diff --git a/example/example_gl2.c b/example/example_gl2.c
index 2ab8e0d..4627882 100644
--- a/example/example_gl2.c
+++ b/example/example_gl2.c
@@ -87,7 +87,10 @@ int main()
 	if (loadDemoData(vg, &data) == -1)
 		return -1;
 
+	glfwSwapInterval(0);
+
 	glfwSetTime(0);
+	prevt = glfwGetTime();
 
 	while (!glfwWindowShouldClose(window))
 	{
diff --git a/example/example_gl3.c b/example/example_gl3.c
index c135b07..3a53d84 100644
--- a/example/example_gl3.c
+++ b/example/example_gl3.c
@@ -91,6 +91,8 @@ int main()
 	if (loadDemoData(vg, &data) == -1)
 		return -1;
 
+	glfwSwapInterval(0);
+
 	glfwSetTime(0);
 	prevt = glfwGetTime();
 
diff --git a/src/nanovg.c b/src/nanovg.c
index af72815..9453404 100644
--- a/src/nanovg.c
+++ b/src/nanovg.c
@@ -941,7 +941,7 @@ static void nvg__flattenPaths(struct NVGcontext* ctx, float m)
 	struct NVGpoint* p1;
 	struct NVGpoint* pts;
 	struct NVGpath* path;
-	int i, j;
+	int i, j, nleft;
 	float* cp1;
 	float* cp2;
 	float* p;
@@ -1035,6 +1035,7 @@ static void nvg__flattenPaths(struct NVGcontext* ctx, float m)
 		path = &cache->paths[j];
 		pts = &cache->points[path->first];
 		path->nbevel = 0;
+		nleft = 0;
 
 		p0 = &pts[path->count-1];
 		p1 = &pts[0];
@@ -1058,12 +1059,15 @@ static void nvg__flattenPaths(struct NVGcontext* ctx, float m)
 			if ((dmr2 * m*m) < 1.0f) {
 				cross = p1->dx * p0->dy - p0->dx * p1->dy;
 				p1->flags |= NVG_BEVEL;
-				if (cross < 0)
+				if (cross < 0) {
 					p1->flags |= NVG_LEFT;
+					nleft++;
+				}
 				path->nbevel++;
 			}
 			p0 = p1++;
 		}
+		path->convex = nleft == 0 ? 1 : 0;
 	}
 }
 
@@ -1077,6 +1081,7 @@ static int nvg__expandStrokeAndFill(struct NVGcontext* ctx, int feats, float w)
 	struct NVGpoint* p0;
 	struct NVGpoint* p1;
 	int nstroke, cverts;
+	int convex = 0;
 	int i, j, s, e;
 	float wo = 0;
 
@@ -1101,6 +1106,7 @@ static int nvg__expandStrokeAndFill(struct NVGcontext* ctx, int feats, float w)
 		path = &cache->paths[i];
 		pts = &cache->points[path->first];
 		nstroke = (path->count + path->nbevel + 1) * 2;
+		convex = 0;
 
 		// Calculate shape vertices.
 		if (feats & NVG_FILL) {
@@ -1138,6 +1144,8 @@ static int nvg__expandStrokeAndFill(struct NVGcontext* ctx, int feats, float w)
 			path->nfill = (int)(dst - verts);
 			verts += path->nfill;
 
+			if (path->convex && cache->npaths == 1)
+				convex = 1;
 		} else {
 			wo = 0.0f;
 			path->fill = 0;
@@ -1146,11 +1154,19 @@ static int nvg__expandStrokeAndFill(struct NVGcontext* ctx, int feats, float w)
 
 		// Calculate fringe
 		if (feats & NVG_STROKE) {
-			float lw = w + wo; float rw = w - wo;
+			float lw = w + wo, rw = w - wo;
+			float u0 = 0, u1 = 1;
 			int loop = ((feats & NVG_CAPS) && path->closed == 0) ? 0 : 1;
 			dst = verts;
 			path->stroke = dst;
 
