Merge "Misc RenderEngine perf fixes" into main (e9bec578) · Commits · e / os / android_frameworks_native

libs/renderengine/include/renderengine/LayerSettings.h

+6 −0

Original line number	Diff line number	Diff line
		@@ -228,6 +228,8 @@ static inline void PrintTo(const Geometry& settings, ::std::ostream* os) {
		*os << "Geometry {";
		*os << "\n .boundaries = ";
		PrintTo(settings.boundaries, os);
		*os << "\n .originalBounds = ";
		PrintTo(settings.originalBounds, os);
		*os << "\n .positionTransform = ";
		PrintMatrix(settings.positionTransform, os);
		*os << "\n .roundedCornersRadiusX = " << settings.roundedCornersRadius.x;
		@@ -328,6 +330,10 @@ static inline void PrintTo(const LayerSettings& settings, ::std::ostream* os) {
		*os << "\n .luts = ";
		PrintTo(settings.luts, os);
		}

		*os << "\n .borderSettings = " << settings.borderSettings.toString();
		*os << "\n .boxShadowSettings = " << settings.boxShadowSettings.toString();

		*os << "\n}";
		}

libs/renderengine/skia/Cache.cpp

+101 −21

Original line number	Diff line number	Diff line
		@@ -67,7 +67,8 @@ const std::array<float, 3> kLayerWhitePoints = {
		};
		} // namespace

		static void drawShadowLayers(SkiaRenderEngine* renderengine, const DisplaySettings& display,
		static void drawElevationShadowLayers(SkiaRenderEngine* renderengine,
		const DisplaySettings& display,
		const std::shared_ptr<ExternalTexture>& dstTexture) {
		// Somewhat arbitrary dimensions, but on screen and slightly shorter, based
		// on actual use.
		@@ -134,6 +135,78 @@ static void drawShadowLayers(SkiaRenderEngine* renderengine, const DisplaySettin
		}
		}

		static void drawBoxShadowLayers(SkiaRenderEngine* renderengine, const DisplaySettings& display,
		const std::shared_ptr<ExternalTexture>& dstTexture) {
		const Rect& displayRect = display.physicalDisplay;

		// The texture must be large enough for two shaders
		// 1. See ComputeBlurredRRectParams in GrBlurUtils.cpp, the dstTexture must be large enough
		// for the blur to be considered nine patcheable.
		// 2. See TesselationPathRenderer, there must be enough GPU work to choose the CPU path.
		LOG_ALWAYS_FATAL_IF(displayRect.width() < 384 \|\| displayRect.height() < 384,
		"dstTexture must be at least 256x256");

		gui::BorderSettings borderSettings;
		borderSettings.strokeWidth = 2.0f;
		borderSettings.color = 666747334;

		gui::BoxShadowSettings boxShadowSettings;
		gui::BoxShadowSettings::BoxShadowParams shadow1;
		shadow1.blurRadius = 28.0f;
		shadow1.spreadRadius = 0.0f;
		shadow1.color = 167772160;
		shadow1.offsetX = 0.0f;
		shadow1.offsetY = 0.0f;
		boxShadowSettings.boxShadows.push_back(shadow1);

		gui::BoxShadowSettings::BoxShadowParams shadow2;
		shadow2.blurRadius = 16.0f;
		shadow2.spreadRadius = 0.0f;
		shadow2.color = 436207616;
		shadow2.offsetX = 0.0f;
		shadow2.offsetY = 4.0f;
		boxShadowSettings.boxShadows.push_back(shadow2);

		FloatRect rect(20, 20, 250, 250);
		LayerSettings layer{
		.geometry =
		Geometry{
		.boundaries = rect,
		.originalBounds = rect,
		.roundedCornersRadius = {32.0f, 32.0f},
		.roundedCornersCrop = rect,
		.otherCrop = FloatRect(-16384, -16384, 16384, 16384),
		},
		.source =
		PixelSource{
		.solidColor = half3(0.f, 0.f, 0.f),
		},
		.alpha = 1,
		// setting this is mandatory for shadows and blurs
		.skipContentDraw = true,
		// drawShadow ignores alpha
		.borderSettings = borderSettings,
		.boxShadowSettings = boxShadowSettings,
		};

		{
		SFTRACE_NAME("RotatedClip");
		// This triggers quite a few shaders, not quite sure what they all are.
		layer.geometry.positionTransform = kFlip;
		renderengine->drawLayers(display, {layer}, dstTexture, base::unique_fd());
		}
		{
		SFTRACE_NAME("RRectBlur_NinePatch");
		layer.geometry.positionTransform = mat4();
		renderengine->drawLayers(display, {layer}, dstTexture, base::unique_fd());
		}
		{
		SFTRACE_NAME("ConcavePath_Tessellate");
		layer.geometry.positionTransform = mat4::scale(vec4(1.5f, 1.5f, 1.0f, 1.0f)) * kFlip;
		renderengine->drawLayers(display, {layer}, dstTexture, base::unique_fd());
		}
		}

		static void drawImageLayers(SkiaRenderEngine* renderengine, const DisplaySettings& display,
		const std::shared_ptr<ExternalTexture>& dstTexture,
		const std::shared_ptr<ExternalTexture>& srcTexture) {
		@@ -143,8 +216,6 @@ static void drawImageLayers(SkiaRenderEngine* renderengine, const DisplaySetting
		.geometry =
		Geometry{
		.boundaries = rect,
		// The position transform doesn't matter when the reduced shader mode
		// in in effect. A matrix transform stage is always included.
		.positionTransform = mat4(),
		.roundedCornersCrop = rect,
		},
		@@ -155,6 +226,8 @@ static void drawImageLayers(SkiaRenderEngine* renderengine, const DisplaySetting
		}},
		};

