Merge branch 'upstream-gallium-0.1' into nouveau-gallium-0.1

Conflicts:

	src/gallium/drivers/Makefile

17 files changed, 300 insertions, 138 deletions

diff --git a/src/gallium/drivers/softpipe/Makefile b/src/gallium/drivers/softpipe/Makefile
index 5479daf8ea..539ffb77f5 100644
--- a/src/gallium/drivers/softpipe/Makefile
+++ b/src/gallium/drivers/softpipe/Makefile
@@ -1,10 +1,9 @@
-
 TOP = ../../../..
 include $(TOP)/configs/current
 
 LIBNAME = softpipe
 
-DRIVER_SOURCES = \
+C_SOURCES = \
 	sp_fs_exec.c \
 	sp_fs_sse.c \
 	sp_fs_llvm.c \
@@ -41,12 +40,6 @@ DRIVER_SOURCES = \
 	sp_tile_cache.c \
 	sp_surface.c 
 
-C_SOURCES = \
-	$(COMMON_SOURCES) \
-	$(DRIVER_SOURCES)
-
-ASM_SOURCES = 
-
 include ../../Makefile.template
 
 symlinks:
diff --git a/src/gallium/drivers/softpipe/sp_context.c b/src/gallium/drivers/softpipe/sp_context.c
index 5e98f190bb..2cdf3c75bf 100644
--- a/src/gallium/drivers/softpipe/sp_context.c
+++ b/src/gallium/drivers/softpipe/sp_context.c
@@ -283,6 +283,7 @@ struct pipe_context *softpipe_create( struct pipe_winsys *pipe_winsys,
    /* textures */
    softpipe->pipe.texture_create = softpipe_texture_create;
    softpipe->pipe.texture_release = softpipe_texture_release;
+   softpipe->pipe.texture_update = softpipe_texture_update;
    softpipe->pipe.get_tex_surface = softpipe_get_tex_surface;
 
    /*
@@ -327,6 +328,18 @@ struct pipe_context *softpipe_create( struct pipe_winsys *pipe_winsys,
       draw_set_rasterize_stage(softpipe->draw, softpipe->setup);
    }
 
+   /* plug in AA line/point stages */
+   draw_install_aaline_stage(softpipe->draw, &softpipe->pipe);
+   draw_install_aapoint_stage(softpipe->draw, &softpipe->pipe);
+
+#if USE_DRAW_STAGE_PSTIPPLE
+   /* Do polygon stipple w/ texture map + frag prog? */
+   draw_install_pstipple_stage(softpipe->draw, &softpipe->pipe);
+#endif
+
+   /* sp_prim_setup can do wide points (don't convert to quads) */
+   draw_convert_wide_points(softpipe->draw, FALSE);
+
    sp_init_surface_functions(softpipe);
 
    return &softpipe->pipe;
diff --git a/src/gallium/drivers/softpipe/sp_context.h b/src/gallium/drivers/softpipe/sp_context.h
index b70d4fea85..feeafc7084 100644
--- a/src/gallium/drivers/softpipe/sp_context.h
+++ b/src/gallium/drivers/softpipe/sp_context.h
@@ -39,6 +39,13 @@
 #include "sp_quad.h"
 
 
+/**
+ * This is a temporary variable for testing draw-stage polygon stipple.
+ * If zero, do stipple in sp_quad_stipple.c
+ */
+#define USE_DRAW_STAGE_PSTIPPLE 1
+
+
 struct softpipe_winsys;
 struct softpipe_vbuf_render;
 struct draw_context;
@@ -68,7 +75,7 @@ struct softpipe_context {
    struct pipe_framebuffer_state framebuffer;
    struct pipe_poly_stipple poly_stipple;
    struct pipe_scissor_state scissor;
-   struct softpipe_texture *texture[PIPE_MAX_SAMPLERS];
+   struct pipe_texture *texture[PIPE_MAX_SAMPLERS];
    struct pipe_viewport_state viewport;
    struct pipe_vertex_buffer vertex_buffer[PIPE_ATTRIB_MAX];
    struct pipe_vertex_element vertex_element[PIPE_ATTRIB_MAX];
diff --git a/src/gallium/drivers/softpipe/sp_fs_exec.c b/src/gallium/drivers/softpipe/sp_fs_exec.c
index 9ad30a7681..8cb0534342 100644
--- a/src/gallium/drivers/softpipe/sp_fs_exec.c
+++ b/src/gallium/drivers/softpipe/sp_fs_exec.c
@@ -44,6 +44,41 @@ struct sp_exec_fragment_shader {
 
 
 
