diff --git a/libs/hwui/Caches.cpp b/libs/hwui/Caches.cpp
index f0f72f9..2883f37 100644
--- a/libs/hwui/Caches.cpp
+++ b/libs/hwui/Caches.cpp
@@ -68,6 +68,7 @@
     mCurrentBuffer = meshBuffer;
     mCurrentIndicesBuffer = 0;
     mCurrentPositionPointer = this;
+    mCurrentPositionStride = 0;
     mCurrentTexCoordsPointer = this;
 
     mTexCoordsArrayEnabled = false;
@@ -340,15 +341,18 @@
 // Meshes and textures
 ///////////////////////////////////////////////////////////////////////////////
 
-void Caches::bindPositionVertexPointer(bool force, GLuint slot, GLvoid* vertices, GLsizei stride) {
-    if (force || vertices != mCurrentPositionPointer) {
+void Caches::bindPositionVertexPointer(bool force, GLvoid* vertices, GLsizei stride) {
+    if (force || vertices != mCurrentPositionPointer || stride != mCurrentPositionStride) {
+        GLuint slot = currentProgram->position;
         glVertexAttribPointer(slot, 2, GL_FLOAT, GL_FALSE, stride, vertices);
         mCurrentPositionPointer = vertices;
+        mCurrentPositionStride = stride;
     }
 }
 
-void Caches::bindTexCoordsVertexPointer(bool force, GLuint slot, GLvoid* vertices) {
+void Caches::bindTexCoordsVertexPointer(bool force, GLvoid* vertices) {
     if (force || vertices != mCurrentTexCoordsPointer) {
+        GLuint slot = currentProgram->texCoords;
         glVertexAttribPointer(slot, 2, GL_FLOAT, GL_FALSE, gMeshStride, vertices);
         mCurrentTexCoordsPointer = vertices;
     }
diff --git a/libs/hwui/Caches.h b/libs/hwui/Caches.h
index 48efd10..9f28409 100644
--- a/libs/hwui/Caches.h
+++ b/libs/hwui/Caches.h
@@ -173,14 +173,13 @@
      * Binds an attrib to the specified float vertex pointer.
      * Assumes a stride of gMeshStride and a size of 2.
      */
-    void bindPositionVertexPointer(bool force, GLuint slot, GLvoid* vertices,
-            GLsizei stride = gMeshStride);
+    void bindPositionVertexPointer(bool force, GLvoid* vertices, GLsizei stride = gMeshStride);
 
     /**
      * Binds an attrib to the specified float vertex pointer.
      * Assumes a stride of gMeshStride and a size of 2.
      */
-    void bindTexCoordsVertexPointer(bool force, GLuint slot, GLvoid* vertices);
+    void bindTexCoordsVertexPointer(bool force, GLvoid* vertices);
 
     /**
      * Resets the vertex pointers.
@@ -295,6 +294,7 @@
     GLuint mCurrentBuffer;
     GLuint mCurrentIndicesBuffer;
     void* mCurrentPositionPointer;
+    GLuint mCurrentPositionStride;
     void* mCurrentTexCoordsPointer;
 
     bool mTexCoordsArrayEnabled;
diff --git a/libs/hwui/FontRenderer.cpp b/libs/hwui/FontRenderer.cpp
index cab68f0..4e97c88 100644
--- a/libs/hwui/FontRenderer.cpp
+++ b/libs/hwui/FontRenderer.cpp
@@ -374,9 +374,8 @@
         int offset = 2;
 
         bool force = caches.unbindMeshBuffer();
-        caches.bindPositionVertexPointer(force, caches.currentProgram->position, buffer);
-        caches.bindTexCoordsVertexPointer(force, caches.currentProgram->texCoords,
-                buffer + offset);
+        caches.bindPositionVertexPointer(force, buffer);
+        caches.bindTexCoordsVertexPointer(force, buffer + offset);
     }
 
     glDrawElements(GL_TRIANGLES, mCurrentQuadIndex * 6, GL_UNSIGNED_SHORT, NULL);
diff --git a/libs/hwui/OpenGLRenderer.cpp b/libs/hwui/OpenGLRenderer.cpp
index 87c3a47..b0328f5 100644
--- a/libs/hwui/OpenGLRenderer.cpp
+++ b/libs/hwui/OpenGLRenderer.cpp
@@ -1248,6 +1248,15 @@
     return !clip.intersects(transformed);
 }
 
+bool OpenGLRenderer::quickRejectPreStroke(float left, float top, float right, float bottom, SkPaint* paint) {
+    if (paint->getStyle() != SkPaint::kFill_Style) {
+        float outset = paint->getStrokeWidth() * 0.5f;
+        return quickReject(left - outset, top - outset, right + outset, bottom + outset);
+    } else {
+        return quickReject(left, top, right, bottom);
+    }
+}
+
 bool OpenGLRenderer::quickReject(float left, float top, float right, float bottom) {
     if (mSnapshot->isIgnored()) {
         return true;
@@ -1490,7 +1499,7 @@
 
 void OpenGLRenderer::setupDrawSimpleMesh() {
     bool force = mCaches.bindMeshBuffer();
-    mCaches.bindPositionVertexPointer(force, mCaches.currentProgram->position, 0);
+    mCaches.bindPositionVertexPointer(force, 0);
     mCaches.unbindIndicesBuffer();
 }
 
