perf(vulkan): draw position hint dots as single rounded quads

The pill and each dot were 1+2*radius scanline fills computed with
math.Sqrt (143 draws for 5x3, 2071 for 16x16). Each is now one
SDF-rounded quad: 1+N draws.

commit: fac869b569de1ba1c42dc7a92903522e4284aa65
author: brice · 35h ago · 2026-10-21 08:05 UTC
parent: 37f59a2
signature: GPG VerifiedCD88F9E564E0102646FB6077A1A07C7C7C3E349D · arch-gpg Verified for the committer email

3 files · +80 -13

M internal/adapters/vulkan/position_hint.go +7 -13


@@ -2,7 +2,6 @@ package vulkan
2	2
3	3		import (
4	4		"image"
5		-	"math"
6	5
7	6		"github.com/bnema/neferwl/internal/ports"
8	7		)

@@ -38,20 +37,15 @@ func (w *sceneWalk) positionHint() {
38	37		w.alpha = alpha
39	38		}
40	39
41		-	// Rounded fills use the existing solid instances and batching pipeline. Each
42		-	// physical scanline is bounded by the scaled radius; no textures or per-frame
43		-	// buffers are allocated. Non-overlapping strips preserve translucent alpha.
	40	+	// hintRound draws one rounded quad (SDF in the shader), so the pill and each
	41	+	// dot cost a single draw whatever the scale. Fills are one translucent draw,
	42	+	// so alpha stays exact.
44	43		func (w *sceneWalk) hintRound(r image.Rectangle, radius int, color [3]uint8) {
45	44		radius = min(radius, r.Dx()/2, r.Dy()/2)
46		-	if radius <= 0 {
47		-	w.fill(r, color)
	45	+	n := len(w.draws)
	46	+	w.fill(r, color)
	47	+	if radius <= 0 \|\| len(w.draws) == n {
48	48		return
49	49		}
50		-	w.fill(image.Rect(r.Min.X, r.Min.Y+radius, r.Max.X, r.Max.Y-radius), color)
51		-	for y := range radius {
52		-	dy := float64(radius-y) - 0.5
53		-	inset := radius - int(math.Sqrt(float64(radiusradius)-dydy)+0.5)
54		-	w.fill(image.Rect(r.Min.X+inset, r.Min.Y+y, r.Max.X-inset, r.Min.Y+y+1), color)
55		-	w.fill(image.Rect(r.Min.X+inset, r.Max.Y-y-1, r.Max.X-inset, r.Max.Y-y), color)
56		-	}
	50	+	roundedDraw(&w.draws[n], r, float32(radius))
57	51		}

 import (
 	"image"
-	"math"
 	"github.com/bnema/neferwl/internal/ports"
 )
 	w.alpha = alpha
 }
-// Rounded fills use the existing solid instances and batching pipeline. Each
+// hintRound draws one rounded quad (SDF in the shader), so the pill and each
-// physical scanline is bounded by the scaled radius; no textures or per-frame
+// dot cost a single draw whatever the scale. Fills are one translucent draw,
-// buffers are allocated. Non-overlapping strips preserve translucent alpha.
+// so alpha stays exact.
 func (w *sceneWalk) hintRound(r image.Rectangle, radius int, color [3]uint8) {
 	radius = min(radius, r.Dx()/2, r.Dy()/2)
-	if radius <= 0 {
+	n := len(w.draws)
-		w.fill(r, color)
+	w.fill(r, color)
+	if radius <= 0 || len(w.draws) == n {
 		return
 	}
-	w.fill(image.Rect(r.Min.X, r.Min.Y+radius, r.Max.X, r.Max.Y-radius), color)
+	roundedDraw(&w.draws[n], r, float32(radius))
-	for y := range radius {
-		dy := float64(radius-y) - 0.5
-		inset := radius - int(math.Sqrt(float64(radius*radius)-dy*dy)+0.5)
-		w.fill(image.Rect(r.Min.X+inset, r.Min.Y+y, r.Max.X-inset, r.Min.Y+y+1), color)
-		w.fill(image.Rect(r.Min.X+inset, r.Max.Y-y-1, r.Max.X-inset, r.Max.Y-y), color)
-	}
 }

