7 files changed, 390 insertions, 102 deletions

diff --git a/src/gallium/drivers/i915simple/i915_context.h b/src/gallium/drivers/i915simple/i915_context.h
index 53fc5ed079..2da90ae49d 100644
--- a/src/gallium/drivers/i915simple/i915_context.h
+++ b/src/gallium/drivers/i915simple/i915_context.h
@@ -245,6 +245,8 @@ struct i915_context
 
    /** Vertex buffer */
    struct pipe_buffer *vbo;
+   size_t vbo_offset;
+   unsigned vbo_flushed;
 
    struct i915_state current;
    unsigned hardware_dirty;
diff --git a/src/gallium/drivers/i915simple/i915_flush.c b/src/gallium/drivers/i915simple/i915_flush.c
index 7d23e6b6b9..4c4718d68e 100644
--- a/src/gallium/drivers/i915simple/i915_flush.c
+++ b/src/gallium/drivers/i915simple/i915_flush.c
@@ -68,6 +68,7 @@ static void i915_flush( struct pipe_context *pipe,
    /* If there are no flags, just flush pending commands to hardware:
     */
    FLUSH_BATCH(fence);
+   i915->vbo_flushed = 1;
 }
 
 
diff --git a/src/gallium/drivers/i915simple/i915_prim_vbuf.c b/src/gallium/drivers/i915simple/i915_prim_vbuf.c
index 7fb2adbb53..faf5971abb 100644
--- a/src/gallium/drivers/i915simple/i915_prim_vbuf.c
+++ b/src/gallium/drivers/i915simple/i915_prim_vbuf.c
@@ -62,8 +62,21 @@ struct i915_vbuf_render {
    /** Vertex size in bytes */
    unsigned vertex_size;
 
+   /** Software primitive */
+   unsigned prim;
+
    /** Hardware primitive */
    unsigned hwprim;
+
+   /** Genereate a vertex list */
+   unsigned fallback;
+
+   /* Stuff for the vbo */
+   struct pipe_buffer *vbo;
+   size_t vbo_size;
+   size_t vbo_offset;
+   void *vbo_ptr;
+   size_t vbo_alloc_size;
 };
 
 
@@ -95,8 +108,8 @@ i915_vbuf_render_get_vertex_info( struct vbuf_render *render )
 
 static void *
 i915_vbuf_render_allocate_vertices( struct vbuf_render *render,
-			            ushort vertex_size,
-				    ushort nr_vertices )
+                                    ushort vertex_size,
+                                    ushort nr_vertices )
 {
    struct i915_vbuf_render *i915_render = i915_vbuf_render(render);
    struct i915_context *i915 = i915_render->i915;
@@ -105,14 +118,31 @@ i915_vbuf_render_allocate_vertices( struct vbuf_render *render,
 
    /* FIXME: handle failure */
    assert(!i915->vbo);
-   i915->vbo = winsys->buffer_create(winsys, 64, I915_BUFFER_USAGE_LIT_VERTEX,
-                                     size);
-   
+
+   if (i915_render->vbo_size > size + i915_render->vbo_offset && !i915->vbo_flushed) {
+   } else {
+      i915->vbo_flushed = 0;
+      pipe_buffer_reference(winsys, &i915_render->vbo, NULL);
+   }
+
+   if (!i915_render->vbo) {
+      i915_render->vbo_size = MAX2(size, i915_render->vbo_alloc_size);
+      i915_render->vbo_offset = 0;
+      i915_render->vbo = winsys->buffer_create(winsys,
+					       64,
+					       I915_BUFFER_USAGE_LIT_VERTEX,
+					       i915_render->vbo_size);
+      i915_render->vbo_ptr = winsys->buffer_map(winsys,
+						i915_render->vbo,
+						PIPE_BUFFER_USAGE_CPU_WRITE);
+      winsys->buffer_unmap(winsys, i915_render->vbo);
+   }
+
+   i915->vbo = i915_render->vbo;
+   i915->vbo_offset = i915_render->vbo_offset;
    i915->dirty |= I915_NEW_VBO;
-   
-   return winsys->buffer_map(winsys, 
-                             i915->vbo, 
-                             PIPE_BUFFER_USAGE_CPU_WRITE);
+
+   return i915_render->vbo_ptr + i915->vbo_offset;
 }
 
