Revert "internal/graphics: remove FilterScreen and use an original Kage program instead"

This reverts these commits

* 3259ef3daf
* 7c55065490
* 5bb70f485e

Reason: iOS crash. Probably we have to prepare a special rendering
pipeline for the final screen.

Closes #2278

internal/graphicscommand/command.go

@@ -345,6 +345,8 @@ func (c *drawTrianglesCommand) String() string {
 		filter = "nearest"
 	case graphicsdriver.FilterLinear:
 		filter = "linear"
+	case graphicsdriver.FilterScreen:
+		filter = "screen"
 	default:
 		panic(fmt.Sprintf("graphicscommand: invalid filter: %d", c.filter))
 	}

internal/graphicsdriver/directx/graphics_windows.go

@@ -1239,6 +1239,10 @@ func (g *Graphics) DrawTriangles(dstID graphicsdriver.ImageID, srcs [graphics.Sh
 		var esBody [16]float32
 		var esTranslate [4]float32
 		colorM.Elements(&esBody, &esTranslate)
+		scale := float32(0)
+		if filter == graphicsdriver.FilterScreen {
+			scale = float32(dst.width) / float32(srcImages[0].width)
+		}
 
 		flattenUniforms = []float32{
 			float32(screenWidth),
@@ -1269,6 +1273,7 @@ func (g *Graphics) DrawTriangles(dstID graphicsdriver.ImageID, srcs [graphics.Sh
 			srcRegion.Y,
 			srcRegion.X + srcRegion.Width,
 			srcRegion.Y + srcRegion.Height,
+			scale,
 		}
 	} else {
 		// TODO: This logic is very similar to Metal's. Let's unify them.

internal/graphicsdriver/directx/pipeline_windows.go

@@ -73,6 +73,9 @@ func (k *builtinPipelineStatesKey) defs() ([]_D3D_SHADER_MACRO, error) {
 	case graphicsdriver.FilterLinear:
 		name := []byte("FILTER_LINEAR\x00")
 		defs = append(defs, _D3D_SHADER_MACRO{&name[0], &defval[0]})
+	case graphicsdriver.FilterScreen:
+		name := []byte("FILTER_SCREEN\x00")
+		defs = append(defs, _D3D_SHADER_MACRO{&name[0], &defval[0]})
 	default:
 		return nil, fmt.Errorf("directx: invalid filter: %d", k.filter)
 	}
@@ -110,6 +113,10 @@ cbuffer ShaderParameter : register(b0) {
   float4x4 color_matrix_body;
   float4 color_matrix_translation;
   float4 source_region;
+
+  // This member should be the last not to create a new sector.
+  // https://docs.microsoft.com/en-us/windows/win32/direct3dhlsl/dx-graphics-hlsl-packing-rules
+  float scale;
 }
 
 PSInput VSMain(float2 position : POSITION, float2 tex : TEXCOORD, float4 color : COLOR) {
@@ -213,6 +220,25 @@ float4 PSMain(PSInput input) : SV_TARGET {
   float4 color = lerp(lerp(c0, c1, rate.x), lerp(c2, c3, rate.x), rate.y);
 #endif // defined(FILTER_LINEAR)
 
+#if defined(FILTER_SCREEN)
+  float2 pos = input.texcoord;
+  float2 texel_size = 1.0 / source_size;
+  float2 half_scaled_texel_size = texel_size / 2.0 / scale;
+
+  float2 p0 = pos - half_scaled_texel_size + (texel_size / 512.0);
+  float2 p1 = pos + half_scaled_texel_size + (texel_size / 512.0);
+
+  float4 c0 = tex.Sample(samp, p0);
+  float4 c1 = tex.Sample(samp, float2(p1.x, p0.y));
+  float4 c2 = tex.Sample(samp, float2(p0.x, p1.y));
+  float4 c3 = tex.Sample(samp, p1);
+  // Texels must be in the source rect, so it is not necessary to check that like linear filter.
+
+  float2 rate_center = float2(1.0, 1.0) - half_scaled_texel_size;
+  float2 rate = clamp(((frac(p0 * source_size) - rate_center) * scale) + rate_center, 0.0, 1.0);
+  float4 color = lerp(lerp(c0, c1, rate.x), lerp(c2, c3, rate.x), rate.y);
+#endif // defined(FILTER_SCREEN)
+
 #if defined(USE_COLOR_MATRIX)
   // Un-premultiply alpha.
   // When the alpha is 0, 1.0 - sign(alpha) is 1.0, which means division does nothing.
@@ -226,6 +252,8 @@ float4 PSMain(PSInput input) : SV_TARGET {
   // Clamp the output.
   color.rgb = min(color.rgb, color.a);
   return color;
+#elif defined(FILTER_SCREEN)
+  return color;
 #else
   return input.color * color;
 #endif // defined(USE_COLOR_MATRIX)

