perf(core): presize scene windows and separators, pin publish and moving-window draw allocations

commit: afb3cdc28db7091677b0e632a05090262f07d8fd
author: brice · 18d ago · 2026-10-04 20:03 UTC
committer: brice · 18d ago · 2026-10-04 20:13 UTC
parent: cca9f08
signature: GPG VerifiedCD88F9E564E0102646FB6077A1A07C7C7C3E349D · arch-gpg Verified for the committer email

5 files · +134 -4

M Makefile +1 -1


@@ -60,7 +60,7 @@ arch:
60	60		$(HOME)/go/bin/hexcheck -hexcheck.config .hexcheck.yaml -hexcheck.root . ./...
61	61		# Fast, mandatory guards for tiled content publications and SHM copies.
62	62		perf-check:
63		-	CGO_ENABLED=0 go test ./internal/core ./internal/app ./internal/adapters/drm ./internal/adapters/vulkan ./internal/adapters/wayland -run '^(TestAnimateAllocations\|TestOutputRoutingAllocations\|TestReportSeenAllocations\|TestFrameLifecycleTransitionsAllocations\|TestFullscreenShownAllocations\|TestFrameDecisionAllocations\|TestCommitFrameAllocations\|TestColorPipelineAllocations\|TestShownBySnapshotAllocations\|TestDueFramesAllocations\|TestRenderSteadyStateAllocations\|TestAccountFlipAllocations\|TestFlipDoneAllocations\|TestSceneWalkUnchangedTiledSHM\|TestTiledCommitPublishAllocations\|TestCapturedCommitApplyAllocations\|TestCapturedViewportApplyAllocations\|TestEffectiveInputEmptyTreeAllocations\|TestHeadlessTiledSHMCallbackAndCopyBudget)$$' -count=1
	63	+	CGO_ENABLED=0 go test ./internal/core ./internal/app ./internal/adapters/drm ./internal/adapters/vulkan ./internal/adapters/wayland -run '^(TestAnimateAllocations\|TestPublishAllocations\|TestOutputRoutingAllocations\|TestReportSeenAllocations\|TestFrameLifecycleTransitionsAllocations\|TestFullscreenShownAllocations\|TestFrameDecisionAllocations\|TestCommitFrameAllocations\|TestColorPipelineAllocations\|TestShownBySnapshotAllocations\|TestDueFramesAllocations\|TestRenderSteadyStateAllocations\|TestAccountFlipAllocations\|TestFlipDoneAllocations\|TestSceneWalkUnchangedTiledSHM\|TestTiledCommitPublishAllocations\|TestCapturedCommitApplyAllocations\|TestCapturedViewportApplyAllocations\|TestEffectiveInputEmptyTreeAllocations\|TestHeadlessTiledSHMCallbackAndCopyBudget)$$' -count=1
64	64		check: vet test arch fakes-check log-check perf-check
65	65
66	66		# On-demand SDR/HDR colour check with examples/testpattern on a headless


