[CK_TILE] Update the file structure part2

example/ck_tile/04_img2col/image_to_column.hpp

@@ -5,7 +5,7 @@
 
 #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"
 #include <string>
 
 #define DefaultConvParams                                                    \

example/ck_tile/06_permute/permute.hpp

@@ -5,7 +5,7 @@
 
 #include "ck_tile/core.hpp"
 #include "ck_tile/host/util/kernel_launch.hpp"
-#include "ck_tile/device/ops/permute.hpp"
+#include "ck_tile/host/ops/permute.hpp"
 #include <string>
 
 struct permute_traits

example/ck_tile/09_topk_softmax/topk_softmax_api.hpp

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

example/ck_tile/10_rmsnorm2d/example_rmsnorm2d_fwd.cpp

 #include "ck_tile/host.hpp"
 #include "ck_tile/core.hpp"
 #include "ck_tile/host/util/kernel_launch.hpp"
-#include "ck_tile/device/ops/rmsnorm2d.hpp"
+#include "ck_tile/host/ops/rmsnorm2d.hpp"
 #include <cstring>
 
 auto create_args(int argc, char* argv[])

example/ck_tile/10_rmsnorm2d/rmsnorm2d_fwd.hpp

@@ -5,7 +5,7 @@
 
 #include "ck_tile/core.hpp"
 #include "ck_tile/host/util/kernel_launch.hpp"
-#include "ck_tile/device/ops/rmsnorm2d.hpp"
+#include "ck_tile/host/ops/rmsnorm2d.hpp"
 #include <string>
 
 template <typename DataType>

example/ck_tile/11_add_rmsnorm2d_rdquant/add_rmsnorm2d_rdquant_fwd.hpp

@@ -5,7 +5,7 @@
 
 #include "ck_tile/core.hpp"
 #include "ck_tile/host/util/kernel_launch.hpp"
-#include "ck_tile/device/ops/add_rmsnorm2d_rdquant.hpp"
+#include "ck_tile/host/ops/add_rmsnorm2d_rdquant.hpp"
 #include <string>
 
 template <typename DataType>

example/ck_tile/11_add_rmsnorm2d_rdquant/example_add_rmsnorm2d_rdquant_fwd.cpp

 #include "ck_tile/host.hpp"
 #include "ck_tile/core.hpp"
 #include "ck_tile/host/util/kernel_launch.hpp"
-#include "ck_tile/device/ops/add_rmsnorm2d_rdquant.hpp"
+#include "ck_tile/host/ops/add_rmsnorm2d_rdquant.hpp"
 #include <cstring>
 
 // different threshold for different dtype

example/ck_tile/12_smoothquant/example_smoothquant.cpp

 #include "ck_tile/host.hpp"
 #include "ck_tile/core.hpp"
 #include "ck_tile/host/util/kernel_launch.hpp"
-#include "ck_tile/device/ops/smoothquant.hpp"
+#include "ck_tile/host/ops/smoothquant.hpp"
 #include <cstring>
 
 // different threshold for different dtype

example/ck_tile/12_smoothquant/smoothquant.hpp

@@ -5,7 +5,7 @@
 
 #include "ck_tile/core.hpp"
 #include "ck_tile/host/util/kernel_launch.hpp"
-#include "ck_tile/device/ops/smoothquant.hpp"
+#include "ck_tile/host/ops/smoothquant.hpp"
 #include <string>
 
 template <typename DataType>

example/ck_tile/13_moe_sorting/moe_sorting_api.hpp

@@ -5,7 +5,7 @@
 #include <string>
 #include "ck_tile/core.hpp"
 #include "ck_tile/host.hpp"
-#include "ck_tile/device/ops/fused_moe.hpp"
+#include "ck_tile/host/ops/fused_moe.hpp"
 
 struct moe_sorting_trait
 {

example/ck_tile/14_moe_smoothquant/moe_smoothquant.hpp

@@ -5,7 +5,7 @@
 
 #include "ck_tile/core.hpp"
 #include "ck_tile/host/util/kernel_launch.hpp"
-#include "ck_tile/device/ops/smoothquant.hpp"
+#include "ck_tile/host/ops/smoothquant.hpp"
 #include <string>
 
 template <typename DataType>

example/ck_tile/15_fused_moe/fused_moegemm.hpp

@@ -5,7 +5,7 @@
 
 #include "ck_tile/core.hpp"
 #include "ck_tile/host/util/kernel_launch.hpp"
-#include "ck_tile/device/ops/fused_moe.hpp"
+#include "ck_tile/host/ops/fused_moe.hpp"
 #include <string>
 
