llvmpipe: reintroduce SET_STATE binner command

But bin lazily only into bins which are receiving geometry.

9 files changed, 100 insertions, 46 deletions

diff --git a/src/gallium/drivers/llvmpipe/lp_rast.c b/src/gallium/drivers/llvmpipe/lp_rast.c
index db9b2f9b128..35e2f731e89 100644
--- a/src/gallium/drivers/llvmpipe/lp_rast.c
+++ b/src/gallium/drivers/llvmpipe/lp_rast.c
@@ -334,7 +334,7 @@ lp_rast_shade_tile(struct lp_rasterizer_task *task,
 {
    const struct lp_scene *scene = task->scene;
    const struct lp_rast_shader_inputs *inputs = arg.shade_tile;
-   const struct lp_rast_state *state = inputs->state;
+   const struct lp_rast_state *state = task->state;
    struct lp_fragment_shader_variant *variant = state->variant;
    const unsigned tile_x = task->x, tile_y = task->y;
    unsigned x, y;
@@ -414,7 +414,7 @@ lp_rast_shade_quads_mask(struct lp_rasterizer_task *task,
                          unsigned x, unsigned y,
                          unsigned mask)
 {
-   const struct lp_rast_state *state = inputs->state;
+   const struct lp_rast_state *state = task->state;
    struct lp_fragment_shader_variant *variant = state->variant;
    const struct lp_scene *scene = task->scene;
    uint8_t *color[PIPE_MAX_COLOR_BUFS];
@@ -490,6 +490,14 @@ lp_rast_end_query(struct lp_rasterizer_task *task,
 }
 
 
+void
+lp_rast_set_state(struct lp_rasterizer_task *task,
+                  const union lp_rast_cmd_arg arg)
+{
+   task->state = arg.state;
+}
+
+
 
 /**
  * Set top row and left column of the tile's pixels to white.  For debugging.
@@ -602,6 +610,7 @@ static lp_rast_cmd_func dispatch[LP_RAST_OP_MAX] =
    lp_rast_shade_tile_opaque,
    lp_rast_begin_query,
    lp_rast_end_query,
+   lp_rast_set_state,
 };
 
 
diff --git a/src/gallium/drivers/llvmpipe/lp_rast.h b/src/gallium/drivers/llvmpipe/lp_rast.h
index e2bcc450168..f74b198a66c 100644
--- a/src/gallium/drivers/llvmpipe/lp_rast.h
+++ b/src/gallium/drivers/llvmpipe/lp_rast.h
@@ -85,8 +85,6 @@ struct lp_rast_shader_inputs {
    float (*a0)[4];
    float (*dadx)[4];
    float (*dady)[4];
-
-   const struct lp_rast_state *state;
 };
 
 /* Note: the order of these values is important as they are loaded by
@@ -154,6 +152,7 @@ union lp_rast_cmd_arg {
       uint32_t value;
       uint32_t mask;
    } clear_zstencil;
+   const struct lp_rast_state *state;
    struct lp_fence *fence;
    struct llvmpipe_query *query_obj;
 };
@@ -245,8 +244,9 @@ lp_rast_arg_null( void )
 #define LP_RAST_OP_SHADE_TILE_OPAQUE 0xe
 #define LP_RAST_OP_BEGIN_QUERY       0xf
 #define LP_RAST_OP_END_QUERY         0x10
+#define LP_RAST_OP_SET_STATE         0x11
 
-#define LP_RAST_OP_MAX               0x11
+#define LP_RAST_OP_MAX               0x12
 #define LP_RAST_OP_MASK              0xff
 
 void
diff --git a/src/gallium/drivers/llvmpipe/lp_rast_debug.c b/src/gallium/drivers/llvmpipe/lp_rast_debug.c
index 6f4ba1c6fef..3113e196c40 100644
--- a/src/gallium/drivers/llvmpipe/lp_rast_debug.c
+++ b/src/gallium/drivers/llvmpipe/lp_rast_debug.c
@@ -12,6 +12,7 @@ static INLINE int u_bit_scan(unsigned *mask)
 struct tile {
    int coverage;
    int overdraw;
+   const struct lp_rast_state *state;
    char data[TILE_SIZE][TILE_SIZE];
 };
 
