rsx: Set up for multi-section inheritance

rpcs3/Emu/RSX/Common/surface_store.h

@@ -188,7 +188,7 @@ namespace rsx
 		template <bool is_depth_surface>
 		void intersect_surface_region(command_list_type cmd, u32 address, surface_type new_surface, surface_type prev_surface)
 		{
-			auto scan_list = [&new_surface, address](const rsx::address_range& mem_range, u64 timestamp_check,
+			auto scan_list = [&new_surface, address](const rsx::address_range& mem_range,
 				std::unordered_map<u32, surface_storage_type>& data) -> std::vector<std::pair<u32, surface_type>>
 			{
 				std::vector<std::pair<u32, surface_type>> result;
@@ -196,7 +196,7 @@ namespace rsx
 				{
 					auto surface = Traits::get(e.second);
 
-					if (e.second->last_use_tag <= timestamp_check ||
+					if (new_surface->last_use_tag > surface->last_use_tag ||
 						new_surface == surface ||
 						address == e.first ||
 						e.second->dirty())
@@ -235,10 +235,8 @@ namespace rsx
 			};
 
 			const rsx::address_range mem_range = new_surface->get_memory_range();
-			const u64 timestamp_check = prev_surface ? prev_surface->last_use_tag : new_surface->last_use_tag;
+			auto list1 = scan_list(mem_range, m_render_targets_storage);
-
+			auto list2 = scan_list(mem_range, m_depth_stencil_storage);
-			auto list1 = scan_list(mem_range, timestamp_check, m_render_targets_storage);
-			auto list2 = scan_list(mem_range, timestamp_check, m_depth_stencil_storage);
 
 			if (prev_surface)
 			{
@@ -279,10 +277,10 @@ namespace rsx
 
 			if (UNLIKELY(surface_info.size() > 1))
 			{
-				// Sort with newest first for early exit
+				// Sort with oldest first for early exit
 				std::sort(surface_info.begin(), surface_info.end(), [](const auto& a, const auto& b)
 				{
-					return (a.second->last_use_tag > b.second->last_use_tag);
+					return (a.second->last_use_tag < b.second->last_use_tag);
 				});
 			}
 
@@ -313,6 +311,12 @@ namespace rsx
 					continue;
 				}
 
+				if (child_w == size.width && child_h == size.height && surface_info.size() > 1)
+				{
+					// If the write covers the whole area, discard anything older
+					new_surface->clear_rw_barrier();
+				}
+
 				// TODO: Eventually need to stack all the overlapping regions, but for now just do the latest rect in the space
 				deferred_clipped_region<surface_type> region;
 				region.src_x = src_offset.x;
@@ -325,7 +329,6 @@ namespace rsx
 				region.target = new_surface;
 
 				new_surface->set_old_contents_region(region, true);
-				break;
 			}
 		}
 
@@ -367,12 +370,6 @@ namespace rsx
 				surface_storage_type &surface = It->second;
 				const bool pitch_compatible = Traits::surface_is_pitch_compatible(surface, pitch);
 
-				if (pitch_compatible)
-				{
-					// Preserve memory outside the area to be inherited if needed
-					split_surface_region<depth>(command_list, address, Traits::get(surface), (u16)width, (u16)height, bpp, antialias);
-				}
-
 				if (Traits::surface_matches_properties(surface, format, width, height, antialias))
 				{
 					if (pitch_compatible)
@@ -386,6 +383,12 @@ namespace rsx
 				}
 				else
 				{
+					if (pitch_compatible)
+					{
+						// Preserve memory outside the area to be inherited if needed
+						split_surface_region<depth>(command_list, address, Traits::get(surface), (u16)width, (u16)height, bpp, antialias);
+					}
+
 					old_surface = Traits::get(surface);
 					old_surface_storage = std::move(surface);
 					primary_storage->erase(It);
@@ -457,12 +460,22 @@ namespace rsx
 				}
 			}
 