@@ -1523,9 +1532,9 @@
         force = mCaches.unbindMeshBuffer();
     }
 
-    mCaches.bindPositionVertexPointer(force, mCaches.currentProgram->position, vertices);
+    mCaches.bindPositionVertexPointer(force, vertices);
     if (mCaches.currentProgram->texCoords >= 0) {
-        mCaches.bindTexCoordsVertexPointer(force, mCaches.currentProgram->texCoords, texCoords);
+        mCaches.bindTexCoordsVertexPointer(force, texCoords);
     }
 
     mCaches.unbindIndicesBuffer();
@@ -1533,16 +1542,15 @@
 
 void OpenGLRenderer::setupDrawMeshIndices(GLvoid* vertices, GLvoid* texCoords) {
     bool force = mCaches.unbindMeshBuffer();
-    mCaches.bindPositionVertexPointer(force, mCaches.currentProgram->position, vertices);
+    mCaches.bindPositionVertexPointer(force, vertices);
     if (mCaches.currentProgram->texCoords >= 0) {
-        mCaches.bindTexCoordsVertexPointer(force, mCaches.currentProgram->texCoords, texCoords);
+        mCaches.bindTexCoordsVertexPointer(force, texCoords);
     }
 }
 
 void OpenGLRenderer::setupDrawVertices(GLvoid* vertices) {
     bool force = mCaches.unbindMeshBuffer();
-    mCaches.bindPositionVertexPointer(force, mCaches.currentProgram->position,
-            vertices, gVertexStride);
+    mCaches.bindPositionVertexPointer(force, vertices, gVertexStride);
     mCaches.unbindIndicesBuffer();
 }
 
@@ -1560,8 +1568,7 @@
 void OpenGLRenderer::setupDrawAALine(GLvoid* vertices, GLvoid* widthCoords,
         GLvoid* lengthCoords, float boundaryWidthProportion, int& widthSlot, int& lengthSlot) {
     bool force = mCaches.unbindMeshBuffer();
-    mCaches.bindPositionVertexPointer(force, mCaches.currentProgram->position,
-            vertices, gAAVertexStride);
+    mCaches.bindPositionVertexPointer(force, vertices, gAAVertexStride);
     mCaches.resetTexCoordsVertexPointer();
     mCaches.unbindIndicesBuffer();
 
@@ -1919,15 +1926,23 @@
 }
 
 /**
- * This function uses a similar approach to that of AA lines in the drawLines() function.
- * We expand the rectangle by a half pixel in screen space on all sides. However, instead of using
- * a fragment shader to compute the translucency of the color from its position, we simply use a
- * varying parameter to define how far a given pixel is into the region.
+ * Renders a convex path via tessellation. For AA paths, this function uses a similar approach to
+ * that of AA lines in the drawLines() function.  We expand the convex path by a half pixel in
+ * screen space in all directions. However, instead of using a fragment shader to compute the
+ * translucency of the color from its position, we simply use a varying parameter to define how far
+ * a given pixel is from the edge. For non-AA paths, the expansion and alpha varying are not used.
+ *
+ * Doesn't yet support joins, caps, or path effects.
  */
-void OpenGLRenderer::drawConvexPath(const SkPath& path, int color, SkXfermode::Mode mode, bool isAA) {
+void OpenGLRenderer::drawConvexPath(const SkPath& path, SkPaint* paint) {
+    int color = paint->getColor();
+    SkPaint::Style style = paint->getStyle();
+    SkXfermode::Mode mode = getXfermode(paint->getXfermode());
+    bool isAA = paint->isAntiAlias();
+
     VertexBuffer vertexBuffer;
     // TODO: try clipping large paths to viewport
-    PathRenderer::convexPathFillVertices(path, mSnapshot->transform, vertexBuffer, isAA);
+    PathRenderer::convexPathVertices(path, paint, mSnapshot->transform, vertexBuffer);
 
     setupDraw();
     setupDrawNoTexture();
@@ -1938,15 +1953,14 @@
     setupDrawShader();
     setupDrawBlending(isAA, mode);
     setupDrawProgram();
-    setupDrawModelViewIdentity(true);
+    setupDrawModelViewIdentity();
     setupDrawColorUniforms();
     setupDrawColorFilterUniforms();
     setupDrawShaderIdentityUniforms();
 
     void* vertices = vertexBuffer.getBuffer();
     bool force = mCaches.unbindMeshBuffer();
-    mCaches.bindPositionVertexPointer(force, mCaches.currentProgram->position,
-                                      vertices, isAA ? gAlphaVertexStride : gVertexStride);
+    mCaches.bindPositionVertexPointer(true, vertices, isAA ? gAlphaVertexStride : gVertexStride);
     mCaches.resetTexCoordsVertexPointer();
     mCaches.unbindIndicesBuffer();
 
@@ -1960,7 +1974,7 @@
         glVertexAttribPointer(alphaSlot, 1, GL_FLOAT, GL_FALSE, gAlphaVertexStride, alphaCoords);
     }
 
-    SkRect bounds = path.getBounds();
+    SkRect bounds = PathRenderer::computePathBounds(path, paint);
     dirtyLayer(bounds.fLeft, bounds.fTop, bounds.fRight, bounds.fBottom, *mSnapshot->transform);
 
     glDrawArrays(GL_TRIANGLE_STRIP, 0, vertexBuffer.getSize());
@@ -2050,7 +2064,7 @@
     setupDrawShader();
     setupDrawBlending(isAA, mode);
     setupDrawProgram();
-    setupDrawModelViewIdentity(true);
+    setupDrawModelViewIdentity();
     setupDrawColorUniforms();
     setupDrawColorFilterUniforms();
     setupDrawShaderIdentityUniforms();
@@ -2330,11 +2344,11 @@
 
 status_t OpenGLRenderer::drawRoundRect(float left, float top, float right, float bottom,
         float rx, float ry, SkPaint* p) {
-    if (mSnapshot->isIgnored() || quickReject(left, top, right, bottom)) {
+    if (mSnapshot->isIgnored() || quickRejectPreStroke(left, top, right, bottom, p)) {
         return DrawGlInfo::kStatusDone;
     }
 