internal/graphicsdriver/filter.go

@@ -19,6 +19,7 @@ type Filter int
 const (
 	FilterNearest Filter = iota
 	FilterLinear
+	FilterScreen
 )
 
 type Address int

internal/graphicsdriver/metal/graphics_darwin.go

@@ -33,6 +33,7 @@ const source = `#include <metal_stdlib>
 
 #define FILTER_NEAREST {{.FilterNearest}}
 #define FILTER_LINEAR {{.FilterLinear}}
+#define FILTER_SCREEN {{.FilterScreen}}
 
 #define ADDRESS_CLAMP_TO_ZERO {{.AddressClampToZero}}
 #define ADDRESS_REPEAT {{.AddressRepeat}}
@@ -183,6 +184,25 @@ struct ColorFromTexel<FILTER_LINEAR, address> {
   }
 };
 
+template<uint8_t address>
+struct ColorFromTexel<FILTER_SCREEN, address> {
+  inline float4 Do(VertexOut v, texture2d<float> texture, constant float2& source_size, constant float4& source_region, float scale) {
+    const float2 texel_size = 1 / source_size;
+
+    float2 p0 = v.tex - texel_size / 2.0 / scale + (texel_size / 512.0);
+    float2 p1 = v.tex + texel_size / 2.0 / scale + (texel_size / 512.0);
+
+    float4 c0 = texture.sample(texture_sampler, p0);
+    float4 c1 = texture.sample(texture_sampler, float2(p1.x, p0.y));
+    float4 c2 = texture.sample(texture_sampler, float2(p0.x, p1.y));
+    float4 c3 = texture.sample(texture_sampler, p1);
+
+    float2 rate_center = float2(1.0, 1.0) - texel_size / 2.0 / scale;
+    float2 rate = clamp(((fract(p0 * source_size) - rate_center) * scale) + rate_center, 0.0, 1.0);
+    return mix(mix(c0, c1, rate.x), mix(c2, c3, rate.x), rate.y);
+  }
+};
+
 template<bool useColorM, uint8_t filter, uint8_t address>
 struct FragmentShaderImpl {
   inline float4 Do(
@@ -207,6 +227,20 @@ struct FragmentShaderImpl {
   }
 };
 
+template<bool useColorM, uint8_t address>
+struct FragmentShaderImpl<useColorM, FILTER_SCREEN, address> {
+  inline float4 Do(
+      VertexOut v,
+      texture2d<float> texture,
+      constant float2& source_size,
+      constant float4x4& color_matrix_body,
+      constant float4& color_matrix_translation,
+      constant float4& source_region,
+      constant float& scale) {
+    return ColorFromTexel<FILTER_SCREEN, address>().Do(v, texture, source_size, source_region, scale);
+  }
+};
+
 // Define Foo and FooCp macros to force macro replacement.
 // See "6.10.3.1 Argument substitution" in ISO/IEC 9899.
 
@@ -239,6 +273,8 @@ FragmentShaderFunc(1, FILTER_LINEAR, ADDRESS_REPEAT)
 FragmentShaderFunc(1, FILTER_NEAREST, ADDRESS_UNSAFE)
 FragmentShaderFunc(1, FILTER_LINEAR, ADDRESS_UNSAFE)
 
+FragmentShaderFunc(0, FILTER_SCREEN, ADDRESS_UNSAFE)
+
 #undef FragmentShaderFuncName
 `
 
