Cell: store current tile status in cur_tile_status_c/z, add TILE_STATUS_GETTING

5 files changed, 87 insertions, 22 deletions

diff --git a/src/mesa/pipe/cell/spu/spu_render.c b/src/mesa/pipe/cell/spu/spu_render.c
index f506095116b..ca54a103bdc 100644
--- a/src/mesa/pipe/cell/spu/spu_render.c
+++ b/src/mesa/pipe/cell/spu/spu_render.c
@@ -95,13 +95,15 @@ static INLINE void
 get_cz_tiles(uint tx, uint ty)
 {
    if (spu.depth_stencil.depth.enabled) {
-      if (tile_status_z[ty][tx] != TILE_STATUS_CLEAR) {
+      if (cur_tile_status_z != TILE_STATUS_CLEAR) {
          get_tile(tx, ty, &ztile, TAG_READ_TILE_Z, 1);
+         cur_tile_status_z = TILE_STATUS_GETTING;
       }
    }
 
-   if (tile_status[ty][tx] != TILE_STATUS_CLEAR) {
+   if (cur_tile_status_c != TILE_STATUS_CLEAR) {
       get_tile(tx, ty, &ctile, TAG_READ_TILE_COLOR, 0);
+      cur_tile_status_c = TILE_STATUS_GETTING;
    }
 }
 
@@ -112,14 +114,24 @@ get_cz_tiles(uint tx, uint ty)
 static INLINE void
 put_cz_tiles(uint tx, uint ty)
 {
-   if (tile_status_z[ty][tx] == TILE_STATUS_DIRTY) {
+   if (cur_tile_status_z == TILE_STATUS_DIRTY) {
+      /* tile was modified and needs to be written back */
       put_tile(tx, ty, &ztile, TAG_WRITE_TILE_Z, 1);
-      tile_status_z[ty][tx] = TILE_STATUS_DEFINED;
+      cur_tile_status_z = TILE_STATUS_DEFINED;
+   }
+   else if (cur_tile_status_z == TILE_STATUS_GETTING) {
+      /* tile was never used */
+      cur_tile_status_z = TILE_STATUS_DEFINED;
    }
 
-   if (tile_status[ty][tx] == TILE_STATUS_DIRTY) {
+   if (cur_tile_status_c == TILE_STATUS_DIRTY) {
+      /* tile was modified and needs to be written back */
       put_tile(tx, ty, &ctile, TAG_WRITE_TILE_COLOR, 0);
-      tile_status[ty][tx] = TILE_STATUS_DEFINED;
+      cur_tile_status_c = TILE_STATUS_DEFINED;
+   }
+   else if (cur_tile_status_c == TILE_STATUS_GETTING) {
+      /* tile was never used */
+      cur_tile_status_c = TILE_STATUS_DEFINED;
    }
 }
 
@@ -238,8 +250,13 @@ cmd_render(const struct cell_command_render *render, uint *pos_incr)
       if (!my_tile(tx, ty))
          continue;
 
+      cur_tile_status_c = tile_status[ty][tx];
+      cur_tile_status_z = tile_status_z[ty][tx];
+
       get_cz_tiles(tx, ty);
 
+      uint drawn = 0;
+
       /* loop over tris */
       for (j = 0; j < render->num_indexes; j += 3) {
          const float *v0, *v1, *v2;
@@ -248,13 +265,18 @@ cmd_render(const struct cell_command_render *render, uint *pos_incr)
          v1 = (const float *) (vertices + indexes[j+1] * vertex_size);
          v2 = (const float *) (vertices + indexes[j+2] * vertex_size);
 
-         tri_draw(v0, v1, v2, tx, ty);
+         drawn += tri_draw(v0, v1, v2, tx, ty);
       }
 
+      //printf("SPU %u: drew %u of %u\n", spu.init.id, drawn, render->num_indexes/3);
+
       /* write color/z tiles back to main framebuffer, if dirtied */
       put_cz_tiles(tx, ty);
 
       wait_put_cz_tiles(); /* XXX seems unnecessary... */
+
+      tile_status[ty][tx] = cur_tile_status_c;
+      tile_status_z[ty][tx] = cur_tile_status_z;
    }
 
    if (Debug)
diff --git a/src/mesa/pipe/cell/spu/spu_tile.c b/src/mesa/pipe/cell/spu/spu_tile.c
index ca1352f9f8d..aea4785bc2f 100644
--- a/src/mesa/pipe/cell/spu/spu_tile.c
+++ b/src/mesa/pipe/cell/spu/spu_tile.c
@@ -37,6 +37,7 @@ tile_t ztile ALIGN16_ATTRIB;
 ubyte tile_status[MAX_HEIGHT/TILE_SIZE][MAX_WIDTH/TILE_SIZE] ALIGN16_ATTRIB;
 ubyte tile_status_z[MAX_HEIGHT/TILE_SIZE][MAX_WIDTH/TILE_SIZE] ALIGN16_ATTRIB;
 