-    if (p->getStyle() != SkPaint::kFill_Style) {
+    if (p->getPathEffect() != 0) {
         mCaches.activeTexture(0);
         const PathTexture* texture = mCaches.roundRectShapeCache.getRoundRect(
                 right - left, bottom - top, rx, ry, p);
@@ -2343,37 +2357,47 @@
 
     SkPath path;
     SkRect rect = SkRect::MakeLTRB(left, top, right, bottom);
+    if (p->getStyle() == SkPaint::kStrokeAndFill_Style) {
+        float outset = p->getStrokeWidth() / 2;
+        rect.outset(outset, outset);
+        rx += outset;
+        ry += outset;
+    }
     path.addRoundRect(rect, rx, ry);
-    drawConvexPath(path, p->getColor(), getXfermode(p->getXfermode()), p->isAntiAlias());
+    drawConvexPath(path, p);
 
     return DrawGlInfo::kStatusDrew;
 }
 
 status_t OpenGLRenderer::drawCircle(float x, float y, float radius, SkPaint* p) {
-    if (mSnapshot->isIgnored() || quickReject(x - radius, y - radius, x + radius, y + radius)) {
+    if (mSnapshot->isIgnored() || quickRejectPreStroke(x - radius, y - radius,
+            x + radius, y + radius, p)) {
         return DrawGlInfo::kStatusDone;
     }
-
-    if (p->getStyle() != SkPaint::kFill_Style) {
+    if (p->getPathEffect() != 0) {
         mCaches.activeTexture(0);
         const PathTexture* texture = mCaches.circleShapeCache.getCircle(radius, p);
         return drawShape(x - radius, y - radius, texture, p);
     }
 
     SkPath path;
-    path.addCircle(x, y, radius);
-    drawConvexPath(path, p->getColor(), getXfermode(p->getXfermode()), p->isAntiAlias());
+    if (p->getStyle() == SkPaint::kStrokeAndFill_Style) {
+        path.addCircle(x, y, radius + p->getStrokeWidth() / 2);
+    } else {
+        path.addCircle(x, y, radius);
+    }
+    drawConvexPath(path, p);
 
     return DrawGlInfo::kStatusDrew;
 }
 
 status_t OpenGLRenderer::drawOval(float left, float top, float right, float bottom,
         SkPaint* p) {
-    if (mSnapshot->isIgnored() || quickReject(left, top, right, bottom)) {
+    if (mSnapshot->isIgnored() || quickRejectPreStroke(left, top, right, bottom, p)) {
         return DrawGlInfo::kStatusDone;
     }
 
-    if (p->getStyle() != SkPaint::kFill_Style) {
+    if (p->getPathEffect() != 0) {
         mCaches.activeTexture(0);
         const PathTexture* texture = mCaches.ovalShapeCache.getOval(right - left, bottom - top, p);
         return drawShape(left, top, texture, p);
@@ -2381,8 +2405,11 @@
 
     SkPath path;
     SkRect rect = SkRect::MakeLTRB(left, top, right, bottom);
+    if (p->getStyle() == SkPaint::kStrokeAndFill_Style) {
+        rect.outset(p->getStrokeWidth() / 2, p->getStrokeWidth() / 2);
+    }
     path.addOval(rect);
-    drawConvexPath(path, p->getColor(), getXfermode(p->getXfermode()), p->isAntiAlias());
+    drawConvexPath(path, p);
 
     return DrawGlInfo::kStatusDrew;
 }
@@ -2402,10 +2429,11 @@
 }
 
 status_t OpenGLRenderer::drawRect(float left, float top, float right, float bottom, SkPaint* p) {
-    if (mSnapshot->isIgnored() || quickReject(left, top, right, bottom)) {
+    if (mSnapshot->isIgnored() || quickRejectPreStroke(left, top, right, bottom, p)) {
         return DrawGlInfo::kStatusDone;
     }
 
+    // only fill style is supported by drawConvexPath, since others have to handle joins
     if (p->getStyle() != SkPaint::kFill_Style) {
         mCaches.activeTexture(0);
         const PathTexture* texture = mCaches.rectShapeCache.getRect(right - left, bottom - top, p);
@@ -2415,7 +2443,7 @@
     if (p->isAntiAlias() && !mSnapshot->transform->isSimple()) {
         SkPath path;
         path.addRect(left, top, right, bottom);
-        drawConvexPath(path, p->getColor(), getXfermode(p->getXfermode()), true);
+        drawConvexPath(path, p);
     } else {
         drawColorRect(left, top, right, bottom, p->getColor(), getXfermode(p->getXfermode()));
     }
diff --git a/libs/hwui/OpenGLRenderer.h b/libs/hwui/OpenGLRenderer.h
index c29e3fb..bc9b693 100644
--- a/libs/hwui/OpenGLRenderer.h
+++ b/libs/hwui/OpenGLRenderer.h
@@ -407,6 +407,11 @@
             Rect& transformed, Rect& clip);
 
