llvmpipe: repartition lp_rasterizer state for threading

Some of the state is per-thread.  Put that state in new lp_rasterizer_task
struct.

5 files changed, 133 insertions, 72 deletions

diff --git a/src/gallium/drivers/llvmpipe/lp_bin.h b/src/gallium/drivers/llvmpipe/lp_bin.h
index 24e599ea66e..b07ff64e623 100644
--- a/src/gallium/drivers/llvmpipe/lp_bin.h
+++ b/src/gallium/drivers/llvmpipe/lp_bin.h
@@ -56,7 +56,9 @@
 
 /* switch to a non-pointer value for this:
  */
-typedef void (*lp_rast_cmd)( struct lp_rasterizer *, const union lp_rast_cmd_arg );
+typedef void (*lp_rast_cmd)( struct lp_rasterizer *,
+                             unsigned thread_index,
+                             const union lp_rast_cmd_arg );
 
 struct cmd_block {
    lp_rast_cmd cmd[CMD_BLOCK_MAX];
diff --git a/src/gallium/drivers/llvmpipe/lp_rast.c b/src/gallium/drivers/llvmpipe/lp_rast.c
index a6192e589dc..37cc28e938e 100644
--- a/src/gallium/drivers/llvmpipe/lp_rast.c
+++ b/src/gallium/drivers/llvmpipe/lp_rast.c
@@ -39,14 +39,18 @@
 struct lp_rasterizer *lp_rast_create( struct pipe_screen *screen )
 {
    struct lp_rasterizer *rast;
+   unsigned i;
 
    rast = CALLOC_STRUCT(lp_rasterizer);
    if(!rast)
       return NULL;
 
    rast->screen = screen;
-   rast->tile.color = align_malloc( TILE_SIZE*TILE_SIZE*4, 16 );
-   rast->tile.depth = align_malloc( TILE_SIZE*TILE_SIZE*4, 16 );
+
+   for (i = 0; i < Elements(rast->tasks); i++) {
+      rast->tasks[i].tile.color = align_malloc( TILE_SIZE*TILE_SIZE*4, 16 );
+      rast->tasks[i].tile.depth = align_malloc( TILE_SIZE*TILE_SIZE*4, 16 );
+   }
 
    return rast;
 }
@@ -153,12 +157,13 @@ lp_rast_end( struct lp_rasterizer *rast )
  */
 static void
 lp_rast_start_tile( struct lp_rasterizer *rast,
+                    unsigned thread_index,
                     unsigned x, unsigned y )
 {
    LP_DBG(DEBUG_RAST, "%s %d,%d\n", __FUNCTION__, x, y);
 
-   rast->x = x;
-   rast->y = y;
+   rast->tasks[thread_index].x = x;
+   rast->tasks[thread_index].y = y;
 }
 
 
@@ -167,9 +172,11 @@ lp_rast_start_tile( struct lp_rasterizer *rast,
  * This is a bin command called during bin processing.
  */
 void lp_rast_clear_color( struct lp_rasterizer *rast,
+                          unsigned thread_index,
                           const union lp_rast_cmd_arg arg )
 {
    const uint8_t *clear_color = arg.clear_color;
+   uint8_t *color_tile = rast->tasks[thread_index].tile.color;
    
    LP_DBG(DEBUG_RAST, "%s 0x%x,0x%x,0x%x,0x%x\n", __FUNCTION__, 
               clear_color[0],
@@ -180,14 +187,14 @@ void lp_rast_clear_color( struct lp_rasterizer *rast,
    if (clear_color[0] == clear_color[1] &&
        clear_color[1] == clear_color[2] &&
        clear_color[2] == clear_color[3]) {
-      memset(rast->tile.color, clear_color[0], TILE_SIZE * TILE_SIZE * 4);
+      memset(color_tile, clear_color[0], TILE_SIZE * TILE_SIZE * 4);
    }
    else {
       unsigned x, y, chan;
       for (y = 0; y < TILE_SIZE; y++)
          for (x = 0; x < TILE_SIZE; x++)
             for (chan = 0; chan < 4; ++chan)
-               TILE_PIXEL(rast->tile.color, x, y, chan) = clear_color[chan];
+               TILE_PIXEL(color_tile, x, y, chan) = clear_color[chan];
    }
 }
 
