vk: Fix CPU frame misalignment bug

- This one has been around for a really long time.
- The frame-based structure was due to translating the original vulkan tutorial to working code.
- While it is not feasible to throw the arch away, we don't need a rigid 2-frame set for acquire-submit semaphores.

rpcs3/Emu/RSX/VK/VKGSRender.cpp

@@ -505,9 +505,6 @@ VKGSRender::VKGSRender(utils::serial* ar) noexcept : GSRender(ar)
 		m_occlusion_query_manager->set_control_flags(VK_QUERY_CONTROL_PRECISE_BIT, 0);
 	}
 
-	VkSemaphoreCreateInfo semaphore_info = {};
-	semaphore_info.sType = VK_STRUCTURE_TYPE_SEMAPHORE_CREATE_INFO;
-
 	// VRAM allocation
 	// This first set is bound persistently, so grow notifications are enabled.
 	m_attrib_ring_info.create(VK_BUFFER_USAGE_UNIFORM_TEXEL_BUFFER_BIT, VK_ATTRIB_RING_BUFFER_SIZE_M * 0x100000, vk::heap_pool_default, "attrib buffer", 0x400000, VK_TRUE);
@@ -570,10 +567,13 @@ VKGSRender::VKGSRender(utils::serial* ar) noexcept : GSRender(ar)
 		rsx_log.warning("Current driver may crash due to memory limitations (%uk)", m_texbuffer_view_size / 1024);
 	}
 
-	for (auto &ctx : frame_context_storage)
+	m_max_async_frames = m_swapchain->get_swap_image_count();
+	m_frame_context_storage.resize(m_max_async_frames);
+	m_current_frame = &m_frame_context_storage[0];
+
+	for (auto& ctx : m_frame_context_storage)
 	{
-		vkCreateSemaphore((*m_device), &semaphore_info, nullptr, &ctx.present_wait_semaphore);
+		ctx.init(*m_device);
-		vkCreateSemaphore((*m_device), &semaphore_info, nullptr, &ctx.acquire_signal_semaphore);
 	}
 
 	const auto& memory_map = m_device->get_memory_mapping();
@@ -612,8 +612,6 @@ VKGSRender::VKGSRender(utils::serial* ar) noexcept : GSRender(ar)
 
 	}
 
-	m_current_frame = &frame_context_storage[0];
-
 	m_texture_cache.initialize((*m_device), m_device->get_graphics_queue(),
 			m_texture_upload_buffer_ring_info);
 
@@ -830,16 +828,17 @@ VKGSRender::~VKGSRender()
 	if (m_current_frame == &m_aux_frame_context)
 	{
 		// Return resources back to the owner
-		m_current_frame = &frame_context_storage[m_current_queue_index];
+		m_current_frame = &m_frame_context_storage[m_current_queue_index];
 		m_current_frame->grab_resources(m_aux_frame_context);
 	}
 
-	// NOTE: aux_context uses descriptor pools borrowed from the main queues and any allocations will be automatically freed when pool is destroyed
+	// CPU frame contexts
-	for (auto &ctx : frame_context_storage)
+	for (auto &ctx : m_frame_context_storage)
 	{
-		vkDestroySemaphore((*m_device), ctx.present_wait_semaphore, nullptr);
+		ctx.destroy(*m_device);
-		vkDestroySemaphore((*m_device), ctx.acquire_signal_semaphore, nullptr);
 	}
+	m_current_frame = nullptr;
+	m_frame_context_storage.clear();
 
 	// Textures
 	m_rtts.destroy();

rpcs3/Emu/RSX/VK/VKGSRender.h

@@ -159,8 +159,9 @@ private:
 	u64 m_texture_parameters_dynamic_offset = 0;
 	u64 m_stipple_array_dynamic_offset = 0;
 
-	std::array<vk::frame_context_t, VK_MAX_ASYNC_FRAMES> frame_context_storage;
+	std::vector<vk::frame_context_t> m_frame_context_storage;
-	//Temp frame context to use if the real frame queue is overburdened. Only used for storage
+	u32 m_max_async_frames = 0u;
+	// Temp frame context to use if the real frame queue is overburdened. Only used for storage
 	vk::frame_context_t m_aux_frame_context;
 
 	u32 m_current_queue_index = 0;

rpcs3/Emu/RSX/VK/VKGSRenderTypes.hpp

@@ -23,7 +23,6 @@
 #define VK_INDEX_RING_BUFFER_SIZE_M 16
 
 #define VK_MAX_ASYNC_CB_COUNT 512
-#define VK_MAX_ASYNC_FRAMES 2
 
 #define FRAME_PRESENT_TIMEOUT 10000000ull // 10 seconds
 #define GENERAL_WAIT_TIMEOUT  2000000ull  // 2 seconds
@@ -186,6 +185,20 @@ namespace vk
 		data_heap_manager::managed_heap_snapshot_t heap_snapshot;
 		u64 last_frame_sync_time = 0;
 