     /**
+     * Performs a quick reject but adjust the bounds to account for stroke width if necessary
+     */
+    bool quickRejectPreStroke(float left, float top, float right, float bottom, SkPaint* paint);
+
+    /**
      * Creates a new layer stored in the specified snapshot.
      *
      * @param snapshot The snapshot associated with the new layer
@@ -513,11 +518,9 @@
      * Renders the convex hull defined by the specified path as a strip of polygons.
      *
      * @param path The hull of the path to draw
-     * @param color The color of the rect
-     * @param mode The blending mode to draw the path
-     * @param isAA True if the drawing should be anti-aliased
+     * @param paint The paint to render with
      */
-    void drawConvexPath(const SkPath& path, int color, SkXfermode::Mode mode, bool isAA);
+    void drawConvexPath(const SkPath& path, SkPaint* paint);
 
     /**
      * Draws a textured rectangle with the specified texture. The specified coordinates
diff --git a/libs/hwui/PathRenderer.cpp b/libs/hwui/PathRenderer.cpp
index d222009..6893f9d 100644
--- a/libs/hwui/PathRenderer.cpp
+++ b/libs/hwui/PathRenderer.cpp
@@ -21,6 +21,7 @@
 #define VERTEX_DEBUG 0
 
 #include <SkPath.h>
+#include <SkPaint.h>
 
 #include <stdlib.h>
 #include <stdint.h>
@@ -39,10 +40,16 @@
 
 #define THRESHOLD 0.5f
 
-void PathRenderer::computeInverseScales(const mat4 *transform,
-        float &inverseScaleX, float& inverseScaleY) {
-    inverseScaleX = 1.0f;
-    inverseScaleY = 1.0f;
+SkRect PathRenderer::computePathBounds(const SkPath& path, const SkPaint* paint) {
+    SkRect bounds = path.getBounds();
+    if (paint->getStyle() != SkPaint::kFill_Style) {
+        float outset = paint->getStrokeWidth() * 0.5f;
+        bounds.outset(outset, outset);
+    }
+    return bounds;
+}
+
+void computeInverseScales(const mat4 *transform, float &inverseScaleX, float& inverseScaleY) {
     if (CC_UNLIKELY(!transform->isPureTranslate())) {
         float m00 = transform->data[Matrix4::kScaleX];
         float m01 = transform->data[Matrix4::kSkewY];
@@ -50,127 +57,275 @@
         float m11 = transform->data[Matrix4::kScaleY];
         float scaleX = sqrt(m00 * m00 + m01 * m01);
         float scaleY = sqrt(m10 * m10 + m11 * m11);
-        inverseScaleX = (scaleX != 0) ? (inverseScaleX / scaleX) : 0;
-        inverseScaleY = (scaleY != 0) ? (inverseScaleY / scaleY) : 0;
+        inverseScaleX = (scaleX != 0) ? (1.0f / scaleX) : 0;
+        inverseScaleY = (scaleY != 0) ? (1.0f / scaleY) : 0;
+    } else {
+        inverseScaleX = 1.0f;
+        inverseScaleY = 1.0f;
     }
 }
 
-void PathRenderer::convexPathFillVertices(const SkPath &path, const mat4 *transform,
-        VertexBuffer &vertexBuffer, bool isAA) {
-    ATRACE_CALL();
-    float inverseScaleX;
-    float inverseScaleY;
-    computeInverseScales(transform, inverseScaleX, inverseScaleY);
+inline void copyVertex(Vertex* destPtr, const Vertex* srcPtr)
+{
+    Vertex::set(destPtr, srcPtr->position[0], srcPtr->position[1]);
+}
 
-    Vector<Vertex> tempVertices;
-    float thresholdx = THRESHOLD * inverseScaleX;
-    float thresholdy = THRESHOLD * inverseScaleY;
-    convexPathVertices(path,
-                       thresholdx * thresholdx,
-                       thresholdy * thresholdy,
-                       tempVertices);
+inline void copyAlphaVertex(AlphaVertex* destPtr, const AlphaVertex* srcPtr)
+{
+    AlphaVertex::set(destPtr, srcPtr->position[0], srcPtr->position[1], srcPtr->alpha);
+}
 
