Beginnings of a demand-filled post-tnl vertex cache.

Probably breaks a bit of stuff, eg unfilled clipping, edgeflags, etc.

4 files changed, 392 insertions, 302 deletions

diff --git a/src/mesa/pipe/draw/draw_context.c b/src/mesa/pipe/draw/draw_context.c
index a97f4883873..4335b47e094 100644
--- a/src/mesa/pipe/draw/draw_context.c
+++ b/src/mesa/pipe/draw/draw_context.c
@@ -59,6 +59,16 @@ struct draw_context *draw_create( void )
 
    draw->vf = vf_create( GL_TRUE );
 
+   /* Statically allocate maximum sized vertices for the cache - could be cleverer...
+    */
+   {
+      int i;
+      char *tmp = malloc(Elements(draw->vcache.vertex) * MAX_VERTEX_SIZE);
+
+      for (i = 0; i < Elements(draw->vcache.vertex); i++)
+	 draw->vcache.vertex[i] = (struct vertex_header *)(tmp + i * MAX_VERTEX_SIZE);
+   }
+
    return draw;
 }
 
@@ -70,6 +80,7 @@ void draw_destroy( struct draw_context *draw )
 
    vf_destroy( draw->vf );
 
+   FREE( draw->vcache.vertex[0] ); /* Frees all the vertices. */
    FREE( draw );
 }
 
diff --git a/src/mesa/pipe/draw/draw_private.h b/src/mesa/pipe/draw/draw_private.h
index 3dfaa0581db..597393afdc9 100644
--- a/src/mesa/pipe/draw/draw_private.h
+++ b/src/mesa/pipe/draw/draw_private.h
@@ -61,12 +61,18 @@ struct vertex_header {
    GLfloat data[][4];		/* Note variable size */
 };
 
+#define MAX_VERTEX_SIZE ((2 + FRAG_ATTRIB_MAX) * 4 * sizeof(GLfloat))
+
+
 
 /**
  * Basic info for a point/line/triangle primitive.
  */
 struct prim_header {
    GLfloat det;                 /**< front/back face determinant */
+   GLuint reset_line_stipple:1;
+   GLuint edgeflags:3;
+   GLuint pad:28;
    struct vertex_header *v[3];  /**< 1 to 3 vertex pointers */
 };
 
@@ -103,6 +109,12 @@ struct draw_stage
 };
 
 
+#define PRIM_QUEUE_LENGTH      16
+#define VCACHE_SIZE            32
+#define VCACHE_OVERFLOW        4
+#define VS_QUEUE_LENGTH        (VCACHE_SIZE + VCACHE_OVERFLOW + 1)	/* can never fill up */
+
+
 /**
  * Private context for the drawing module.
  */
@@ -141,7 +153,45 @@ struct draw_context
    GLuint nr_vertices;
    GLboolean in_vb;
 
+   void *elts;
+
+   struct vertex_header *(*get_vertex)( struct draw_context *draw,
+					GLuint i );
+
+   /* Post-tnl vertex cache:
+    */
+   struct {
+      GLuint referenced;
+      GLuint idx[VCACHE_SIZE + VCACHE_OVERFLOW];
+      struct vertex_header *vertex[VCACHE_SIZE + VCACHE_OVERFLOW];
+      GLuint overflow;
+   } vcache;
+
+   /* Vertex shader queue:
+    */
+   struct {
+      struct {
+	 GLuint elt;
+	 struct vertex_header *dest;
+      } queue[VS_QUEUE_LENGTH];
+      GLuint queue_nr;
+   } vs;
+
+   /* Prim pipeline queue:
+    */
+   struct {
+
+      /* Need to queue up primitives until their vertices have been
+       * transformed by a vs queue flush.
+       */
+      struct prim_header queue[PRIM_QUEUE_LENGTH];
+      GLuint queue_nr;
+   } pq;
+
+
+
    GLenum prim;   /**< GL_POINTS, GL_LINE_STRIP, GL_QUADS, etc */
+   unsigned reduced_prim;
 