@@ -197,15 +204,17 @@ void lp_rast_clear_color( struct lp_rasterizer *rast,
  * This is a bin command called during bin processing.
  */
 void lp_rast_clear_zstencil( struct lp_rasterizer *rast,
+                             unsigned thread_index,
                              const union lp_rast_cmd_arg arg)
 {
    unsigned i, j;
+   uint32_t *depth_tile = rast->tasks[thread_index].tile.depth;
    
    LP_DBG(DEBUG_RAST, "%s 0x%x\n", __FUNCTION__, arg.clear_zstencil);
 
    for (i = 0; i < TILE_SIZE; i++)
       for (j = 0; j < TILE_SIZE; j++)
-	 rast->tile.depth[i*TILE_SIZE + j] = arg.clear_zstencil;
+	 depth_tile[i*TILE_SIZE + j] = arg.clear_zstencil;
 }
 
 
@@ -214,6 +223,7 @@ void lp_rast_clear_zstencil( struct lp_rasterizer *rast,
  * This is a bin command called during bin processing.
  */
 void lp_rast_load_color( struct lp_rasterizer *rast,
+                         unsigned thread_index,
                          const union lp_rast_cmd_arg arg)
 {
    LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
@@ -227,6 +237,7 @@ void lp_rast_load_color( struct lp_rasterizer *rast,
  * This is a bin command called during bin processing.
  */
 void lp_rast_load_zstencil( struct lp_rasterizer *rast,
+                            unsigned thread_index,
                             const union lp_rast_cmd_arg arg )
 {
    LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
@@ -236,6 +247,7 @@ void lp_rast_load_zstencil( struct lp_rasterizer *rast,
 
 
 void lp_rast_set_state( struct lp_rasterizer *rast,
+                        unsigned thread_index,
                         const union lp_rast_cmd_arg arg )
 {
    const struct lp_rast_state *state = arg.set_state;
@@ -243,7 +255,7 @@ void lp_rast_set_state( struct lp_rasterizer *rast,
    LP_DBG(DEBUG_RAST, "%s %p\n", __FUNCTION__, (void *) state);
 
    /* just set the current state pointer for this rasterizer */
-   rast->current_state = state;
+   rast->tasks[thread_index].current_state = state;
 }
 
 
@@ -257,9 +269,12 @@ void lp_rast_set_state( struct lp_rasterizer *rast,
  * This is a bin command called during bin processing.
  */
 void lp_rast_shade_tile( struct lp_rasterizer *rast,
+                         unsigned thread_index,
                          const union lp_rast_cmd_arg arg )
 {
    const struct lp_rast_shader_inputs *inputs = arg.shade_tile;
+   const unsigned tile_x = rast->tasks[thread_index].x;
+   const unsigned tile_y = rast->tasks[thread_index].y;
    const unsigned mask = ~0;
    unsigned x, y;
 
@@ -269,7 +284,12 @@ void lp_rast_shade_tile( struct lp_rasterizer *rast,
     */
    for (y = 0; y < TILE_SIZE; y += 4)
       for (x = 0; x < TILE_SIZE; x += 4)
-         lp_rast_shade_quads( rast, inputs, rast->x + x, rast->y + y, mask);
+         lp_rast_shade_quads( rast,
+                              thread_index,
+                              inputs,
+                              tile_x + x,
+                              tile_y + y,
+                              mask);
 }
 
