Cell: use depth/stencil state to enable ztest

Move z-test code into do_depth_test().
Add ZSIZE symbol to support 2 or 4-byte Z values.

5 files changed, 113 insertions, 62 deletions

diff --git a/src/mesa/pipe/cell/spu/spu_main.c b/src/mesa/pipe/cell/spu/spu_main.c
index bff098f06e..30c39aab68 100644
--- a/src/mesa/pipe/cell/spu/spu_main.c
+++ b/src/mesa/pipe/cell/spu/spu_main.c
@@ -239,8 +239,8 @@ cmd_render(const struct cell_command_render *render)
    /* how much vertex data */
    vertex_bytes = render->num_verts * vertex_size;
    index_bytes = render->num_indexes * sizeof(ushort);
-   if (index_bytes < 8)
-      index_bytes = 8;
+   if (index_bytes < 16)
+      index_bytes = 16;
    else
       index_bytes = (index_bytes + 15) & ~0xf; /* multiple of 16 */
 
@@ -297,7 +297,7 @@ cmd_render(const struct cell_command_render *render)
       /* Start fetching color/z tiles.  We'll wait for completion when
        * we need read/write to them later in triangle rasterization.
        */
-      if (spu.fb.depth_format == PIPE_FORMAT_Z16_UNORM) {
+      if (spu.depth_stencil.depth.enabled) {
          if (tile_status_z[ty][tx] != TILE_STATUS_CLEAR) {
             get_tile(tx, ty, (uint *) ztile, TAG_READ_TILE_Z, 1);
          }
@@ -325,7 +325,7 @@ cmd_render(const struct cell_command_render *render)
          put_tile(tx, ty, (uint *) ctile, TAG_WRITE_TILE_COLOR, 0);
          tile_status[ty][tx] = TILE_STATUS_DEFINED;
       }
-      if (spu.fb.depth_format == PIPE_FORMAT_Z16_UNORM) {
+      if (spu.depth_stencil.depth.enabled) {
          if (tile_status_z[ty][tx] == TILE_STATUS_DIRTY) {
             put_tile(tx, ty, (uint *) ztile, TAG_WRITE_TILE_Z, 1);
             tile_status_z[ty][tx] = TILE_STATUS_DEFINED;
@@ -334,7 +334,7 @@ cmd_render(const struct cell_command_render *render)
 
       /* XXX move these... */
       wait_on_mask(1 << TAG_WRITE_TILE_COLOR);
-      if (spu.fb.depth_format == PIPE_FORMAT_Z16_UNORM) {
+      if (spu.depth_stencil.depth.enabled) {
          wait_on_mask(1 << TAG_WRITE_TILE_Z);
       }
    }
@@ -365,6 +365,14 @@ cmd_framebuffer(const struct cell_command_framebuffer *cmd)
    spu.fb.height = cmd->height;
    spu.fb.width_tiles = (spu.fb.width + TILE_SIZE - 1) / TILE_SIZE;
    spu.fb.height_tiles = (spu.fb.height + TILE_SIZE - 1) / TILE_SIZE;
+
+   if (cmd->depth_format == PIPE_FORMAT_Z16_UNORM) {
+      ASSERT(ZSIZE == 2);
+   }
+   else if (cmd->depth_format == PIPE_FORMAT_Z32_UNORM) {
+      ASSERT(ZSIZE == 4);
+   }
+
 }
 
 
@@ -375,9 +383,9 @@ cmd_state_depth_stencil(const struct pipe_depth_stencil_alpha_state *state)
       printf("SPU %u: DEPTH_STENCIL: ztest %d\n",
              spu.init.id,
              state->depth.enabled);
-   /*
+
    memcpy(&spu.depth_stencil, state, sizeof(*state));
-   */
+
 }
 
 
diff --git a/src/mesa/pipe/cell/spu/spu_main.h b/src/mesa/pipe/cell/spu/spu_main.h
index f39f375d24..cd2afbe9bf 100644
--- a/src/mesa/pipe/cell/spu/spu_main.h
+++ b/src/mesa/pipe/cell/spu/spu_main.h
@@ -32,6 +32,8 @@
 #include "pipe/cell/common.h"
 #include "pipe/p_state.h"
 