    /* Helper for tnl:
     */
diff --git a/src/mesa/pipe/draw/draw_unfilled.c b/src/mesa/pipe/draw/draw_unfilled.c
index 05242d8d107..e0e486ebe6f 100644
--- a/src/mesa/pipe/draw/draw_unfilled.c
+++ b/src/mesa/pipe/draw/draw_unfilled.c
@@ -91,9 +91,9 @@ static void points( struct draw_stage *stage,
    struct vertex_header *v1 = header->v[1];
    struct vertex_header *v2 = header->v[2];
 
-   if (v0->edgeflag) point( stage, v0 );
-   if (v1->edgeflag) point( stage, v1 );
-   if (v2->edgeflag) point( stage, v2 );
+   if (header->edgeflags & 0x1) point( stage, v0 );
+   if (header->edgeflags & 0x2) point( stage, v1 );
+   if (header->edgeflags & 0x4) point( stage, v2 );
 }
 
 
@@ -104,9 +104,9 @@ static void lines( struct draw_stage *stage,
    struct vertex_header *v1 = header->v[1];
    struct vertex_header *v2 = header->v[2];
 
-   if (v0->edgeflag) line( stage, v0, v1 );
-   if (v1->edgeflag) line( stage, v1, v2 );
-   if (v2->edgeflag) line( stage, v2, v0 );
+   if (header->edgeflags & 0x1) line( stage, v0, v1 );
+   if (header->edgeflags & 0x2) line( stage, v1, v2 );
+   if (header->edgeflags & 0x4) line( stage, v2, v0 );
 }
 
 
diff --git a/src/mesa/pipe/draw/draw_vb.c b/src/mesa/pipe/draw/draw_vb.c
index f9c10e5f970..ef32f02ec10 100644
--- a/src/mesa/pipe/draw/draw_vb.c
+++ b/src/mesa/pipe/draw/draw_vb.c
@@ -40,356 +40,377 @@
 #include "draw_context.h"
 
 
-/* This file is a temporary set of hooks to allow us to use the tnl/
- * and vf/ modules until we have replacements in pipe.
- */
+#define RP_NONE  0
+#define RP_POINT 1
+#define RP_LINE  2
+#define RP_TRI   3
+
+static unsigned reduced_prim[GL_POLYGON + 1] = {
+   RP_POINT,
+   RP_LINE,
+   RP_LINE,
+   RP_LINE,
+   RP_TRI,
+   RP_TRI,
+   RP_TRI,
+   RP_TRI,
+   RP_TRI,
+   RP_TRI
+};
 
 
-static struct vertex_header *get_vertex( struct draw_context *pipe,
-					       GLuint i )
-{
-   return (struct vertex_header *)(pipe->verts + i * pipe->vertex_size);
-}
 
 
+/* This file is a temporary set of hooks to allow us to use the tnl/
+ * and vf/ modules until we have replacements in pipe.
+ */
 
-static void draw_allocate_vertices( struct draw_context *draw,
-				    GLuint nr_vertices )
+static void vs_flush( struct draw_context *draw )
 {
-   draw->nr_vertices = nr_vertices;
-   draw->verts = (GLubyte *) malloc( nr_vertices * draw->vertex_size );
+   unsigned i;
 
-   draw->pipeline.first->begin( draw->pipeline.first );
+   /* We're not really running a vertex shader yet, so flushing the vs
+    * queue is just a matter of building the vertices and returning.
+    */ 
+   /* Actually, I'm cheating even more and pre-building them still
+    * with the mesa/vf module.  So it's very easy...
+    */
+   for (i = 0; i < draw->vs.queue_nr; i++) {
+      /* Would do the following steps here:
+       *
+       * 1) Loop over vertex element descriptors, fetch data from each
+       *    to build the pre-tnl vertex.  This might require a new struct
+       *    to represent the pre-tnl vertex.
+       * 
+       * 2) Bundle groups of upto 4 pre-tnl vertices together and pass
+       *    to vertex shader.  
+       *
+       * 3) Do any necessary unswizzling, make sure vertex headers are
+       *    correctly populated, store resulting post-transformed
+       *    vertices in vcache.
+       *
+       * In this version, just do the last step:
+       */
+      unsigned elt = draw->vs.queue[i].elt;
+      struct vertex_header *dest = draw->vs.queue[i].dest;
+
+      /* Magic:
+       */
+      memcpy(dest, 
+	     draw->verts + elt * draw->vertex_size,
+	     draw->vertex_size);
+	     
+   }
+   draw->vs.queue_nr = 0;
 }
 