@@ -572,6 +608,7 @@ func (g *Graphics) Initialize() error {
 	replaces := map[string]string{
 		"{{.FilterNearest}}":      fmt.Sprintf("%d", graphicsdriver.FilterNearest),
 		"{{.FilterLinear}}":       fmt.Sprintf("%d", graphicsdriver.FilterLinear),
+		"{{.FilterScreen}}":       fmt.Sprintf("%d", graphicsdriver.FilterScreen),
 		"{{.AddressClampToZero}}": fmt.Sprintf("%d", graphicsdriver.AddressClampToZero),
 		"{{.AddressRepeat}}":      fmt.Sprintf("%d", graphicsdriver.AddressRepeat),
 		"{{.AddressUnsafe}}":      fmt.Sprintf("%d", graphicsdriver.AddressUnsafe),
@@ -589,6 +626,27 @@ func (g *Graphics) Initialize() error {
 	if err != nil {
 		return err
 	}
+	fs, err := lib.MakeFunction(
+		fmt.Sprintf("FragmentShader_%d_%d_%d", 0, graphicsdriver.FilterScreen, graphicsdriver.AddressUnsafe))
+	if err != nil {
+		return err
+	}
+	rpld := mtl.RenderPipelineDescriptor{
+		VertexFunction:   vs,
+		FragmentFunction: fs,
+	}
+	rpld.ColorAttachments[0].PixelFormat = g.view.colorPixelFormat()
+	rpld.ColorAttachments[0].BlendingEnabled = true
+	rpld.ColorAttachments[0].DestinationAlphaBlendFactor = mtl.BlendFactorZero
+	rpld.ColorAttachments[0].DestinationRGBBlendFactor = mtl.BlendFactorZero
+	rpld.ColorAttachments[0].SourceAlphaBlendFactor = mtl.BlendFactorOne
+	rpld.ColorAttachments[0].SourceRGBBlendFactor = mtl.BlendFactorOne
+	rpld.ColorAttachments[0].WriteMask = mtl.ColorWriteMaskAll
+	rps, err := g.view.getMTLDevice().MakeRenderPipelineState(rpld)
+	if err != nil {
+		return err
+	}
+	g.screenRPS = rps
 
 	for _, screen := range []bool{false, true} {
 		for _, cm := range []bool{false, true} {
@@ -813,6 +871,9 @@ func (g *Graphics) DrawTriangles(dstID graphicsdriver.ImageID, srcIDs [graphics.
 	rpss := map[stencilMode]mtl.RenderPipelineState{}
 	var uniformVars [][]float32
 	if shaderID == graphicsdriver.InvalidShaderID {
+		if dst.screen && filter == graphicsdriver.FilterScreen {
+			rpss[noStencil] = g.screenRPS
+		} else {
 			for _, stencil := range []stencilMode{
 				prepareStencil,
 				drawWithStencil,
@@ -827,6 +888,7 @@ func (g *Graphics) DrawTriangles(dstID graphicsdriver.ImageID, srcIDs [graphics.
 					stencilMode:   stencil,
 				}]
 			}
+		}
 
 		w, h := dst.internalSize()
 		sourceSize := []float32{0, 0}
@@ -838,6 +900,10 @@ func (g *Graphics) DrawTriangles(dstID graphicsdriver.ImageID, srcIDs [graphics.
 		var esBody [16]float32
 		var esTranslate [4]float32
 		colorM.Elements(&esBody, &esTranslate)
+		scale := float32(0)
+		if filter == graphicsdriver.FilterScreen {
+			scale = float32(dst.width) / float32(srcs[0].width)
+		}
 		uniformVars = [][]float32{
 			{float32(w), float32(h)},
 			sourceSize,
@@ -849,6 +915,7 @@ func (g *Graphics) DrawTriangles(dstID graphicsdriver.ImageID, srcIDs [graphics.
 				srcRegion.X + srcRegion.Width,
 				srcRegion.Y + srcRegion.Height,
 			},
+			{scale},
 		}
 	} else {
 		for _, stencil := range []stencilMode{

internal/graphicsdriver/opengl/defaultshader.go

@@ -83,6 +83,8 @@ func fragmentShaderStr(useColorM bool, filter graphicsdriver.Filter, address gra
 		defs = append(defs, "#define FILTER_NEAREST")
 	case graphicsdriver.FilterLinear:
 		defs = append(defs, "#define FILTER_LINEAR")
+	case graphicsdriver.FilterScreen:
+		defs = append(defs, "#define FILTER_SCREEN")
 	default:
 		panic(fmt.Sprintf("opengl: invalid filter: %d", filter))
 	}
@@ -149,6 +151,10 @@ uniform vec4 color_matrix_translation;
 
 uniform highp vec2 source_size;
 
+#if defined(FILTER_SCREEN)
+uniform highp float scale;
+#endif
+
 varying highp vec2 varying_tex;
 varying highp vec4 varying_color_scale;
 
@@ -229,6 +235,27 @@ void main(void) {
   color = mix(mix(c0, c1, rate.x), mix(c2, c3, rate.x), rate.y);
 #endif  // defined(FILTER_LINEAR)
 
