fix bug, function pass.

include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_v1.hpp

@@ -365,10 +365,10 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v1<BlockGemmPipelineScheduler::I
                     // B VGPR->VGPR dequant
                     b_thread_dequant_copy.Run(b_block_desc_n0_n1_k0_k1,
                                               b_block_origin_idx,
-                                              b_thread_bufs(mfma_reg_buf),
+                                              b_thread_bufs(local_read_buf),
                                               b_thread_desc_,
                                               make_tuple(I0, I0, I0, I0),
-                                              b_thread_dequant_bufs(mfma_reg_buf));
+                                              b_thread_dequant_bufs(local_read_buf));
 
                     HotLoopScheduler();
                     __builtin_amdgcn_sched_barrier(0);

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_b_preshuffle.hpp

@@ -1620,7 +1620,7 @@ struct GridwiseGemm_xdl_cshuffle_v3_b_preshuffle
             decltype(b_block_desc_bk0_n_bk1),
             tensor_operation::element_wise::PassThrough,
             Sequence<Number<NXdlPerWave>{}, I1, Number<KRepeat>{}, Number<BK1Value>{}>,
-            Sequence<0, 1, 2, 3>,
+            Sequence<1, 2, 0, 3>,
             3,
             BK1Number>(b_element_op);