diff --git a/backend/goglbackend/fill.go b/backend/goglbackend/fill.go
index f9b6b08..39d67c2 100644
--- a/backend/goglbackend/fill.go
+++ b/backend/goglbackend/fill.go
@@ -245,6 +245,10 @@ func (b *GoGLBackend) drawBlurred(size float64, min, max vec) {
 	min[1] -= fsize * 3
 	max[0] += fsize * 3
 	max[1] += fsize * 3
+	min[0] = math.Max(0.0, math.Min(b.fw, min[0]))
+	min[1] = math.Max(0.0, math.Min(b.fh, min[1]))
+	max[0] = math.Max(0.0, math.Min(b.fw, max[0]))
+	max[1] = math.Max(0.0, math.Min(b.fh, max[1]))
 
 	gl.BindBuffer(gl.ARRAY_BUFFER, b.shadowBuf)
 	data := [16]float32{
@@ -276,27 +280,27 @@ func (b *GoGLBackend) drawBlurred(size float64, min, max vec) {
 
 	gl.ClearColor(0, 0, 0, 0)
 
-	gl.BindTexture(gl.TEXTURE_2D, b.offscr1.tex)
 	b.enableTextureRenderTarget(&b.offscr2)
-	gl.Clear(gl.COLOR_BUFFER_BIT)
-	b.box3(sizea, false)
-	gl.BindTexture(gl.TEXTURE_2D, b.offscr2.tex)
+	gl.BindTexture(gl.TEXTURE_2D, b.offscr1.tex)
+	gl.Clear(gl.COLOR_BUFFER_BIT | gl.STENCIL_BUFFER_BIT)
+	b.box3(sizea, 0, false)
 	b.enableTextureRenderTarget(&b.offscr1)
-	b.box3(sizeb, false)
-	gl.BindTexture(gl.TEXTURE_2D, b.offscr1.tex)
-	b.enableTextureRenderTarget(&b.offscr2)
-	b.box3(sizec, false)
 	gl.BindTexture(gl.TEXTURE_2D, b.offscr2.tex)
-	b.enableTextureRenderTarget(&b.offscr1)
-	b.box3(sizea, true)
-	gl.BindTexture(gl.TEXTURE_2D, b.offscr1.tex)
+	b.box3(sizeb, -0.5, false)
 	b.enableTextureRenderTarget(&b.offscr2)
-	b.box3(sizeb, true)
+	gl.BindTexture(gl.TEXTURE_2D, b.offscr1.tex)
+	b.box3(sizec, 0, false)
+	b.enableTextureRenderTarget(&b.offscr1)
+	gl.BindTexture(gl.TEXTURE_2D, b.offscr2.tex)
+	b.box3(sizea, 0, true)
+	b.enableTextureRenderTarget(&b.offscr2)
+	gl.BindTexture(gl.TEXTURE_2D, b.offscr1.tex)
+	b.box3(sizeb, -0.5, true)
 	gl.Enable(gl.BLEND)
 	gl.BlendFunc(gl.ONE, gl.ONE_MINUS_SRC_ALPHA)
-	gl.BindTexture(gl.TEXTURE_2D, b.offscr2.tex)
 	b.disableTextureRenderTarget()
-	b.box3(sizec, true)
+	gl.BindTexture(gl.TEXTURE_2D, b.offscr2.tex)
+	b.box3(sizec, 0, true)
 
 	gl.DisableVertexAttribArray(b.shd.Vertex)
 	gl.DisableVertexAttribArray(b.shd.TexCoord)
@@ -304,7 +308,7 @@ func (b *GoGLBackend) drawBlurred(size float64, min, max vec) {
 	gl.BlendFunc(gl.SRC_ALPHA, gl.ONE_MINUS_SRC_ALPHA)
 }
 
