Keep masked barriers local under cluster gbar

hw/rtl/core/VX_tensor_ctrl_unit.sv

@@ -91,7 +91,7 @@ module VX_tensor_ctrl_unit import VX_gpu_pkg::*; #(
     assign barrier.id = req_rs1_data[`NB_WIDTH-1:0];
 `ifdef GBAR_ENABLE
 `ifdef GBAR_CLUSTER_ENABLE
-    assign barrier.is_global = 1'b1;
+    assign barrier.is_global = is_bar_mask ? 1'b0 : 1'b1;
 `else
     assign barrier.is_global = req_rs1_data[31];
 `endif

hw/rtl/core/VX_wctl_unit.sv

@@ -113,9 +113,9 @@ module VX_wctl_unit import VX_gpu_pkg::*; #(
     assign barrier.id       = rs1_data[`NB_WIDTH-1:0];
 `ifdef GBAR_ENABLE
 `ifdef GBAR_CLUSTER_ENABLE
-    // all barriers are cluster-wide barriers, which is implemented by
-    // modifying the global barrier logic
-    assign barrier.is_global = 1'b1;
+    // Explicit warp masks are core-local; rs2 carries the mask, not a warp
+    // count for the cluster global-barrier protocol.
+    assign barrier.is_global = is_bar_mask ? 1'b0 : 1'b1;
 `else
     assign barrier.is_global = rs1_data[31];
 `endif