softpipe: move all depth/stencil/alpha pixel processing into one stage

1 files changed, 544 insertions, 115 deletions

diff --git a/src/gallium/drivers/softpipe/sp_quad_depth_test.c b/src/gallium/drivers/softpipe/sp_quad_depth_test.c
index 8f223a7eaeb..bf65799a818 100644
--- a/src/gallium/drivers/softpipe/sp_quad_depth_test.c
+++ b/src/gallium/drivers/softpipe/sp_quad_depth_test.c
@@ -31,61 +31,109 @@
 
 #include "pipe/p_defines.h"
 #include "util/u_memory.h"
+#include "tgsi/tgsi_scan.h"
 #include "sp_context.h"
 #include "sp_quad.h"
 #include "sp_surface.h"
 #include "sp_quad_pipe.h"
 #include "sp_tile_cache.h"
+#include "sp_state.h"           /* for sp_fragment_shader */
 
 
-/**
- * Do depth testing for a quad.
- * Not static since it's used by the stencil code.
- */
+struct depth_data {
+   struct pipe_surface *ps;
+   enum pipe_format format;
+   unsigned bzzzz[QUAD_SIZE];  /**< Z values fetched from depth buffer */
+   unsigned qzzzz[QUAD_SIZE];  /**< Z values from the quad */
+   ubyte stencilVals[QUAD_SIZE];
+   struct softpipe_cached_tile *tile;
+};
 
-/*
- * To increase efficiency, we should probably have multiple versions
- * of this function that are specifically for Z16, Z32 and FP Z buffers.
- * Try to effectively do that with codegen...
- */
 
-boolean
-sp_depth_test_quad(struct quad_stage *qs, struct quad_header *quad)
+
+static void
+get_depth_stencil_values( struct depth_data *data,
+                          const struct quad_header *quad )
 {
-   struct softpipe_context *softpipe = qs->softpipe;
-   struct pipe_surface *ps = softpipe->framebuffer.zsbuf;
-   const enum pipe_format format = ps->format;
-   unsigned bzzzz[QUAD_SIZE];  /**< Z values fetched from depth buffer */
-   unsigned qzzzz[QUAD_SIZE];  /**< Z values from the quad */
-   unsigned zmask = 0;
    unsigned j;
-   struct softpipe_cached_tile *tile
-      = sp_get_cached_tile(softpipe->zsbuf_cache, quad->input.x0, quad->input.y0);
+   const struct softpipe_cached_tile *tile = data->tile;
+
+   switch (data->format) {
+   case PIPE_FORMAT_Z16_UNORM:
+      for (j = 0; j < QUAD_SIZE; j++) {
+         int x = quad->input.x0 % TILE_SIZE + (j & 1);
+         int y = quad->input.y0 % TILE_SIZE + (j >> 1);
+         data->bzzzz[j] = tile->data.depth16[y][x];
+      }
+      break;
+   case PIPE_FORMAT_Z32_UNORM:
+      for (j = 0; j < QUAD_SIZE; j++) {
+         int x = quad->input.x0 % TILE_SIZE + (j & 1);
+         int y = quad->input.y0 % TILE_SIZE + (j >> 1);
+         data->bzzzz[j] = tile->data.depth32[y][x];
+      }
+      break;
+   case PIPE_FORMAT_X8Z24_UNORM:
+   case PIPE_FORMAT_S8Z24_UNORM:
+      for (j = 0; j < QUAD_SIZE; j++) {
+         int x = quad->input.x0 % TILE_SIZE + (j & 1);
+         int y = quad->input.y0 % TILE_SIZE + (j >> 1);
+         data->bzzzz[j] = tile->data.depth32[y][x] & 0xffffff;
+         data->stencilVals[j] = tile->data.depth32[y][x] >> 24;
+      }
+   break;
+   case PIPE_FORMAT_Z24X8_UNORM:
+   case PIPE_FORMAT_Z24S8_UNORM:
+      for (j = 0; j < QUAD_SIZE; j++) {
+         int x = quad->input.x0 % TILE_SIZE + (j & 1);
+         int y = quad->input.y0 % TILE_SIZE + (j >> 1);
+         data->bzzzz[j] = tile->data.depth32[y][x] >> 8;
+         data->stencilVals[j] = tile->data.depth32[y][x] & 0xff;
+      }
+      break;
+   default:
+      assert(0);
+   }
+}
 