 
@@ -278,13 +298,14 @@ void lp_rast_shade_tile( struct lp_rasterizer *rast,
  * This is a bin command called during bin processing.
  */
 void lp_rast_shade_quads( struct lp_rasterizer *rast,
+                          unsigned thread_index,
                           const struct lp_rast_shader_inputs *inputs,
                           unsigned x, unsigned y,
                           unsigned mask)
 {
 #if 1
-   const struct lp_rast_state *state = rast->current_state;
-   struct lp_rast_tile *tile = &rast->tile;
+   const struct lp_rast_state *state = rast->tasks[thread_index].current_state;
+   struct lp_rast_tile *tile = &rast->tasks[thread_index].tile;
    void *color;
    void *depth;
    uint32_t ALIGN16_ATTRIB masks[2][2][2][2];
@@ -388,10 +409,11 @@ void lp_rast_shade_quads( struct lp_rasterizer *rast,
 /**
  * Write the rasterizer's color tile to the framebuffer.
  */
-static void lp_rast_store_color( struct lp_rasterizer *rast )
+static void lp_rast_store_color( struct lp_rasterizer *rast,
+                                 unsigned thread_index)
 {
-   const unsigned x = rast->x;
-   const unsigned y = rast->y;
+   const unsigned x = rast->tasks[thread_index].x;
+   const unsigned y = rast->tasks[thread_index].y;
    unsigned w = TILE_SIZE;
    unsigned h = TILE_SIZE;
 
@@ -404,7 +426,7 @@ static void lp_rast_store_color( struct lp_rasterizer *rast )
    LP_DBG(DEBUG_RAST, "%s %d,%d %dx%d\n", __FUNCTION__, x, y, w, h);
 
    lp_tile_write_4ub(rast->cbuf_transfer->format,
-                     rast->tile.color,
+                     rast->tasks[thread_index].tile.color,
                      rast->cbuf_map, 
                      rast->cbuf_transfer->stride,
                      x, y,
@@ -430,10 +452,11 @@ lp_tile_write_z32(const uint32_t *src, uint8_t *dst, unsigned dst_stride,
 /**
  * Write the rasterizer's z/stencil tile to the framebuffer.
  */
-static void lp_rast_store_zstencil( struct lp_rasterizer *rast )
+static void lp_rast_store_zstencil( struct lp_rasterizer *rast,
+                                    unsigned thread_index )
 {
-   const unsigned x = rast->x;
-   const unsigned y = rast->y;
+   const unsigned x = rast->tasks[thread_index].x;
+   const unsigned y = rast->tasks[thread_index].y;
    unsigned w = TILE_SIZE;
    unsigned h = TILE_SIZE;
 
@@ -446,7 +469,7 @@ static void lp_rast_store_zstencil( struct lp_rasterizer *rast )
    LP_DBG(DEBUG_RAST, "%s %d,%d %dx%d\n", __FUNCTION__, x, y, w, h);
 
    assert(rast->zsbuf_transfer->format == PIPE_FORMAT_Z32_UNORM);
-   lp_tile_write_z32(rast->tile.depth,
+   lp_tile_write_z32(rast->tasks[thread_index].tile.depth,
                      rast->zsbuf_map, 
                      rast->zsbuf_transfer->stride,
                      x, y, w, h);
@@ -457,15 +480,16 @@ static void lp_rast_store_zstencil( struct lp_rasterizer *rast )
  * Write the rasterizer's tiles to the framebuffer.
  */
 static void
-lp_rast_end_tile( struct lp_rasterizer *rast )
+lp_rast_end_tile( struct lp_rasterizer *rast,
+                  unsigned thread_index )
 {
    LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
 
    if (rast->state.write_color)
-      lp_rast_store_color(rast);
+      lp_rast_store_color(rast, thread_index);
 
    if (rast->state.write_zstencil)
-      lp_rast_store_zstencil(rast);
+      lp_rast_store_zstencil(rast, thread_index);
 }
 
 
@@ -476,6 +500,7 @@ lp_rast_end_tile( struct lp_rasterizer *rast )
  */
 static void
 rasterize_bin( struct lp_rasterizer *rast,
+               unsigned thread_index,
                const struct cmd_bin *bin,
                int x, int y)
 {
@@ -483,16 +508,16 @@ rasterize_bin( struct lp_rasterizer *rast,
    struct cmd_block *block;
    unsigned k;
 
-   lp_rast_start_tile( rast, x, y );
+   lp_rast_start_tile( rast, thread_index, x, y );
 
    /* 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]( rast, block->arg[k] );
+         block->cmd[k]( rast, 0, block->arg[k] );
       }
    }
 