-#if VERTEX_DEBUG
-    for (unsigned int i = 0; i < tempVertices.size(); i++) {
-        ALOGD("orig path: point at %f %f",
-              tempVertices[i].position[0],
-              tempVertices[i].position[1]);
-    }
-#endif
+void getFillVerticesFromPerimeter(const Vector<Vertex>& perimeter, VertexBuffer& vertexBuffer) {
+    Vertex* buffer = vertexBuffer.alloc<Vertex>(perimeter.size());
+
     int currentIndex = 0;
-    if (!isAA) {
-        Vertex* buffer = vertexBuffer.alloc<Vertex>(tempVertices.size());
-
-        // zig zag between all previous points on the inside of the hull to create a
-        // triangle strip that fills the hull
-        int srcAindex = 0;
-        int srcBindex = tempVertices.size() - 1;
-        while (srcAindex <= srcBindex) {
-            Vertex::set(&buffer[currentIndex++],
-                        tempVertices.editArray()[srcAindex].position[0],
-                        tempVertices.editArray()[srcAindex].position[1]);
-            if (srcAindex == srcBindex) break;
-            Vertex::set(&buffer[currentIndex++],
-                        tempVertices.editArray()[srcBindex].position[0],
-                        tempVertices.editArray()[srcBindex].position[1]);
-            srcAindex++;
-            srcBindex--;
-        }
-        return;
+    // zig zag between all previous points on the inside of the hull to create a
+    // triangle strip that fills the hull
+    int srcAindex = 0;
+    int srcBindex = perimeter.size() - 1;
+    while (srcAindex <= srcBindex) {
+        copyVertex(&buffer[currentIndex++], &perimeter[srcAindex]);
+        if (srcAindex == srcBindex) break;
+        copyVertex(&buffer[currentIndex++], &perimeter[srcBindex]);
+        srcAindex++;
+        srcBindex--;
     }
-    AlphaVertex* buffer = vertexBuffer.alloc<AlphaVertex>(tempVertices.size() * 3 + 2);
+}
+
+void getStrokeVerticesFromPerimeter(const Vector<Vertex>& perimeter, float halfStrokeWidth,
+        VertexBuffer& vertexBuffer, float inverseScaleX, float inverseScaleY) {
+    Vertex* buffer = vertexBuffer.alloc<Vertex>(perimeter.size() * 2 + 2);
+
+    int currentIndex = 0;
+    const Vertex* last = &(perimeter[perimeter.size() - 1]);
+    const Vertex* current = &(perimeter[0]);
+    vec2 lastNormal(current->position[1] - last->position[1],
+            last->position[0] - current->position[0]);
+    lastNormal.normalize();
+    for (unsigned int i = 0; i < perimeter.size(); i++) {
+        const Vertex* next = &(perimeter[i + 1 >= perimeter.size() ? 0 : i + 1]);
+        vec2 nextNormal(next->position[1] - current->position[1],
+                current->position[0] - next->position[0]);
+        nextNormal.normalize();
+
+        // offset each point by its normal, out and in, by appropriate stroke offset
+        vec2 totalOffset = (lastNormal + nextNormal);
+        totalOffset.normalize();
+        if (halfStrokeWidth == 0.0f) {
+            // hairline - compensate for scale
+            totalOffset.x *= 0.5f * inverseScaleX;
+            totalOffset.y *= 0.5f * inverseScaleY;
+        } else {
+            totalOffset *= halfStrokeWidth;
+        }
+
+        Vertex::set(&buffer[currentIndex++],
+                current->position[0] + totalOffset.x,
+                current->position[1] + totalOffset.y);
+
+        Vertex::set(&buffer[currentIndex++],
+                current->position[0] - totalOffset.x,
+                current->position[1] - totalOffset.y);
+
+        last = current;
+        current = next;
+        lastNormal = nextNormal;
+    }
+
+    // wrap around to beginning
+    copyVertex(&buffer[currentIndex++], &buffer[0]);
+    copyVertex(&buffer[currentIndex++], &buffer[1]);
+}
+
+void getFillVerticesFromPerimeterAA(const Vector<Vertex>& perimeter, VertexBuffer& vertexBuffer,
+         float inverseScaleX, float inverseScaleY) {
+    AlphaVertex* buffer = vertexBuffer.alloc<AlphaVertex>(perimeter.size() * 3 + 2);
 
     // generate alpha points - fill Alpha vertex gaps in between each point with
     // alpha 0 vertex, offset by a scaled normal.
-    Vertex* last = &(tempVertices.editArray()[tempVertices.size()-1]);
-
-    for (unsigned int i = 0; i<tempVertices.size(); i++) {
-        Vertex* current = &(tempVertices.editArray()[i]);
-        Vertex* next = &(tempVertices.editArray()[i + 1 >= tempVertices.size() ? 0 : i + 1]);
-
-        vec2 lastNormal(current->position[1] - last->position[1],
-                        last->position[0] - current->position[0]);
-        lastNormal.normalize();
+    int currentIndex = 0;
+    const Vertex* last = &(perimeter[perimeter.size() - 1]);
+    const Vertex* current = &(perimeter[0]);
+    vec2 lastNormal(current->position[1] - last->position[1],
+            last->position[0] - current->position[0]);
+    lastNormal.normalize();
+    for (unsigned int i = 0; i < perimeter.size(); i++) {
+        const Vertex* next = &(perimeter[i + 1 >= perimeter.size() ? 0 : i + 1]);
         vec2 nextNormal(next->position[1] - current->position[1],
-                        current->position[0] - next->position[0]);
+                current->position[0] - next->position[0]);
         nextNormal.normalize();
 
         // AA point offset from original point is that point's normal, such that
         // each side is offset by .5 pixels
-        vec2 totalOffset = (lastNormal + nextNormal) / (2 * (1 + lastNormal.dot(nextNormal)));
-        totalOffset.x *= inverseScaleX;
-        totalOffset.y *= inverseScaleY;
+        vec2 totalOffset = (lastNormal + nextNormal);
+        totalOffset.normalize();
+        totalOffset.x *= inverseScaleX * 0.5f;
+        totalOffset.y *= inverseScaleY * 0.5f;
 
         AlphaVertex::set(&buffer[currentIndex++],
-                         current->position[0] + totalOffset.x,
-                         current->position[1] + totalOffset.y,
-                         0.0f);
+                current->position[0] + totalOffset.x,
+                current->position[1] + totalOffset.y,
+                0.0f);
         AlphaVertex::set(&buffer[currentIndex++],
-                         current->position[0] - totalOffset.x,
-                         current->position[1] - totalOffset.y,
-                         1.0f);
+                current->position[0] - totalOffset.x,
+                current->position[1] - totalOffset.y,
+                1.0f);
+
         last = current;
+        current = next;
+        lastNormal = nextNormal;
     }
 
     // wrap around to beginning
-    AlphaVertex::set(&buffer[currentIndex++],
-                     buffer[0].position[0],
-                     buffer[0].position[1], 0.0f);
-    AlphaVertex::set(&buffer[currentIndex++],
-                     buffer[1].position[0],
-                     buffer[1].position[1], 1.0f);
+    copyAlphaVertex(&buffer[currentIndex++], &buffer[0]);
+    copyAlphaVertex(&buffer[currentIndex++], &buffer[1]);
 