@@ -47,6 +48,7 @@ static const char *cmd_names[LP_RAST_OP_MAX] =
    "shade_tile_opaque",
    "begin_query",
    "end_query",
+   "set_state",
 };
 
 static const char *cmd_name(unsigned cmd)
@@ -56,31 +58,31 @@ static const char *cmd_name(unsigned cmd)
 }
 
 static const struct lp_fragment_shader_variant *
-get_variant(  const struct cmd_block *block,
-              int k )
+get_variant( const struct lp_rast_state *state,
+             const struct cmd_block *block,
+             int k )
 {
    if (block->cmd[k] == LP_RAST_OP_SHADE_TILE ||
-       block->cmd[k] == LP_RAST_OP_SHADE_TILE_OPAQUE)
-      return  block->arg[k].shade_tile->state->variant;
-
-   if (block->cmd[k] == LP_RAST_OP_TRIANGLE_1 ||
+       block->cmd[k] == LP_RAST_OP_SHADE_TILE_OPAQUE ||
+       block->cmd[k] == LP_RAST_OP_TRIANGLE_1 ||
        block->cmd[k] == LP_RAST_OP_TRIANGLE_2 ||
        block->cmd[k] == LP_RAST_OP_TRIANGLE_3 ||
        block->cmd[k] == LP_RAST_OP_TRIANGLE_4 ||
        block->cmd[k] == LP_RAST_OP_TRIANGLE_5 ||
        block->cmd[k] == LP_RAST_OP_TRIANGLE_6 ||
        block->cmd[k] == LP_RAST_OP_TRIANGLE_7)
-      return block->arg[k].triangle.tri->inputs.state->variant;
+      return state->variant;
 
    return NULL;
 }
 
 
 static boolean
-is_blend( const struct cmd_block *block,
+is_blend( const struct lp_rast_state *state,
+          const struct cmd_block *block,
           int k )
 {
-   const struct lp_fragment_shader_variant *variant = get_variant(block, k);
+   const struct lp_fragment_shader_variant *variant = get_variant(state, block, k);
 
    if (variant)
       return  variant->key.blend.rt[0].blend_enable;
@@ -93,6 +95,7 @@ is_blend( const struct cmd_block *block,
 static void
 debug_bin( const struct cmd_bin *bin )
 {
+   const struct lp_rast_state *state;
    const struct cmd_block *head = bin->head;
    int i, j = 0;
 
@@ -100,9 +103,12 @@ debug_bin( const struct cmd_bin *bin )
                 
    while (head) {
       for (i = 0; i < head->count; i++, j++) {
+         if (head->cmd[i] == LP_RAST_OP_SET_STATE)
+            state = head->arg[i].state;
+
          debug_printf("%d: %s %s\n", j,
                       cmd_name(head->cmd[i]),
-                      is_blend(head, i) ? "blended" : "");
+                      is_blend(state, head, i) ? "blended" : "");
       }
       head = head->next;
    }
@@ -134,7 +140,7 @@ debug_shade_tile(int x, int y,
                  char val)
 {
    const struct lp_rast_shader_inputs *inputs = arg.shade_tile;
-   boolean blend = inputs->state->variant->key.blend.rt[0].blend_enable;
+   boolean blend = tile->state->variant->key.blend.rt[0].blend_enable;
    unsigned i,j;
 
    if (inputs->disable)
@@ -176,7 +182,7 @@ debug_triangle(int tilex, int tiley,
    int x, y;
    int count = 0;
    unsigned i, nr_planes = 0;
-   boolean blend = tri->inputs.state->variant->key.blend.rt[0].blend_enable;
+   boolean blend = tile->state->variant->key.blend.rt[0].blend_enable;
 