 
@@ -121,16 +151,36 @@ i915_vbuf_render_set_primitive( struct vbuf_render *render,
                                 unsigned prim )
 {
    struct i915_vbuf_render *i915_render = i915_vbuf_render(render);
-   
+   i915_render->prim = prim;
+
    switch(prim) {
    case PIPE_PRIM_POINTS:
       i915_render->hwprim = PRIM3D_POINTLIST;
+      i915_render->fallback = 0;
       return TRUE;
    case PIPE_PRIM_LINES:
       i915_render->hwprim = PRIM3D_LINELIST;
+      i915_render->fallback = 0;
+      return TRUE;
+   case PIPE_PRIM_LINE_STRIP:
+      i915_render->hwprim = PRIM3D_LINESTRIP;
+      i915_render->fallback = 0;
       return TRUE;
    case PIPE_PRIM_TRIANGLES:
       i915_render->hwprim = PRIM3D_TRILIST;
+      i915_render->fallback = 0;
+      return TRUE;
+   case PIPE_PRIM_TRIANGLE_STRIP:
+      i915_render->hwprim = PRIM3D_TRISTRIP;
+      i915_render->fallback = 0;
+      return TRUE;
+   case PIPE_PRIM_QUADS:
+      i915_render->hwprim = PRIM3D_TRILIST;
+      i915_render->fallback = PIPE_PRIM_QUADS;
+      return TRUE;
+   case PIPE_PRIM_QUAD_STRIP:
+      i915_render->hwprim = PRIM3D_TRILIST;
+      i915_render->fallback = PIPE_PRIM_QUAD_STRIP;
       return TRUE;
    default:
       /* Actually, can handle a lot more just fine...  Fixme.
@@ -140,6 +190,180 @@ i915_vbuf_render_set_primitive( struct vbuf_render *render,
 }
 