+			// Create only half a fringe for convex shapes so that
+			// the shape can be rendered without stenciling.
+			if (convex) {
+				lw = wo;	// This should generate the same vertex as fill inset above.
+				u1 = 0.5f;	// Set outline fade at middle.
+			}
+
 			if (loop) {
 				// Looping
 				p0 = &pts[path->count-1];
@@ -1173,10 +1189,10 @@ static int nvg__expandStrokeAndFill(struct NVGcontext* ctx, int feats, float w)
 				nvg__normalize(&dx, &dy);
 				dlx = dy;
 				dly = -dx;
-				nvg__vset(dst, p0->x + dlx*rw - dx*NVG_AA, p0->y + dly*rw - dy*NVG_AA, 0,0); dst++;
-				nvg__vset(dst, p0->x - dlx*lw - dx*NVG_AA, p0->y - dly*lw - dy*NVG_AA, 1,0); dst++;
-				nvg__vset(dst, p0->x + dlx*rw, p0->y + dly * rw, 0,1); dst++;
-				nvg__vset(dst, p0->x - dlx*lw, p0->y - dly * lw, 1,1); dst++;
+				nvg__vset(dst, p0->x + dlx*rw - dx*NVG_AA, p0->y + dly*rw - dy*NVG_AA, u0,0); dst++;
+				nvg__vset(dst, p0->x - dlx*lw - dx*NVG_AA, p0->y - dly*lw - dy*NVG_AA, u1,0); dst++;
+				nvg__vset(dst, p0->x + dlx*rw, p0->y + dly * rw, u0,1); dst++;
+				nvg__vset(dst, p0->x - dlx*lw, p0->y - dly * lw, u1,1); dst++;
 			}
 
 			for (j = s; j < e; ++j) {
@@ -1192,10 +1208,10 @@ static int nvg__expandStrokeAndFill(struct NVGcontext* ctx, int feats, float w)
 						float ry1 = p1->y + dly1 * rw;
 						float lx = p1->x - p1->dmx * lw;
 						float ly = p1->y - p1->dmy * lw;
-						nvg__vset(dst, rx0, ry0, 0,1); dst++;
-						nvg__vset(dst, lx, ly, 1,1); dst++;
-						nvg__vset(dst, rx1, ry1, 0,1); dst++;
-						nvg__vset(dst, lx, ly, 1,1); dst++;
+						nvg__vset(dst, rx0, ry0, u0,1); dst++;
+						nvg__vset(dst, lx, ly, u1,1); dst++;
+						nvg__vset(dst, rx1, ry1, u0,1); dst++;
+						nvg__vset(dst, lx, ly, u1,1); dst++;
 					} else {
 						float rx = p1->x + p1->dmx * rw;
 						float ry = p1->y + p1->dmy * rw;
@@ -1203,22 +1219,22 @@ static int nvg__expandStrokeAndFill(struct NVGcontext* ctx, int feats, float w)
 						float ly0 = p1->y - dly0 * lw;
 						float lx1 = p1->x - dlx1 * lw;
 						float ly1 = p1->y - dly1 * lw;
-						nvg__vset(dst, rx, ry, 0,1); dst++;
-						nvg__vset(dst, lx0, ly0, 1,1); dst++;
-						nvg__vset(dst, rx, ry, 0,1); dst++;
-						nvg__vset(dst, lx1, ly1, 1,1); dst++;
+						nvg__vset(dst, rx, ry, u0,1); dst++;
+						nvg__vset(dst, lx0, ly0, u1,1); dst++;
+						nvg__vset(dst, rx, ry, u0,1); dst++;
+						nvg__vset(dst, lx1, ly1, u1,1); dst++;
 					}
 				} else {
-					nvg__vset(dst, p1->x + (p1->dmx * rw), p1->y + (p1->dmy * rw), 0,1); dst++;
-					nvg__vset(dst, p1->x - (p1->dmx * lw), p1->y - (p1->dmy * lw), 1,1); dst++;
+					nvg__vset(dst, p1->x + (p1->dmx * rw), p1->y + (p1->dmy * rw), u0,1); dst++;
+					nvg__vset(dst, p1->x - (p1->dmx * lw), p1->y - (p1->dmy * lw), u1,1); dst++;
 				}
 				p0 = p1++;
 			}
 
