN_DEBUG
			else
			{
				LOGI("  %s: %llu (raw ns)\n", stage_tag, static_cast<unsigned long long>(stage.time));
			}
#endif

			present_timing.present_stage = stage.stage;
			present_timing.reference_time = stage.time;
		}

		auto itr = std::find_if(
				present_timing.error_stats.begin(), present_timing.error_stats.end(),
				[&](const ErrorStats &err) {
					return err.present_id == timing.presentId;
				});

		present_timing.presentation_time_error = 0;

		if (itr != present_timing.error_stats.end())
		{
			// Retire the pending compensation.
			present_timing.pending_compensation -= itr->compensation;

			if (present_timing.present_done_host_time && itr->target_absolute)
			{
				auto err = int64_t(present_timing.present_done_host_time) - int64_t(itr->target_absolute);
				present_timing.presentation_time_error = err;
#ifdef VULKAN_DEBUG
				LOGI("    Error: %.3f ms\n", 1e-6 * double(err));
#endif
			}

			present_timing.error_stats.erase(itr);
		}

		frr_pacer.set_frame_time_ns(present_timing.refresh_duration);

		if (present_timing.present_done_host_time && present_timing.gpu_done_host_time)
		{
			frr_pacer.update_feedback(present_timing.present_id, present_timing.gpu_done_host_time,
			                          present_timing.present_done_host_time);
		}
	}
}

void WSI::set_present_timing_request(VkPresentTimingInfoEXT &timing)
{
	// No stable way to set targets yet.
	if (present_timing.refresh_duration == 0 || present_timing.reference_time == 0 ||
	    present_timing.present_done_host_time == 0)
		return;

	// Presentation timing is only meaningful for FIFO.
	if (active_present_mode != VK_PRESENT_MODE_FIFO_KHR &&
	    active_present_mode != VK_PRESENT_MODE_FIFO_RELAXED_KHR &&
	    active_present_mode != VK_PRESENT_MODE_FIFO_LATEST_READY_KHR)
		return;

	// Not supported.
	if (!supports_present_timing.absolute && !supports_present_timing.relative)
		return;

	// No request to set time.
	if (present_timing.target_absolute_time == 0 && present_timing.target_relative_time == 0)
		return;

	uint64_t relative_time = present_timing.target_relative_time;
	if (relative_time && !present_timing.force_vrr && present_timing.refresh_mode != RefreshMode::VRR)
	{
		// If we keep accumulating relative time in a non-locked way, we'll get terrible pacing.
		// Realign the relative time to boundary unless we're in VRR mode.

		uint64_t interval = present_timing.refresh_duration;
		relative_time = std::max<uint64_t>(relative_time, interval);

		// The refresh interval represents the alignment of a refresh cycle.
		// Duration is a multiple of interval.
		if (present_timing.refresh_interval)
			interval = present_timing.refresh_interval;

		auto cycles = (relative_time + interval - 1) / interval;
		relative_time = interval * cycles;
	}

	// VRR does not have to align to boundaries, so rounding is somewhat meaningless.
	if (present_timing.refresh_mode != RefreshMode::VRR && !present_timing.force_vrr)
		timing.flags |= VK_PRESENT_TIMING_INFO_PRESENT_AT_NEAREST_REFRESH_CYCLE_BIT_EXT;

	if (present_timing.time_domain == VK_TIME_DOMAIN_PRESENT_STAGE_LOCAL_EXT)
		timing.targetTimeDomainPresentStage = present_timing.present_stage;

	uint64_t minimum_interval = present_timing.refresh_duration;

	if (supports_present_timing.relative && relative_time)
	{
		// Relative time is very nice, since it's not our job to align frames :)
		timing.flags |= VK_PRESENT_TIMING_INFO_PRESENT_AT_RELATIVE_TIME_BIT_EXT;
		timing.targetTime = relative_time;

		// If we're using pure relative timing, we don't care about error accumulation and compensation,
		// since it's expected that application will be driven by feedback rather than opposite.
	}
	else
	{
		if (!supports_present_timing.absolute)
		{
			// Ensure that we don't attempt to compensate for errors when absolute times are very tight.
			// The emulated relative time must never be smaller than duration.
			relative_time = std::max<uint64_t>(relative_time, present_timing.refresh_duration);
		}

		// If presentations get sufficiently delayed, we will need to catch up with our internal accumulator.
		// If we're using present interval > 1, we'll catch up in a few frames.
		uint64_t estimated_minimum_time = present_timing.present_done_host_time +
		                                  (present_last_id - present_timing.present_id) * minimum_interval;

		present_timing.last_absolute_target_time =
				std::max<uint64_t>(present_timing.last_absolute_target_time, estimated_minimum_time);

		// Compute the target absolute timing.
		uint64_t next_absolute_time = std::max<uint64_t>(
				present_timing.last_absolute_target_time + relative_time,
				present_timing.target_absolute_time);

		int64_t compensation = 0;

		if (supports_present_timing.absolute)
		{
			timing.targetTime = next_absolute_time;
		}
		else
		{
			if (present_timing.last_absolute_target_time)
			{
				// This is kinda crude. We're emulating absolute timestamp with relative.
				timing.targetTime = std::max<uint64_t>(next_absolute_time, present_timing.last_absolute_target_time) -
				                    present_timing.last_absolute_target_time;

				// Safety against deadlocks.
				timing.targetTime = std::min<uint64_t>(timing.targetTime, 1000 * 1000 * 1000);

				// If we lost frames, relative timing will remain delayed, so pull back the relative timing
				// to realign with absolute time.
				// The value can be negative in theory if we got back frames too early,
				// but that shouldn't really happen.
				int64_t in_flight_error = present_timing.presentation_time_error - present_timing.pending_compensation;

				VK_ASSERT(timing.targetTime >= present_timing.refresh_duration);

				// Don't aim to compensate all error in one go. That seems to create some unfortunate feedback loop effects,
				// especially on Windows. We'll accept some error as long as it means more stable pacing.
				compensation = std::min<int64_t>(timing.targetTime - present_timing.refresh_duration, in_flight_error / 2);
				timing.targetTime -= compensation;

#ifdef _WIN32
				if (device->get_device_features().driver_id == VK_DRIVER_ID_NVIDIA_PROPRIETARY &&
				    (timing.flags & VK_PRESENT_TIMING_INFO_PRESENT_AT_NEAREST_REFRESH_CYCLE_BIT_EXT) != 0)
				{
					// The driver seems very temperamental here, and it seems to round down the relative time
					// to DXGI present intervals or something (which is a bug) ... Realign the next_absolute_time exactly.
					uint64_t interval = present_timing.refresh_interval ?
						present_timing.refresh_interval : present_timing.refresh_duration;

					auto cycles = (timing.targetTime + interval / 2) / interval;
					auto new_relative_time = cycles * interval;
					auto adj = new_relative_time - timing.targetTime;
					next_absolute_time += adj;
					compensation -= adj;
					timing.targetTime = new_relative_time;
				}
#endif

				present_timing.pending_compensation += compensation;

#ifdef VULKAN_DEBUG
				LOGI("Relative target time: %.3f ms.\n", timing.targetTime * 1e-6);
				LOGI("  Abs target time: %.3f ms.\n", next_absolute_time * 1e-6);
				LOGI("  Compensation offset: %.3f ms\n", compensation * 1e-6);
#endif
			}

			timing.flags |= VK_PRESENT_TIMING_INFO_PRESENT_AT_RELATIVE_TIME_BIT_EXT;
		}

		present_timing.last_absolute_target_time = next_absolute_time;
		present_timing.error_stats.push_back({ next_present_id, next_absolute_time, compensation });

		// If we keep accumulating time through relative time emulation, we need to account for clock drift.
		if (supports_present_timing.absolute && relative_time &&
		    present_timing.refresh_mode != RefreshMode::VRR && !present_timing.force_vrr)
		{
			uint64_t interval = present_timing.refresh_interval ?
			                    present_timing.refresh_interval : present_timing.refresh_duration;

			uint64_t align =
					(present_timing.last_absolute_target_time - present_timing.present_done_host_time) % interval;

			// Try to quickly-ish align to a refresh cycle.
			if (align > interval / 2)
				present_timing.last_absolute_target_time += (interval - align) / 8;
			else
				present_timing.last_absolute_target_time -= align / 8;
		}
	}

	if ((timing.flags & VK_PRESENT_TIMING_INFO_PRESENT_AT_RELATIVE_TIME_BIT_EXT) == 0)
	{
		bool has_calibrated_time = false;
		size_t n = std::min<size_t>(present_timing.time_domain_ids.size(), present_timing.calibration.size());

		for (size_t i = 0; i < n; i++)
		{
			if (present_timing.time_domain_ids[i] == timing.timeDomainId)
			{
				auto &c = present_timing.calibration[i];