    if (tri->inputs.disable) {
       /* This triangle was partially binned and has been disabled */
@@ -236,12 +242,15 @@ do_debug_bin( struct tile *tile,
 
    for (block = bin->head; block; block = block->next) {
       for (k = 0; k < block->count; k++, j++) {
-         boolean blend = is_blend(block, k);
+         boolean blend = is_blend(tile->state, block, k);
          char val = get_label(j);
          int count = 0;
             
          if (print_cmds)
             debug_printf("%c: %15s", val, cmd_name(block->cmd[k]));
+
+         if (block->cmd[k] == LP_RAST_OP_SET_STATE)
+            tile->state = block->arg[k].state;
          
          if (block->cmd[k] == LP_RAST_OP_CLEAR_COLOR ||
              block->cmd[k] == LP_RAST_OP_CLEAR_ZSTENCIL)
diff --git a/src/gallium/drivers/llvmpipe/lp_rast_priv.h b/src/gallium/drivers/llvmpipe/lp_rast_priv.h
index 104000a040c..7ffd735def2 100644
--- a/src/gallium/drivers/llvmpipe/lp_rast_priv.h
+++ b/src/gallium/drivers/llvmpipe/lp_rast_priv.h
@@ -77,6 +77,7 @@ struct cmd_bin;
 struct lp_rasterizer_task
 {
    const struct cmd_bin *bin;
+   const struct lp_rast_state *state;
 
    struct lp_scene *scene;
    unsigned x, y;          /**< Pos of this tile in framebuffer, in pixels */
@@ -244,7 +245,7 @@ lp_rast_shade_quads_all( struct lp_rasterizer_task *task,
                          unsigned x, unsigned y )
 {
    const struct lp_scene *scene = task->scene;
-   const struct lp_rast_state *state = inputs->state;
+   const struct lp_rast_state *state = task->state;
    struct lp_fragment_shader_variant *variant = state->variant;
    uint8_t *color[PIPE_MAX_COLOR_BUFS];
    void *depth;
@@ -298,6 +299,10 @@ void lp_rast_triangle_4_16( struct lp_rasterizer_task *,
                             const union lp_rast_cmd_arg );
 
 void
+lp_rast_set_state(struct lp_rasterizer_task *task,
+                  const union lp_rast_cmd_arg arg);
+ 
+void
 lp_debug_bin( const struct cmd_bin *bin );
 
 #endif
diff --git a/src/gallium/drivers/llvmpipe/lp_scene.c b/src/gallium/drivers/llvmpipe/lp_scene.c
index 8b504f23a33..a4fdf7cff36 100644
--- a/src/gallium/drivers/llvmpipe/lp_scene.c
+++ b/src/gallium/drivers/llvmpipe/lp_scene.c
@@ -203,7 +203,9 @@ lp_scene_end_rasterization(struct lp_scene *scene )
    for (i = 0; i < scene->tiles_x; i++) {
       for (j = 0; j < scene->tiles_y; j++) {
          struct cmd_bin *bin = lp_scene_get_bin(scene, i, j);
-         bin->head = bin->tail = NULL;
+         bin->head = NULL;
+         bin->tail = NULL;
+         bin->last_state = NULL;
       }
    }
 
diff --git a/src/gallium/drivers/llvmpipe/lp_scene.h b/src/gallium/drivers/llvmpipe/lp_scene.h
index dbef7692e42..622c522f11a 100644
--- a/src/gallium/drivers/llvmpipe/lp_scene.h
+++ b/src/gallium/drivers/llvmpipe/lp_scene.h
@@ -41,6 +41,7 @@
 #include "lp_debug.h"
 
 struct lp_scene_queue;
+struct lp_rast_state;
 