-static void draw_set_prim( struct draw_context *draw,
-			   GLenum prim )
+
+static void draw_flush( struct draw_context *draw )
 {
-   draw->prim = prim;
+   struct draw_stage *first = draw->pipeline.first;
+   unsigned i;
 
-   /* Not done yet - need to force edgeflags to 1 in strip/fan
-    * primitives.
+   /* Make sure all vertices are available:
     */
-#if 0
-   switch (prim) {
-   case GL_TRIANGLES:
-   case GL_POLYGON:
-   case GL_QUADS:
-   case GL_QUAD_STRIP:		/* yes, we need this */
-      respect_edgeflags( pipe, GL_TRUE );
+   vs_flush( draw );
+
+
+   switch (draw->reduced_prim) {
+   case RP_TRI:
+      for (i = 0; i < draw->pq.queue_nr; i++) {
+	 if (draw->pq.queue[i].reset_line_stipple)
+	    first->reset_stipple_counter( first );
+
+	 first->tri( first, &draw->pq.queue[i] );
+      }
       break;
+   case RP_LINE:
+      for (i = 0; i < draw->pq.queue_nr; i++) {
+	 if (draw->pq.queue[i].reset_line_stipple)
+	    first->reset_stipple_counter( first );
 
-   default:
-      respect_edgeflags( pipe, GL_FALSE );
+	 first->line( first, &draw->pq.queue[i] );
+      }
+      break;
+   case RP_POINT:
+      first->reset_stipple_counter( first );
+      for (i = 0; i < draw->pq.queue_nr; i++)
+	 first->point( first, &draw->pq.queue[i] );
       break;
    }
-#endif
-}
-			  
 
+   draw->pq.queue_nr = 0;
+   draw->vcache.referenced = 0;
+   draw->vcache.overflow = 0;
+}
 
-static void do_quad( struct draw_stage *first,
-		     struct vertex_header *v0,
-		     struct vertex_header *v1,
-		     struct vertex_header *v2,
-		     struct vertex_header *v3 )
+static void draw_invalidate_vcache( struct draw_context *draw )
 {
-   struct prim_header prim;
-
-   {
-      GLuint tmp = v1->edgeflag;
-      v1->edgeflag = 0;
-
-      prim.v[0] = v0;
-      prim.v[1] = v1;
-      prim.v[2] = v3;
-      first->tri( first, &prim );
+   unsigned i;
+
+   assert(draw->pq.queue_nr == 0);
+   assert(draw->vs.queue_nr == 0);
+   assert(draw->vcache.referenced == 0);
+   
+   for (i = 0; i < Elements( draw->vcache.idx ); i++)
+      draw->vcache.idx[i] = ~0;
+}
 
-      v1->edgeflag = tmp;
-   }
 
-   {
-      GLuint tmp = v3->edgeflag;
-      v3->edgeflag = 0;
+/* Return a pointer to a freshly queued primitive header.  Ensure that
+ * there is room in the vertex cache for a maximum of "nr_verts" new
+ * vertices.  Flush primitive and/or vertex queues if necessary to
+ * make space.
+ */
+static struct prim_header *get_queued_prim( struct draw_context *draw,
+					    GLuint nr_verts )
+{
+   if (draw->pq.queue_nr + 1 >= PRIM_QUEUE_LENGTH ||
+       draw->vcache.overflow + nr_verts >= VCACHE_OVERFLOW) 
+      draw_flush( draw );
 
-      prim.v[0] = v1;
-      prim.v[1] = v2;
-      prim.v[2] = v3;
-      first->tri( first, &prim );
+   /* The vs queue is sized so that this can never happen:
+    */
+   assert(draw->vs.queue_nr + nr_verts < VS_QUEUE_LENGTH);
 
-      v3->edgeflag = tmp;
-   }
+   return &draw->pq.queue[draw->pq.queue_nr++];
 }
 
