Bug fix

library/include/ck/library/tensor_operation_instance/gpu/gemm_universal.hpp

@@ -347,7 +347,7 @@ void add_device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_comp_kpadding_instances(
         DeviceGemmV2<Row, Row, Row, F8, F8, BF16, PassThrough, PassThrough, PassThrough>>>&
         instances);
 
-void add_device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_comp_kpadding_instances(
+void add_device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_comp_nkpadding_instances(
     std::vector<std::unique_ptr<
         DeviceGemmV2<Row, Row, Row, F8, F8, BF16, PassThrough, PassThrough, PassThrough>>>&
         instances);
@@ -362,7 +362,7 @@ void add_device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_v1_kpadding_instances
         DeviceGemmV2<Row, Row, Row, F8, F8, BF16, PassThrough, PassThrough, PassThrough>>>&
         instances);
 
-void add_device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_v1_kpadding_instances(
+void add_device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_v1_nkpadding_instances(
     std::vector<std::unique_ptr<
         DeviceGemmV2<Row, Row, Row, F8, F8, BF16, PassThrough, PassThrough, PassThrough>>>&
         instances);
@@ -377,7 +377,7 @@ void add_device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_v2_kpadding_instances
         DeviceGemmV2<Row, Row, Row, F8, F8, BF16, PassThrough, PassThrough, PassThrough>>>&
         instances);
 
-void add_device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_v2_kpadding_instances(
+void add_device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_v2_nkpadding_instances(
     std::vector<std::unique_ptr<
         DeviceGemmV2<Row, Row, Row, F8, F8, BF16, PassThrough, PassThrough, PassThrough>>>&
         instances);
@@ -629,18 +629,18 @@ struct DeviceOperationInstanceFactory<
             {
                 add_device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_comp_default_instances(op_ptrs);
                 add_device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_comp_kpadding_instances(op_ptrs);
-                add_device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_comp_kpadding_instances(op_ptrs);
+                add_device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_comp_nkpadding_instances(op_ptrs);
 
                 add_device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_v1_default_instances(op_ptrs);
                 add_device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_v1_kpadding_instances(
                     op_ptrs);
-                add_device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_v1_kpadding_instances(
+                add_device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_v1_nkpadding_instances(
                     op_ptrs);
 
                 add_device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_v2_default_instances(op_ptrs);
                 add_device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_v2_kpadding_instances(
                     op_ptrs);
-                add_device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_v2_kpadding_instances(
+                add_device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_v2_nkpadding_instances(
                     op_ptrs);
             }
             else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&

library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_f8_f8_bf16/device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_comp_nkpadding_instance.cpp

@@ -8,7 +8,7 @@ namespace tensor_operation {
 namespace device {
 namespace instance {
 
-void add_device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_comp_kpadding_instances(
+void add_device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_comp_nkpadding_instances(
     std::vector<std::unique_ptr<
         DeviceGemmV2<Row, Row, Row, F8, F8, BF16, PassThrough, PassThrough, PassThrough>>>&
         instances)

library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_f8_f8_bf16/device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_v1_nkpadding_instance.cpp

@@ -8,7 +8,7 @@ namespace tensor_operation {
 namespace device {
 namespace instance {
 
-void add_device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_v1_kpadding_instances(
+void add_device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_v1_nkpadding_instances(
     std::vector<std::unique_ptr<
         DeviceGemmV2<Row, Row, Row, F8, F8, BF16, PassThrough, PassThrough, PassThrough>>>&
         instances)

library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_f8_f8_bf16/device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_v2_nkpadding_instance.cpp

@@ -8,7 +8,7 @@ namespace tensor_operation {
 namespace device {
 namespace instance {
 
