rsx: Fixes

- vk: Always reopen primary command buffers. They should only be closed in flush_command_queue
- If uploading a texture and there are collisions with protected buffers, do not rebuild the cache
- Perform writes via flush before reprotecting pages that were not trampled
- Only flush no pages once

rpcs3/Emu/RSX/Common/texture_cache.h

@@ -273,13 +273,14 @@ namespace rsx
 			if (trampled_set.size() > 0)
 			{
 				// Rebuild the cache by only destroying ranges that need to be destroyed to unlock this page
-				const auto to_reprotect = std::remove_if(trampled_set.begin(), trampled_set.end(),
+				const auto to_reprotect = !rebuild_cache? trampled_set.end() :
+				std::remove_if(trampled_set.begin(), trampled_set.end(),
 				[&](const std::pair<section_storage_type*, ranged_storage*>& obj)
 				{
 					if (!is_writing && obj.first->get_protection() != utils::protection::no)
 						return true;
 
-					if (!rebuild_cache && !obj.first->is_flushable())
+					if (!obj.first->is_flushable())
 						return false;
 
 					const std::pair<u32, u32> null_check = std::make_pair(UINT32_MAX, 0);
@@ -309,6 +310,20 @@ namespace rsx
 					obj.second->remove_one();
 				}
 
+				if (allow_flush)
+				{
+					// Flush here before 'reprotecting' since flushing will write the whole span
+					for (const auto &tex : sections_to_flush)
+					{
+						if (!tex->flush(std::forward<Args>(extras)...))
+						{
+							//Missed address, note this
+							//TODO: Lower severity when successful to keep the cache from overworking
+							record_cache_miss(*tex);
+						}
+					}
+				}
+
 				for (auto It = to_reprotect; It != trampled_set.end(); It++)
 				{
 					auto obj = *It;
@@ -319,26 +334,9 @@ namespace rsx
 					obj.first->set_dirty(false);
 				}
 
-				if (discard_only)
+				if (discard_only || allow_flush)
 					return{ true, {} };
 
-				//trampled_set.erase(to_reprotect, trampled_set.end());
-
-				if (allow_flush)
-				{
-					for (const auto &tex : sections_to_flush)
-					{
-						if (!tex->flush(std::forward<Args>(extras)...))
-						{
-							//Missed address, note this
-							//TODO: Lower severity when successful to keep the cache from overworking
-							record_cache_miss(*tex);
-						}
-					}
-
-					return{ true, {} };
-				}
-
 				return std::make_pair(true, sections_to_flush);
 			}
 
@@ -348,7 +346,7 @@ namespace rsx
 		template <typename ...Args>
 		std::pair<bool, std::vector<section_storage_type*>> invalidate_range_impl(u32 address, u32 range, bool is_writing, bool discard, bool allow_flush, Args&&... extras)
 		{
-			return invalidate_range_impl_base(address, range, is_writing, discard, false, allow_flush, std::forward<Args>(extras)...);
+			return invalidate_range_impl_base(address, range, is_writing, discard, true, allow_flush, std::forward<Args>(extras)...);
 		}
 
 		bool is_hw_blit_engine_compatible(const u32 format) const
@@ -930,7 +928,8 @@ namespace rsx
 			auto subresources_layout = get_subresources_layout(tex);
 			auto remap_vector = tex.decoded_remap();
 
-			invalidate_range_impl(texaddr, tex_size, false, false, true, std::forward<Args>(extras)...);
+			//Invalidate with writing=false, discard=false, rebuild=false, native_flush=true
+			invalidate_range_impl_base(texaddr, tex_size, false, false, false, true, std::forward<Args>(extras)...);
 
 			m_texture_memory_in_use += (tex_pitch * tex_height);
 			return upload_image_from_cpu(cmd, texaddr, tex_width, tex_height, depth, tex.get_exact_mipmap_count(), tex_pitch, format,

rpcs3/Emu/RSX/GL/GLTextureCache.h

@@ -386,6 +386,8 @@ namespace gl
 
 		bool flush()
 		{
+			if (flushed) return true; //Already written, ignore
+
 			if (!copied)
 			{
 				LOG_WARNING(RSX, "Cache miss at address 0x%X. This is gonna hurt...", cpu_address_base);
@@ -399,7 +401,6 @@ namespace gl
 				}
 			}
 
-			protect(utils::protection::rw);
 			m_fence.wait_for_signal();
 			flushed = true;
 

rpcs3/Emu/RSX/VK/VKGSRender.cpp

@@ -567,6 +567,7 @@ VKGSRender::VKGSRender() : GSRender()
 	//Create secondary command_buffer for parallel operations
 	m_secondary_command_buffer_pool.create((*m_device));
 	m_secondary_command_buffer.create(m_secondary_command_buffer_pool);
+	m_secondary_command_buffer.access_hint = vk::command_buffer::access_type_hint::all;
 	
 	//Precalculated stuff
 	m_render_passes = get_precomputed_render_passes(*m_device, m_optimal_tiling_supported_formats);

rpcs3/Emu/RSX/VK/VKHelpers.h

@@ -1013,6 +1013,14 @@ namespace vk
 		vk::command_pool *pool = nullptr;
 		VkCommandBuffer commands = nullptr;
 
+	public:
+		enum access_type_hint
+		{
+			flush_only, //Only to be submitted/opened/closed via command flush
+			all         //Auxilliary, can be sumitted/opened/closed at any time
+		}
+		access_hint = flush_only;
+
 	public:
 		command_buffer() {}
 		~command_buffer() {}

rpcs3/Emu/RSX/VK/VKTextureCache.h

@@ -20,6 +20,7 @@ namespace vk
 		//DMA relevant data
 		VkFence dma_fence = VK_NULL_HANDLE;
 		bool synchronized = false;
+		bool flushed = false;
 		u64 sync_timestamp = 0;
 		u64 last_use_timestamp = 0;
 		vk::render_device* m_device = nullptr;
@@ -62,6 +63,7 @@ namespace vk
 			//Even if we are managing the same vram section, we cannot guarantee contents are static
 			//The create method is only invoked when a new mangaged session is required
 			synchronized = false;
+			flushed = false;
 			sync_timestamp = 0ull;
 			last_use_timestamp = get_system_time();
 		}
@@ -179,6 +181,9 @@ namespace vk
 				CHECK_RESULT(vkWaitForFences(*m_device, 1, &dma_fence, VK_TRUE, UINT64_MAX));
 				CHECK_RESULT(vkResetCommandBuffer(cmd, 0));
 				CHECK_RESULT(vkResetFences(*m_device, 1, &dma_fence));
+
+				if (cmd.access_hint != vk::command_buffer::access_type_hint::all)
+					cmd.begin();
 			}
 
 			synchronized = true;
@@ -204,6 +209,8 @@ namespace vk
 
 		bool flush(vk::render_device& dev, vk::command_buffer& cmd, vk::memory_type_mapping& memory_types, VkQueue submit_queue)
 		{
+			if (flushed) return true;
+
 			if (m_device == nullptr)
 				m_device = &dev;
 
@@ -217,7 +224,7 @@ namespace vk
 				result = false;
 			}
 
-			protect(utils::protection::rw);
+			flushed = true;
 
 			void* pixels_src = dma_buffer->map(0, cpu_address_range);
 			void* pixels_dst = vm::base(cpu_address_base);