-   assert(ps); /* shouldn't get here if there's no zbuffer */
+/* If the shader has not been run, interpolate the depth values
+ * ourselves.
+ */
+static void
+interpolate_quad_depth( struct quad_header *quad )
+{
+   const float fx = (float) quad->input.x0;
+   const float fy = (float) quad->input.y0;
+   const float dzdx = quad->posCoef->dadx[2];
+   const float dzdy = quad->posCoef->dady[2];
+   const float z0 = quad->posCoef->a0[2] + dzdx * fx + dzdy * fy;
+
+   quad->output.depth[0] = z0;
+   quad->output.depth[1] = z0 + dzdx;
+   quad->output.depth[2] = z0 + dzdy;
+   quad->output.depth[3] = z0 + dzdx + dzdy;
+}
 
-   /*
-    * Convert quad's float depth values to int depth values (qzzzz).
+
+static void
+convert_quad_depth( struct depth_data *data, 
+                    const struct quad_header *quad )
+{
+   unsigned j;
+
+   /* Convert quad's float depth values to int depth values (qzzzz).
     * If the Z buffer stores integer values, we _have_ to do the depth
     * compares with integers (not floats).  Otherwise, the float->int->float
     * conversion of Z values (which isn't an identity function) will cause
     * Z-fighting errors.
-    *
-    * Also, get the zbuffer values (bzzzz) from the cached tile.
     */
-   switch (format) {
+   switch (data->format) {
    case PIPE_FORMAT_Z16_UNORM:
       {
          float scale = 65535.0;
 
          for (j = 0; j < QUAD_SIZE; j++) {
-            qzzzz[j] = (unsigned) (quad->output.depth[j] * scale);
-         }
-
-         for (j = 0; j < QUAD_SIZE; j++) {
-            int x = quad->input.x0 % TILE_SIZE + (j & 1);
-            int y = quad->input.y0 % TILE_SIZE + (j >> 1);
-            bzzzz[j] = tile->data.depth16[y][x];
+            data->qzzzz[j] = (unsigned) (quad->output.depth[j] * scale);
          }
       }
       break;
@@ -94,47 +142,247 @@ sp_depth_test_quad(struct quad_stage *qs, struct quad_header *quad)
          double scale = (double) (uint) ~0UL;
 
          for (j = 0; j < QUAD_SIZE; j++) {
-            qzzzz[j] = (unsigned) (quad->output.depth[j] * scale);
-         }
-
-         for (j = 0; j < QUAD_SIZE; j++) {
-            int x = quad->input.x0 % TILE_SIZE + (j & 1);
-            int y = quad->input.y0 % TILE_SIZE + (j >> 1);
-            bzzzz[j] = tile->data.depth32[y][x];
+            data->qzzzz[j] = (unsigned) (quad->output.depth[j] * scale);
          }
       }
       break;
    case PIPE_FORMAT_X8Z24_UNORM:
-      /* fall-through */
    case PIPE_FORMAT_S8Z24_UNORM:
       {
          float scale = (float) ((1 << 24) - 1);
 
          for (j = 0; j < QUAD_SIZE; j++) {
-            qzzzz[j] = (unsigned) (quad->output.depth[j] * scale);
-         }
-
-         for (j = 0; j < QUAD_SIZE; j++) {
-            int x = quad->input.x0 % TILE_SIZE + (j & 1);
-            int y = quad->input.y0 % TILE_SIZE + (j >> 1);
-            bzzzz[j] = tile->data.depth32[y][x] & 0xffffff;
+            data->qzzzz[j] = (unsigned) (quad->output.depth[j] * scale);
          }
       }
       break;
    case PIPE_FORMAT_Z24X8_UNORM:
-      /* fall-through */
    case PIPE_FORMAT_Z24S8_UNORM:
       {
          float scale = (float) ((1 << 24) - 1);
 
