perf: optimize and benchmark terminal frame damage detection

Extract frame damage detection into a private reusable implementation and add release benchmarks for representative GUI damage patterns at 720p, 1080p, and 4K. Use tile-major traversal for smaller framebuffers and cache-friendly row-major traversal above 16 MiB, with reusable scratch storage and no whole-frame equality precheck. Determine full-frame updates from the completed damage bounds instead of prematurely falling back based on dirty-tile area. Present benchmark results as a table showing each strategy, the selected production path, its gap to the fastest candidate, and resulting transfer size.
2026-09-22 13:01:05 +03:00 · 2026-09-10 12:03:15 -03:00
parent fcc3a412fb
commit 6ccb433dd5
5 changed files with 425 additions and 55 deletions
--- a/src/benchmarks/terminal_frame_damage_benchmark.cpp
+++ b/src/benchmarks/terminal_frame_damage_benchmark.cpp
@@ -0,0 +1,292 @@
+#include "../tests/unit_tests/utest.hpp"
+
+#include "../eepp/window/terminal/framedamage.hpp"
+
+#include <eepp/core/string.hpp>
+#include <eepp/system/clock.hpp>
+#include <tabulate/tabulate.hpp>
+
+#include <algorithm>
+#include <array>
+#include <cstring>
+#include <functional>
+#include <vector>
+
+using namespace EE;
+using namespace EE::System;
+using namespace EE::Window::Private;
+
+namespace {
+
+struct DamageRectangle {
+	Int32 x;
+	Int32 y;
+	Int32 width;
+	Int32 height;
+};
+
+struct BenchmarkResult {
+	double microsecondsPerFrame;
+	size_t transferPixels;
+	Uint64 checksum;
+};
+
+using DamageFunction =
+	std::function<FrameDamageResult( const std::vector<Uint8>&, const std::vector<Uint8>&,
+									 const Math::Sizei&, std::vector<DamageRectangle>& )>;
+
+FrameDamageResult findFrameDamageWithEqualityCheck( const std::vector<Uint8>& current,
+													const std::vector<Uint8>& previous,
+													const Math::Sizei& size,
+													std::vector<DamageRectangle>& rectangles ) {
+	if ( 0 == std::memcmp( current.data(), previous.data(), current.size() ) ) {
+		rectangles.clear();
+		return FrameDamageResult::None;
+	}
+	return findFrameDamage( current, previous, size, rectangles );
+}
+
+FrameDamageResult findFrameDamageRowMajor( const std::vector<Uint8>& current,
+										   const std::vector<Uint8>& previous,
+										   const Math::Sizei& size,
+										   std::vector<DamageRectangle>& rectangles,
+										   std::vector<Uint8>& changedTiles ) {
+	rectangles.clear();
+	const Int32 tilesX = ( size.x + FrameDamageTileSize - 1 ) / FrameDamageTileSize;
+	const Int32 tilesY = ( size.y + FrameDamageTileSize - 1 ) / FrameDamageTileSize;
+	const size_t rowBytes = static_cast<size_t>( size.x ) * 3;
+	Int32 left = size.x;
+	Int32 top = size.y;
+	Int32 right = 0;
+	Int32 bottom = 0;
+	changedTiles.resize( static_cast<size_t>( tilesX ) );
+
+	for ( Int32 tileY = 0; tileY < tilesY; ++tileY ) {
+		const Int32 y = tileY * FrameDamageTileSize;
+		const Int32 height = eemin( FrameDamageTileSize, size.y - y );
+		std::fill( changedTiles.begin(), changedTiles.end(), 0 );
+		Int32 remainingTiles = tilesX;
+		for ( Int32 row = 0; row < height && remainingTiles; ++row ) {
+			const Int32 sourceRow = size.y - 1 - ( y + row );
+			const Uint8* currentRow = current.data() + static_cast<size_t>( sourceRow ) * rowBytes;
