rsx: Fix temporary subresource caching behaviour

- Do not cache if a gathered subresource contains a bound RTT
- Change op to dynamic copy if parent is still bound

rpcs3/Emu/RSX/Common/texture_cache.h

@@ -178,6 +178,7 @@ namespace rsx
 			u16 width = 0;
 			u16 height = 0;
 			u16 depth = 1;
+			bool do_not_cache = false;
 
 			deferred_subresource()
 			{}
@@ -1506,20 +1507,23 @@ namespace rsx
 
 		image_view_type create_temporary_subresource(commandbuffer_type &cmd, deferred_subresource& desc)
 		{
-			const auto found = m_temporary_subresource_cache.equal_range(desc.base_address);
+			if (!desc.do_not_cache)
-			for (auto It = found.first; It != found.second; ++It)
 			{
-				const auto& found_desc = It->second.first;
+				const auto found = m_temporary_subresource_cache.equal_range(desc.base_address);
-				if (found_desc.external_handle != desc.external_handle ||
+				for (auto It = found.first; It != found.second; ++It)
-					found_desc.op != desc.op ||
+				{
-					found_desc.x != desc.x || found_desc.y != desc.y ||
+					const auto& found_desc = It->second.first;
-					found_desc.width != desc.width || found_desc.height != desc.height)
+					if (found_desc.external_handle != desc.external_handle ||
-					continue;
+						found_desc.op != desc.op ||
+						found_desc.x != desc.x || found_desc.y != desc.y ||
+						found_desc.width != desc.width || found_desc.height != desc.height)
+						continue;
 
-				if (desc.op == deferred_request_command::copy_image_dynamic)
+					if (desc.op == deferred_request_command::copy_image_dynamic)
-					update_image_contents(cmd, It->second.second, desc.external_handle, desc.width, desc.height);
+						update_image_contents(cmd, It->second.second, desc.external_handle, desc.width, desc.height);
 
-				return It->second.second;
+					return It->second.second;
+				}
 			}
 
 			image_view_type result = 0;
@@ -1600,9 +1604,19 @@ namespace rsx
 			return result;
 		}
 
-		void notify_surface_changed(u32 base_address)
+		void notify_surface_changed(const utils::address_range& range)
 		{
-			m_temporary_subresource_cache.erase(base_address);
+			for (auto It = m_temporary_subresource_cache.begin(); It != m_temporary_subresource_cache.end();)
+			{
+				if (range.overlaps(It->first))
+				{
+					It = m_temporary_subresource_cache.erase(It);
+				}
+				else
+				{
+					++It;
+				}
+			}
 		}
 
 		template<typename surface_store_list_type>
@@ -2271,6 +2285,25 @@ namespace rsx
 						// TODO: Overlapped section persistance is required for framebuffer resources to work with this!
 						// Yellow filter in SCV is because of a 384x384 surface being reused as 160x90 (and likely not getting written to)
 						// Its then sampled again here as 384x384 and this does not work! (obviously)
+
+						// Optionally disallow caching if resource is being written to as it is being read from
+						for (const auto &section : overlapping_fbos)
+						{
+							if (m_rtts.address_is_bound(section.base_address))
+							{
+								if (result.external_subresource_desc.op == deferred_request_command::copy_image_static)
+								{
+									result.external_subresource_desc.op = deferred_request_command::copy_image_dynamic;
+								}
+								else
+								{
+									result.external_subresource_desc.do_not_cache = true;
+								}
+
+								break;
+							}
+						}
+
 						return result;
 					}
 					else
@@ -2798,23 +2831,26 @@ namespace rsx
 				typeless_info.dst_context = texture_upload_context::blit_engine_dst;
 			}
 
+			// Calculate number of bytes actually modified
+			u32 mem_length;
+			const u32 mem_base = dst_address - dst.rsx_address;
+			if (dst.clip_height == 1)
+			{
+				mem_length = dst.clip_width * dst_bpp;
+			}
+			else
+			{
+				const u32 mem_excess = mem_base % dst.pitch;
+				mem_length = (dst.pitch * dst.clip_height) - mem_excess;
+			}
+
+			// Invalidate any cached subresources in modified range
+			notify_surface_changed(utils::address_range::start_length(dst_address, mem_length));
+
 			if (cached_dest)
 			{
 				lock.upgrade();
 