A internal/adapters/vulkan/position_hint_bench_test.go +35 -0

@@ -0,0 +1,35 @@
	1	+	package vulkan
	2	+
	3	+	import (
	4	+	"image"
	5	+	"testing"
	6	+
	7	+	"github.com/bnema/neferwl/internal/ports"
	8	+	)
	9	+
	10	+	func hintGrid(cols, rows int) ports.PositionHint {
	11	+	h := ports.PositionHint{Opacity: 1, Columns: uint8(cols), ActiveCol: 1, ActiveRow: 1}
	12	+	for c := range cols {
	13	+	h.Counts[c] = uint8(rows)
	14	+	}
	15	+	return h
	16	+	}
	17	+
	18	+	func BenchmarkPositionHintDraws(b *testing.B) {
	19	+	for _, g := range []struct {
	20	+	name string
	21	+	cols, rows int
	22	+	}{{"5x3", 5, 3}, {"16x16", 16, 16}} {
	23	+	b.Run(g.name, func(b *testing.B) {
	24	+	r := &Renderer{width: 1920, height: 1080}
	25	+	s := ports.Scene{OutputWidth: 1920, OutputHeight: 1080, Scale: 1, PositionHint: hintGrid(g.cols, g.rows)}
	26	+	dmg := newDamage(&target{}, s, image.Rect(0, 0, r.width, r.height))
	27	+	b.ReportAllocs()
	28	+	var n int
	29	+	for b.Loop() {
	30	+	n = len(r.draws(s, nil, dmg))
	31	+	}
	32	+	b.ReportMetric(float64(n), "draws")
	33	+	})
	34	+	}
	35	+	}

@@ -0,0 +1,35 @@
	1	+	package vulkan
	2	+
	3	+	import (
	4	+	"image"
	5	+	"testing"
	6	+
	7	+	"github.com/bnema/neferwl/internal/ports"
	8	+	)
	9	+
	10	+	func hintGrid(cols, rows int) ports.PositionHint {
	11	+	h := ports.PositionHint{Opacity: 1, Columns: uint8(cols), ActiveCol: 1, ActiveRow: 1}
	12	+	for c := range cols {
	13	+	h.Counts[c] = uint8(rows)
	14	+	}
	15	+	return h
	16	+	}
	17	+
	18	+	func BenchmarkPositionHintDraws(b *testing.B) {
	19	+	for _, g := range []struct {
	20	+	name string
	21	+	cols, rows int
	22	+	}{{"5x3", 5, 3}, {"16x16", 16, 16}} {
	23	+	b.Run(g.name, func(b *testing.B) {
	24	+	r := &Renderer{width: 1920, height: 1080}
	25	+	s := ports.Scene{OutputWidth: 1920, OutputHeight: 1080, Scale: 1, PositionHint: hintGrid(g.cols, g.rows)}
	26	+	dmg := newDamage(&target{}, s, image.Rect(0, 0, r.width, r.height))
	27	+	b.ReportAllocs()
	28	+	var n int
	29	+	for b.Loop() {
	30	+	n = len(r.draws(s, nil, dmg))
	31	+	}
	32	+	b.ReportMetric(float64(n), "draws")
	33	+	})
	34	+	}
	35	+	}

M internal/adapters/vulkan/position_hint_test.go +38 -0


@@ -4,6 +4,7 @@ import (
4	4		"github.com/bnema/neferwl/internal/ports"
5	5		"github.com/stretchr/testify/require"
6	6		"image"
	7	+	"math"
7	8		"testing"
8	9		)
9	10

@@ -69,3 +70,40 @@ func TestPositionHintDamageTransitions(t *testing.T) {
69	70		require.False(t, ok, "scale change requires full target redraw; geometry scaling is checked separately")
70	71		})
71	72		}
	73	+
	74	+	func TestPositionHintRoundedQuads(t *testing.T) {
	75	+	for _, scale := range []float64{1, 2} {
	76	+	r := &Renderer{width: int(640 * scale), height: int(480 * scale)}
	77	+	h := hintGrid(3, 2)
	78	+	s := ports.Scene{OutputWidth: 640, OutputHeight: 480, Scale: scale, PositionHint: h}
	79	+	dmg := newDamage(&target{}, s, image.Rect(0, 0, r.width, r.height))
	80	+	draws := r.draws(s, nil, dmg)
	81	+	require.Len(t, draws, 1+6)
	82	+	w := sceneWalk{scale: scale}
	83	+	b := h.Bounds(640, 480)
	84	+	pill := w.physRect(b.X, b.Y, b.W, b.H)
	85	+	for i, d := range draws {
	86	+	require.NotZero(t, d.pc.misc[1]&flagRounded, "draw %d", i)
	87	+	require.Zero(t, d.pc.misc[1]&flagRing)
	88	+	require.Equal(t, roundRect(image.Rect(int(d.pc.rect[0]), int(d.pc.rect[1]), int(d.pc.rect[2]), int(d.pc.rect[3]))), d.round)
	89	+	}
	90	+	require.Equal(t, pill, image.Rect(int(draws[0].pc.rect[0]), int(draws[0].pc.rect[1]), int(draws[0].pc.rect[2]), int(draws[0].pc.rect[3])))
	91	+	require.Equal(t, float32(w.phys(ports.PositionHintRadius)), math.Float32frombits(draws[0].radii[0]))
	92	+	// Active dot (col 1,row 1) is the larger one: radius = phys(active diameter)/2.
	93	+	active := draws[1+1*2+1]
	94	+	require.Equal(t, float32(w.phys(ports.PositionHintActiveDiameter)/2), math.Float32frombits(active.radii[0]))
	95	+	other := draws[1]
	96	+	require.Equal(t, float32(w.phys(ports.PositionHintDiameter)/2), math.Float32frombits(other.radii[0]))
	97	+	}
	98	+	}
	99	+
	100	+	func TestPositionHintRoundedOrient(t *testing.T) {
	101	+	r := &Renderer{width: 640, height: 480}
	102	+	s := ports.Scene{OutputWidth: 640, OutputHeight: 480, Scale: 1, PositionHint: hintGrid(2, 2)}
	103	+	dmg := newDamage(&target{}, s, image.Rect(0, 0, r.width, r.height))
	104	+	ds := r.draws(s, nil, dmg)
	105	+	r.orient(ds, ports.BufferTransform(1))
	106	+	for _, d := range ds {
	107	+	require.Equal(t, roundRect(image.Rect(int(d.pc.rect[0]), int(d.pc.rect[1]), int(d.pc.rect[2]), int(d.pc.rect[3]))), d.round)
	108	+	}
	109	+	}


