rsx: Speed up slice gathering for large datasets

- We can cut down by an order of magnitude the amount of checks we do to construct objects

rpcs3/Emu/RSX/Common/texture_cache_helpers.h

@@ -262,18 +262,24 @@ namespace rsx
 				u64 tag;   // Timestamp
 				u32 list;  // List source, 0 = fbo, 1 = local
 				u32 index; // Index in list
+
+				utils::address_range32 bounds;
 			};
 
-			std::vector<sort_helper> sort_list;
+			const u32 available_slices = fbos.size() + local.size();
+			bool unordered_list = false;
+
+			rsx::simple_array<sort_helper> sort_list;
+			rsx::simple_array<utils::address_range32> sort_ranges;
+			sort_list.reserve(available_slices);
+			sort_ranges.reserve(available_slices);
 
-			if (!fbos.empty() && !local.empty())
-			{
 			// Generate sorting tree if both resources are available and overlapping
-				sort_list.reserve(fbos.size() + local.size());
-
 			for (u32 index = 0; index < fbos.size(); ++index)
 			{
-					sort_list.push_back({ fbos[index].surface->last_use_tag, 0, index });
+				const auto range = fbos[index].surface->get_memory_range();
+				sort_ranges.push_back(range);
+				sort_list.push_back({ fbos[index].surface->last_use_tag, 0, index, range });
 			}
 
 			for (u32 index = 0; index < local.size(); ++index)
@@ -281,13 +287,32 @@ namespace rsx
 				if (local[index]->get_context() != rsx::texture_upload_context::blit_engine_dst)
 					continue;
 
-					sort_list.push_back({ local[index]->last_write_tag, 1, index });
+				const auto range = local[index]->get_section_range();
+				sort_ranges.push_back(range);
+				sort_list.push_back({ local[index]->last_write_tag, 1, index, range });
 			}
 
-				std::sort(sort_list.begin(), sort_list.end(), FN(x.tag < y.tag));
+			if (!fbos.empty() && !local.empty())
+			{
+				sort_list.sort(FN(x.tag < y.tag));
 			}
 
-			auto add_rtt_resource = [&](auto& section, u16 slice)
+			// Check if ordered
+			for (u32 i = 0; i < sort_list.size(); ++i)
+			{
+				if (i == 0)
+				{
+					continue;
+				}
+
+				if (sort_ranges[i].start < sort_ranges[i - 1].end)
+				{
+					unordered_list = true;
+					break;
+				}
+			}
+
+			auto add_rtt_resource = [&](auto& section, u16 slice) -> std::pair<bool, bool> // [ input fully consumed, output fully covered ]
 			{
 				const u32 slice_begin = (slice * attr.slice_h);
 				const u32 slice_end = (slice_begin + attr.height);
@@ -296,7 +321,7 @@ namespace rsx
 				if (section.dst_area.y >= slice_end || section_end <= slice_begin)
 				{
 					// Belongs to a different slice
-					return;
+					return { section_end <= slice_begin, false };
 				}
 
 				// How much of this slice to read?
@@ -346,9 +371,11 @@ namespace rsx
 					.dst_w = dst_width,
 					.dst_h = dst_height
 				});
+
+				return { section_end <= slice_end, section_end >= slice_end };
 			};
 
-			auto add_local_resource = [&](auto& section, u32 address, u16 slice, bool scaling = true)
+			auto add_local_resource = [&](auto& section, u32 address, u16 slice, bool scaling = true) -> std::pair<bool, bool> // [ input fully consumed, output fully covered ]
 			{
 				// Intersect this resource with the original one.
 				// Note that intersection takes place in a normalized coordinate space (bpp = 1)
@@ -364,7 +391,7 @@ namespace rsx
 				if (!dimensions.width || !dimensions.height)
 				{
 					// Out of bounds, invalid intersection
-					return;
+					return { false, false };
 				}
 
 				// The intersection takes place in a normalized coordinate space. Now we convert back to domain-specific
@@ -383,7 +410,7 @@ namespace rsx
 				if (dst_y >= dst_slice_end || write_section_end <= dst_slice_begin)
 				{
 					// Belongs to a different slice
-					return;
+					return { write_section_end <= dst_slice_begin, false };
 				}
 
 				const u16 dst_w = static_cast<u16>(dst_size.width);
@@ -411,9 +438,10 @@ namespace rsx
 						.dst_w = _dst_w,
 						.dst_h = _dst_h
 					});
+
+					return { write_section_end <= dst_slice_end, write_section_end >= dst_slice_end };
 				}
-				else
+
-				{
 				out.push_back
 				({
 					.src = section->get_raw_texture(),
@@ -429,7 +457,8 @@ namespace rsx
 					.dst_w = dst_w,
 					.dst_h = height
 				});
-				}
+
+				return { write_section_end <= dst_slice_end, write_section_end >= dst_slice_end };
 			};
 
 			u32 current_address = attr.address;
@@ -442,34 +471,36 @@ namespace rsx
 
 			for (u16 slice = 0; slice < count; ++slice)
 			{
-				auto num_surface = out.size();
+				const auto num_surface = out.size();
+				const auto slice_range = utils::address_range32::start_length(current_address, slice_size);
 
-				if (local.empty()) [[likely]]
+				for (auto& e : sort_list)
 				{
-					for (auto& section : fbos)
+					if (e.index == umax || !slice_range.overlaps(e.bounds))
 					{
-						add_rtt_resource(section, slice);
+						continue;
 					}
-				}
+
-				else if (fbos.empty())
+					bool remove = false, slice_complete = false;
-				{
-					for (auto& section : local)
-					{
-						add_local_resource(section, current_address, slice, false);
-					}
-				}
-				else
-				{
-					for (const auto& e : sort_list)
-					{
 					if (e.list == 0)
 					{
-							add_rtt_resource(fbos[e.index], slice);
+						std::tie(remove, slice_complete) = add_rtt_resource(fbos[e.index], slice);
 					}
 					else
 					{
-							add_local_resource(local[e.index], current_address, slice);
+						std::tie(remove, slice_complete) = add_local_resource(local[e.index], current_address, slice);
 					}
+
+					if (remove)
+					{
+						// If we got here, the section has been fully ingested by the current slice and will never match again.
+						e.index = umax;
+					}
+
+					if (slice_complete && !unordered_list)
+					{
+						// Reached the end of the current slice and we are guaranteed to not match any other section since they lie after this one in memory.
+						break;
 					}
 				}