				// Spec seems to suggest that targetTime is always in terms of FIRST_PIXEL_VISIBLE,
				// but we can only use timestamps based on the latest present stage we get feedback for.
				uint64_t stage_time = c.stage_times[Util::trailing_zeroes(present_timing.present_stage)];

				if (stage_time == 0)
				{
					LOGW("Cannot compute targetTime.\n");
				}
				else
				{
					// Recompute the absolute time to target the domain.
					timing.targetTime = (timing.targetTime - c.host_time) + stage_time;

					if (timing.targetTime >= (1ull << 63))
						LOGW("Detected strange overflow, ignoring time request.\n");
					else
						has_calibrated_time = true;
				}

				break;
			}
		}

		if (!has_calibrated_time)
			timing.targetTime = 0;
	}

	// Completely meaningless to keep targeting absolute.
	present_timing.target_absolute_time = 0;
}

bool WSI::get_presentation_stats(PresentationStats &stats) const
{
	if (!present_timing.reference_time)
		return false;

	stats.feedback_present_id = present_timing.present_id;
	stats.gpu_done_ts = present_timing.gpu_done_host_time;
	stats.present_done_ts = present_timing.present_done_host_time;
	stats.error = present_timing.presentation_time_error;
	return true;
}

bool WSI::get_refresh_rate_info(RefreshRateInfo &info) const
{
	if (present_timing.refresh_duration == 0)
		return false;

	info.refresh_duration = present_timing.refresh_duration;
	info.refresh_interval = present_timing.refresh_interval;
	info.mode = present_timing.refresh_mode;
	return true;
}

uint64_t WSI::get_last_submitted_present_id() const
{
	return present_last_id;
}

bool WSI::set_target_presentation_time(uint64_t absolute_time_ns, uint64_t relative_time_ns, bool force_vrr)
{
	present_timing.target_absolute_time = absolute_time_ns;
	present_timing.target_relative_time = relative_time_ns;
	present_timing.force_vrr = force_vrr;

	if (absolute_time_ns || relative_time_ns)
		present_feedback_enable = true;

	if (active_present_mode != VK_PRESENT_MODE_FIFO_KHR &&
	    active_present_mode != VK_PRESENT_MODE_FIFO_RELAXED_KHR &&
	    active_present_mode != VK_PRESENT_MODE_FIFO_LATEST_READY_KHR)
		return false;

	return present_timing.refresh_duration != 0 && present_timing.reference_time != 0 &&
	       present_timing.present_done_host_time != 0 &&
	       (supports_present_timing.relative || supports_present_timing.absolute);
}

void WSI::set_enable_timing_feedback(bool enable)
{
	present_feedback_enable = enable;
}

void WSI::set_fixed_rate_low_latency_pacer(bool enable)
{
	frr_pacer_enable = enable;
	if (!frr_pacer_enable)
		frr_pacer.reset();
}

FixedRefreshRatePacer &WSI::get_fixed_rate_pacer()
{
	return frr_pacer;
}

bool WSI::begin_frame()
{
	if (frame_is_external)
		return begin_frame_external();

#ifdef VULKAN_WSI_TIMING_DEBUG
	auto next_frame_start = Util::get_current_time_nsecs();
#endif

	device->next_frame_context();
	external_release.reset();

#ifdef VULKAN_WSI_TIMING_DEBUG
	auto next_frame_end = Util::get_current_time_nsecs();
	LOGI("Waited for vacant frame context for %.3f ms.\n", (next_frame_end - next_frame_start) * 1e-6);
#endif

#ifdef HAVE_WSI_DXGI_INTEROP
	if (dxgi)
	{
		if (platform->should_resize())
			update_framebuffer(platform->get_surface_width(), platform->get_surface_height());

		if (has_acquired_swapchain_index)
			return true;
		return begin_frame_dxgi();
	}
	else
#endif
	{
		if (swapchain == VK_NULL_HANDLE || platform->should_resize() || swapchain_is_suboptimal)
			update_framebuffer(platform->get_surface_width(), platform->get_surface_height());
		if (has_acquired_swapchain_index)
			return true;
	}

	if (swapchain == VK_NULL_HANDLE)
	{
		LOGE("Completely lost swapchain. Cannot continue.\n");
		return false;
	}

	VkResult result;
	do
	{
		auto acquire = device->request_semaphore(VK_SEMAPHORE_TYPE_BINARY);

#ifdef VULKAN_WSI_TIMING_DEBUG
		auto acquire_start = Util::get_current_time_nsecs();
#endif

		Fence fence;

		// TODO: Improve this with fancier approaches as needed.
		if (low_latency_mode_enable_present &&
		    !device->get_device_features().present_wait_features.presentWait &&
		    !supports_present_wait2 &&
		    current_present_mode == PresentMode::SyncToVBlank)
		{
			fence = device->request_legacy_fence();
		}

		auto acquire_ts = device->write_calibrated_timestamp();
		result = table->vkAcquireNextImageKHR(context->get_device(), swapchain, UINT64_MAX, acquire->get_semaphore(),
		                                      fence ? fence->get_fence() : VK_NULL_HANDLE, &swapchain_index);
		device->register_time_interval("WSI", std::move(acquire_ts), device->write_calibrated_timestamp(), "acquire");

		if (fence)
			fence->wait();

#if defined(ANDROID)
		// Android 10 can return suboptimal here, only because of pre-transform.
		// We don't care about that, and treat this as success.
		if (result == VK_SUBOPTIMAL_KHR && !support_prerotate)
			result = VK_SUCCESS;
#endif

		if (result == VK_ERROR_FULL_SCREEN_EXCLUSIVE_MODE_LOST_EXT)
		{
			LOGE("Lost exclusive full-screen ...\n");
		}

#ifdef VULKAN_WSI_TIMING_DEBUG
		auto acquire_end = Util::get_current_time_nsecs();
		LOGI("vkAcquireNextImageKHR took %.3f ms.\n", (acquire_end - acquire_start) * 1e-6);
#endif

		if (result == VK_SUBOPTIMAL_KHR)
		{
#ifdef VULKAN_DEBUG
			LOGI("AcquireNextImageKHR is suboptimal, will recreate.\n");
#endif
			swapchain_is_suboptimal = true;
			LOGW("Swapchain suboptimal.\n");
		}

		if (result >= 0)
		{
			has_acquired_swapchain_index = true;
			acquire->signal_external();

			// WSI signals this, which exists outside the domain of our Vulkan queues.
			acquire->set_signal_is_foreign_queue();

			wait_swapchain_latency();

			if (supports_present_timing.feedback)
			{
				update_present_timing_properties();
				poll_present_timing_feedback();
			}

			if (frr_pacer_enable)
			{
				auto pace_start_ts = device->write_calibrated_timestamp();
				frr_pacer.begin_frame_submission(next_present_id);
				device->register_time_interval(
					"WSI", std::move(pace_start_ts), device->write_calibrated_timestamp(),
					"pacer-wait");
			}

			auto frame_time = platform->get_frame_timer().frame();
			auto elapsed_time = platform->get_frame_timer().get_elapsed();

			smooth_frame_time = frame_time;
			smooth_elapsed_time = elapsed_time;

			// Poll after acquire as well for optimal latency.
			platform->poll_input();
			platform->event_frame_tick(frame_time, elapsed_time);

			platform->event_swapchain_index(device.get(), swapchain_index);

			device->set_acquire_semaphore(swapchain_index, acquire);
			if (device->get_device_features().present_id_features.presentId ||
			    device->get_device_features().present_id2_features.presentId2)
			{
				device->set_present_id(swapchain, next_present_id);
			}
		}
		else if (result == VK_ERROR_OUT_OF_DATE_KHR || result == VK_ERROR_FULL_SCREEN_EXCLUSIVE_MODE_LOST_EXT)
		{
			LOGW("Swapchain out of date.\n");
			VK_ASSERT(swapchain_width != 0);
			VK_ASSERT(swapchain_height != 0);

			tear_down_swapchain();

			if (!blocking_init_swapchain(swapchain_width, swapchain_height))
				return false;
			device->init_swapchain(swapchain_images, swapchain_width, swapchain_height,
			                       swapchain_surface_format.format, swapchain_current_prerotate,
			                       current_extra_usage | VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT);
		}
		else
		{
			return false;
		}
	} while (result < 0);
	return true;
}