 
+
+/**
+ * Used for fallbacks in draw_arrays
+ */
+static void
+draw_arrays_generate_indices( struct vbuf_render *render,
+                              unsigned start, uint nr,
+                              unsigned type )
+{
+   struct i915_vbuf_render *i915_render = i915_vbuf_render(render);
+   struct i915_context *i915 = i915_render->i915;
+   unsigned i;
+   unsigned end = start + nr;
+   switch(type) {
+   case 0:
+      for (i = start; i+1 < end; i += 2)
+	 OUT_BATCH( (i+0) | (i+1) << 16 );
+      if (i < end)
+	 OUT_BATCH( i );
+      break;
+   case PIPE_PRIM_QUADS:
+      for (i = start; i + 3 < end; i += 4) {
+	 OUT_BATCH( (i+0) | (i+1) << 16 );
+	 OUT_BATCH( (i+3) | (i+1) << 16 );
+	 OUT_BATCH( (i+2) | (i+3) << 16 );
+      }
+      break;
+   case PIPE_PRIM_QUAD_STRIP:
+      for (i = start; i + 3 < end; i += 2) {
+	 OUT_BATCH( (i+0) | (i+1) << 16 );
+	 OUT_BATCH( (i+3) | (i+2) << 16 );
+	 OUT_BATCH( (i+0) | (i+3) << 16 );
+      }
+      break;
+   default:
+      assert(0);
+   }
+}
+
+static unsigned
+draw_arrays_calc_nr_indices( uint nr, unsigned type )
+{
+   switch (type) {
+   case 0:
+      return nr;
+   case PIPE_PRIM_QUADS:
+      return (nr / 4) * 6;
+   case PIPE_PRIM_QUAD_STRIP:
+      return ((nr - 2) / 2) * 6;
+   default:
+      assert(0);
+      return 0;
+   }
+}
+
+static void
+draw_arrays_fallback( struct vbuf_render *render,
+                      unsigned start,
+                      uint nr )
+{
+   struct i915_vbuf_render *i915_render = i915_vbuf_render(render);
+   struct i915_context *i915 = i915_render->i915;
+   struct pipe_winsys *winsys = i915->pipe.winsys;
+   unsigned nr_indices;
+
+   if (i915->dirty)
+      i915_update_derived( i915 );
+
+   if (i915->hardware_dirty)
+      i915_emit_hardware_state( i915 );
+
+   nr_indices = draw_arrays_calc_nr_indices( nr, i915_render->fallback );
+
+   if (!BEGIN_BATCH( 1 + (nr_indices + 1)/2, 1 )) {
+      FLUSH_BATCH(NULL);
+
+      /* Make sure state is re-emitted after a flush:
+       */
+      i915_update_derived( i915 );
+      i915_emit_hardware_state( i915 );
+      i915->vbo_flushed = 1;
+
+      if (!BEGIN_BATCH( 1 + (nr_indices + 1)/2, 1 )) {
+	 assert(0);
+	 goto out;
+      }
+   }
+   OUT_BATCH( _3DPRIMITIVE |
+	      PRIM_INDIRECT |
+	      i915_render->hwprim |
+	      PRIM_INDIRECT_ELTS |
+	      nr_indices );
+
+   draw_arrays_generate_indices( render, start, nr, i915_render->fallback );
+
+out:
+   return;
+}
+
+static void
+i915_vbuf_render_draw_arrays( struct vbuf_render *render,
+                              unsigned start,
+                              uint nr )
+{
+   struct i915_vbuf_render *i915_render = i915_vbuf_render(render);
+
+   if (i915_render->fallback) {
+      draw_arrays_fallback( render, start, nr );
+      return;
+   }
+
+   /* JB: TODO submit direct cmds */
+   draw_arrays_fallback( render, start, nr );
+}
+
+/**
+ * Used for normal and fallback emitting of indices
+ * If type is zero normal operation assumed.
+ */
+static void
+draw_generate_indices( struct vbuf_render *render,
+                       const ushort *indices,
+                       uint nr_indices,
+                       unsigned type )
+{
+   struct i915_vbuf_render *i915_render = i915_vbuf_render(render);
+   struct i915_context *i915 = i915_render->i915;
+   unsigned i;
+
+   switch(type) {
+   case 0:
+      for (i = 0; i + 1 < nr_indices; i += 2) {
+	 OUT_BATCH( indices[i] | indices[i+1] << 16 );
+      }
+      if (i < nr_indices) {
+	 OUT_BATCH( indices[i] );
+      }
+      break;
+   case PIPE_PRIM_QUADS:
+      for (i = 0; i + 3 < nr_indices; i += 4) {
+	 OUT_BATCH( indices[i+0] | indices[i+1] << 16 );
+	 OUT_BATCH( indices[i+3] | indices[i+1] << 16 );
+	 OUT_BATCH( indices[i+2] | indices[i+3] << 16 );
+      }
+      break;
+   case PIPE_PRIM_QUAD_STRIP:
+      for (i = 0; i + 3 < nr_indices; i += 2) {
+	 OUT_BATCH( indices[i+0] | indices[i+1] << 16 );
+	 OUT_BATCH( indices[i+3] | indices[i+2] << 16 );
+	 OUT_BATCH( indices[i+0] | indices[i+3] << 16 );
+      }
+      break;
+   default:
+      assert(0);
+      break;
+   }
+}
+
+static unsigned
+draw_calc_nr_indices( uint nr_indices, unsigned type )
+{
+   switch (type) {
+   case 0:
+      return nr_indices;
+   case PIPE_PRIM_QUADS:
+      return (nr_indices / 4) * 6;
+   case PIPE_PRIM_QUAD_STRIP:
+      return ((nr_indices - 2) / 2) * 6;
+   default:
+      assert(0);
+      return 0;
+   }
+}
+
 static void 
 i915_vbuf_render_draw( struct vbuf_render *render,
                        const ushort *indices,
@@ -147,13 +371,14 @@ i915_vbuf_render_draw( struct vbuf_render *render,
 {
    struct i915_vbuf_render *i915_render = i915_vbuf_render(render);
    struct i915_context *i915 = i915_render->i915;
-   unsigned i;
+   struct pipe_winsys *winsys = i915->pipe.winsys;
+   unsigned save_nr_indices;
+
+   save_nr_indices = nr_indices;
+   nr_indices = draw_calc_nr_indices( nr_indices, i915_render->fallback );
 
    assert(nr_indices);
 