 // this is only a convenient structure for creating an example

example/ck_tile/15_fused_moe/fused_moesorting.hpp

@@ -5,7 +5,7 @@
 #include <string>
 #include "ck_tile/core.hpp"
 #include "ck_tile/host.hpp"
-#include "ck_tile/device/ops/fused_moe.hpp"
+#include "ck_tile/host/ops/fused_moe.hpp"
 
 struct fused_moesorting_trait
 {

example/ck_tile/15_fused_moe/instances/fused_moegemm_api_internal.hpp

@@ -4,7 +4,7 @@
 #pragma once
 
 #include "fused_moegemm_api_traits.hpp"
-#include "ck_tile/device/ops/fused_moe.hpp"
+#include "ck_tile/host/ops/fused_moe.hpp"
 #include <iostream>
 
 template <ck_tile::index_t... Is>

include/ck_tile/device/ops/layernorm2d/pipeline/layernorm2d_fwd_pipeline_default_policy.hpp

@@ -4,8 +4,8 @@
 #pragma once
 
 #include "ck_tile/core.hpp"
-#include "ck_tile/device/ops/norm_reduce/block/block_norm_reduce_problem.hpp"
-#include "ck_tile/device/ops/norm_reduce/block/block_norm_reduce.hpp"
+#include "ck_tile/host/ops/norm_reduce/block/block_norm_reduce_problem.hpp"
+#include "ck_tile/host/ops/norm_reduce/block/block_norm_reduce.hpp"
 
 namespace ck_tile {
 

include/ck_tile/host/ops/add_rmsnorm2d_rdquant/pipeline/add_rmsnorm2d_rdquant_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/add_rmsnorm2d_rdquant/pipeline/add_rmsnorm2d_rdquant_fwd_pipeline_three_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/flatmm/block/flatmm_32x512x128_1x4x1_16x16x32.hpp

@@ -5,7 +5,7 @@
 
 #include "ck_tile/core.hpp"
 #include "ck_tile/device/ops/gemm/warp/warp_gemm.hpp"
-#include "ck_tile/device/ops/flatmm/block/flatmm_uk_config.hpp"
+#include "ck_tile/host/ops/flatmm/block/flatmm_uk_config.hpp"
 
 namespace ck_tile {
 

include/ck_tile/host/ops/flatmm/block/flatmm_sn_32x128x512_1x4x1_16x16x32.hpp

@@ -5,7 +5,7 @@
 
 #include "ck_tile/core.hpp"
 #include "ck_tile/device/ops/gemm/warp/warp_gemm.hpp"
-#include "ck_tile/device/ops/flatmm/block/flatmm_uk_config.hpp"
+#include "ck_tile/host/ops/flatmm/block/flatmm_uk_config.hpp"
 
 namespace ck_tile {
 

include/ck_tile/host/ops/flatmm/block/flatmm_sn_32x128x512_1x4x1_16x16x32_itl.hpp

@@ -5,8 +5,8 @@
 
 #include "ck_tile/core.hpp"
 #include "ck_tile/device/ops/gemm/warp/warp_gemm.hpp"
-#include "ck_tile/device/ops/flatmm/block/flatmm_uk_config.hpp"
-#include "ck_tile/device/ops/flatmm/block/flatmm_sn_32x128x512_1x4x1_16x16x32.hpp"
+#include "ck_tile/host/ops/flatmm/block/flatmm_uk_config.hpp"
+#include "ck_tile/host/ops/flatmm/block/flatmm_sn_32x128x512_1x4x1_16x16x32.hpp"
 
 namespace ck_tile {