llvmpipe: Fix alignment.
[mesa.git] / src / gallium / drivers / llvmpipe / lp_rast.c
index a00a592f2fe120d6fe4b7f272a1244f7eb02daef..40a8591cabe14c8b4738409a070b9d84188fd460 100644 (file)
 #include <limits.h>
 #include "util/u_memory.h"
 #include "util/u_math.h"
+#include "util/u_rect.h"
 #include "util/u_surface.h"
+#include "util/u_pack_color.h"
+
+#include "os/os_time.h"
 
 #include "lp_scene_queue.h"
 #include "lp_debug.h"
 #include "lp_fence.h"
 #include "lp_perf.h"
+#include "lp_query.h"
 #include "lp_rast.h"
 #include "lp_rast_priv.h"
-#include "lp_tile_soa.h"
 #include "gallivm/lp_bld_debug.h"
 #include "lp_scene.h"
+#include "lp_tex_sample.h"
+
+
+#ifdef DEBUG
+int jit_line = 0;
+const struct lp_rast_state *jit_state = NULL;
+const struct lp_rasterizer_task *jit_task = NULL;
+#endif
 
 
 /**
@@ -49,44 +61,12 @@ static void
 lp_rast_begin( struct lp_rasterizer *rast,
                struct lp_scene *scene )
 {
-   const struct pipe_framebuffer_state *fb = &scene->fb;
-   int i;
 
    rast->curr_scene = scene;
 
    LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
 
-   rast->state.nr_cbufs = scene->fb.nr_cbufs;
-   
-   for (i = 0; i < rast->state.nr_cbufs; i++) {
-      struct pipe_surface *cbuf = scene->fb.cbufs[i];
-      rast->cbuf[i].format = cbuf->texture->format;
-      rast->cbuf[i].tiles_per_row = align(cbuf->width, TILE_SIZE) / TILE_SIZE;
-      rast->cbuf[i].blocksize = 
-         util_format_get_blocksize(cbuf->texture->format);
-      rast->cbuf[i].map = llvmpipe_resource_map(cbuf->texture,
-                                               cbuf->face,
-                                               cbuf->level,
-                                               cbuf->zslice,
-                                               LP_TEX_USAGE_READ_WRITE,
-                                               LP_TEX_LAYOUT_NONE);
-   }
-
-   if (fb->zsbuf) {
-      struct pipe_surface *zsbuf = scene->fb.zsbuf;
-      rast->zsbuf.stride = llvmpipe_resource_stride(zsbuf->texture, zsbuf->level);
-      rast->zsbuf.blocksize = 
-         util_format_get_blocksize(zsbuf->texture->format);
-
-      rast->zsbuf.map = llvmpipe_resource_map(zsbuf->texture,
-                                             zsbuf->face,
-                                             zsbuf->level,
-                                             zsbuf->zslice,
-                                             LP_TEX_USAGE_READ_WRITE,
-                                             LP_TEX_LAYOUT_NONE);
-      assert(rast->zsbuf.map);
-   }
-
+   lp_scene_begin_rasterization( scene );
    lp_scene_bin_iter_begin( scene );
 }
 
@@ -94,38 +74,9 @@ lp_rast_begin( struct lp_rasterizer *rast,
 static void
 lp_rast_end( struct lp_rasterizer *rast )
 {
-   struct lp_scene *scene = rast->curr_scene;
-   unsigned i;
-
-   /* Unmap color buffers */
-   for (i = 0; i < rast->state.nr_cbufs; i++) {
-      struct pipe_surface *cbuf = scene->fb.cbufs[i];
-      llvmpipe_resource_unmap(cbuf->texture,
-                             cbuf->face,
-                             cbuf->level,
-                             cbuf->zslice);
-      rast->cbuf[i].map = NULL;
-   }
-
-   /* Unmap z/stencil buffer */
-   if (rast->zsbuf.map) {
-      struct pipe_surface *zsbuf = scene->fb.zsbuf;
-      llvmpipe_resource_unmap(zsbuf->texture,
-                             zsbuf->face,
-                             zsbuf->level,
-                             zsbuf->zslice);
-      rast->zsbuf.map = NULL;
-   }
-
-   lp_scene_reset( rast->curr_scene );
+   lp_scene_end_rasterization( rast->curr_scene );
 
    rast->curr_scene = NULL;
-
-#ifdef DEBUG
-   if (0)
-      debug_printf("Post render scene: tile unswizzle: %u tile swizzle: %u\n",
-                   lp_tile_unswizzle_count, lp_tile_swizzle_count);
-#endif
 }
 