#ifdef HAVE_WSI_DXGI_INTEROP
bool WSI::end_frame_dxgi()
{
	auto release = device->consume_release_semaphore();
	VK_ASSERT(release);
	VK_ASSERT(release->is_signalled());
	VK_ASSERT(!release->is_pending_wait());
	return dxgi->present(std::move(release), current_present_mode == PresentMode::SyncToVBlank);
}
#endif

void WSI::set_frame_duplication_aware(bool enable, uint32_t target_image_count)
{
	frame_dupe_aware = enable;
	frame_dupe_target_images = target_image_count;

	if (!has_acquired_swapchain_index && (current_frame_dupe_aware != frame_dupe_aware ||
	                                      current_frame_dupe_target_images != frame_dupe_target_images))
	{
		current_frame_dupe_aware = frame_dupe_aware;
		current_frame_dupe_target_images = frame_dupe_target_images;
		update_framebuffer(swapchain_width, swapchain_height);
	}
}

void WSI::set_next_present_is_duplicated()
{
	next_present_is_dupe = true;
}

bool WSI::end_frame()
{
	device->end_frame_context();

	// Take ownership of the release semaphore so that the external user can use it.
	if (frame_is_external)
	{
		// If we didn't render into the swapchain this frame, we will return a blank semaphore.
		external_release = device->consume_release_semaphore();
		VK_ASSERT(!external_release || external_release->is_signalled());
		frame_is_external = false;
	}
	else
	{
		if (!device->swapchain_touched())
			return true;

		emit_end_of_frame_markers();
		has_acquired_swapchain_index = false;

#ifdef HAVE_WSI_DXGI_INTEROP
		if (dxgi)
			return end_frame_dxgi();
#endif

		auto release = device->consume_release_semaphore();
		VK_ASSERT(release);
		VK_ASSERT(release->is_signalled());
		VK_ASSERT(!release->is_pending_wait());

		auto release_semaphore = release->get_semaphore();
		VK_ASSERT(release_semaphore != VK_NULL_HANDLE);

		VkResult result = VK_SUCCESS;
		VkPresentInfoKHR info = { VK_STRUCTURE_TYPE_PRESENT_INFO_KHR };
		info.waitSemaphoreCount = 1;
		info.pWaitSemaphores = &release_semaphore;
		info.swapchainCount = 1;
		info.pSwapchains = &swapchain;
		info.pImageIndices = &swapchain_index;
		info.pResults = &result;

		VkSwapchainPresentFenceInfoKHR present_fence = { VK_STRUCTURE_TYPE_SWAPCHAIN_PRESENT_FENCE_INFO_KHR };
		VkSwapchainPresentModeInfoKHR present_mode_info = { VK_STRUCTURE_TYPE_SWAPCHAIN_PRESENT_MODE_INFO_KHR };
		VkPresentIdKHR present_id_info = { VK_STRUCTURE_TYPE_PRESENT_ID_KHR };
		VkPresentId2KHR present_id2_info = { VK_STRUCTURE_TYPE_PRESENT_ID_2_KHR };
		VkPresentTimingsInfoEXT timings_info = { VK_STRUCTURE_TYPE_PRESENT_TIMINGS_INFO_EXT };
		VkPresentTimingInfoEXT timing_info = { VK_STRUCTURE_TYPE_PRESENT_TIMING_INFO_EXT };

		if (supports_present_wait2)
		{
			present_id2_info.swapchainCount = 1;
			present_id2_info.pPresentIds = &next_present_id;
			present_id2_info.pNext = info.pNext;
			info.pNext = &present_id2_info;
		}
		else if (device->get_device_features().present_id_features.presentId)
		{
			present_id_info.swapchainCount = 1;
			present_id_info.pPresentIds = &next_present_id;
			present_id_info.pNext = info.pNext;
			info.pNext = &present_id_info;
		}

		// If we can, just promote the new presentation mode right away.
		update_active_presentation_mode(present_mode);

		if (device->get_device_features().swapchain_maintenance1_features.swapchainMaintenance1)
		{
			if (!device->get_workarounds().broken_present_fence)
			{
				last_present_fence = device->request_legacy_fence();
				present_fence.swapchainCount = 1;
				present_fence.pFences = &last_present_fence->get_fence();
				present_fence.pNext = const_cast<void *>(info.pNext);
				info.pNext = &present_fence;
			}

			present_mode_info.swapchainCount = 1;
			present_mode_info.pPresentModes = &active_present_mode;
			present_mode_info.pNext = const_cast<void *>(info.pNext);
			info.pNext = &present_mode_info;
		}

		if (supports_present_timing.feedback && present_feedback_enable)
		{
			timing_info.presentStageQueries = supports_present_timing.feedback;
			timing_info.timeDomainId = present_timing.time_domain_id;
			set_present_timing_request(timing_info);

			timings_info.swapchainCount = 1;
			timings_info.pTimingInfos = &timing_info;
			timings_info.pNext = info.pNext;
			info.pNext = &timings_info;
		}

#ifdef VULKAN_WSI_TIMING_DEBUG
		auto present_start = Util::get_current_time_nsecs();
#endif

		auto present_ts = device->write_calibrated_timestamp();

		device->external_queue_lock();
		emit_marker_pre_present();
#if defined(ANDROID) && defined(HAVE_SWAPPY)
		VkResult overall = SwappyVk_queuePresent(device->get_current_present_queue(), &info);
#else
		VkResult overall = table->vkQueuePresentKHR(device->get_current_present_queue(), &info);
#endif
		emit_marker_post_present();
		device->external_queue_unlock();

		device->register_time_interval("WSI", std::move(present_ts), device->write_calibrated_timestamp(), "present");

#if defined(ANDROID)
		// Android 10 can return suboptimal here, only because of pre-transform.
		// We don't care about that, and treat this as success.
		if (overall == VK_SUBOPTIMAL_KHR && !support_prerotate)
			overall = VK_SUCCESS;
		if (result == VK_SUBOPTIMAL_KHR && !support_prerotate)
			result = VK_SUCCESS;
#endif

		if (overall == VK_ERROR_FULL_SCREEN_EXCLUSIVE_MODE_LOST_EXT ||
		    result == VK_ERROR_FULL_SCREEN_EXCLUSIVE_MODE_LOST_EXT)
		{
			LOGE("Lost exclusive full-screen ...\n");
		}

#ifdef VULKAN_WSI_TIMING_DEBUG
		auto present_end = Util::get_current_time_nsecs();
		LOGI("vkQueuePresentKHR took %.3f ms.\n", (present_end - present_start) * 1e-6);
#endif

		bool dupes_frame = next_present_is_dupe && current_frame_dupe_aware && !low_latency_mode_enable_present;

		// The presentID only seems to get updated if QueuePresent returns success.
		// This makes sense I guess. Record the latest present ID which was successfully presented
		// so we don't risk deadlock.
		if ((result == VK_SUCCESS || result == VK_SUBOPTIMAL_KHR) &&
		    (device->get_device_features().present_id_features.presentId || supports_present_wait2) &&
		    !dupes_frame)
		{
			present_last_id = next_present_id;
		}

		next_present_id++;
		next_present_is_dupe = false;

		if (dupes_frame)
		{
			duplicated_frames++;
		}
		else
		{
			last_duplicated_frames = duplicated_frames;
			duplicated_frames = 0;
		}

		if (overall == VK_SUBOPTIMAL_KHR || result == VK_SUBOPTIMAL_KHR)
		{
#ifdef VULKAN_DEBUG
			LOGI("QueuePresent is suboptimal, will recreate.\n");
#endif
			swapchain_is_suboptimal = true;
		}

		// The present semaphore is consumed even on OUT_OF_DATE, etc.
		release->wait_external();

		if (!device->get_workarounds().broken_present_fence &&
		    device->get_device_features().swapchain_maintenance1_features.swapchainMaintenance1)
		{
			deferred_semaphore.push_back({ std::move(release), last_present_fence });
			release = {};
		}

		if (overall < 0 || result < 0)
		{
			LOGE("vkQueuePresentKHR failed.\n");
			release.reset();
			tear_down_swapchain();
			return false;
		}
		else
		{
			// Cannot release the WSI wait semaphore until we observe that the image has been
			// waited on again.
			// Could make this a bit tighter with swapchain_maintenance1, but not that important here.
			release_semaphores[swapchain_index] = std::move(release);
		}

