rsx: improve memory coherency

- Avoid tagging and rely on read/write barriers and the dirty flag mechanism. Testing is done with a weak 8-byte memory test
- Introducing new data when tagging breaks applications with race conditions where tags can overwrite flushed data

rpcs3/Emu/RSX/Common/surface_store.h

@@ -92,10 +92,10 @@ namespace rsx
 	struct render_target_descriptor
 	{
 		u64 last_use_tag = 0;         // tag indicating when this block was last confirmed to have been written to
-		u32 tag_address = 0;
+		u32 memory_tag_address = 0u;  // memory address of the start of the ROP block
+		u64 memory_tag_sample = 0ull; // memory sample taken at the memory_tag_address for change testing
 
 		bool dirty = false;
-		bool needs_tagging = false;
 		image_storage_type old_contents = nullptr;
 		rsx::surface_antialiasing read_aa_mode = rsx::surface_antialiasing::center_1_sample;
 
@@ -119,31 +119,26 @@ namespace rsx
 			write_aa_mode = read_aa_mode = rsx::surface_antialiasing::center_1_sample;
 		}
 
-		void tag()
+		bool test() const
 		{
-			auto ptr = vm::get_super_ptr<atomic_t<u32>>(tag_address);
+			if (dirty)
-			*ptr = tag_address;
-
-			needs_tagging = false;
-		}
-
-		bool test()
-		{
-			if (needs_tagging && dirty)
 			{
 				// TODO
 				// Should RCB or mem-sync (inherit previous mem) to init memory
 				LOG_TODO(RSX, "Resource used before memory initialization");
 			}
 
-			auto ptr = vm::get_super_ptr<atomic_t<u32>>(tag_address);
+			return (memory_tag_sample == *vm::get_super_ptr<u64>(memory_tag_address));
-			return (*ptr == tag_address);
 		}
 
 		void queue_tag(u32 address)
 		{
-			tag_address = address;
+			memory_tag_address = address;
-			needs_tagging = true;
+		}
+
+		void sync_tag()
+		{
+			memory_tag_sample = *vm::get_super_ptr<u64>(memory_tag_address);
 		}
 
 		void on_write(u64 write_tag = 0)
@@ -154,10 +149,8 @@ namespace rsx
 				last_use_tag = write_tag;
 			}
 
-			if (needs_tagging)
+			// Tag unconditionally without introducing new data
-			{
+			sync_tag();
-				tag();
-			}
 
 			read_aa_mode = write_aa_mode;
 			dirty = false;

rpcs3/Emu/RSX/Common/texture_cache.h

@@ -372,18 +372,6 @@ namespace rsx
 			return true;
 		}
 
-		void tag_framebuffer(u32 texaddr)
-		{
-			auto ptr = vm::get_super_ptr<atomic_t<u32>>(texaddr);
-			*ptr = texaddr;
-		}
-
-		bool test_framebuffer(u32 texaddr)
-		{
-			auto ptr = vm::get_super_ptr<atomic_t<u32>>(texaddr);
-			return *ptr == texaddr;
-		}
-
 		/**
 		 * Section invalidation
 		 */
@@ -1157,7 +1145,6 @@ namespace rsx
 
 			region.create(width, height, 1, 1, image, pitch, false, std::forward<Args>(extras)...);
 			region.reprotect(utils::protection::no, { 0, rsx_range.length() });
-			tag_framebuffer(region.get_section_base());
 
 			region.set_dirty(false);
 			region.touch(m_cache_update_tag);
@@ -1703,8 +1690,8 @@ namespace rsx
 				// TODO: When framebuffer Y compression is properly handled, this section can be removed. A more accurate framebuffer storage check exists below this block
 				if (auto texptr = m_rtts.get_texture_from_render_target_if_applicable(texaddr))
 				{
-					const bool is_active = m_rtts.address_is_bound(texaddr, false);
+					if (const bool is_active = m_rtts.address_is_bound(texaddr, false);
-					if (texptr->test() || is_active)
+						is_active || texptr->test())
 					{
 						return process_framebuffer_resource(cmd, texptr, texaddr, tex.format(), m_rtts,
 								tex_width, tex_height, depth, tex_pitch, extended_dimension, false, is_active,
@@ -1719,8 +1706,8 @@ namespace rsx
 