 
@@ -136,47 +87,27 @@ lp_rast_end( struct lp_rasterizer *rast )
  */
 static void
 lp_rast_tile_begin(struct lp_rasterizer_task *task,
-                   unsigned x, unsigned y)
+                   const struct cmd_bin *bin)
 {
-   struct lp_rasterizer *rast = task->rast;
-   struct lp_scene *scene = rast->curr_scene;
+   const struct lp_scene *scene = task->scene;
    enum lp_texture_usage usage;
-   unsigned buf;
-
-   LP_DBG(DEBUG_RAST, "%s %d,%d\n", __FUNCTION__, x, y);
-
-   assert(x % TILE_SIZE == 0);
-   assert(y % TILE_SIZE == 0);
-
-   task->x = x;
-   task->y = y;
-
-   if (scene->has_color_clear)
-      usage = LP_TEX_USAGE_WRITE_ALL;
-   else
-      usage = LP_TEX_USAGE_READ_WRITE;
-
-   /* get pointers to color tile(s) */
-   for (buf = 0; buf < rast->state.nr_cbufs; buf++) {
-      struct pipe_surface *cbuf = rast->curr_scene->fb.cbufs[buf];
-      struct llvmpipe_resource *lpt;
-      assert(cbuf);
-      lpt = llvmpipe_resource(cbuf->texture);
-      task->color_tiles[buf] = llvmpipe_get_texture_tile(lpt,
-                                                         cbuf->face + cbuf->zslice,
-                                                         cbuf->level,
-                                                         usage,
-                                                         x, y);
-      assert(task->color_tiles[buf]);
-   }
+
+   LP_DBG(DEBUG_RAST, "%s %d,%d\n", __FUNCTION__, bin->x, bin->y);
+
+   task->bin = bin;
+   task->x = bin->x * TILE_SIZE;
+   task->y = bin->y * TILE_SIZE;
+
+   /* reset pointers to color tile(s) */
+   memset(task->color_tiles, 0, sizeof(task->color_tiles));
 
    /* get pointer to depth/stencil tile */
    {
-      struct pipe_surface *zsbuf = rast->curr_scene->fb.zsbuf;
+      struct pipe_surface *zsbuf = task->scene->fb.zsbuf;
       if (zsbuf) {
          struct llvmpipe_resource *lpt = llvmpipe_resource(zsbuf->texture);
 
-         if (scene->has_depth_clear)
+         if (scene->has_depthstencil_clear)
             usage = LP_TEX_USAGE_WRITE_ALL;
          else
             usage = LP_TEX_USAGE_READ_WRITE;
@@ -185,14 +116,17 @@ lp_rast_tile_begin(struct lp_rasterizer_task *task,
           * and update the tile's layout info.
           */
          (void) llvmpipe_get_texture_tile(lpt,
-                                          zsbuf->face + zsbuf->zslice,
-                                          zsbuf->level,
+                                          zsbuf->u.tex.first_layer,
+                                          zsbuf->u.tex.level,
                                           usage,
-                                          x, y);
+                                          task->x,
+                                          task->y);
          /* Get actual pointer to the tile data.  Note that depth/stencil
           * data is tiled differently than color data.
           */
-         task->depth_tile = lp_rast_get_depth_block_pointer(rast, x, y);
+         task->depth_tile = lp_rast_get_depth_block_pointer(task,
+                                                            task->x,
+                                                            task->y);
 
          assert(task->depth_tile);
       }
@@ -207,78 +141,73 @@ lp_rast_tile_begin(struct lp_rasterizer_task *task,
  * Clear the rasterizer's current color tile.
  * This is a bin command called during bin processing.
  */
-void
+static void
 lp_rast_clear_color(struct lp_rasterizer_task *task,
                     const union lp_rast_cmd_arg arg)
 {
-   struct lp_rasterizer *rast = task->rast;
-   const uint8_t *clear_color = arg.clear_color;
+   const struct lp_scene *scene = task->scene;
+   uint8_t clear_color[4];
 
    unsigned i;
 
-   LP_DBG(DEBUG_RAST, "%s 0x%x,0x%x,0x%x,0x%x\n", __FUNCTION__, 
+   for (i = 0; i < 4; ++i) {
+      clear_color[i] = float_to_ubyte(arg.clear_color[i]);
+   }
+
+   LP_DBG(DEBUG_RAST, "%s 0x%x,0x%x,0x%x,0x%x\n", __FUNCTION__,
               clear_color[0],
               clear_color[1],
               clear_color[2],
               clear_color[3]);
 
