Allow building CK for specific data types and split off last remaining DL instances. (#830)

* properly split conv_nd_bwd_data instances

* split conv2d_fwd instance data types

* split the gemm, conv2d_fwd and batched_gemm_softamx_gemm

* split the tests by data types where possible

* filter examples by DTYPES

* split few remaining examples by DTYPES

* filter most instances by DTYPES

* add new lines at end of headers, fix grouped_gemm profiler

* fix syntax

* split the ckprofiler instances by DTYPES

* split the conv2d and quantization DL and XDL instances

* fix the splitting of conv2d DL instances

* split softmax and pool_fwd tests for fp16 and fp32 types

* fix syntax

* fix the dl_int8 quantization instances isolation

example/01_gemm/CMakeLists.txt

-add_custom_target(example_gemm_dl)
+if(DL_KERNELS)
+  add_custom_target(example_gemm_dl)
 
-add_example_executable(example_gemm_dl_fp32 gemm_dl_fp32.cpp)
-add_example_executable(example_gemm_dl_fp16 gemm_dl_fp16.cpp)
+  add_example_executable(example_gemm_dl_fp32 gemm_dl_fp32.cpp)
+  add_dependencies(example_gemm_dl example_gemm_dl_fp32)
+  if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_gemm_dl_fp16 gemm_dl_fp16.cpp)
+    add_dependencies(example_gemm_dl example_gemm_dl_fp16)
+  endif()
+  if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+    add_example_executable(example_gemm_dl_int8 gemm_dl_int8.cpp)
+    add_dependencies(example_gemm_dl example_gemm_dl_int8)
+  endif()
 
-add_dependencies(example_gemm_dl example_gemm_dl_fp32)
-add_dependencies(example_gemm_dl example_gemm_dl_fp16)
-
-if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
-   add_example_executable(example_gemm_dl_int8 gemm_dl_int8.cpp)
-   add_dependencies(example_gemm_dl example_gemm_dl_int8)
+  if(USE_BITINT_EXTENSION_INT4)
+    add_example_executable(example_gemm_dl_int4 gemm_dl_int4.cpp)
+    add_dependencies(example_gemm_dl example_gemm_dl_int4)
+  endif(USE_BITINT_EXTENSION_INT4)
 endif()
 
-if(USE_BITINT_EXTENSION_INT4)
-  add_example_executable(example_gemm_dl_int4 gemm_dl_int4.cpp)
-  add_dependencies(example_gemm_dl example_gemm_dl_int4)
-endif(USE_BITINT_EXTENSION_INT4)
-
-
 add_custom_target(example_gemm_xdl)
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+  add_example_executable(example_gemm_xdl_fp16 gemm_xdl_fp16.cpp)
+  add_example_executable(example_gemm_xdl_wavelet_fp16 gemm_xdl_wavelet_fp16.cpp)
+  add_dependencies(example_gemm_xdl example_gemm_xdl_fp16)
+  add_dependencies(example_gemm_xdl example_gemm_xdl_wavelet_fp16)
+  add_example_executable(example_gemm_xdl_skip_b_lds_fp16 gemm_xdl_skip_b_lds_fp16.cpp)
+  add_dependencies(example_gemm_xdl example_gemm_xdl_skip_b_lds_fp16)
 
-add_example_executable(example_gemm_xdl_fp16 gemm_xdl_fp16.cpp)
-add_example_executable(example_gemm_xdl_wavelet_fp16 gemm_xdl_wavelet_fp16.cpp)
-add_example_executable(example_gemm_xdl_bf16 gemm_xdl_bf16.cpp)
+  if(GPU_TARGETS MATCHES "gfx1100" OR GPU_TARGETS MATCHES "gfx1101" OR GPU_TARGETS MATCHES "gfx1102")
+    add_custom_target(example_gemm_wmma)
+    add_example_executable(example_gemm_wmma_fp16 gemm_wmma_fp16.cpp)
+    add_dependencies(example_gemm_wmma example_gemm_wmma_fp16)
+  endif()
 
-add_dependencies(example_gemm_xdl example_gemm_xdl_fp16)
-add_dependencies(example_gemm_xdl example_gemm_xdl_bf16)
-add_dependencies(example_gemm_xdl example_gemm_xdl_wavelet_fp16)
+endif()
+
+if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+  add_example_executable(example_gemm_xdl_bf16 gemm_xdl_bf16.cpp)
+  add_dependencies(example_gemm_xdl example_gemm_xdl_bf16)
+endif()
 
 if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
   add_example_executable(example_gemm_xdl_int8 gemm_xdl_int8.cpp)
@@ -37,22 +50,17 @@ if(USE_BITINT_EXTENSION_INT4)
   add_dependencies(example_gemm_xdl example_gemm_xdl_int4)
 endif(USE_BITINT_EXTENSION_INT4)
 