 /* We're limited to 2K by 2K for 32bit fixed point rasterization.
  * Will need a 64-bit version for larger framebuffers.
@@ -94,6 +95,7 @@ struct data_block {
 struct cmd_bin {
    ushort x;
    ushort y;
+   const struct lp_rast_state *last_state;       /* most recent state set in bin */
    struct cmd_block *head;
    struct cmd_block *tail;
 };
@@ -297,7 +299,7 @@ lp_scene_bin_command( struct lp_scene *scene,
 
    assert(x < scene->tiles_x);
    assert(y < scene->tiles_y);
-   assert(cmd <= LP_RAST_OP_END_QUERY);
+   assert(cmd < LP_RAST_OP_MAX);
 
    if (tail == NULL || tail->count == CMD_BLOCK_MAX) {
       tail = lp_scene_new_cmd_block( scene, bin );
@@ -318,6 +320,30 @@ lp_scene_bin_command( struct lp_scene *scene,
 }
 
 
+static INLINE boolean
+lp_scene_bin_cmd_with_state( struct lp_scene *scene,
+                             unsigned x, unsigned y,
+                             const struct lp_rast_state *state,
+                             unsigned cmd,
+                             union lp_rast_cmd_arg arg )
+{
+   struct cmd_bin *bin = lp_scene_get_bin(scene, x, y);
+
+   if (state != bin->last_state) {
+      bin->last_state = state;
+      if (!lp_scene_bin_command(scene, x, y,
+                                LP_RAST_OP_SET_STATE,
+                                lp_rast_arg_state(state)))
+         return FALSE;
+   }
+
+   if (!lp_scene_bin_command( scene, x, y, cmd, arg ))
+      return FALSE;
+
+   return TRUE;
+}
+
+
 /* Add a command to all active bins.
  */
 static INLINE boolean
diff --git a/src/gallium/drivers/llvmpipe/lp_setup_line.c b/src/gallium/drivers/llvmpipe/lp_setup_line.c
index 693ac281752..e4cff9aa42c 100644
--- a/src/gallium/drivers/llvmpipe/lp_setup_line.c
+++ b/src/gallium/drivers/llvmpipe/lp_setup_line.c
@@ -597,7 +597,6 @@ try_setup_line( struct lp_setup_context *setup,
    setup_line_coefficients( setup, line, &info); 
 
    line->inputs.facing = 1.0F;
-   line->inputs.state = setup->fs.stored;
    line->inputs.disable = FALSE;
    line->inputs.opaque = FALSE;
 
diff --git a/src/gallium/drivers/llvmpipe/lp_setup_point.c b/src/gallium/drivers/llvmpipe/lp_setup_point.c
index 64b24a88d54..93c3efe347e 100644
--- a/src/gallium/drivers/llvmpipe/lp_setup_point.c
+++ b/src/gallium/drivers/llvmpipe/lp_setup_point.c
@@ -374,7 +374,6 @@ try_setup_point( struct lp_setup_context *setup,
    setup_point_coefficients(setup, point, &info);
 
    point->inputs.facing = 1.0F;
-   point->inputs.state = setup->fs.stored;
    point->inputs.disable = FALSE;
    point->inputs.opaque = FALSE;
 
diff --git a/src/gallium/drivers/llvmpipe/lp_setup_tri.c b/src/gallium/drivers/llvmpipe/lp_setup_tri.c
index 8fd034666c3..bc48eb8d1b5 100644
--- a/src/gallium/drivers/llvmpipe/lp_setup_tri.c
+++ b/src/gallium/drivers/llvmpipe/lp_setup_tri.c
@@ -200,14 +200,16 @@ lp_setup_whole_tile(struct lp_setup_context *setup,
       }
 