		// Re-init swapchain.
		if (present_mode != current_present_mode ||
		    has_backbuffer_format_delta() ||
		    extra_usage != current_extra_usage ||
		    compression.type != current_compression.type ||
		    compression.fixed_rates != current_compression.fixed_rates ||
		    frame_dupe_aware != current_frame_dupe_aware ||
		    frame_dupe_target_images != current_frame_dupe_target_images)
		{
			current_present_mode = present_mode;
			current_backbuffer_format = backbuffer_format;
			current_extra_usage = extra_usage;
			current_compression = compression;
			current_custom_backbuffer_format = custom_backbuffer_format;
			current_frame_dupe_aware = frame_dupe_aware;
			current_frame_dupe_target_images = frame_dupe_target_images;
			update_framebuffer(swapchain_width, swapchain_height);
		}
	}

	nonblock_delete_swapchain_resources();
	return true;
}

bool WSI::has_backbuffer_format_delta() const
{
	bool has_format_delta = backbuffer_format != current_backbuffer_format;
	if (!has_format_delta && backbuffer_format == BackbufferFormat::Custom)
	{
		has_format_delta = current_custom_backbuffer_format.format != custom_backbuffer_format.format ||
		                   current_custom_backbuffer_format.colorSpace != custom_backbuffer_format.colorSpace;
	}

	return has_format_delta;
}

void WSI::update_framebuffer(unsigned width, unsigned height)
{
	if (context && device)
	{
#ifdef HAVE_WSI_DXGI_INTEROP
		if (dxgi)
		{
			if (!init_surface_swapchain_dxgi(width, height))
				LOGE("Failed to resize DXGI swapchain.\n");
		}
		else
#endif
		{
			drain_swapchain(false);
			if (blocking_init_swapchain(width, height))
			{
				device->init_swapchain(swapchain_images, swapchain_width, swapchain_height,
				                       swapchain_surface_format.format, swapchain_current_prerotate,
				                       current_extra_usage | VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT);
			}
		}
	}

	if (platform)
		platform->notify_current_swapchain_dimensions(swapchain_width, swapchain_height);
}

bool WSI::update_active_presentation_mode(PresentMode mode)
{
	if (current_present_mode == mode)
		return true;

#ifdef HAVE_WSI_DXGI_INTEROP
	// We set this on Present time.
	if (dxgi)
	{
		current_present_mode = mode;
		return true;
	}
#endif

	for (auto m : present_mode_compat_group)
	{
		bool match = false;
		switch (m)
		{
		case VK_PRESENT_MODE_FIFO_KHR:
			match = mode == PresentMode::SyncToVBlank;
			break;

		case VK_PRESENT_MODE_IMMEDIATE_KHR:
			match = mode == PresentMode::UnlockedMaybeTear ||
			        mode == PresentMode::UnlockedForceTearing;
			break;

		case VK_PRESENT_MODE_MAILBOX_KHR:
			match = mode == PresentMode::UnlockedNoTearing ||
			        mode == PresentMode::UnlockedMaybeTear;
			break;

		default:
			break;
		}

		if (match)
		{
			active_present_mode = m;
			current_present_mode = mode;
			return true;
		}
	}

	return false;
}

void WSI::set_present_mode(PresentMode mode)
{
	present_mode = mode;
	if (!has_acquired_swapchain_index && present_mode != current_present_mode)
	{
		if (!update_active_presentation_mode(present_mode))
		{
			current_present_mode = present_mode;
			update_framebuffer(swapchain_width, swapchain_height);
		}
	}
}

void WSI::set_extra_usage_flags(VkImageUsageFlags usage)
{
	extra_usage = usage;
	if (!has_acquired_swapchain_index && extra_usage != current_extra_usage)
	{
		current_extra_usage = extra_usage;
		update_framebuffer(swapchain_width, swapchain_height);
	}
}

void WSI::set_backbuffer_format(BackbufferFormat format)
{
	backbuffer_format = format;

	if (!has_acquired_swapchain_index && has_backbuffer_format_delta())
	{
		current_backbuffer_format = backbuffer_format;
		current_custom_backbuffer_format = custom_backbuffer_format;
		update_framebuffer(swapchain_width, swapchain_height);
	}
}

void WSI::set_image_compression_control(const ImageCompression &comp)
{
	if (device && !device->get_device_features().image_compression_control_swapchain_features.imageCompressionControlSwapchain)
		return;

	compression = comp;
	if (!has_acquired_swapchain_index &&
	    (compression.type != current_compression.type ||
	     compression.fixed_rates != current_compression.fixed_rates))
	{
		current_compression = compression;
		update_framebuffer(swapchain_width, swapchain_height);
	}
}

void WSI::set_custom_backbuffer_format(VkSurfaceFormatKHR format)
{
	custom_backbuffer_format = format;
	set_backbuffer_format(BackbufferFormat::Custom);
}

void WSI::set_backbuffer_srgb(bool enable)
{
	set_backbuffer_format(enable ? BackbufferFormat::sRGB : BackbufferFormat::UNORM);
}

void WSI::teardown()
{
	low_latency_semaphore.reset();

	if (platform)
		platform->release_resources();

	if (context)
		tear_down_swapchain();

	if (surface != VK_NULL_HANDLE)
	{
		platform->destroy_surface(context->get_instance(), surface);
		surface = VK_NULL_HANDLE;
	}

	if (platform)
		platform->event_device_destroyed();
	external_release.reset();
	external_acquire.reset();
	external_swapchain_images.clear();
	device.reset();
	context.reset();
}

bool WSI::blocking_init_swapchain(unsigned width, unsigned height)
{
	SwapchainError err;
	unsigned retry_counter = 0;
	do
	{
		swapchain_aspect_ratio = platform->get_aspect_ratio();
		err = init_swapchain(width, height);

		if (err != SwapchainError::None)
			platform->notify_current_swapchain_dimensions(0, 0);

		if (err == SwapchainError::Error)
		{
			if (++retry_counter > 3)
				return false;

			// Try to not reuse the swapchain.
			tear_down_swapchain();
		}
		else if (err == SwapchainError::NoSurface)
		{
			LOGW("WSI cannot make forward progress due to minimization, blocking ...\n");
			device->set_enable_async_thread_frame_context(true);
			platform->block_until_wsi_forward_progress(*this);
			device->set_enable_async_thread_frame_context(false);
			LOGW("Woke up!\n");
		}
	} while (err != SwapchainError::None);

	frr_pacer.reset();
	return swapchain != VK_NULL_HANDLE;
}

VkSurfaceFormatKHR WSI::find_suitable_present_format(const std::vector<VkSurfaceFormatKHR> &formats, BackbufferFormat desired_format) const
{
	size_t format_count = formats.size();
	VkSurfaceFormatKHR format = { VK_FORMAT_UNDEFINED };

	VkFormatFeatureFlags features = VK_FORMAT_FEATURE_COLOR_ATTACHMENT_BIT |
	                                VK_FORMAT_FEATURE_COLOR_ATTACHMENT_BLEND_BIT;
	if ((current_extra_usage & VK_IMAGE_USAGE_STORAGE_BIT) != 0)
		features |= VK_FORMAT_FEATURE_STORAGE_IMAGE_BIT;

	if (format_count == 0)
	{
		LOGE("Surface has no formats?\n");
		return format;
	}

	for (size_t i = 0; i < format_count; i++)
	{
		if (!device->image_format_is_supported(formats[i].format, features))
			continue;

		if (desired_format == BackbufferFormat::Custom)
		{
			if (formats[i].colorSpace == current_custom_backbuffer_format.colorSpace &&
			    formats[i].format == current_custom_backbuffer_format.format)
			{
				format = formats[i];
				break;
			}
		}
		else if (desired_format == BackbufferFormat::DisplayP3)
		{
			if (formats[i].colorSpace == VK_COLOR_SPACE_DISPLAY_P3_NONLINEAR_EXT &&
			    (formats[i].format == VK_FORMAT_A2B10G10R10_UNORM_PACK32 ||
			     formats[i].format == VK_FORMAT_A2R10G10B10_UNORM_PACK32))
			{
				format = formats[i];
				break;
			}
		}
		else if (desired_format == BackbufferFormat::UNORMPassthrough)
		{
			if (formats[i].colorSpace == VK_COLOR_SPACE_PASS_THROUGH_EXT &&
			    (formats[i].format == VK_FORMAT_R8G8B8A8_UNORM ||
			     formats[i].format == VK_FORMAT_B8G8R8A8_UNORM ||
			     formats[i].format == VK_FORMAT_A2B10G10R10_UNORM_PACK32 ||
			     formats[i].format == VK_FORMAT_A2R10G10B10_UNORM_PACK32))
			{
				format = formats[i];
				break;
			}
		}
		else if (desired_format == BackbufferFormat::HDR10)
		{
			if (formats[i].colorSpace == VK_COLOR_SPACE_HDR10_ST2084_EXT &&
			    (formats[i].format == VK_FORMAT_A2B10G10R10_UNORM_PACK32 ||
			     formats[i].format == VK_FORMAT_A2R10G10B10_UNORM_PACK32))
			{
				format = formats[i];
				break;
			}
		}
		else if (desired_format == BackbufferFormat::scRGB)
		{
			if (formats[i].colorSpace == VK_COLOR_SPACE_EXTENDED_SRGB_LINEAR_EXT &&
			    formats[i].format == VK_FORMAT_R16G16B16A16_SFLOAT)
			{
				format = formats[i];
				break;
			}
		}
		else if (desired_format == BackbufferFormat::sRGB)
		{
			if (formats[i].colorSpace == VK_COLOR_SPACE_SRGB_NONLINEAR_KHR &&
			    (formats[i].format == VK_FORMAT_R8G8B8A8_SRGB ||
			     formats[i].format == VK_FORMAT_B8G8R8A8_SRGB ||
			     formats[i].format == VK_FORMAT_A8B8G8R8_SRGB_PACK32))
			{
				format = formats[i];
				break;
			}
		}
		else
		{
			if (formats[i].colorSpace == VK_COLOR_SPACE_SRGB_NONLINEAR_KHR &&
			    (formats[i].format == VK_FORMAT_R8G8B8A8_UNORM ||
			     formats[i].format == VK_FORMAT_B8G8R8A8_UNORM ||
			     formats[i].format == VK_FORMAT_A2B10G10R10_UNORM_PACK32 ||
			     formats[i].format == VK_FORMAT_A2R10G10B10_UNORM_PACK32 ||
			     formats[i].format == VK_FORMAT_A8B8G8R8_UNORM_PACK32))
			{
				format = formats[i];
				break;
			}
		}
	}