          for (j = 0; j < QUAD_SIZE; j++) {
-            qzzzz[j] = (unsigned) (quad->output.depth[j] * scale);
+            data->qzzzz[j] = (unsigned) (quad->output.depth[j] * scale);
          }
+      }
+      break;
+   default:
+      assert(0);
+   }
+}
 
-         for (j = 0; j < QUAD_SIZE; j++) {
-            int x = quad->input.x0 % TILE_SIZE + (j & 1);
-            int y = quad->input.y0 % TILE_SIZE + (j >> 1);
-            bzzzz[j] = tile->data.depth32[y][x] >> 8;
+
+
+static void
+write_depth_stencil_values( struct depth_data *data,
+                            struct quad_header *quad )
+{
+   struct softpipe_cached_tile *tile = data->tile;
+   unsigned j;
+
+   /* put updated Z values back into cached tile */
+   switch (data->format) {
+   case PIPE_FORMAT_Z16_UNORM:
+      for (j = 0; j < QUAD_SIZE; j++) {
+         int x = quad->input.x0 % TILE_SIZE + (j & 1);
+         int y = quad->input.y0 % TILE_SIZE + (j >> 1);
+         tile->data.depth16[y][x] = (ushort) data->bzzzz[j];
+      }
+      break;
+   case PIPE_FORMAT_X8Z24_UNORM:
+   case PIPE_FORMAT_Z32_UNORM:
+      for (j = 0; j < QUAD_SIZE; j++) {
+         int x = quad->input.x0 % TILE_SIZE + (j & 1);
+         int y = quad->input.y0 % TILE_SIZE + (j >> 1);
+         tile->data.depth32[y][x] = data->bzzzz[j];
+      }
+      break;
+   case PIPE_FORMAT_S8Z24_UNORM:
+      for (j = 0; j < QUAD_SIZE; j++) {
+         int x = quad->input.x0 % TILE_SIZE + (j & 1);
+         int y = quad->input.y0 % TILE_SIZE + (j >> 1);
+         tile->data.depth32[y][x] = (data->stencilVals[j] << 24) | data->bzzzz[j];
+      }
+      break;
+   case PIPE_FORMAT_Z24S8_UNORM:
+      for (j = 0; j < QUAD_SIZE; j++) {
+         int x = quad->input.x0 % TILE_SIZE + (j & 1);
+         int y = quad->input.y0 % TILE_SIZE + (j >> 1);
+         tile->data.depth32[y][x] = (data->bzzzz[j] << 8) | data->stencilVals[j];
+      }
+      break;
+   case PIPE_FORMAT_Z24X8_UNORM:
+      for (j = 0; j < QUAD_SIZE; j++) {
+         int x = quad->input.x0 % TILE_SIZE + (j & 1);
+         int y = quad->input.y0 % TILE_SIZE + (j >> 1);
+         tile->data.depth32[y][x] = data->bzzzz[j] << 8;
+      }
+      break;
+   default:
+      assert(0);
+   }
+}
+
+
+
+
+/** Only 8-bit stencil supported */
+#define STENCIL_MAX 0xff
+
+
+/**
+ * Do the basic stencil test (compare stencil buffer values against the
+ * reference value.
+ *
+ * \param data->stencilVals  the stencil values from the stencil buffer
+ * \param func  the stencil func (PIPE_FUNC_x)
+ * \param ref  the stencil reference value
+ * \param valMask  the stencil value mask indicating which bits of the stencil
+ *                 values and ref value are to be used.
+ * \return mask indicating which pixels passed the stencil test
+ */
+static unsigned
+do_stencil_test(struct depth_data *data,
+                unsigned func,
+                unsigned ref, unsigned valMask)
+{
+   unsigned passMask = 0x0;
+   unsigned j;
+
+   ref &= valMask;
+
+   switch (func) {
+   case PIPE_FUNC_NEVER:
+      /* passMask = 0x0 */
+      break;
+   case PIPE_FUNC_LESS:
+      for (j = 0; j < QUAD_SIZE; j++) {
+         if (ref < (data->stencilVals[j] & valMask)) {
+            passMask |= (1 << j);
+         }
+      }
+      break;
+   case PIPE_FUNC_EQUAL:
+      for (j = 0; j < QUAD_SIZE; j++) {