+/**
+ * Compute quad X,Y,Z,W for the four fragments in a quad.
+ *
+ * This should really be part of the compiled shader.
+ */
+void
+sp_setup_pos_vector(const struct tgsi_interp_coef *coef,
+		    float x, float y,
+		    struct tgsi_exec_vector *quadpos)
+{
+   uint chan;
+   /* do X */
+   quadpos->xyzw[0].f[0] = x;
+   quadpos->xyzw[0].f[1] = x + 1;
+   quadpos->xyzw[0].f[2] = x;
+   quadpos->xyzw[0].f[3] = x + 1;
+
+   /* do Y */
+   quadpos->xyzw[1].f[0] = y;
+   quadpos->xyzw[1].f[1] = y;
+   quadpos->xyzw[1].f[2] = y + 1;
+   quadpos->xyzw[1].f[3] = y + 1;
+
+   /* do Z and W for all fragments in the quad */
+   for (chan = 2; chan < 4; chan++) {
+      const float dadx = coef->dadx[chan];
+      const float dady = coef->dady[chan];
+      const float a0 = coef->a0[chan] + dadx * x + dady * y;
+      quadpos->xyzw[chan].f[0] = a0;
+      quadpos->xyzw[chan].f[1] = a0 + dadx;
+      quadpos->xyzw[chan].f[2] = a0 + dady;
+      quadpos->xyzw[chan].f[3] = a0 + dadx + dady;
+   }
+}
+
 
 static void
 exec_prepare( struct sp_fragment_shader *base,
diff --git a/src/gallium/drivers/softpipe/sp_fs_sse.c b/src/gallium/drivers/softpipe/sp_fs_sse.c
index d90066e025..8095d662ee 100644
--- a/src/gallium/drivers/softpipe/sp_fs_sse.c
+++ b/src/gallium/drivers/softpipe/sp_fs_sse.c
@@ -42,7 +42,7 @@
 
 #if defined(__i386__) || defined(__386__)
 
-#include "x86/rtasm/x86sse.h"
+#include "rtasm/rtasm_x86sse.h"
 
 /* Surely this should be defined somewhere in a tgsi header:
  */
@@ -63,41 +63,6 @@ struct sp_sse_fragment_shader {
 };
 
 
-/**
- * Compute quad X,Y,Z,W for the four fragments in a quad.
- *
- * This should really be part of the compiled shader.
- */
-void
-sp_setup_pos_vector(const struct tgsi_interp_coef *coef,
-		    float x, float y,
-		    struct tgsi_exec_vector *quadpos)
-{
-   uint chan;
-   /* do X */
-   quadpos->xyzw[0].f[0] = x;
-   quadpos->xyzw[0].f[1] = x + 1;
-   quadpos->xyzw[0].f[2] = x;
-   quadpos->xyzw[0].f[3] = x + 1;
-
-   /* do Y */
-   quadpos->xyzw[1].f[0] = y;
-   quadpos->xyzw[1].f[1] = y;
-   quadpos->xyzw[1].f[2] = y + 1;
-   quadpos->xyzw[1].f[3] = y + 1;
-
-   /* do Z and W for all fragments in the quad */
-   for (chan = 2; chan < 4; chan++) {
-      const float dadx = coef->dadx[chan];
-      const float dady = coef->dady[chan];
-      const float a0 = coef->a0[chan] + dadx * x + dady * y;
-      quadpos->xyzw[chan].f[0] = a0;
-      quadpos->xyzw[chan].f[1] = a0 + dadx;
-      quadpos->xyzw[chan].f[2] = a0 + dady;
-      quadpos->xyzw[chan].f[3] = a0 + dadx + dady;
-   }
-}
-
 
 static void
 fs_sse_prepare( struct sp_fragment_shader *base,
@@ -124,6 +89,9 @@ fs_sse_run( struct sp_fragment_shader *base,
 		       (float)quad->x0, (float)quad->y0, 
 		       machine->Temps);
 