+			const Uint8* previousRow =
+				previous.data() + static_cast<size_t>( sourceRow ) * rowBytes;
+			for ( Int32 tileX = 0; tileX < tilesX; ++tileX ) {
+				if ( changedTiles[static_cast<size_t>( tileX )] )
+					continue;
+				const Int32 x = tileX * FrameDamageTileSize;
+				const Int32 width = eemin( FrameDamageTileSize, size.x - x );
+				const size_t offset = static_cast<size_t>( x ) * 3;
+				if ( 0 != std::memcmp( currentRow + offset, previousRow + offset,
+									   static_cast<size_t>( width ) * 3 ) ) {
+					changedTiles[static_cast<size_t>( tileX )] = 1;
+					--remainingTiles;
+					left = eemin( left, x );
+					top = eemin( top, y );
+					right = eemax( right, x + width );
+					bottom = eemax( bottom, y + height );
+				}
+			}
+		}
+	}
+
+	if ( left < right && top < bottom ) {
+		rectangles.push_back( { left, top, right - left, bottom - top } );
+		return left == 0 && top == 0 && right == size.x && bottom == size.y
+				   ? FrameDamageResult::Full
+				   : FrameDamageResult::Rectangle;
+	}
+	return FrameDamageResult::None;
+}
+
+void changeRectangle( std::vector<Uint8>& pixels, const Math::Sizei& size, Int32 x, Int32 y,
+					  Int32 width, Int32 height ) {
+	const size_t rowBytes = static_cast<size_t>( size.x ) * 3;
+	for ( Int32 row = y; row < y + height; ++row ) {
+		Uint8* pixel =
+			pixels.data() + static_cast<size_t>( row ) * rowBytes + static_cast<size_t>( x ) * 3;
+		for ( Int32 column = 0; column < width * 3; ++column )
+			pixel[column] ^= static_cast<Uint8>( 0x5A + ( column & 7 ) );
+	}
+}
+
+void changeRandomTiles( std::vector<Uint8>& pixels, const Math::Sizei& size, Int32 percentage ) {
+	Uint32 state = 0x12345678;
+	const Int32 tilesX = ( size.x + FrameDamageTileSize - 1 ) / FrameDamageTileSize;
+	const Int32 tilesY = ( size.y + FrameDamageTileSize - 1 ) / FrameDamageTileSize;
+	for ( Int32 tileY = 0; tileY < tilesY; ++tileY ) {
+		for ( Int32 tileX = 0; tileX < tilesX; ++tileX ) {
+			state ^= state << 13;
+			state ^= state >> 17;
+			state ^= state << 5;
+			if ( static_cast<Int32>( state % 100 ) >= percentage )
+				continue;
+			const Int32 x = tileX * FrameDamageTileSize;
+			const Int32 y = tileY * FrameDamageTileSize;
+			changeRectangle( pixels, size, x, y, eemin( FrameDamageTileSize, size.x - x ),
+							 eemin( FrameDamageTileSize, size.y - y ) );
+		}
+	}
+}
+
+BenchmarkResult runBenchmark( const DamageFunction& function, const std::vector<Uint8>& current,
+							  const std::vector<Uint8>& previous, const Math::Sizei& size,
+							  Int32 iterations ) {
+	static constexpr size_t SampleCount = 5;
+	std::vector<DamageRectangle> rectangles;
+	FrameDamageResult result = FrameDamageResult::None;
+	Uint64 checksum = 0;
+	function( current, previous, size, rectangles );
+	std::array<double, SampleCount> samples;
+	for ( double& sample : samples ) {
+		Clock clock;
+		for ( Int32 iteration = 0; iteration < iterations; ++iteration ) {
+			result = function( current, previous, size, rectangles );
+			checksum += 1 + static_cast<Uint8>( result ) + rectangles.size();
+		}
+		sample = static_cast<double>( clock.getElapsedTime().asMicroseconds() ) / iterations;
+	}
+	std::sort( samples.begin(), samples.end() );
+	size_t transferPixels = 0;