-void add_device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_v2_kpadding_instances(
+void add_device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_v2_nkpadding_instances(
     std::vector<std::unique_ptr<
         DeviceGemmV2<Row, Row, Row, F8, F8, BF16, PassThrough, PassThrough, PassThrough>>>&
         instances)

profiler/src/CMakeLists.txt

 # ckProfiler
 set(PROFILER_SOURCES
     profiler.cpp
-    # profile_gemm.cpp
-    # profile_reduce.cpp
-    # profile_groupnorm_bwd_data.cpp
-    # profile_groupnorm_fwd.cpp
-    # profile_layernorm_bwd_data.cpp
-    # profile_layernorm_bwd_gamma_beta.cpp
-    # profile_groupnorm_bwd_gamma_beta.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_conv_tensor_rearrange.cpp
-    # profile_transpose.cpp
-    # profile_permute_scale.cpp
+    profile_gemm.cpp
+    profile_reduce.cpp
+    profile_groupnorm_bwd_data.cpp
+    profile_groupnorm_fwd.cpp
+    profile_layernorm_bwd_data.cpp
+    profile_layernorm_bwd_gamma_beta.cpp
+    profile_groupnorm_bwd_gamma_beta.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_conv_tensor_rearrange.cpp
+    profile_transpose.cpp
+    profile_permute_scale.cpp
 )
 
 if(GPU_TARGETS MATCHES "gfx9")
-#   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)
-#   endif()
-#   if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
-#     list(APPEND PROFILER_SOURCES profile_gemm_reduce.cpp)
-#     list(APPEND PROFILER_SOURCES profile_batched_gemm_gemm.cpp)
-#     list(APPEND PROFILER_SOURCES profile_batched_gemm_add_relu_gemm_add.cpp)
-#     list(APPEND PROFILER_SOURCES profile_gemm_add.cpp)
-#     list(APPEND PROFILER_SOURCES profile_gemm_add_add_fastgelu.cpp)
-#     list(APPEND PROFILER_SOURCES profile_gemm_add_fastgelu.cpp)
-#     list(APPEND PROFILER_SOURCES profile_grouped_gemm.cpp)
-#     list(APPEND PROFILER_SOURCES profile_gemm_streamk.cpp)
-#     list(APPEND PROFILER_SOURCES profile_gemm_fastgelu.cpp)
-#     list(APPEND PROFILER_SOURCES profile_gemm_add_relu.cpp)
-#     list(APPEND PROFILER_SOURCES profile_gemm_add_silu.cpp)
-#     list(APPEND PROFILER_SOURCES profile_gemm_add_relu_add_layernorm.cpp)
-#     list(APPEND PROFILER_SOURCES profile_grouped_gemm_fixed_nk.cpp)
-#     list(APPEND PROFILER_SOURCES profile_grouped_gemm_two_stage.cpp)
-#     list(APPEND PROFILER_SOURCES profile_grouped_gemm_fastgelu.cpp)
-#     list(APPEND PROFILER_SOURCES profile_grouped_gemm_tile_loop.cpp)
-#     list(APPEND PROFILER_SOURCES profile_grouped_gemm_multiply_tile_loop.cpp)
-#   endif()
-#   list(APPEND PROFILER_SOURCES profile_gemm_multiply_add.cpp)
-#   if(GPU_TARGETS MATCHES "gfx94")
-#     list(APPEND PROFILER_SOURCES profile_gemm_multiply_multiply.cpp)
-#     list(APPEND PROFILER_SOURCES profile_gemm_ab_scale.cpp)
-#   endif()
-#   list(APPEND PROFILER_SOURCES profile_batched_gemm.cpp)
-#   list(APPEND PROFILER_SOURCES profile_batched_gemm_reduce.cpp)
-#   list(APPEND PROFILER_SOURCES profile_gemm_add_multiply.cpp)