-func (b *GoGLBackend) box3(size int, vertical bool) {
+func (b *GoGLBackend) box3(size int, offset float32, vertical bool) {
 	gl.Uniform1i(b.shd.BoxSize, int32(size))
 	if vertical {
 		gl.Uniform1i(b.shd.BoxVertical, 1)
@@ -313,5 +317,6 @@ func (b *GoGLBackend) box3(size int, vertical bool) {
 		gl.Uniform1i(b.shd.BoxVertical, 0)
 		gl.Uniform1f(b.shd.BoxScale, 1/float32(b.fw))
 	}
+	gl.Uniform1f(b.shd.BoxOffset, offset)
 	gl.DrawArrays(gl.TRIANGLE_FAN, 0, 4)
 }
diff --git a/backend/goglbackend/gogl.go b/backend/goglbackend/gogl.go
index f85f695..e9c4896 100644
--- a/backend/goglbackend/gogl.go
+++ b/backend/goglbackend/gogl.go
@@ -139,6 +139,7 @@ func New(x, y, w, h int, ctx *GLContext) (*GoGLBackend, error) {
 	}
 	b.disableTextureRenderTarget = func() {
 		gl.BindFramebuffer(gl.FRAMEBUFFER, 0)
+		gl.Viewport(int32(b.x), int32(b.y), int32(b.w), int32(b.h))
 	}
 
 	return b, nil
@@ -288,28 +289,29 @@ func colorGoToGL(c color.RGBA) glColor {
 }
 
 func (b *GoGLBackend) useShader(style *backendbase.FillStyle, useAlpha bool, alphaTexSlot int32) (vertexLoc, alphaTexCoordLoc uint32) {
-	var alphaVal int32
+	gl.UseProgram(b.shd.ID)
+	gl.Uniform2f(b.shd.CanvasSize, float32(b.fw), float32(b.fh))
 	if useAlpha {
-		alphaVal = 1
+		gl.Uniform1i(b.shd.UseAlphaTex, 1)
+		gl.Uniform1i(b.shd.AlphaTex, alphaTexSlot)
+	} else {
+		gl.Uniform1i(b.shd.UseAlphaTex, 0)
 	}
+	gl.Uniform1f(b.shd.GlobalAlpha, float32(style.Color.A)/255)
+
 	if lg := style.LinearGradient; lg != nil {
 		lg := lg.(*LinearGradient)
 		gl.ActiveTexture(gl.TEXTURE0)
 		gl.BindTexture(gl.TEXTURE_2D, lg.tex)
-		gl.UseProgram(b.shd.ID)
 		from := vec{style.Gradient.X0, style.Gradient.Y0}
 		to := vec{style.Gradient.X1, style.Gradient.Y1}
 		dir := to.sub(from)
 		length := dir.len()
 		dir = dir.scale(1 / length)
-		gl.Uniform2f(b.shd.CanvasSize, float32(b.fw), float32(b.fh))
 		gl.Uniform2f(b.shd.From, float32(from[0]), float32(from[1]))
 		gl.Uniform2f(b.shd.Dir, float32(dir[0]), float32(dir[1]))
 		gl.Uniform1f(b.shd.Len, float32(length))
 		gl.Uniform1i(b.shd.Gradient, 0)
-		gl.Uniform1f(b.shd.GlobalAlpha, float32(style.Color.A)/255)
-		gl.Uniform1i(b.shd.AlphaTex, alphaTexSlot)
-		gl.Uniform1i(b.shd.UseAlphaTex, alphaVal)
 		gl.Uniform1i(b.shd.Func, shdFuncLinearGradient)
 		return b.shd.Vertex, b.shd.TexCoord
 	}
@@ -317,28 +319,21 @@ func (b *GoGLBackend) useShader(style *backendbase.FillStyle, useAlpha bool, alp
 		rg := rg.(*RadialGradient)
 		gl.ActiveTexture(gl.TEXTURE0)
 		gl.BindTexture(gl.TEXTURE_2D, rg.tex)
-		gl.UseProgram(b.shd.ID)
 		from := vec{style.Gradient.X0, style.Gradient.Y0}
 		to := vec{style.Gradient.X1, style.Gradient.Y1}
-		gl.Uniform2f(b.shd.CanvasSize, float32(b.fw), float32(b.fh))
 		gl.Uniform2f(b.shd.From, float32(from[0]), float32(from[1]))
 		gl.Uniform2f(b.shd.To, float32(to[0]), float32(to[1]))
 		gl.Uniform1f(b.shd.RadFrom, float32(style.Gradient.RadFrom))
 		gl.Uniform1f(b.shd.RadTo, float32(style.Gradient.RadTo))
 		gl.Uniform1i(b.shd.Gradient, 0)
-		gl.Uniform1f(b.shd.GlobalAlpha, float32(style.Color.A)/255)
-		gl.Uniform1i(b.shd.AlphaTex, alphaTexSlot)
-		gl.Uniform1i(b.shd.UseAlphaTex, alphaVal)
 		gl.Uniform1i(b.shd.Func, shdFuncRadialGradient)
 		return b.shd.Vertex, b.shd.TexCoord
 	}
 	if ip := style.ImagePattern; ip != nil {
 		ipd := ip.(*ImagePattern).data
 		img := ipd.Image.(*Image)
-		gl.UseProgram(b.shd.ID)
 		gl.ActiveTexture(gl.TEXTURE0)
 		gl.BindTexture(gl.TEXTURE_2D, img.tex)
-		gl.Uniform2f(b.shd.CanvasSize, float32(b.fw), float32(b.fh))
 		gl.Uniform2f(b.shd.ImageSize, float32(img.w), float32(img.h))
 		gl.Uniform1i(b.shd.Image, 0)
 		var f32mat [9]float32
@@ -356,20 +351,13 @@ func (b *GoGLBackend) useShader(style *backendbase.FillStyle, useAlpha bool, alp
 		case backendbase.NoRepeat:
 			gl.Uniform2f(b.shd.Repeat, 0, 0)
 		}
-		gl.Uniform1f(b.shd.GlobalAlpha, float32(style.Color.A)/255)
-		gl.Uniform1i(b.shd.AlphaTex, alphaTexSlot)
-		gl.Uniform1i(b.shd.UseAlphaTex, alphaVal)
 		gl.Uniform1i(b.shd.Func, shdFuncImagePattern)
 		return b.shd.Vertex, b.shd.TexCoord
 	}
 