 
-
-
-static void draw_indexed_prim( struct draw_context *draw,
-			       const GLuint *elts,
-			       GLuint count )
+/* Check if vertex is in cache, otherwise add it.  It won't go through
+ * VS yet, not until there is a flush operation or the VS queue fills up.  
+ */
+static struct vertex_header *get_vertex( struct draw_context *draw,
+					 GLuint i )
 {
-   struct draw_stage * const first = draw->pipeline.first;
-   struct prim_header prim;
-   GLuint i;
+   unsigned slot = (i + (i>>5)) & 31;
+   
+   /* Cache miss?
+    */
+   if (draw->vcache.idx[slot] != i) {
 
-   prim.det = 0;		/* valid from cull stage onwards */
-   prim.v[0] = 0;
-   prim.v[1] = 0;
-   prim.v[2] = 0;
+      /* If slot is in use, use the overflow area:
+       */
+      if (draw->vcache.referenced & (1<<slot)) 
+	 slot = draw->vcache.overflow++;
 
-   switch (draw->prim) {
-   case GL_POINTS:
-      for (i = 0; i < count; i ++) {
-	 prim.v[0] = get_vertex( draw, elts[i] );
+      draw->vcache.idx[slot] = i;
 
-	 first->point( first, &prim );
-      }
-      break;
+      /* Add to vertex shader queue:
+       */
+      draw->vs.queue[draw->vs.queue_nr].dest = draw->vcache.vertex[slot];
+      draw->vs.queue[draw->vs.queue_nr].elt = i;
+      draw->vs.queue_nr++;
+   }
 
-   case GL_LINES:
-      for (i = 0; i+1 < count; i += 2) {
-	 prim.v[0] = get_vertex( draw, elts[i + 0] );
-	 prim.v[1] = get_vertex( draw, elts[i + 1] );
-      
-         first->reset_stipple_counter( first );
-	 first->line( first, &prim );
-      }
-      break;
+   /* Mark slot as in-use:
+    */
+   draw->vcache.referenced |= (1<<slot);
+   return draw->vcache.vertex[slot];
+}
 
-   case GL_LINE_LOOP:  
-      if (count >= 2) {
-         first->reset_stipple_counter( first );
-	 for (i = 1; i < count; i++) {
-	    prim.v[0] = get_vertex( draw, elts[i-1] );
-	    prim.v[1] = get_vertex( draw, elts[i] );	    
-	    first->line( first, &prim );
-	 }
 
-	 prim.v[0] = get_vertex( draw, elts[count-1] );
-	 prim.v[1] = get_vertex( draw, elts[0] );	    
-	 first->line( first, &prim );
-      }
-      break;
+static struct vertex_header *get_uint_elt_vertex( struct draw_context *draw,
+						    GLuint i )
+{
+   const GLuint *elts = (const GLuint *)draw->elts;
+   return get_vertex( draw, elts[i] );
+}
 
-   case GL_LINE_STRIP:
-      /* I'm guessing it will be necessary to have something like a
-       * render->reset_line_stipple() method to properly support
-       * splitting strips into primitives like this.  Alternately we
-       * could just scan ahead to find individual clipped lines and
-       * otherwise leave the strip intact - that might be better, but
-       * require more complex code here.
-       */
-      if (count >= 2) {
-         first->reset_stipple_counter( first );
-	 prim.v[0] = 0;
-	 prim.v[1] = get_vertex( draw, elts[0] );
-	 
-	 for (i = 1; i < count; i++) {
-	    prim.v[0] = prim.v[1];
-	    prim.v[1] = get_vertex( draw, elts[i] );
-	    
-	    first->line( first, &prim );
-	 }
-      }
-      break;
+#if 0
+static struct vertex_header *get_ushort_elt_vertex( struct draw_context *draw,
+						    const void *elts,
+						    GLuint i )
+{
+   const GLushort *elts = (const GLushort *)draw->elts;
+   return get_vertex( draw, elts[i] );
+}
 
-   case GL_TRIANGLES:
-      for (i = 0; i+2 < count; i += 3) {
-	 prim.v[0] = get_vertex( draw, elts[i + 0] );
-	 prim.v[1] = get_vertex( draw, elts[i + 1] );
-	 prim.v[2] = get_vertex( draw, elts[i + 2] );
-      
-	 first->tri( first, &prim );
-      }
-      break;
+static struct vertex_header *get_ubyte_elt_vertex( struct draw_context *draw,
+						    const void *elts,
+						    GLuint i )
+{
+   const GLubyte *elts = (const GLubyte *)draw->elts;
+   return get_vertex( draw, elts[i] );
+}
+#endif
 
