diff --git a/demos/gtk-demo/iconscroll.c b/demos/gtk-demo/iconscroll.c
index 2eb9aa6e92..8212818516 100644
--- a/demos/gtk-demo/iconscroll.c
+++ b/demos/gtk-demo/iconscroll.c
@@ -116,20 +116,25 @@ populate_emoji_text (void)
   GtkWidget *textview;
   GtkTextBuffer *buffer;
   GString *s;
+  GtkTextIter iter;
 
-  s = g_string_sized_new (1000 * 30 * 4);
+  s = g_string_sized_new (500 * 30 * 4);
 
-  for (int i = 0; i < 1000; i++)
+  for (int i = 0; i < 500; i++)
     {
       if (i % 2)
-        g_string_append (s, "x");
+        g_string_append (s, "<span underline=\"single\" underline_color=\"red\">x</span>");
       for (int j = 0; j < 30; j++)
-        g_string_append (s, "💓x");
+        {
+          g_string_append (s, "💓");
+          g_string_append (s, "<span underline=\"single\" underline_color=\"red\">x</span>");
+        }
       g_string_append (s, "\n");
     }
 
   buffer = gtk_text_buffer_new (NULL);
-  gtk_text_buffer_set_text (buffer, s->str, s->len);
+  gtk_text_buffer_get_start_iter (buffer, &iter);
+  gtk_text_buffer_insert_markup (buffer, &iter, s->str, s->len);
 
   g_string_free (s, TRUE);
 
diff --git a/gsk/ngl/gsknglcommandqueue.c b/gsk/ngl/gsknglcommandqueue.c
index bdfef2f35c..ca84d04774 100644
--- a/gsk/ngl/gsknglcommandqueue.c
+++ b/gsk/ngl/gsknglcommandqueue.c
@@ -169,7 +169,7 @@ gsk_ngl_command_queue_print_batch (GskNglCommandQueue       *self,
       for (guint i = 0; i < batch->draw.bind_count; i++)
         {
           const GskNglCommandBind *bind = &self->batch_binds.items[batch->draw.bind_offset + i];
-          g_print ("      Bind[%d]: %u\n", bind->texture, bind->id);
+          g_printerr ("      Bind[%d]: %u\n", bind->texture, bind->id);
         }
 
       for (guint i = 0; i < batch->draw.uniform_count; i++)
@@ -951,6 +951,7 @@ gsk_ngl_command_queue_execute (GskNglCommandQueue   *self,
   guint n_binds = 0;
   guint n_fbos = 0;
   guint n_uniforms = 0;
+  guint n_programs = 0;
   guint vao_id;
   guint vbo_id;
   int textures[4];
@@ -1062,6 +1063,8 @@ gsk_ngl_command_queue_execute (GskNglCommandQueue   *self,
             {
               program = batch->any.program;
               glUseProgram (program);
+
+              n_programs++;
             }
 
           if (apply_framebuffer (&framebuffer, batch->draw.framebuffer))
@@ -1144,6 +1147,7 @@ gsk_ngl_command_queue_execute (GskNglCommandQueue   *self,
   gdk_profiler_set_int_counter (self->metrics.n_binds, n_binds);
   gdk_profiler_set_int_counter (self->metrics.n_uniforms, n_uniforms);
   gdk_profiler_set_int_counter (self->metrics.n_fbos, n_fbos);
+  gdk_profiler_set_int_counter (self->metrics.n_programs, n_programs);
   gdk_profiler_set_int_counter (self->metrics.n_uploads, self->n_uploads);
   gdk_profiler_set_int_counter (self->metrics.queue_depth, self->batches.len);
 
@@ -1415,6 +1419,7 @@ gsk_ngl_command_queue_set_profiler (GskNglCommandQueue *self,
       self->metrics.n_fbos = gdk_profiler_define_int_counter ("fbos", "Number of framebuffers attached");
       self->metrics.n_uniforms = gdk_profiler_define_int_counter ("uniforms", "Number of uniforms changed");
       self->metrics.n_uploads = gdk_profiler_define_int_counter ("uploads", "Number of texture uploads");
+      self->metrics.n_programs = gdk_profiler_define_int_counter ("programs", "Number of program changes");
       self->metrics.queue_depth = gdk_profiler_define_int_counter ("gl-queue-depth", "Depth of GL command batches");
     }
 #endif
diff --git a/gsk/ngl/gsknglcommandqueueprivate.h b/gsk/ngl/gsknglcommandqueueprivate.h
index bcd7c8312d..a05095dbf0 100644
--- a/gsk/ngl/gsknglcommandqueueprivate.h
+++ b/gsk/ngl/gsknglcommandqueueprivate.h
@@ -250,6 +250,7 @@ struct _GskNglCommandQueue
     guint n_fbos;
     guint n_uniforms;
     guint n_uploads;
+    guint n_programs;
     guint queue_depth;
   } metrics;
 
diff --git a/gsk/ngl/gsknglrenderjob.c b/gsk/ngl/gsknglrenderjob.c
index b1bda44790..69c92ac80f 100644
--- a/gsk/ngl/gsknglrenderjob.c
+++ b/gsk/ngl/gsknglrenderjob.c
@@ -1360,16 +1360,54 @@ blur_node (GskNglRenderJob       *job,
   *max_y = job->offset_y + node->bounds.origin.y + node->bounds.size.height + half_blur_extra;
 }
 