-   /* this seems to be bogus, since we validate state right after this */
-   /*assert((i915->dirty & ~I915_NEW_VBO) == 0);*/
-   
    if (i915->dirty)
       i915_update_derived( i915 );
 
@@ -167,25 +392,26 @@ i915_vbuf_render_draw( struct vbuf_render *render,
        */
       i915_update_derived( i915 );
       i915_emit_hardware_state( i915 );
+      i915->vbo_flushed = 1;
 
       if (!BEGIN_BATCH( 1 + (nr_indices + 1)/2, 1 )) {
 	 assert(0);
-	 return;
+     goto out;
       }
    }
 
    OUT_BATCH( _3DPRIMITIVE |
-              PRIM_INDIRECT |
-              i915_render->hwprim |
-   	      PRIM_INDIRECT_ELTS |
-   	      nr_indices );
-   for (i = 0; i + 1 < nr_indices; i += 2) {
-      OUT_BATCH( indices[i] |
-                 (indices[i + 1] << 16) );
-   }
-   if (i < nr_indices) {
-      OUT_BATCH( indices[i] );
-   }
+	      PRIM_INDIRECT |
+	      i915_render->hwprim |
+	      PRIM_INDIRECT_ELTS |
+	      nr_indices );
+   draw_generate_indices( render,
+			  indices,
+			  save_nr_indices,
+			  i915_render->fallback );
+
+out:
+   return;
 }
 
 
@@ -198,10 +424,13 @@ i915_vbuf_render_release_vertices( struct vbuf_render *render,
    struct i915_vbuf_render *i915_render = i915_vbuf_render(render);
    struct i915_context *i915 = i915_render->i915;
    struct pipe_winsys *winsys = i915->pipe.winsys;
+   size_t size = (size_t)vertex_size * (size_t)vertices_used;
 
    assert(i915->vbo);
-   winsys->buffer_unmap(winsys, i915->vbo);
-   pipe_buffer_reference(winsys, &i915->vbo, NULL);
+
+   i915_render->vbo_offset += size;
+   i915->vbo = NULL;
+   i915->dirty |= I915_NEW_VBO;
 }
 
 
@@ -220,6 +449,7 @@ static struct vbuf_render *
 i915_vbuf_render_create( struct i915_context *i915 )
 {
    struct i915_vbuf_render *i915_render = CALLOC_STRUCT(i915_vbuf_render);
+   struct pipe_winsys *winsys = i915->pipe.winsys;
 
    i915_render->i915 = i915;
    
@@ -229,14 +459,27 @@ i915_vbuf_render_create( struct i915_context *i915 )
     * batch buffer.
     */
    i915_render->base.max_indices = 16*1024;
-   
+
    i915_render->base.get_vertex_info = i915_vbuf_render_get_vertex_info;
    i915_render->base.allocate_vertices = i915_vbuf_render_allocate_vertices;
    i915_render->base.set_primitive = i915_vbuf_render_set_primitive;
    i915_render->base.draw = i915_vbuf_render_draw;
+   i915_render->base.draw_arrays = i915_vbuf_render_draw_arrays;
    i915_render->base.release_vertices = i915_vbuf_render_release_vertices;
    i915_render->base.destroy = i915_vbuf_render_destroy;
-   
+
+   i915_render->vbo_alloc_size = 128 * 4096;
+   i915_render->vbo_size = i915_render->vbo_alloc_size;
+   i915_render->vbo_offset = 0;
+   i915_render->vbo = winsys->buffer_create(winsys,
+					    64,
+					    I915_BUFFER_USAGE_LIT_VERTEX,
+					    i915_render->vbo_size);
+   i915_render->vbo_ptr = winsys->buffer_map(winsys,
+					     i915_render->vbo,
+					     PIPE_BUFFER_USAGE_CPU_WRITE);
+   winsys->buffer_unmap(winsys, i915_render->vbo);
+
    return &i915_render->base;
 }
 