     // zig zag between all previous points on the inside of the hull to create a
     // triangle strip that fills the hull, repeating the first inner point to
     // create degenerate tris to start inside path
     int srcAindex = 0;
-    int srcBindex = tempVertices.size() - 1;
+    int srcBindex = perimeter.size() - 1;
     while (srcAindex <= srcBindex) {
-        AlphaVertex::set(&buffer[currentIndex++],
-                         buffer[srcAindex * 2 + 1].position[0],
-                         buffer[srcAindex * 2 + 1].position[1],
-                         1.0f);
+        copyAlphaVertex(&buffer[currentIndex++], &buffer[srcAindex * 2 + 1]);
         if (srcAindex == srcBindex) break;
-        AlphaVertex::set(&buffer[currentIndex++],
-                         buffer[srcBindex * 2 + 1].position[0],
-                         buffer[srcBindex * 2 + 1].position[1],
-                         1.0f);
+        copyAlphaVertex(&buffer[currentIndex++], &buffer[srcBindex * 2 + 1]);
         srcAindex++;
         srcBindex--;
     }
 
 #if VERTEX_DEBUG
-    for (unsigned int i = 0; i < vertexBuffer.mSize; i++) {
-        ALOGD("point at %f %f",
-              buffer[i].position[0],
-              buffer[i].position[1]);
+    for (unsigned int i = 0; i < vertexBuffer.getSize(); i++) {
+        ALOGD("point at %f %f", buffer[i].position[0], buffer[i].position[1]);
     }
 #endif
 }
 
+void getStrokeVerticesFromPerimeterAA(const Vector<Vertex>& perimeter, float halfStrokeWidth,
+        VertexBuffer& vertexBuffer, float inverseScaleX, float inverseScaleY) {
+    AlphaVertex* buffer = vertexBuffer.alloc<AlphaVertex>(6 * perimeter.size() + 8);
 
-void PathRenderer::convexPathVertices(const SkPath &path, float thresholdx, float thresholdy,
-        Vector<Vertex> &outputVertices) {
+    int offset = 2 * perimeter.size() + 3;
+    int currentAAOuterIndex = 0;
+    int currentStrokeIndex = offset;
+    int currentAAInnerIndex = offset * 2;
+
+    const Vertex* last = &(perimeter[perimeter.size() - 1]);
+    const Vertex* current = &(perimeter[0]);
+    vec2 lastNormal(current->position[1] - last->position[1],
+            last->position[0] - current->position[0]);
+    lastNormal.normalize();
+    for (unsigned int i = 0; i < perimeter.size(); i++) {
+        const Vertex* next = &(perimeter[i + 1 >= perimeter.size() ? 0 : i + 1]);
+        vec2 nextNormal(next->position[1] - current->position[1],
+                current->position[0] - next->position[0]);
+        nextNormal.normalize();
+
+        vec2 pointNormal = (lastNormal + nextNormal);
+        pointNormal.normalize();
+        vec2 AAOffset = pointNormal * 0.5f;
+        AAOffset.x *= inverseScaleX;
+        AAOffset.y *= inverseScaleY;
+
+        vec2 innerOffset = pointNormal;
+        if (halfStrokeWidth == 0.0f) {
+            // hairline! - compensate for scale
+            innerOffset.x *= 0.5f * inverseScaleX;
+            innerOffset.y *= 0.5f * inverseScaleY;
+        } else {
+            innerOffset *= halfStrokeWidth;
+        }
+        vec2 outerOffset = innerOffset + AAOffset;
+        innerOffset -= AAOffset;
+
+        AlphaVertex::set(&buffer[currentAAOuterIndex++],
+                current->position[0] + outerOffset.x,
+                current->position[1] + outerOffset.y,
+                0.0f);
+        AlphaVertex::set(&buffer[currentAAOuterIndex++],
+                current->position[0] + innerOffset.x,
+                current->position[1] + innerOffset.y,
+                1.0f);
+
+        AlphaVertex::set(&buffer[currentStrokeIndex++],
+                current->position[0] + innerOffset.x,
+                current->position[1] + innerOffset.y,
+                1.0f);
+        AlphaVertex::set(&buffer[currentStrokeIndex++],
+                current->position[0] - innerOffset.x,
+                current->position[1] - innerOffset.y,
+                1.0f);
+
+        AlphaVertex::set(&buffer[currentAAInnerIndex++],
+                current->position[0] - innerOffset.x,
+                current->position[1] - innerOffset.y,
+                1.0f);
+        AlphaVertex::set(&buffer[currentAAInnerIndex++],
+                current->position[0] - outerOffset.x,
+                current->position[1] - outerOffset.y,
+                0.0f);
+
+        // TODO: current = next, copy last normal instead of recalculate
+        last = current;
+        current = next;
+        lastNormal = nextNormal;
+    }
+
+    // wrap each strip around to beginning, creating degenerate tris to bridge strips
+    copyAlphaVertex(&buffer[currentAAOuterIndex++], &buffer[0]);
+    copyAlphaVertex(&buffer[currentAAOuterIndex++], &buffer[1]);
+    copyAlphaVertex(&buffer[currentAAOuterIndex++], &buffer[1]);
+
+    copyAlphaVertex(&buffer[currentStrokeIndex++], &buffer[offset]);
+    copyAlphaVertex(&buffer[currentStrokeIndex++], &buffer[offset + 1]);
+    copyAlphaVertex(&buffer[currentStrokeIndex++], &buffer[offset + 1]);
+
+    copyAlphaVertex(&buffer[currentAAInnerIndex++], &buffer[2 * offset]);
+    copyAlphaVertex(&buffer[currentAAInnerIndex++], &buffer[2 * offset + 1]);
+    // don't need to create last degenerate tri
+}
+
+void PathRenderer::convexPathVertices(const SkPath &path, const SkPaint* paint,
+        const mat4 *transform, VertexBuffer& vertexBuffer) {
+    ATRACE_CALL();
+
+    SkPaint::Style style = paint->getStyle();
+    bool isAA = paint->isAntiAlias();
+
+    float inverseScaleX, inverseScaleY;
+    computeInverseScales(transform, inverseScaleX, inverseScaleY);
+
+    Vector<Vertex> tempVertices;
+    convexPathPerimeterVertices(path, inverseScaleX * inverseScaleX, inverseScaleY * inverseScaleY,
+            tempVertices);
+
+#if VERTEX_DEBUG
+    for (unsigned int i = 0; i < tempVertices.size(); i++) {
+        ALOGD("orig path: point at %f %f", tempVertices[i].position[0], tempVertices[i].position[1]);
+    }
+#endif
+
+    if (style == SkPaint::kStroke_Style) {
+        float halfStrokeWidth = paint->getStrokeWidth() * 0.5f;
+        if (!isAA) {
+            getStrokeVerticesFromPerimeter(tempVertices, halfStrokeWidth, vertexBuffer,
+                    inverseScaleX, inverseScaleY);
+        } else {
+            getStrokeVerticesFromPerimeterAA(tempVertices, halfStrokeWidth, vertexBuffer,
+                    inverseScaleX, inverseScaleY);
+        }
+    } else {
+        // For kStrokeAndFill style, the path should be adjusted externally, as it will be treated as a fill here.
+        if (!isAA) {
+            getFillVerticesFromPerimeter(tempVertices, vertexBuffer);
+        } else {
+            getFillVerticesFromPerimeterAA(tempVertices, vertexBuffer, inverseScaleX, inverseScaleY);
+        }
+    }
+}
+
+
+void PathRenderer::convexPathPerimeterVertices(const SkPath& path,
+        float sqrInvScaleX, float sqrInvScaleY, Vector<Vertex>& outputVertices) {
     ATRACE_CALL();
 