-#   list(APPEND PROFILER_SOURCES profile_gemm_bias_add_reduce.cpp)
-#   list(APPEND PROFILER_SOURCES profile_gemm_splitk.cpp)
+  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)
+  endif()
+  if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    list(APPEND PROFILER_SOURCES profile_gemm_reduce.cpp)
+    list(APPEND PROFILER_SOURCES profile_batched_gemm_gemm.cpp)
+    list(APPEND PROFILER_SOURCES profile_batched_gemm_add_relu_gemm_add.cpp)
+    list(APPEND PROFILER_SOURCES profile_gemm_add.cpp)
+    list(APPEND PROFILER_SOURCES profile_gemm_add_add_fastgelu.cpp)
+    list(APPEND PROFILER_SOURCES profile_gemm_add_fastgelu.cpp)
+    list(APPEND PROFILER_SOURCES profile_grouped_gemm.cpp)
+    list(APPEND PROFILER_SOURCES profile_gemm_streamk.cpp)
+    list(APPEND PROFILER_SOURCES profile_gemm_fastgelu.cpp)
+    list(APPEND PROFILER_SOURCES profile_gemm_add_relu.cpp)
+    list(APPEND PROFILER_SOURCES profile_gemm_add_silu.cpp)
+    list(APPEND PROFILER_SOURCES profile_gemm_add_relu_add_layernorm.cpp)
+    list(APPEND PROFILER_SOURCES profile_grouped_gemm_fixed_nk.cpp)
+    list(APPEND PROFILER_SOURCES profile_grouped_gemm_two_stage.cpp)
+    list(APPEND PROFILER_SOURCES profile_grouped_gemm_fastgelu.cpp)
+    list(APPEND PROFILER_SOURCES profile_grouped_gemm_tile_loop.cpp)
+    list(APPEND PROFILER_SOURCES profile_grouped_gemm_multiply_tile_loop.cpp)
+  endif()
+  list(APPEND PROFILER_SOURCES profile_gemm_multiply_add.cpp)
+  if(GPU_TARGETS MATCHES "gfx94")
+    list(APPEND PROFILER_SOURCES profile_gemm_multiply_multiply.cpp)
+    list(APPEND PROFILER_SOURCES profile_gemm_ab_scale.cpp)
+  endif()
+  list(APPEND PROFILER_SOURCES profile_batched_gemm.cpp)
+  list(APPEND PROFILER_SOURCES profile_batched_gemm_reduce.cpp)
+  list(APPEND PROFILER_SOURCES profile_gemm_add_multiply.cpp)
+  list(APPEND PROFILER_SOURCES profile_gemm_bias_add_reduce.cpp)
+  list(APPEND PROFILER_SOURCES profile_gemm_splitk.cpp)
   list(APPEND PROFILER_SOURCES profile_gemm_universal.cpp)
-#   list(APPEND PROFILER_SOURCES profile_gemm_universal_reduce.cpp)
-#   list(APPEND PROFILER_SOURCES profile_gemm_universal_streamk.cpp)
-#   list(APPEND PROFILER_SOURCES profile_conv_fwd_bias_relu.cpp)
-#   list(APPEND PROFILER_SOURCES profile_conv_fwd_bias_relu_add.cpp)
-#   list(APPEND PROFILER_SOURCES profile_conv_bwd_data.cpp)
-#   list(APPEND PROFILER_SOURCES profile_conv_fwd.cpp)
-#   list(APPEND PROFILER_SOURCES profile_grouped_conv_fwd_outelementop.cpp)
+  list(APPEND PROFILER_SOURCES profile_gemm_universal_reduce.cpp)
+  list(APPEND PROFILER_SOURCES profile_gemm_universal_streamk.cpp)
+  list(APPEND PROFILER_SOURCES profile_conv_fwd_bias_relu.cpp)
+  list(APPEND PROFILER_SOURCES profile_conv_fwd_bias_relu_add.cpp)
+  list(APPEND PROFILER_SOURCES profile_conv_bwd_data.cpp)
+  list(APPEND PROFILER_SOURCES profile_conv_fwd.cpp)
+  list(APPEND PROFILER_SOURCES profile_grouped_conv_fwd_outelementop.cpp)
 
