From d4d4b9c070dc1568ebc6f5e553cc987d3d8fe1d7 Mon Sep 17 00:00:00 2001
From: Hajime Hoshi <hajimehoshi@gmail.com>
Date: Fri, 4 Mar 2022 03:47:55 +0900
Subject: [PATCH] internal/graphicsdriver/metal,
 internal/graphicsdriver/opengl: more efficient modulo

---
 image_test.go                                 | 81 +++++++++++++++++++
 .../graphicsdriver/metal/graphics_darwin.go   |  8 +-
 .../graphicsdriver/opengl/defaultshader.go    |  9 +--
 3 files changed, 85 insertions(+), 13 deletions(-)

diff --git a/image_test.go b/image_test.go
index 6ed2f31c1..5d56d6c12 100644
--- a/image_test.go
+++ b/image_test.go
@@ -1318,6 +1318,87 @@ func TestImageAddressRepeat(t *testing.T) {
 	}
 }
 
+func TestImageAddressRepeatNegativePosition(t *testing.T) {
+	const w, h = 16, 16
+	src := ebiten.NewImage(w, h)
+	dst := ebiten.NewImage(w, h)
+	pix := make([]byte, 4*w*h)
+	for j := 0; j < h; j++ {
+		for i := 0; i < w; i++ {
+			idx := 4 * (i + j*w)
+			if 4 <= i && i < 8 && 4 <= j && j < 8 {
+				pix[idx] = byte(i-4) * 0x10
+				pix[idx+1] = byte(j-4) * 0x10
+				pix[idx+2] = 0
+				pix[idx+3] = 0xff
+			} else {
+				pix[idx] = 0
+				pix[idx+1] = 0
+				pix[idx+2] = 0xff
+				pix[idx+3] = 0xff
+			}
+		}
+	}
+	src.ReplacePixels(pix)
+
+	vs := []ebiten.Vertex{
+		{
+			DstX:   0,
+			DstY:   0,
+			SrcX:   -w,
+			SrcY:   -h,
+			ColorR: 1,
+			ColorG: 1,
+			ColorB: 1,
+			ColorA: 1,
+		},
+		{
+			DstX:   w,
+			DstY:   0,
+			SrcX:   0,
+			SrcY:   -h,
+			ColorR: 1,
+			ColorG: 1,
+			ColorB: 1,
+			ColorA: 1,
+		},
+		{
+			DstX:   0,
+			DstY:   h,
+			SrcX:   -w,
+			SrcY:   0,
+			ColorR: 1,
+			ColorG: 1,
+			ColorB: 1,
+			ColorA: 1,
+		},
+		{
+			DstX:   w,
+			DstY:   h,
+			SrcX:   0,
+			SrcY:   0,
+			ColorR: 1,
+			ColorG: 1,
+			ColorB: 1,
+			ColorA: 1,
+		},
+	}
+	is := []uint16{0, 1, 2, 1, 2, 3}
+	op := &ebiten.DrawTrianglesOptions{}
+	op.Address = ebiten.AddressRepeat
+	dst.DrawTriangles(vs, is, src.SubImage(image.Rect(4, 4, 8, 8)).(*ebiten.Image), op)
+
+	for j := 0; j < h; j++ {
+		for i := 0; i < w; i++ {
+			got := dst.At(i, j).(color.RGBA)
+			want := color.RGBA{byte(i%4) * 0x10, byte(j%4) * 0x10, 0, 0xff}
+			if !sameColors(got, want, 1) {
+				t.Errorf("dst.At(%d, %d): got %v, want: %v", i, j, got, want)
+			}
+		}
+	}
+}
+
 func TestImageReplacePixelsAfterClear(t *testing.T) {
 	const w, h = 256, 256
 	img := ebiten.NewImage(w, h)
diff --git a/internal/graphicsdriver/metal/graphics_darwin.go b/internal/graphicsdriver/metal/graphics_darwin.go
index 8acf297e0..f79c59eec 100644
--- a/internal/graphicsdriver/metal/graphics_darwin.go
+++ b/internal/graphicsdriver/metal/graphics_darwin.go
@@ -92,10 +92,8 @@ vertex VertexOut VertexShader(
   return out;
 }
 
-float FloorMod(float x, float y) {
-  if (x < 0.0) {
-    return y - (-x - y * floor(-x/y));
-  }
+float EuclideanMod(float x, float y) {
+  // Assume that y is always positive.
   return x - y * floor(x/y);
 }
 
@@ -111,7 +109,7 @@ template<>
 inline float2 AdjustTexelByAddress<ADDRESS_REPEAT>(float2 p, float4 source_region) {
   float2 o = float2(source_region[0], source_region[1]);
   float2 size = float2(source_region[2] - source_region[0], source_region[3] - source_region[1]);
-  return float2(FloorMod((p.x - o.x), size.x) + o.x, FloorMod((p.y - o.y), size.y) + o.y);
+  return float2(EuclideanMod((p.x - o.x), size.x) + o.x, EuclideanMod((p.y - o.y), size.y) + o.y);
 }
 
 template<uint8_t filter, uint8_t address>
diff --git a/internal/graphicsdriver/opengl/defaultshader.go b/internal/graphicsdriver/opengl/defaultshader.go
index cd999caf6..b63d18259 100644
--- a/internal/graphicsdriver/opengl/defaultshader.go
+++ b/internal/graphicsdriver/opengl/defaultshader.go
@@ -158,13 +158,6 @@ uniform highp float scale;
 varying highp vec2 varying_tex;
 varying highp vec4 varying_color_scale;
 
-highp float floorMod(highp float x, highp float y) {
-  if (x < 0.0) {
-    return y - (-x - y * floor(-x/y));
-  }
-  return x - y * floor(x/y);
-}
-
 highp vec2 adjustTexelByAddress(highp vec2 p, highp vec4 source_region) {
 #if defined(ADDRESS_CLAMP_TO_ZERO)
   return p;
@@ -173,7 +166,7 @@ highp vec2 adjustTexelByAddress(highp vec2 p, highp vec4 source_region) {
 #if defined(ADDRESS_REPEAT)
   highp vec2 o = vec2(source_region[0], source_region[1]);
   highp vec2 size = vec2(source_region[2] - source_region[0], source_region[3] - source_region[1]);
-  return vec2(floorMod((p.x - o.x), size.x) + o.x, floorMod((p.y - o.y), size.y) + o.y);
+  return vec2(mod((p.x - o.x), size.x) + o.x, mod((p.y - o.y), size.y) + o.y);
 #endif
 
 #if defined(ADDRESS_UNSAFE)