+/** XXX temp bytes/z value */
+#define ZSIZE 2
 
 struct spu_framebuffer {
    void *color_start;              /**< addr of color surface in main memory */
diff --git a/src/mesa/pipe/cell/spu/spu_tile.c b/src/mesa/pipe/cell/spu/spu_tile.c
index 1505cb322b..13fc5e345f 100644
--- a/src/mesa/pipe/cell/spu/spu_tile.c
+++ b/src/mesa/pipe/cell/spu/spu_tile.c
@@ -32,7 +32,11 @@
 
 
 uint ctile[TILE_SIZE][TILE_SIZE] ALIGN16_ATTRIB;
+#if ZSIZE == 2
 ushort ztile[TILE_SIZE][TILE_SIZE] ALIGN16_ATTRIB;
+#else
+uint ztile[TILE_SIZE][TILE_SIZE] ALIGN16_ATTRIB;
+#endif
 
 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;
@@ -43,7 +47,7 @@ void
 get_tile(uint tx, uint ty, uint *tile, int tag, int zBuf)
 {
    const uint offset = ty * spu.fb.width_tiles + tx;
-   const uint bytesPerTile = TILE_SIZE * TILE_SIZE * (zBuf ? 2 : 4);
+   const uint bytesPerTile = TILE_SIZE * TILE_SIZE * (zBuf ? ZSIZE : 4);
    const ubyte *src = zBuf ? spu.fb.depth_start : spu.fb.color_start;
 
    src += offset * bytesPerTile;
@@ -68,7 +72,7 @@ void
 put_tile(uint tx, uint ty, const uint *tile, int tag, int zBuf)
 {
    const uint offset = ty * spu.fb.width_tiles + tx;
-   const uint bytesPerTile = TILE_SIZE * TILE_SIZE * (zBuf ? 2 : 4);
+   const uint bytesPerTile = TILE_SIZE * TILE_SIZE * (zBuf ? ZSIZE : 4);
    ubyte *dst = zBuf ? spu.fb.depth_start : spu.fb.color_start;
 
    dst += offset * bytesPerTile;
@@ -81,7 +85,6 @@ put_tile(uint tx, uint ty, const uint *tile, int tag, int zBuf)
           spu.init.id,
           tile, (unsigned int) dst, bytesPerTile);
    */
-
    mfc_put((void *) tile,  /* src in local memory */
            (unsigned int) dst,  /* dst in main memory */
            bytesPerTile,
@@ -103,7 +106,13 @@ clear_tile(uint tile[TILE_SIZE][TILE_SIZE], uint value)
 }
 
 void
-clear_tile_z(ushort tile[TILE_SIZE][TILE_SIZE], uint value)
+clear_tile_z(
+#if ZSIZE == 2
+             ushort tile[TILE_SIZE][TILE_SIZE],
+#else
+             uint tile[TILE_SIZE][TILE_SIZE],
+#endif
+             uint value)
 {
    uint i, j;
    for (i = 0; i < TILE_SIZE; i++) {
diff --git a/src/mesa/pipe/cell/spu/spu_tile.h b/src/mesa/pipe/cell/spu/spu_tile.h
index 637923764b..7c288a1f91 100644
--- a/src/mesa/pipe/cell/spu/spu_tile.h
+++ b/src/mesa/pipe/cell/spu/spu_tile.h
@@ -40,7 +40,11 @@
 
 
 extern uint ctile[TILE_SIZE][TILE_SIZE] ALIGN16_ATTRIB;
+#if ZSIZE == 2
 extern ushort ztile[TILE_SIZE][TILE_SIZE] ALIGN16_ATTRIB;
+#else
+extern uint ztile[TILE_SIZE][TILE_SIZE] ALIGN16_ATTRIB;
+#endif
 
 
 #define TILE_STATUS_CLEAR   1
@@ -61,7 +65,13 @@ void
 clear_tile(uint tile[TILE_SIZE][TILE_SIZE], uint value);
 
 void
-clear_tile_z(ushort tile[TILE_SIZE][TILE_SIZE], uint value);
+clear_tile_z(
+#if ZSIZE == 2
+             ushort tile[TILE_SIZE][TILE_SIZE],
+#else
+             uint tile[TILE_SIZE][TILE_SIZE],
+#endif
+             uint value);
 
 
 #endif /* SPU_TILE_H */
diff --git a/src/mesa/pipe/cell/spu/spu_tri.c b/src/mesa/pipe/cell/spu/spu_tri.c
index ddd5e662d2..0b8533092e 100644
--- a/src/mesa/pipe/cell/spu/spu_tri.c
+++ b/src/mesa/pipe/cell/spu/spu_tri.c
@@ -217,8 +217,8 @@ static INLINE void
 eval_z( struct setup_stage *setup,
         float x, float y, float result[4])
 {
-   uint slot = 0;
-   uint i = 2;
+   const uint slot = 0;
+   const uint i = 2;
    const float *dadx = setup->coef[slot].dadx;
    const float *dady = setup->coef[slot].dady;
 