+ubyte cur_tile_status_c, cur_tile_status_z;
 
 
 void
diff --git a/src/mesa/pipe/cell/spu/spu_tile.h b/src/mesa/pipe/cell/spu/spu_tile.h
index 18d1b3c1173..1f123a2b7b4 100644
--- a/src/mesa/pipe/cell/spu/spu_tile.h
+++ b/src/mesa/pipe/cell/spu/spu_tile.h
@@ -51,12 +51,16 @@ extern tile_t ztile ALIGN16_ATTRIB;
 
 
 #define TILE_STATUS_CLEAR   1
-#define TILE_STATUS_DEFINED 2  /**< defined pixel data */
-#define TILE_STATUS_DIRTY   3  /**< modified, but not put back yet */
+#define TILE_STATUS_DEFINED 2  /**< defined in FB, but not in local store */
+#define TILE_STATUS_CLEAN   3  /**< in local store, but not changed */
+#define TILE_STATUS_DIRTY   4  /**< modified locally, but not put back yet */
+#define TILE_STATUS_GETTING 5  /**< mfc_get() called but not yet arrived */
 
 extern ubyte tile_status[MAX_HEIGHT/TILE_SIZE][MAX_WIDTH/TILE_SIZE] ALIGN16_ATTRIB;
 extern ubyte tile_status_z[MAX_HEIGHT/TILE_SIZE][MAX_WIDTH/TILE_SIZE] ALIGN16_ATTRIB;
 
+extern ubyte cur_tile_status_c, cur_tile_status_z;
+
 
 void
 get_tile(uint tx, uint ty, tile_t *tile, int tag, int zBuf);
diff --git a/src/mesa/pipe/cell/spu/spu_tri.c b/src/mesa/pipe/cell/spu/spu_tri.c
index 08b8bf0c9c7..a32878d9178 100644
--- a/src/mesa/pipe/cell/spu/spu_tri.c
+++ b/src/mesa/pipe/cell/spu/spu_tri.c
@@ -299,16 +299,23 @@ do_depth_test(int x, int y, unsigned int mask)
 
    zvals.v = eval_z((float) x, (float) y);
 
-   if (tile_status_z[setup.ty][setup.tx] == TILE_STATUS_CLEAR) {
+   if (cur_tile_status_c == TILE_STATUS_CLEAR) {
       /* now, _really_ clear the tile */
       clear_z_tile(&ztile);
+      cur_tile_status_z = TILE_STATUS_DIRTY;
    }
-   else if (tile_status_z[setup.ty][setup.tx] != TILE_STATUS_DIRTY) {
+
+#if 0
+   if (cur_tile_status_z == TILE_STATUS_CLEAR) {
+      /* now, _really_ clear the tile */
+      clear_z_tile(&ztile);
+   }
+   else if (cur_tile_status_z != TILE_STATUS_DIRTY) {
       /* make sure we've got the tile from main mem */
       wait_on_mask(1 << TAG_READ_TILE_Z);
    }
-   tile_status_z[setup.ty][setup.tx] = TILE_STATUS_DIRTY;
-
+   cur_tile_status_z = TILE_STATUS_DIRTY;
+#endif
 
    if (spu.fb.depth_format == PIPE_FORMAT_Z16_UNORM) {
       zvals.v = spu_mul(zvals.v, zscale16.v);
@@ -380,6 +387,9 @@ do_depth_test(int x, int y, unsigned int mask)
       }
    }
 
+   if (mask)
+      cur_tile_status_z = TILE_STATUS_DIRTY;
+
    return mask;
 }
 
@@ -397,15 +407,15 @@ do_depth_test_simd(int x, int y, vector unsigned int quadmask)
 
    zvals.v = eval_z((float) x, (float) y);
 