+	if ( result == FrameDamageResult::Full ) {
+		transferPixels = static_cast<size_t>( size.x ) * size.y;
+	} else {
+		for ( const DamageRectangle& rectangle : rectangles )
+			transferPixels += static_cast<size_t>( rectangle.width ) * rectangle.height;
+	}
+	return { samples[SampleCount / 2], transferPixels, checksum };
+}
+
+} // namespace
+
+UTEST( Benchmark, TerminalFrameDamage ) {
+	struct Resolution {
+		Math::Sizei size;
+		Int32 iterations;
+	};
+	struct Scenario {
+		const char* name;
+		std::function<void( std::vector<Uint8>&, const Math::Sizei& )> change;
+	};
+
+	const std::vector<Resolution> resolutions = {
+		{ { 1280, 720 }, 150 }, { { 1920, 1080 }, 80 }, { { 3840, 2160 }, 20 } };
+	const std::vector<Scenario> scenarios = {
+		{ "identical", []( auto&, const auto& ) {} },
+		{ "caret",
+		  []( auto& pixels, const auto& size ) {
+			  changeRectangle( pixels, size, size.x / 2, size.y / 2, 3, 24 );
+		  } },
+		{ "full-width-row",
+		  []( auto& pixels, const auto& size ) {
+			  changeRectangle( pixels, size, 0, size.y / 2, size.x, 32 );
+		  } },
+		{ "distant-corners",
+		  []( auto& pixels, const auto& size ) {
+			  changeRectangle( pixels, size, 0, 0, 32, 32 );
+			  changeRectangle( pixels, size, size.x - 32, size.y - 32, 32, 32 );
+		  } },
+		{ "scroll-region",
+		  []( auto& pixels, const auto& size ) {
+			  changeRectangle( pixels, size, size.x / 5, size.y / 5, size.x * 3 / 5,
+							   size.y * 3 / 5 );
+		  } },
+		{ "random-10pct",
+		  []( auto& pixels, const auto& size ) { changeRandomTiles( pixels, size, 10 ); } },
+		{ "random-50pct",
+		  []( auto& pixels, const auto& size ) { changeRandomTiles( pixels, size, 50 ); } },
+		{ "random-100pct",
+		  []( auto& pixels, const auto& size ) { changeRandomTiles( pixels, size, 100 ); } },
+	};
+
+	std::vector<Uint8> changedTiles;
+	const DamageFunction currentAlgorithm = []( const auto& current, const auto& previous,
+												const auto& size, auto& rectangles ) {
+		return findFrameDamage( current, previous, size, rectangles );
+	};
+	const DamageFunction equalityFastPath = findFrameDamageWithEqualityCheck;
+	const DamageFunction rowMajor = [&changedTiles]( const auto& current, const auto& previous,
+													 const auto& size, auto& rectangles ) {
+		return findFrameDamageRowMajor( current, previous, size, rectangles, changedTiles );
+	};
+	const DamageFunction equalityRowMajor = [&changedTiles]( const auto& current,
+															 const auto& previous, const auto& size,
+															 auto& rectangles ) {
+		if ( 0 == std::memcmp( current.data(), previous.data(), current.size() ) ) {
+			rectangles.clear();
+			return FrameDamageResult::None;
+		}
+		return findFrameDamageRowMajor( current, previous, size, rectangles, changedTiles );
+	};
+	const DamageFunction adaptive = [&changedTiles]( const auto& current, const auto& previous,
+													 const auto& size, auto& rectangles ) {
+		return findFrameDamageAdaptive( current, previous, size, rectangles, changedTiles );
+	};
+
+	tabulate::Table table;
+	table.add_row( { "Resolution", "Scenario", "Tile-major us", "memcmp + tile us", "Row-major us",
+					 "memcmp + rows us", "Production", "Production us", "Gap to fastest",
+					 "Transfer MiB" } );