+#if defined(FILTER_SCREEN)
+  highp vec2 texel_size = 1.0 / source_size;
+  highp vec2 half_scaled_texel_size = texel_size / 2.0 / scale;
+
+  highp vec2 p0 = pos - half_scaled_texel_size + (texel_size / 512.0);
+  highp vec2 p1 = pos + half_scaled_texel_size + (texel_size / 512.0);
+
+  vec4 c0 = texture2D(T0, p0);
+  vec4 c1 = texture2D(T0, vec2(p1.x, p0.y));
+  vec4 c2 = texture2D(T0, vec2(p0.x, p1.y));
+  vec4 c3 = texture2D(T0, p1);
+  // Texels must be in the source rect, so it is not necessary to check that like linear filter.
+
+  vec2 rate_center = vec2(1.0, 1.0) - half_scaled_texel_size;
+  vec2 rate = clamp(((fract(p0 * source_size) - rate_center) * scale) + rate_center, 0.0, 1.0);
+  gl_FragColor = mix(mix(c0, c1, rate.x), mix(c2, c3, rate.x), rate.y);
+
+  // Assume that a color matrix and color vector values are not used with FILTER_SCREEN.
+
+#else
+
 # if defined(USE_COLOR_MATRIX)
   // Un-premultiply alpha.
   // When the alpha is 0, 1.0 - sign(alpha) is 1.0, which means division does nothing.
@@ -248,6 +275,9 @@ void main(void) {
 # endif  // defined(USE_COLOR_MATRIX)
 
   gl_FragColor = color;
+
+#endif  // defined(FILTER_SCREEN)
+
 }
 `
 )

internal/graphicsdriver/opengl/graphics.go

@@ -244,6 +244,15 @@ func (g *Graphics) DrawTriangles(dstID graphicsdriver.ImageID, srcIDs [graphics.
 				typ:   shaderir.Type{Main: shaderir.Vec2},
 			})
 		}
+
+		if filter == graphicsdriver.FilterScreen {
+			scale := float32(destination.width) / float32(g.images[srcIDs[0]].width)
+			g.uniformVars = append(g.uniformVars, uniformVariable{
+				name:  "scale",
+				value: []float32{scale},
+				typ:   shaderir.Type{Main: shaderir.Float},
+			})
+		}
 	} else {
 		shader := g.shaders[shaderID]
 		program = shader.p

internal/graphicsdriver/opengl/program.go

@@ -191,6 +191,7 @@ func (s *openGLState) reset(context *context) error {
 			for _, f := range []graphicsdriver.Filter{
 				graphicsdriver.FilterNearest,
 				graphicsdriver.FilterLinear,
+				graphicsdriver.FilterScreen,
 			} {
 				shaderFragmentColorMatrixNative, err := context.newFragmentShader(fragmentShaderStr(c, f, a))
 				if err != nil {

internal/mipmap/mipmap.go

@@ -65,7 +65,7 @@ func (m *Mipmap) DrawTriangles(srcs [graphics.ShaderImageCount]*Mipmap, vertices
 
 	level := 0
 	// TODO: Do we need to check all the sources' states of being volatile?
-	if !canSkipMipmap && srcs[0] != nil && !srcs[0].volatile {
+	if !canSkipMipmap && srcs[0] != nil && !srcs[0].volatile && filter != graphicsdriver.FilterScreen {
 		level = math.MaxInt32
 		for i := 0; i < len(indices)/3; i++ {
 			const n = graphics.VertexFloatCount
@@ -226,6 +226,10 @@ func (m *Mipmap) disposeMipmaps() {
 func mipmapLevelFromDistance(dx0, dy0, dx1, dy1, sx0, sy0, sx1, sy1 float32, filter graphicsdriver.Filter) int {
 	const maxLevel = 6
 
+	if filter == graphicsdriver.FilterScreen {
+		return 0
+	}
+
 	d := (dx1-dx0)*(dx1-dx0) + (dy1-dy0)*(dy1-dy0)
 	s := (sx1-sx0)*(sx1-sx0) + (sy1-sy0)*(sy1-sy0)
 	if s == 0 {

internal/ui/context.go

@@ -30,41 +30,6 @@ import (
 	"github.com/hajimehoshi/ebiten/v2/internal/hooks"
 )
 