+         if (ref == (data->stencilVals[j] & valMask)) {
+            passMask |= (1 << j);
+         }
+      }
+      break;
+   case PIPE_FUNC_LEQUAL:
+      for (j = 0; j < QUAD_SIZE; j++) {
+         if (ref <= (data->stencilVals[j] & valMask)) {
+            passMask |= (1 << j);
+         }
+      }
+      break;
+   case PIPE_FUNC_GREATER:
+      for (j = 0; j < QUAD_SIZE; j++) {
+         if (ref > (data->stencilVals[j] & valMask)) {
+            passMask |= (1 << j);
+         }
+      }
+      break;
+   case PIPE_FUNC_NOTEQUAL:
+      for (j = 0; j < QUAD_SIZE; j++) {
+         if (ref != (data->stencilVals[j] & valMask)) {
+            passMask |= (1 << j);
+         }
+      }
+      break;
+   case PIPE_FUNC_GEQUAL:
+      for (j = 0; j < QUAD_SIZE; j++) {
+         if (ref >= (data->stencilVals[j] & valMask)) {
+            passMask |= (1 << j);
+         }
+      }
+      break;
+   case PIPE_FUNC_ALWAYS:
+      passMask = MASK_ALL;
+      break;
+   default:
+      assert(0);
+   }
+
+   return passMask;
+}
+
+
+/**
+ * Apply the stencil operator to stencil values.
+ *
+ * \param data->stencilVals  the stencil buffer values (read and written)
+ * \param mask  indicates which pixels to update
+ * \param op  the stencil operator (PIPE_STENCIL_OP_x)
+ * \param ref  the stencil reference value
+ * \param wrtMask  writemask controlling which bits are changed in the
+ *                 stencil values
+ */
+static void
+apply_stencil_op(struct depth_data *data,
+                 unsigned mask, unsigned op, ubyte ref, ubyte wrtMask)
+{
+   unsigned j;
+   ubyte newstencil[QUAD_SIZE];
+
+   for (j = 0; j < QUAD_SIZE; j++) {
+      newstencil[j] = data->stencilVals[j];
+   }
+
+   switch (op) {
+   case PIPE_STENCIL_OP_KEEP:
+      /* no-op */
+      break;
+   case PIPE_STENCIL_OP_ZERO:
+      for (j = 0; j < QUAD_SIZE; j++) {
+         if (mask & (1 << j)) {
+            newstencil[j] = 0;
+         }
+      }
+      break;
+   case PIPE_STENCIL_OP_REPLACE:
+      for (j = 0; j < QUAD_SIZE; j++) {
+         if (mask & (1 << j)) {
+            newstencil[j] = ref;
+         }
+      }
+      break;
+   case PIPE_STENCIL_OP_INCR:
+      for (j = 0; j < QUAD_SIZE; j++) {
+         if (mask & (1 << j)) {
+            if (data->stencilVals[j] < STENCIL_MAX) {
+               newstencil[j] = data->stencilVals[j] + 1;
+            }
+         }
+      }
+      break;
+   case PIPE_STENCIL_OP_DECR:
+      for (j = 0; j < QUAD_SIZE; j++) {
+         if (mask & (1 << j)) {
+            if (data->stencilVals[j] > 0) {
+               newstencil[j] = data->stencilVals[j] - 1;
+            }
+         }
+      }
+      break;
+   case PIPE_STENCIL_OP_INCR_WRAP:
+      for (j = 0; j < QUAD_SIZE; j++) {
+         if (mask & (1 << j)) {
+            newstencil[j] = data->stencilVals[j] + 1;
+         }
+      }
+      break;
+   case PIPE_STENCIL_OP_DECR_WRAP:
+      for (j = 0; j < QUAD_SIZE; j++) {
+         if (mask & (1 << j)) {
+            newstencil[j] = data->stencilVals[j] - 1;
+         }
+      }
+      break;
+   case PIPE_STENCIL_OP_INVERT:
+      for (j = 0; j < QUAD_SIZE; j++) {
+         if (mask & (1 << j)) {
+            newstencil[j] = ~data->stencilVals[j];
          }
       }
       break;
@@ -142,6 +390,39 @@ sp_depth_test_quad(struct quad_stage *qs, struct quad_header *quad)
       assert(0);
    }
 