+	for ( size_t column = 0; column < table[0].size(); ++column ) {
+		table[0][column]
+			.format()
+			.font_align( tabulate::FontAlign::center )
+			.font_style( { tabulate::FontStyle::bold } );
+	}
+	for ( const Resolution& resolution : resolutions ) {
+		const size_t frameBytes = static_cast<size_t>( resolution.size.x ) * resolution.size.y * 3;
+		std::vector<Uint8> previous( frameBytes, 0x35 );
+		for ( const Scenario& scenario : scenarios ) {
+			std::vector<Uint8> current = previous;
+			scenario.change( current, resolution.size );
+			std::vector<DamageRectangle> expectedRectangles;
+			std::vector<DamageRectangle> candidateRectangles;
+			const FrameDamageResult expected =
+				currentAlgorithm( current, previous, resolution.size, expectedRectangles );
+			for ( const DamageFunction* candidate :
+				  { &equalityFastPath, &rowMajor, &equalityRowMajor, &adaptive } ) {
+				const FrameDamageResult actual =
+					( *candidate )( current, previous, resolution.size, candidateRectangles );
+				EXPECT_EQ( static_cast<int>( expected ), static_cast<int>( actual ) );
+				EXPECT_EQ( expectedRectangles.size(), candidateRectangles.size() );
+				if ( !expectedRectangles.empty() ) {
+					EXPECT_EQ( expectedRectangles[0].x, candidateRectangles[0].x );
+					EXPECT_EQ( expectedRectangles[0].y, candidateRectangles[0].y );
+					EXPECT_EQ( expectedRectangles[0].width, candidateRectangles[0].width );
+					EXPECT_EQ( expectedRectangles[0].height, candidateRectangles[0].height );
+				}
+			}
+			const BenchmarkResult baseline = runBenchmark( currentAlgorithm, current, previous,
+														   resolution.size, resolution.iterations );
+			const BenchmarkResult equality = runBenchmark( equalityFastPath, current, previous,
+														   resolution.size, resolution.iterations );
+			const BenchmarkResult rows =
+				runBenchmark( rowMajor, current, previous, resolution.size, resolution.iterations );
+			const BenchmarkResult equalityRows = runBenchmark(
+				equalityRowMajor, current, previous, resolution.size, resolution.iterations );
+			const bool productionUsesRows = frameBytes >= FrameDamageRowMajorThreshold;
+			const double productionTime =
+				productionUsesRows ? rows.microsecondsPerFrame : baseline.microsecondsPerFrame;
+			const double fastestTime =
+				eemin( eemin( baseline.microsecondsPerFrame, equality.microsecondsPerFrame ),
+					   eemin( rows.microsecondsPerFrame, equalityRows.microsecondsPerFrame ) );
+			const double gapToFastest = ( productionTime / fastestTime - 1.0 ) * 100.0;
+			const double transferMiB =
+				static_cast<double>( baseline.transferPixels * 3 ) / ( 1024.0 * 1024.0 );
+			table.add_row( { String::format( "%dx%d", resolution.size.x, resolution.size.y ),
+							 scenario.name, String::format( "%.2f", baseline.microsecondsPerFrame ),
+							 String::format( "%.2f", equality.microsecondsPerFrame ),
+							 String::format( "%.2f", rows.microsecondsPerFrame ),
+							 String::format( "%.2f", equalityRows.microsecondsPerFrame ),
+							 productionUsesRows ? "row-major" : "tile-major",
+							 String::format( "%.2f", productionTime ),
+							 String::format( "%.1f%%", gapToFastest ),
+							 String::format( "%.2f", transferMiB ) } );
+			EXPECT_GT(
+				baseline.checksum + equality.checksum + rows.checksum + equalityRows.checksum, 0u );
