From c87fea5c483f7649ccea3698f24c427202d40f09 Mon Sep 17 00:00:00 2001
From: abnerhexu <hexu@rochexp.com>
Date: Wed, 27 May 2026 09:07:47 +0800
Subject: [PATCH] Keep masked barriers local under cluster gbar

---
 hw/rtl/core/VX_tensor_ctrl_unit.sv | 2 +-
 hw/rtl/core/VX_wctl_unit.sv        | 6 +++---
 2 files changed, 4 insertions(+), 4 deletions(-)

diff --git a/hw/rtl/core/VX_tensor_ctrl_unit.sv b/hw/rtl/core/VX_tensor_ctrl_unit.sv
index 50f47095..e70c0f1d 100644
--- a/hw/rtl/core/VX_tensor_ctrl_unit.sv
+++ b/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
diff --git a/hw/rtl/core/VX_wctl_unit.sv b/hw/rtl/core/VX_wctl_unit.sv
index b0f64f72..8e946f16 100644
--- a/hw/rtl/core/VX_wctl_unit.sv
+++ b/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