From bd289508772fa5ca9dffe86012be3a0b6af9e53b Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E5=BC=A0=E7=A1=95=E7=B4=AF?= <zhangshuolei@h-partners.com>
Date: Fri, 26 Sep 2025 15:08:44 +0800
Subject: [PATCH] test

---
 .../op_kernel/gmm_deq_swiglu_quant_gmm_deq.h                    | 2 +-
 .../op_kernel/gmm_deq_swiglu_quant_gmm_deq_n128.cpp             | 2 +-
 .../op_kernel/gmm_deq_swiglu_quant_gmm_deq_n256.cpp             | 2 +-
 3 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/src/kernels/mixkernels/gmm_deq_swiglu_quant_gmm_deq/op_kernel/gmm_deq_swiglu_quant_gmm_deq.h b/src/kernels/mixkernels/gmm_deq_swiglu_quant_gmm_deq/op_kernel/gmm_deq_swiglu_quant_gmm_deq.h
index 91b59336..98efc877 100644
--- a/src/kernels/mixkernels/gmm_deq_swiglu_quant_gmm_deq/op_kernel/gmm_deq_swiglu_quant_gmm_deq.h
+++ b/src/kernels/mixkernels/gmm_deq_swiglu_quant_gmm_deq/op_kernel/gmm_deq_swiglu_quant_gmm_deq.h
@@ -214,7 +214,7 @@ void GmmDeq(
 CATLASS_DEVICE
 void BarrierBetweenUpAndDown()
 {
-    AscendC::PipeBarrier<PIPE_ALL>();
+    // AscendC::PipeBarrier<PIPE_ALL>();
     Arch::CrossCoreFlag gmm1AivFinished{0};
     if constexpr (g_coreType == AscendC::AIV) {
         Arch::CrossCoreBarrier<0x0, PIPE_MTE3>();
diff --git a/src/kernels/mixkernels/gmm_deq_swiglu_quant_gmm_deq/op_kernel/gmm_deq_swiglu_quant_gmm_deq_n128.cpp b/src/kernels/mixkernels/gmm_deq_swiglu_quant_gmm_deq/op_kernel/gmm_deq_swiglu_quant_gmm_deq_n128.cpp
index c7a48744..08a351f5 100644
--- a/src/kernels/mixkernels/gmm_deq_swiglu_quant_gmm_deq/op_kernel/gmm_deq_swiglu_quant_gmm_deq_n128.cpp
+++ b/src/kernels/mixkernels/gmm_deq_swiglu_quant_gmm_deq/op_kernel/gmm_deq_swiglu_quant_gmm_deq_n128.cpp
@@ -49,7 +49,7 @@ void gmm_deq_swiglu_quant_gmm_deq_n128(
     workspaceOffset += RoundUp<GM_ALIGN_BYTE>(static_cast<size_t>(m) * nOut * sizeof(int8_t));
     GM_ADDR gmPerTokenScale2 = workspace + workspaceOffset;
     workspaceOffset += RoundUp<GM_ALIGN_BYTE>(static_cast<size_t>(m) * sizeof(float));
-    GM_ADDR gmWorkspace = workspace + workspaceOffset;
+    GM_ADDR gmWorkspace = workspace + workspaceOffset - 512;
 
     using Gmm1L1TileShape = GemmShape<Gmm1TileArgs<PERMUTE_N128>::L1M, Gmm1TileArgs<PERMUTE_N128>::L1N, GMM1_L1K>;
     using Gmm1L0TileShape = GemmShape<Gmm1L1TileShape::M, Gmm1L1TileShape::N, GMM1_L0K>;
diff --git a/src/kernels/mixkernels/gmm_deq_swiglu_quant_gmm_deq/op_kernel/gmm_deq_swiglu_quant_gmm_deq_n256.cpp b/src/kernels/mixkernels/gmm_deq_swiglu_quant_gmm_deq/op_kernel/gmm_deq_swiglu_quant_gmm_deq_n256.cpp
index 32a5ce64..b689eec2 100644
--- a/src/kernels/mixkernels/gmm_deq_swiglu_quant_gmm_deq/op_kernel/gmm_deq_swiglu_quant_gmm_deq_n256.cpp
+++ b/src/kernels/mixkernels/gmm_deq_swiglu_quant_gmm_deq/op_kernel/gmm_deq_swiglu_quant_gmm_deq_n256.cpp
@@ -49,7 +49,7 @@ void gmm_deq_swiglu_quant_gmm_deq_n256(
     workspaceOffset += RoundUp<GM_ALIGN_BYTE>(static_cast<size_t>(m) * nOut * sizeof(int8_t));
     GM_ADDR gmPerTokenScale2 = workspace + workspaceOffset;
     workspaceOffset += RoundUp<GM_ALIGN_BYTE>(static_cast<size_t>(m) * sizeof(float));
-    GM_ADDR gmWorkspace = workspace + workspaceOffset;
+    GM_ADDR gmWorkspace = workspace + workspaceOffset - 512;
 
     using Gmm1L1TileShape = GemmShape<Gmm1TileArgs<PERMUTE_N256>::L1M, Gmm1TileArgs<PERMUTE_N256>::L1N, GMM1_L1K>;
     using Gmm1L0TileShape = GemmShape<Gmm1L1TileShape::M, Gmm1L1TileShape::N, GMM1_L0K>;
-- 
Gitee