+		}
+	}
+	for ( size_t column = 2; column < table[0].size(); ++column )
+		table.column( column ).format().font_align( tabulate::FontAlign::right );
+	UTEST_PRINT_INFO( ( "\n" + table.str() ).c_str() );
+}
--- a/src/eepp/graphics/textureatlasloader.cpp
+++ b/src/eepp/graphics/textureatlasloader.cpp
@@ -156,6 +156,7 @@ void TextureAtlasLoader::loadFromStream( IOStream& IOS ) {

 				std::string name( &tTextureHdr.Name[0] );
 				std::string path( FileSystem::fileRemoveFileName( mTextureAtlasPath ) + name );
+				std::string completePath( path );
 				FileSystem::filePathRemoveProcessPath( path );

 				//! Checks if the texture is already loaded
@@ -173,8 +174,10 @@ void TextureAtlasLoader::loadFromStream( IOStream& IOS ) {
 							mTempAtlass[textureIndex].LoadedTexture = std::move( texture );
 						} );
 					} else {
-						mRL.add( [this, textureIndex, path = std::move( path )] {
-							TexturePtr texture = TextureFactory::instance()->loadFromFile( path );
+						mRL.add( [this, textureIndex, path = std::move( path ),
+								  completePath = std::move( completePath )] {
+							TexturePtr texture =
+								TextureFactory::instance()->loadFromFile( completePath );
 							if ( texture )
 								mResourceScope->publishLocal( path, texture );
 							mTempAtlass[textureIndex].LoadedTexture = std::move( texture );
--- a/src/eepp/window/kittyframepresenter.cpp
+++ b/src/eepp/window/kittyframepresenter.cpp
@@ -6,6 +6,8 @@
 #include <eepp/window/terminal/terminalruntime.hpp>
 #include <eepp/window/window.hpp>

+#include "terminal/framedamage.hpp"
+
 #include <array>
 #include <cstdio>
 #include <cstdlib>
@@ -17,11 +19,9 @@ namespace EE { namespace Window {

 namespace {

-constexpr Int32 TileSize = 32;
 constexpr size_t CompressionThreshold = 4096;
 constexpr Uint32 StreamImageId = 0x45455050;
 constexpr Uint32 StreamPlacementId = 1;
-enum class DamageResult : Uint8 { None, Rectangle, Full };

 bool environmentFlag( const char* name, bool defaultValue ) {
 	const char* value = std::getenv( name );
@@ -53,53 +53,6 @@ int environmentCompressionLevel() {
 	return end != value && *end == '\0' && level >= 0 && level <= 9 ? static_cast<int>( level ) : 1;
 }

-template <typename Rectangle>
-DamageResult findDamage( const std::vector<Uint8>& current, const std::vector<Uint8>& previous,
-						 const Math::Sizei& size, std::vector<Rectangle>& rectangles ) {
-	rectangles.clear();
-	const Int32 tilesX = ( size.x + TileSize - 1 ) / TileSize;
-	const Int32 tilesY = ( size.y + TileSize - 1 ) / TileSize;
-	const size_t rowBytes = static_cast<size_t>( size.x ) * 3;
-	const size_t totalPixels = static_cast<size_t>( size.x ) * size.y;
-	size_t changedPixels = 0;
-	Int32 left = size.x;
-	Int32 top = size.y;
-	Int32 right = 0;
-	Int32 bottom = 0;
-
-	for ( Int32 tileY = 0; tileY < tilesY; ++tileY ) {
-		const Int32 y = tileY * TileSize;
-		const Int32 height = eemin( TileSize, size.y - y );
-		for ( Int32 tileX = 0; tileX < tilesX; ++tileX ) {
-			bool changed = false;
-			const Int32 x = tileX * TileSize;
-			const Int32 width = eemin( TileSize, size.x - x );
-			for ( Int32 row = 0; row < height && !changed; ++row ) {
-				const Int32 sourceRow = size.y - 1 - ( y + row );
-				const size_t offset =