-   if (tile_status_z[setup.ty][setup.tx] == TILE_STATUS_CLEAR) {
+   if (cur_tile_status_z == TILE_STATUS_CLEAR) {
       /* now, _really_ clear the tile */
       clear_z_tile(&ztile);
    }
-   else if (tile_status_z[setup.ty][setup.tx] != TILE_STATUS_DIRTY) {
+   else if (cur_tile_status_z != TILE_STATUS_DIRTY) {
       /* make sure we've got the tile from main mem */
       wait_on_mask(1 << TAG_READ_TILE_Z);
    }
-   tile_status_z[setup.ty][setup.tx] = TILE_STATUS_DIRTY;
+   cur_tile_status_z = TILE_STATUS_DIRTY;
 
    /* XXX fetch Z value sooner to hide latency here */
    zmask = spu_cmpgt(ztile.f4[ix][iy].v, zvals.v);
@@ -462,15 +472,23 @@ emit_quad( int x, int y, mask_t mask )
    if (mask)
 #endif
    {
-      if (tile_status[setup.ty][setup.tx] == TILE_STATUS_CLEAR) {
+      if (cur_tile_status_c == TILE_STATUS_CLEAR) {
          /* now, _really_ clear the tile */
          clear_c_tile(&ctile);
       }
-      else if (tile_status[setup.ty][setup.tx] != TILE_STATUS_DIRTY) {
+
+#if 0
+      if (cur_tile_status_c == TILE_STATUS_CLEAR) {
+         /* now, _really_ clear the tile */
+         clear_c_tile(&ctile);
+         cur_tile_status_c = TILE_STATUS_DIRTY;
+      }
+      else if (cur_tile_status_c != TILE_STATUS_DIRTY) {
          /* make sure we've got the tile from main mem */
          wait_on_mask(1 << TAG_READ_TILE_COLOR);
       }
-      tile_status[setup.ty][setup.tx] = TILE_STATUS_DIRTY;
+#endif
+      cur_tile_status_c = TILE_STATUS_DIRTY;
 
 #if SIMD_Z
       if (spu_extract(mask, 0))
@@ -970,7 +988,7 @@ static void subtriangle( struct edge *eleft,
  * Draw triangle into tile at (tx, ty) (tile coords)
  * The tile data should have already been fetched.
  */
-void
+boolean
 tri_draw(const float *v0, const float *v1, const float *v2, uint tx, uint ty)
 {
    setup.tx = tx;
@@ -985,7 +1003,7 @@ tri_draw(const float *v0, const float *v1, const float *v2, uint tx, uint ty)
    if (!setup_sort_vertices((struct vertex_header *) v0,
                             (struct vertex_header *) v1,
                             (struct vertex_header *) v2)) {
-      return; /* totally clipped */
+      return FALSE; /* totally clipped */
    }
 
    setup_tri_coefficients();
@@ -999,6 +1017,24 @@ tri_draw(const float *v0, const float *v1, const float *v2, uint tx, uint ty)
 
    /*   init_constant_attribs( setup ); */
       
+   if (cur_tile_status_c == TILE_STATUS_GETTING) {
+      /* wait for mfc_get() to complete */
+      wait_on_mask(1 << TAG_READ_TILE_COLOR);
+      cur_tile_status_c = TILE_STATUS_CLEAN;
+   }
+
+   ASSERT(cur_tile_status_c != TILE_STATUS_DEFINED);
+
+   if (spu.depth_stencil.depth.enabled) {
+      if (cur_tile_status_z == TILE_STATUS_GETTING) {
+         /* wait for mfc_get() to complete */
+         wait_on_mask(1 << TAG_READ_TILE_Z);
+         cur_tile_status_z = TILE_STATUS_CLEAN;
+      }
+   ASSERT(cur_tile_status_z != TILE_STATUS_DEFINED);
+   }
+
+
    if (setup.oneoverarea < 0.0) {
       /* emaj on left:
        */
@@ -1013,4 +1049,6 @@ tri_draw(const float *v0, const float *v1, const float *v2, uint tx, uint ty)
    }
 
    flush_spans();
+
+   return TRUE;
 }
diff --git a/src/mesa/pipe/cell/spu/spu_tri.h b/src/mesa/pipe/cell/spu/spu_tri.h
index 86c42b63394..aa694dd7c93 100644
--- a/src/mesa/pipe/cell/spu/spu_tri.h
+++ b/src/mesa/pipe/cell/spu/spu_tri.h
@@ -30,7 +30,7 @@
 #define SPU_TRI_H
 
 
-extern void
+extern boolean
 tri_draw(const float *v0, const float *v1, const float *v2, uint tx, uint ty);