-add_example_executable(example_gemm_xdl_skip_b_lds_fp16 gemm_xdl_skip_b_lds_fp16.cpp)
-# FIXME: re-enable this exampe as test when SWDEV-335738 is fixed
-add_example_executable_no_testing(example_gemm_xdl_fp64 gemm_xdl_fp64.cpp)
-
-add_dependencies(example_gemm_xdl example_gemm_xdl_skip_b_lds_fp16)
-add_dependencies(example_gemm_xdl example_gemm_xdl_fp64)
-
-if(GPU_TARGETS MATCHES "gfx1100" OR GPU_TARGETS MATCHES "gfx1101" OR GPU_TARGETS MATCHES "gfx1102")
-  add_custom_target(example_gemm_wmma)
-  add_example_executable(example_gemm_wmma_fp16 gemm_wmma_fp16.cpp)
-  add_dependencies(example_gemm_wmma example_gemm_wmma_fp16)
+if(DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES)
+  # FIXME: re-enable this exampe as test when SWDEV-335738 is fixed
+  add_example_executable_no_testing(example_gemm_xdl_fp64 gemm_xdl_fp64.cpp)
+  add_dependencies(example_gemm_xdl example_gemm_xdl_fp64)
 endif()
 
 add_example_executable(example_gemm_xdl_streamk gemm_xdl_streamk.cpp)
 
-if(GPU_TARGETS MATCHES "gfx940" OR GPU_TARGETS MATCHES "gfx941" OR GPU_TARGETS MATCHES "gfx942")
-  add_example_executable(example_gemm_xdl_f8 gemm_xdl_f8.cpp)
-  add_dependencies(example_gemm_xdl example_gemm_xdl_f8)
+if(DTYPES MATCHES "fp8" OR NOT DEFINED DTYPES)
+  if(GPU_TARGETS MATCHES "gfx940" OR GPU_TARGETS MATCHES "gfx941" OR GPU_TARGETS MATCHES "gfx942")
+    add_example_executable(example_gemm_xdl_f8 gemm_xdl_f8.cpp)
+    add_dependencies(example_gemm_xdl example_gemm_xdl_f8)
+  endif()
 endif()

example/02_gemm_bilinear/CMakeLists.txt

+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
 list(APPEND gpu_list1 gfx1100 gfx1101 gfx1102)
 list(APPEND gpu_list2 gfx908 gfx90a gfx940 gfx941 gfx942)
 set(target 0)
@@ -15,3 +16,4 @@ foreach(gpu IN LISTS GPU_TARGETS)
    set(target 1)
  endif()
 endforeach()
+endif()

example/03_gemm_bias_relu/CMakeLists.txt

+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
 list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
 set(target 0)
 foreach(gpu IN LISTS GPU_TARGETS)
@@ -6,3 +7,4 @@ foreach(gpu IN LISTS GPU_TARGETS)
    set(target 1)
  endif()
 endforeach()
+endif()

example/04_gemm_add_add_fastgelu/CMakeLists.txt

@@ -3,22 +3,26 @@ set(target 0)
 foreach(gpu IN LISTS GPU_TARGETS)
  if(gpu IN_LIST gpu_list AND target EQUAL 0)
     add_custom_target(example_gemm_add_add_fastgelu_xdl)
-
-    add_example_executable(example_gemm_add_add_fastgelu_xdl_bf16 gemm_add_add_fastgelu_xdl_bf16.cpp)
-    add_example_executable(example_gemm_add_add_fastgelu_xdl_fp16 gemm_add_add_fastgelu_xdl_fp16.cpp)
-    add_example_executable(example_gemm_add_add_fastgelu_xdl_fp32 gemm_add_add_fastgelu_xdl_fp32.cpp)
+    if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+      add_example_executable(example_gemm_add_add_fastgelu_xdl_bf16 gemm_add_add_fastgelu_xdl_bf16.cpp)
+      add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_bf16)
+    endif()
+    if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+      add_example_executable(example_gemm_add_add_fastgelu_xdl_fp16 gemm_add_add_fastgelu_xdl_fp16.cpp)
+      add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_fp16)
+    endif()
+    if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+      add_example_executable(example_gemm_add_add_fastgelu_xdl_fp32 gemm_add_add_fastgelu_xdl_fp32.cpp)
+      add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_fp32)
+    endif()
     if(USE_BITINT_EXTENSION_INT4)
        add_example_executable(example_gemm_add_add_fastgelu_xdl_int4 gemm_add_add_fastgelu_xdl_int4.cpp)
+       add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_int4)
     endif(USE_BITINT_EXTENSION_INT4)
-    add_example_executable(example_gemm_add_add_fastgelu_xdl_int8 gemm_add_add_fastgelu_xdl_int8.cpp)
-
-    add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_bf16)
-    add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_fp16)
-    add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_fp32)
-    if(USE_BITINT_EXTENSION_INT4)
-        add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_int4)
-    endif(USE_BITINT_EXTENSION_INT4)
-    add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_int8)
+    if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+      add_example_executable(example_gemm_add_add_fastgelu_xdl_int8 gemm_add_add_fastgelu_xdl_int8.cpp)
+      add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_int8)
+    endif()
    set(target 1)
  endif()
 endforeach()
\ No newline at end of file