-					static_cast<size_t>( sourceRow ) * rowBytes + static_cast<size_t>( x ) * 3;
-				changed = 0 != std::memcmp( current.data() + offset, previous.data() + offset,
-											static_cast<size_t>( width ) * 3 );
-			}
-			if ( changed ) {
-				changedPixels += static_cast<size_t>( width ) * height;
-				left = eemin( left, x );
-				top = eemin( top, y );
-				right = eemax( right, x + width );
-				bottom = eemax( bottom, y + height );
-				if ( changedPixels * 5 >= totalPixels * 3 )
-					return DamageResult::Full;
-			}
-		}
-	}
-
-	if ( left < right && top < bottom ) {
-		rectangles.push_back( { left, top, right - left, bottom - top } );
-		return DamageResult::Rectangle;
-	}
-	return DamageResult::None;
-}
-
 bool compressPixels( const std::vector<Uint8>& input, std::vector<Uint8>& output, int level ) {
 	if ( level == 0 || input.size() < CompressionThreshold )
 		return false;
@@ -221,13 +174,13 @@ bool KittyFramePresenter::sendFrame( const Frame& frame ) {
 		return sendTransfer( mTransferPixels, full, true );
 	}

-	const DamageResult damage =
-		findDamage( frame.pixels, mPresented.pixels, frame.size, mDamageRectangles );
-	if ( damage == DamageResult::Full ) {
+	const Private::FrameDamageResult damage = Private::findFrameDamageAdaptive(
+		frame.pixels, mPresented.pixels, frame.size, mDamageRectangles, mDamageTiles );
+	if ( damage == Private::FrameDamageResult::Full ) {
 		extractRectangle( frame, full );
 		return sendTransfer( mTransferPixels, full, false );
 	}
-	if ( damage == DamageResult::None )
+	if ( damage == Private::FrameDamageResult::None )
 		return true;
 	for ( const DamageRectangle& rectangle : mDamageRectangles ) {
 		extractRectangle( frame, rectangle );
--- a/src/eepp/window/terminal/framedamage.hpp
+++ b/src/eepp/window/terminal/framedamage.hpp
@@ -0,0 +1,121 @@
+#ifndef EE_WINDOW_TERMINAL_FRAMEDAMAGE_HPP
+#define EE_WINDOW_TERMINAL_FRAMEDAMAGE_HPP
+
+#include <eepp/math/size.hpp>
+
+#include <cstring>
+#include <vector>
+
+namespace EE { namespace Window { namespace Private {
+
+constexpr Int32 FrameDamageTileSize = 32;
+constexpr size_t FrameDamageRowMajorThreshold = 16 * 1024 * 1024;
+
+enum class FrameDamageResult : Uint8 { None, Rectangle, Full };
+
+/** Finds the bounding rectangle of changed RGB24 tiles in bottom-up frame buffers. */
+template <typename Rectangle>
+FrameDamageResult findFrameDamage( const std::vector<Uint8>& current,
+								   const std::vector<Uint8>& previous, const Math::Sizei& size,
+								   std::vector<Rectangle>& rectangles ) {
+	rectangles.clear();
+	const Int32 tilesX = ( size.x + FrameDamageTileSize - 1 ) / FrameDamageTileSize;
+	const Int32 tilesY = ( size.y + FrameDamageTileSize - 1 ) / FrameDamageTileSize;
+	const size_t rowBytes = static_cast<size_t>( size.x ) * 3;
+	Int32 left = size.x;
+	Int32 top = size.y;
+	Int32 right = 0;
+	Int32 bottom = 0;
+
+	for ( Int32 tileY = 0; tileY < tilesY; ++tileY ) {
+		const Int32 y = tileY * FrameDamageTileSize;
+		const Int32 height = eemin( FrameDamageTileSize, size.y - y );
+		for ( Int32 tileX = 0; tileX < tilesX; ++tileX ) {
+			bool changed = false;
+			const Int32 x = tileX * FrameDamageTileSize;
+			const Int32 width = eemin( FrameDamageTileSize, size.x - x );
+			for ( Int32 row = 0; row < height && !changed; ++row ) {