+#define ATLAS_SIZE 512
+
 static inline void
 gsk_ngl_render_job_visit_color_node (GskNglRenderJob     *job,
                                      const GskRenderNode *node)
 {
   guint16 color[4];
+  GskNglProgram *program;
+  GskNglCommandBatch *batch;
 
   rgba_to_half (gsk_color_node_get_color (node), color);
-  gsk_ngl_render_job_begin_draw (job, CHOOSE_PROGRAM (job, color));
-  gsk_ngl_render_job_draw_rect_with_color (job, &node->bounds, color);
-  gsk_ngl_render_job_end_draw (job);
+
+  /* Avoid switching away from the coloring program for
+   * rendering a solid color.
+   */
+  program = CHOOSE_PROGRAM (job, coloring);
+  batch = gsk_ngl_command_queue_get_batch (job->command_queue);
+
+  if (batch->any.kind == GSK_NGL_COMMAND_KIND_DRAW &&
+      batch->any.program == program->id)
+    {
+      GskNglRenderOffscreen offscreen = {0};
+
+      gsk_ngl_render_job_begin_draw (job, program);
+
+      /* The top left few pixels in our atlases are always
+       * solid white, so we can use it here, without
+       * having to choose any particular atlas texture.
+       */
+      offscreen.was_offscreen = FALSE;
+      offscreen.area.x = 1.f / ATLAS_SIZE;
+      offscreen.area.y = 1.f / ATLAS_SIZE;
+      offscreen.area.x2 = 2.f / ATLAS_SIZE;
+      offscreen.area.y2 = 2.f / ATLAS_SIZE;
+
+      gsk_ngl_render_job_draw_offscreen_with_color (job,
+                                                    &node->bounds,
+                                                    &offscreen,
+                                                    color);
+
+      gsk_ngl_render_job_end_draw (job);
+    }
+  else
+    {
+      gsk_ngl_render_job_begin_draw (job, CHOOSE_PROGRAM (job, color));
+      gsk_ngl_render_job_draw_rect_with_color (job, &node->bounds, color);
+      gsk_ngl_render_job_end_draw (job);
+    }
 }
 
 static inline void
diff --git a/gsk/ngl/gskngltexturelibrary.c b/gsk/ngl/gskngltexturelibrary.c
index 85449b3fa2..2de6e492c5 100644
--- a/gsk/ngl/gskngltexturelibrary.c
+++ b/gsk/ngl/gskngltexturelibrary.c
@@ -20,6 +20,7 @@
 
 #include "config.h"
 
+#include <gdk/gdkglcontextprivate.h>
 #include <gsk/gskdebugprivate.h>
 
 #include "gsknglcommandqueueprivate.h"
@@ -239,6 +240,51 @@ gsk_ngl_texture_atlas_pack (GskNglTextureAtlas *self,
   return rect.was_packed;
 }
 
+static void
+gsk_ngl_texture_atlas_initialize (GskNglDriver       *driver,
+                                  GskNglTextureAtlas *atlas)
+{
+  /* Insert a single pixel at 0,0 for use in coloring */
+
+  gboolean packed;
+  int x, y;
+  guint gl_format;
+  guint gl_type;
+  guint8 pixel_data[4 * 3 * 3];
+
+  gdk_gl_context_push_debug_group_printf (gdk_gl_context_get_current (),
+                                          "Initializing Atlas");
+
+  packed = gsk_ngl_texture_atlas_pack (atlas, 3, 3, &x, &y);
+  g_assert (packed);
+  g_assert (x == 0 && y == 0);
+
+  memset (pixel_data, 255, sizeof pixel_data);
+
+  if (gdk_gl_context_get_use_es (gdk_gl_context_get_current ()))
+    {
+      gl_format = GL_RGBA;
+      gl_type = GL_UNSIGNED_BYTE;
+    }
+  else
+    {
+      gl_format = GL_BGRA;
+      gl_type = GL_UNSIGNED_INT_8_8_8_8_REV;
+    }
+
+  glBindTexture (GL_TEXTURE_2D, atlas->texture_id);
+
+  glTexSubImage2D (GL_TEXTURE_2D, 0,
+                   0, 0,
+                   3, 3,
+                   gl_format, gl_type,
+                   pixel_data);
+
+  gdk_gl_context_pop_debug_group (gdk_gl_context_get_current ());
+
+  driver->command_queue->n_uploads++;
+}
+
 static void
 gsk_ngl_texture_atlases_pack (GskNglDriver        *driver,
                               int                  width,
@@ -265,6 +311,8 @@ gsk_ngl_texture_atlases_pack (GskNglDriver        *driver,
       /* No atlas has enough space, so create a new one... */
       atlas = gsk_ngl_driver_create_atlas (driver);
 
+      gsk_ngl_texture_atlas_initialize (driver, atlas);
+
       /* Pack it onto that one, which surely has enough space... */
       if (!gsk_ngl_texture_atlas_pack (atlas, width, height, &x, &y))
         g_assert_not_reached ();