+   /* init kill mask */
+   machine->Temps[TGSI_EXEC_TEMP_KILMASK_I].xyzw[TGSI_EXEC_TEMP_KILMASK_C].u[0] = 0x0;
+
    shader->func( machine->Inputs,
 		 machine->Outputs,
 		 machine->Consts,
diff --git a/src/gallium/drivers/softpipe/sp_prim_setup.c b/src/gallium/drivers/softpipe/sp_prim_setup.c
index d73521ccbe..7b1e131ee1 100644
--- a/src/gallium/drivers/softpipe/sp_prim_setup.c
+++ b/src/gallium/drivers/softpipe/sp_prim_setup.c
@@ -476,33 +476,33 @@ static void tri_persp_coeff( struct setup_stage *setup,
  * We could do a bit less work if we'd examine gl_FragCoord's swizzle mask.
  */
 static void
-setup_fragcoord_coeff(struct setup_stage *setup)
+setup_fragcoord_coeff(struct setup_stage *setup, uint slot)
 {
    /*X*/
-   setup->coef[0].a0[0] = 0;
-   setup->coef[0].dadx[0] = 1.0;
-   setup->coef[0].dady[0] = 0.0;
+   setup->coef[slot].a0[0] = 0;
+   setup->coef[slot].dadx[0] = 1.0;
+   setup->coef[slot].dady[0] = 0.0;
    /*Y*/
    if (setup->softpipe->rasterizer->origin_lower_left) {
       /* y=0=bottom */
       const int winHeight = setup->softpipe->framebuffer.cbufs[0]->height;
-      setup->coef[0].a0[1] = (float) (winHeight - 1);
-      setup->coef[0].dady[1] = -1.0;
+      setup->coef[slot].a0[1] = (float) (winHeight - 1);
+      setup->coef[slot].dady[1] = -1.0;
    }
    else {
       /* y=0=top */
-      setup->coef[0].a0[1] = 0.0;
-      setup->coef[0].dady[1] = 1.0;
+      setup->coef[slot].a0[1] = 0.0;
+      setup->coef[slot].dady[1] = 1.0;
    }
-   setup->coef[0].dadx[1] = 0.0;
+   setup->coef[slot].dadx[1] = 0.0;
    /*Z*/
-   setup->coef[0].a0[2] = setup->posCoef.a0[2];
-   setup->coef[0].dadx[2] = setup->posCoef.dadx[2];
-   setup->coef[0].dady[2] = setup->posCoef.dady[2];
+   setup->coef[slot].a0[2] = setup->posCoef.a0[2];
+   setup->coef[slot].dadx[2] = setup->posCoef.dadx[2];
+   setup->coef[slot].dady[2] = setup->posCoef.dady[2];
    /*W*/
-   setup->coef[0].a0[3] = setup->posCoef.a0[3];
-   setup->coef[0].dadx[3] = setup->posCoef.dadx[3];
-   setup->coef[0].dady[3] = setup->posCoef.dady[3];
+   setup->coef[slot].a0[3] = setup->posCoef.a0[3];
+   setup->coef[slot].dadx[3] = setup->posCoef.dadx[3];
+   setup->coef[slot].dady[3] = setup->posCoef.dady[3];
 }
 
 
@@ -543,8 +543,7 @@ static void setup_tri_coefficients( struct setup_stage *setup )
             tri_persp_coeff(setup, &setup->coef[fragSlot], vertSlot, j);
          break;
       case INTERP_POS:
-         assert(fragSlot == 0);
-         setup_fragcoord_coeff(setup);
+         setup_fragcoord_coeff(setup, fragSlot);
          break;
       default:
          assert(0);
@@ -798,9 +797,7 @@ setup_line_coefficients(struct setup_stage *setup, struct prim_header *prim)
             line_persp_coeff(setup, &setup->coef[fragSlot], vertSlot, j);
          break;
       case INTERP_POS:
-         assert(fragSlot == 0);
-         assert(0); /* XXX fix this: */
-         setup_fragcoord_coeff(setup);
+         setup_fragcoord_coeff(setup, fragSlot);
          break;
       default:
          assert(0);
@@ -1022,9 +1019,7 @@ setup_point(struct draw_stage *stage, struct prim_header *prim)
                               &setup->coef[fragSlot], vertSlot, j);
          break;
       case INTERP_POS:
-         assert(fragSlot == 0);
-         assert(0); /* XXX fix this: */
-         setup_fragcoord_coeff(setup);
+         setup_fragcoord_coeff(setup, fragSlot);
          break;
       default:
          assert(0);
diff --git a/src/gallium/drivers/softpipe/sp_quad.c b/src/gallium/drivers/softpipe/sp_quad.c
index 6bd468a51c..15b5594547 100644
--- a/src/gallium/drivers/softpipe/sp_quad.c
+++ b/src/gallium/drivers/softpipe/sp_quad.c
@@ -112,7 +112,9 @@ sp_build_quad_pipeline(struct softpipe_context *sp)
       sp_push_quad_first( sp, sp->quad.earlyz );
    }
 
+#if !USE_DRAW_STAGE_PSTIPPLE
    if (sp->rasterizer->poly_stipple_enable) {
       sp_push_quad_first( sp, sp->quad.polygon_stipple );
    }
+#endif
 }
diff --git a/src/gallium/drivers/softpipe/sp_quad_fs.c b/src/gallium/drivers/softpipe/sp_quad_fs.c
index cf1b1eff75..2f40e09d5c 100644
--- a/src/gallium/drivers/softpipe/sp_quad_fs.c
+++ b/src/gallium/drivers/softpipe/sp_quad_fs.c
@@ -142,7 +142,7 @@ static void shade_begin(struct quad_stage *qs)
    /* set TGSI sampler state that varies */
    for (i = 0; i < PIPE_MAX_SAMPLERS; i++) {
       qss->samplers[i].state = softpipe->sampler[i];
-      qss->samplers[i].texture = &softpipe->texture[i]->base;
+      qss->samplers[i].texture = softpipe->texture[i];
    }
 