 endif()
 
-# if(GPU_TARGETS MATCHES "gfx11" OR GPU_TARGETS MATCHES "gfx12" OR GPU_TARGETS MATCHES "gfx9")
-#   if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
-#     list(APPEND PROFILER_SOURCES profile_gemm_bilinear.cpp)
-#   endif()
-#   list(APPEND PROFILER_SOURCES profile_grouped_conv_fwd.cpp)
-#   list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_data.cpp)
-#   list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_weight.cpp)
-# endif()
+if(GPU_TARGETS MATCHES "gfx11" OR GPU_TARGETS MATCHES "gfx12" OR GPU_TARGETS MATCHES "gfx9")
+  if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    list(APPEND PROFILER_SOURCES profile_gemm_bilinear.cpp)
+  endif()
+  list(APPEND PROFILER_SOURCES profile_grouped_conv_fwd.cpp)
+  list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_data.cpp)
+  list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_weight.cpp)
+endif()
 
-# if(DL_KERNELS)
-#   list(APPEND PROFILER_SOURCES profile_batched_gemm_multi_d.cpp)
-#   list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_weight.cpp)
-# endif()
+if(DL_KERNELS)
+  list(APPEND PROFILER_SOURCES profile_batched_gemm_multi_d.cpp)
+  list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_weight.cpp)
+endif()
 
 set(PROFILER_EXECUTABLE ckProfiler)
 
@@ -91,85 +91,85 @@ if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 600241132)
 endif()
 
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE utility getopt::getopt)
-# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_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(${PROFILER_EXECUTABLE} PRIVATE device_normalization_bwd_gamma_beta_instance)
-# target_link_libraries(${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_image_to_column_instance)
-# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_column_to_image_instance)
-# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_transpose_instance)
-# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_permute_scale_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_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(${PROFILER_EXECUTABLE} PRIVATE device_normalization_bwd_gamma_beta_instance)
+target_link_libraries(${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_image_to_column_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_column_to_image_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_transpose_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_permute_scale_instance)
 
 if(GPU_TARGETS MATCHES "gfx9")
-  # 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)
-  # endif()
-  # if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
-  #   target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_instance)
-  #   target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_add_fastgelu_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(${PROFILER_EXECUTABLE} PRIVATE device_gemm_streamk_instance)
-  #   target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_fastgelu_instance)
-  #   target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_instance)
-  #   target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_silu_instance)
-  #   target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_add_layernorm_instance)
-  #   target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fixed_nk_instance)
-  #   target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fastgelu_instance)
-  #   target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_tile_loop_instance)
-  # endif()
-  # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_instance)
-  # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_reduce_instance)
-  # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_add_instance)
-  # if(GPU_TARGETS MATCHES "gfx94")
-  #   target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_multiply_instance)
-  #   target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_ab_scale_instance)
-  # endif()
-  # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_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)
+  endif()
+  if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_instance)
+    target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_add_fastgelu_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(${PROFILER_EXECUTABLE} PRIVATE device_gemm_streamk_instance)
+    target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_fastgelu_instance)
+    target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_instance)
+    target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_silu_instance)
+    target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_add_layernorm_instance)
+    target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fixed_nk_instance)
+    target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fastgelu_instance)
+    target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_tile_loop_instance)
+  endif()
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_instance)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_reduce_instance)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_add_instance)
+  if(GPU_TARGETS MATCHES "gfx94")
+    target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_multiply_instance)
+    target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_ab_scale_instance)
+  endif()
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance)
   target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_instance)
-  # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_reduce_instance)
-  # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_streamk_instance)
-  # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_multiply_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(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_instance)
-  # target_link_libraries(${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_grouped_conv1d_fwd_instance)
-  # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv1d_bwd_data_instance)
-  # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv3d_bwd_data_instance)
-  # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_bwd_data_instance)
-  # target_link_libraries(${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_fwd_convscale_instance)
-  # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_convinvscale_instance)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_reduce_instance)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_streamk_instance)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_multiply_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(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_instance)
+  target_link_libraries(${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_grouped_conv1d_fwd_instance)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv1d_bwd_data_instance)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv3d_bwd_data_instance)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_bwd_data_instance)
+  target_link_libraries(${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_fwd_convscale_instance)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_convinvscale_instance)
 endif()
 
-# if(GPU_TARGETS MATCHES "gfx9" OR GPU_TARGETS MATCHES "gfx11" OR GPU_TARGETS MATCHES "gfx12")
-#   if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
-#     target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bilinear_instance)
-#   endif()
-#   target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_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(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_fwd_instance)
-#   target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_weight_instance)
-# endif()
+if(GPU_TARGETS MATCHES "gfx9" OR GPU_TARGETS MATCHES "gfx11" OR GPU_TARGETS MATCHES "gfx12")
+  if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bilinear_instance)
+  endif()
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_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(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_fwd_instance)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_weight_instance)
+endif()
 
-# if(DL_KERNELS)
-#   target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_multi_d_instance)
-#   target_link_libraries(${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)
-# endif()
+if(DL_KERNELS)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_multi_d_instance)
+  target_link_libraries(${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)
+endif()
 
 rocm_install(TARGETS ${PROFILER_EXECUTABLE} COMPONENT profiler)