[Mesa-dev] [PATCH 1/2] i965: Implement guardband clipping on Sandybridge.

[Kenneth Graunke]

Improves performance in Citybench:
- 320x240:  19.8008% +/- 0.937818%
- 1280x480: 6.53856% +/- 0.859083%

No apparent difference in OpenArena nor Xonotic.

Signed-off-by: Kenneth Graunke <kenneth at whitecape.org>
---
 src/mesa/drivers/dri/i965/gen6_clip_state.c     |    1 +
 src/mesa/drivers/dri/i965/gen6_viewport_state.c |   24 +++++++++++++----------
 2 files changed, 15 insertions(+), 10 deletions(-)

diff --git a/src/mesa/drivers/dri/i965/gen6_clip_state.c b/src/mesa/drivers/dri/i965/gen6_clip_state.c
index abf1d76..5d88978 100644
--- a/src/mesa/drivers/dri/i965/gen6_clip_state.c
+++ b/src/mesa/drivers/dri/i965/gen6_clip_state.c
@@ -74,6 +74,7 @@ upload_clip_state(struct brw_context *brw)
 	     GEN6_CLIP_MODE_NORMAL |
              nonperspective_barycentric_enable_flag |
 	     GEN6_CLIP_XY_TEST |
+	     GEN6_CLIP_GB_TEST |
 	     userclip << GEN6_USER_CLIP_CLIP_DISTANCES_SHIFT |
 	     depth_clamp |
 	     provoking);
diff --git a/src/mesa/drivers/dri/i965/gen6_viewport_state.c b/src/mesa/drivers/dri/i965/gen6_viewport_state.c
index f787ac7..95243f0 100644
--- a/src/mesa/drivers/dri/i965/gen6_viewport_state.c
+++ b/src/mesa/drivers/dri/i965/gen6_viewport_state.c
@@ -33,31 +33,35 @@
 
 /* The clip VP defines the guardband region where expensive clipping is skipped
  * and fragments are allowed to be generated and clipped out cheaply by the SF.
- *
- * By setting it to NDC bounds of [-1,1], we don't do GB clipping.  It's
- * supposed to cause seams to become visible in apps due to shared edges taking
- * different clip/no clip paths depending on whether the rest of the prim ends
- * up in the guardband or not.
  */
 static void
 gen6_upload_clip_vp(struct brw_context *brw)
 {
+   struct gl_context *ctx = &brw->intel.ctx;
    struct brw_clipper_viewport *vp;
 
    vp = brw_state_batch(brw, AUB_TRACE_CLIP_VP_STATE,
 			sizeof(*vp), 32, &brw->clip.vp_offset);
 
-   vp->xmin = -1.0;
-   vp->xmax = 1.0;
-   vp->ymin = -1.0;
-   vp->ymax = 1.0;
+   /* According to the Sandybridge PRM, Volume 2, Part 1, Section 6.3.8
+    * "Vertex X,Y Clamping and Quantization", the screen-aligned 2D
+    * bounding-box of an object must not exceed 16K pixels in either X or Y.
+    */
+   const float maximum_post_clamp_delta = 16384;
+   float gbx = maximum_post_clamp_delta / (float) ctx->Viewport.Width;
+   float gby = maximum_post_clamp_delta / (float) ctx->Viewport.Height;
+
+   vp->xmin = -gbx;
+   vp->xmax = gbx;
+   vp->ymin = -gby;
+   vp->ymax = gby;
 
    brw->state.dirty.cache |= CACHE_NEW_CLIP_VP;
 }
 
 const struct brw_tracked_state gen6_clip_vp = {
    .dirty = {
-      .mesa = 0,
+      .mesa = _NEW_VIEWPORT,
       .brw = BRW_NEW_BATCH,
       .cache = 0,
    },
-- 
1.7.10.1