-			// Check if old_surface is 'new' and avoid intersection
+			bool do_intersection_test = true;
+
+			// Check if old_surface is 'new' and hopefully avoid intersection
 			if (old_surface && old_surface->last_use_tag >= write_tag)
 			{
+				const auto new_area = new_surface->get_normalized_memory_area();
+				const auto old_area = old_surface->get_normalized_memory_area();
+
+				if (new_area.x2 <= old_area.x2 && new_area.y2 <= old_area.y2)
+				{
+					do_intersection_test = false;
 					new_surface->set_old_contents(old_surface);
 				}
-			else
+			}
+
+			if (do_intersection_test)
 			{
 				intersect_surface_region<depth>(command_list, address, new_surface, old_surface);
 			}

rpcs3/Emu/RSX/Common/surface_utils.h

@@ -51,6 +51,9 @@ namespace rsx
 	template <typename surface_type>
 	struct deferred_clipped_region
 	{
+		// Chain
+		deferred_clipped_region<surface_type>* next_ptr = nullptr;
+
 		u16 src_x, src_y, dst_x, dst_y, width, height;
 		f32 transfer_scale_x, transfer_scale_y;
 		surface_type target;
@@ -130,8 +133,7 @@ namespace rsx
 		u64 last_use_tag = 0;         // tag indicating when this block was last confirmed to have been written to
 		std::array<std::pair<u32, u64>, 5> memory_tag_samples;
 
-		// Obsolete, requires updating
+		std::vector<deferred_clipped_region<image_storage_type>> old_contents;
-		deferred_clipped_region<image_storage_type> old_contents{};
 
 		// Surface properties
 		u16 rsx_pitch = 0;
@@ -161,7 +163,7 @@ namespace rsx
 
 		virtual ~render_target_descriptor()
 		{
-			if (old_contents)
+			if (!old_contents.empty())
 			{
 				// Cascade resource derefs
 				LOG_ERROR(RSX, "Resource was destroyed whilst holding a resource reference!");
@@ -284,7 +286,7 @@ namespace rsx
 
 		bool dirty() const
 		{
-			return (state_flags != rsx::surface_state_flags::ready) || old_contents;
+			return (state_flags != rsx::surface_state_flags::ready) || !old_contents.empty();
 		}
 
 		bool test() const
@@ -311,45 +313,37 @@ namespace rsx
 
 		void clear_rw_barrier()
 		{
-			release_ref(old_contents.source);
+			for (auto &e : old_contents)
-			old_contents = {};
+			{
+				release_ref(e.source);
+			}
+
+			old_contents.clear();
 		}
 
 		template<typename T>
 		void set_old_contents(T* other)
 		{
-			verify(HERE), !old_contents;
+			verify(HERE), old_contents.empty();
 
 			if (!other || other->get_rsx_pitch() != this->get_rsx_pitch())
 			{
-				old_contents = {};
 				return;
 			}
 
-			old_contents = {};
+			old_contents.emplace_back();
-			old_contents.source = other;
+			old_contents.back().source = other;
 			other->add_ref();
 		}
 
 		template<typename T>
 		void set_old_contents_region(const T& region, bool normalized)
 		{
-			if (old_contents)
-			{
-				// This can happen when doing memory splits
-				auto old_surface = static_cast<decltype(region.source)>(old_contents.source);
-				if (old_surface->last_use_tag > region.source->last_use_tag)
-				{
-					return;
-				}
-
-				clear_rw_barrier();
-			}
-
 			// NOTE: This method will not perform pitch verification!
-			verify(HERE), !old_contents, region.source, region.source != this;
+			verify(HERE), region.source, region.source != this;
 
-			old_contents = region.template cast<image_storage_type>();
+			old_contents.push_back(region.template cast<image_storage_type>());
+			auto &slice = old_contents.back();
 			region.source->add_ref();
 
 			// Reverse normalization process if needed
@@ -357,39 +351,39 @@ namespace rsx
 			{
 				const u16 bytes_to_texels_x = region.source->get_bpp() * region.source->samples_x;
 				const u16 rows_to_texels_y = region.source->samples_y;
-				old_contents.src_x /= bytes_to_texels_x;
+				slice.src_x /= bytes_to_texels_x;
-				old_contents.src_y /= rows_to_texels_y;
+				slice.src_y /= rows_to_texels_y;
-				old_contents.width /= bytes_to_texels_x;
+				slice.width /= bytes_to_texels_x;
-				old_contents.height /= rows_to_texels_y;
+				slice.height /= rows_to_texels_y;
 
 				const u16 bytes_to_texels_x2 = (get_bpp() * samples_x);
 				const u16 rows_to_texels_y2 = samples_y;
-				old_contents.dst_x /= bytes_to_texels_x2;
+				slice.dst_x /= bytes_to_texels_x2;
-				old_contents.dst_y /= rows_to_texels_y2;
+				slice.dst_y /= rows_to_texels_y2;
 