-   case GL_TRIANGLE_STRIP:
-      for (i = 0; i+2 < count; i++) {
-	 if (i & 1) {
-	    prim.v[0] = get_vertex( draw, elts[i + 1] );
-	    prim.v[1] = get_vertex( draw, elts[i + 0] );
-	    prim.v[2] = get_vertex( draw, elts[i + 2] );
-	 }
-	 else {
-	    prim.v[0] = get_vertex( draw, elts[i + 0] );
-	    prim.v[1] = get_vertex( draw, elts[i + 1] );
-	    prim.v[2] = get_vertex( draw, elts[i + 2] );
-	 }
-	 
-	 first->tri( first, &prim );
-      }
-      break;
 
-   case GL_TRIANGLE_FAN:
-      if (count >= 3) {
-	 prim.v[0] = get_vertex( draw, elts[0] );
-	 prim.v[1] = 0;
-	 prim.v[2] = get_vertex( draw, elts[1] );
-	 
-	 for (i = 0; i+2 < count; i++) {
-	    prim.v[1] = prim.v[2];
-	    prim.v[2] = get_vertex( draw, elts[i+2] );
-      
-	    first->tri( first, &prim );
-	 }
-      }
-      break;
+static void draw_set_prim( struct draw_context *draw,
+			   GLenum prim )
+{
+   if (reduced_prim[prim] != draw->reduced_prim) {
+      draw_flush( draw );
+      draw->reduced_prim = reduced_prim[prim];
+   }
 
-   case GL_QUADS:
-      for (i = 0; i+3 < count; i += 4) {
-	 do_quad( first,
-		  get_vertex( draw, elts[i + 0] ),
-		  get_vertex( draw, elts[i + 1] ),
-		  get_vertex( draw, elts[i + 2] ),
-		  get_vertex( draw, elts[i + 3] ));
-      }
-      break;
+   draw->prim = prim;
+}
 
-   case GL_QUAD_STRIP:
-      for (i = 0; i+3 < count; i += 2) {
-	 do_quad( first,
-		  get_vertex( draw, elts[i + 2] ),
-		  get_vertex( draw, elts[i + 0] ),
-		  get_vertex( draw, elts[i + 1] ),
-		  get_vertex( draw, elts[i + 3] ));
-      }
-      break;
 
+static void do_point( struct draw_context *draw,
+		      GLuint i0 )
+{
+   struct prim_header *prim = get_queued_prim( draw, 1 );
+   
+   prim->reset_line_stipple = 0;
+   prim->edgeflags = 1;
+   prim->pad = 0;
+   prim->v[0] = draw->get_vertex( draw, i0 );
+}
 
-   case GL_POLYGON:
-      if (count >= 3) {
-         int e1save, e2save;
-	 prim.v[0] = 0;
-	 prim.v[1] = get_vertex( draw, elts[1] );
-	 prim.v[2] = get_vertex( draw, elts[0] );
-	 e2save = prim.v[2]->edgeflag;
-	 
-	 for (i = 0; i+2 < count; i++) {
-	    prim.v[0] = prim.v[1];
-	    prim.v[1] = get_vertex( draw, elts[i+2] );
-      
-            /* save v1 edge flag, and clear if not last triangle */
-            e1save = prim.v[1]->edgeflag;
-            if (i + 3 < count)
-               prim.v[1]->edgeflag = 0;
 
-            /* draw */
-	    first->tri( first, &prim );
+static void do_line( struct draw_context *draw,
+		     GLboolean reset_stipple,
+		     GLuint i0,
+		     GLuint i1 )
+{
+   struct prim_header *prim = get_queued_prim( draw, 2 );
+   
+   prim->reset_line_stipple = reset_stipple;
+   prim->edgeflags = 1;
+   prim->pad = 0;
+   prim->v[0] = draw->get_vertex( draw, i0 );
+   prim->v[1] = draw->get_vertex( draw, i1 );
+}
 