 			if (loop) {
 				// Loop it
-				nvg__vset(dst, verts[0].x, verts[0].y, 0,1); dst++;
-				nvg__vset(dst, verts[1].x, verts[1].y, 1,1); dst++;
+				nvg__vset(dst, verts[0].x, verts[0].y, u0,1); dst++;
+				nvg__vset(dst, verts[1].x, verts[1].y, u1,1); dst++;
 			} else {
 				// Add cap
 				float dx, dy, dlx, dly;
@@ -1227,10 +1243,10 @@ static int nvg__expandStrokeAndFill(struct NVGcontext* ctx, int feats, float w)
 				nvg__normalize(&dx, &dy);
 				dlx = dy;
 				dly = -dx;
-				nvg__vset(dst, p1->x + dlx*rw, p1->y + dly * rw, 0,1); dst++;
-				nvg__vset(dst, p1->x - dlx*lw, p1->y - dly * lw, 1,1); dst++;
-				nvg__vset(dst, p1->x + dlx*rw + dx*NVG_AA, p1->y + dly*rw + dy*NVG_AA, 0,0); dst++;
-				nvg__vset(dst, p1->x - dlx*lw + dx*NVG_AA, p1->y - dly*lw + dy*NVG_AA, 1,0); dst++;
+				nvg__vset(dst, p1->x + dlx*rw, p1->y + dly * rw, u0,1); dst++;
+				nvg__vset(dst, p1->x - dlx*lw, p1->y - dly * lw, u1,1); dst++;
+				nvg__vset(dst, p1->x + dlx*rw + dx*NVG_AA, p1->y + dly*rw + dy*NVG_AA, u0,0); dst++;
+				nvg__vset(dst, p1->x - dlx*lw + dx*NVG_AA, p1->y - dly*lw + dy*NVG_AA, u1,0); dst++;
 			}
 
 			path->nstroke = (int)(dst - verts);
diff --git a/src/nanovg.h b/src/nanovg.h
index f408b59..fabe79c 100644
--- a/src/nanovg.h
+++ b/src/nanovg.h
@@ -387,6 +387,7 @@ struct NVGpath {
 	struct NVGvertex* stroke;
 	int nstroke;
 	int winding;
+	int convex;
 };
 
 struct NVGparams {
diff --git a/src/nanovg_gl2.h b/src/nanovg_gl2.h
index 8473c75..25272e8 100644
--- a/src/nanovg_gl2.h
+++ b/src/nanovg_gl2.h
@@ -486,85 +486,104 @@ static void glnvg__renderFill(void* uptr, struct NVGpaint* paint, struct NVGscis
 	if (gl->gradShader.prog == 0)
 		return;
 
-	glEnable(GL_CULL_FACE);
+	if (npaths == 1 && paths[0].convex) {
 
-	glEnableClientState(GL_VERTEX_ARRAY);
+		glEnable(GL_CULL_FACE);
 
-	// Draw shapes
-	glDisable(GL_BLEND);
-	glEnable(GL_STENCIL_TEST);
-	glStencilMask(0xff);
-	glStencilFunc(GL_ALWAYS, 0, ~0);
-	glColorMask(GL_FALSE, GL_FALSE, GL_FALSE, GL_FALSE);
+		glEnableClientState(GL_VERTEX_ARRAY);
 
-	glStencilOpSeparate(GL_FRONT, GL_KEEP, GL_KEEP, GL_INCR_WRAP);
-	glStencilOpSeparate(GL_BACK, GL_KEEP, GL_KEEP, GL_DECR_WRAP);
-	glDisable(GL_CULL_FACE);
-	for (i = 0; i < npaths; i++) {
-		path = &paths[i];
-		glVertexPointer(2, GL_FLOAT, sizeof(struct NVGvertex), &path->fill[0].x);
-		glDrawArrays(GL_TRIANGLE_FAN, 0, path->nfill);
-	}
+		glnvg__setupPaint(gl, paint, scissor, 1.0001f, aasize);
 
-/*	glStencilOp(GL_KEEP, GL_KEEP, GL_INCR_WRAP);
-	glCullFace(GL_BACK);
+		// Fill
+		glDisable(GL_CULL_FACE);
+		glTexCoord2f(0.5f,1.0f);
+		for (i = 0; i < npaths; i++) {
+			path = &paths[i];
+			glVertexPointer(2, GL_FLOAT, sizeof(struct NVGvertex), &path->fill[0].x);
+			glDrawArrays(GL_TRIANGLE_FAN, 0, path->nfill);
+		}
 
-	for (i = 0; i < npaths; i++) {
-		path = &paths[i];
-		glVertexPointer(2, GL_FLOAT, sizeof(struct NVGvertex), &path->fill[0].x);
-		glDrawArrays(GL_TRIANGLE_FAN, 0, path->nfill);
-	}
+		glEnableClientState(GL_TEXTURE_COORD_ARRAY);
+		glEnable(GL_CULL_FACE);
 
-	glStencilOp(GL_KEEP, GL_KEEP, GL_DECR_WRAP);
-	glCullFace(GL_FRONT);
+		// Draw fringes
+		for (i = 0; i < npaths; i++) {
+			path = &paths[i];
+			glVertexPointer(2, GL_FLOAT, sizeof(struct NVGvertex), &path->stroke[0].x);
+			glTexCoordPointer(2, GL_FLOAT, sizeof(struct NVGvertex), &path->stroke[0].u);
+			glDrawArrays(GL_TRIANGLE_STRIP, 0, path->nstroke);
+		}
 
