[CK_TILE] Update the file structure part2

include/ck_tile/host/ops/fused_moe/kernel/fused_moegemm_kernel.hpp

@@ -5,7 +5,7 @@
 
 #include "ck_tile/core.hpp"
 #include "ck_tile/device/ops/common.hpp"
-#include "ck_tile/device/ops/elementwise.hpp"
+#include "ck_tile/host/ops/elementwise.hpp"
 #include <string>
 #include <type_traits>
 

include/ck_tile/host/ops/fused_moe/kernel/moe_sorting_kernel.hpp

@@ -5,7 +5,7 @@
 
 #include "ck_tile/core.hpp"
 #include "ck_tile/device/ops/common.hpp"
-#include "ck_tile/device/ops/elementwise.hpp"
+#include "ck_tile/host/ops/elementwise.hpp"
 #include "ck_tile/host/util/hip_check_error.hpp"
 #include <string>
 #include <type_traits>

include/ck_tile/host/ops/fused_moe/pipeline/fused_moegemm_pipeline_flatmm_ex.hpp

@@ -5,7 +5,7 @@
 
 #include "ck_tile/core.hpp"
 #include "ck_tile/device/ops/common/tensor_layout.hpp"
-#include "ck_tile/device/ops/fused_moe/pipeline/fused_moegemm_pipeline_flatmm_policy.hpp"
+#include "ck_tile/host/ops/fused_moe/pipeline/fused_moegemm_pipeline_flatmm_policy.hpp"
 
 namespace ck_tile {
 

include/ck_tile/host/ops/fused_moe/pipeline/fused_moegemm_pipeline_flatmm_policy.hpp

@@ -4,8 +4,8 @@
 #pragma once
 
 #include "ck_tile/core.hpp"
-#include "ck_tile/device/ops/fused_moe/pipeline/fused_moegemm_traits.hpp"
-#include "ck_tile/device/ops/flatmm.hpp"
+#include "ck_tile/host/ops/fused_moe/pipeline/fused_moegemm_traits.hpp"
+#include "ck_tile/host/ops/flatmm.hpp"
 #include "ck_tile/device/ops/gemm/warp/warp_gemm.hpp"
 #include "ck_tile/device/ops/gemm/warp/warp_gemm_dispatcher.hpp"
 

include/ck_tile/host/ops/fused_moe/pipeline/fused_moegemm_pipeline_flatmm_uk.hpp

@@ -5,7 +5,7 @@
 
 #include "ck_tile/core.hpp"
 #include "ck_tile/device/ops/common/tensor_layout.hpp"
-#include "ck_tile/device/ops/fused_moe/pipeline/fused_moegemm_pipeline_flatmm_policy.hpp"
+#include "ck_tile/host/ops/fused_moe/pipeline/fused_moegemm_pipeline_flatmm_policy.hpp"
 
 namespace ck_tile {
 

include/ck_tile/host/ops/fused_moe/pipeline/moe_sorting_pipeline.hpp

@@ -4,7 +4,7 @@
 #pragma once
 
 #include "ck_tile/core.hpp"
-#include "ck_tile/device/ops/fused_moe/pipeline/moe_sorting_policy.hpp"
+#include "ck_tile/host/ops/fused_moe/pipeline/moe_sorting_policy.hpp"
 #include <string>
 #include <type_traits>
 

include/ck_tile/host/ops/fused_moe/pipeline/moe_sorting_policy.hpp

@@ -4,8 +4,8 @@
 #pragma once
 
 #include "ck_tile/core.hpp"
-#include "ck_tile/device/ops/softmax.hpp"
-#include "ck_tile/device/ops/topk.hpp"
+#include "ck_tile/host/ops/softmax.hpp"
+#include "ck_tile/host/ops/topk.hpp"
 
 namespace ck_tile {
 

include/ck_tile/host/ops/norm_reduce/block/block_norm_reduce.hpp

@@ -4,7 +4,7 @@
 #pragma once
 
 #include "ck_tile/core.hpp"
-#include "ck_tile/device/ops/norm_reduce/thread/thread_welford.hpp"
+#include "ck_tile/host/ops/norm_reduce/thread/thread_welford.hpp"
 
 namespace ck_tile {
 

include/ck_tile/host/ops/permute/kernel/generic_permute_kernel.hpp

@@ -5,7 +5,7 @@
 
 #include "ck_tile/core.hpp"
 #include "ck_tile/device/ops/common.hpp"
-// #include "ck_tile/device/ops/permute/pipeline/generic_petmute_problem.hpp"
+// #include "ck_tile/host/ops/permute/pipeline/generic_petmute_problem.hpp"
 