-            prim.v[1]->edgeflag = e1save; /* restore */
-            prim.v[2]->edgeflag = 0; /* disable edge after 1st tri */
-	 }
-         prim.v[2]->edgeflag = e2save;
-      }
-      break;
+static void do_triangle( struct draw_context *draw,
+			 GLuint i0,
+			 GLuint i1,
+			 GLuint i2 )
+{
+   struct prim_header *prim = get_queued_prim( draw, 3 );
+   
+   prim->reset_line_stipple = 1;
+   prim->edgeflags = ~0;
+   prim->pad = 0;
+   prim->v[0] = draw->get_vertex( draw, i0 );
+   prim->v[1] = draw->get_vertex( draw, i1 );
+   prim->v[2] = draw->get_vertex( draw, i2 );
+}
+			  
+static void do_ef_triangle( struct draw_context *draw,
+			    GLboolean reset_stipple,
+			    GLuint ef_mask,
+			    GLuint i0,
+			    GLuint i1,
+			    GLuint i2 )
+{
+   struct prim_header *prim = get_queued_prim( draw, 3 );
+   struct vertex_header *v0 = draw->get_vertex( draw, i0 );
+   struct vertex_header *v1 = draw->get_vertex( draw, i1 );
+   struct vertex_header *v2 = draw->get_vertex( draw, i2 );
+   
+   prim->reset_line_stipple = reset_stipple;
+
+   prim->edgeflags = ef_mask & ((v0->edgeflag << 0) | 
+				(v1->edgeflag << 1) | 
+				(v2->edgeflag << 2));
+   prim->pad = 0;
+   prim->v[0] = v0;
+   prim->v[1] = v1;
+   prim->v[2] = v2;
+}
 
-   default:
-      assert(0);
-      break;
-   }
+
+static void do_quad( struct draw_context *draw,
+		     unsigned v0,
+		     unsigned v1,
+		     unsigned v2,
+		     unsigned v3 )
+{
+   do_ef_triangle( draw, 1, ~(1<<0), v0, v1, v3 );
+   do_ef_triangle( draw, 0, ~(1<<1), v1, v2, v3 );
 }
 
+
 static void draw_prim( struct draw_context *draw,
 		       GLuint start,
 		       GLuint count )
 {
-   struct draw_stage * const first = draw->pipeline.first;
-   struct prim_header prim;
    GLuint i;
 
 //   _mesa_printf("%s (%d) %d/%d\n", __FUNCTION__, draw->prim, start, count );
 
-   prim.det = 0;		/* valid from cull stage onwards */
-   prim.v[0] = 0;
-   prim.v[1] = 0;
-   prim.v[2] = 0;
-
    switch (draw->prim) {
    case GL_POINTS:
       for (i = 0; i < count; i ++) {
-	 prim.v[0] = get_vertex( draw, start + i );
-	 first->point( first, &prim );
+	 do_point( draw,
+		   start + i );
       }
       break;
 
    case GL_LINES:
       for (i = 0; i+1 < count; i += 2) {
-	 prim.v[0] = get_vertex( draw, start + i + 0 );
-	 prim.v[1] = get_vertex( draw, start + i + 1 );
-
-         first->reset_stipple_counter( first );
-	 first->line( first, &prim );
+	 do_line( draw, 
+		  TRUE,
+		  start + i + 0,
+		  start + i + 1);
       }
       break;
 
    case GL_LINE_LOOP:  
       if (count >= 2) {
-         first->reset_stipple_counter( first );
 	 for (i = 1; i < count; i++) {
-	    prim.v[0] = get_vertex( draw, start + i - 1 );
-	    prim.v[1] = get_vertex( draw, start + i );	    
-	    first->line( first, &prim );
+	    do_line( draw, 
+		     i == 1, 	/* XXX: only if vb not split */
+		     start + i - 1,
+		     start + i );
 	 }
 
