Merge branch 'develop' into uif2-initial

library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/CMakeLists.txt

 set(GROUPED_CONV1D_BWD_WEIGHT
-    device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f16_instance.cpp
-    device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f32_instance.cpp
-    device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp)
+    xdl/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f16_instance.cpp
+    xdl/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f32_instance.cpp
+    xdl/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp)
 
 if(DL_KERNELS)
     list(APPEND GROUPED_CONV1D_BWD_WEIGHT
-        device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_f16_instance.cpp
-        device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_f32_instance.cpp
-        device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_bf16_instance.cpp
-        device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_f16_instance.cpp
-        device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_f32_instance.cpp
-        device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_bf16_instance.cpp)
+        dl/device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_f16_instance.cpp
+        dl/device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_f32_instance.cpp
+        dl/device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_bf16_instance.cpp
+        dl/device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_f16_instance.cpp
+        dl/device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_f32_instance.cpp
+        dl/device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_bf16_instance.cpp)
 endif()
 
 add_instance_library(device_grouped_conv1d_bwd_weight_instance ${GROUPED_CONV1D_BWD_WEIGHT})

library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/CMakeLists.txt

 set(GROUPED_CONV2D_BWD_WEIGHT
-    device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
-    device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
-    device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
-    device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
-    device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
-    device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp)
+    xdl/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
+    xdl/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
+    xdl/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
+    xdl/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
+    xdl/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
+    xdl/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp)
 
 if(DL_KERNELS)
     list(APPEND GROUPED_CONV2D_BWD_WEIGHT
-        device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
-        device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
-        device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
-        device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
-        device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
-        device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp)
+        dl/device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
+        dl/device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
+        dl/device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
+        dl/device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
+        dl/device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
+        dl/device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp)
 endif()
 
 add_instance_library(device_grouped_conv2d_bwd_weight_instance ${GROUPED_CONV2D_BWD_WEIGHT})