vk: Update async compute (fast) to use the new sync API

rpcs3/Emu/RSX/VK/VKAsyncScheduler.cpp

@@ -10,7 +10,7 @@
 
 namespace vk
 {
-	AsyncTaskScheduler::AsyncTaskScheduler([[maybe_unused]] vk_gpu_scheduler_mode mode)
+	AsyncTaskScheduler::AsyncTaskScheduler(vk_gpu_scheduler_mode mode, const VkDependencyInfoKHR& queue_dependency)
 	{
 		if (g_cfg.video.renderer != video_renderer::vulkan || !g_cfg.video.vk.asynchronous_texture_streaming)
 		{
@@ -20,7 +20,7 @@ namespace vk
 			return;
 		}
 
-		init_config_options(mode);
+		init_config_options(mode, queue_dependency);
 	}
 
 	AsyncTaskScheduler::~AsyncTaskScheduler()
@@ -32,7 +32,7 @@ namespace vk
 		}
 	}
 
-	void AsyncTaskScheduler::init_config_options(vk_gpu_scheduler_mode mode)
+	void AsyncTaskScheduler::init_config_options(vk_gpu_scheduler_mode mode, const VkDependencyInfoKHR& queue_dependency)
 	{
 		std::lock_guard lock(m_config_mutex);
 		if (std::exchange(m_options_initialized, true))
@@ -43,6 +43,8 @@ namespace vk
 
 		m_use_host_scheduler = (mode == vk_gpu_scheduler_mode::safe) || g_cfg.video.strict_rendering_mode;
 		rsx_log.notice("Asynchronous task scheduler is active running in %s mode", m_use_host_scheduler? "'Safe'" : "'Fast'");
+
+		m_dependency_info = queue_dependency;
 	}
 
 	void AsyncTaskScheduler::delayed_init()
@@ -76,7 +78,7 @@ namespace vk
 		auto& sync_label = m_events_pool[m_next_event_id++ % events_pool_size];
 
 		sync_label->reset();
-		sync_label->signal(*m_current_cb, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, 0);
+		sync_label->signal(*m_current_cb, m_dependency_info);
 		m_sync_label = sync_label.get();
 	}
 

rpcs3/Emu/RSX/VK/VKAsyncScheduler.h

@@ -29,6 +29,7 @@ namespace vk
 		// Sync
 		event* m_sync_label = nullptr;
 		atomic_t<bool> m_sync_required = false;
+		VkDependencyInfoKHR m_dependency_info{};
 
 		static constexpr u32 events_pool_size = 16384;
 		std::vector<std::unique_ptr<vk::event>> m_events_pool;
@@ -39,12 +40,12 @@ namespace vk
 
 		shared_mutex m_submit_mutex;
 
-		void init_config_options(vk_gpu_scheduler_mode mode);
+		void init_config_options(vk_gpu_scheduler_mode mode, const VkDependencyInfoKHR& queue_dependency);
 		void delayed_init();
 		void insert_sync_event();
 
 	public:
-		AsyncTaskScheduler(vk_gpu_scheduler_mode mode); // This ctor stops default initialization by fxo
+		AsyncTaskScheduler(vk_gpu_scheduler_mode mode, const VkDependencyInfoKHR& queue_dependency);
 		~AsyncTaskScheduler();
 
 		command_buffer* get_current();

rpcs3/Emu/RSX/VK/VKDraw.cpp

@@ -478,7 +478,7 @@ void VKGSRender::load_texture_env()
 			// Sync any async scheduler tasks
 			if (auto ev = async_task_scheduler.get_primary_sync_label())
 			{
-				ev->gpu_wait(*m_current_command_buffer, { .sType = VK_STRUCTURE_TYPE_DEPENDENCY_INFO_KHR });
+				ev->gpu_wait(*m_current_command_buffer, m_async_compute_dependency_info);
 			}
 		}
 	}

rpcs3/Emu/RSX/VK/VKGSRender.cpp

@@ -822,8 +822,25 @@ VKGSRender::VKGSRender(utils::serial* ar) noexcept : GSRender(ar)
 
 	if (backend_config.supports_asynchronous_compute)
 	{
+		m_async_compute_memory_barrier =
+		{
+			.sType = VK_STRUCTURE_TYPE_MEMORY_BARRIER_2_KHR,
+			.pNext = nullptr,
+			.srcStageMask = VK_PIPELINE_STAGE_2_TRANSFER_BIT_KHR | VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT_KHR,
+			.srcAccessMask = VK_ACCESS_2_MEMORY_WRITE_BIT_KHR,
+			.dstStageMask = VK_PIPELINE_STAGE_2_FRAGMENT_SHADER_BIT_KHR | VK_PIPELINE_STAGE_2_VERTEX_SHADER_BIT_KHR,
+			.dstAccessMask = VK_ACCESS_2_SHADER_SAMPLED_READ_BIT_KHR
+		};
+
+		m_async_compute_dependency_info =
+		{
+			.sType = VK_STRUCTURE_TYPE_DEPENDENCY_INFO_KHR,
+			.memoryBarrierCount = 1,
+			.pMemoryBarriers = &m_async_compute_memory_barrier
+		};
+
 		// Run only if async compute can be used.
-		g_fxo->init<vk::AsyncTaskScheduler>(g_cfg.video.vk.asynchronous_scheduler);
+		g_fxo->init<vk::AsyncTaskScheduler>(g_cfg.video.vk.asynchronous_scheduler, m_async_compute_dependency_info);
 	}
 
 	if (backend_config.supports_host_gpu_labels)

rpcs3/Emu/RSX/VK/VKGSRender.h

@@ -91,6 +91,9 @@ private:
 	std::unique_ptr<vk::buffer_view> m_volatile_attribute_storage;
 	std::unique_ptr<vk::buffer_view> m_vertex_layout_storage;
 
+	VkDependencyInfoKHR m_async_compute_dependency_info{};
+	VkMemoryBarrier2KHR m_async_compute_memory_barrier{};
+
 public:
 	//vk::fbo draw_fbo;
 	std::unique_ptr<vk::vertex_cache> m_vertex_cache;