-   if (clear_color[0] == clear_color[1] &&
-       clear_color[1] == clear_color[2] &&
-       clear_color[2] == clear_color[3]) {
-      /* clear to grayscale value {x, x, x, x} */
-      for (i = 0; i < rast->state.nr_cbufs; i++) {
-         uint8_t *ptr = task->color_tiles[i];
-        memset(ptr, clear_color[0], TILE_SIZE * TILE_SIZE * 4);
-      }
-   }
-   else {
-      /* Non-gray color.
-       * Note: if the swizzled tile layout changes (see TILE_PIXEL) this code
-       * will need to change.  It'll be pretty obvious when clearing no longer
-       * works.
-       */
-      const unsigned chunk = TILE_SIZE / 4;
-      for (i = 0; i < rast->state.nr_cbufs; i++) {
-         uint8_t *c = task->color_tiles[i];
-         unsigned j;
-
-         for (j = 0; j < 4 * TILE_SIZE; j++) {
-            memset(c, clear_color[0], chunk);
-            c += chunk;
-            memset(c, clear_color[1], chunk);
-            c += chunk;
-            memset(c, clear_color[2], chunk);
-            c += chunk;
-            memset(c, clear_color[3], chunk);
-            c += chunk;
-         }
-      }
+   for (i = 0; i < scene->fb.nr_cbufs; i++) {
+      const struct lp_scene *scene = task->scene;
+      union util_color uc;
+
+      util_pack_color(arg.clear_color,
+                      scene->fb.cbufs[i]->format, &uc);
+
+      util_fill_rect(scene->cbufs[i].map,
+                     scene->fb.cbufs[i]->format,
+                     scene->cbufs[i].stride,
+                     task->x,
+                     task->y,
+                     TILE_SIZE,
+                     TILE_SIZE,
+                     &uc);
    }
 
    LP_COUNT(nr_color_tile_clear);
 }
 
 
+
+
+
+
 /**
  * Clear the rasterizer's current z/stencil tile.
  * This is a bin command called during bin processing.
  */
-void
+static void
 lp_rast_clear_zstencil(struct lp_rasterizer_task *task,
                        const union lp_rast_cmd_arg arg)
 {
-   struct lp_rasterizer *rast = task->rast;
+   const struct lp_scene *scene = task->scene;
+   uint32_t clear_value = arg.clear_zstencil.value;
+   uint32_t clear_mask = arg.clear_zstencil.mask;
    const unsigned height = TILE_SIZE / TILE_VECTOR_HEIGHT;
    const unsigned width = TILE_SIZE * TILE_VECTOR_HEIGHT;
-   const unsigned block_size = rast->zsbuf.blocksize;
-   const unsigned dst_stride = rast->zsbuf.stride * TILE_VECTOR_HEIGHT;
+   const unsigned block_size = scene->zsbuf.blocksize;
+   const unsigned dst_stride = scene->zsbuf.stride * TILE_VECTOR_HEIGHT;
    uint8_t *dst;
    unsigned i, j;
 
-   LP_DBG(DEBUG_RAST, "%s 0x%x\n", __FUNCTION__, arg.clear_zstencil);
+   LP_DBG(DEBUG_RAST, "%s: value=0x%08x, mask=0x%08x\n",
+           __FUNCTION__, clear_value, clear_mask);
 
    /*
-    * Clear the aera of the swizzled depth/depth buffer matching this tile, in
+    * Clear the area of the swizzled depth/depth buffer matching this tile, in
     * stripes of TILE_VECTOR_HEIGHT x TILE_SIZE at a time.
     *
     * The swizzled depth format is such that the depths for
@@ -287,26 +216,51 @@ lp_rast_clear_zstencil(struct lp_rasterizer_task *task,
 
    dst = task->depth_tile;
 
-   assert(dst == lp_rast_get_depth_block_pointer(rast, task->x, task->y));
+   clear_value &= clear_mask;
 
    switch (block_size) {
    case 1:
-      memset(dst, (uint8_t) arg.clear_zstencil, height * width);
+      assert(clear_mask == 0xff);
+      memset(dst, (uint8_t) clear_value, height * width);
       break;
    case 2:
-      for (i = 0; i < height; i++) {
-         uint16_t *row = (uint16_t *)dst;
-         for (j = 0; j < width; j++)
-            *row++ = (uint16_t) arg.clear_zstencil;
-         dst += dst_stride;
+      if (clear_mask == 0xffff) {
+         for (i = 0; i < height; i++) {
+            uint16_t *row = (uint16_t *)dst;
+            for (j = 0; j < width; j++)
+               *row++ = (uint16_t) clear_value;
+            dst += dst_stride;
+         }
+      }
+      else {
+         for (i = 0; i < height; i++) {
+            uint16_t *row = (uint16_t *)dst;
+            for (j = 0; j < width; j++) {
+               uint16_t tmp = ~clear_mask & *row;
+               *row++ = clear_value | tmp;
+            }
+            dst += dst_stride;
+         }
       }
       break;
    case 4:
-      for (i = 0; i < height; i++) {
-         uint32_t *row = (uint32_t *)dst;
-         for (j = 0; j < width; j++)
-            *row++ = arg.clear_zstencil;
-         dst += dst_stride;
+      if (clear_mask == 0xffffffff) {
+         for (i = 0; i < height; i++) {
+            uint32_t *row = (uint32_t *)dst;
+            for (j = 0; j < width; j++)
+               *row++ = clear_value;
+            dst += dst_stride;
+         }
+      }
+      else {
+         for (i = 0; i < height; i++) {
+            uint32_t *row = (uint32_t *)dst;
+            for (j = 0; j < width; j++) {
+               uint32_t tmp = ~clear_mask & *row;
+               *row++ = clear_value | tmp;
+            }
+            dst += dst_stride;
+         }
       }
       break;
    default:
@@ -316,158 +270,119 @@ lp_rast_clear_zstencil(struct lp_rasterizer_task *task,
 }
 