-   lp_rast_end_tile( rast );
+   lp_rast_end_tile( rast, thread_index );
 }
 
 
@@ -522,7 +547,7 @@ lp_rasterize_bins( struct lp_rasterizer *rast,
       for (i = 0; i < bins->tiles_x; i++) {
          for (j = 0; j < bins->tiles_y; j++) {
             struct cmd_bin *bin = lp_get_bin(bins, i, j);
-            rasterize_bin( rast, bin, i * TILE_SIZE, j * TILE_SIZE );
+            rasterize_bin( rast, 0, bin, i * TILE_SIZE, j * TILE_SIZE );
          }
       }
    }
@@ -534,7 +559,7 @@ lp_rasterize_bins( struct lp_rasterizer *rast,
       lp_bin_iter_begin( bins );
 
       while ((bin = lp_bin_iter_next(bins, &x, &y))) {
-         rasterize_bin( rast, bin, x * TILE_SIZE, y * TILE_SIZE);
+         rasterize_bin( rast, 0, bin, x * TILE_SIZE, y * TILE_SIZE);
       }
    }
 #endif
@@ -550,10 +575,16 @@ lp_rasterize_bins( struct lp_rasterizer *rast,
  */
 void lp_rast_destroy( struct lp_rasterizer *rast )
 {
+   unsigned i;
+
    pipe_surface_reference(&rast->state.cbuf, NULL);
    pipe_surface_reference(&rast->state.zsbuf, NULL);
-   align_free(rast->tile.depth);
-   align_free(rast->tile.color);
+
+   for (i = 0; i < Elements(rast->tasks); i++) {
+      align_free(rast->tasks[i].tile.depth);
+      align_free(rast->tasks[i].tile.color);
+   }
+
    FREE(rast);
 }
 
diff --git a/src/gallium/drivers/llvmpipe/lp_rast.h b/src/gallium/drivers/llvmpipe/lp_rast.h
index e77c77b7762..25e7f8e0086 100644
--- a/src/gallium/drivers/llvmpipe/lp_rast.h
+++ b/src/gallium/drivers/llvmpipe/lp_rast.h
@@ -193,24 +193,31 @@ lp_rast_arg_null( void )
  */
 
 void lp_rast_clear_color( struct lp_rasterizer *, 
+                          unsigned thread_index,
                           const union lp_rast_cmd_arg );
 
 void lp_rast_clear_zstencil( struct lp_rasterizer *, 
+                             unsigned thread_index,
                              const union lp_rast_cmd_arg );
 
 void lp_rast_load_color( struct lp_rasterizer *, 
+                         unsigned thread_index,
                          const union lp_rast_cmd_arg );
 
 void lp_rast_load_zstencil( struct lp_rasterizer *, 
+                            unsigned thread_index,
                             const union lp_rast_cmd_arg );
 
 void lp_rast_set_state( struct lp_rasterizer *, 
+                        unsigned thread_index,
                         const union lp_rast_cmd_arg );
 
 void lp_rast_triangle( struct lp_rasterizer *, 
+                       unsigned thread_index,
                        const union lp_rast_cmd_arg );
 
 void lp_rast_shade_tile( struct lp_rasterizer *,
+                         unsigned thread_index,
                          const union lp_rast_cmd_arg );
 
 
diff --git a/src/gallium/drivers/llvmpipe/lp_rast_priv.h b/src/gallium/drivers/llvmpipe/lp_rast_priv.h
index 98111edff72..9e7cbd7912e 100644
--- a/src/gallium/drivers/llvmpipe/lp_rast_priv.h
+++ b/src/gallium/drivers/llvmpipe/lp_rast_priv.h
@@ -30,6 +30,10 @@
 
 #include "lp_rast.h"
 
+
+#define MAX_THREADS 4  /* XXX probably temporary here */
+
+
 struct pipe_transfer;
 struct pipe_screen;
 