    /* find output slots for depth, color */
diff --git a/src/gallium/drivers/softpipe/sp_state_clip.c b/src/gallium/drivers/softpipe/sp_state_clip.c
index c797c0dd3b..4946c776e3 100644
--- a/src/gallium/drivers/softpipe/sp_state_clip.c
+++ b/src/gallium/drivers/softpipe/sp_state_clip.c
@@ -42,24 +42,16 @@ void softpipe_set_clip_state( struct pipe_context *pipe,
 }
 
 
-
-/* Called when driver state tracker notices changes to the viewport
- * matrix:
- */
 void softpipe_set_viewport_state( struct pipe_context *pipe,
                                   const struct pipe_viewport_state *viewport )
 {
    struct softpipe_context *softpipe = softpipe_context(pipe);
 
-   softpipe->viewport = *viewport; /* struct copy */
-   softpipe->dirty |= SP_NEW_VIEWPORT;
-
    /* pass the viewport info to the draw module */
    draw_set_viewport_state(softpipe->draw, viewport);
 
-   /* Using tnl/ and vf/ modules is temporary while getting started.
-    * Full pipe will have vertex shader, vertex fetch of its own.
-    */
+   softpipe->viewport = *viewport; /* struct copy */
+   softpipe->dirty |= SP_NEW_VIEWPORT;
 }
 
 
@@ -68,7 +60,9 @@ void softpipe_set_scissor_state( struct pipe_context *pipe,
 {
    struct softpipe_context *softpipe = softpipe_context(pipe);
 
-   memcpy( &softpipe->scissor, scissor, sizeof(*scissor) );
+   draw_flush(softpipe->draw);
+
+   softpipe->scissor = *scissor; /* struct copy */
    softpipe->dirty |= SP_NEW_SCISSOR;
 }
 
@@ -78,6 +72,8 @@ void softpipe_set_polygon_stipple( struct pipe_context *pipe,
 {
    struct softpipe_context *softpipe = softpipe_context(pipe);
 
-   memcpy( &softpipe->poly_stipple, stipple, sizeof(*stipple) );
+   draw_flush(softpipe->draw);
+
+   softpipe->poly_stipple = *stipple; /* struct copy */
    softpipe->dirty |= SP_NEW_STIPPLE;
 }
diff --git a/src/gallium/drivers/softpipe/sp_state_derived.c b/src/gallium/drivers/softpipe/sp_state_derived.c
index 9d8fd8b750..f9f2c5eaa8 100644
--- a/src/gallium/drivers/softpipe/sp_state_derived.c
+++ b/src/gallium/drivers/softpipe/sp_state_derived.c
@@ -44,7 +44,8 @@
  * condition that users shouldn't hit anyway.
  */
 static int
-find_vs_output(const struct pipe_shader_state *vs,
+find_vs_output(struct softpipe_context *sp,
+               const struct pipe_shader_state *vs,
                uint semantic_name,
                uint semantic_index)
 {
@@ -54,7 +55,9 @@ find_vs_output(const struct pipe_shader_state *vs,
           vs->output_semantic_index[i] == semantic_index)
          return i;
    }
-   return 0;
+
+   /* See if the draw module is introducing a new attribute... */
+   return draw_find_vs_output(sp->draw, semantic_name, semantic_index);
 }
 
 
@@ -111,24 +114,24 @@ softpipe_get_vertex_info(struct softpipe_context *softpipe)
          int src;
          switch (fs->input_semantic_name[i]) {
          case TGSI_SEMANTIC_POSITION:
-            src = find_vs_output(vs, TGSI_SEMANTIC_POSITION, 0);
+            src = find_vs_output(softpipe, vs, TGSI_SEMANTIC_POSITION, 0);
             draw_emit_vertex_attr(vinfo, EMIT_4F, INTERP_POS, src);
             break;
 
          case TGSI_SEMANTIC_COLOR:
-            src = find_vs_output(vs, TGSI_SEMANTIC_COLOR, 
+            src = find_vs_output(softpipe, vs, TGSI_SEMANTIC_COLOR, 
                                  fs->input_semantic_index[i]);
             draw_emit_vertex_attr(vinfo, EMIT_4F, colorInterp, src);
             break;
 
          case TGSI_SEMANTIC_FOG:
-            src = find_vs_output(vs, TGSI_SEMANTIC_FOG, 0);
+            src = find_vs_output(softpipe, vs, TGSI_SEMANTIC_FOG, 0);
             draw_emit_vertex_attr(vinfo, EMIT_4F, INTERP_PERSPECTIVE, src);
             break;
 
          case TGSI_SEMANTIC_GENERIC:
             /* this includes texcoords and varying vars */
-            src = find_vs_output(vs, TGSI_SEMANTIC_GENERIC,
+            src = find_vs_output(softpipe, vs, TGSI_SEMANTIC_GENERIC,
                                  fs->input_semantic_index[i]);
             draw_emit_vertex_attr(vinfo, EMIT_4F, INTERP_PERSPECTIVE, src);
             break;
@@ -138,7 +141,7 @@ softpipe_get_vertex_info(struct softpipe_context *softpipe)
          }
       }
 