       LP_COUNT(nr_shade_opaque_64);
-      return lp_scene_bin_command( scene, tx, ty,
-                                   LP_RAST_OP_SHADE_TILE_OPAQUE,
-                                   lp_rast_arg_inputs(inputs) );
+      return lp_scene_bin_cmd_with_state( scene, tx, ty,
+                                          setup->fs.stored,
+                                          LP_RAST_OP_SHADE_TILE_OPAQUE,
+                                          lp_rast_arg_inputs(inputs) );
    } else {
       LP_COUNT(nr_shade_64);
-      return lp_scene_bin_command( scene, tx, ty,
-                                   LP_RAST_OP_SHADE_TILE,
-                                   lp_rast_arg_inputs(inputs) );
+      return lp_scene_bin_cmd_with_state( scene, tx, ty,
+                                          setup->fs.stored, 
+                                          LP_RAST_OP_SHADE_TILE,
+                                          lp_rast_arg_inputs(inputs) );
    }
 }
 
@@ -320,7 +322,6 @@ do_triangle_ccw(struct lp_setup_context *setup,
    tri->inputs.facing = frontfacing ? 1.0F : -1.0F;
    tri->inputs.disable = FALSE;
    tri->inputs.opaque = setup->fs.current.variant->opaque;
-   tri->inputs.state = setup->fs.stored;
 
   
    for (i = 0; i < 3; i++) {
@@ -491,34 +492,36 @@ lp_setup_bin_triangle( struct lp_setup_context *setup,
          {
             /* Triangle is contained in a single 4x4 stamp:
              */
-
-            return lp_scene_bin_command( scene, ix0, iy0,
-                                         LP_RAST_OP_TRIANGLE_3_4,
-                                         lp_rast_arg_triangle(tri, mask) );
+            return lp_scene_bin_cmd_with_state( scene, ix0, iy0,
+                                                setup->fs.stored,
+                                                LP_RAST_OP_TRIANGLE_3_4,
+                                                lp_rast_arg_triangle(tri, mask) );
          }
 
          if (sz < 16)
          {
             /* Triangle is contained in a single 16x16 block:
              */
-            return lp_scene_bin_command( scene, ix0, iy0,
-                                         LP_RAST_OP_TRIANGLE_3_16,
-                                         lp_rast_arg_triangle(tri, mask) );
+            return lp_scene_bin_cmd_with_state( scene, ix0, iy0,
+                                                setup->fs.stored,
+                                                LP_RAST_OP_TRIANGLE_3_16,
+                                                lp_rast_arg_triangle(tri, mask) );
          }
       }
       else if (nr_planes == 4 && sz < 16) 
       {
-         return lp_scene_bin_command( scene, ix0, iy0,
-                                      LP_RAST_OP_TRIANGLE_4_16,
-                                      lp_rast_arg_triangle(tri, mask) );
+         return lp_scene_bin_cmd_with_state(scene, ix0, iy0,
+                                            setup->fs.stored,
+                                            LP_RAST_OP_TRIANGLE_4_16,
+                                            lp_rast_arg_triangle(tri, mask) );
       }
 
 
       /* Triangle is contained in a single tile:
        */
-      return lp_scene_bin_command( scene, ix0, iy0,
-                                   lp_rast_tri_tab[nr_planes], 
-                                   lp_rast_arg_triangle(tri, (1<<nr_planes)-1) );
+      return lp_scene_bin_cmd_with_state( scene, ix0, iy0, setup->fs.stored,
+                                          lp_rast_tri_tab[nr_planes], 
+                                          lp_rast_arg_triangle(tri, (1<<nr_planes)-1) );
    }
    else
    {
@@ -584,9 +587,11 @@ lp_setup_bin_triangle( struct lp_setup_context *setup,
                 */
                int count = util_bitcount(partial);
                in = TRUE;
-               if (!lp_scene_bin_command( scene, x, y,
-                                          lp_rast_tri_tab[count], 
-                                          lp_rast_arg_triangle(tri, partial) ))
+               
+               if (!lp_scene_bin_cmd_with_state( scene, x, y,
+                                                 setup->fs.stored,
+                                                 lp_rast_tri_tab[count], 
+                                                 lp_rast_arg_triangle(tri, partial) ))
                   goto fail;
 
                LP_COUNT(nr_partially_covered_64);