 
-/**
- * Load tile color from the framebuffer surface.
- * This is a bin command called during bin processing.
- */
-#if 0
-void
-lp_rast_load_color(struct lp_rasterizer_task *task,
-                   const union lp_rast_cmd_arg arg)
-{
-   struct lp_rasterizer *rast = task->rast;
-   unsigned buf;
-   enum lp_texture_usage usage;
-
-   LP_DBG(DEBUG_RAST, "%s at %u, %u\n", __FUNCTION__, x, y);
-
-   if (scene->has_color_clear)
-      usage = LP_TEX_USAGE_WRITE_ALL;
-   else
-      usage = LP_TEX_USAGE_READ_WRITE;
-
-   /* Get pointers to color tile(s).
-    * This will convert linear data to tiled if needed.
-    */
-   for (buf = 0; buf < rast->state.nr_cbufs; buf++) {
-      struct pipe_surface *cbuf = rast->curr_scene->fb.cbufs[buf];
-      struct llvmpipe_texture *lpt;
-      assert(cbuf);
-      lpt = llvmpipe_texture(cbuf->texture);
-      task->color_tiles[buf] = llvmpipe_get_texture_tile(lpt,
-                                                         cbuf->face + cbuf->zslice,
-                                                         cbuf->level,
-                                                         usage,
-                                                         task->x, task->y);
-      assert(task->color_tiles[buf]);
-   }
-}
-#endif
-
-
-/**
- * Convert the color tile from tiled to linear layout.
- * This is generally only done when we're flushing the scene just prior to
- * SwapBuffers.  If we didn't do this here, we'd have to convert the entire
- * tiled color buffer to linear layout in the llvmpipe_texture_unmap()
- * function.  It's better to do it here to take advantage of
- * threading/parallelism.
- * This is a bin command which is stored in all bins.
- */
-void
-lp_rast_store_color( struct lp_rasterizer_task *task,
-                     const union lp_rast_cmd_arg arg)
-{
-   struct lp_rasterizer *rast = task->rast;
-   struct lp_scene *scene = rast->curr_scene;
-   unsigned buf;
-
-   for (buf = 0; buf < rast->state.nr_cbufs; buf++) {
-      struct pipe_surface *cbuf = scene->fb.cbufs[buf];
-      const unsigned face = cbuf->face, level = cbuf->level;
-      struct llvmpipe_resource *lpt = llvmpipe_resource(cbuf->texture);
-      /* this will convert the tiled data to linear if needed */
-      (void) llvmpipe_get_texture_tile_linear(lpt, face, level,
-                                              LP_TEX_USAGE_READ,
-                                              task->x, task->y);
-   }
-}
-
-
-/**
- * This is a bin command called during bin processing.
- */
-void
-lp_rast_set_state(struct lp_rasterizer_task *task,
-                  const union lp_rast_cmd_arg arg)
-{
-   const struct lp_rast_state *state = arg.set_state;
-
-   LP_DBG(DEBUG_RAST, "%s %p\n", __FUNCTION__, (void *) state);
-
-   /* just set the current state pointer for this rasterizer */
-   task->current_state = state;
-}
-
 
 /**
  * Run the shader on all blocks in a tile.  This is used when a tile is
  * completely contained inside a triangle.
  * This is a bin command called during bin processing.
  */
-void
+static void
 lp_rast_shade_tile(struct lp_rasterizer_task *task,
                    const union lp_rast_cmd_arg arg)
 {
-   struct lp_rasterizer *rast = task->rast;
-   const struct lp_rast_state *state = task->current_state;
+   const struct lp_scene *scene = task->scene;
    const struct lp_rast_shader_inputs *inputs = arg.shade_tile;
+   const struct lp_rast_state *state;
+   struct lp_fragment_shader_variant *variant;
    const unsigned tile_x = task->x, tile_y = task->y;
    unsigned x, y;
 
+   if (inputs->disable) {
+      /* This command was partially binned and has been disabled */
+      return;
+   }
+
    LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
 
+   state = task->state;
+   assert(state);
+   if (!state) {
+      return;
+   }
+   variant = state->variant;
+
    /* render the whole 64x64 tile in 4x4 chunks */
    for (y = 0; y < TILE_SIZE; y += 4){
       for (x = 0; x < TILE_SIZE; x += 4) {
          uint8_t *color[PIPE_MAX_COLOR_BUFS];
+         unsigned stride[PIPE_MAX_COLOR_BUFS];
          uint32_t *depth;
          unsigned i;
 