	return format;
}

struct SurfaceInfo
{
	VkPhysicalDeviceSurfaceInfo2KHR surface_info;
	VkSurfacePresentModeKHR present_mode;
	VkSurfaceCapabilitiesKHR surface_capabilities;
	std::vector<VkSurfaceFormatKHR> formats;
	VkSwapchainPresentModesCreateInfoKHR present_modes_info;
	VkImageCompressionControlEXT compression_control;
	VkImageCompressionFixedRateFlagsEXT compression_control_fixed_rates;
	VkSurfaceCapabilitiesPresentId2KHR present_id2;
	VkSurfaceCapabilitiesPresentWait2KHR present_wait2;
	VkPresentTimingSurfaceCapabilitiesEXT present_timing;
	std::vector<VkPresentModeKHR> present_mode_compat_group;
	const void *swapchain_pnext;
	VkSwapchainLatencyCreateInfoNV latency_create_info;
#ifdef _WIN32
	VkSurfaceFullScreenExclusiveInfoEXT exclusive_info;
	VkSurfaceFullScreenExclusiveWin32InfoEXT exclusive_info_win32;
#endif
};

static bool init_surface_info(Device &device, WSIPlatform &platform,
	VkSurfaceKHR surface, BackbufferFormat format,
	const WSI::ImageCompression &compression,
	PresentMode present_mode, SurfaceInfo &info, bool low_latency_mode_enable)
{
	if (surface == VK_NULL_HANDLE)
	{
		LOGE("Cannot create swapchain with surface == VK_NULL_HANDLE.\n");
		return false;
	}

	info.surface_info = { VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SURFACE_INFO_2_KHR };
	info.surface_info.surface = surface;
	info.swapchain_pnext = nullptr;

	auto &ext = device.get_device_features();

#ifdef _WIN32
	if (ext.supports_full_screen_exclusive)
	{
		info.exclusive_info = { VK_STRUCTURE_TYPE_SURFACE_FULL_SCREEN_EXCLUSIVE_INFO_EXT };
		auto monitor = reinterpret_cast<HMONITOR>(platform.get_fullscreen_monitor());
		info.swapchain_pnext = &info.exclusive_info;
		info.surface_info.pNext = &info.exclusive_info;

		if (monitor != nullptr)
		{
			info.exclusive_info_win32 = { VK_STRUCTURE_TYPE_SURFACE_FULL_SCREEN_EXCLUSIVE_WIN32_INFO_EXT };
			info.exclusive_info.pNext = &info.exclusive_info_win32;
			info.exclusive_info_win32.hmonitor = monitor;
			LOGI("Win32: Got a full-screen monitor.\n");
		}
		else
			LOGI("Win32: Not running full-screen.\n");

		bool prefer_exclusive = Util::get_environment_bool("GRANITE_EXCLUSIVE_FULL_SCREEN", false) || low_latency_mode_enable;
		if (ext.driver_id == VK_DRIVER_ID_INTEL_PROPRIETARY_WINDOWS)
			prefer_exclusive = false; // Broken on Intel Windows

		if (ext.driver_id == VK_DRIVER_ID_AMD_PROPRIETARY &&
		    (format == BackbufferFormat::HDR10 || format == BackbufferFormat::scRGB))
		{
			LOGI("Win32: HDR requested on AMD Windows. Forcing exclusive fullscreen, or HDR will not work properly.\n");
			prefer_exclusive = true;
		}

		if (prefer_exclusive && monitor != nullptr)
		{
			LOGI("Win32: Opting in to exclusive full-screen!\n");
			info.exclusive_info.fullScreenExclusive = VK_FULL_SCREEN_EXCLUSIVE_ALLOWED_EXT;

			// Try to promote this to application controlled exclusive.
			VkSurfaceCapabilities2KHR surface_capabilities2 = { VK_STRUCTURE_TYPE_SURFACE_CAPABILITIES_2_KHR };
			VkSurfaceCapabilitiesFullScreenExclusiveEXT capability_full_screen_exclusive = {
				VK_STRUCTURE_TYPE_SURFACE_CAPABILITIES_FULL_SCREEN_EXCLUSIVE_EXT
			};
			surface_capabilities2.pNext = &capability_full_screen_exclusive;

			if (vkGetPhysicalDeviceSurfaceCapabilities2KHR(device.get_physical_device(), &info.surface_info,
			                                               &surface_capabilities2) != VK_SUCCESS)
				return false;

			if (capability_full_screen_exclusive.fullScreenExclusiveSupported)
			{
				LOGI("Win32: Opting for exclusive fullscreen access.\n");
				info.exclusive_info.fullScreenExclusive = VK_FULL_SCREEN_EXCLUSIVE_APPLICATION_CONTROLLED_EXT;
			}
		}
		else
		{
			LOGI("Win32: Opting out of exclusive full-screen!\n");
			info.exclusive_info.fullScreenExclusive =
				prefer_exclusive ? VK_FULL_SCREEN_EXCLUSIVE_ALLOWED_EXT : VK_FULL_SCREEN_EXCLUSIVE_DISALLOWED_EXT;
		}
	}
#else
	(void)platform;
	(void)format;
#endif

	std::vector<VkPresentModeKHR> present_modes;
	uint32_t num_present_modes = 0;
	auto gpu = device.get_physical_device();

#ifdef _WIN32
	if (ext.supports_surface_capabilities2 && ext.supports_full_screen_exclusive)
	{
		if (vkGetPhysicalDeviceSurfacePresentModes2EXT(gpu, &info.surface_info, &num_present_modes, nullptr) !=
		    VK_SUCCESS)
		{
			return false;
		}
		present_modes.resize(num_present_modes);
		if (vkGetPhysicalDeviceSurfacePresentModes2EXT(gpu, &info.surface_info, &num_present_modes,
		                                               present_modes.data()) != VK_SUCCESS)
		{
			return false;
		}
	}
	else
#endif
	{
		if (vkGetPhysicalDeviceSurfacePresentModesKHR(gpu, surface, &num_present_modes, nullptr) != VK_SUCCESS)
			return false;
		present_modes.resize(num_present_modes);
		if (vkGetPhysicalDeviceSurfacePresentModesKHR(gpu, surface, &num_present_modes, present_modes.data()) != VK_SUCCESS)
			return false;
	}

	auto swapchain_present_mode = VK_PRESENT_MODE_FIFO_KHR;
	bool use_vsync = present_mode == PresentMode::SyncToVBlank;
	if (!use_vsync)
	{
		bool allow_mailbox = present_mode != PresentMode::UnlockedForceTearing;
		bool allow_immediate = present_mode != PresentMode::UnlockedNoTearing;

#ifdef _WIN32
		// If we're trying to go exclusive full-screen,
		// we need to ban certain types of present modes which apparently do not work as we expect.
		if (info.exclusive_info.fullScreenExclusive == VK_FULL_SCREEN_EXCLUSIVE_APPLICATION_CONTROLLED_EXT)
			allow_mailbox = false;
#endif

		for (auto &mode : present_modes)
		{
			if ((allow_immediate && mode == VK_PRESENT_MODE_IMMEDIATE_KHR) ||
			    (allow_mailbox && mode == VK_PRESENT_MODE_MAILBOX_KHR))
			{
				swapchain_present_mode = mode;
				break;
			}
		}
	}

	if (swapchain_present_mode == VK_PRESENT_MODE_FIFO_KHR && low_latency_mode_enable)
		for (auto mode : present_modes)
			if (mode == VK_PRESENT_MODE_FIFO_RELAXED_KHR)
				swapchain_present_mode = mode;