+   /*
+    * update the stencil values
+    */
+   if (wrtMask != STENCIL_MAX) {
+      /* apply bit-wise stencil buffer writemask */
+      for (j = 0; j < QUAD_SIZE; j++) {
+         data->stencilVals[j] = (wrtMask & newstencil[j]) | (~wrtMask & data->stencilVals[j]);
+      }
+   }
+   else {
+      for (j = 0; j < QUAD_SIZE; j++) {
+         data->stencilVals[j] = newstencil[j];
+      }
+   }
+}
+
+   
+
+/*
+ * To increase efficiency, we should probably have multiple versions
+ * of this function that are specifically for Z16, Z32 and FP Z buffers.
+ * Try to effectively do that with codegen...
+ */
+
+static boolean
+depth_test_quad(struct quad_stage *qs, 
+                struct depth_data *data,
+                struct quad_header *quad)
+{
+   struct softpipe_context *softpipe = qs->softpipe;
+   unsigned zmask = 0;
+   unsigned j;
+
    switch (softpipe->depth_stencil->depth.func) {
    case PIPE_FUNC_NEVER:
       /* zmask = 0 */
@@ -151,37 +432,37 @@ sp_depth_test_quad(struct quad_stage *qs, struct quad_header *quad)
        * Like this:  quad->mask &= (quad->outputs.depth < zzzz);
        */
       for (j = 0; j < QUAD_SIZE; j++) {
-	 if (qzzzz[j] < bzzzz[j]) 
+	 if (data->qzzzz[j] < data->bzzzz[j]) 
 	    zmask |= 1 << j;
       }
       break;
    case PIPE_FUNC_EQUAL:
       for (j = 0; j < QUAD_SIZE; j++) {
-	 if (qzzzz[j] == bzzzz[j]) 
+	 if (data->qzzzz[j] == data->bzzzz[j]) 
 	    zmask |= 1 << j;
       }
       break;
    case PIPE_FUNC_LEQUAL:
       for (j = 0; j < QUAD_SIZE; j++) {
-	 if (qzzzz[j] <= bzzzz[j]) 
+	 if (data->qzzzz[j] <= data->bzzzz[j]) 
 	    zmask |= (1 << j);
       }
       break;
    case PIPE_FUNC_GREATER:
       for (j = 0; j < QUAD_SIZE; j++) {
-	 if (qzzzz[j] > bzzzz[j]) 
+	 if (data->qzzzz[j] > data->bzzzz[j]) 
 	    zmask |= (1 << j);
       }
       break;
    case PIPE_FUNC_NOTEQUAL:
       for (j = 0; j < QUAD_SIZE; j++) {
-	 if (qzzzz[j] != bzzzz[j]) 
+	 if (data->qzzzz[j] != data->bzzzz[j]) 
 	    zmask |= (1 << j);
       }
       break;
    case PIPE_FUNC_GEQUAL:
       for (j = 0; j < QUAD_SIZE; j++) {
-	 if (qzzzz[j] >= bzzzz[j]) 
+	 if (data->qzzzz[j] >= data->bzzzz[j]) 
 	    zmask |= (1 << j);
       }
       break;
@@ -196,83 +477,231 @@ sp_depth_test_quad(struct quad_stage *qs, struct quad_header *quad)
    if (quad->inout.mask == 0)
       return FALSE;
 