     SkPath::Iter iter(path, true);
@@ -189,31 +344,30 @@
                     break;
                 case SkPath::kLine_Verb:
                     ALOGV("kLine_Verb %f %f -> %f %f",
-                          pts[0].x(), pts[0].y(),
-                          pts[1].x(), pts[1].y());
+                            pts[0].x(), pts[0].y(),
+                            pts[1].x(), pts[1].y());
 
                     // TODO: make this not yuck
                     outputVertices.push();
-                    newVertex = &(outputVertices.editArray()[outputVertices.size()-1]);
+                    newVertex = &(outputVertices.editArray()[outputVertices.size() - 1]);
                     Vertex::set(newVertex, pts[1].x(), pts[1].y());
                     break;
                 case SkPath::kQuad_Verb:
                     ALOGV("kQuad_Verb");
                     recursiveQuadraticBezierVertices(
-                        pts[0].x(), pts[0].y(),
-                        pts[2].x(), pts[2].y(),
-                        pts[1].x(), pts[1].y(),
-                        thresholdx, thresholdy,
-                        outputVertices);
+                            pts[0].x(), pts[0].y(),
+                            pts[2].x(), pts[2].y(),
+                            pts[1].x(), pts[1].y(),
+                            sqrInvScaleX, sqrInvScaleY, outputVertices);
                     break;
                 case SkPath::kCubic_Verb:
                     ALOGV("kCubic_Verb");
                     recursiveCubicBezierVertices(
-                        pts[0].x(), pts[0].y(),
-                        pts[1].x(), pts[1].y(),
-                        pts[3].x(), pts[3].y(),
-                        pts[2].x(), pts[2].y(),
-                        thresholdx, thresholdy, outputVertices);
+                            pts[0].x(), pts[0].y(),
+                            pts[1].x(), pts[1].y(),
+                            pts[3].x(), pts[3].y(),
+                            pts[2].x(), pts[2].y(),
+                        sqrInvScaleX, sqrInvScaleY, outputVertices);
                     break;
                 default:
                     break;
@@ -224,18 +378,20 @@
 void PathRenderer::recursiveCubicBezierVertices(
         float p1x, float p1y, float c1x, float c1y,
         float p2x, float p2y, float c2x, float c2y,
-        float thresholdx, float thresholdy, Vector<Vertex> &outputVertices) {
+        float sqrInvScaleX, float sqrInvScaleY, Vector<Vertex>& outputVertices) {
     float dx = p2x - p1x;
     float dy = p2y - p1y;
     float d1 = fabs((c1x - p2x) * dy - (c1y - p2y) * dx);
     float d2 = fabs((c2x - p2x) * dy - (c2y - p2y) * dx);
     float d = d1 + d2;
 