	LOGI("Using present mode: %u.\n", swapchain_present_mode);

	// First, query minImageCount without any present mode.
	// Avoid opting for present mode compat that is pathological in nature,
	// e.g. Xorg MAILBOX where minImageCount shoots up to 5 for stupid reasons.
	if (ext.supports_surface_capabilities2)
	{
		VkSurfaceCapabilities2KHR surface_capabilities2 = { VK_STRUCTURE_TYPE_SURFACE_CAPABILITIES_2_KHR };

		if (ext.present_id2_features.presentId2 && ext.present_wait2_features.presentWait2)
		{
			info.present_wait2 = { VK_STRUCTURE_TYPE_SURFACE_CAPABILITIES_PRESENT_WAIT_2_KHR };
			info.present_id2 = { VK_STRUCTURE_TYPE_SURFACE_CAPABILITIES_PRESENT_ID_2_KHR };
			info.present_wait2.pNext = &info.present_id2;
			surface_capabilities2.pNext = &info.present_wait2;
		}

		if (ext.present_timing_features.presentTiming)
		{
			info.present_timing = { VK_STRUCTURE_TYPE_PRESENT_TIMING_SURFACE_CAPABILITIES_EXT };
			info.present_timing.pNext = surface_capabilities2.pNext;
			surface_capabilities2.pNext = &info.present_timing;
		}

		if (vkGetPhysicalDeviceSurfaceCapabilities2KHR(gpu, &info.surface_info, &surface_capabilities2) != VK_SUCCESS)
			return false;
		info.surface_capabilities = surface_capabilities2.surfaceCapabilities;
	}
	else
	{
		if (vkGetPhysicalDeviceSurfaceCapabilitiesKHR(gpu, surface, &info.surface_capabilities) != VK_SUCCESS)
			return false;
	}

	// Make sure we query surface caps tied to the present mode for correct results.
	if (ext.swapchain_maintenance1_features.swapchainMaintenance1 &&
	    ext.supports_surface_capabilities2)
	{
		VkSurfaceCapabilities2KHR surface_capabilities2 = { VK_STRUCTURE_TYPE_SURFACE_CAPABILITIES_2_KHR };
		VkSurfacePresentModeCompatibilityKHR present_mode_caps =
		    { VK_STRUCTURE_TYPE_SURFACE_PRESENT_MODE_COMPATIBILITY_KHR };
		std::vector<VkPresentModeKHR> present_mode_compat_group;

		present_mode_compat_group.resize(32);
		present_mode_caps.presentModeCount = present_mode_compat_group.size();
		present_mode_caps.pPresentModes = present_mode_compat_group.data();

		info.present_mode.pNext = const_cast<void *>(info.surface_info.pNext);
		info.surface_info.pNext = &info.present_mode;
		info.present_mode = { VK_STRUCTURE_TYPE_SURFACE_PRESENT_MODE_KHR };
		info.present_mode.presentMode = swapchain_present_mode;

		surface_capabilities2.pNext = &present_mode_caps;
		if (vkGetPhysicalDeviceSurfaceCapabilities2KHR(gpu, &info.surface_info, &surface_capabilities2) != VK_SUCCESS)
			return false;
		surface_capabilities2.pNext = present_mode_caps.pNext;

		info.surface_capabilities.minImageCount = surface_capabilities2.surfaceCapabilities.minImageCount;
		present_mode_compat_group.resize(present_mode_caps.presentModeCount);
		info.present_mode_compat_group.reserve(present_mode_caps.presentModeCount);
		info.present_mode_compat_group.push_back(swapchain_present_mode);

		for (auto mode : present_mode_compat_group)
		{
			if (mode == swapchain_present_mode)
				continue;

			// Only allow sensible present modes that we know of.
			if (mode != VK_PRESENT_MODE_FIFO_KHR &&
			    mode != VK_PRESENT_MODE_FIFO_RELAXED_KHR &&
			    mode != VK_PRESENT_MODE_IMMEDIATE_KHR &&
			    mode != VK_PRESENT_MODE_MAILBOX_KHR)
			{
				continue;
			}

			info.present_mode.presentMode = mode;
			if (vkGetPhysicalDeviceSurfaceCapabilities2KHR(gpu, &info.surface_info, &surface_capabilities2) != VK_SUCCESS)
				return false;

			// Accept the present mode if it does not modify minImageCount.
			// If image count changes, we should probably recreate the swapchain.
			// If we have present wait we're at no risk of adding more latency, so just go ahead.
			if (surface_capabilities2.surfaceCapabilities.minImageCount == info.surface_capabilities.minImageCount ||
			    device.get_device_features().present_wait_features.presentWait ||
			    info.present_wait2.presentWait2Supported)
			{
				info.present_mode_compat_group.push_back(mode);
				info.surface_capabilities.minImageCount =
						std::max<uint32_t>(info.surface_capabilities.minImageCount,
						                   surface_capabilities2.surfaceCapabilities.minImageCount);
			}
		}
	}

	uint32_t format_count = 0;
	if (ext.supports_surface_capabilities2)
	{
		if (vkGetPhysicalDeviceSurfaceFormats2KHR(device.get_physical_device(),
		                                          &info.surface_info, &format_count,
		                                          nullptr) != VK_SUCCESS)
		{
			return false;
		}

		std::vector<VkSurfaceFormat2KHR> formats2(format_count);

		for (auto &f : formats2)
		{
			f = {};
			f.sType = VK_STRUCTURE_TYPE_SURFACE_FORMAT_2_KHR;
		}

		if (vkGetPhysicalDeviceSurfaceFormats2KHR(gpu, &info.surface_info, &format_count, formats2.data()) != VK_SUCCESS)
			return false;

		info.formats.reserve(format_count);
		for (auto &f : formats2)
			info.formats.push_back(f.surfaceFormat);
	}
	else
	{
		if (vkGetPhysicalDeviceSurfaceFormatsKHR(gpu, surface, &format_count, nullptr) != VK_SUCCESS)
			return false;
		info.formats.resize(format_count);
		if (vkGetPhysicalDeviceSurfaceFormatsKHR(gpu, surface, &format_count, info.formats.data()) != VK_SUCCESS)
			return false;
	}

	// Ensure that 10-bit formats come before other formats.
	std::sort(info.formats.begin(), info.formats.end(), [](const VkSurfaceFormatKHR &a, const VkSurfaceFormatKHR &b) {
		const auto qual = [](VkFormat fmt) {
			// Prefer a consistent ordering so Fossilize caches are more effective.
			if (fmt == VK_FORMAT_A2B10G10R10_UNORM_PACK32)
				return 3;
			else if (fmt == VK_FORMAT_A2R10G10B10_UNORM_PACK32)
				return 2;
			else if (fmt == VK_FORMAT_B8G8R8A8_UNORM)
				return 1;
			else
				return 0;
		};
		return qual(a.format) > qual(b.format);
	});

	// Allow for seamless toggle between presentation modes.
	if (ext.swapchain_maintenance1_features.swapchainMaintenance1)
	{
		info.present_modes_info = { VK_STRUCTURE_TYPE_SWAPCHAIN_PRESENT_MODES_CREATE_INFO_KHR };
		info.present_modes_info.pNext = const_cast<void *>(info.swapchain_pnext);
		info.present_modes_info.presentModeCount = info.present_mode_compat_group.size();
		info.present_modes_info.pPresentModes = info.present_mode_compat_group.data();
		info.swapchain_pnext = &info.present_modes_info;
	}

	info.present_mode.presentMode = swapchain_present_mode;

	if (ext.image_compression_control_swapchain_features.imageCompressionControlSwapchain &&
	    compression.type != VK_IMAGE_COMPRESSION_DEFAULT_EXT)
	{
		// There is no VU that we cannot just pass in whatever we want here,
		// but we might not be honored if we pass in something unsupported.
		// That's fine for now.
		info.compression_control = { VK_STRUCTURE_TYPE_IMAGE_COMPRESSION_CONTROL_EXT };
		info.compression_control.pNext = info.swapchain_pnext;
		info.compression_control.flags = compression.type;
		if (compression.type == VK_IMAGE_COMPRESSION_FIXED_RATE_EXPLICIT_EXT)
		{
			info.compression_control_fixed_rates = compression.fixed_rates;
			info.compression_control.pFixedRateFlags = &info.compression_control_fixed_rates;
			info.compression_control.compressionControlPlaneCount = 1;
			LOGI("Using fixed-rate compression for swapchain (flags #%08x).\n", compression.fixed_rates);
		}
		else if (compression.type == VK_IMAGE_COMPRESSION_FIXED_RATE_DEFAULT_EXT)
			LOGI("Using default fixed-rate compression for swapchain.\n");
		else if (compression.type == VK_IMAGE_COMPRESSION_DISABLED_EXT)
			LOGI("Disabling compression for swapchain.\n");

		info.swapchain_pnext = &info.compression_control;
	}

	if (ext.supports_low_latency2_nv)
	{
		info.latency_create_info = { VK_STRUCTURE_TYPE_SWAPCHAIN_LATENCY_CREATE_INFO_NV };
		info.latency_create_info.latencyModeEnable = VK_TRUE;
		info.latency_create_info.pNext = info.swapchain_pnext;
		info.swapchain_pnext = &info.latency_create_info;
	}

	return true;
}

WSI::SwapchainError WSI::init_swapchain(unsigned width, unsigned height)
{
	SurfaceInfo surface_info = {};
	if (!init_surface_info(*device, *platform, surface, current_backbuffer_format, current_compression,
	                       current_present_mode, surface_info, low_latency_mode_enable_present))
	{
		return SwapchainError::Error;
	}
	const auto &caps = surface_info.surface_capabilities;

