iris: Add Tile Cache Flush for Unified Cache.

src/gallium/drivers/iris/iris_context.h

@@ -222,6 +222,7 @@ enum pipe_control_flags
    PIPE_CONTROL_STATE_CACHE_INVALIDATE          = (1 << 22),
    PIPE_CONTROL_STALL_AT_SCOREBOARD             = (1 << 23),
    PIPE_CONTROL_DEPTH_CACHE_FLUSH               = (1 << 24),
+   PIPE_CONTROL_TILE_CACHE_FLUSH                = (1 << 25),
 };
 
 #define PIPE_CONTROL_CACHE_FLUSH_BITS \

src/gallium/drivers/iris/iris_state.c

@@ -6788,6 +6788,23 @@ iris_emit_raw_pipe_control(struct iris_batch *batch,
       flags |= PIPE_CONTROL_CS_STALL;
    }
 
+   if (GEN_GEN >= 12 && ((flags & PIPE_CONTROL_RENDER_TARGET_FLUSH) ||
+                         (flags & PIPE_CONTROL_DEPTH_CACHE_FLUSH))) {
+      /* From the PIPE_CONTROL instruction table, bit 28 (Tile Cache Flush
+       * Enable):
+       *
+       *    Unified Cache (Tile Cache Disabled):
+       *
+       *    When the Color and Depth (Z) streams are enabled to be cached in
+       *    the DC space of L2, Software must use "Render Target Cache Flush
+       *    Enable" and "Depth Cache Flush Enable" along with "Tile Cache
+       *    Flush" for getting the color and depth (Z) write data to be
+       *    globally observable.  In this mode of operation it is not required
+       *    to set "CS Stall" upon setting "Tile Cache Flush" bit.
+       */
+      flags |= PIPE_CONTROL_TILE_CACHE_FLUSH;
+   }
+
    if (GEN_GEN == 9 && devinfo->gt == 4) {
       /* TODO: The big Skylake GT4 post sync op workaround */
    }
@@ -6910,6 +6927,9 @@ iris_emit_raw_pipe_control(struct iris_batch *batch,
    }
 
    iris_emit_cmd(batch, GENX(PIPE_CONTROL), pc) {
+#if GEN_GEN >= 12
+      pc.TileCacheFlushEnable = flags & PIPE_CONTROL_TILE_CACHE_FLUSH;
+#endif
       pc.LRIPostSyncOperation = NoLRIOperation;
       pc.PipeControlFlushEnable = flags & PIPE_CONTROL_FLUSH_ENABLE;
       pc.DCFlushEnable = flags & PIPE_CONTROL_DATA_CACHE_FLUSH;