          /* color buffer */
-         for (i = 0; i < rast->state.nr_cbufs; i++)
-            color[i] = lp_rast_get_color_block_pointer(task, i,
-                                                       tile_x + x, tile_y + y);
+         for (i = 0; i < scene->fb.nr_cbufs; i++){
+            stride[i] = scene->cbufs[i].stride;
+
+            color[i] = lp_rast_get_unswizzled_color_block_pointer(task, i, tile_x + x, tile_y + y);
+         }
 
          /* depth buffer */
-         depth = lp_rast_get_depth_block_pointer(rast, tile_x + x, tile_y + y);
+         depth = lp_rast_get_depth_block_pointer(task, tile_x + x, tile_y + y);
 
          /* run shader on 4x4 block */
-         state->jit_function[RAST_WHOLE]( &state->jit_context,
-                                          tile_x + x, tile_y + y,
-                                          inputs->facing,
-                                          inputs->a0,
-                                          inputs->dadx,
-                                          inputs->dady,
-                                          color,
-                                          depth,
-                                          INT_MIN, INT_MIN, INT_MIN,
-                                          NULL, NULL, NULL );
+         BEGIN_JIT_CALL(state, task);
+         variant->jit_function[RAST_WHOLE]( &state->jit_context,
+                                            tile_x + x, tile_y + y,
+                                            inputs->frontfacing,
+                                            GET_A0(inputs),
+                                            GET_DADX(inputs),
+                                            GET_DADY(inputs),
+                                            color,
+                                            depth,
+                                            0xffff,
+                                            &task->vis_counter,
+                                            stride);
+         END_JIT_CALL();
       }
    }
 }
 
 
 /**
- * Compute shading for a 4x4 block of pixels.
+ * Run the shader on all blocks in a tile.  This is used when a tile is
+ * completely contained inside a triangle, and the shader is opaque.
+ * This is a bin command called during bin processing.
+ */
+static void
+lp_rast_shade_tile_opaque(struct lp_rasterizer_task *task,
+                          const union lp_rast_cmd_arg arg)
+{
+   LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
+
+   assert(task->state);
+   if (!task->state) {
+      return;
+   }
+
+   lp_rast_shade_tile(task, arg);
+}
+
+
+/**
+ * Compute shading for a 4x4 block of pixels inside a triangle.
  * This is a bin command called during bin processing.
  * \param x  X position of quad in window coords
  * \param y  Y position of quad in window coords
  */
-void lp_rast_shade_quads( struct lp_rasterizer_task *task,
-                          const struct lp_rast_shader_inputs *inputs,
-                          unsigned x, unsigned y,
-                          int32_t c1, int32_t c2, int32_t c3)
+void
+lp_rast_shade_quads_mask(struct lp_rasterizer_task *task,
+                         const struct lp_rast_shader_inputs *inputs,
+                         unsigned x, unsigned y,
+                         unsigned mask)
 {
-   const struct lp_rast_state *state = task->current_state;
-   struct lp_rasterizer *rast = task->rast;
+   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];
+   unsigned stride[PIPE_MAX_COLOR_BUFS];
    void *depth;
    unsigned i;
 
    assert(state);
 
    /* Sanity checks */
+   assert(x < scene->tiles_x * TILE_SIZE);
+   assert(y < scene->tiles_y * TILE_SIZE);
    assert(x % TILE_VECTOR_WIDTH == 0);
    assert(y % TILE_VECTOR_HEIGHT == 0);
 
@@ -475,86 +390,103 @@ void lp_rast_shade_quads( struct lp_rasterizer_task *task,
    assert((y % 4) == 0);
 
    /* color buffer */
-   for (i = 0; i < rast->state.nr_cbufs; i++) {
-      color[i] = lp_rast_get_color_block_pointer(task, i, x, y);
-      assert(lp_check_alignment(color[i], 16));
+   for (i = 0; i < scene->fb.nr_cbufs; i++) {
+      stride[i] = scene->cbufs[i].stride;
+
+      color[i] = lp_rast_get_unswizzled_color_block_pointer(task, i, x, y);
    }
 
    /* depth buffer */
-   depth = lp_rast_get_depth_block_pointer(rast, x, y);
+   depth = lp_rast_get_depth_block_pointer(task, x, y);
 
 
-   assert(lp_check_alignment(state->jit_context.blend_color, 16));
-
-   assert(lp_check_alignment(inputs->step[0], 16));
-   assert(lp_check_alignment(inputs->step[1], 16));
-   assert(lp_check_alignment(inputs->step[2], 16));
+   assert(lp_check_alignment(state->jit_context.u8_blend_color, 16));
 
