Revert disabled ckProfiler operators.

profiler/src/CMakeLists.txt

 # ckProfiler
 set(PROFILER_SOURCES
     profiler.cpp
-    # profile_gemm.cpp
-    # profile_gemm_splitk.cpp
-    # profile_gemm_bias_add_reduce.cpp
-    # profile_gemm_add_multiply.cpp
-    # profile_gemm_multiply_add.cpp
-    # profile_gemm_reduce.cpp
-    # profile_batched_gemm.cpp
-    # profile_batched_gemm_reduce.cpp
-    # profile_conv_fwd.cpp
-    # profile_conv_fwd_bias_relu.cpp
-    # profile_conv_fwd_bias_relu_add.cpp
-    # profile_conv_bwd_data.cpp
-    # profile_grouped_conv_fwd.cpp
-    # profile_grouped_conv_bwd_weight.cpp
-    # profile_reduce.cpp
-    # profile_groupnorm_bwd_data.cpp
-    # profile_groupnorm_fwd.cpp
-    # profile_layernorm_bwd_data.cpp
-    # profile_layernorm_fwd.cpp
-    # profile_max_pool3d_fwd.cpp
-    # profile_avg_pool3d_bwd.cpp
-    # profile_max_pool3d_bwd.cpp
-    # profile_softmax.cpp
-    # profile_batchnorm_fwd.cpp
-    # profile_batchnorm_bwd.cpp
-    # profile_batchnorm_infer.cpp
-    # profile_grouped_conv_bwd_data.cpp
-    # profile_conv_tensor_rearrange.cpp
+    profile_gemm.cpp
+    profile_gemm_splitk.cpp
+    profile_gemm_bias_add_reduce.cpp
+    profile_gemm_add_multiply.cpp
+    profile_gemm_multiply_add.cpp
+    profile_gemm_reduce.cpp
+    profile_batched_gemm.cpp
+    profile_batched_gemm_reduce.cpp
+    profile_conv_fwd.cpp
+    profile_conv_fwd_bias_relu.cpp
+    profile_conv_fwd_bias_relu_add.cpp
+    profile_conv_bwd_data.cpp
+    profile_grouped_conv_fwd.cpp
+    profile_grouped_conv_bwd_weight.cpp
+    profile_reduce.cpp
+    profile_groupnorm_bwd_data.cpp
+    profile_groupnorm_fwd.cpp
+    profile_layernorm_bwd_data.cpp
+    profile_layernorm_fwd.cpp
+    profile_max_pool3d_fwd.cpp
+    profile_avg_pool3d_bwd.cpp
+    profile_max_pool3d_bwd.cpp
+    profile_softmax.cpp
+    profile_batchnorm_fwd.cpp
+    profile_batchnorm_bwd.cpp
+    profile_batchnorm_infer.cpp
+    profile_grouped_conv_bwd_data.cpp
+    profile_conv_tensor_rearrange.cpp
 )
 
 if(DL_KERNELS)
@@ -36,22 +36,22 @@ if(DL_KERNELS)
 endif()
 
 if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
-  # list(APPEND PROFILER_SOURCES profile_batched_gemm_gemm.cpp)
-  # list(APPEND PROFILER_SOURCES profile_gemm_fastgelu.cpp)
-  # list(APPEND PROFILER_SOURCES profile_gemm_streamk.cpp)
-  # list(APPEND PROFILER_SOURCES profile_gemm_bilinear.cpp)
-  # list(APPEND PROFILER_SOURCES profile_gemm_add_fastgelu.cpp)
-  # list(APPEND PROFILER_SOURCES profile_gemm_add_add_fastgelu.cpp)
-  # list(APPEND PROFILER_SOURCES profile_gemm_add_relu_add_layernorm.cpp)
-  # list(APPEND PROFILER_SOURCES profile_batched_gemm_add_relu_gemm_add.cpp)
+  list(APPEND PROFILER_SOURCES profile_batched_gemm_gemm.cpp)
+  list(APPEND PROFILER_SOURCES profile_gemm_fastgelu.cpp)
+  list(APPEND PROFILER_SOURCES profile_gemm_streamk.cpp)
+  list(APPEND PROFILER_SOURCES profile_gemm_bilinear.cpp)
+  list(APPEND PROFILER_SOURCES profile_gemm_add_fastgelu.cpp)
+  list(APPEND PROFILER_SOURCES profile_gemm_add_add_fastgelu.cpp)
+  list(APPEND PROFILER_SOURCES profile_gemm_add_relu_add_layernorm.cpp)
+  list(APPEND PROFILER_SOURCES profile_batched_gemm_add_relu_gemm_add.cpp)
   list(APPEND PROFILER_SOURCES profile_grouped_gemm.cpp)
-  # list(APPEND PROFILER_SOURCES profile_grouped_gemm_fastgelu.cpp)
+  list(APPEND PROFILER_SOURCES profile_grouped_gemm_fastgelu.cpp)
   list(APPEND PROFILER_SOURCES profile_grouped_gemm_multiple_d_splitk.cpp)
 endif()
 