-	for (i = 0; i < npaths; i++) {
-		path = &paths[i];
-		glVertexPointer(2, GL_FLOAT, sizeof(struct NVGvertex), &path->fill[0].x);
-		glDrawArrays(GL_TRIANGLE_FAN, 0, path->nfill);
-	}
-	glCullFace(GL_BACK);*/
+		glDisableClientState(GL_VERTEX_ARRAY);
+		glDisableClientState(GL_TEXTURE_COORD_ARRAY);
 
-	glEnable(GL_CULL_FACE);
+		glUseProgram(0);
+
+		glDisable(GL_TEXTURE_2D);
+
+	} else {
+		
+		glEnable(GL_CULL_FACE);
+
+		glEnableClientState(GL_VERTEX_ARRAY);
+
+		// Draw shapes
+		glDisable(GL_BLEND);
+		glEnable(GL_STENCIL_TEST);
+		glStencilMask(0xff);
+		glStencilFunc(GL_ALWAYS, 0, ~0);
+		glColorMask(GL_FALSE, GL_FALSE, GL_FALSE, GL_FALSE);
+
+		glStencilOpSeparate(GL_FRONT, GL_KEEP, GL_KEEP, GL_INCR_WRAP);
+		glStencilOpSeparate(GL_BACK, GL_KEEP, GL_KEEP, GL_DECR_WRAP);
+		glDisable(GL_CULL_FACE);
+		for (i = 0; i < npaths; i++) {
+			path = &paths[i];
+			glVertexPointer(2, GL_FLOAT, sizeof(struct NVGvertex), &path->fill[0].x);
+			glDrawArrays(GL_TRIANGLE_FAN, 0, path->nfill);
+		}
 
-    // Draw aliased off-pixels
-	glColorMask(GL_TRUE, GL_TRUE, GL_TRUE, GL_TRUE);
-	glEnable(GL_BLEND);
+		glEnable(GL_CULL_FACE);
 
-	glStencilFunc(GL_EQUAL, 0x00, 0xff);
-	glStencilOp(GL_KEEP, GL_KEEP, GL_KEEP);
+	    // Draw aliased off-pixels
+		glColorMask(GL_TRUE, GL_TRUE, GL_TRUE, GL_TRUE);
+		glEnable(GL_BLEND);
 
-	glEnableClientState(GL_TEXTURE_COORD_ARRAY);
+		glStencilFunc(GL_EQUAL, 0x00, 0xff);
+		glStencilOp(GL_KEEP, GL_KEEP, GL_KEEP);
 
-	glnvg__setupPaint(gl, paint, scissor, 1.0001f, aasize);
+		glEnableClientState(GL_TEXTURE_COORD_ARRAY);
 
-	// Draw fringes
-	for (i = 0; i < npaths; i++) {
-		path = &paths[i];
-		glVertexPointer(2, GL_FLOAT, sizeof(struct NVGvertex), &path->stroke[0].x);
-		glTexCoordPointer(2, GL_FLOAT, sizeof(struct NVGvertex), &path->stroke[0].u);
-		glDrawArrays(GL_TRIANGLE_STRIP, 0, path->nstroke);
-	}
+		glnvg__setupPaint(gl, paint, scissor, 1.0001f, aasize);
 
-	glDisableClientState(GL_VERTEX_ARRAY);
-	glDisableClientState(GL_TEXTURE_COORD_ARRAY);
+		// Draw fringes
+		for (i = 0; i < npaths; i++) {
+			path = &paths[i];
+			glVertexPointer(2, GL_FLOAT, sizeof(struct NVGvertex), &path->stroke[0].x);
+			glTexCoordPointer(2, GL_FLOAT, sizeof(struct NVGvertex), &path->stroke[0].u);
+			glDrawArrays(GL_TRIANGLE_STRIP, 0, path->nstroke);
+		}
 
-	// Draw fill
-	glStencilFunc(GL_NOTEQUAL, 0x0, 0xff);
-	glStencilOp(GL_ZERO, GL_ZERO, GL_ZERO);
+		glDisableClientState(GL_VERTEX_ARRAY);
+		glDisableClientState(GL_TEXTURE_COORD_ARRAY);
 