@@ -47,14 +51,34 @@ struct lp_rast_tile
 
 
 /**
- * This is the state required while rasterizing a tile.
- * The tile size is TILE_SIZE x TILE_SIZE pixels.
+ * Per-thread rasterization state
  */
-struct lp_rasterizer
+struct lp_rasterizer_task
 {
    struct lp_rast_tile tile;   /** Tile color/z/stencil memory */
 
    unsigned x, y;          /**< Pos of this tile in framebuffer, in pixels */
+
+   /* Pixel blocks produced during rasterization
+    */
+   unsigned nr_blocks;
+   struct {
+      unsigned x;
+      unsigned y;
+      unsigned mask;
+   } blocks[256];
+
+   const struct lp_rast_state *current_state;
+};
+
+
+/**
+ * This is the state required while rasterizing tiles.
+ * Note that this contains per-thread information too.
+ * The tile size is TILE_SIZE x TILE_SIZE pixels.
+ */
+struct lp_rasterizer
+{
    unsigned width, height; /**< Size of framebuffer, in pixels */
 
    boolean clipped_tile;
@@ -78,20 +102,13 @@ struct lp_rasterizer
       char clear_stencil;
    } state;
 
-   /* Pixel blocks produced during rasterization
-    */
-   unsigned nr_blocks;
-   struct {
-      unsigned x;
-      unsigned y;
-      unsigned mask;
-   } blocks[256];
-
-   const struct lp_rast_state *current_state;
+   /** A task object for each rasterization thread */
+   struct lp_rasterizer_task tasks[MAX_THREADS];
 };
 
 
 void lp_rast_shade_quads( struct lp_rasterizer *rast,
+                          unsigned thread_index,
                           const struct lp_rast_shader_inputs *inputs,
                           unsigned x, unsigned y,
                           unsigned masks);
diff --git a/src/gallium/drivers/llvmpipe/lp_rast_tri.c b/src/gallium/drivers/llvmpipe/lp_rast_tri.c
index 81a9c1c1422..6c96010c52f 100644
--- a/src/gallium/drivers/llvmpipe/lp_rast_tri.c
+++ b/src/gallium/drivers/llvmpipe/lp_rast_tri.c
@@ -40,15 +40,15 @@
  * All pixels are known to be inside the triangle's bounds.
  */
 static void
-block_full_4( struct lp_rasterizer *rast, int x, int y )
+block_full_4( struct lp_rasterizer_task *rast_task, int x, int y )
 {
-   const unsigned i = rast->nr_blocks;
+   const unsigned i = rast_task->nr_blocks;
    assert(x % 4 == 0);
    assert(y % 4 == 0);
-   rast->blocks[i].x = x;
-   rast->blocks[i].y = y;
-   rast->blocks[i].mask = ~0;
-   rast->nr_blocks++;
+   rast_task->blocks[i].x = x;
+   rast_task->blocks[i].y = y;
+   rast_task->blocks[i].mask = ~0;
+   rast_task->nr_blocks++;
 }
 
 
@@ -57,14 +57,14 @@ block_full_4( struct lp_rasterizer *rast, int x, int y )
  * All pixels are known to be inside the triangle's bounds.
  */
 static void
-block_full_16( struct lp_rasterizer *rast, int x, int y )
+block_full_16( struct lp_rasterizer_task *rast_task, int x, int y )
 {
    unsigned ix, iy;
    assert(x % 16 == 0);
    assert(y % 16 == 0);
    for (iy = 0; iy < 16; iy += 4)
       for (ix = 0; ix < 16; ix += 4)
-	 block_full_4(rast, x + ix, y + iy);
+	 block_full_4(rast_task, x + ix, y + iy);
 }
 
 
@@ -74,7 +74,7 @@ block_full_16( struct lp_rasterizer *rast, int x, int y )
  * Generate a mask of in/out flags and add the block to the blocks list.
  */
 static void
-do_block_4( struct lp_rasterizer *rast,
+do_block_4( struct lp_rasterizer_task *rast_task,
 	    const struct lp_rast_triangle *tri,
 	    int x, int y,
 	    int c1,
@@ -97,11 +97,11 @@ do_block_4( struct lp_rasterizer *rast,
    /* As we do trivial reject already, masks should rarely be all zero:
     */
    if (mask) {
-      const unsigned i = rast->nr_blocks;
-      rast->blocks[i].x = x;
-      rast->blocks[i].y = y;
-      rast->blocks[i].mask = mask;
-      rast->nr_blocks++;
+      const unsigned i = rast_task->nr_blocks;
+      rast_task->blocks[i].x = x;
+      rast_task->blocks[i].y = y;
+      rast_task->blocks[i].mask = mask;
+      rast_task->nr_blocks++;
    }
 }
 