-				old_contents.transfer_scale_x = f32(bytes_to_texels_x) / bytes_to_texels_x2;
+				slice.transfer_scale_x = f32(bytes_to_texels_x) / bytes_to_texels_x2;
-				old_contents.transfer_scale_y = f32(rows_to_texels_y) / rows_to_texels_y2;
+				slice.transfer_scale_y = f32(rows_to_texels_y) / rows_to_texels_y2;
 			}
 
 			// Apply resolution scale if needed
 			if (g_cfg.video.resolution_scale_percent != 100)
 			{
-				auto src_width = rsx::apply_resolution_scale(old_contents.width, true, old_contents.source->width());
+				auto src_width = rsx::apply_resolution_scale(slice.width, true, slice.source->width());
-				auto src_height = rsx::apply_resolution_scale(old_contents.height, true, old_contents.source->height());
+				auto src_height = rsx::apply_resolution_scale(slice.height, true, slice.source->height());
 
-				auto dst_width = rsx::apply_resolution_scale(old_contents.width, true, old_contents.target->width());
+				auto dst_width = rsx::apply_resolution_scale(slice.width, true, slice.target->width());
-				auto dst_height = rsx::apply_resolution_scale(old_contents.height, true, old_contents.target->height());
+				auto dst_height = rsx::apply_resolution_scale(slice.height, true, slice.target->height());
 
-				old_contents.transfer_scale_x *= f32(dst_width) / src_width;
+				slice.transfer_scale_x *= f32(dst_width) / src_width;
-				old_contents.transfer_scale_y *= f32(dst_height) / src_height;
+				slice.transfer_scale_y *= f32(dst_height) / src_height;
 
-				old_contents.width = src_width;
+				slice.width = src_width;
-				old_contents.height = src_height;
+				slice.height = src_height;
 
-				old_contents.src_x = rsx::apply_resolution_scale(old_contents.src_x, false, old_contents.source->width());
+				slice.src_x = rsx::apply_resolution_scale(slice.src_x, false, slice.source->width());
-				old_contents.src_y = rsx::apply_resolution_scale(old_contents.src_y, false, old_contents.source->height());
+				slice.src_y = rsx::apply_resolution_scale(slice.src_y, false, slice.source->height());
-				old_contents.dst_x = rsx::apply_resolution_scale(old_contents.dst_x, false, old_contents.target->width());
+				slice.dst_x = rsx::apply_resolution_scale(slice.dst_x, false, slice.target->width());
-				old_contents.dst_y = rsx::apply_resolution_scale(old_contents.dst_y, false, old_contents.target->height());
+				slice.dst_y = rsx::apply_resolution_scale(slice.dst_y, false, slice.target->height());
 			}
 		}
 
@@ -457,7 +451,7 @@ namespace rsx
 				msaa_flags = resolve_flags;
 			}
 
-			if (old_contents.source)
+			if (!old_contents.empty())
 			{
 				clear_rw_barrier();
 			}

rpcs3/Emu/RSX/GL/GLGSRender.cpp

@@ -213,19 +213,18 @@ void GLGSRender::end()
 	gl::render_target *ds = std::get<1>(m_rtts.m_bound_depth_stencil);
 
 	// Handle special memory barrier for ARGB8->D24S8 in an active DSV
-	if (ds && ds->old_contents &&
+	if (ds && ds->old_contents.size() == 1 &&
-		ds->old_contents.source->get_internal_format() == gl::texture::internal_format::rgba8 &&
+		ds->old_contents[0].source->get_internal_format() == gl::texture::internal_format::rgba8)
-		rsx::pitch_compatible(ds, gl::as_rtt(ds->old_contents.source)))
 	{
 		gl_state.enable(GL_FALSE, GL_SCISSOR_TEST);
 