	// Happens on Windows when you minimize a window.
	if (caps.maxImageExtent.width == 0 && caps.maxImageExtent.height == 0)
		return SwapchainError::NoSurface;

	if (current_extra_usage && support_prerotate)
	{
		LOGW("Disabling prerotate support due to extra usage flags in swapchain.\n");
		support_prerotate = false;
	}

	if (current_extra_usage & ~caps.supportedUsageFlags)
	{
		LOGW("Attempting to use unsupported usage flags 0x%x for swapchain.\n", current_extra_usage);
		current_extra_usage &= caps.supportedUsageFlags;
		extra_usage = current_extra_usage;
	}

	auto attempt_backbuffer_format = current_backbuffer_format;
	auto surface_format = find_suitable_present_format(surface_info.formats, attempt_backbuffer_format);

	if (surface_format.format == VK_FORMAT_UNDEFINED &&
	    (attempt_backbuffer_format == BackbufferFormat::HDR10 ||
	     attempt_backbuffer_format == BackbufferFormat::scRGB ||
	     attempt_backbuffer_format == BackbufferFormat::DisplayP3 ||
	     attempt_backbuffer_format == BackbufferFormat::UNORMPassthrough ||
	     attempt_backbuffer_format == BackbufferFormat::Custom))
	{
		LOGW("Could not find suitable present format for HDR. Attempting fallback to UNORM.\n");
		attempt_backbuffer_format = BackbufferFormat::UNORM;
		surface_format = find_suitable_present_format(surface_info.formats, attempt_backbuffer_format);
	}

	if (surface_format.format == VK_FORMAT_UNDEFINED)
	{
		LOGW("Could not find supported format for swapchain usage flags 0x%x.\n", current_extra_usage);
		current_extra_usage = 0;
		extra_usage = 0;
		surface_format = find_suitable_present_format(surface_info.formats, attempt_backbuffer_format);
	}

	if (surface_format.format == VK_FORMAT_UNDEFINED)
	{
		LOGE("Failed to find any suitable format for swapchain.\n");
		return SwapchainError::Error;
	}

	static const char *transform_names[] = {
		"IDENTITY_BIT_KHR",
		"ROTATE_90_BIT_KHR",
		"ROTATE_180_BIT_KHR",
		"ROTATE_270_BIT_KHR",
		"HORIZONTAL_MIRROR_BIT_KHR",
		"HORIZONTAL_MIRROR_ROTATE_90_BIT_KHR",
		"HORIZONTAL_MIRROR_ROTATE_180_BIT_KHR",
		"HORIZONTAL_MIRROR_ROTATE_270_BIT_KHR",
		"INHERIT_BIT_KHR",
	};

	LOGI("Current transform is enum 0x%x.\n", unsigned(caps.currentTransform));

	for (unsigned i = 0; i <= 8; i++)
	{
		if (caps.supportedTransforms & (1u << i))
			LOGI("Supported transform 0x%x: %s.\n", 1u << i, transform_names[i]);
	}

	VkSurfaceTransformFlagBitsKHR pre_transform;
	if (!support_prerotate && (caps.supportedTransforms & VK_SURFACE_TRANSFORM_IDENTITY_BIT_KHR) != 0)
		pre_transform = VK_SURFACE_TRANSFORM_IDENTITY_BIT_KHR;
	else
	{
		// Only attempt to use prerotate if we can deal with it purely using a XY clip fixup.
		// For horizontal flip we need to start flipping front-face as well ...
		if ((caps.currentTransform & (
				VK_SURFACE_TRANSFORM_ROTATE_90_BIT_KHR |
				VK_SURFACE_TRANSFORM_ROTATE_180_BIT_KHR |
				VK_SURFACE_TRANSFORM_ROTATE_270_BIT_KHR)) != 0)
			pre_transform = caps.currentTransform;
		else
			pre_transform = VK_SURFACE_TRANSFORM_IDENTITY_BIT_KHR;
	}

	if (pre_transform != caps.currentTransform)
	{
		LOGW("surfaceTransform (0x%x) != currentTransform (0x%u). Might get performance penalty.\n",
		     unsigned(pre_transform), unsigned(caps.currentTransform));
	}

	swapchain_current_prerotate = pre_transform;

	VkExtent2D swapchain_size;
	LOGI("Swapchain current extent: %d x %d\n",
	     int(caps.currentExtent.width),
	     int(caps.currentExtent.height));

	if (width == 0)
	{
		if (caps.currentExtent.width != ~0u)
			width = caps.currentExtent.width;
		else
			width = 1280;
		LOGI("Auto selected width = %u.\n", width);
	}

	if (height == 0)
	{
		if (caps.currentExtent.height != ~0u)
			height = caps.currentExtent.height;
		else
			height = 720;
		LOGI("Auto selected height = %u.\n", height);
	}

	// Try to match the swapchain size up with what we expect w.r.t. aspect ratio.
	float target_aspect_ratio = float(width) / float(height);
	if ((swapchain_aspect_ratio > 1.0f && target_aspect_ratio < 1.0f) ||
	    (swapchain_aspect_ratio < 1.0f && target_aspect_ratio > 1.0f))
	{
		std::swap(width, height);
	}

	// If we are using pre-rotate of 90 or 270 degrees, we need to flip width and height again.
	if (swapchain_current_prerotate &
	    (VK_SURFACE_TRANSFORM_ROTATE_90_BIT_KHR | VK_SURFACE_TRANSFORM_ROTATE_270_BIT_KHR |
	     VK_SURFACE_TRANSFORM_HORIZONTAL_MIRROR_ROTATE_90_BIT_KHR |
	     VK_SURFACE_TRANSFORM_HORIZONTAL_MIRROR_ROTATE_270_BIT_KHR))
	{
		std::swap(width, height);
	}

	// Clamp the target width, height to boundaries.
	swapchain_size.width =
	    std::max(std::min(width, caps.maxImageExtent.width), caps.minImageExtent.width);
	swapchain_size.height =
	    std::max(std::min(height, caps.maxImageExtent.height), caps.minImageExtent.height);

	uint32_t desired_swapchain_images =
		low_latency_mode_enable_present && current_present_mode == PresentMode::SyncToVBlank ? 2 : 3;

	supports_present_wait2 =
	    surface_info.present_id2.presentId2Supported && surface_info.present_wait2.presentWait2Supported;

	// Need a deeper swapchain to avoid potential stalls when duping frames.
	// We only do this when present wait is supported, so latency should not be compromised.
	if (current_frame_dupe_aware && (device->get_device_features().present_wait_features.presentWait || supports_present_wait2))
		desired_swapchain_images = current_frame_dupe_target_images;

	desired_swapchain_images = Util::get_environment_uint("GRANITE_VULKAN_SWAPCHAIN_IMAGES", desired_swapchain_images);
	LOGI("Targeting %u swapchain images.\n", desired_swapchain_images);

	if (desired_swapchain_images < caps.minImageCount)
		desired_swapchain_images = caps.minImageCount;

	if ((caps.maxImageCount > 0) && (desired_swapchain_images > caps.maxImageCount))
		desired_swapchain_images = caps.maxImageCount;

	VkCompositeAlphaFlagBitsKHR composite_mode = VK_COMPOSITE_ALPHA_OPAQUE_BIT_KHR;
	if (caps.supportedCompositeAlpha & VK_COMPOSITE_ALPHA_OPAQUE_BIT_KHR)
		composite_mode = VK_COMPOSITE_ALPHA_OPAQUE_BIT_KHR;
	else if (caps.supportedCompositeAlpha & VK_COMPOSITE_ALPHA_INHERIT_BIT_KHR)
		composite_mode = VK_COMPOSITE_ALPHA_INHERIT_BIT_KHR;
	else if (caps.supportedCompositeAlpha & VK_COMPOSITE_ALPHA_POST_MULTIPLIED_BIT_KHR)
		composite_mode = VK_COMPOSITE_ALPHA_POST_MULTIPLIED_BIT_KHR;
	else if (caps.supportedCompositeAlpha & VK_COMPOSITE_ALPHA_PRE_MULTIPLIED_BIT_KHR)
		composite_mode = VK_COMPOSITE_ALPHA_PRE_MULTIPLIED_BIT_KHR;
	else
		LOGW("No sensible composite mode supported?\n");