-	glBegin(GL_QUADS);
-	glTexCoord2f(0.5f,1.0f);
-	glVertex2f(bounds[0], bounds[3]);
-	glVertex2f(bounds[2], bounds[3]);
-	glVertex2f(bounds[2], bounds[1]);
-	glVertex2f(bounds[0], bounds[1]);
-	glEnd();
+		// Draw fill
+		glStencilFunc(GL_NOTEQUAL, 0x0, 0xff);
+		glStencilOp(GL_ZERO, GL_ZERO, GL_ZERO);
 
-	glUseProgram(0);
+		glBegin(GL_QUADS);
+		glTexCoord2f(0.5f,1.0f);
+		glVertex2f(bounds[0], bounds[3]);
+		glVertex2f(bounds[2], bounds[3]);
+		glVertex2f(bounds[2], bounds[1]);
+		glVertex2f(bounds[0], bounds[1]);
+		glEnd();
 
-	glDisable(GL_TEXTURE_2D);
-	glDisable(GL_STENCIL_TEST);
+		glUseProgram(0);
+
+		glDisable(GL_TEXTURE_2D);
+		glDisable(GL_STENCIL_TEST);
+	}
 }
 
 static void glnvg__renderStroke(void* uptr, struct NVGpaint* paint, struct NVGscissor* scissor, float aasize,
diff --git a/src/nanovg_gl3.h b/src/nanovg_gl3.h
index 7134567..1975f5c 100644
--- a/src/nanovg_gl3.h
+++ b/src/nanovg_gl3.h
@@ -231,15 +231,6 @@ static void glnvg__getUniforms(struct GLNVGshader* shader)
 	shader->loc[GLNVG_LOC_TEXTYPE] = glGetUniformLocation(shader->prog, "texType");
 }
 
-
-static void glnvg__resizeVertexBuffer(struct GLNVGcontext* gl, int size)
-{
-    glBindVertexArray(gl->vertArr);
-	glBindBuffer(GL_ARRAY_BUFFER, gl->vertBuf);
-	glBufferData(GL_ARRAY_BUFFER, size * sizeof(struct NVGvertex), 0, GL_STREAM_DRAW);
-	gl->vertArrSize = size;
-}
-
 static int glnvg__renderCreate(void* uptr)
 {
 	struct GLNVGcontext* gl = (struct GLNVGcontext*)uptr;
@@ -577,97 +568,159 @@ static int glnvg__maxVertCount(const struct NVGpath* paths, int npaths)
 {
 	int i, count = 0;
 	for (i = 0; i < npaths; i++) {
-		if (paths[i].nfill > count) count = paths[i].nfill;
-		if (paths[i].nstroke > count) count = paths[i].nstroke;
+		count += paths[i].nfill;
+		count += paths[i].nstroke;
 	}
 	return count;
 }
 
+static void glnvg__uploadPaths(const struct NVGpath* paths, int npaths)
+{
+	const struct NVGpath* path;
+	int i, n = 0;
+	for (i = 0; i < npaths; i++) {
+		path = &paths[i];
+		if (path->nfill > 0) {
+			glBufferSubData(GL_ARRAY_BUFFER, n*sizeof(struct NVGvertex), path->nfill * sizeof(struct NVGvertex), &path->fill[0].x);
+			n += path->nfill;
+		}
+		if (path->nstroke > 0) {
+			glBufferSubData(GL_ARRAY_BUFFER, n*sizeof(struct NVGvertex), path->nstroke * sizeof(struct NVGvertex), &path->stroke[0].x);
+			n += path->nstroke;
+		}
+	}
+}
+
 static void glnvg__renderFill(void* uptr, struct NVGpaint* paint, struct NVGscissor* scissor, float aasize,
 							  const float* bounds, const struct NVGpath* paths, int npaths)
 {
 	struct GLNVGcontext* gl = (struct GLNVGcontext*)uptr;
 	const struct NVGpath* path;
-	int i, maxCount;
-
-	maxCount = glnvg__maxVertCount(paths, npaths);
-	if (maxCount > gl->vertArrSize)
-		glnvg__resizeVertexBuffer(gl, maxCount);
+	int i, n, offset, maxCount;
 
 	if (gl->gradShader.prog == 0)
 		return;
 