 namespace ck_tile {
 

include/ck_tile/host/ops/rmsnorm2d/pipeline/rmsnorm2d_fwd_pipeline_one_pass.hpp

@@ -4,7 +4,7 @@
 #pragma once
 
 #include "ck_tile/core.hpp"
-#include "ck_tile/device/ops/rmsnorm2d/pipeline/rmsnorm2d_fwd_pipeline_default_policy.hpp"
+#include "ck_tile/host/ops/rmsnorm2d/pipeline/rmsnorm2d_fwd_pipeline_default_policy.hpp"
 #include <string>
 #include <type_traits>
 

include/ck_tile/host/ops/rmsnorm2d/pipeline/rmsnorm2d_fwd_pipeline_two_pass.hpp

@@ -4,7 +4,7 @@
 #pragma once
 
 #include "ck_tile/core.hpp"
-#include "ck_tile/device/ops/rmsnorm2d/pipeline/rmsnorm2d_fwd_pipeline_default_policy.hpp"
+#include "ck_tile/host/ops/rmsnorm2d/pipeline/rmsnorm2d_fwd_pipeline_default_policy.hpp"
 #include <string>
 #include <type_traits>
 

include/ck_tile/host/ops/smoothquant/pipeline/smoothquant_pipeline_one_pass.hpp

@@ -4,7 +4,7 @@
 #pragma once
 
 #include "ck_tile/core.hpp"
-#include "ck_tile/device/ops/rmsnorm2d/pipeline/rmsnorm2d_fwd_pipeline_default_policy.hpp"
+#include "ck_tile/host/ops/rmsnorm2d/pipeline/rmsnorm2d_fwd_pipeline_default_policy.hpp"
 #include <string>
 #include <type_traits>
 

include/ck_tile/host/ops/smoothquant/pipeline/smoothquant_pipeline_two_pass.hpp

@@ -4,7 +4,7 @@
 #pragma once
 
 #include "ck_tile/core.hpp"
-#include "ck_tile/device/ops/rmsnorm2d/pipeline/rmsnorm2d_fwd_pipeline_default_policy.hpp"
+#include "ck_tile/host/ops/rmsnorm2d/pipeline/rmsnorm2d_fwd_pipeline_default_policy.hpp"
 #include <string>
 #include <type_traits>
 

include/ck_tile/host/ops/topk_softmax/kernel/topk_softmax_kernel.hpp

@@ -5,7 +5,7 @@
 
 #include "ck_tile/core.hpp"
 #include "ck_tile/device/ops/common.hpp"
-#include "ck_tile/device/ops/elementwise.hpp"
+#include "ck_tile/host/ops/elementwise.hpp"
 #include "ck_tile/host/util/hip_check_error.hpp"
 #include <string>
 #include <type_traits>

include/ck_tile/host/ops/topk_softmax/pipeline/topk_softmax_warp_per_row_pipeline.hpp

@@ -4,7 +4,7 @@
 #pragma once
 
 #include "ck_tile/core.hpp"
-#include "ck_tile/device/ops/topk_softmax/pipeline/topk_softmax_warp_per_row_policy.hpp"
+#include "ck_tile/host/ops/topk_softmax/pipeline/topk_softmax_warp_per_row_policy.hpp"
 #include <string>
 #include <type_traits>
 

include/ck_tile/host/ops/topk_softmax/pipeline/topk_softmax_warp_per_row_policy.hpp

@@ -4,8 +4,8 @@
 #pragma once
 
 #include "ck_tile/core.hpp"
-#include "ck_tile/device/ops/softmax.hpp"
-#include "ck_tile/device/ops/topk.hpp"
+#include "ck_tile/host/ops/softmax.hpp"
+#include "ck_tile/host/ops/topk.hpp"
 
 namespace ck_tile {
 

test/ck_tile/image_to_column/test_tile_image_to_column.cpp

@@ -7,7 +7,7 @@
 #include "ck_tile/host.hpp"
 #include "ck_tile/core.hpp"
 #include "ck_tile/host/util/kernel_launch.hpp"
-#include "ck_tile/device/ops/image_to_column.hpp"
+#include "ck_tile/host/ops/image_to_column.hpp"
 
 // Host API implementation
 template <typename DataType>