+   /* Update our internal copy only if writemask set.  Even if
+    * depth.writemask is FALSE, may still need to write out buffer
+    * data due to stencil changes.
+    */
    if (softpipe->depth_stencil->depth.writemask) {
-      
-      /* This is also efficient with sse / spe instructions: 
-       */
       for (j = 0; j < QUAD_SIZE; j++) {
-	 if (quad->inout.mask & (1 << j)) {
-	    bzzzz[j] = qzzzz[j];
-	 }
+         if (quad->inout.mask & (1 << j)) {
+            data->bzzzz[j] = data->qzzzz[j];
+         }
       }
+   }
 
-      /* put updated Z values back into cached tile */
-      switch (format) {
-      case PIPE_FORMAT_Z16_UNORM:
-         for (j = 0; j < QUAD_SIZE; j++) {
-            int x = quad->input.x0 % TILE_SIZE + (j & 1);
-            int y = quad->input.y0 % TILE_SIZE + (j >> 1);
-            tile->data.depth16[y][x] = (ushort) bzzzz[j];
-         }
-         break;
-      case PIPE_FORMAT_X8Z24_UNORM:
-         /* fall-through */
-         /* (yes, this falls through to a different case than above) */
-      case PIPE_FORMAT_Z32_UNORM:
-         for (j = 0; j < QUAD_SIZE; j++) {
-            int x = quad->input.x0 % TILE_SIZE + (j & 1);
-            int y = quad->input.y0 % TILE_SIZE + (j >> 1);
-            tile->data.depth32[y][x] = bzzzz[j];
-         }
-         break;
-      case PIPE_FORMAT_S8Z24_UNORM:
-         for (j = 0; j < QUAD_SIZE; j++) {
-            int x = quad->input.x0 % TILE_SIZE + (j & 1);
-            int y = quad->input.y0 % TILE_SIZE + (j >> 1);
-            uint s8z24 = tile->data.depth32[y][x];
-            s8z24 = (s8z24 & 0xff000000) | bzzzz[j];
-            tile->data.depth32[y][x] = s8z24;
-         }
-         break;
-      case PIPE_FORMAT_Z24S8_UNORM:
-         for (j = 0; j < QUAD_SIZE; j++) {
-            int x = quad->input.x0 % TILE_SIZE + (j & 1);
-            int y = quad->input.y0 % TILE_SIZE + (j >> 1);
-            uint z24s8 = tile->data.depth32[y][x];
-            z24s8 = (z24s8 & 0xff) | (bzzzz[j] << 8);
-            tile->data.depth32[y][x] = z24s8;
+   return TRUE;
+}
+
+
+
+/**
+ * Do stencil (and depth) testing.  Stenciling depends on the outcome of
+ * depth testing.
+ */
+static boolean
+depth_stencil_test_quad(struct quad_stage *qs, 
+                        struct depth_data *data,
+                        struct quad_header *quad)
+{
+   struct softpipe_context *softpipe = qs->softpipe;
+   unsigned func, zFailOp, zPassOp, failOp;