 		// TODO: Stencil transfer
 		gl::g_hw_blitter->fast_clear_image(cmd, ds, 1.f, 0xFF);
-		ds->old_contents.init_transfer(ds);
+		ds->old_contents[0].init_transfer(ds);
 
-		m_depth_converter.run(ds->old_contents.src_rect(),
+		m_depth_converter.run(ds->old_contents[0].src_rect(),
-			ds->old_contents.dst_rect(),
+			ds->old_contents[0].dst_rect(),
-			ds->old_contents.source, ds);
+			ds->old_contents[0].source, ds);
 
 		ds->on_write();
 	}

rpcs3/Emu/RSX/GL/GLRenderTargets.cpp

@@ -1,4 +1,4 @@
-#include "stdafx.h"
+#include "stdafx.h"
 #include "GLGSRender.h"
 #include "Emu/System.h"
 
@@ -609,7 +609,7 @@ void gl::render_target::memory_barrier(gl::command_context& cmd, bool force_init
 		state_flags &= ~rsx::surface_state_flags::erase_bkgnd;
 	};
 
-	if (!old_contents)
+	if (old_contents.empty())
 	{
 		// No memory to inherit
 		if (dirty() && (force_init || state_flags & rsx::surface_state_flags::erase_bkgnd))
@@ -623,15 +623,9 @@ void gl::render_target::memory_barrier(gl::command_context& cmd, bool force_init
 		return;
 	}
 
-	auto src_texture = gl::as_rtt(old_contents.source);
+	for (auto &section : old_contents)
-	if (!rsx::pitch_compatible(this, src_texture))
 	{
-		LOG_TRACE(RSX, "Pitch mismatch, could not transfer inherited memory");
+		auto src_texture = gl::as_rtt(section.source);
-
-		clear_rw_barrier();
-		return;
-	}
-
 		const auto src_bpp = src_texture->get_bpp();
 		const auto dst_bpp = get_bpp();
 		rsx::typeless_xfer typeless_info{};
@@ -656,11 +650,11 @@ void gl::render_target::memory_barrier(gl::command_context& cmd, bool force_init
 		}
 
 		const bool dst_is_depth = !!(aspect() & gl::image_aspect::depth);
-	old_contents.init_transfer(this);
+		section.init_transfer(this);
 
 		if (state_flags & rsx::surface_state_flags::erase_bkgnd)
 		{
-		const auto area = old_contents.dst_rect();
+			const auto area = section.dst_rect();
 			if (area.x1 > 0 || area.y1 > 0 || unsigned(area.x2) < width() || unsigned(area.y2) < height())
 			{
 				clear_surface_impl();
@@ -671,10 +665,11 @@ void gl::render_target::memory_barrier(gl::command_context& cmd, bool force_init
 			}
 		}
 
-	gl::g_hw_blitter->scale_image(cmd, old_contents.source, this,
+		gl::g_hw_blitter->scale_image(cmd, section.source, this,
-		old_contents.src_rect(),
+			section.src_rect(),
-		old_contents.dst_rect(),
+			section.dst_rect(),
 			!dst_is_depth, dst_is_depth, typeless_info);
+	}
 
 	// Memory has been transferred, discard old contents and update memory flags
 	// TODO: Preserve memory outside surface clip region

rpcs3/Emu/RSX/GL/GLRenderTargets.h

@@ -269,7 +269,7 @@ struct gl_render_target_traits
 	static
 	void notify_surface_invalidated(const std::unique_ptr<gl::render_target>& surface)
 	{
-		if (surface->old_contents)
+		if (!surface->old_contents.empty())
 		{
 			// TODO: Retire the deferred writes
 			surface->clear_rw_barrier();

rpcs3/Emu/RSX/VK/VKGSRender.cpp

@@ -1212,8 +1212,8 @@ void VKGSRender::end()
 	// Check for data casts
 	// NOTE: This is deprecated and will be removed soon. The memory barrier invoked before rendering does this better
 	auto ds = std::get<1>(m_rtts.m_bound_depth_stencil);
-	if (ds && ds->old_contents &&
+	if (ds && ds->old_contents.size() == 1 &&
-		ds->old_contents.source->info.format == VK_FORMAT_B8G8R8A8_UNORM)
+		ds->old_contents[0].source->info.format == VK_FORMAT_B8G8R8A8_UNORM)
 	{
 		auto key = vk::get_renderpass_key(ds->info.format);
 		auto render_pass = vk::get_renderpass(*m_device, key);
@@ -1223,7 +1223,7 @@ void VKGSRender::end()
 		VkImageSubresourceRange range = { VK_IMAGE_ASPECT_DEPTH_BIT | VK_IMAGE_ASPECT_STENCIL_BIT, 0, 1, 0, 1 };
 