+				const Int32 sourceRow = size.y - 1 - ( y + row );
+				const size_t offset =
+					static_cast<size_t>( sourceRow ) * rowBytes + static_cast<size_t>( x ) * 3;
+				changed = 0 != std::memcmp( current.data() + offset, previous.data() + offset,
+											static_cast<size_t>( width ) * 3 );
+			}
+			if ( changed ) {
+				left = eemin( left, x );
+				top = eemin( top, y );
+				right = eemax( right, x + width );
+				bottom = eemax( bottom, y + height );
+			}
+		}
+	}
+
+	if ( left < right && top < bottom ) {
+		rectangles.push_back( { left, top, right - left, bottom - top } );
+		return left == 0 && top == 0 && right == size.x && bottom == size.y
+				   ? FrameDamageResult::Full
+				   : FrameDamageResult::Rectangle;
+	}
+	return FrameDamageResult::None;
+}
+
+/** Uses sequential row traversal for frame buffers too large for cache-friendly tile traversal. */
+template <typename Rectangle>
+FrameDamageResult
+findFrameDamageAdaptive( const std::vector<Uint8>& current, const std::vector<Uint8>& previous,
+						 const Math::Sizei& size, std::vector<Rectangle>& rectangles,
+						 std::vector<Uint8>& changedTiles ) {
+	if ( current.size() < FrameDamageRowMajorThreshold )
+		return findFrameDamage( current, previous, size, rectangles );
+
+	rectangles.clear();
+	const Int32 tilesX = ( size.x + FrameDamageTileSize - 1 ) / FrameDamageTileSize;
+	const Int32 tilesY = ( size.y + FrameDamageTileSize - 1 ) / FrameDamageTileSize;
+	const size_t rowBytes = static_cast<size_t>( size.x ) * 3;
+	Int32 left = size.x;
+	Int32 top = size.y;
+	Int32 right = 0;
+	Int32 bottom = 0;
+	changedTiles.resize( static_cast<size_t>( tilesX ) );
+
+	for ( Int32 tileY = 0; tileY < tilesY; ++tileY ) {
+		const Int32 y = tileY * FrameDamageTileSize;
+		const Int32 height = eemin( FrameDamageTileSize, size.y - y );
+		std::memset( changedTiles.data(), 0, changedTiles.size() );
+		Int32 remainingTiles = tilesX;
+		for ( Int32 row = 0; row < height && remainingTiles; ++row ) {
+			const Int32 sourceRow = size.y - 1 - ( y + row );
+			const Uint8* currentRow = current.data() + static_cast<size_t>( sourceRow ) * rowBytes;
+			const Uint8* previousRow =
+				previous.data() + static_cast<size_t>( sourceRow ) * rowBytes;
+			for ( Int32 tileX = 0; tileX < tilesX; ++tileX ) {
+				if ( changedTiles[static_cast<size_t>( tileX )] )
+					continue;
+				const Int32 x = tileX * FrameDamageTileSize;
+				const Int32 width = eemin( FrameDamageTileSize, size.x - x );
+				const size_t offset = static_cast<size_t>( x ) * 3;
+				if ( 0 != std::memcmp( currentRow + offset, previousRow + offset,
+									   static_cast<size_t>( width ) * 3 ) ) {
+					changedTiles[static_cast<size_t>( tileX )] = 1;
+					--remainingTiles;
+					left = eemin( left, x );
+					top = eemin( top, y );
+					right = eemax( right, x + width );
+					bottom = eemax( bottom, y + height );
+				}
+			}
+		}
+	}
+
+	if ( left < right && top < bottom ) {
+		rectangles.push_back( { left, top, right - left, bottom - top } );
+		return left == 0 && top == 0 && right == size.x && bottom == size.y
+				   ? FrameDamageResult::Full
+				   : FrameDamageResult::Rectangle;
+	}
+	return FrameDamageResult::None;
+}
+
+}}} // namespace EE::Window::Private
+
+#endif