-	gl.UseProgram(b.shd.ID)
-	gl.Uniform2f(b.shd.CanvasSize, float32(b.fw), float32(b.fh))
 	c := colorGoToGL(style.Color)
 	gl.Uniform4f(b.shd.Color, float32(c.r), float32(c.g), float32(c.b), float32(c.a))
 	gl.Uniform1f(b.shd.GlobalAlpha, 1)
-	gl.Uniform1i(b.shd.AlphaTex, alphaTexSlot)
-	gl.Uniform1i(b.shd.UseAlphaTex, alphaVal)
 	gl.Uniform1i(b.shd.Func, shdFuncSolid)
 	return b.shd.Vertex, b.shd.TexCoord
 }
diff --git a/backend/goglbackend/shaders.go b/backend/goglbackend/shaders.go
index fa6b183..ae29d70 100755
--- a/backend/goglbackend/shaders.go
+++ b/backend/goglbackend/shaders.go
@@ -41,6 +41,7 @@ uniform sampler2D alphaTex;
 uniform int boxSize;
 uniform bool boxVertical;
 uniform float boxScale;
+uniform float boxOffset;
 
 bool isNaN(float v) {
   return v < 0.0 || 0.0 < v || v == 0.0 ? false : true;
@@ -52,17 +53,17 @@ void main() {
 	if (func == 5) {
 		vec4 sum = vec4(0.0);
 		if (boxVertical) {
-			vec2 start = v_tc - vec2(0.0, (float)(boxSize) * boxScale);
-			for (int i=0; i<boxSize*2; i++) {
-				sum += texture2D(image, start + vec2(0.0, (float)(i) * boxScale));
+			vec2 start = v_tc - vec2(0.0, (float(boxSize) * 0.5 + boxOffset) * boxScale);
+			for (int i=0; i <= boxSize; i++) {
+				sum += texture2D(image, start + vec2(0.0, float(i) * boxScale));
 			}
 		} else {
-			vec2 start = v_tc - vec2((float)(boxSize) * boxScale, 0.0);
-			for (int i=0; i<boxSize*2; i++) {
-				sum += texture2D(image, start + vec2((float)(i) * boxScale, 0.0));
+			vec2 start = v_tc - vec2((float(boxSize) * 0.5 + boxOffset) * boxScale, 0.0);
+			for (int i=0; i <= boxSize; i++) {
+				sum += texture2D(image, start + vec2(float(i) * boxScale, 0.0));
 			}
 		}
-		gl_FragColor = sum / float(boxSize * 2);
+		gl_FragColor = sum / float(boxSize+1);
 		return;
 	}
 
@@ -149,4 +150,5 @@ type unifiedShader struct {
 	BoxSize     int32
 	BoxVertical int32
 	BoxScale    int32
+	BoxOffset   int32
 }
diff --git a/testdata/Shadow.png b/testdata/Shadow.png
index a8719bc..82c4aa4 100755
Binary files a/testdata/Shadow.png and b/testdata/Shadow.png differ