-      softpipe->psize_slot = find_vs_output(vs, TGSI_SEMANTIC_PSIZE, 0);
+      softpipe->psize_slot = find_vs_output(softpipe, vs, TGSI_SEMANTIC_PSIZE, 0);
       if (softpipe->psize_slot > 0) {
          draw_emit_vertex_attr(vinfo, EMIT_4F, INTERP_CONSTANT,
                                softpipe->psize_slot);
diff --git a/src/gallium/drivers/softpipe/sp_state_sampler.c b/src/gallium/drivers/softpipe/sp_state_sampler.c
index 460adccec4..18669a1c6e 100644
--- a/src/gallium/drivers/softpipe/sp_state_sampler.c
+++ b/src/gallium/drivers/softpipe/sp_state_sampler.c
@@ -30,6 +30,7 @@
  */
 
 #include "pipe/p_util.h"
+#include "pipe/p_inlines.h"
 
 #include "draw/draw_context.h"
 
@@ -82,9 +83,9 @@ softpipe_set_sampler_texture(struct pipe_context *pipe,
    draw_flush(softpipe->draw);
 
    assert(unit < PIPE_MAX_SAMPLERS);
-   softpipe->texture[unit] = softpipe_texture(texture);  /* ptr, not struct */
+   pipe_texture_reference(pipe, &softpipe->texture[unit], texture);
 
-   sp_tile_cache_set_texture(softpipe->tex_cache[unit], texture);
+   sp_tile_cache_set_texture(pipe, softpipe->tex_cache[unit], texture);
 
    softpipe->dirty |= SP_NEW_TEXTURE;
 }
diff --git a/src/gallium/drivers/softpipe/sp_state_surface.c b/src/gallium/drivers/softpipe/sp_state_surface.c
index e2c6893e9f..124b18b708 100644
--- a/src/gallium/drivers/softpipe/sp_state_surface.c
+++ b/src/gallium/drivers/softpipe/sp_state_surface.c
@@ -27,7 +27,7 @@
 
 /* Authors:  Keith Whitwell <keith@tungstengraphics.com>
  */
-#include "p_inlines.h"
+#include "pipe/p_inlines.h"
 
 #include "sp_context.h"
 #include "sp_state.h"
diff --git a/src/gallium/drivers/softpipe/sp_tex_sample.c b/src/gallium/drivers/softpipe/sp_tex_sample.c
index c54e9d385c..43d5085895 100644
--- a/src/gallium/drivers/softpipe/sp_tex_sample.c
+++ b/src/gallium/drivers/softpipe/sp_tex_sample.c
@@ -332,6 +332,61 @@ linear_texcoord(unsigned wrapMode, float s, unsigned size,
 }
 
 
+/**
+ * For RECT textures / unnormalized texcoords
+ * Only a subset of wrap modes supported.
+ */
+static INLINE int
+nearest_texcoord_unnorm(unsigned wrapMode, float s, unsigned size)
+{
+   int i;
+   switch (wrapMode) {
+   case PIPE_TEX_WRAP_CLAMP:
+      i = ifloor(s);
+      return CLAMP(i, 0, size-1);
+   case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
+      /* fall-through */
+   case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
+      return ifloor( CLAMP(s, 0.5F, (float) size - 0.5F) );
+   default:
+      assert(0);
+      return 0;
+   }
+}
+
+
+/**
+ * For RECT textures / unnormalized texcoords.
+ * Only a subset of wrap modes supported.
+ */
+static INLINE void
+linear_texcoord_unnorm(unsigned wrapMode, float s, unsigned size,
+                       int *i0, int *i1, float *a)
+{
+   switch (wrapMode) {
+   case PIPE_TEX_WRAP_CLAMP:
+      /* Not exactly what the spec says, but it matches NVIDIA output */
+      s = CLAMP(s - 0.5F, 0.0, (float) size - 1.0);
+      *i0 = ifloor(s);
+      *i1 = *i0 + 1;
+      break;
+   case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
+      /* fall-through */
+   case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
+      s = CLAMP(s, 0.5F, (float) size - 0.5F);
+      s -= 0.5F;
+      *i0 = ifloor(s);
+      *i1 = *i0 + 1;
+      if (*i1 > size - 1)
+         *i1 = size - 1;
+      break;
+   default:
+      assert(0);
+   }
+   *a = FRAC(s);
+}
+
+
 static unsigned
 choose_cube_face(float rx, float ry, float rz, float *newS, float *newT)
 {
@@ -415,15 +470,15 @@ compute_lambda(struct tgsi_sampler *sampler,
 {
    float rho, lambda;
 