-	glEnable(GL_CULL_FACE);
-
+	maxCount = glnvg__maxVertCount(paths, npaths);
 	glBindVertexArray(gl->vertArr);
 	glBindBuffer(GL_ARRAY_BUFFER, gl->vertBuf);
+	glBufferData(GL_ARRAY_BUFFER, maxCount * sizeof(struct NVGvertex), NULL, GL_STREAM_DRAW);
+	glnvg__uploadPaths(paths, npaths);
+
+	if (npaths == 1 && paths[0].convex) {
+
+		glEnable(GL_CULL_FACE);
+
+		glEnableVertexAttribArray(0);
+		glEnableVertexAttribArray(1);
+		glnvg__setupPaint(gl, paint, scissor, 1.0001f, aasize);
+
+		glDisable(GL_CULL_FACE);
+		n = 0;
+		for (i = 0; i < npaths; i++) {
+			path = &paths[i];
+			offset = n * sizeof(struct NVGvertex);
+			glVertexAttribPointer(0, 2, GL_FLOAT, GL_FALSE, sizeof(struct NVGvertex), (const GLvoid*)(size_t)offset);
+			glVertexAttribPointer(1, 2, GL_FLOAT, GL_FALSE, sizeof(struct NVGvertex), (const GLvoid*)(offset + 2*sizeof(float)));
+			glDrawArrays(GL_TRIANGLE_FAN, 0, path->nfill);
+			n += path->nfill + path->nstroke;
+		}
 
-	// Draw shapes
-	glDisable(GL_BLEND);
-	glEnable(GL_STENCIL_TEST);
-	glStencilMask(0xff);
-	glStencilFunc(GL_ALWAYS, 0, ~0);
-	glColorMask(GL_FALSE, GL_FALSE, GL_FALSE, GL_FALSE);
+		glEnable(GL_CULL_FACE);
+
+		// Draw fringes
+		n = 0;
+		for (i = 0; i < npaths; i++) {
+			path = &paths[i];
+			offset = (n + path->nfill) * sizeof(struct NVGvertex);
+			glVertexAttribPointer(0, 2, GL_FLOAT, GL_FALSE, sizeof(struct NVGvertex), (const GLvoid*)(size_t)offset);
+			glVertexAttribPointer(1, 2, GL_FLOAT, GL_FALSE, sizeof(struct NVGvertex), (const GLvoid*)(offset + 2*sizeof(float)));
+			glDrawArrays(GL_TRIANGLE_STRIP, 0, path->nstroke);
+			n += path->nfill + path->nstroke;
+		}
 
-	glUseProgram(gl->solidShader.prog);
-	glUniform2f(gl->solidShader.loc[GLNVG_LOC_VIEWPOS], 0, 0);
-	glUniform2f(gl->solidShader.loc[GLNVG_LOC_VIEWSIZE], gl->viewWidth, gl->viewHeight);
-	glnvg__checkError("fill solid loc");
+		glUseProgram(0);
+		glDisableVertexAttribArray(0);
+		glDisableVertexAttribArray(1);
 
-	glEnableVertexAttribArray(0);
+	} else {
 
-	glStencilOpSeparate(GL_FRONT, GL_KEEP, GL_KEEP, GL_INCR_WRAP);
-	glStencilOpSeparate(GL_BACK, GL_KEEP, GL_KEEP, GL_DECR_WRAP);
-	glDisable(GL_CULL_FACE);
-	for (i = 0; i < npaths; i++) {
-		path = &paths[i];
-		glBufferSubData(GL_ARRAY_BUFFER, 0, path->nfill * sizeof(struct NVGvertex), &path->fill[0].x);
-		glVertexAttribPointer(0, 2, GL_FLOAT, GL_FALSE, sizeof(struct NVGvertex), (const GLvoid*)0);
-		glDrawArrays(GL_TRIANGLE_FAN, 0, path->nfill);
-	}
+		glEnable(GL_CULL_FACE);
+
+		glBindVertexArray(gl->vertArr);
+		glBindBuffer(GL_ARRAY_BUFFER, gl->vertBuf);
+
+		// Draw shapes
+		glDisable(GL_BLEND);
+		glEnable(GL_STENCIL_TEST);
+		glStencilMask(0xff);
+		glStencilFunc(GL_ALWAYS, 0, ~0);
+		glColorMask(GL_FALSE, GL_FALSE, GL_FALSE, GL_FALSE);
+
+		glUseProgram(gl->solidShader.prog);
+		glUniform2f(gl->solidShader.loc[GLNVG_LOC_VIEWPOS], 0, 0);
+		glUniform2f(gl->solidShader.loc[GLNVG_LOC_VIEWSIZE], gl->viewWidth, gl->viewHeight);
+		glnvg__checkError("fill solid loc");
+
+		glEnableVertexAttribArray(0);
+
+		glStencilOpSeparate(GL_FRONT, GL_KEEP, GL_KEEP, GL_INCR_WRAP);
+		glStencilOpSeparate(GL_BACK, GL_KEEP, GL_KEEP, GL_DECR_WRAP);
+		glDisable(GL_CULL_FACE);
+		n = 0;
+		for (i = 0; i < npaths; i++) {
+			path = &paths[i];
+			offset = n * sizeof(struct NVGvertex);
+			glVertexAttribPointer(0, 2, GL_FLOAT, GL_FALSE, sizeof(struct NVGvertex), (const GLvoid*)(size_t)offset);
+			glDrawArrays(GL_TRIANGLE_FAN, 0, path->nfill);
+			n += path->nfill + path->nstroke;
+		}
 