@@ -60,7 +60,7 @@ arch:
60		$(HOME)/go/bin/hexcheck -hexcheck.config .hexcheck.yaml -hexcheck.root . ./...	60		$(HOME)/go/bin/hexcheck -hexcheck.config .hexcheck.yaml -hexcheck.root . ./...
61		# Fast, mandatory guards for tiled content publications and SHM copies.	61		# Fast, mandatory guards for tiled content publications and SHM copies.
62		perf-check:	62		perf-check:
63	-	CGO_ENABLED=0 go test ./internal/core ./internal/app ./internal/adapters/drm ./internal/adapters/vulkan ./internal/adapters/wayland -run '^(TestAnimateAllocations\|TestOutputRoutingAllocations\|TestReportSeenAllocations\|TestFrameLifecycleTransitionsAllocations\|TestFullscreenShownAllocations\|TestFrameDecisionAllocations\|TestCommitFrameAllocations\|TestColorPipelineAllocations\|TestShownBySnapshotAllocations\|TestDueFramesAllocations\|TestRenderSteadyStateAllocations\|TestAccountFlipAllocations\|TestFlipDoneAllocations\|TestSceneWalkUnchangedTiledSHM\|TestTiledCommitPublishAllocations\|TestCapturedCommitApplyAllocations\|TestCapturedViewportApplyAllocations\|TestEffectiveInputEmptyTreeAllocations\|TestHeadlessTiledSHMCallbackAndCopyBudget)$$' -count=1	63	+	CGO_ENABLED=0 go test ./internal/core ./internal/app ./internal/adapters/drm ./internal/adapters/vulkan ./internal/adapters/wayland -run '^(TestAnimateAllocations\|TestPublishAllocations\|TestOutputRoutingAllocations\|TestReportSeenAllocations\|TestFrameLifecycleTransitionsAllocations\|TestFullscreenShownAllocations\|TestFrameDecisionAllocations\|TestCommitFrameAllocations\|TestColorPipelineAllocations\|TestShownBySnapshotAllocations\|TestDueFramesAllocations\|TestRenderSteadyStateAllocations\|TestAccountFlipAllocations\|TestFlipDoneAllocations\|TestSceneWalkUnchangedTiledSHM\|TestTiledCommitPublishAllocations\|TestCapturedCommitApplyAllocations\|TestCapturedViewportApplyAllocations\|TestEffectiveInputEmptyTreeAllocations\|TestHeadlessTiledSHMCallbackAndCopyBudget)$$' -count=1
64		check: vet test arch fakes-check log-check perf-check	64		check: vet test arch fakes-check log-check perf-check
65			65
66		# On-demand SDR/HDR colour check with examples/testpattern on a headless	66		# On-demand SDR/HDR colour check with examples/testpattern on a headless

M internal/adapters/vulkan/scene_perf_test.go +37 -1


@@ -9,7 +9,10 @@ import (
9	9		)
10	10
11	11		// TestRenderSteadyStateAllocations measures the cost of a small stable scene,
12		-	// on a normal and on a rotated output (orient works in place).
	12	+	// on a normal and on a rotated output (orient works in place), and that
	13	+	// building the draws of a scene with a moving window (a new Seq and window
	14	+	// slice every frame, as an animation publishes) allocates nothing.
	15	+	// TODO(A2): add the partial redraw seeded by sceneDelta once it is merged.
13	16		func TestRenderSteadyStateAllocations(t *testing.T) {
14	17		for _, tc := range []struct {
15	18		name string

@@ -41,6 +44,39 @@ func TestRenderSteadyStateAllocations(t *testing.T) {
41	44		} else {
42	45		t.Logf("small steady-state Render: %.1f allocs/frame", allocs)
43	46		}
	47	+	// A moving window: scenes are prebuilt (core allocates them, not
	48	+	// the renderer), each with a fresh Seq, so every frame draws
	49	+	// everything again. The draw building must not allocate.
	50	+	var moving [4]ports.Scene
	51	+	for i := range moving {
	52	+	moving[i] = ports.Scene{Seq: uint64(100 + i), Transform: tc.transform, Scale: 1, Windows: []ports.SceneWindow{{ID: 1, Rect: ports.Rect{X: i, W: tc.sceneW - 4, H: tc.sceneH - 4}}}}
	53	+	}
	54	+	bounds := image.Rect(0, 0, r.width, r.height)
	55	+	n := 0
	56	+	frameDraws := func() {
	57	+	n++
	58	+	s := moving[n%len(moving)]
	59	+	_ = r.draws(s, contents, r.frameDamage(r.last, s, bounds))
	60	+	}
	61	+	frameDraws()
	62	+	if allocs := testing.AllocsPerRun(50, frameDraws); allocs != 0 {
	63	+	t.Errorf("moving window draws: %.1f allocs/frame, want 0", allocs)
	64	+	}
	65	+	// A content change under an unchanged scene Seq: partial redraw
	66	+	// from the window damage, clipped to the changed region.
	67	+	changed := [2]ports.SurfaceContent{content, content}
	68	+	changed[1].Seq, changed[1].Version = 2, 2
	69	+	same := scene
	70	+	same.Seq = r.last.sceneSeq
	71	+	partial := func() {
	72	+	n++
	73	+	contents[1] = changed[n%2]
	74	+	_ = r.draws(same, contents, r.frameDamage(r.last, same, bounds))
	75	+	}
	76	+	partial()
	77	+	if allocs := testing.AllocsPerRun(50, partial); allocs != 0 {
	78	+	t.Errorf("content-damage draws: %.1f allocs/frame, want 0", allocs)
	79	+	}
44	80		})
45	81		}
46	82		}