@@ -252,6 +252,72 @@ pack_color(const float color[4])
 }
 
 
+static uint
+do_depth_test(struct setup_stage *setup, int x, int y, unsigned mask)
+{
+   int ix = x - setup->cliprect_minx;
+   int iy = y - setup->cliprect_miny;
+   float zvals[4];
+   float zscale = 65535.0;
+
+   if (ZSIZE == 2) {
+      ASSERT(spu.fb.depth_format == PIPE_FORMAT_Z16_UNORM);
+   }
+   else {
+      ASSERT(spu.fb.depth_format == PIPE_FORMAT_Z32_UNORM);
+   }
+   ASSERT(sizeof(ztile[0][0]) == ZSIZE);
+
+
+   eval_z(setup, (float) x, (float) y, zvals);
+
+   if (tile_status_z[setup->ty][setup->tx] == TILE_STATUS_CLEAR) {
+      /* now, _really_ clear the tile */
+      clear_tile_z(ztile, spu.fb.depth_clear_value);
+   }
+   else {
+      /* 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;
+
+
+   if (mask & MASK_TOP_LEFT) {
+      uint z = (uint) (zvals[0] * zscale);
+      if (z < ztile[iy][ix])
+         ztile[iy][ix] = z;
+      else
+         mask &= ~MASK_TOP_LEFT;
+   }
+
+   if (mask & MASK_TOP_RIGHT) {
+      uint z = (uint) (zvals[1] * zscale);
+      if (z < ztile[iy][ix+1])
+         ztile[iy][ix+1] = z;
+      else
+         mask &= ~MASK_TOP_RIGHT;
+   }
+
+   if (mask & MASK_BOTTOM_LEFT) {
+      uint z = (uint) (zvals[2] * zscale);
+      if (z < ztile[iy+1][ix])
+         ztile[iy+1][ix] = z;
+      else
+         mask &= ~MASK_BOTTOM_LEFT;
+   }
+
+   if (mask & MASK_BOTTOM_RIGHT) {
+      uint z = (uint) (zvals[3] * zscale);
+      if (z < ztile[iy+1][ix+1])
+         ztile[iy+1][ix+1] = z;
+      else
+         mask &= ~MASK_BOTTOM_RIGHT;
+   }
+
+   return mask;
+}
+
+
 /**
  * Emit a quad (pass to next stage).  No clipping is done.
  */
@@ -266,58 +332,14 @@ emit_quad( struct setup_stage *setup, int x, int y, unsigned mask )
    sp->quad.first->run(sp->quad.first, &setup->quad);
 #else
    /* Cell: "write" quad fragments to the tile by setting prim color */
-   int ix = x - setup->cliprect_minx;
-   int iy = y - setup->cliprect_miny;
+   const int ix = x - setup->cliprect_minx;
+   const int iy = y - setup->cliprect_miny;
    float colors[4][4];
-   uint z;
 
    eval_coeff(setup, 1, (float) x, (float) y, colors);
 
-   if (spu.fb.depth_format == PIPE_FORMAT_Z16_UNORM) {
-      float zvals[4];
-      eval_z(setup, (float) x, (float) y, zvals);
-
-      if (tile_status_z[setup->ty][setup->tx] == TILE_STATUS_CLEAR) {
-         /* now, _really_ clear the tile */
-         clear_tile_z(ztile, spu.fb.depth_clear_value);
-      }
-      else {
-         /* 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;
-
-      if (mask & MASK_TOP_LEFT) {
-         z = (uint) (zvals[0] * 65535.0);
-         if (z < ztile[iy][ix])
-            ztile[iy][ix] = z;
-         else
-            mask &= ~MASK_TOP_LEFT;
-      }
-
-      if (mask & MASK_TOP_RIGHT) {
-         z = (uint) (zvals[1] * 65535.0);
-         if (z < ztile[iy][ix+1])
-            ztile[iy][ix+1] = z;
-         else
-            mask &= ~MASK_TOP_RIGHT;
-      }
-
-      if (mask & MASK_BOTTOM_LEFT) {
-         z = (uint) (zvals[2] * 65535.0);
-         if (z < ztile[iy+1][ix])
-            ztile[iy+1][ix] = z;
-         else
-            mask &= ~MASK_BOTTOM_LEFT;
-      }
-
-      if (mask & MASK_BOTTOM_RIGHT) {
-         z = (uint) (zvals[3] * 65535.0);
-         if (z < ztile[iy+1][ix+1])
-            ztile[iy+1][ix+1] = z;
-         else
-            mask &= ~MASK_BOTTOM_RIGHT;
-      }
+   if (spu.depth_stencil.depth.enabled) {
+      mask &= do_depth_test(setup, x, y, mask);
    }
 
    if (mask) {