+		void init(VkDevice dev)
+		{
+			VkSemaphoreCreateInfo semaphore_info = {};
+			semaphore_info.sType = VK_STRUCTURE_TYPE_SEMAPHORE_CREATE_INFO;
+			vkCreateSemaphore(dev, &semaphore_info, nullptr, &present_wait_semaphore);
+			vkCreateSemaphore(dev, &semaphore_info, nullptr, &acquire_signal_semaphore);
+		}
+
+		void destroy(VkDevice dev)
+		{
+			vkDestroySemaphore(dev, present_wait_semaphore, nullptr);
+			vkDestroySemaphore(dev, acquire_signal_semaphore, nullptr);
+		}
+
 		// Copy shareable information
 		void grab_resources(frame_context_t& other)
 		{

rpcs3/Emu/RSX/VK/VKPresent.cpp

@@ -52,7 +52,7 @@ void VKGSRender::reinitialize_swapchain()
 	m_current_command_buffer->reset();
 	m_current_command_buffer->begin();
 
-	for (auto &ctx : frame_context_storage)
+	for (auto &ctx : m_frame_context_storage)
 	{
 		if (ctx.present_image == umax)
 			continue;
@@ -67,6 +67,16 @@ void VKGSRender::reinitialize_swapchain()
 	// Drain all the queues
 	vkDeviceWaitIdle(*m_device);
 
+	// Reset frame context storage
+	for (auto& ctx : m_frame_context_storage)
+	{
+		ctx.destroy(*m_device);
+	}
+	m_current_frame = nullptr;
+	m_max_async_frames = 0;
+	m_current_queue_index = 0;
+	m_frame_context_storage.clear();
+
 	// Rebuild swapchain. Old swapchain destruction is handled by the init_swapchain call
 	if (!m_swapchain->init(m_swapchain_dims.width, m_swapchain_dims.height))
 	{
@@ -75,6 +85,16 @@ void VKGSRender::reinitialize_swapchain()
 		return;
 	}
 
+	// Re-initialize CPU frame contexts
+	m_max_async_frames = m_swapchain->get_swap_image_count();
+	m_frame_context_storage.resize(m_max_async_frames);
+	for (auto& ctx : m_frame_context_storage)
+	{
+		ctx.init(*m_device);
+	}
+	m_current_queue_index = 0;
+	m_current_frame = &m_frame_context_storage[0];
+
 	// Prepare new swapchain images for use
 	for (u32 i = 0; i < m_swapchain->get_swap_image_count(); ++i)
 	{
@@ -158,10 +178,10 @@ void VKGSRender::advance_queued_frames()
 	m_current_frame->tag_frame_end();
 
 	m_queued_frames.push_back(m_current_frame);
-	ensure(m_queued_frames.size() <= VK_MAX_ASYNC_FRAMES);
+	ensure(m_queued_frames.size() <= m_max_async_frames);
 
-	m_current_queue_index = (m_current_queue_index + 1) % VK_MAX_ASYNC_FRAMES;
+	m_current_queue_index = (m_current_queue_index + 1) % m_max_async_frames;
-	m_current_frame = &frame_context_storage[m_current_queue_index];
+	m_current_frame = &m_frame_context_storage[m_current_queue_index];
 	m_current_frame->flags |= frame_context_state::dirty;
 
 	vk::advance_frame_counter();
@@ -398,7 +418,7 @@ void VKGSRender::flip(const rsx::display_flip_info_t& info)
 
 	if (m_current_frame == &m_aux_frame_context)
 	{
-		m_current_frame = &frame_context_storage[m_current_queue_index];
+		m_current_frame = &m_frame_context_storage[m_current_queue_index];
 		if (m_current_frame->swap_command_buffer)
 		{
 			// Its possible this flip request is triggered by overlays and the flip queue is in undefined state
@@ -520,7 +540,7 @@ void VKGSRender::flip(const rsx::display_flip_info_t& info)
 	ensure(m_current_frame->present_image == umax);
 	ensure(m_current_frame->swap_command_buffer == nullptr);
 
-	u64 timeout = m_swapchain->get_swap_image_count() <= VK_MAX_ASYNC_FRAMES? 0ull: 100000000ull;
+	u64 timeout = m_swapchain->get_swap_image_count() <= 2? 0ull: 100000000ull;
 	while (VkResult status = m_swapchain->acquire_next_swapchain_image(m_current_frame->acquire_signal_semaphore, timeout, &m_current_frame->present_image))
 	{
 		switch (status)