@@ -258,6 +501,7 @@ struct draw_stage *i915_draw_vbuf_stage( struct i915_context *i915 )
       render->destroy(render);
       return NULL;
    }
-    
+   draw_set_render(i915->draw, render);
+
    return stage;
 }
diff --git a/src/gallium/drivers/i915simple/i915_state.c b/src/gallium/drivers/i915simple/i915_state.c
index e09c225900..39d5d5e9d6 100644
--- a/src/gallium/drivers/i915simple/i915_state.c
+++ b/src/gallium/drivers/i915simple/i915_state.c
@@ -41,7 +41,6 @@
 #include "i915_state_inlines.h"
 #include "i915_fpc.h"
 
-
 /* The i915 (and related graphics cores) do not support GL_CLAMP.  The
  * Intel drivers for "other operating systems" implement GL_CLAMP as
  * GL_CLAMP_TO_EDGE, so the same is done here.
@@ -178,6 +177,7 @@ static void i915_bind_blend_state(struct pipe_context *pipe,
                                   void *blend)
 {
    struct i915_context *i915 = i915_context(pipe);
+   draw_flush(i915->draw);
 
    i915->blend = (struct i915_blend_state*)blend;
 
@@ -194,6 +194,7 @@ static void i915_set_blend_color( struct pipe_context *pipe,
 			     const struct pipe_blend_color *blend_color )
 {
    struct i915_context *i915 = i915_context(pipe);
+   draw_flush(i915->draw);
 
    i915->blend_color = *blend_color;
 
@@ -280,6 +281,8 @@ static void i915_bind_sampler_states(struct pipe_context *pipe,
        !memcmp(i915->sampler, sampler, num * sizeof(void *)))
       return;
 
+   draw_flush(i915->draw);
+
    for (i = 0; i < num; ++i)
       i915->sampler[i] = sampler[i];
    for (i = num; i < PIPE_MAX_SAMPLERS; ++i)
@@ -398,6 +401,7 @@ static void i915_bind_depth_stencil_state(struct pipe_context *pipe,
                                           void *depth_stencil)
 {
    struct i915_context *i915 = i915_context(pipe);
+   draw_flush(i915->draw);
 
    i915->depth_stencil = (const struct i915_depth_stencil_state *)depth_stencil;
 
@@ -415,6 +419,7 @@ static void i915_set_scissor_state( struct pipe_context *pipe,
                                  const struct pipe_scissor_state *scissor )
 {
    struct i915_context *i915 = i915_context(pipe);
+   draw_flush(i915->draw);
 
    memcpy( &i915->scissor, scissor, sizeof(*scissor) );
    i915->dirty |= I915_NEW_SCISSOR;
@@ -451,6 +456,7 @@ static void
 i915_bind_fs_state(struct pipe_context *pipe, void *shader)
 {
    struct i915_context *i915 = i915_context(pipe);
+   draw_flush(i915->draw);
 
    i915->fs = (struct i915_fragment_shader*) shader;
 
@@ -506,6 +512,7 @@ static void i915_set_constant_buffer(struct pipe_context *pipe,
 {
    struct i915_context *i915 = i915_context(pipe);
    struct pipe_winsys *ws = pipe->winsys;
+   draw_flush(i915->draw);
 
    assert(shader < PIPE_SHADER_TYPES);
    assert(index == 0);
@@ -574,6 +581,7 @@ static void i915_set_framebuffer_state(struct pipe_context *pipe,
 				       const struct pipe_framebuffer_state *fb)
 {
    struct i915_context *i915 = i915_context(pipe);
+   draw_flush(i915->draw);
 
    i915->framebuffer = *fb; /* struct copy */
 
@@ -586,6 +594,7 @@ static void i915_set_clip_state( struct pipe_context *pipe,
 			     const struct pipe_clip_state *clip )
 {
    struct i915_context *i915 = i915_context(pipe);
+   draw_flush(i915->draw);
 
    draw_set_clip_state(i915->draw, clip);
 