+   assert(sampler->state->normalized_coords);
+
    assert(s);
    {
       float dsdx = s[QUAD_BOTTOM_RIGHT] - s[QUAD_BOTTOM_LEFT];
       float dsdy = s[QUAD_TOP_LEFT]     - s[QUAD_BOTTOM_LEFT];
       dsdx = FABSF(dsdx);
       dsdy = FABSF(dsdy);
-      rho = MAX2(dsdx, dsdy);
-      if (sampler->state->normalized_coords)
-         rho *= sampler->texture->width[0];
+      rho = MAX2(dsdx, dsdy) * sampler->texture->width[0];
    }
    if (t) {
       float dtdx = t[QUAD_BOTTOM_RIGHT] - t[QUAD_BOTTOM_LEFT];
@@ -431,9 +486,7 @@ compute_lambda(struct tgsi_sampler *sampler,
       float max;
       dtdx = FABSF(dtdx);
       dtdy = FABSF(dtdy);
-      max = MAX2(dtdx, dtdy);
-      if (sampler->state->normalized_coords)
-         max *= sampler->texture->height[0];
+      max = MAX2(dtdx, dtdy) * sampler->texture->height[0];
       rho = MAX2(rho, max);
    }
    if (p) {
@@ -442,9 +495,7 @@ compute_lambda(struct tgsi_sampler *sampler,
       float max;
       dpdx = FABSF(dpdx);
       dpdy = FABSF(dpdy);
-      max = MAX2(dpdx, dpdy);
-      if (sampler->state->normalized_coords)
-         max *= sampler->texture->depth[0];
+      max = MAX2(dpdx, dpdy) * sampler->texture->depth[0];
       rho = MAX2(rho, max);
    }
 
@@ -628,13 +679,10 @@ sp_get_samples_2d_common(struct tgsi_sampler *sampler,
    choose_mipmap_levels(sampler, s, t, p, lodbias,
                         &level0, &level1, &levelBlend, &imgFilter);
 
-   if (sampler->state->normalized_coords) {
-      width = sampler->texture->width[level0];
-      height = sampler->texture->height[level0];
-   }
-   else {
-      width = height = 1;
-   }
+   assert(sampler->state->normalized_coords);
+
+   width = sampler->texture->width[level0];
+   height = sampler->texture->height[level0];
 
    assert(width > 0);
 
@@ -765,14 +813,11 @@ sp_get_samples_3d(struct tgsi_sampler *sampler,
    choose_mipmap_levels(sampler, s, t, p, lodbias,
                         &level0, &level1, &levelBlend, &imgFilter);
 
-   if (sampler->state->normalized_coords) {
-      width = sampler->texture->width[level0];
-      height = sampler->texture->height[level0];
-      depth = sampler->texture->depth[level0];
-   }
-   else {
-      width = height = depth = 1;
-   }
+   assert(sampler->state->normalized_coords);
+
+   width = sampler->texture->width[level0];
+   height = sampler->texture->height[level0];
+   depth = sampler->texture->depth[level0];
 
    assert(width > 0);
    assert(height > 0);
@@ -889,6 +934,73 @@ sp_get_samples_cube(struct tgsi_sampler *sampler,
 }
 