-	 prim.v[0] = get_vertex( draw, start + count - 1 );
-	 prim.v[1] = get_vertex( draw, start + 0 );	    
-	 first->line( first, &prim );
+	 do_line( draw, 
+		  0,
+		  start + count - 1,
+		  start + 0 );
       }
       break;
 
    case GL_LINE_STRIP:
       if (count >= 2) {
-         first->reset_stipple_counter( first );
-	 prim.v[0] = 0;
-	 prim.v[1] = get_vertex( draw, start + 0 );
-	 
 	 for (i = 1; i < count; i++) {
-	    prim.v[0] = prim.v[1];
-	    prim.v[1] = get_vertex( draw, start + i );
-	    
-	    first->line( first, &prim );
+	    do_line( draw,
+		     i == 1,
+		     start + i - 1,
+		     start + i );
 	 }
       }
       break;
 
    case GL_TRIANGLES:
       for (i = 0; i+2 < count; i += 3) {
-	 prim.v[0] = get_vertex( draw, start + i + 0 );
-	 prim.v[1] = get_vertex( draw, start + i + 1 );
-	 prim.v[2] = get_vertex( draw, start + i + 2 );
-      
-	 first->tri( first, &prim );
+	 do_ef_triangle( draw,
+			 1, 
+			 ~0,
+			 start + i + 0,
+			 start + i + 1,
+			 start + i + 2 );
       }
       break;
 
    case GL_TRIANGLE_STRIP:
       for (i = 0; i+2 < count; i++) {
 	 if (i & 1) {
-	    prim.v[0] = get_vertex( draw, start + i + 1 );
-	    prim.v[1] = get_vertex( draw, start + i + 0 );
-	    prim.v[2] = get_vertex( draw, start + i + 2 );
+	    do_triangle( draw,
+			 start + i + 1,
+			 start + i + 0,
+			 start + i + 2 );
 	 }
 	 else {
-	    prim.v[0] = get_vertex( draw, start + i + 0 );
-	    prim.v[1] = get_vertex( draw, start + i + 1 );
-	    prim.v[2] = get_vertex( draw, start + i + 2 );
+	    do_triangle( draw,
+			 start + i + 0,
+			 start + i + 1,
+			 start + i + 2 );
 	 }
-	 
-	 first->tri( first, &prim );
       }
       break;
 
    case GL_TRIANGLE_FAN:
       if (count >= 3) {
-	 prim.v[0] = get_vertex( draw, start + 0 );
-	 prim.v[1] = 0;
-	 prim.v[2] = get_vertex( draw, start + 1 );
-	 
 	 for (i = 0; i+2 < count; i++) {
-	    prim.v[1] = prim.v[2];
-	    prim.v[2] = get_vertex( draw, start + i + 2 );
-      
-	    first->tri( first, &prim );
+	    do_triangle( draw,
+			 start + 0,
+			 start + i + 1,
+			 start + i + 2 );
 	 }
       }
       break;
@@ -397,48 +418,42 @@ static void draw_prim( struct draw_context *draw,
 
    case GL_QUADS:
       for (i = 0; i+3 < count; i += 4) {
-	 do_quad( first,
-		  get_vertex( draw, start + i + 0 ),
-		  get_vertex( draw, start + i + 1 ),
-		  get_vertex( draw, start + i + 2 ),
-		  get_vertex( draw, start + i + 3 ));
+	 do_quad( draw,
+		  start + i + 0,
+		  start + i + 1,
+		  start + i + 2,
+		  start + i + 3);
       }
       break;
 
    case GL_QUAD_STRIP:
       for (i = 0; i+3 < count; i += 2) {
-	 do_quad( first,
-		  get_vertex( draw, start + i + 2 ),
-		  get_vertex( draw, start + i + 0 ),
-		  get_vertex( draw, start + i + 1 ),
-		  get_vertex( draw, start + i + 3 ));
+	 do_quad( draw,
+		  start + i + 2,
+		  start + i + 0,
+		  start + i + 1,
+		  start + i + 3);
       }
       break;
 
    case GL_POLYGON:
       if (count >= 3) {
-         int e1save, e2save;
-	 prim.v[0] = 0;
-	 prim.v[1] = get_vertex( draw, start + 1 );
-	 prim.v[2] = get_vertex( draw, start + 0 );
-	 e2save = prim.v[2]->edgeflag;
+	 unsigned ef_mask = (1<<2) | (1<<0);
 
 	 for (i = 0; i+2 < count; i++) {
-	    prim.v[0] = prim.v[1];
-	    prim.v[1] = get_vertex( draw, start + i + 2 );
 
-            /* save v1 edge flag, and clear if not last triangle */
-            e1save = prim.v[1]->edgeflag;
-            if (i + 3 < count)
-               prim.v[1]->edgeflag = 0;
+            if (i + 3 >= count)
+	       ef_mask |= (1<<1);
 