example/09_convnd_fwd/CMakeLists.txt

@@ -2,16 +2,34 @@ list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
 set(target 0)
 foreach(gpu IN LISTS GPU_TARGETS)
  if(gpu IN_LIST gpu_list AND target EQUAL 0)
-    add_example_executable(example_convnd_fwd_xdl_fp32 convnd_fwd_xdl_fp32.cpp)
-    add_example_executable(example_convnd_fwd_xdl_fp16 convnd_fwd_xdl_fp16.cpp)
-    add_example_executable(example_convnd_fwd_xdl_bf16 convnd_fwd_xdl_bf16.cpp)
-    add_example_executable(example_convnd_fwd_xdl_int8 convnd_fwd_xdl_int8.cpp)
+    if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+      add_example_executable(example_convnd_fwd_xdl_fp32 convnd_fwd_xdl_fp32.cpp)
+    endif()
+    if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+      add_example_executable(example_convnd_fwd_xdl_fp16 convnd_fwd_xdl_fp16.cpp)
+    endif()
+    if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+      add_example_executable(example_convnd_fwd_xdl_bf16 convnd_fwd_xdl_bf16.cpp)
+    endif()
+    if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+      add_example_executable(example_convnd_fwd_xdl_int8 convnd_fwd_xdl_int8.cpp)
+    endif()
     # FIXME: re-enable this exampe as test when SWDEV-335738 is fixed
-    add_example_executable_no_testing(example_convnd_fwd_xdl_fp64 convnd_fwd_xdl_fp64.cpp)
+    if(DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES)
+      add_example_executable_no_testing(example_convnd_fwd_xdl_fp64 convnd_fwd_xdl_fp64.cpp)
+    endif()
    set(target 1)
  endif()
 endforeach()
 
-add_example_executable(example_convnd_fwd_dl_fp16 convnd_fwd_dl_fp16.cpp)
-add_example_executable(example_convnd_fwd_dl_fp32 convnd_fwd_dl_fp32.cpp)
-add_example_executable(example_convnd_fwd_dl_int8 convnd_fwd_dl_int8.cpp)
+if(DL_KERNELS)
+  if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_convnd_fwd_dl_fp16 convnd_fwd_dl_fp16.cpp)
+  endif()
+  if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+    add_example_executable(example_convnd_fwd_dl_fp32 convnd_fwd_dl_fp32.cpp)
+  endif()
+  if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+    add_example_executable(example_convnd_fwd_dl_int8 convnd_fwd_dl_int8.cpp)
+  endif()
+endif()

example/10_convnd_fwd_multiple_d_multiple_reduce/CMakeLists.txt

@@ -3,14 +3,22 @@ set(target 0)
 foreach(gpu IN LISTS GPU_TARGETS)
  if(gpu IN_LIST gpu_list AND target EQUAL 0)
    add_custom_target(example_convnd_fwd_reduce_xdl)
-   add_example_executable(example_convnd_fwd_max_xdl_int8 convnd_fwd_max_xdl_int8.cpp)
-   add_example_executable_no_testing(example_convnd_fwd_max_xdl_bf16 convnd_fwd_max_xdl_bf16.cpp)
-   add_example_executable_no_testing(example_convnd_fwd_max_xdl_fp16 convnd_fwd_max_xdl_fp16.cpp)
-   add_example_executable(example_convnd_fwd_max_xdl_fp32 convnd_fwd_max_xdl_fp32.cpp)
-   add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_int8)
-   add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_bf16)
-   add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_fp16)
-   add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_fp32)
+   if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+    add_example_executable(example_convnd_fwd_max_xdl_int8 convnd_fwd_max_xdl_int8.cpp)
+    add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_int8)
+   endif()
+   if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+    add_example_executable_no_testing(example_convnd_fwd_max_xdl_bf16 convnd_fwd_max_xdl_bf16.cpp)
+    add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_bf16)
+   endif()
+   if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable_no_testing(example_convnd_fwd_max_xdl_fp16 convnd_fwd_max_xdl_fp16.cpp)
+    add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_fp16)
+   endif()
+   if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+    add_example_executable(example_convnd_fwd_max_xdl_fp32 convnd_fwd_max_xdl_fp32.cpp)
+    add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_fp32)
+   endif()
    if(USE_BITINT_EXTENSION_INT4)
       add_example_executable(example_convnd_fwd_max_xdl_int4 convnd_fwd_max_xdl_int4.cpp)
       add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_int4)

