rsx: Clean up surface cache routines around RTT invalidate

rpcs3/Emu/RSX/Common/surface_store.h

@@ -1,6 +1,7 @@
 #pragma once
 
 #include "surface_utils.h"
+#include "simple_array.hpp"
 #include "../gcm_enums.h"
 #include "../rsx_utils.h"
 #include <list>
@@ -56,10 +57,9 @@ namespace rsx
 		rsx::surface_raster_type m_active_raster_type = rsx::surface_raster_type::linear;
 
 	public:
-		std::pair<u8, u8> m_bound_render_targets_config = {};
+		rsx::simple_array<u8> m_bound_render_target_ids = {};
 		std::array<std::pair<u32, surface_type>, 4> m_bound_render_targets = {};
 		std::pair<u32, surface_type> m_bound_depth_stencil = {};
-		u8 m_bound_buffers_count = 0;
 
 		// List of sections derived from a section that has been split and invalidated
 		std::vector<surface_type> orphaned_surfaces;
@@ -840,23 +840,19 @@ namespace rsx
 			cache_tag = rsx::get_shared_tag();
 			m_invalidate_on_write = (antialias != rsx::surface_antialiasing::center_1_sample);
 			m_active_raster_type = raster_type;
-			m_bound_buffers_count = 0;
 
 			// Make previous RTTs sampleable
-			for (int i = m_bound_render_targets_config.first, count = 0;
-				count < m_bound_render_targets_config.second;
-				++i, ++count)
+			for (const auto& i : m_bound_render_target_ids)
 			{
 				auto &rtt = m_bound_render_targets[i];
 				Traits::prepare_surface_for_sampling(command_list, std::get<1>(rtt));
 				rtt = std::make_pair(0, nullptr);
 			}
 
-			const auto rtt_indices = utility::get_rtt_indexes(set_surface_target);
-			if (!rtt_indices.empty()) [[likely]]
+			m_bound_render_target_ids.clear();
+			if (const auto rtt_indices = utility::get_rtt_indexes(set_surface_target);
+				!rtt_indices.empty()) [[likely]]
 			{
-				m_bound_render_targets_config = { rtt_indices.front(), 0 };
-
 				// Create/Reuse requested rtts
 				for (u8 surface_index : rtt_indices)
 				{
@@ -867,14 +863,9 @@ namespace rsx
 						bind_address_as_render_targets(command_list, surface_addresses[surface_index], color_format, antialias,
 							clip_width, clip_height, surface_pitch[surface_index], std::forward<Args>(extra_params)...));
 
-					++m_bound_render_targets_config.second;
-					++m_bound_buffers_count;
+					m_bound_render_target_ids.push_back(surface_index);
 				}
 			}
-			else
-			{
-				m_bound_render_targets_config = { 0, 0 };
-			}
 
 			// Same for depth buffer
 			if (std::get<1>(m_bound_depth_stencil) != nullptr)
@@ -887,8 +878,6 @@ namespace rsx
 				m_bound_depth_stencil = std::make_pair(address_z,
 					bind_address_as_depth_stencil(command_list, address_z, depth_format, antialias,
 						clip_width, clip_height, zeta_pitch, std::forward<Args>(extra_params)...));
-
-				++m_bound_buffers_count;
 			}
 			else
 			{
@@ -898,7 +887,7 @@ namespace rsx
 
 		u8 get_color_surface_count() const
 		{
-			return m_bound_render_targets_config.second;
+			return static_cast<u8>(m_bound_render_target_ids.size());
 		}
 
 		surface_type get_surface_at(u32 address)
@@ -947,14 +936,15 @@ namespace rsx
 			}
 		}
 
-		bool address_is_bound(u32 address) const
+		inline bool address_is_bound(u32 address) const
 		{
-			for (int i = m_bound_render_targets_config.first, count = 0;
-				count < m_bound_render_targets_config.second;
-				++i, ++count)
+			ensure(address);
+			for (int i = 0; i < 4; ++i)
 			{
 				if (m_bound_render_targets[i].first == address)
+				{
 					return true;
+				}
 			}
 
 			return (m_bound_depth_stencil.first == address);
@@ -1127,21 +1117,33 @@ namespace rsx
 			// TODO: Take WCB/WDB into account. Should speed this up a bit by skipping sync_tag calls
 			write_tag = rsx::get_shared_tag();
 