@@ -111,7 +111,7 @@ do_block_4( struct lp_rasterizer *rast,
  * of the triangle's bounds.
  */
 static void
-do_block_16( struct lp_rasterizer *rast,
+do_block_16( struct lp_rasterizer_task *rast_task,
              const struct lp_rast_triangle *tri,
              int x, int y,
              int c1,
@@ -146,11 +146,11 @@ do_block_16( struct lp_rasterizer *rast,
 		  cx2 + ei2 > 0 &&
 		  cx3 + ei3 > 0) {
             /* the block is completely inside the triangle */
-	    block_full_4(rast, x+ix, y+iy);
+	    block_full_4(rast_task, x+ix, y+iy);
 	 }
 	 else {
             /* the block is partially in/out of the triangle */
-	    do_block_4(rast, tri, x+ix, y+iy, cx1, cx2, cx3);
+	    do_block_4(rast_task, tri, x+ix, y+iy, cx1, cx2, cx3);
 	 }
       }
    }
@@ -163,12 +163,14 @@ do_block_16( struct lp_rasterizer *rast,
  */
 void
 lp_rast_triangle( struct lp_rasterizer *rast,
+                  unsigned thread_index,
                   const union lp_rast_cmd_arg arg )
 {
+   struct lp_rasterizer_task *rast_task = &rast->tasks[thread_index];
    const struct lp_rast_triangle *tri = arg.triangle;
 
-   int x = rast->x;
-   int y = rast->y;
+   int x = rast_task->x;
+   int y = rast_task->y;
    int ix, iy;
    unsigned i = 0;
 
@@ -184,11 +186,11 @@ lp_rast_triangle( struct lp_rasterizer *rast,
    int eo2 = tri->eo2 * 16;
    int eo3 = tri->eo3 * 16;
 
-   assert(Elements(rast->blocks) == (TILE_SIZE * TILE_SIZE) / (4*4));
+   assert(Elements(rast_task->blocks) == (TILE_SIZE * TILE_SIZE) / (4*4));
 
    LP_DBG(DEBUG_RAST, "lp_rast_triangle\n");
 
-   rast->nr_blocks = 0;
+   rast_task->nr_blocks = 0;
 
    /* Walk over the tile to build a list of 4x4 pixel blocks which will
     * be filled/shaded.  We do this at two granularities: 16x16 blocks
@@ -209,21 +211,23 @@ lp_rast_triangle( struct lp_rasterizer *rast,
 		  cx2 + ei2 > 0 &&
 		  cx3 + ei3 > 0) {
             /* the block is completely inside the triangle */
-	    block_full_16(rast, x+ix, y+iy);
+	    block_full_16(rast_task, x+ix, y+iy);
 	 }
 	 else {
             /* the block is partially in/out of the triangle */
-	    do_block_16(rast, tri, x+ix, y+iy, cx1, cx2, cx3);
+	    do_block_16(rast_task, tri, x+ix, y+iy, cx1, cx2, cx3);
 	 }
       }
    }
 
-   assert(rast->nr_blocks <= Elements(rast->blocks));
+   assert(rast_task->nr_blocks <= Elements(rast_task->blocks));
 
    /* Shade the 4x4 pixel blocks */
-   for (i = 0; i < rast->nr_blocks; i++) 
-      lp_rast_shade_quads(rast, &tri->inputs, 
-			  rast->blocks[i].x,
-			  rast->blocks[i].y,
-			  rast->blocks[i].mask);
+   for (i = 0; i < rast_task->nr_blocks; i++) 
+      lp_rast_shade_quads(rast,
+                          thread_index,
+                          &tri->inputs, 
+			  rast_task->blocks[i].x,
+			  rast_task->blocks[i].y,
+			  rast_task->blocks[i].mask);
 }