 
+static void
+sp_get_samples_rect(struct tgsi_sampler *sampler,
+                    const float s[QUAD_SIZE],
+                    const float t[QUAD_SIZE],
+                    const float p[QUAD_SIZE],
+                    float lodbias,
+                    float rgba[NUM_CHANNELS][QUAD_SIZE])
+{
+   //sp_get_samples_2d_common(sampler, s, t, p, lodbias, rgba, faces);
+   static const uint face = 0;
+   const uint compare_func = sampler->state->compare_func;
+   unsigned level0, level1, j, imgFilter;
+   int width, height;
+   float levelBlend;
+
+   choose_mipmap_levels(sampler, s, t, p, lodbias,
+                        &level0, &level1, &levelBlend, &imgFilter);
+
+   /* texture RECTS cannot be mipmapped */
+   assert(level0 == level1);
+
+   width = sampler->texture->width[level0];
+   height = sampler->texture->height[level0];
+
+   assert(width > 0);
+
+   switch (imgFilter) {
+   case PIPE_TEX_FILTER_NEAREST:
+      for (j = 0; j < QUAD_SIZE; j++) {
+         int x = nearest_texcoord_unnorm(sampler->state->wrap_s, s[j], width);
+         int y = nearest_texcoord_unnorm(sampler->state->wrap_t, t[j], height);
+         get_texel(sampler, face, level0, x, y, 0, rgba, j);
+         if (sampler->state->compare_mode == PIPE_TEX_COMPARE_R_TO_TEXTURE) {
+            shadow_compare(compare_func, rgba, p, j);
+         }
+      }
+      break;
+   case PIPE_TEX_FILTER_LINEAR:
+      for (j = 0; j < QUAD_SIZE; j++) {
+         float tx[4][4], a, b;
+         int x0, y0, x1, y1, c;
+         linear_texcoord_unnorm(sampler->state->wrap_s, s[j], width,  &x0, &x1, &a);
+         linear_texcoord_unnorm(sampler->state->wrap_t, t[j], height, &y0, &y1, &b);
+         get_texel(sampler, face, level0, x0, y0, 0, tx, 0);
+         get_texel(sampler, face, level0, x1, y0, 0, tx, 1);
+         get_texel(sampler, face, level0, x0, y1, 0, tx, 2);
+         get_texel(sampler, face, level0, x1, y1, 0, tx, 3);
+         if (sampler->state->compare_mode == PIPE_TEX_COMPARE_R_TO_TEXTURE) {
+            shadow_compare(compare_func, tx, p, 0);
+            shadow_compare(compare_func, tx, p, 1);
+            shadow_compare(compare_func, tx, p, 2);
+            shadow_compare(compare_func, tx, p, 3);
+         }
+
+         for (c = 0; c < 4; c++) {
+            rgba[c][j] = lerp_2d(a, b, tx[c][0], tx[c][1], tx[c][2], tx[c][3]);
+         }
+      }
+      break;
+   default:
+      assert(0);
+   }
+}
+
+
+
+
 /**
  * Called via tgsi_sampler::get_samples()
  * Use the sampler's state setting to get a filtered RGBA value
@@ -914,15 +1026,21 @@ sp_get_samples(struct tgsi_sampler *sampler,
 
    switch (sampler->texture->target) {
    case PIPE_TEXTURE_1D:
+      assert(sampler->state->normalized_coords);
       sp_get_samples_1d(sampler, s, t, p, lodbias, rgba);
       break;
    case PIPE_TEXTURE_2D:
-      sp_get_samples_2d(sampler, s, t, p, lodbias, rgba);
+      if (sampler->state->normalized_coords)
+         sp_get_samples_2d(sampler, s, t, p, lodbias, rgba);
+      else
+         sp_get_samples_rect(sampler, s, t, p, lodbias, rgba);
       break;
    case PIPE_TEXTURE_3D:
+      assert(sampler->state->normalized_coords);
       sp_get_samples_3d(sampler, s, t, p, lodbias, rgba);
       break;
    case PIPE_TEXTURE_CUBE:
+      assert(sampler->state->normalized_coords);
       sp_get_samples_cube(sampler, s, t, p, lodbias, rgba);
       break;
    default:
diff --git a/src/gallium/drivers/softpipe/sp_texture.c b/src/gallium/drivers/softpipe/sp_texture.c
index 6de7a9b543..6ba0f09e0a 100644
--- a/src/gallium/drivers/softpipe/sp_texture.c
+++ b/src/gallium/drivers/softpipe/sp_texture.c
@@ -39,6 +39,7 @@
 #include "sp_context.h"
 #include "sp_state.h"
 #include "sp_texture.h"
+#include "sp_tile_cache.h"
 
 
 /* Simple, maximally packed layout.
@@ -88,6 +89,7 @@ softpipe_texture_create(struct pipe_context *pipe,
       return NULL;
 
    spt->base = *templat;
+   spt->base.refcount = 1;
 
    softpipe_texture_layout(spt);
 
@@ -99,6 +101,8 @@ softpipe_texture_create(struct pipe_context *pipe,
       return NULL;
    }
 
+   assert(spt->base.refcount == 1);
+
    return &spt->base;
 }
 
@@ -128,6 +132,20 @@ softpipe_texture_release(struct pipe_context *pipe, struct pipe_texture **pt)
 }
 
 
+void
+softpipe_texture_update(struct pipe_context *pipe,
+                        struct pipe_texture *texture)
+{
+   struct softpipe_context *softpipe = softpipe_context(pipe);
+   uint unit;
+   for (unit = 0; unit < PIPE_MAX_SAMPLERS; unit++) {
+      if (softpipe->texture[unit] == texture) {
+         sp_flush_tile_cache(softpipe, softpipe->tex_cache[unit]);
+      }
+   }
+}
+
+
 /**
  * Called via pipe->get_tex_surface()
  */
diff --git a/src/gallium/drivers/softpipe/sp_texture.h b/src/gallium/drivers/softpipe/sp_texture.h
index fa646c0de9..50fc100427 100644
--- a/src/gallium/drivers/softpipe/sp_texture.h
+++ b/src/gallium/drivers/softpipe/sp_texture.h
@@ -62,6 +62,10 @@ softpipe_texture_create(struct pipe_context *pipe,
 extern void
 softpipe_texture_release(struct pipe_context *pipe, struct pipe_texture **pt);
 