		for (mat4 transform : {mat4(), kFlip}) {
		layer.geometry.positionTransform = transform;
		for (auto dataspace : {kDestDataSpace, kOtherDataSpace}) {
		layer.sourceDataspace = dataspace;
		// Cache shaders for both rects and round rects.
		@@ -173,6 +246,7 @@ static void drawImageLayers(SkiaRenderEngine* renderengine, const DisplaySetting
		}
		}
		}
		}

		static void drawSolidLayers(SkiaRenderEngine* renderengine, const DisplaySettings& display,
		const std::shared_ptr<ExternalTexture>& dstTexture) {
		@@ -768,7 +842,7 @@ void Cache::primeShaderCache(SkiaRenderEngine* renderengine, PrimeCacheConfig co
		// The loop is beneficial for debugging and should otherwise be optimized out by the compiler.
		// Adding additional bounds to the loop is useful for verifying that the size of the dst buffer
		// does not impact the shader compilation counts by triggering different behaviors in RE/Skia.
		for (SkSize bounds : {SkSize::Make(128, 128), /SkSize::Make(1080, 2340)/}) {
		for (SkSize bounds : {SkSize::Make(384, 384), /SkSize::Make(1080, 2340)/}) {
		const nsecs_t timeBefore = systemTime();
		// The dimensions should not matter, so long as we draw inside them.
		const Rect displayRect(0, 0, bounds.fWidth, bounds.fHeight);
		@@ -851,9 +925,15 @@ void Cache::primeShaderCache(SkiaRenderEngine* renderengine, PrimeCacheConfig co
		}

		if (config.cacheShadowLayers) {
		{
		SFTRACE_NAME("cacheShadowLayers");
		drawShadowLayers(renderengine, display, srcTexture);
		drawShadowLayers(renderengine, p3Display, srcTexture);
		drawElevationShadowLayers(renderengine, display, srcTexture);
		drawElevationShadowLayers(renderengine, p3Display, srcTexture);
		}
		{
		SFTRACE_NAME("cacheBoxShadows");
		drawBoxShadowLayers(renderengine, display, srcTexture);
		}
		}

		if (renderengine->supportsBackgroundBlur()) {

libs/renderengine/skia/SkiaGLRenderEngine.cpp

+6 −0

Original line number	Diff line number	Diff line
		@@ -314,6 +314,12 @@ SkiaRenderEngine::Contexts SkiaGLRenderEngine::createContexts() {
		return contexts;
		}

		bool SkiaGLRenderEngine::supportsFastRotatedClipRRectAA() const {
		// clipRRect can take 1-10ms depending on the rrect size due to generating the
		// clip mask on the CPU.
		return false;
		}

		bool SkiaGLRenderEngine::supportsForwardPixelKill() const {
		// ARM gpu support this since 2013
		constexpr std::string kArm = "ARM";

libs/renderengine/skia/SkiaGLRenderEngine.h

+1 −0

Original line number	Diff line number	Diff line
		@@ -60,6 +60,7 @@ protected:
		// rendering backend
		virtual SkiaRenderEngine::Contexts createContexts();
		bool supportsForwardPixelKill() const override;
		bool supportsFastRotatedClipRRectAA() const override;
		bool supportsProtectedContentImpl() const override;
		bool useProtectedContextImpl(GrProtected isProtected) override;
		void waitFence(SkiaGpuContext* context, base::borrowed_fd fenceFd) override;

libs/renderengine/skia/SkiaRenderEngine.cpp

+9 −4

Original line number	Diff line number	Diff line
		@@ -874,7 +874,8 @@ void SkiaRenderEngine::drawLayersInternal(
		SkData::MakeWithCString(layerSettings.str().c_str()));
		}
		// Layers have a local transform that should be applied to them
		canvas->concat(getSkM44(layer.geometry.positionTransform).asM33());
		SkMatrix positionTransform = getSkM44(layer.geometry.positionTransform).asM33();
		canvas->concat(positionTransform);

		const auto [bounds, roundRectClip] =
		getBoundsAndClip(layer.geometry.boundaries, layer.geometry.roundedCornersCrop,
		@@ -957,6 +958,9 @@ void SkiaRenderEngine::drawLayersInternal(
		}
		}

		bool enableAntiAlias =
		supportsFastRotatedClipRRectAA() \|\| positionTransform.rectStaysRect();

		{
		SFTRACE_NAME("OutsetRendering");
		SkRRect otherCrop;
		@@ -966,7 +970,7 @@ void SkiaRenderEngine::drawLayersInternal(
		// Outset rendering needs to be clipped by parent.
		SkAutoCanvasRestore acr(canvas, true);
		if (!otherCrop.isEmpty()) {
		canvas->clipRRect(otherCrop, true);
		canvas->clipRRect(otherCrop, enableAntiAlias);
		}

		if (layer.shadow.length > 0) {
		@@ -1018,7 +1022,8 @@ void SkiaRenderEngine::drawLayersInternal(
		layer.borderSettings.strokeWidth);

		SkPaint paint;
		paint.setAntiAlias(true);
		// When rotated / scaling the lack of AA is imperceptible for the outline.
		paint.setAntiAlias(enableAntiAlias);
		paint.setColor(layer.borderSettings.color);
		paint.setStyle(SkPaint::kFill_Style);
		canvas->drawDRRect(outlineRect, preferredOriginalBounds, paint);
		@@ -1291,7 +1296,7 @@ void SkiaRenderEngine::drawLayersInternal(
		}

		if (!roundRectClip.isEmpty()) {
		canvas->clipRRect(roundRectClip, true);
		canvas->clipRRect(roundRectClip, enableAntiAlias);
		}

		if (!bounds.isRect()) {