fix CI build error

Author: linqun

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

@@ -793,31 +793,31 @@ struct GridwiseGemm_xdl_cshuffle_v3_b_preshuffle
                        Number<BK1Value>{})); //??? BK1Value same as KPack?
     }
 
-    using BlockwiseGemmPipe =
-        remove_cvref_t<decltype(BlockGemmBPreshufflePipeline_Selector<
-                                BlkGemmPipelineVer,
-                                BlkGemmPipeSched,
-                                BlockSize,
-                                ADataType,
-                                BDataType,
-                                ComputeTypeA,
-                                AccDataType,
-                                decltype(GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()),
-                                decltype(GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()),
-                                decltype(MakeAMmaTileDescriptor_M0_M1_M2_K(
-                                    GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1())),
-                                decltype(MakeBMmaTileDescriptor_N0_N1_N2_K(
-                                    GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1())),
-                                ABlockTransferSrcScalarPerVector,
-                                BBlockTransferSrcScalarPerVector,
-                                MPerBlock,
-                                NPerBlock,
-                                KPerBlock,
-                                MPerXdl,
-                                NPerXdl,
-                                MXdlPerWave,
-                                NXdlPerWave,
-                                KPack>())>;
+    using BlockwiseGemmPipe = remove_cvref_t<
+        decltype(BlockGemmBPreshufflePipeline_Selector<
+                 BlkGemmPipelineVer,
+                 BlkGemmPipeSched,
+                 BlockSize,
+                 ADataType,
+                 BDataType,
+                 ComputeTypeA,
+                 AccDataType,
+                 decltype(GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1(get_device_arch())),
+                 decltype(GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()),
+                 decltype(MakeAMmaTileDescriptor_M0_M1_M2_K(
+                     GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1(get_device_arch()))),
+                 decltype(MakeBMmaTileDescriptor_N0_N1_N2_K(
+                     GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1())),
+                 ABlockTransferSrcScalarPerVector,
+                 BBlockTransferSrcScalarPerVector,
+                 MPerBlock,
+                 NPerBlock,
+                 KPerBlock,
+                 MPerXdl,
+                 NPerXdl,
+                 MXdlPerWave,
+                 NXdlPerWave,
+                 KPack>())>;
 
     IS_VALID_COMPILATION_PARAMETER_IMPL(CDataType)