@@ -9,7 +9,10 @@ import (
9		)	9		)
10			10
11		// TestRenderSteadyStateAllocations measures the cost of a small stable scene,	11		// TestRenderSteadyStateAllocations measures the cost of a small stable scene,
12	-	// on a normal and on a rotated output (orient works in place).	12	+	// on a normal and on a rotated output (orient works in place), and that
			13	+	// building the draws of a scene with a moving window (a new Seq and window
			14	+	// slice every frame, as an animation publishes) allocates nothing.
			15	+	// TODO(A2): add the partial redraw seeded by sceneDelta once it is merged.
13		func TestRenderSteadyStateAllocations(t *testing.T) {	16		func TestRenderSteadyStateAllocations(t *testing.T) {
14		for _, tc := range []struct {	17		for _, tc := range []struct {
15		name string	18		name string

@@ -41,6 +44,39 @@ func TestRenderSteadyStateAllocations(t *testing.T) {
41		} else {	44		} else {
42		t.Logf("small steady-state Render: %.1f allocs/frame", allocs)	45		t.Logf("small steady-state Render: %.1f allocs/frame", allocs)
43		}	46		}
			47	+	// A moving window: scenes are prebuilt (core allocates them, not
			48	+	// the renderer), each with a fresh Seq, so every frame draws
			49	+	// everything again. The draw building must not allocate.
			50	+	var moving [4]ports.Scene
			51	+	for i := range moving {
			52	+	moving[i] = ports.Scene{Seq: uint64(100 + i), Transform: tc.transform, Scale: 1, Windows: []ports.SceneWindow{{ID: 1, Rect: ports.Rect{X: i, W: tc.sceneW - 4, H: tc.sceneH - 4}}}}
			53	+	}
			54	+	bounds := image.Rect(0, 0, r.width, r.height)
			55	+	n := 0
			56	+	frameDraws := func() {
			57	+	n++
			58	+	s := moving[n%len(moving)]
			59	+	_ = r.draws(s, contents, r.frameDamage(r.last, s, bounds))
			60	+	}
			61	+	frameDraws()
			62	+	if allocs := testing.AllocsPerRun(50, frameDraws); allocs != 0 {
			63	+	t.Errorf("moving window draws: %.1f allocs/frame, want 0", allocs)
			64	+	}
			65	+	// A content change under an unchanged scene Seq: partial redraw
			66	+	// from the window damage, clipped to the changed region.
			67	+	changed := [2]ports.SurfaceContent{content, content}
			68	+	changed[1].Seq, changed[1].Version = 2, 2
			69	+	same := scene
			70	+	same.Seq = r.last.sceneSeq
			71	+	partial := func() {
			72	+	n++
			73	+	contents[1] = changed[n%2]
			74	+	_ = r.draws(same, contents, r.frameDamage(r.last, same, bounds))
			75	+	}
			76	+	partial()
			77	+	if allocs := testing.AllocsPerRun(50, partial); allocs != 0 {
			78	+	t.Errorf("content-damage draws: %.1f allocs/frame, want 0", allocs)
			79	+	}
44		})	80		})
45		}	81		}
46		}	82		}

M internal/core/core.go +9 -2


@@ -67,6 +67,8 @@ type binding struct {
67	67		key string
68	68		}
69	69		type Core struct {
	70	+	// overviewReal is publish's scratch map of the real layouts under an open overview.
	71	+	overviewReal map[WindowID]Placement
70	72		security ports.SecurityState
71	73		lockSurfaces []ports.LockSurfacePlacement
72	74		lockFocus WindowID