 				if (auto texptr = m_rtts.get_texture_from_depth_stencil_if_applicable(texaddr))
 				{
-					const bool is_active = m_rtts.address_is_bound(texaddr, true);
+					if (const bool is_active = m_rtts.address_is_bound(texaddr, false);
-					if (texptr->test() || is_active)
+						is_active || texptr->test())
 					{
 						return process_framebuffer_resource(cmd, texptr, texaddr, tex.format(), m_rtts,
 								tex_width, tex_height, depth, tex_pitch, extended_dimension, true, is_active,
@@ -1949,6 +1936,8 @@ namespace rsx
 
 			if (src_is_render_target)
 			{
+				src_subres.surface->read_barrier(cmd);
+
 				const auto surf = src_subres.surface;
 				auto src_bpp = surf->get_native_pitch() / surf->get_surface_width();
 				auto expected_bpp = src_is_argb8 ? 4 : 2;
@@ -1972,6 +1961,9 @@ namespace rsx
 
 			if (dst_is_render_target)
 			{
+				// Full barrier is required in case of partial transfers
+				dst_subres.surface->read_barrier(cmd);
+
 				auto dst_bpp = dst_subres.surface->get_native_pitch() / dst_subres.surface->get_surface_width();
 				auto expected_bpp = dst_is_argb8 ? 4 : 2;
 				if (dst_bpp != expected_bpp)
@@ -2411,7 +2403,6 @@ namespace rsx
 							AUDIT(section.get_memory_read_flags() == memory_read_flags::flush_always);
 
 							section.reprotect(utils::protection::no);
-							tag_framebuffer(section.get_section_base());
 							update_tag = true;
 						}
 					}

rpcs3/Emu/RSX/GL/GLGSRender.cpp

@@ -309,6 +309,7 @@ void GLGSRender::end()
 		// Program is not ready, skip drawing this
 		std::this_thread::yield();
 		execute_nop_draw();
+		m_rtts.on_write();
 		rsx::thread::end();
 		return;
 	}

rpcs3/Emu/RSX/GL/GLTextureCache.h

@@ -261,6 +261,12 @@ namespace gl
 				baseclass::on_speculative_flush();
 			}
 
+			if (context == rsx::texture_upload_context::framebuffer_storage)
+			{
+				auto as_rtt = static_cast<gl::render_target*>(vram_texture);
+				if (as_rtt->dirty) as_rtt->read_barrier(cmd);
+			}
+
 			if (!pbo_id)
 			{
 				init_buffer();
@@ -403,7 +409,6 @@ namespace gl
 			glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
 			glBindBuffer(GL_PIXEL_PACK_BUFFER, GL_NONE);
 
-
 			// Shuffle
 			bool require_manual_shuffle = false;
 			if (pack_unpack_swap_bytes)
@@ -424,7 +429,6 @@ namespace gl
 			}
 			else if (pack_unpack_swap_bytes && ::gl::get_driver_caps().vendor_AMD)
 			{
-
 				//AMD driver bug - cannot use pack_swap_bytes
 				//Manually byteswap texel data
 				switch (type)
@@ -474,6 +478,12 @@ namespace gl
 				}
 				}
 			}
+
+			if (context == rsx::texture_upload_context::framebuffer_storage)
+			{
+				// Update memory tag
+				static_cast<gl::render_target*>(vram_texture)->sync_tag();
+			}
 		}
 
 		/**

rpcs3/Emu/RSX/VK/VKGSRender.cpp

@@ -1504,6 +1504,7 @@ void VKGSRender::end()
 		// Program is not ready, skip drawing this
 		std::this_thread::yield();
 		execute_nop_draw();
+		m_rtts.on_write();
 		rsx::thread::end();
 		return;
 	}

rpcs3/Emu/RSX/VK/VKTextureCache.h

@@ -180,6 +180,12 @@ namespace vk
 				cmd.begin();
 			}
 
+			if (context == rsx::texture_upload_context::framebuffer_storage)
+			{
+				auto as_rtt = static_cast<vk::render_target*>(vram_texture);
+				if (as_rtt->dirty) as_rtt->read_barrier(cmd);
+			}
+
 			vk::image *target = vram_texture;
 			real_pitch = vk::get_format_texel_width(vram_texture->info.format) * vram_texture->width();
 
@@ -333,6 +339,12 @@ namespace vk
 		void finish_flush()
 		{
 			dma_buffer->unmap();
+
+			if (context == rsx::texture_upload_context::framebuffer_storage)
+			{
+				// Update memory tag
+				static_cast<vk::render_target*>(vram_texture)->sync_tag();
+			}
 		}