 		// Initialize source
-		auto src = vk::as_rtt(ds->old_contents.source);
+		auto src = vk::as_rtt(ds->old_contents[0].source);
 		src->read_barrier(*m_current_command_buffer);
 
 		switch (src->current_layout)
@@ -1244,10 +1244,10 @@ void VKGSRender::end()
 		if (!preinitialized) ds->pop_layout(*m_current_command_buffer);
 
 		// TODO: Stencil transfer
-		ds->old_contents.init_transfer(ds);
+		ds->old_contents[0].init_transfer(ds);
 		m_depth_converter->run(*m_current_command_buffer,
-			ds->old_contents.src_rect(),
+			ds->old_contents[0].src_rect(),
-			ds->old_contents.dst_rect(),
+			ds->old_contents[0].dst_rect(),
 			src->get_view(0xAAE4, rsx::default_remap_vector),
 			ds, render_pass);
 

rpcs3/Emu/RSX/VK/VKRenderTargets.h

@@ -281,13 +281,7 @@ namespace vk
 				get_resolve_target();
 			}
 
-			if (old_contents && !rsx::pitch_compatible(this, static_cast<vk::render_target*>(old_contents.source)))
+			if (LIKELY(old_contents.empty()))
-			{
-				LOG_TRACE(RSX, "Pitch mismatch, could not transfer inherited memory");
-				clear_rw_barrier();
-			}
-
-			if (LIKELY(!old_contents))
 			{
 				if (state_flags & rsx::surface_state_flags::erase_bkgnd)
 				{
@@ -321,7 +315,10 @@ namespace vk
 				return;
 			}
 
-			auto src_texture = static_cast<vk::render_target*>(old_contents.source);
+			bool optimize_copy = true;
+			for (auto &section : old_contents)
+			{
+				auto src_texture = static_cast<vk::render_target*>(section.source);
 				src_texture->read_barrier(cmd);
 
 				const auto src_bpp = src_texture->get_bpp();
@@ -346,10 +343,10 @@ namespace vk
 				}
 
 				vk::blitter hw_blitter;
-			old_contents.init_transfer(this);
+				section.init_transfer(this);
 
-			auto src_area = old_contents.src_rect();
+				auto src_area = section.src_rect();
-			auto dst_area = old_contents.dst_rect();
+				auto dst_area = section.dst_rect();
 
 				if (g_cfg.video.antialiasing_level != msaa_level::none)
 				{
@@ -390,7 +387,10 @@ namespace vk
 					dst_area,
 					/*linear?*/false, /*depth?(unused)*/false, typeless_info);
 
-			on_write_copy(0, !memory_load);
+				optimize_copy = optimize_copy && !memory_load;
+			}
+
+			on_write_copy(0, optimize_copy);
 
 			if (!read_access && samples() > 1)
 			{
@@ -639,7 +639,7 @@ namespace rsx
 			surface->frame_tag = vk::get_current_frame_id();
 			if (!surface->frame_tag) surface->frame_tag = 1;
 
-			if (surface->old_contents)
+			if (!surface->old_contents.empty())
 			{
 				// TODO: Retire the deferred writes
 				surface->clear_rw_barrier();

rpcs3/Emu/RSX/rsx_utils.h

@@ -1,4 +1,4 @@
-#pragma once
+#pragma once
 
 #include "../System.h"
 #include "Utilities/address_range.h"
@@ -606,7 +606,7 @@ namespace rsx
 	template <typename SurfaceType>
 	std::tuple<u16, u16, u16, u16> get_transferable_region(const SurfaceType* surface)
 	{
-		auto src = static_cast<const SurfaceType*>(surface->old_contents.source);
+		auto src = static_cast<const SurfaceType*>(surface->old_contents[0].source);
 		auto area1 = src->get_normalized_memory_area();
 		auto area2 = surface->get_normalized_memory_area();