@@ -559,13 +561,18 @@ func (c *Core) publish(ctx context.Context) error {
559	561		if frame != (Rect{W: o.W, H: o.H}) {
560	562		clip = frame
561	563		}
562		-	scene := ports.Scene{Security: c.security, Output: sc.name(), OutputWidth: o.W, OutputHeight: o.H, WorkspaceClip: clip, Scale: sc.scale, Transform: sc.transform, Off: sc.off, Background: c.cfg.Background.Color, Border: ports.Border{Width: c.cfg.Border.Width, Active: c.cfg.Border.Active, Inactive: c.cfg.Border.Inactive}, Windows: make([]ports.SceneWindow, 0), Layers: shownLayers(sc)}
563	564		// layout is what is drawn; settled (same indexes) is where the
564	565		// windows are going, and alone sizes the configures.
565	566		layout, settled := sc.shown, sc.settledLayout
	567	+	scene := ports.Scene{Security: c.security, Output: sc.name(), OutputWidth: o.W, OutputHeight: o.H, WorkspaceClip: clip, Scale: sc.scale, Transform: sc.transform, Off: sc.off, Background: c.cfg.Background.Color, Border: ports.Border{Width: c.cfg.Border.Width, Active: c.cfg.Border.Active, Inactive: c.cfg.Border.Inactive}, Windows: make([]ports.SceneWindow, 0, len(layout)+len(c.popupOrder)), Layers: shownLayers(sc)}
566	568		var real map[WindowID]Placement
567	569		if sc.mon.ov.open {
568		-	real = make(map[WindowID]Placement)
	570	+	// Local to this screen's build: reused, cleared each time.
	571	+	if c.overviewReal == nil {
	572	+	c.overviewReal = make(map[WindowID]Placement)
	573	+	}
	574	+	clear(c.overviewReal)
	575	+	real = c.overviewReal
569	576		for w := range sc.mon.all() {
570	577		for _, p := range w.Layout() {
571	578		real[p.ID] = p

 	key  string
 }
 type Core struct {
+	// overviewReal is publish's scratch map of the real layouts under an open overview.
+	overviewReal    map[WindowID]Placement
 	security        ports.SecurityState
 	lockSurfaces    []ports.LockSurfacePlacement
 	lockFocus       WindowID
 		if frame != (Rect{W: o.W, H: o.H}) {
 			clip = frame
 		}
-		scene := ports.Scene{Security: c.security, Output: sc.name(), OutputWidth: o.W, OutputHeight: o.H, WorkspaceClip: clip, Scale: sc.scale, Transform: sc.transform, Off: sc.off, Background: c.cfg.Background.Color, Border: ports.Border{Width: c.cfg.Border.Width, Active: c.cfg.Border.Active, Inactive: c.cfg.Border.Inactive}, Windows: make([]ports.SceneWindow, 0), Layers: shownLayers(sc)}
 		// layout is what is drawn; settled (same indexes) is where the
 		// windows are going, and alone sizes the configures.
 		layout, settled := sc.shown, sc.settledLayout
+		scene := ports.Scene{Security: c.security, Output: sc.name(), OutputWidth: o.W, OutputHeight: o.H, WorkspaceClip: clip, Scale: sc.scale, Transform: sc.transform, Off: sc.off, Background: c.cfg.Background.Color, Border: ports.Border{Width: c.cfg.Border.Width, Active: c.cfg.Border.Active, Inactive: c.cfg.Border.Inactive}, Windows: make([]ports.SceneWindow, 0, len(layout)+len(c.popupOrder)), Layers: shownLayers(sc)}
 		var real map[WindowID]Placement
 		if sc.mon.ov.open {
-			real = make(map[WindowID]Placement)
+			// Local to this screen's build: reused, cleared each time.
+			if c.overviewReal == nil {
+				c.overviewReal = make(map[WindowID]Placement)
+			}
+			clear(c.overviewReal)
+			real = c.overviewReal
 			for w := range sc.mon.all() {
 				for _, p := range w.Layout() {
 					real[p.ID] = p