+   ubyte ref, wrtMask, valMask;
+   uint face = quad->input.facing;
+
+   if (!softpipe->depth_stencil->stencil[1].enabled) {
+      /* single-sided stencil test, use front (face=0) state */
+      face = 0;
+   }
+
+   /* choose front or back face function, operator, etc */
+   /* XXX we could do these initializations once per primitive */
+   func    = softpipe->depth_stencil->stencil[face].func;
+   failOp  = softpipe->depth_stencil->stencil[face].fail_op;
+   zFailOp = softpipe->depth_stencil->stencil[face].zfail_op;
+   zPassOp = softpipe->depth_stencil->stencil[face].zpass_op;
+   ref     = softpipe->depth_stencil->stencil[face].ref_value;
+   wrtMask = softpipe->depth_stencil->stencil[face].writemask;
+   valMask = softpipe->depth_stencil->stencil[face].valuemask;
+
+
+   /* do the stencil test first */
+   {
+      unsigned passMask, failMask;
+      passMask = do_stencil_test(data, func, ref, valMask);
+      failMask = quad->inout.mask & ~passMask;
+      quad->inout.mask &= passMask;
+
+      if (failOp != PIPE_STENCIL_OP_KEEP) {
+         apply_stencil_op(data, failMask, failOp, ref, wrtMask);
+      }
+   }
+
+   if (quad->inout.mask) {
+      /* now the pixels that passed the stencil test are depth tested */
+      if (softpipe->depth_stencil->depth.enabled) {
+         const unsigned origMask = quad->inout.mask;
+
+         depth_test_quad(qs, data, quad);  /* quad->mask is updated */
+
+         /* update stencil buffer values according to z pass/fail result */
+         if (zFailOp != PIPE_STENCIL_OP_KEEP) {
+            const unsigned failMask = origMask & ~quad->inout.mask;
+            apply_stencil_op(data, failMask, zFailOp, ref, wrtMask);
          }
-         break;
-      case PIPE_FORMAT_Z24X8_UNORM:
-         for (j = 0; j < QUAD_SIZE; j++) {
-            int x = quad->input.x0 % TILE_SIZE + (j & 1);
-            int y = quad->input.y0 % TILE_SIZE + (j >> 1);
-            tile->data.depth32[y][x] = bzzzz[j] << 8;
+
+         if (zPassOp != PIPE_STENCIL_OP_KEEP) {
+            const unsigned passMask = origMask & quad->inout.mask;
+            apply_stencil_op(data, passMask, zPassOp, ref, wrtMask);
          }
-         break;
-      default:
-         assert(0);
+      }
+      else {
+         /* no depth test, apply Zpass operator to stencil buffer values */
+         apply_stencil_op(data, quad->inout.mask, zPassOp, ref, wrtMask);
       }
    }
 
