llvmpipe: fastpath for interpolated z16 less depthtesting

Because this is interpolated (ie. early) depth, we can build in an
assumption about the quads emitted by triangle setup, ie that they
are actually linear spans.  Interpolate z over those spans in z16
format to save on math & conversion.

src/gallium/drivers/llvmpipe/lp_quad_depth_test.c

@@ -646,9 +646,9 @@ static unsigned mask_count[0x8] =
 
 
 static void
-depth_test_quads(struct quad_stage *qs, 
-                 struct quad_header *quads[],
-                 unsigned nr)
+depth_test_quads_fallback(struct quad_stage *qs, 
+                          struct quad_header *quads[],
+                          unsigned nr)
 {
    unsigned i, pass = 0;
    const struct lp_fragment_shader *fs = qs->llvmpipe->fs;
@@ -704,9 +704,144 @@ depth_test_quads(struct quad_stage *qs,
       qs->next->run(qs->next, quads, nr);
 }
 
+/* XXX: this function assumes setup function actually emits linear
+ * spans of quads.  It seems a lot more natural to do (early)
+ * depth-testing on spans rather than quads.
+ */
+static void
+depth_interp_z16_less_write(struct quad_stage *qs, 
+                            struct quad_header *quads[],
+                            unsigned nr)
+{
+   unsigned i, pass = 0;
+   const unsigned ix = quads[0]->input.x0;
+   const unsigned iy = quads[0]->input.y0;
+   const float fx = (float) ix;
+   const float fy = (float) iy;
+   const float dzdx = quads[0]->posCoef->dadx[2];
+   const float dzdy = quads[0]->posCoef->dady[2];
+   const float z0 = quads[0]->posCoef->a0[2] + dzdx * fx + dzdy * fy;
+   struct llvmpipe_cached_tile *tile;
+   ushort (*depth16)[TILE_SIZE];
+   ushort idepth[4], depth_step;
+   const float scale = 65535.0;
+
+   idepth[0] = (ushort)((z0) * scale);
+   idepth[1] = (ushort)((z0 + dzdx) * scale);
+   idepth[2] = (ushort)((z0 + dzdy) * scale);
+   idepth[3] = (ushort)((z0 + dzdx + dzdy) * scale);
+
+   depth_step = (ushort)(dzdx * 2 * scale);
+
+   tile = lp_get_cached_tile(qs->llvmpipe->zsbuf_cache, ix, iy);
+
+   depth16 = (ushort (*)[TILE_SIZE])
+      &tile->data.depth16[iy % TILE_SIZE][ix % TILE_SIZE];
+
+   for (i = 0; i < nr; i++) {
+      unsigned outmask = quads[i]->inout.mask;
+      unsigned mask = 0;
+      
+      if ((outmask & 1) && idepth[0] < depth16[0][0]) {
+         depth16[0][0] = idepth[0];
+         mask |= (1 << 0);
+      }
+
+      if ((outmask & 2) && idepth[1] < depth16[0][1]) {
+         depth16[0][1] = idepth[1];
+         mask |= (1 << 1);
+      }
+
+      if ((outmask & 4) && idepth[2] < depth16[1][0]) {
+         depth16[1][0] = idepth[2];
+         mask |= (1 << 2);
+      }
+
+      if ((outmask & 8) && idepth[3] < depth16[1][1]) {
+         depth16[1][1] = idepth[3];
+         mask |= (1 << 3);
+      }
+
+      idepth[0] += depth_step;
+      idepth[1] += depth_step;
+      idepth[2] += depth_step;
+      idepth[3] += depth_step;
+
+      depth16 = (ushort (*)[TILE_SIZE]) &depth16[0][2];
+
+      quads[i]->inout.mask = mask;
+      if (quads[i]->inout.mask)
+         quads[pass++] = quads[i];
+   }
+
+   if (pass)
+      qs->next->run(qs->next, quads, pass);
+
+}
+
+
+static void
+depth_noop(struct quad_stage *qs, 
+           struct quad_header *quads[],
+           unsigned nr)
+{
+   qs->next->run(qs->next, quads, nr);
+}
+
+
+
+static void
+choose_depth_test(struct quad_stage *qs, 
+                  struct quad_header *quads[],
+                  unsigned nr)
+{
+   boolean interp_depth = !qs->llvmpipe->fs->info.writes_z;
+
+   boolean alpha = qs->llvmpipe->depth_stencil->alpha.enabled;
+
+   boolean depth = (qs->llvmpipe->framebuffer.zsbuf && 
+                    qs->llvmpipe->depth_stencil->depth.enabled);
+
+   unsigned depthfunc = qs->llvmpipe->depth_stencil->depth.func;
+
+   boolean stencil = qs->llvmpipe->depth_stencil->stencil[0].enabled;
+
+   boolean depthwrite = qs->llvmpipe->depth_stencil->depth.writemask;
+
+
+   qs->run = depth_test_quads_fallback;
+
+   if (!alpha &&
+       !depth &&
+       !stencil) {
+      qs->run = depth_noop;
+   }
+   else if (!alpha && 
+            interp_depth && 
+            depth && 
+            depthfunc == PIPE_FUNC_LESS && 
+            depthwrite && 
+            !stencil) 
+   {
+      switch (qs->llvmpipe->framebuffer.zsbuf->format) {
+      case PIPE_FORMAT_Z16_UNORM:
+         qs->run = depth_interp_z16_less_write;
+         break;
+      default:
+         break;
+      }
+   }
+
+   qs->run( qs, quads, nr );
+}
+
+
+
+
 
 static void depth_test_begin(struct quad_stage *qs)
 {
+   qs->run = choose_depth_test;
    qs->next->begin(qs->next);
 }
 
@@ -723,7 +858,7 @@ struct quad_stage *lp_quad_depth_test_stage( struct llvmpipe_context *llvmpipe )
 
    stage->llvmpipe = llvmpipe;
    stage->begin = depth_test_begin;
-   stage->run = depth_test_quads;
+   stage->run = choose_depth_test;
    stage->destroy = depth_test_destroy;
 
    return stage;