@@ -698,6 +707,10 @@ static void i915_set_vertex_buffers(struct pipe_context *pipe,
                                     const struct pipe_vertex_buffer *buffers)
 {
    struct i915_context *i915 = i915_context(pipe);
+   /* Because we change state before the draw_set_vertex_buffers call
+    * we need a flush here, just to be sure.
+    */
+   draw_flush(i915->draw);
 
    memcpy(i915->vertex_buffer, buffers, count * sizeof(buffers[0]));
    i915->num_vertex_buffers = count;
@@ -711,16 +724,27 @@ static void i915_set_vertex_elements(struct pipe_context *pipe,
                                      const struct pipe_vertex_element *elements)
 {
    struct i915_context *i915 = i915_context(pipe);
+   /* Because we change state before the draw_set_vertex_buffers call
+    * we need a flush here, just to be sure.
+    */
+   draw_flush(i915->draw);
+
    i915->num_vertex_elements = count;
    /* pass-through to draw module */
    draw_set_vertex_elements(i915->draw, count, elements);
 }
 
 
+static void i915_set_edgeflags(struct pipe_context *pipe,
+                               const unsigned *bitfield)
+{
+   /* TODO do something here */
+}
 
 void
 i915_init_state_functions( struct i915_context *i915 )
 {
+   i915->pipe.set_edgeflags = i915_set_edgeflags;
    i915->pipe.create_blend_state = i915_create_blend_state;
    i915->pipe.bind_blend_state = i915_bind_blend_state;
    i915->pipe.delete_blend_state = i915_delete_blend_state;
diff --git a/src/gallium/drivers/i915simple/i915_state_emit.c b/src/gallium/drivers/i915simple/i915_state_emit.c
index 6f947d4346..8bcc26ad35 100644
--- a/src/gallium/drivers/i915simple/i915_state_emit.c
+++ b/src/gallium/drivers/i915simple/i915_state_emit.c
@@ -217,6 +217,7 @@ i915_emit_hardware_state(struct i915_context *i915 )
 
 	 OUT_BATCH(BUF_3D_ID_COLOR_BACK | 
 		   BUF_3D_PITCH(pitch) |  /* pitch in bytes */
+//		   BUF_3D_TILED_SURFACE); /* JB: Used to force tileing */
 		   BUF_3D_USE_FENCE);
 
 	 OUT_RELOC(cbuf_surface->buffer,
diff --git a/src/gallium/drivers/i915simple/i915_state_immediate.c b/src/gallium/drivers/i915simple/i915_state_immediate.c
index dfbbcab624..704ea4d838 100644
--- a/src/gallium/drivers/i915simple/i915_state_immediate.c
+++ b/src/gallium/drivers/i915simple/i915_state_immediate.c
@@ -54,7 +54,7 @@ static void upload_S0S1(struct i915_context *i915)
 
    /* INTEL_NEW_VBO */
    /* TODO: re-use vertex buffers here? */
-   LIS0 = 0;
+   LIS0 = i915->vbo_offset;
 
    /* INTEL_NEW_VERTEX_SIZE -- do this where the vertex size is calculated! 
     */
diff --git a/src/gallium/drivers/i915simple/i915_texture.c b/src/gallium/drivers/i915simple/i915_texture.c
index 3e23e540f9..4c0c7ce3f3 100644
--- a/src/gallium/drivers/i915simple/i915_texture.c
+++ b/src/gallium/drivers/i915simple/i915_texture.c
@@ -104,52 +104,50 @@ i915_miptree_set_image_offset(struct i915_texture *tex,
    */
 }
 