-				u32 mem_length;
-				const u32 mem_base = dst_address - cached_dest->get_section_base();
-
-				if (dst.clip_height == 1)
-				{
-					mem_length = dst.clip_width * dst_bpp;
-				}
-				else
-				{
-					const u32 mem_excess = mem_base % dst.pitch;
-					mem_length = (dst.pitch * dst.clip_height) - mem_excess;
-				}
-
 				verify(HERE), (mem_base + mem_length) <= cached_dest->get_section_size();
 
 				cached_dest->reprotect(utils::protection::no, { mem_base, mem_length });
@@ -2863,7 +2899,6 @@ namespace rsx
 
 			typeless_info.analyse();
 			blitter.scale_image(cmd, vram_texture, dest_texture, src_area, dst_area, interpolate, is_depth_blit, typeless_info);
-			notify_surface_changed(dst.rsx_address);
 
 			blit_op_result result = true;
 			result.is_depth = is_depth_blit;

rpcs3/Emu/RSX/GL/GLRenderTargets.cpp

@@ -261,7 +261,7 @@ void GLGSRender::init_buffers(rsx::framebuffer_creation_context context, bool sk
 
 			rtt->tile = find_tile(color_offsets[i], color_locations[i]);
 			rtt->write_aa_mode = layout.aa_mode;
-			m_gl_texture_cache.notify_surface_changed(m_surface_info[i].address);
+			m_gl_texture_cache.notify_surface_changed(m_surface_info[i].get_memory_range(layout.aa_factors));
 		}
 		else
 		{
@@ -270,18 +270,18 @@ void GLGSRender::init_buffers(rsx::framebuffer_creation_context context, bool sk
 		}
 	}
 
+	if (m_depth_surface_info.pitch && g_cfg.video.write_depth_buffer)
+	{
+		auto bpp = m_depth_surface_info.pitch / m_depth_surface_info.width;
+		auto old_format = (bpp == 2) ? gl::texture::format::depth : gl::texture::format::depth_stencil;
+
+		const utils::address_range surface_range = m_depth_surface_info.get_memory_range();
+		m_gl_texture_cache.set_memory_read_flags(surface_range, rsx::memory_read_flags::flush_once);
+		m_gl_texture_cache.flush_if_cache_miss_likely(cmd, surface_range);
+	}
+
 	if (std::get<0>(m_rtts.m_bound_depth_stencil))
 	{
-		if (m_depth_surface_info.pitch && g_cfg.video.write_depth_buffer)
-		{
-			auto bpp = m_depth_surface_info.pitch / m_depth_surface_info.width;
-			auto old_format = (bpp == 2) ? gl::texture::format::depth : gl::texture::format::depth_stencil;
-
-			const utils::address_range surface_range = m_depth_surface_info.get_memory_range();
-			m_gl_texture_cache.set_memory_read_flags(surface_range, rsx::memory_read_flags::flush_once);
-			m_gl_texture_cache.flush_if_cache_miss_likely(cmd, surface_range);
-		}
-
 		auto ds = std::get<1>(m_rtts.m_bound_depth_stencil);
 		depth_stencil_target = ds->id();
 
@@ -289,7 +289,7 @@ void GLGSRender::init_buffers(rsx::framebuffer_creation_context context, bool sk
 		m_depth_surface_info = { layout.zeta_address, layout.actual_zeta_pitch, true, layout.color_format, layout.depth_format, layout.width, layout.height, depth_bpp };
 
 		ds->write_aa_mode = layout.aa_mode;
-		m_gl_texture_cache.notify_surface_changed(layout.zeta_address);
+		m_gl_texture_cache.notify_surface_changed(m_depth_surface_info.get_memory_range(layout.aa_factors));
 	}
 	else
 	{
@@ -387,7 +387,7 @@ void GLGSRender::init_buffers(rsx::framebuffer_creation_context context, bool sk
 	{
 		if (!m_surface_info[i].address || !m_surface_info[i].pitch) continue;
 