-	glEnable(GL_CULL_FACE);
+		glEnable(GL_CULL_FACE);
 
-    // Draw aliased off-pixels
-	glColorMask(GL_TRUE, GL_TRUE, GL_TRUE, GL_TRUE);
-	glEnable(GL_BLEND);
+		// Draw aliased off-pixels
+		glColorMask(GL_TRUE, GL_TRUE, GL_TRUE, GL_TRUE);
+		glEnable(GL_BLEND);
 
-	glStencilFunc(GL_EQUAL, 0x00, 0xff);
-	glStencilOp(GL_KEEP, GL_KEEP, GL_KEEP);
+		glStencilFunc(GL_EQUAL, 0x00, 0xff);
+		glStencilOp(GL_KEEP, GL_KEEP, GL_KEEP);
 
-	glEnableVertexAttribArray(1);
+		glEnableVertexAttribArray(1);
 
-	glnvg__setupPaint(gl, paint, scissor, 1.0001f, aasize);
+		glnvg__setupPaint(gl, paint, scissor, 1.0001f, aasize);
 
-	// Draw fringes
-	for (i = 0; i < npaths; i++) {
-		path = &paths[i];
-		glBufferSubData(GL_ARRAY_BUFFER, 0, path->nstroke * sizeof(struct NVGvertex), &path->stroke[0].x);
-		glVertexAttribPointer(0, 2, GL_FLOAT, GL_FALSE, sizeof(struct NVGvertex), (const GLvoid*)0);
-		glVertexAttribPointer(1, 2, GL_FLOAT, GL_FALSE, sizeof(struct NVGvertex), (const GLvoid*)(2 * sizeof(float)));
-		glDrawArrays(GL_TRIANGLE_STRIP, 0, path->nstroke);
-	}
+		// Draw fringes
+		n = 0;
+		for (i = 0; i < npaths; i++) {
+			path = &paths[i];
+			offset = (n + path->nfill) * sizeof(struct NVGvertex);
+			glVertexAttribPointer(0, 2, GL_FLOAT, GL_FALSE, sizeof(struct NVGvertex), (const GLvoid*)(size_t)offset);
+			glVertexAttribPointer(1, 2, GL_FLOAT, GL_FALSE, sizeof(struct NVGvertex), (const GLvoid*)(offset + 2*sizeof(float)));
+			glDrawArrays(GL_TRIANGLE_STRIP, 0, path->nstroke);
+			n += path->nfill + path->nstroke;
+		}
 
-	// Draw fill
-	glStencilFunc(GL_NOTEQUAL, 0x0, 0xff);
-	glStencilOp(GL_ZERO, GL_ZERO, GL_ZERO);
+		// Draw fill
+		glStencilFunc(GL_NOTEQUAL, 0x0, 0xff);
+		glStencilOp(GL_ZERO, GL_ZERO, GL_ZERO);
 
-	glDisableVertexAttribArray(1);
+		glDisableVertexAttribArray(1);
 
-	float quad[6*2] = {
-		bounds[0], bounds[3], bounds[2], bounds[3], bounds[2], bounds[1],
-		bounds[0], bounds[3], bounds[2], bounds[1], bounds[0], bounds[1],
-	};
-	glBufferSubData(GL_ARRAY_BUFFER, 0, 6 * 2*sizeof(float), quad);
-	glVertexAttribPointer(0, 2, GL_FLOAT, GL_FALSE, 2*sizeof(float), (const GLvoid*)0);
-	glVertexAttrib2f(1, 0.5f, 1.0f);
-	glDrawArrays(GL_TRIANGLES, 0, 6);
+		float quad[6*2] = {
+			bounds[0], bounds[3], bounds[2], bounds[3], bounds[2], bounds[1],
+			bounds[0], bounds[3], bounds[2], bounds[1], bounds[0], bounds[1],
+		};
+		glBufferSubData(GL_ARRAY_BUFFER, 0, 6 * 2*sizeof(float), quad);
+		glVertexAttribPointer(0, 2, GL_FLOAT, GL_FALSE, 2*sizeof(float), (const GLvoid*)0);
+		glVertexAttrib2f(1, 0.5f, 1.0f);
+		glDrawArrays(GL_TRIANGLES, 0, 6);
 