+static unsigned
+power_of_two(unsigned x)
+{
+   int value = 1;
+   while (value <= x)
+      value = value << 1;
+   return value;
+}
+
+static unsigned
+round_up(unsigned n, unsigned multiple)
+{
+   return (n + multiple - 1) & ~(multiple - 1);
+}
 
-/* Hack it up to use the old winsys->surface_alloc_storage()
- * method for now:
+/**
+ * Special case to deal with display targets.
  */
 static boolean
 i915_displaytarget_layout(struct pipe_screen *screen,
                           struct i915_texture *tex)
 {
-   struct pipe_winsys *ws = screen->winsys;
-   struct pipe_surface surf;
-   unsigned flags = (PIPE_BUFFER_USAGE_CPU_READ |
-                     PIPE_BUFFER_USAGE_CPU_WRITE |
-                     PIPE_BUFFER_USAGE_GPU_READ |
-                     PIPE_BUFFER_USAGE_GPU_WRITE);
-
-
-   memset(&surf, 0, sizeof(surf));
-
-   ws->surface_alloc_storage( ws, 
-                              &surf,
-                              tex->base.width[0], 
-                              tex->base.height[0],
-                              tex->base.format,
-                              flags,
-                              tex->base.tex_usage);
-      
-   /* Now extract the goodies: 
-    */
-   i915_miptree_set_image_offset( tex, 0, 0, 0, 0 );
-   i915_miptree_set_level_info( tex, 0, 0, 0, 0, 
+   struct pipe_texture *pt = &tex->base;
+
+   if (pt->last_level > 0 || pt->cpp != 4)
+      return 0;
+
+   i915_miptree_set_level_info( tex, 0, 1, 0, 0,
                                 tex->base.width[0],
                                 tex->base.height[0],
                                 1 );
+   i915_miptree_set_image_offset( tex, 0, 0, 0, 0 );
 
-   tex->buffer = surf.buffer;
-   tex->pitch = surf.pitch;
-   tex->total_height = 0;
-
+#if 0
+   tex->pitch = MAX2(512, power_of_two(tex->base.width[0] * pt->cpp)) / pt->cpp;
+   tex->total_height = round_up(tex->base.height[0], 8);
+#else
+   tex->pitch = round_up(tex->base.width[0], 64 / pt->cpp);
+   tex->total_height = tex->base.height[0];
+#endif
 
-   return tex->buffer != NULL;
+   return 1;
 }
 
-
-
-
-
 static void
 i945_miptree_layout_2d( struct i915_texture *tex )
 {
@@ -179,7 +177,7 @@ i945_miptree_layout_2d( struct i915_texture *tex )
    /* Pitch must be a whole number of dwords, even though we
     * express it in texels.
     */
-   tex->pitch = align_int(tex->pitch * pt->cpp, 4) / pt->cpp;
+   tex->pitch = align_int(tex->pitch * pt->cpp, 64) / pt->cpp;
    tex->total_height = 0;
 
    for (level = 0; level <= pt->last_level; level++) {
@@ -525,54 +523,47 @@ i945_miptree_layout(struct i915_texture * tex)
 
 
 static struct pipe_texture *
-i915_texture_create_screen(struct pipe_screen *screen,
-                           const struct pipe_texture *templat)
+i915_texture_create(struct pipe_screen *screen,
+                    const struct pipe_texture *templat)
 {
    struct i915_screen *i915screen = i915_screen(screen);
    struct pipe_winsys *ws = screen->winsys;
    struct i915_texture *tex = CALLOC_STRUCT(i915_texture);
 
-   if (!tex) 
+   if (!tex)
       return NULL;
 
    tex->base = *templat;
    tex->base.refcount = 1;
    tex->base.screen = screen;
 