-   return TRUE;
+   return quad->inout.mask != 0;
 }
 
 
+#define ALPHATEST( FUNC, COMP )                                         \
+   static int                                                          \
+   alpha_test_quads_##FUNC( struct quad_stage *qs,                      \
+                           struct quad_header *quads[],                 \
+                           unsigned nr )                                \
+   {                                                                    \
+      const float ref = qs->softpipe->depth_stencil->alpha.ref_value;   \
+      const uint cbuf = 0; /* only output[0].alpha is tested */         \
+      unsigned pass_nr = 0;                                             \
+      unsigned i;                                                       \
+                                                                        \
+      for (i = 0; i < nr; i++) {                                        \
+         const float *aaaa = quads[i]->output.color[cbuf][3];           \
+         unsigned passMask = 0;                                         \
+                                                                        \
+         if (aaaa[0] COMP ref) passMask |= (1 << 0);                    \
+         if (aaaa[1] COMP ref) passMask |= (1 << 1);                    \
+         if (aaaa[2] COMP ref) passMask |= (1 << 2);                    \
+         if (aaaa[3] COMP ref) passMask |= (1 << 3);                    \
+                                                                        \
+         quads[i]->inout.mask &= passMask;                              \
+                                                                        \
+         if (quads[i]->inout.mask)                                      \
+            quads[pass_nr++] = quads[i];                                \
+      }                                                                 \
+                                                                        \
+      return pass_nr;                                                   \
+   }
+
+
+ALPHATEST( LESS,     < )
+ALPHATEST( EQUAL,    == )
+ALPHATEST( LEQUAL,   <= )
+ALPHATEST( GREATER,  > )
+ALPHATEST( NOTEQUAL, != )
+ALPHATEST( GEQUAL,   >= )
+
+
+/* XXX: Incorporate into shader using KILP.
+ */
+static int
+alpha_test_quads(struct quad_stage *qs, 
+                 struct quad_header *quads[], 
+                 unsigned nr)
+{
+   switch (qs->softpipe->depth_stencil->alpha.func) {
+   case PIPE_FUNC_LESS:
+      return alpha_test_quads_LESS( qs, quads, nr );
+   case PIPE_FUNC_EQUAL:
+      return alpha_test_quads_EQUAL( qs, quads, nr );
+      break;
+   case PIPE_FUNC_LEQUAL:
+      return alpha_test_quads_LEQUAL( qs, quads, nr );
+   case PIPE_FUNC_GREATER:
+      return alpha_test_quads_GREATER( qs, quads, nr );
+   case PIPE_FUNC_NOTEQUAL:
+      return alpha_test_quads_NOTEQUAL( qs, quads, nr );
+   case PIPE_FUNC_GEQUAL:
+      return alpha_test_quads_GEQUAL( qs, quads, nr );
+   case PIPE_FUNC_ALWAYS:
+      return nr;
+   case PIPE_FUNC_NEVER:
+   default:
+      return 0;
+   }
+}
+
+static unsigned mask_count[0x8] = 
+{
+   0,                           /* 0x0 */
+   1,                           /* 0x1 */
+   1,                           /* 0x2 */
+   2,                           /* 0x3 */
+   1,                           /* 0x4 */
+   2,                           /* 0x5 */
+   2,                           /* 0x6 */
+   3,                           /* 0x7 */
+};
+
+
+
 static void
 depth_test_quads(struct quad_stage *qs, 
                  struct quad_header *quads[],
                  unsigned nr)
 {
    unsigned i, pass = 0;
+   const struct sp_fragment_shader *fs = qs->softpipe->fs;
+   boolean interp_depth = !fs->info.writes_z;
+   struct depth_data data;
 
-   for (i = 0; i < nr; i++) {
-      if (sp_depth_test_quad(qs, quads[i]))
+
+   if (qs->softpipe->depth_stencil->alpha.enabled) {
+      nr = alpha_test_quads(qs, quads, nr);
+   }
+
+   if (qs->softpipe->framebuffer.zsbuf && 
+       (qs->softpipe->depth_stencil->depth.enabled ||
+        qs->softpipe->depth_stencil->stencil[0].enabled)) {
+
+      data.ps = qs->softpipe->framebuffer.zsbuf;
+      data.format = data.ps->format;
+      data.tile = sp_get_cached_tile(qs->softpipe->zsbuf_cache, 
+                                     quads[0]->input.x0, 
+                                     quads[0]->input.y0);
+
+      for (i = 0; i < nr; i++) {
+         get_depth_stencil_values(&data, quads[i]);
+
+         if (qs->softpipe->depth_stencil->depth.enabled) {
+            if (interp_depth)
+               interpolate_quad_depth(quads[i]);
+
+            convert_quad_depth(&data, quads[i]);
+         }
+
+         if (qs->softpipe->depth_stencil->stencil[0].enabled) {
+            if (!depth_stencil_test_quad(qs, &data, quads[i]))
+               continue;
+         }
+         else {
+            if (!depth_test_quad(qs, &data, quads[i]))
+               continue;
+         }
+
+         if (qs->softpipe->depth_stencil->stencil[0].enabled ||
+             qs->softpipe->depth_stencil->depth.writemask)
+            write_depth_stencil_values(&data, quads[i]);
+
+         qs->softpipe->occlusion_count += mask_count[quads[i]->inout.mask];
          quads[pass++] = quads[i];
+      }
+
+      nr = pass;
    }
-   
-   if (pass)
-      qs->next->run(qs->next, quads, pass);
+
+   if (nr)
+      qs->next->run(qs->next, quads, nr);
 }