-	glUseProgram(0);
+		glUseProgram(0);
 
-	glDisableVertexAttribArray(0);
+		glDisableVertexAttribArray(0);
 
-	glDisable(GL_STENCIL_TEST);
+		glDisable(GL_STENCIL_TEST);
+	}
 }
 
 static void glnvg__renderStroke(void* uptr, struct NVGpaint* paint, struct NVGscissor* scissor, float aasize,
@@ -675,11 +728,7 @@ static void glnvg__renderStroke(void* uptr, struct NVGpaint* paint, struct NVGsc
 {
 	struct GLNVGcontext* gl = (struct GLNVGcontext*)uptr;
 	const struct NVGpath* path;
-	int i, maxCount;
-
-	maxCount = glnvg__maxVertCount(paths, npaths);
-	if (maxCount > gl->vertArrSize)
-		glnvg__resizeVertexBuffer(gl, maxCount);
+	int i, n, offset, maxCount;
 
 	if (gl->gradShader.prog == 0)
 		return;
@@ -688,19 +737,24 @@ static void glnvg__renderStroke(void* uptr, struct NVGpaint* paint, struct NVGsc
 
 	glEnable(GL_CULL_FACE);
 
+	maxCount = glnvg__maxVertCount(paths, npaths);
 	glBindVertexArray(gl->vertArr);
 	glBindBuffer(GL_ARRAY_BUFFER, gl->vertBuf);
+	glBufferData(GL_ARRAY_BUFFER, maxCount * sizeof(struct NVGvertex), NULL, GL_STREAM_DRAW);
+	glnvg__uploadPaths(paths, npaths);
 
 	glEnableVertexAttribArray(0);
 	glEnableVertexAttribArray(1);
 
 	// Draw Strokes
+	n = 0;
 	for (i = 0; i < npaths; i++) {
 		path = &paths[i];
-		glBufferSubData(GL_ARRAY_BUFFER, 0, path->nstroke * sizeof(struct NVGvertex), &path->stroke[0].x);
-		glVertexAttribPointer(0, 2, GL_FLOAT, GL_FALSE, sizeof(struct NVGvertex), (const GLvoid*)0);
-		glVertexAttribPointer(1, 2, GL_FLOAT, GL_FALSE, sizeof(struct NVGvertex), (const GLvoid*)(2 * sizeof(float)));
+		offset = (n + path->nfill) * sizeof(struct NVGvertex);
+		glVertexAttribPointer(0, 2, GL_FLOAT, GL_FALSE, sizeof(struct NVGvertex), (const GLvoid*)(size_t)offset);
+		glVertexAttribPointer(1, 2, GL_FLOAT, GL_FALSE, sizeof(struct NVGvertex), (const GLvoid*)(offset + 2*sizeof(float)));
 		glDrawArrays(GL_TRIANGLE_STRIP, 0, path->nstroke);
+		n += path->nfill + path->nstroke;
 	}
 
 	glDisableVertexAttribArray(0);
@@ -716,9 +770,6 @@ static void glnvg__renderTriangles(void* uptr, struct NVGpaint* paint, struct NV
 	struct GLNVGtexture* tex = glnvg__findTexture(gl, image);
 	float color[4];
 
-	if (nverts > gl->vertArrSize)
-		glnvg__resizeVertexBuffer(gl, nverts);
-
 	if (tex != NULL) {
 		glBindTexture(GL_TEXTURE_2D, tex->tex);
 	}
@@ -733,7 +784,8 @@ static void glnvg__renderTriangles(void* uptr, struct NVGpaint* paint, struct NV
 
 	glBindVertexArray(gl->vertArr);
 	glBindBuffer(GL_ARRAY_BUFFER, gl->vertBuf);
-	glBufferSubData(GL_ARRAY_BUFFER, 0, nverts * sizeof(struct NVGvertex), verts);
+	glBufferData(GL_ARRAY_BUFFER, nverts * sizeof(struct NVGvertex), verts, GL_STREAM_DRAW);
+
 	glVertexAttribPointer(0, 2, GL_FLOAT, GL_FALSE, sizeof(struct NVGvertex), (const GLvoid*)0);
 	glVertexAttribPointer(1, 2, GL_FLOAT, GL_FALSE, sizeof(struct NVGvertex), (const GLvoid*)(2 * sizeof(float)));
 	glEnableVertexAttribArray(0);