	VkSwapchainKHR old_swapchain = swapchain;

	VkSwapchainCreateInfoKHR info = { VK_STRUCTURE_TYPE_SWAPCHAIN_CREATE_INFO_KHR };
	info.surface = surface;
	info.pNext = surface_info.swapchain_pnext;
	info.minImageCount = desired_swapchain_images;
	info.imageFormat = surface_format.format;
	info.imageColorSpace = surface_format.colorSpace;
	info.imageExtent.width = swapchain_size.width;
	info.imageExtent.height = swapchain_size.height;
	info.imageArrayLayers = 1;
	info.imageUsage = VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT | current_extra_usage;
	info.imageSharingMode = VK_SHARING_MODE_EXCLUSIVE;
	info.preTransform = pre_transform;
	info.compositeAlpha = composite_mode;
	info.presentMode = surface_info.present_mode.presentMode;
	info.clipped = VK_TRUE;
	info.oldSwapchain = old_swapchain;

	// Defer the deletion instead.
	if (device->get_device_features().swapchain_maintenance1_features.swapchainMaintenance1 &&
	    old_swapchain != VK_NULL_HANDLE && !device->get_workarounds().broken_present_fence)
	{
		deferred_swapchains.push_back({ old_swapchain, last_present_fence });
		old_swapchain = VK_NULL_HANDLE;
	}

	platform->event_swapchain_destroyed();

	supports_present_timing.feedback = surface_info.present_timing.presentTimingSupported ?
			surface_info.present_timing.presentStageQueries : 0;

	// Only ackknowledge present stages we understand in case there are future extensions.
	supports_present_timing.feedback &=
			VK_PRESENT_STAGE_QUEUE_OPERATIONS_END_BIT_EXT |
			VK_PRESENT_STAGE_REQUEST_DEQUEUED_BIT_EXT |
			VK_PRESENT_STAGE_IMAGE_FIRST_PIXEL_OUT_BIT_EXT |
			VK_PRESENT_STAGE_IMAGE_FIRST_PIXEL_VISIBLE_BIT_EXT;

	supports_present_timing.absolute = surface_info.present_timing.presentTimingSupported &&
	                                   surface_info.present_timing.presentAtAbsoluteTimeSupported;

	supports_present_timing.relative = surface_info.present_timing.presentTimingSupported &&
	                                   surface_info.present_timing.presentAtRelativeTimeSupported;

	if (supports_present_wait2)
		info.flags |= VK_SWAPCHAIN_CREATE_PRESENT_ID_2_BIT_KHR | VK_SWAPCHAIN_CREATE_PRESENT_WAIT_2_BIT_KHR;
	else
		supports_present_timing.feedback = 0;

	if (supports_present_timing.feedback)
	{
		info.flags |= VK_SWAPCHAIN_CREATE_PRESENT_TIMING_BIT_EXT;
		LOGI("Enabling present timing, queries #%x, absolute %u, relative %u.\n",
		     supports_present_timing.feedback,
		     supports_present_timing.absolute,
		     supports_present_timing.relative);
	}

	auto res = table->vkCreateSwapchainKHR(context->get_device(), &info, nullptr, &swapchain);
	if (res < 0)
		swapchain = VK_NULL_HANDLE;

	if (res == VK_SUCCESS && supports_present_timing.feedback)
	{
		table->vkSetSwapchainPresentTimingQueueSizeEXT(context->get_device(), swapchain, PresentTimingQueueSize);
		present_timing = {};
		update_present_timing_properties();
		update_time_domain_properties();
		recalibrate_present_timing_domains();
	}

	platform->destroy_swapchain_resources(old_swapchain);
	table->vkDestroySwapchainKHR(context->get_device(), old_swapchain, nullptr);
	has_acquired_swapchain_index = false;
	next_present_id = 1;
	present_last_id = 0;
	device->set_present_id(VK_NULL_HANDLE, 0);

	if (res == VK_SUCCESS && device->get_device_features().supports_low_latency2_nv)
	{
		VkLatencySleepModeInfoNV sleep_mode_info = { VK_STRUCTURE_TYPE_LATENCY_SLEEP_MODE_INFO_NV };
		sleep_mode_info.lowLatencyBoost = low_latency_mode_enable_gpu_submit;
		sleep_mode_info.lowLatencyMode = low_latency_mode_enable_gpu_submit;
		if (table->vkSetLatencySleepModeNV(context->get_device(), swapchain, &sleep_mode_info) != VK_SUCCESS)
			LOGE("Failed to set low latency sleep mode.\n");
	}

	active_present_mode = info.presentMode;
	present_mode_compat_group = std::move(surface_info.present_mode_compat_group);

#ifdef _WIN32
	if (surface_info.exclusive_info.fullScreenExclusive == VK_FULL_SCREEN_EXCLUSIVE_APPLICATION_CONTROLLED_EXT)
	{
		bool success = vkAcquireFullScreenExclusiveModeEXT(context->get_device(), swapchain) == VK_SUCCESS;
		if (success)
			LOGI("Successfully acquired exclusive full-screen.\n");
		else
			LOGI("Failed to acquire exclusive full-screen. Using borderless windowed.\n");
	}
#endif

	if (res != VK_SUCCESS)
	{
		LOGE("Failed to create swapchain (code: %d)\n", int(res));
		swapchain = VK_NULL_HANDLE;
		return SwapchainError::Error;
	}

	swapchain_width = swapchain_size.width;
	swapchain_height = swapchain_size.height;
	swapchain_surface_format = surface_format;
	swapchain_is_suboptimal = false;

	LOGI("Created swapchain %u x %u (fmt: %u, transform: %u).\n", swapchain_width, swapchain_height,
	     unsigned(swapchain_surface_format.format), unsigned(swapchain_current_prerotate));

	uint32_t image_count;
	if (table->vkGetSwapchainImagesKHR(context->get_device(), swapchain, &image_count, nullptr) != VK_SUCCESS)
		return SwapchainError::Error;
	swapchain_images.resize(image_count);
	release_semaphores.resize(image_count);
	if (table->vkGetSwapchainImagesKHR(context->get_device(), swapchain, &image_count, swapchain_images.data()) != VK_SUCCESS)
		return SwapchainError::Error;

	LOGI("Got %u swapchain images.\n", image_count);

	platform->event_swapchain_created(device.get(), swapchain, swapchain_width, swapchain_height,
	                                  swapchain_aspect_ratio, image_count,
	                                  swapchain_surface_format.format,
	                                  swapchain_surface_format.colorSpace,
	                                  swapchain_current_prerotate);

	if (swapchain_surface_format.colorSpace == VK_COLOR_SPACE_HDR10_ST2084_EXT &&
	    valid_hdr_metadata && device->get_device_features().supports_hdr_metadata)
	{
		table->vkSetHdrMetadataEXT(device->get_device(), 1, &swapchain, &hdr_metadata);
	}

	return SwapchainError::None;
}

void WSI::set_support_prerotate(bool enable)
{
	support_prerotate = enable;
}

VkSurfaceTransformFlagBitsKHR WSI::get_current_prerotate() const
{
	return swapchain_current_prerotate;
}

CommandBuffer::Type WSI::get_current_present_queue_type() const
{
	return device->get_current_present_queue_type();
}

WSI::~WSI()
{
	teardown();
}

void WSIPlatform::event_device_created(Device *) {}
void WSIPlatform::event_device_destroyed() {}
void WSIPlatform::event_swapchain_created(Device *, VkSwapchainKHR, unsigned, unsigned, float, size_t,
                                          VkFormat, VkColorSpaceKHR,
                                          VkSurfaceTransformFlagBitsKHR) {}
void WSIPlatform::event_swapchain_destroyed() {}
void WSIPlatform::destroy_swapchain_resources(VkSwapchainKHR) {}
void WSIPlatform::event_frame_tick(double, double) {}
void WSIPlatform::event_swapchain_index(Device *, unsigned) {}
void WSIPlatform::begin_drop_event() {}
void WSIPlatform::begin_soft_keyboard(const std::string &) {}
void WSIPlatform::end_soft_keyboard() {}
void WSIPlatform::show_message_box(const std::string &, Vulkan::WSIPlatform::MessageType) {}
}