-const screenShader = `package main
-
-var Scale float
-
-func Fragment(position vec4, texCoord vec2, color vec4) vec4 {
-	sourceSize := imageSrcTextureSize()
-	// texelSize is one pixel size in texel sizes.
-	texelSize := 1 / sourceSize
-	halfScaledTexelSize := texelSize / 2 / Scale
-
-	// Shift 1/512 [texel] to avoid the tie-breaking issue.
-	// As all the vertex positions are aligned to 1/16 [pixel], this shiting should work in most cases.
-	pos := texCoord
-	p0 := pos - halfScaledTexelSize + (texelSize / 512)
-	p1 := pos + halfScaledTexelSize + (texelSize / 512)
-
-	// Texels must be in the source rect, so it is not necessary to check.
-	c0 := imageSrc0UnsafeAt(p0)
-	c1 := imageSrc0UnsafeAt(vec2(p1.x, p0.y))
-	c2 := imageSrc0UnsafeAt(vec2(p0.x, p1.y))
-	c3 := imageSrc0UnsafeAt(p1)
-
-	// p is the p1 value in one pixel assuming that the pixel's upper-left is (0, 0) and the lower-right is (1, 1).
-	p := fract(p1 * sourceSize)
-
-	// rate indicates how much the 4 colors are mixed. rate is in between [0, 1].
-	//
-	//     0 <= p <= 1/Scale: The rate is in between [0, 1]
-	//     1/Scale < p:       Don't care. Adjacent colors (e.g. c0 vs c1 in an X direction) should be the same.
-	rate := clamp(p*Scale, 0, 1)
-
-	return mix(mix(c0, c1, rate.x), mix(c2, c3, rate.x), rate.y)
-}
-`
-
 type Game interface {
 	NewOffscreenImage(width, height int) *Image
 	Layout(outsideWidth, outsideHeight int) (int, int)
@@ -84,8 +49,6 @@ type context struct {
 	outsideWidth  float64
 	outsideHeight float64
 
-	screenShader *Shader
-
 	m sync.Mutex
 }
 
@@ -145,15 +108,6 @@ func (c *context) updateFrameImpl(graphicsDriver graphicsdriver.Graphics, update
 		}
 	}()
 
-	// Create a shader for the screen if necessary.
-	if c.screenShader == nil {
-		ir, err := graphics.CompileShader([]byte(screenShader))
-		if err != nil {
-			return err
-		}
-		c.screenShader = NewShader(ir)
-	}
-
 	// ForceUpdate can be invoked even if the context is not initialized yet (#1591).
 	if w, h := c.layoutGame(outsideWidth, outsideHeight, deviceScaleFactor); w == 0 || h == 0 {
 		return nil
@@ -232,16 +186,15 @@ func (c *context) drawGame(graphicsDriver graphicsdriver.Graphics) {
 	gty += offsetY
 
 	var filter graphicsdriver.Filter
-	var screenFilter bool
 	switch {
 	case !theGlobalState.isScreenFilterEnabled():
 		filter = graphicsdriver.FilterNearest
 	case math.Floor(s) == s:
 		filter = graphicsdriver.FilterNearest
 	case s > 1:
-		screenFilter = true
+		filter = graphicsdriver.FilterScreen
 	default:
-		// screenShader works with >=1 scale, but does not well with <1 scale.
+		// FilterScreen works with >=1 scale, but does not well with <1 scale.
 		// Use regular FilterLinear instead so far (#669).
 		filter = graphicsdriver.FilterLinear
 	}
@@ -260,17 +213,7 @@ func (c *context) drawGame(graphicsDriver graphicsdriver.Graphics) {
 	is := graphics.QuadIndices()
 
 	srcs := [graphics.ShaderImageCount]*Image{c.offscreen}
-	var shader *Shader
+	c.screen.DrawTriangles(srcs, vs, is, affine.ColorMIdentity{}, graphicsdriver.CompositeModeCopy, filter, graphicsdriver.AddressUnsafe, dstRegion, graphicsdriver.Region{}, [graphics.ShaderImageCount - 1][2]float32{}, nil, nil, false, true)
-	var uniforms [][]float32
-	if screenFilter {
-		shader = c.screenShader
-		dstWidth := c.screen.width
-		srcWidth := c.offscreen.width
-		uniforms = shader.ConvertUniforms(map[string]interface{}{
-			"Scale": float32(dstWidth) / float32(srcWidth),
-		})
-	}
-	c.screen.DrawTriangles(srcs, vs, is, affine.ColorMIdentity{}, graphicsdriver.CompositeModeCopy, filter, graphicsdriver.AddressUnsafe, dstRegion, graphicsdriver.Region{}, [graphics.ShaderImageCount - 1][2]float32{}, shader, uniforms, false, true)
 }
 
 func (c *context) layoutGame(outsideWidth, outsideHeight float64, deviceScaleFactor float64) (int, int) {