 if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES)
-  # list(APPEND PROFILER_SOURCES profile_contraction_bilinear.cpp)
-  # list(APPEND PROFILER_SOURCES profile_contraction_scale.cpp)
+  list(APPEND PROFILER_SOURCES profile_contraction_bilinear.cpp)
+  list(APPEND PROFILER_SOURCES profile_contraction_scale.cpp)
 endif()
 
 set(PROFILER_EXECUTABLE ckProfiler)
@@ -60,42 +60,42 @@ add_executable(${PROFILER_EXECUTABLE} ${PROFILER_SOURCES})
 target_compile_options(${PROFILER_EXECUTABLE} PRIVATE -Wno-global-constructors)
 
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE utility)
-# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_instance)
-# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance)
-# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_multiply_instance)
-# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_add_instance)
-# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_reduce_instance)
-# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bias_add_reduce_instance)
-# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_instance)
-# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_reduce_instance)
-# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_instance)
-# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_fwd_instance)
-# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_fwd_instance)
-# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_instance)
-# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv1d_bwd_data_instance)
-# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_bwd_data_instance)
-# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv3d_bwd_data_instance)
-# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_bwd_weight_instance)
-# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_weight_instance)
-# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_weight_instance)
-# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_instance)
-# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_add_instance)
-# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_fwd_instance)
-# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_bwd_data_instance)
-# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_softmax_instance)
-# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_reduce_instance)
-# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batchnorm_instance)
-# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool3d_fwd_instance)
-# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_avg_pool3d_bwd_instance)
-# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_max_pool_bwd_instance)
-# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_data_instance)
-# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_data_instance)
-# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_image_to_column_instance)
-# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_column_to_image_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_multiply_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_add_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_reduce_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bias_add_reduce_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_reduce_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_fwd_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_fwd_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv1d_bwd_data_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_bwd_data_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv3d_bwd_data_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_bwd_weight_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_weight_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_weight_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_add_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_fwd_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_bwd_data_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_softmax_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_reduce_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batchnorm_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool3d_fwd_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_avg_pool3d_bwd_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_max_pool_bwd_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_data_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_data_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_image_to_column_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_column_to_image_instance)
 
 if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES)
-  # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_bilinear_instance)
-  # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_scale_instance)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_bilinear_instance)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_scale_instance)
 endif()
 
 
@@ -105,16 +105,16 @@ if(DL_KERNELS)
 endif()
 
 if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
-  # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_fastgelu_instance)
-  # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_add_layernorm_instance)
-  # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bilinear_instance)
-  # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_add_fastgelu_instance)
-  # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_streamk_instance)
-  # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_fastgelu_instance)
-  # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_gemm_instance)
-  # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_add_relu_gemm_add_instance)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_fastgelu_instance)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_add_layernorm_instance)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bilinear_instance)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_add_fastgelu_instance)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_streamk_instance)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_fastgelu_instance)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_gemm_instance)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_add_relu_gemm_add_instance)
   target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_instance)
-  # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fastgelu_instance)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fastgelu_instance)
   target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_multiple_d_instance)
 endif()