-            /* draw */
-	    first->tri( first, &prim );
+	    do_ef_triangle( draw,
+			    i == 0,
+			    ef_mask,
+			    start + i + 1,
+			    start + i + 2,
+			    start + i + 0);
 
-            prim.v[1]->edgeflag = e1save; /* restore */
-            prim.v[2]->edgeflag = 0; /* disable edge after 1st tri */
+	    ef_mask &= ~(1<<2);
 	 }
-         prim.v[2]->edgeflag = e2save;
       }
       break;
 
@@ -449,10 +464,18 @@ static void draw_prim( struct draw_context *draw,
 }
 
 
-static void draw_release_vertices( struct draw_context *draw )
+static void draw_allocate_vertices( struct draw_context *draw,
+				    GLuint nr_vertices )
 {
-   draw->pipeline.first->end( draw->pipeline.first );
+   draw->nr_vertices = nr_vertices;
+   draw->verts = (GLubyte *) malloc( nr_vertices * draw->vertex_size );
+   draw_invalidate_vcache( draw );
+}
+
+
 
+static void draw_release_vertices( struct draw_context *draw )
+{
    free(draw->verts);
    draw->verts = NULL;
 }
@@ -598,6 +621,8 @@ void draw_vb(struct draw_context *draw,
 
    draw->in_vb = 1;
 
+   draw->pipeline.first->begin( draw->pipeline.first );
+
    /* Allocate the vertices:
     */
    draw_allocate_vertices( draw, VB->Count );
@@ -605,11 +630,14 @@ void draw_vb(struct draw_context *draw,
    /* Bind the vb outputs:
     */
    vf_set_sources( draw->vf, VB->AttribPtr, 0 );
-
-   /* Build the hardware or prim-pipe vertices: 
-    */
    vf_emit_vertices( draw->vf, VB->Count, draw->verts );
 
+   draw->elts = VB->Elts;
+
+   if (VB->Elts) 
+      draw->get_vertex = get_uint_elt_vertex;
+   else
+      draw->get_vertex = get_vertex;
 
    for (i = 0; i < VB->PrimitiveCount; i++) {
 
@@ -628,21 +656,15 @@ void draw_vb(struct draw_context *draw,
       if (draw->prim != mode) 
 	 draw_set_prim( draw, mode );
 
-      if (VB->Elts) {
-	 draw_indexed_prim( draw, 
-			    VB->Elts + start,
-			    length );
-      }
-      else {
-	 draw_prim( draw, 
-		    start,
-		    length );
-      }	 
+      draw_prim( draw, start, length );
    }
 
+   draw_flush(draw);
+   draw->pipeline.first->end( draw->pipeline.first );
    draw_release_vertices( draw );
    draw->verts = NULL;
    draw->in_vb = 0;
+   draw->elts = NULL;
 }
 
 
@@ -661,11 +683,17 @@ draw_vertices(struct draw_context *draw,
 
    assert(mode <= GL_POLYGON);
 
+   draw->get_vertex = get_vertex;
    draw->vertex_size
       = sizeof(struct vertex_header) + numAttrs * 4 * sizeof(GLfloat);
 
+
+
+
    /*draw_prim_info(mode, &first, &incr);*/
    draw_allocate_vertices( draw, numVerts );
+   draw->pipeline.first->begin( draw->pipeline.first );
+
    if (draw->prim != mode) 
       draw_set_prim( draw, mode );
 
@@ -694,6 +722,9 @@ draw_vertices(struct draw_context *draw,
 
    /* draw */
    draw_prim(draw, 0, numVerts);
+   draw_flush(draw);
+   draw->pipeline.first->end( draw->pipeline.first );
+
 
    /* clean up */
    draw_release_vertices( draw );
@@ -759,8 +790,6 @@ void draw_set_vertex_attributes( struct draw_context *draw,
 }
 			    
 
-#define MAX_VERTEX_SIZE ((2 + FRAG_ATTRIB_MAX) * 4 * sizeof(GLfloat))
-
 void draw_alloc_tmps( struct draw_stage *stage, GLuint nr )
 {
    stage->nr_tmps = nr;