-   if (tex->base.tex_usage & PIPE_TEXTURE_USAGE_DISPLAY_TARGET) {
-      if (!i915_displaytarget_layout(screen, tex))
-         goto fail;
-   }
-   else {
-      if (i915screen->is_i945) {
-         if (!i945_miptree_layout(tex))
-            goto fail;
-      }
-      else {
-         if (!i915_miptree_layout(tex))
-            goto fail;
-      }
-      
-      tex->buffer = ws->buffer_create(ws, 64,
-                                      PIPE_BUFFER_USAGE_PIXEL,
-                                      tex->pitch * tex->base.cpp *
-                                      tex->total_height);
-
-      if (!tex->buffer) 
-         goto fail;
+   if (i915screen->is_i945) {
+      if (!i945_miptree_layout(tex))
+	 goto fail;
+   } else {
+      if (!i915_miptree_layout(tex))
+	 goto fail;
    }
 
+   tex->buffer = ws->buffer_create(ws, 64,
+				   PIPE_BUFFER_USAGE_PIXEL,
+				   tex->pitch * tex->base.cpp *
+				   tex->total_height);
+
+   if (!tex->buffer)
+      goto fail;
+
    return &tex->base;
 
- fail:
+fail:
    FREE(tex);
    return NULL;
 }
 
 
 static void
-i915_texture_release_screen(struct pipe_screen *screen,
-                            struct pipe_texture **pt)
+i915_texture_release(struct pipe_screen *screen,
+                     struct pipe_texture **pt)
 {
    if (!*pt)
       return;
@@ -606,10 +597,10 @@ i915_texture_release_screen(struct pipe_screen *screen,
  * XXX note: same as code in sp_surface.c
  */
 static struct pipe_surface *
-i915_get_tex_surface_screen(struct pipe_screen *screen,
-                            struct pipe_texture *pt,
-                            unsigned face, unsigned level, unsigned zslice,
-                            unsigned flags)
+i915_get_tex_surface(struct pipe_screen *screen,
+                     struct pipe_texture *pt,
+                     unsigned face, unsigned level, unsigned zslice,
+                     unsigned flags)
 {
    struct i915_texture *tex = (struct i915_texture *)pt;
    struct pipe_winsys *ws = screen->winsys;
@@ -629,10 +620,11 @@ i915_get_tex_surface_screen(struct pipe_screen *screen,
       assert(zslice == 0);
    }
 
-   ps = ws->surface_alloc(ws);
+   ps = CALLOC_STRUCT(pipe_surface);//ws->surface_alloc(ws);
    if (ps) {
       assert(ps->refcount);
       assert(ps->winsys);
+      pipe_texture_reference(&ps->texture, pt);
       pipe_buffer_reference(ws, &ps->buffer, tex->buffer);
       ps->format = pt->format;
       ps->cpp = pt->cpp;
@@ -652,11 +644,35 @@ i915_init_texture_functions(struct i915_context *i915)
 //   i915->pipe.texture_update = i915_texture_update;
 }
 
+static void
+i915_tex_surface_release(struct pipe_screen *screen,
+                         struct pipe_surface **surface)
+{
+   struct pipe_surface *surf = *surface;
+
+   if (--surf->refcount == 0) {
+
+      /* This really should not be possible, but it's actually
+       * happening quite a bit...  Will fix.
+       */
+      if (surf->status == PIPE_SURFACE_STATUS_CLEAR) {
+         debug_printf("XXX destroying a surface with pending clears...\n");
+         assert(0);
+      }
+
+      pipe_texture_reference(&surf->texture, NULL);
+      pipe_buffer_reference(screen->winsys, &surf->buffer, NULL);
+      FREE(surf);
+   }
+
+   *surface = NULL;
+}
 
 void
 i915_init_screen_texture_functions(struct pipe_screen *screen)
 {
-   screen->texture_create = i915_texture_create_screen;
-   screen->texture_release = i915_texture_release_screen;
-   screen->get_tex_surface = i915_get_tex_surface_screen;
+   screen->texture_create = i915_texture_create;
+   screen->texture_release = i915_texture_release;
+   screen->get_tex_surface = i915_get_tex_surface;
+   screen->tex_surface_release = i915_tex_surface_release;
 }