diff --git a/impl/matmul/matmul_impl.h b/impl/matmul/matmul_impl.h
index 09d44c628e464b2ec1a91422f074e1733c7589ea..24e89bbc2348b4d4fb088161020a8520710b333b 100644
--- a/impl/matmul/matmul_impl.h
+++ b/impl/matmul/matmul_impl.h
@@ -7939,6 +7939,7 @@ __aicore__ inline void MatmulImpl<A_TYPE, B_TYPE, C_TYPE, BIAS_TYPE, MM_CFG, MM_
     }
     MatmulInstr::ssAmatrixTranspose1_ = var.isTransposeA_;
     MatmulInstr::ssBmatrixTranspose1_ = var.isTransposeB_;
+    LocalTensor<BiasT> bias;
     for (int k = 0; k < var.kIter_; k++) {                                       // start reduce K axis
         var.baseUseK_ = ToMatmulConfig(MM_CFG).basicK;                                       //  Disassemble into main tail block.
         constexpr uint16_t blockUseK = ToMatmulConfig(MM_CFG).basicK / c0Size_;
@@ -7958,7 +7959,6 @@ __aicore__ inline void MatmulImpl<A_TYPE, B_TYPE, C_TYPE, BIAS_TYPE, MM_CFG, MM_
         }
         if constexpr (BIAS_TYPE::isTrans) {
             if (k == 0) {
-                LocalTensor<BiasT> bias;
                 if constexpr (A_TYPE::layout == LayoutMode::NONE || ToMatmulConfig(MM_CFG).batchMode ==
                     BatchMode::SINGLE_LARGE_THAN_L1) {
                     bias = var.qidBias_.template DeQue<BiasT>();
@@ -7966,16 +7966,16 @@ __aicore__ inline void MatmulImpl<A_TYPE, B_TYPE, C_TYPE, BIAS_TYPE, MM_CFG, MM_
                     bias.SetAddr(var.inputBias_);
                     bias = bias[var.curN_ * var.tiling_.GetBaseN()];
                 }
-                MatmulInstr::Compute();
+                MatmulInstr::Compute(a1, b1, var.cMatrix_, bias);
                 if constexpr (A_TYPE::layout == LayoutMode::NONE || ToMatmulConfig(MM_CFG).batchMode ==
                     BatchMode::SINGLE_LARGE_THAN_L1) {
                     var.qidBias_.FreeTensor(bias);
                 }
             } else {
-                MatmulInstr::Compute();
+                MatmulInstr::Compute(a1, b1, var.cMatrix_, bias);
             }
         } else {
-            MatmulInstr::Compute();
+            MatmulInstr::Compute(a1, b1, var.cMatrix_, bias);
         }
         if constexpr (!PhyPosIsL1(A_TYPE::pos) && (A_TYPE::layout == LayoutMode::NONE ||
             ToMatmulConfig(MM_CFG).batchMode == BatchMode::SINGLE_LARGE_THAN_L1)) {