-			for (u8 i = m_bound_render_targets_config.first, count = 0;
-				count < m_bound_render_targets_config.second;
-				++count, ++i)
+			for (const auto& i : m_bound_render_target_ids)
 			{
-				if (auto surface = m_bound_render_targets[i].second;
-					surface && color_mrt_writes_enabled[i])
+				if (color_mrt_writes_enabled[i])
 				{
-					surface->on_write(write_tag);
+					auto surface = m_bound_render_targets[i].second;
+					if (surface->last_use_tag > cache_tag) [[ likely ]]
+					{
+						surface->on_write_fast(write_tag);
+					}
+					else
+					{
+						surface->on_write(write_tag, rsx::surface_state_flags::require_resolve, m_active_raster_type);
+					}
 				}
 			}
 
 			if (auto zsurface = m_bound_depth_stencil.second;
 				zsurface && depth_stencil_writes_enabled)
 			{
-				zsurface->on_write(write_tag);
+				if (zsurface->last_use_tag > cache_tag) [[ likely ]]
+				{
+					zsurface->on_write_fast(write_tag);
+				}
+				else
+				{
+					zsurface->on_write(write_tag, rsx::surface_state_flags::require_resolve, m_active_raster_type);
+				}
 			}
 		}
 
@@ -1164,7 +1166,7 @@ namespace rsx
 			ensure(m_active_memory_used == 0);
 
 			m_bound_depth_stencil = std::make_pair(0, nullptr);
-			m_bound_render_targets_config = { 0, 0 };
+			m_bound_render_target_ids.clear();
 			for (auto &rtt : m_bound_render_targets)
 			{
 				rtt = std::make_pair(0, nullptr);

rpcs3/Emu/RSX/Common/surface_utils.h

@@ -623,11 +623,14 @@ namespace rsx
 			}
 		}
 
-		void on_invalidate_children()
+		inline void on_write_fast(u64 write_tag)
 		{
+			ensure(write_tag);
+			last_use_tag = write_tag;
+
 			if (resolve_surface)
 			{
-				msaa_flags = rsx::surface_state_flags::require_resolve;
+				msaa_flags |= rsx::surface_state_flags::require_resolve;
 			}
 		}
 

rpcs3/Emu/RSX/GL/GLGSRender.cpp

@@ -644,16 +644,15 @@ void GLGSRender::clear_surface(u32 arg)
 			gl_state.clear_color(clear_r, clear_g, clear_b, clear_a);
 			mask |= GLenum(gl::buffers::color);
 
-			for (u8 index = m_rtts.m_bound_render_targets_config.first, count = 0;
-				 count < m_rtts.m_bound_render_targets_config.second;
-				 ++count, ++index)
+			int hw_index = 0;
+			for (const auto& index : m_rtts.m_bound_render_target_ids)
 			{
 				if (!full_frame)
 				{
 					m_rtts.m_bound_render_targets[index].second->write_barrier(cmd);
 				}
 
-				gl_state.color_maski(count, colormask);
+				gl_state.color_maski(hw_index++, colormask);
 			}
 
 			update_color = true;

rpcs3/Emu/RSX/VK/VKGSRender.cpp

@@ -1382,9 +1382,7 @@ void VKGSRender::clear_surface(u32 mask)
 				if (!use_fast_clear || !full_frame)
 				{
 					// If we're not clobber all the memory, a barrier is required
-					for (u8 index = m_rtts.m_bound_render_targets_config.first, count = 0;
-						count < m_rtts.m_bound_render_targets_config.second;
-						++count, ++index)
+					for (const auto& index : m_rtts.m_bound_render_target_ids)
 					{
 						m_rtts.m_bound_render_targets[index].second->write_barrier(*m_current_command_buffer);
 					}