    /* run shader on 4x4 block */
-   state->jit_function[RAST_EDGE_TEST]( &state->jit_context,
-                                        x, y,
-                                        inputs->facing,
-                                        inputs->a0,
-                                        inputs->dadx,
-                                        inputs->dady,
-                                        color,
-                                        depth,
-                                        c1, c2, c3,
-                                        inputs->step[0],
-                                        inputs->step[1],
-                                        inputs->step[2]);
+   BEGIN_JIT_CALL(state, task);
+   variant->jit_function[RAST_EDGE_TEST](&state->jit_context,
+                                         x, y,
+                                         inputs->frontfacing,
+                                         GET_A0(inputs),
+                                         GET_DADX(inputs),
+                                         GET_DADY(inputs),
+                                         color,
+                                         depth,
+                                         mask,
+                                         &task->vis_counter,
+                                         stride);
+   END_JIT_CALL();
 }
 
 
+
 /**
- * Set top row and left column of the tile's pixels to white.  For debugging.
+ * Begin a new occlusion query.
+ * This is a bin command put in all bins.
+ * Called per thread.
  */
 static void
-outline_tile(uint8_t *tile)
+lp_rast_begin_query(struct lp_rasterizer_task *task,
+                    const union lp_rast_cmd_arg arg)
 {
-   const uint8_t val = 0xff;
-   unsigned i;
+   struct llvmpipe_query *pq = arg.query_obj;
 
-   for (i = 0; i < TILE_SIZE; i++) {
-      TILE_PIXEL(tile, i, 0, 0) = val;
-      TILE_PIXEL(tile, i, 0, 1) = val;
-      TILE_PIXEL(tile, i, 0, 2) = val;
-      TILE_PIXEL(tile, i, 0, 3) = val;
+   assert(task->query[pq->type] == NULL);
 
-      TILE_PIXEL(tile, 0, i, 0) = val;
-      TILE_PIXEL(tile, 0, i, 1) = val;
-      TILE_PIXEL(tile, 0, i, 2) = val;
-      TILE_PIXEL(tile, 0, i, 3) = val;
+   switch (pq->type) {
+   case PIPE_QUERY_OCCLUSION_COUNTER:
+      task->vis_counter = 0;
+      break;
+   case PIPE_QUERY_TIME_ELAPSED:
+      task->query_start = os_time_get_nano();
+      break;
+   default:
+      assert(0);
+      break;
    }
+
+   task->query[pq->type] = pq;
 }
 
 
 /**
- * Draw grid of gray lines at 16-pixel intervals across the tile to
- * show the sub-tile boundaries.  For debugging.
+ * End the current occlusion query.
+ * This is a bin command put in all bins.
+ * Called per thread.
  */
 static void
-outline_subtiles(uint8_t *tile)
+lp_rast_end_query(struct lp_rasterizer_task *task,
+                  const union lp_rast_cmd_arg arg)
 {
-   const uint8_t val = 0x80;
-   const unsigned step = 16;
-   unsigned i, j;
+   struct llvmpipe_query *pq = arg.query_obj;
+   assert(task->query[pq->type] == pq || pq->type == PIPE_QUERY_TIMESTAMP);
 
-   for (i = 0; i < TILE_SIZE; i += step) {
-      for (j = 0; j < TILE_SIZE; j++) {
-         TILE_PIXEL(tile, i, j, 0) = val;
-         TILE_PIXEL(tile, i, j, 1) = val;
-         TILE_PIXEL(tile, i, j, 2) = val;
-         TILE_PIXEL(tile, i, j, 3) = val;
-
-         TILE_PIXEL(tile, j, i, 0) = val;
-         TILE_PIXEL(tile, j, i, 1) = val;
-         TILE_PIXEL(tile, j, i, 2) = val;
-         TILE_PIXEL(tile, j, i, 3) = val;
-      }
+   switch (pq->type) {
+   case PIPE_QUERY_OCCLUSION_COUNTER:
+      pq->count[task->thread_index] += task->vis_counter;
+      break;
+   case PIPE_QUERY_TIME_ELAPSED:
+      pq->count[task->thread_index] = os_time_get_nano() - task->query_start;
+      break;
+   case PIPE_QUERY_TIMESTAMP:
+      pq->count[task->thread_index] = os_time_get_nano();
+      break;
+   default:
+      assert(0);
+      break;
+   }
+
+   if (task->query[pq->type] == pq) {
+      task->query[pq->type] = NULL;
    }
+}
 
-   outline_tile(tile);
+
+void
+lp_rast_set_state(struct lp_rasterizer_task *task,
+                  const union lp_rast_cmd_arg arg)
+{
+   task->state = arg.state;
 }
 
 
@@ -565,42 +497,60 @@ outline_subtiles(uint8_t *tile)
 static void
 lp_rast_tile_end(struct lp_rasterizer_task *task)
 {
-#if DEBUG
-   struct lp_rasterizer *rast = task->rast;
-   unsigned buf;
-
-   for (buf = 0; buf < rast->state.nr_cbufs; buf++) {
-      uint8_t *color = lp_rast_get_color_block_pointer(task, buf,
-                                                       task->x, task->y);
+   unsigned i;
 