-    if (d * d < (thresholdx * (dx * dx) + thresholdy * (dy * dy))) {
+    // multiplying by sqrInvScaleY/X equivalent to multiplying in dimensional scale factors
+
+    if (d * d < THRESHOLD * THRESHOLD * (dx * dx * sqrInvScaleY + dy * dy * sqrInvScaleX)) {
         // below thresh, draw line by adding endpoint
         // TODO: make this not yuck
         outputVertices.push();
-        Vertex* newVertex = &(outputVertices.editArray()[outputVertices.size()-1]);
+        Vertex* newVertex = &(outputVertices.editArray()[outputVertices.size() - 1]);
         Vertex::set(newVertex, p2x, p2y);
     } else {
         float p1c1x = (p1x + c1x) * 0.5f;
@@ -258,13 +414,11 @@
         recursiveCubicBezierVertices(
                 p1x, p1y, p1c1x, p1c1y,
                 mx, my, p1c1c2x, p1c1c2y,
-                thresholdx, thresholdy,
-                outputVertices);
+                sqrInvScaleX, sqrInvScaleY, outputVertices);
         recursiveCubicBezierVertices(
                 mx, my, p2c1c2x, p2c1c2y,
                 p2x, p2y, p2c2x, p2c2y,
-                thresholdx, thresholdy,
-                outputVertices);
+                sqrInvScaleX, sqrInvScaleY, outputVertices);
     }
 }
 
@@ -272,16 +426,16 @@
         float ax, float ay,
         float bx, float by,
         float cx, float cy,
-        float thresholdx, float thresholdy, Vector<Vertex> &outputVertices) {
+        float sqrInvScaleX, float sqrInvScaleY, Vector<Vertex>& outputVertices) {
     float dx = bx - ax;
     float dy = by - ay;
     float d = (cx - bx) * dy - (cy - by) * dx;
 
-    if (d * d < (thresholdx * (dx * dx) + thresholdy * (dy * dy))) {
+    if (d * d < THRESHOLD * THRESHOLD * (dx * dx * sqrInvScaleY + dy * dy * sqrInvScaleX)) {
         // below thresh, draw line by adding endpoint
         // TODO: make this not yuck
         outputVertices.push();
-        Vertex* newVertex = &(outputVertices.editArray()[outputVertices.size()-1]);
+        Vertex* newVertex = &(outputVertices.editArray()[outputVertices.size() - 1]);
         Vertex::set(newVertex, bx, by);
     } else {
         float acx = (ax + cx) * 0.5f;
@@ -294,9 +448,9 @@
         float my = (acy + bcy) * 0.5f;
 
         recursiveQuadraticBezierVertices(ax, ay, mx, my, acx, acy,
-                thresholdx, thresholdy, outputVertices);
+                sqrInvScaleX, sqrInvScaleY, outputVertices);
         recursiveQuadraticBezierVertices(mx, my, bx, by, bcx, bcy,
-                thresholdx, thresholdy, outputVertices);
+                sqrInvScaleX, sqrInvScaleY, outputVertices);
     }
 }
 
diff --git a/libs/hwui/PathRenderer.h b/libs/hwui/PathRenderer.h
index 1354f16..28a5b90 100644
--- a/libs/hwui/PathRenderer.h
+++ b/libs/hwui/PathRenderer.h
@@ -35,15 +35,13 @@
         mCleanupMethod(0)
     {}
 
-    ~VertexBuffer()
-    {
+    ~VertexBuffer() {
         if (mCleanupMethod)
             mCleanupMethod(mBuffer);
     }
 
     template <class TYPE>
-    TYPE* alloc(int size)
-    {
+    TYPE* alloc(int size) {
         mSize = size;
         mBuffer = (void*)new TYPE[size];
         mCleanupMethod = &(cleanup<TYPE>);
@@ -56,8 +54,7 @@
 
 private:
     template <class TYPE>
-    static void cleanup(void* buffer)
-    {
+    static void cleanup(void* buffer) {
         delete[] (TYPE*)buffer;
     }
 
@@ -68,17 +65,15 @@
 
 class PathRenderer {
 public:
-    static void computeInverseScales(
-        const mat4 *transform, float &inverseScaleX, float& inverseScaleY);
+    static SkRect computePathBounds(const SkPath& path, const SkPaint* paint);
 
-    static void convexPathFillVertices(
-        const SkPath &path, const mat4 *transform,
-        VertexBuffer &vertexBuffer, bool isAA);
+    static void convexPathVertices(const SkPath& path, const SkPaint* paint,
+            const mat4 *transform, VertexBuffer& vertexBuffer);
 
 private:
-    static void convexPathVertices(
+    static void convexPathPerimeterVertices(
         const SkPath &path,
-        float thresholdx, float thresholdy,
+        float sqrInvScaleX, float sqrInvScaleY,
         Vector<Vertex> &outputVertices);
 
 /*
@@ -86,23 +81,23 @@
   control c
  */
     static void recursiveQuadraticBezierVertices(
-        float ax, float ay,
-        float bx, float by,
-        float cx, float cy,
-        float thresholdx, float thresholdy,
-        Vector<Vertex> &outputVertices);
+            float ax, float ay,
+            float bx, float by,
+            float cx, float cy,
+            float sqrInvScaleX, float sqrInvScaleY,
+            Vector<Vertex> &outputVertices);
 
 /*
   endpoints p1, p2
   control c1, c2
  */
     static void recursiveCubicBezierVertices(
-        float p1x, float p1y,
-        float c1x, float c1y,
-        float p2x, float p2y,
-        float c2x, float c2y,
-        float thresholdx, float thresholdy,
-        Vector<Vertex> &outputVertices);
+            float p1x, float p1y,
+            float c1x, float c1y,
+            float p2x, float p2y,
+            float c2x, float c2y,
+            float sqrInvScaleX, float sqrInvScaleY,
+            Vector<Vertex> &outputVertices);
 };
 
 }; // namespace uirenderer