brw: switch mesh/task URB fence prior to EOT to GPU

Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Cc: mesa-stable
Reviewed-by: Sagar Ghuge <sagar.ghuge@intel.com>
Reviewed-by: Rohan Garg <rohan.garg@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/30849>
(cherry picked from commit 93fba40389)

.pick_status.json

@@ -524,7 +524,7 @@
         "description": "brw: switch mesh/task URB fence prior to EOT to GPU",
         "nominated": true,
         "nomination_type": 0,
-        "resolution": 0,
+        "resolution": 1,
         "main_sha": null,
         "because_sha": null,
         "notes": null

src/intel/compiler/brw_fs_visitor.cpp

@@ -964,7 +964,18 @@ fs_visitor::emit_urb_fence()
                              brw_imm_ud(true),
                              brw_imm_ud(0));
    fence->sfid = BRW_SFID_URB;
-   fence->desc = lsc_fence_msg_desc(devinfo, LSC_FENCE_LOCAL,
+   /* The logical thing here would likely be a THREADGROUP fence but that's
+    * still failing some tests like in dEQP-VK.mesh_shader.ext.query.*
+    *
+    * Gfx12.5 has a comment about this on BSpec 53533 :
+    *
+    *    "If fence scope is Local or Threadgroup, HW ignores the flush type
+    *     and operates as if it was set to None (no flush)"
+    *
+    * Software workaround from HSD-22014129519 indicates that a GPU fence
+    * resolves the issue.
+    */
+   fence->desc = lsc_fence_msg_desc(this->devinfo, LSC_FENCE_GPU,
                                     LSC_FLUSH_TYPE_NONE, true);
 
    bld.exec_all().group(1, 0).emit(FS_OPCODE_SCHEDULING_FENCE,