[965] Improve performance by allocating CURBE buffers a page at a time.

Since each one is only 64b, and kernel allocations are a page anyway, this
lets us reduce buffer allocation by packing many CURBEs into one buffer, for
each batchbuffer submitted.  Improves openarena performance by around 10%.

1 files changed, 3 insertions, 0 deletions

diff --git a/src/mesa/drivers/dri/i965/brw_vtbl.c b/src/mesa/drivers/dri/i965/brw_vtbl.c
index e9fed4dae1..126e655839 100644
--- a/src/mesa/drivers/dri/i965/brw_vtbl.c
+++ b/src/mesa/drivers/dri/i965/brw_vtbl.c
@@ -94,6 +94,9 @@ static void brw_new_batch( struct intel_context *intel )
    /* Check that we didn't just wrap our batchbuffer at a bad time. */
    assert(!brw->no_batch_wrap);
 
+   dri_bo_unreference(brw->curbe.curbe_bo);
+   brw->curbe.curbe_bo = NULL;
+
    /* Mark all context state as needing to be re-emitted.
     * This is probably not as severe as on 915, since almost all of our state
     * is just in referenced buffers.