A internal/core/publish_alloc_test.go +82 -0

@@ -0,0 +1,82 @@
	1	+	package core
	2	+
	3	+	import (
	4	+	"context"
	5	+	"testing"
	6	+	"time"
	7	+
	8	+	"github.com/bnema/neferwl/internal/ports"
	9	+	)
	10	+
	11	+	// publishAllocBudget is the allocations of one animation frame (step) on two
	12	+	// screens of six windows each, with camera and rect motions running. Every
	13	+	// scene sent is freshly built (immutable once sent), and the layouts are
	14	+	// rebuilt for every publish, so the floor is not 0. Measured: 43 before
	15	+	// presizing the scene windows and separators, 36 after.
	16	+	const publishAllocBudget = 36
	17	+
	18	+	// TestPublishAllocations pins the cost of publishing an animation frame.
	19	+	func TestPublishAllocations(t *testing.T) {
	20	+	// No Clock port: a mock clock would count its own bookkeeping, and the
	21	+	// fallback timer then is the system one.
	22	+	var cfg ports.Config
	23	+	cfg.Keyboard.CmdKey = "super"
	24	+	cfg.Layout.MaxColumns = 2
	25	+	cfg.Border.Width = 2
	26	+	cfg.Animations.On = true
	27	+	c, err := New(cfg, Channels{Scenes: make(chan []ports.Scene, 1), Commands: make(chan ports.ClientCommand, 64)})
	28	+	if err != nil {
	29	+	t.Fatal(err)
	30	+	}
	31	+	c.addScreen(ports.OutputInfo{Name: "A", Width: 300, Height: 200})
	32	+	c.addScreen(ports.OutputInfo{Name: "B", Width: 300, Height: 200})
	33	+	t0 := time.Now()
	34	+	next := WindowID(1)
	35	+	for _, sc := range c.screens {
	36	+	for range 6 {
	37	+	sc.mon.AddWindow(next)
	38	+	next++
	39	+	}
	40	+	sc.arrange()
	41	+	}
	42	+	c.refreshShown()
	43	+	for _, sc := range c.screens {
	44	+	m := sc.mon
	45	+	ws := m.Current()
	46	+	ws.motion = c.spring(viewSpring(100, 0), t0)
	47	+	ws.shift = 100
	48	+	sc.rects = make(map[WindowID]rectMotion, 6)
	49	+	sp := viewSpring(100, 0)
	50	+	for _, p := range sc.settledLayout {
	51	+	sc.rects[p.ID] = rectMotion{
	52	+	x: c.spring(sp, t0), y: c.spring(sp, t0),
	53	+	w: c.spring(sp, t0), h: c.spring(sp, t0),
	54	+	dx: 100, dy: 100, dw: 100, dh: 100,
	55	+	}
	56	+	}
	57	+	sc.rectsWS = ws
	58	+	}
	59	+	ctx := context.Background()
	60	+	step := func() {
	61	+	if err := c.step(ctx, nil); err != nil {
	62	+	t.Fatal(err)
	63	+	}
	64	+	<-c.ch.Scenes
	65	+	}
	66	+	step()
	67	+	for _, sc := range c.screens {
	68	+	if len(sc.rects) != 6 \|\| !sc.mon.Current().motion.on \|\| len(sc.shown) < 6 {
	69	+	t.Fatalf("setup: rects %d camera %v shown %d", len(sc.rects), sc.mon.Current().motion.on, len(sc.shown))
	70	+	}
	71	+	}
	72	+	n := testing.AllocsPerRun(50, step)
	73	+	t.Logf("allocs per step = %v", n)
	74	+	for _, sc := range c.screens {
	75	+	if len(sc.rects) == 0 \|\| !sc.mon.Current().motion.on {
	76	+	t.Fatalf("motions settled during the measure: rects %d", len(sc.rects))
	77	+	}
	78	+	}
	79	+	if n > publishAllocBudget {
	80	+	t.Errorf("publish allocs per frame = %v, budget %d", n, publishAllocBudget)
	81	+	}
	82	+	}