-		const auto surface_range = m_surface_info[i].get_memory_range(layout.aa_factors);
+		const auto surface_range = m_surface_info[i].get_memory_range();
 		if (g_cfg.video.write_color_buffers)
 		{
 			// Mark buffer regions as NO_ACCESS on Cell-visible side
@@ -402,7 +402,7 @@ void GLGSRender::init_buffers(rsx::framebuffer_creation_context context, bool sk
 
 	if (m_depth_surface_info.address && m_depth_surface_info.pitch)
 	{
-		const auto surface_range = m_depth_surface_info.get_memory_range(layout.aa_factors);
+		const auto surface_range = m_depth_surface_info.get_memory_range();
 		if (g_cfg.video.write_depth_buffer)
 		{
 			const auto depth_format_gl = rsx::internals::surface_depth_format_to_gl(layout.depth_format);

rpcs3/Emu/RSX/VK/VKGSRender.cpp

@@ -2872,7 +2872,7 @@ void VKGSRender::prepare_rtts(rsx::framebuffer_creation_context context)
 
 	for (u8 i = 0; i < rsx::limits::color_buffers_count; ++i)
 	{
-		//Flush old address if we keep missing it
+		// Flush old address if we keep missing it
 		if (m_surface_info[i].pitch && g_cfg.video.write_color_buffers)
 		{
 			if (old_format == VK_FORMAT_UNDEFINED)
@@ -2925,7 +2925,7 @@ void VKGSRender::prepare_rtts(rsx::framebuffer_creation_context context)
 			verify("Pitch mismatch!" HERE), surface->rsx_pitch == layout.actual_color_pitch[index];
 
 			surface->write_aa_mode = layout.aa_mode;
-			m_texture_cache.notify_surface_changed(layout.color_addresses[index]);
+			m_texture_cache.notify_surface_changed(m_surface_info[index].get_memory_range(layout.aa_factors));
 			m_draw_buffers.push_back(index);
 		}
 	}
@@ -2940,7 +2940,7 @@ void VKGSRender::prepare_rtts(rsx::framebuffer_creation_context context)
 		verify("Pitch mismatch!" HERE), ds->rsx_pitch == layout.actual_zeta_pitch;
 
 		ds->write_aa_mode = layout.aa_mode;
-		m_texture_cache.notify_surface_changed(layout.zeta_address);
+		m_texture_cache.notify_surface_changed(m_depth_surface_info.get_memory_range(layout.aa_factors));
 	}
 
 	// Before messing with memory properties, flush command queue if there are dma transfers queued up
@@ -2954,7 +2954,7 @@ void VKGSRender::prepare_rtts(rsx::framebuffer_creation_context context)
 	{
 		if (!m_surface_info[index].address || !m_surface_info[index].pitch) continue;
 
-		const utils::address_range surface_range = m_surface_info[index].get_memory_range(layout.aa_factors);
+		const utils::address_range surface_range = m_surface_info[index].get_memory_range();
 		if (g_cfg.video.write_color_buffers)
 		{
 			m_texture_cache.lock_memory_region(*m_current_command_buffer, std::get<1>(m_rtts.m_bound_render_targets[index]), surface_range,
@@ -2968,7 +2968,7 @@ void VKGSRender::prepare_rtts(rsx::framebuffer_creation_context context)
 
 	if (m_depth_surface_info.address && m_depth_surface_info.pitch)
 	{
-		const utils::address_range surface_range = m_depth_surface_info.get_memory_range(layout.aa_factors);
+		const utils::address_range surface_range = m_depth_surface_info.get_memory_range();
 		if (g_cfg.video.write_depth_buffer)
 		{
 			const u32 gcm_format = (m_depth_surface_info.depth_format != rsx::surface_depth_format::z16) ? CELL_GCM_TEXTURE_DEPTH16 : CELL_GCM_TEXTURE_DEPTH24_D8;

rpcs3/Emu/RSX/rsx_utils.h

@@ -81,6 +81,7 @@ namespace rsx
 
 		address_range get_memory_range() const
 		{
+			verify(HERE), range.start == address;
 			return range;
 		}
 	};