@@ -4,6 +4,7 @@ import (
4	"github.com/bnema/neferwl/internal/ports"	4		"github.com/bnema/neferwl/internal/ports"
5	"github.com/stretchr/testify/require"	5		"github.com/stretchr/testify/require"
6	"image"	6		"image"
		7	+	"math"
7	"testing"	8		"testing"
8	)	9		)
9		10

@@ -69,3 +70,40 @@ func TestPositionHintDamageTransitions(t *testing.T) {
69	require.False(t, ok, "scale change requires full target redraw; geometry scaling is checked separately")	70		require.False(t, ok, "scale change requires full target redraw; geometry scaling is checked separately")
70	})	71		})
71	}	72		}
		73	+
		74	+	func TestPositionHintRoundedQuads(t *testing.T) {
		75	+	for _, scale := range []float64{1, 2} {
		76	+	r := &Renderer{width: int(640 * scale), height: int(480 * scale)}
		77	+	h := hintGrid(3, 2)
		78	+	s := ports.Scene{OutputWidth: 640, OutputHeight: 480, Scale: scale, PositionHint: h}
		79	+	dmg := newDamage(&target{}, s, image.Rect(0, 0, r.width, r.height))
		80	+	draws := r.draws(s, nil, dmg)
		81	+	require.Len(t, draws, 1+6)
		82	+	w := sceneWalk{scale: scale}
		83	+	b := h.Bounds(640, 480)
		84	+	pill := w.physRect(b.X, b.Y, b.W, b.H)
		85	+	for i, d := range draws {
		86	+	require.NotZero(t, d.pc.misc[1]&flagRounded, "draw %d", i)
		87	+	require.Zero(t, d.pc.misc[1]&flagRing)
		88	+	require.Equal(t, roundRect(image.Rect(int(d.pc.rect[0]), int(d.pc.rect[1]), int(d.pc.rect[2]), int(d.pc.rect[3]))), d.round)
		89	+	}
		90	+	require.Equal(t, pill, image.Rect(int(draws[0].pc.rect[0]), int(draws[0].pc.rect[1]), int(draws[0].pc.rect[2]), int(draws[0].pc.rect[3])))
		91	+	require.Equal(t, float32(w.phys(ports.PositionHintRadius)), math.Float32frombits(draws[0].radii[0]))
		92	+	// Active dot (col 1,row 1) is the larger one: radius = phys(active diameter)/2.
		93	+	active := draws[1+1*2+1]
		94	+	require.Equal(t, float32(w.phys(ports.PositionHintActiveDiameter)/2), math.Float32frombits(active.radii[0]))
		95	+	other := draws[1]
		96	+	require.Equal(t, float32(w.phys(ports.PositionHintDiameter)/2), math.Float32frombits(other.radii[0]))
		97	+	}
		98	+	}
		99	+
		100	+	func TestPositionHintRoundedOrient(t *testing.T) {
		101	+	r := &Renderer{width: 640, height: 480}
		102	+	s := ports.Scene{OutputWidth: 640, OutputHeight: 480, Scale: 1, PositionHint: hintGrid(2, 2)}
		103	+	dmg := newDamage(&target{}, s, image.Rect(0, 0, r.width, r.height))
		104	+	ds := r.draws(s, nil, dmg)
		105	+	r.orient(ds, ports.BufferTransform(1))
		106	+	for _, d := range ds {
		107	+	require.Equal(t, roundRect(image.Rect(int(d.pc.rect[0]), int(d.pc.rect[1]), int(d.pc.rect[2]), int(d.pc.rect[3]))), d.round)
		108	+	}
		109	+	}