-      if (LP_DEBUG & DEBUG_SHOW_SUBTILES)
-         outline_subtiles(color);
-      else if (LP_DEBUG & DEBUG_SHOW_TILES)
-         outline_tile(color);
+   for (i = 0; i < PIPE_QUERY_TYPES; ++i) {
+      if (task->query[i]) {
+         lp_rast_end_query(task, lp_rast_arg_query(task->query[i]));
+      }
    }
-#else
-   (void) outline_subtiles;
-#endif
 
    /* debug */
    memset(task->color_tiles, 0, sizeof(task->color_tiles));
    task->depth_tile = NULL;
+
+   task->bin = NULL;
 }
 
+static lp_rast_cmd_func dispatch[LP_RAST_OP_MAX] =
+{
+   lp_rast_clear_color,
+   lp_rast_clear_zstencil,
+   lp_rast_triangle_1,
+   lp_rast_triangle_2,
+   lp_rast_triangle_3,
+   lp_rast_triangle_4,
+   lp_rast_triangle_5,
+   lp_rast_triangle_6,
+   lp_rast_triangle_7,
+   lp_rast_triangle_8,
+   lp_rast_triangle_3_4,
+   lp_rast_triangle_3_16,
+   lp_rast_triangle_4_16,
+   lp_rast_shade_tile,
+   lp_rast_shade_tile_opaque,
+   lp_rast_begin_query,
+   lp_rast_end_query,
+   lp_rast_set_state,
+};
 
 
-/**
- * Signal on a fence.  This is called during bin execution/rasterization.
- * Called per thread.
- */
-void
-lp_rast_fence(struct lp_rasterizer_task *task,
-              const union lp_rast_cmd_arg arg)
+static void
+do_rasterize_bin(struct lp_rasterizer_task *task,
+                 const struct cmd_bin *bin)
 {
-   struct lp_fence *fence = arg.fence;
-   lp_fence_signal(fence);
-}
+   const struct cmd_block *block;
+   unsigned k;
 
+   if (0)
+      lp_debug_bin(bin);
+
+   for (block = bin->head; block; block = block->next) {
+      for (k = 0; k < block->count; k++) {
+         dispatch[block->cmd[k]]( task, block->arg[k] );
+      }
+   }
+}
 
 
 
@@ -612,66 +562,26 @@ lp_rast_fence(struct lp_rasterizer_task *task,
  */
 static void
 rasterize_bin(struct lp_rasterizer_task *task,
-              const struct cmd_bin *bin,
-              int x, int y)
+              const struct cmd_bin *bin )
 {
-   const struct cmd_block_list *commands = &bin->commands;
-   struct cmd_block *block;
-   unsigned k;
-
-   lp_rast_tile_begin( task, x * TILE_SIZE, y * TILE_SIZE );
+   lp_rast_tile_begin( task, bin );
 
-   /* simply execute each of the commands in the block list */
-   for (block = commands->head; block; block = block->next) {
-      for (k = 0; k < block->count; k++) {
-         block->cmd[k]( task, block->arg[k] );
-      }
-   }
+   do_rasterize_bin(task, bin);
 
    lp_rast_tile_end(task);
 
-   /* Free data for this bin.
-    */
-   lp_scene_bin_reset( task->rast->curr_scene, x, y);
-}
 
-
-#define RAST(x) { lp_rast_##x, #x }
-
-static struct {
-   lp_rast_cmd cmd;
-   const char *name;
-} cmd_names[] = 
-{
-   RAST(clear_color),
-   RAST(clear_zstencil),
-   RAST(triangle),
-   RAST(shade_tile),
-   RAST(set_state),
-   RAST(store_color),
-   RAST(fence),
-};
-
-static void
-debug_bin( const struct cmd_bin *bin )
-{
-   const struct cmd_block *head = bin->commands.head;
-   int i, j;
-
-   for (i = 0; i < head->count; i++) {
-      debug_printf("%d: ", i);
-      for (j = 0; j < Elements(cmd_names); j++) {
-         if (head->cmd[i] == cmd_names[j].cmd) {
-            debug_printf("%s\n", cmd_names[j].name);
-            break;
-         }
-      }
-      if (j == Elements(cmd_names))
-         debug_printf("...other\n");
+   /* Debug/Perf flags:
+    */
+   if (bin->head->count == 1) {
+      if (bin->head->cmd[0] == LP_RAST_OP_SHADE_TILE_OPAQUE)
+         LP_COUNT(nr_pure_shade_opaque_64);
+      else if (bin->head->cmd[0] == LP_RAST_OP_SHADE_TILE)
+         LP_COUNT(nr_pure_shade_64);
    }
-
 }
 