@@ -0,0 +1,82 @@
	1	+	package core
	2	+
	3	+	import (
	4	+	"context"
	5	+	"testing"
	6	+	"time"
	7	+
	8	+	"github.com/bnema/neferwl/internal/ports"
	9	+	)
	10	+
	11	+	// publishAllocBudget is the allocations of one animation frame (step) on two
	12	+	// screens of six windows each, with camera and rect motions running. Every
	13	+	// scene sent is freshly built (immutable once sent), and the layouts are
	14	+	// rebuilt for every publish, so the floor is not 0. Measured: 43 before
	15	+	// presizing the scene windows and separators, 36 after.
	16	+	const publishAllocBudget = 36
	17	+
	18	+	// TestPublishAllocations pins the cost of publishing an animation frame.
	19	+	func TestPublishAllocations(t *testing.T) {
	20	+	// No Clock port: a mock clock would count its own bookkeeping, and the
	21	+	// fallback timer then is the system one.
	22	+	var cfg ports.Config
	23	+	cfg.Keyboard.CmdKey = "super"
	24	+	cfg.Layout.MaxColumns = 2
	25	+	cfg.Border.Width = 2
	26	+	cfg.Animations.On = true
	27	+	c, err := New(cfg, Channels{Scenes: make(chan []ports.Scene, 1), Commands: make(chan ports.ClientCommand, 64)})
	28	+	if err != nil {
	29	+	t.Fatal(err)
	30	+	}
	31	+	c.addScreen(ports.OutputInfo{Name: "A", Width: 300, Height: 200})
	32	+	c.addScreen(ports.OutputInfo{Name: "B", Width: 300, Height: 200})
	33	+	t0 := time.Now()
	34	+	next := WindowID(1)
	35	+	for _, sc := range c.screens {
	36	+	for range 6 {
	37	+	sc.mon.AddWindow(next)
	38	+	next++
	39	+	}
	40	+	sc.arrange()
	41	+	}
	42	+	c.refreshShown()
	43	+	for _, sc := range c.screens {
	44	+	m := sc.mon
	45	+	ws := m.Current()
	46	+	ws.motion = c.spring(viewSpring(100, 0), t0)
	47	+	ws.shift = 100
	48	+	sc.rects = make(map[WindowID]rectMotion, 6)
	49	+	sp := viewSpring(100, 0)
	50	+	for _, p := range sc.settledLayout {
	51	+	sc.rects[p.ID] = rectMotion{
	52	+	x: c.spring(sp, t0), y: c.spring(sp, t0),
	53	+	w: c.spring(sp, t0), h: c.spring(sp, t0),
	54	+	dx: 100, dy: 100, dw: 100, dh: 100,
	55	+	}
	56	+	}
	57	+	sc.rectsWS = ws
	58	+	}
	59	+	ctx := context.Background()
	60	+	step := func() {
	61	+	if err := c.step(ctx, nil); err != nil {
	62	+	t.Fatal(err)
	63	+	}
	64	+	<-c.ch.Scenes
	65	+	}
	66	+	step()
	67	+	for _, sc := range c.screens {
	68	+	if len(sc.rects) != 6 \|\| !sc.mon.Current().motion.on \|\| len(sc.shown) < 6 {
	69	+	t.Fatalf("setup: rects %d camera %v shown %d", len(sc.rects), sc.mon.Current().motion.on, len(sc.shown))
	70	+	}
	71	+	}
	72	+	n := testing.AllocsPerRun(50, step)
	73	+	t.Logf("allocs per step = %v", n)
	74	+	for _, sc := range c.screens {
	75	+	if len(sc.rects) == 0 \|\| !sc.mon.Current().motion.on {
	76	+	t.Fatalf("motions settled during the measure: rects %d", len(sc.rects))
	77	+	}
	78	+	}
	79	+	if n > publishAllocBudget {
	80	+	t.Errorf("publish allocs per frame = %v, budget %d", n, publishAllocBudget)
	81	+	}
	82	+	}

M internal/core/separators.go +5 -0


@@ -38,6 +38,11 @@ func separators(ps []Placement, width, gap int, o Rect, lit bool) []ports.Separa
38	38		if p.Floating {
39	39		id = p.ID
40	40		}
	41	+	if out == nil {
	42	+	// One line per side of every tile, and the focused one's four
	43	+	// lit lines: the common case grows no further.
	44	+	out = make([]ports.Separator, 0, 4*len(ps)+4)
	45	+	}
41	46		out = append(out, ports.Separator{Rect: r, Active: active, Window: id})
42	47		}
43	48		var focused *Placement


@@ -38,6 +38,11 @@ func separators(ps []Placement, width, gap int, o Rect, lit bool) []ports.Separa
38	if p.Floating {	38		if p.Floating {
39	id = p.ID	39		id = p.ID
40	}	40		}
		41	+	if out == nil {
		42	+	// One line per side of every tile, and the focused one's four
		43	+	// lit lines: the common case grows no further.
		44	+	out = make([]ports.Separator, 0, 4*len(ps)+4)
		45	+	}
41	out = append(out, ports.Separator{Rect: r, Active: active, Window: id})	46		out = append(out, ports.Separator{Rect: r, Active: active, Window: id})
42	}	47		}
43	var focused *Placement	48		var focused *Placement