vk: Implement state tracking for command buffers

- Only tracks the heavy items for now. Binding the same pipeline or descriptor set causes some drivers to perform expensive operations (RADV, AMDVLK)

rpcs3/Emu/RSX/VK/VKGSRenderTypes.hpp

@@ -74,7 +74,7 @@ namespace vk
 			}
 
 			++reset_id;
-			CHECK_RESULT(vkResetCommandBuffer(commands, 0));
+			vk::command_buffer::reset();
 		}
 
 		bool poke()

rpcs3/Emu/RSX/VK/VKProgramPipeline.cpp

@@ -414,8 +414,8 @@ namespace vk
 				bind_sets[count++] = set.commit();   // Commit variable changes and return handle to the new set
 			}
 
-			vkCmdBindPipeline(cmd, bind_point, m_pipeline);
+			cmd.bind_pipeline(m_pipeline, bind_point);
-			vkCmdBindDescriptorSets(cmd, bind_point, m_pipeline_layout, 0, count, bind_sets, 0, nullptr);
+			cmd.bind_descriptor_sets({ bind_sets, count }, bind_point, m_pipeline_layout);
 			return *this;
 		}
 

rpcs3/Emu/RSX/VK/VKTextureCache.cpp

@@ -1339,7 +1339,7 @@ namespace vk
 			cmd.submit(submit_info, VK_TRUE);
 			vk::wait_for_fence(&submit_fence, GENERAL_WAIT_TIMEOUT);
 
-			CHECK_RESULT(vkResetCommandBuffer(cmd, 0));
+			cmd.reset();
 			cmd.begin();
 		}
 		else

rpcs3/Emu/RSX/VK/vkutils/commands.cpp

@@ -70,6 +70,16 @@ namespace vk
 		}
 	}
 
+	void command_buffer::reset()
+	{
+		// Nuke the state cache
+		m_bound_pipeline = VK_NULL_HANDLE;
+		m_bound_descriptor_sets[0] = VK_NULL_HANDLE;
+
+		// Do the driver reset
+		CHECK_RESULT(vkResetCommandBuffer(commands, 0));
+	}
+
 	void command_buffer::begin()
 	{
 		if (m_submit_fence && is_pending)
@@ -79,7 +89,7 @@ namespace vk
 
 			//CHECK_RESULT(vkResetFences(pool->get_owner(), 1, &m_submit_fence));
 			m_submit_fence->reset();
-			CHECK_RESULT(vkResetCommandBuffer(commands, 0));
+			reset();
 		}
 
 		if (is_open)
@@ -129,4 +139,51 @@ namespace vk
 		queue_submit(submit_info, flush);
 		clear_flags();
 	}
+
+	void command_buffer::bind_pipeline(VkPipeline pipeline, VkPipelineBindPoint bind_point) const
+	{
+		if (m_bound_pipeline == pipeline)
+		{
+			return;
+		}
+
+		m_bound_pipeline = pipeline;
+		vkCmdBindPipeline(commands, bind_point, pipeline);
+	}
+
+	void command_buffer::bind_descriptor_sets(
+		const std::span<VkDescriptorSet>& sets,
+		const std::span<u32>& dynamic_offsets,
+		VkPipelineBindPoint bind_point,
+		VkPipelineLayout pipe_layout) const
+	{
+		const u32 num_sets = ::size32(sets);
+		ensure(num_sets <= 2);
+
+		if (dynamic_offsets.empty() &&
+			!memcmp(sets.data(), m_bound_descriptor_sets.data(), sets.size_bytes()))
+		{
+			return;
+		}
+
+		std::memcpy(m_bound_descriptor_sets.data(), sets.data(), sets.size_bytes());
+		vkCmdBindDescriptorSets(commands, bind_point, pipe_layout, 0, num_sets, sets.data(), ::size32(dynamic_offsets), dynamic_offsets.data());
+	}
+
+	void command_buffer::bind_descriptor_sets(
+		const std::span<VkDescriptorSet>& sets,
+		VkPipelineBindPoint bind_point,
+		VkPipelineLayout pipe_layout) const
+	{
+		const u32 num_sets = ::size32(sets);
+		ensure(num_sets <= 2);
+
+		if (!memcmp(sets.data(), m_bound_descriptor_sets.data(), sets.size_bytes()))
+		{
+			return;
+		}
+
+		std::memcpy(m_bound_descriptor_sets.data(), sets.data(), sets.size_bytes());
+		vkCmdBindDescriptorSets(commands, bind_point, pipe_layout, 0, num_sets, sets.data(), 0, nullptr);
+	}
 }

rpcs3/Emu/RSX/VK/vkutils/commands.h

@@ -71,6 +71,10 @@ namespace vk
 		command_pool* pool = nullptr;
 		VkCommandBuffer commands = nullptr;
 
+		// State cache
+		mutable std::array<VkDescriptorSet, 2> m_bound_descriptor_sets {{ VK_NULL_HANDLE }};
+		mutable VkPipeline m_bound_pipeline = VK_NULL_HANDLE;
+
 	public:
 		enum access_type_hint
 		{
@@ -97,40 +101,28 @@ namespace vk
 
 		void create(command_pool& cmd_pool);
 		void destroy();
+		void reset();
 
 		void begin();
 		void end();
 		void submit(queue_submit_t& submit_info, VkBool32 flush = VK_FALSE);
 
+		// Abstractions with caching
+		void bind_pipeline(VkPipeline pipeline, VkPipelineBindPoint bind_point) const;
+		void bind_descriptor_sets(const std::span<VkDescriptorSet>& sets, VkPipelineBindPoint bind_point, VkPipelineLayout pipe_layout) const;
+		void bind_descriptor_sets(const std::span<VkDescriptorSet>& sets, const std::span<u32>& dynamic_offsets, VkPipelineBindPoint bind_point, VkPipelineLayout pipe_layout) const;
+
 		// Properties
-		command_pool& get_command_pool() const
+		command_pool& get_command_pool() const { return *pool; }
-		{
+		u32 get_queue_family() const { return pool->get_queue_family(); }
-			return *pool;
+		bool is_recording() const { return is_open; }
-		}
 
-		u32 get_queue_family() const
+		void clear_flags() { flags = 0; }
-		{
+		void set_flag(command_buffer_data_flag flag) { flags |= flag; }
-			return pool->get_queue_family();
-		}
-
-		void clear_flags()
-		{
-			flags = 0;
-		}
-
-		void set_flag(command_buffer_data_flag flag)
-		{
-			flags |= flag;
-		}
 
 		operator VkCommandBuffer() const
 		{
 			return commands;
 		}
-
-		bool is_recording() const
-		{
-			return is_open;
-		}
 	};
 }

rpcs3/Emu/RSX/VK/vkutils/descriptors.cpp

@@ -453,7 +453,8 @@ namespace vk
 		// Notify
 		on_bind();
 
-		vkCmdBindDescriptorSets(cmd, bind_point, layout, 0, 1, &m_handle, ::size32(m_dynamic_offsets), m_dynamic_offsets.data());
+		VkDescriptorSet sets[1] = { m_handle };
+		cmd.bind_descriptor_sets(sets, m_dynamic_offsets, bind_point, layout);
 	}
 
 	void descriptor_set::flush()