+
 /* An empty bin is one that just loads the contents of the tile and
  * stores them again unchanged.  This typically happens when bins have
  * been flushed for some reason in the middle of a frame, or when
@@ -682,34 +592,10 @@ debug_bin( const struct cmd_bin *bin )
 static boolean
 is_empty_bin( const struct cmd_bin *bin )
 {
-   const struct cmd_block *head = bin->commands.head;
-   int i;
-   
-   if (0)
-      debug_bin(bin);
-   
-   /* We emit at most two load-tile commands at the start of the first
-    * command block.  In addition we seem to emit a couple of
-    * set-state commands even in empty bins.
-    *
-    * As a heuristic, if a bin has more than 4 commands, consider it
-    * non-empty.
-    */
-   if (head->next != NULL ||
-       head->count > 4) {
-      return FALSE;
-   }
-
-   for (i = 0; i < head->count; i++)
-      if (head->cmd[i] != lp_rast_set_state) {
-         return FALSE;
-      }
-
-   return TRUE;
+   return bin->head == NULL;
 }
 
 
-
 /**
  * Rasterize/execute all bins within a scene.
  * Called per thread.
@@ -718,29 +604,39 @@ static void
 rasterize_scene(struct lp_rasterizer_task *task,
                 struct lp_scene *scene)
 {
-   /* loop over scene bins, rasterize each */
+   task->scene = scene;
+
+   if (!task->rast->no_rast) {
+      /* loop over scene bins, rasterize each */
 #if 0
-   {
-      unsigned i, j;
-      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);
-            rasterize_bin(task, bin, i, j);
+      {
+         unsigned i, j;
+         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);
+               rasterize_bin(task, bin, i, j);
+            }
          }
       }
-   }
 #else
-   {
-      struct cmd_bin *bin;
-      int x, y;
+      {
+         struct cmd_bin *bin;
 
-      assert(scene);
-      while ((bin = lp_scene_bin_iter_next(scene, &x, &y))) {
-         if (!is_empty_bin( bin ))
-            rasterize_bin(task, bin, x, y);
+         assert(scene);
+         while ((bin = lp_scene_bin_iter_next(scene))) {
+            if (!is_empty_bin( bin ))
+               rasterize_bin(task, bin);
+         }
       }
-   }
 #endif
+   }
+
+
+   if (scene->fence) {
+      lp_fence_signal(scene->fence);
+   }
+
+   task->scene = NULL;
 }
 
 
@@ -760,8 +656,6 @@ lp_rast_queue_scene( struct lp_rasterizer *rast,
 
       rasterize_scene( &rast->tasks[0], scene );
 
-      lp_scene_reset( scene );
-
       lp_rast_end( rast );
 
       rast->curr_scene = NULL;
@@ -806,7 +700,7 @@ lp_rast_finish( struct lp_rasterizer *rast )
  *   2. do work
  *   3. signal that we're done
  */
-static PIPE_THREAD_ROUTINE( thread_func, init_data )
+static PIPE_THREAD_ROUTINE( thread_function, init_data )
 {
    struct lp_rasterizer_task *task = (struct lp_rasterizer_task *) init_data;
    struct lp_rasterizer *rast = task->rast;
@@ -874,7 +768,7 @@ create_rast_threads(struct lp_rasterizer *rast)
    for (i = 0; i < rast->num_threads; i++) {
       pipe_semaphore_init(&rast->tasks[i].work_ready, 0);
       pipe_semaphore_init(&rast->tasks[i].work_done, 0);
-      rast->threads[i] = pipe_thread_create(thread_func,
+      rast->threads[i] = pipe_thread_create(thread_function,
                                             (void *) &rast->tasks[i]);
    }
 }
@@ -893,10 +787,14 @@ lp_rast_create( unsigned num_threads )
    unsigned i;
 
    rast = CALLOC_STRUCT(lp_rasterizer);
-   if(!rast)
-      return NULL;
+   if (!rast) {
+      goto no_rast;
+   }
 
    rast->full_scenes = lp_scene_queue_create();
+   if (!rast->full_scenes) {
+      goto no_full_scenes;
+   }
 
    for (i = 0; i < Elements(rast->tasks); i++) {
       struct lp_rasterizer_task *task = &rast->tasks[i];
@@ -906,12 +804,23 @@ lp_rast_create( unsigned num_threads )
 
    rast->num_threads = num_threads;
 
+   rast->no_rast = debug_get_bool_option("LP_NO_RAST", FALSE);
+
    create_rast_threads(rast);
 
    /* for synchronizing rasterization threads */
    pipe_barrier_init( &rast->barrier, rast->num_threads );
 
+   memset(lp_swizzled_cbuf, 0, sizeof lp_swizzled_cbuf);
+
+   memset(lp_dummy_tile, 0, sizeof lp_dummy_tile);
+
    return rast;
+
+no_full_scenes:
+   FREE(rast);
+no_rast:
+   return NULL;
 }
 
 
@@ -956,3 +865,5 @@ lp_rast_get_num_threads( struct lp_rasterizer *rast )
 {
    return rast->num_threads;
 }
+
+