diff --git a/c++/contourgl/cl/contour-base.cl b/c++/contourgl/cl/contour-base.cl
index a29a829..fa16b9c 100644
--- a/c++/contourgl/cl/contour-base.cl
+++ b/c++/contourgl/cl/contour-base.cl
@@ -124,11 +124,6 @@ kernel void fill(
 	//int ialpha;
 	int icover = 0, c0 = bounds.s0, c1 = bounds.s0;
 	while(c0 < bounds.s2) {
-		c1 = min(c1, bounds.s2);
-		mark = &row[c1];
-		m = *mark;
-		*mark = (int4)(0, 0, c1 | (c1 + 1), 0); 
-		
 		//ialpha = abs(icover);
 		//ialpha = evenodd ? ONE - abs((ialpha % TWO) - ONE)
 		//				 : min(ialpha, ONE);
@@ -139,6 +134,10 @@ kernel void fill(
 	
 		if (c1 >= bounds.s2) return;
 		
+		mark = &row[c1];
+		m = *mark;
+		*mark = (int4)(0, 0, c1 | (c1 + 1), 0); 
+		
 		//ialpha = abs(mark.x + icover);
 		//ialpha = evenodd ? ONE - abs((ialpha % TWO) - ONE)
 		//				 : min(ialpha, ONE);
@@ -150,7 +149,7 @@ kernel void fill(
 				           min(pixel->w + alpha, 1.f) );
 		
 		c0 = c1 + 1;
-		c1 = m.z;
+		c1 = min(m.z, bounds.s2);
 		icover += m.y;
 	}
 }
diff --git a/c++/contourgl/clcontext.cpp b/c++/contourgl/clcontext.cpp
index 1d35e13..5d1b7bb 100644
--- a/c++/contourgl/clcontext.cpp
+++ b/c++/contourgl/clcontext.cpp
@@ -100,17 +100,42 @@ ClContext::ClContext():
     max_group_size = max_group_sizes.front();
     //cout << "Device " << device_index << " max group size " << max_group_size << endl;
 
-    // context
+	size_t timer_resolution;
+    err |= clGetDeviceInfo(device, CL_DEVICE_PROFILING_TIMER_RESOLUTION, sizeof(timer_resolution), &timer_resolution, NULL);
+    assert(!err);
+    //cout << "Device " << device_index << " timer resolution " << timer_resolution << " ns" << endl;
+
+    unsigned long long global_mem_size;
+    err |= clGetDeviceInfo(device, CL_DEVICE_GLOBAL_MEM_SIZE, sizeof(global_mem_size), &global_mem_size, NULL);
+    assert(!err);
+    //cout << "Device " << device_index << " global mem size " << global_mem_size << endl;
+
+    unsigned long long local_mem_size;
+    err |= clGetDeviceInfo(device, CL_DEVICE_LOCAL_MEM_SIZE, sizeof(local_mem_size), &local_mem_size, NULL);
+    assert(!err);
+    //cout << "Device " << device_index << " local mem size " << local_mem_size << endl;
+
+    unsigned long long max_constant_buffer_size;
+    err |= clGetDeviceInfo(device, CL_DEVICE_MAX_CONSTANT_BUFFER_SIZE, sizeof(max_constant_buffer_size), &max_constant_buffer_size, NULL);
+    assert(!err);
+    //cout << "Device " << device_index << " max constant buffer size " << max_constant_buffer_size << endl;
+
+	// context
 
     cl_context_properties context_props[] = {
-    	CL_CONTEXT_PLATFORM, (cl_context_properties)platform,
+    	CL_CONTEXT_PLATFORM,       (cl_context_properties)platform,
 		CL_NONE };
     context = clCreateContext(context_props, 1, &device, callback, NULL, &err);
     assert(context);
 
 	// command queue
 
-	queue = clCreateCommandQueue(context, device, CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE, NULL);
+    cl_command_queue_properties props = 0
+    	| CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE
+    	//| CL_QUEUE_PROFILING_ENABLE
+    	| 0;
+	queue = clCreateCommandQueue(
+		context, device, props, NULL);
 	assert(queue);
 
 	//hello();