+extern void
+softpipe_texture_update(struct pipe_context *pipe,
+                        struct pipe_texture *texture);
+
 extern struct pipe_surface *
 softpipe_get_tex_surface(struct pipe_context *pipe,
                          struct pipe_texture *pt,
diff --git a/src/gallium/drivers/softpipe/sp_tile_cache.c b/src/gallium/drivers/softpipe/sp_tile_cache.c
index dde3fabc81..da30dd6c48 100644
--- a/src/gallium/drivers/softpipe/sp_tile_cache.c
+++ b/src/gallium/drivers/softpipe/sp_tile_cache.c
@@ -212,14 +212,15 @@ sp_tile_cache_unmap_surfaces(struct softpipe_tile_cache *tc)
  * Specify the texture to cache.
  */
 void
-sp_tile_cache_set_texture(struct softpipe_tile_cache *tc,
+sp_tile_cache_set_texture(struct pipe_context *pipe,
+                          struct softpipe_tile_cache *tc,
                           struct pipe_texture *texture)
 {
    uint i;
 
    assert(!tc->surface);
 
-   tc->texture = texture;
+   pipe_texture_reference(pipe, &tc->texture, texture);
 
    if (tc->tex_surf_map) {
       pipe_surface_unmap(tc->tex_surf);
@@ -358,30 +359,37 @@ sp_flush_tile_cache(struct softpipe_context *softpipe,
    struct pipe_surface *ps = tc->surface;
    int inuse = 0, pos;
 
-   if (!ps || !ps->buffer)
-      return;
-
-   for (pos = 0; pos < NUM_ENTRIES; pos++) {
-      struct softpipe_cached_tile *tile = tc->entries + pos;
-      if (tile->x >= 0) {
-         if (tc->depth_stencil) {
-            pipe_put_tile_raw(pipe, ps,
-                           tile->x, tile->y, TILE_SIZE, TILE_SIZE,
-                           tile->data.depth32, 0/*STRIDE*/);
-         }
-         else {
-            pipe_put_tile_rgba(pipe, ps,
-                               tile->x, tile->y, TILE_SIZE, TILE_SIZE,
-                               (float *) tile->data.color);
+   if (ps && ps->buffer) {
+      /* caching a drawing surface */
+      for (pos = 0; pos < NUM_ENTRIES; pos++) {
+         struct softpipe_cached_tile *tile = tc->entries + pos;
+         if (tile->x >= 0) {
+            if (tc->depth_stencil) {
+               pipe_put_tile_raw(pipe, ps,
+                              tile->x, tile->y, TILE_SIZE, TILE_SIZE,
+                              tile->data.depth32, 0/*STRIDE*/);
+            }
+            else {
+               pipe_put_tile_rgba(pipe, ps,
+                                  tile->x, tile->y, TILE_SIZE, TILE_SIZE,
+                                  (float *) tile->data.color);
+            }
+            tile->x = tile->y = -1;  /* mark as empty */
+            inuse++;
          }
-         tile->x = tile->y = -1;  /* mark as empty */
-         inuse++;
       }
-   }
 
 #if TILE_CLEAR_OPTIMIZATION
-   sp_tile_cache_flush_clear(&softpipe->pipe, tc);
+      sp_tile_cache_flush_clear(&softpipe->pipe, tc);
 #endif
+   }
+   else if (tc->texture) {
+      /* caching a texture, mark all entries as embpy */
+      for (pos = 0; pos < NUM_ENTRIES; pos++) {
+         tc->entries[pos].x = -1;
+      }
+      tc->tex_face = -1;
+   }
 
 #if 0
    debug_printf("flushed tiles in use: %d\n", inuse);
diff --git a/src/gallium/drivers/softpipe/sp_tile_cache.h b/src/gallium/drivers/softpipe/sp_tile_cache.h
index 7fd1081286..2631e29a3a 100644
--- a/src/gallium/drivers/softpipe/sp_tile_cache.h
+++ b/src/gallium/drivers/softpipe/sp_tile_cache.h
@@ -80,7 +80,8 @@ extern void
 sp_tile_cache_unmap_surfaces(struct softpipe_tile_cache *tc);
 
 extern void
-sp_tile_cache_set_texture(struct softpipe_tile_cache *tc,
+sp_tile_cache_set_texture(struct pipe_context *pipe,
+                          struct softpipe_tile_cache *tc,
                           struct pipe_texture *texture);
 
 extern void