fix bug

include/ck/tensor_operation/cpu/grid/gridwise_gemm_bias_activation_add_avx2.hpp

@@ -623,10 +623,6 @@ struct GridwiseGemmBiasActivationAddAvx2_MxN
                             c_threadwise_copy.SetSrc2SliceOrigin(c_block_desc,
                                                                  GetCIndex(i_mc, i_nc));
 
-                            _mm_prefetch(reinterpret_cast<const float*>(c1_grid_buf.p_data_) +
-                                             c_threadwise_copy.src2_offset,
-                                         _MM_HINT_T1);
-
                             if constexpr(!UseCLocalBuffer)
                             {
                                 c_threadwise_copy.SetSrcSliceOrigin(c_block_desc,

include/ck/tensor_operation/cpu/thread/threadwise_tensor_slice_transfer_avx2_specialization.hpp

@@ -2403,7 +2403,7 @@ struct ThreadwiseTensorSliceTransferAvx2Specialization_MatC_Store_Bias_Residual_
     // dst_slice_origin_step_idx need to be known at compile-time, for performance reason
     void MoveDstSliceWindow(const DstDesc&, const Index&) {}
 
-    // private:
+    private:
     const ElementwiseOperation element_op_;
 
     intptr_t i_dst_gemm_m;