example/13_pool2d_fwd/CMakeLists.txt

-add_example_executable(example_pool2d_fwd_fp16 pool2d_fwd_fp16.cpp)
-add_example_executable(example_pool2d_fwd_fp32 pool2d_fwd_fp32.cpp)
-
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_pool2d_fwd_fp16 pool2d_fwd_fp16.cpp)
+endif()
+if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+    add_example_executable(example_pool2d_fwd_fp32 pool2d_fwd_fp32.cpp)
+endif()

example/14_gemm_quantization/CMakeLists.txt

 if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
 # dlops
-add_example_executable(example_gemm_dl_quantization_int8 gemm_dl_quantization_int8.cpp)
+if(DL_KERNELS)
+  add_example_executable(example_gemm_dl_quantization_int8 gemm_dl_quantization_int8.cpp)
+endif()
 
 # xdlops
 list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)

example/15_grouped_gemm/CMakeLists.txt

 add_custom_target(example_grouped_gemm_xdl)
-
-add_example_executable(example_grouped_gemm_xdl_fp32 grouped_gemm_xdl_fp32.cpp)
-add_example_executable(example_grouped_gemm_xdl_fp16 grouped_gemm_xdl_fp16.cpp)
-add_example_executable(example_grouped_gemm_xdl_bfp16 grouped_gemm_xdl_bfp16.cpp)
-add_example_executable(example_grouped_gemm_xdl_int8 grouped_gemm_xdl_int8.cpp)
-add_example_executable(example_grouped_gemm_multiple_d_dl_fp16 grouped_gemm_multiple_d_dl_fp16.cpp)
-add_example_executable(example_grouped_gemm_xdl_splitk_fp16 grouped_gemm_xdl_splitk_fp16.cpp)
-
-
-add_dependencies(example_grouped_gemm_xdl
-                 example_grouped_gemm_xdl_fp32
-                 example_grouped_gemm_xdl_fp16
-                 example_grouped_gemm_xdl_bfp16
-                 example_grouped_gemm_xdl_int8
-                 example_grouped_gemm_multiple_d_dl_fp16
-                 example_grouped_gemm_xdl_splitk_fp16)
-
+if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+  add_example_executable(example_grouped_gemm_xdl_fp32 grouped_gemm_xdl_fp32.cpp)
+  add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_fp32)
+endif()
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+  add_example_executable(example_grouped_gemm_xdl_fp16 grouped_gemm_xdl_fp16.cpp)
+  add_example_executable(example_grouped_gemm_multiple_d_dl_fp16 grouped_gemm_multiple_d_dl_fp16.cpp)
+  add_example_executable(example_grouped_gemm_xdl_splitk_fp16 grouped_gemm_xdl_splitk_fp16.cpp)
+  add_dependencies(example_grouped_gemm_xdl 
+                   example_grouped_gemm_xdl_fp16
+                   example_grouped_gemm_multiple_d_dl_fp16
+                   example_grouped_gemm_xdl_splitk_fp16)
+endif()
+if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+  add_example_executable(example_grouped_gemm_xdl_bfp16 grouped_gemm_xdl_bfp16.cpp)
+  add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_bfp16)
+endif()
+if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+  add_example_executable(example_grouped_gemm_xdl_int8 grouped_gemm_xdl_int8.cpp)
+  add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_int8)
+endif()
 if(USE_BITINT_EXTENSION_INT4)
   add_example_executable(example_grouped_gemm_xdl_int4 grouped_gemm_xdl_int4.cpp)
   add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_int4)

example/16_gemm_multi_d_multi_reduces/CMakeLists.txt

@@ -6,33 +6,33 @@ foreach(gpu IN LISTS GPU_TARGETS)
    add_custom_target(example_gemm_reduce_xdl_max)
    add_custom_target(example_gemm_reduce_xdl_mean_meansquare)
    add_custom_target(example_gemm_add_add_mean_meansquare_xdl)
-
-   add_example_executable(example_gemm_max_xdl_fp16 gemm_max_xdl_fp16.cpp)
-   add_example_executable(example_gemm_max_xdl_int8 gemm_max_xdl_int8.cpp)
-   add_example_executable(example_gemm_max_xdl_fp32 gemm_max_xdl_fp32.cpp)
-   add_example_executable(example_gemm_max_xdl_bf16 gemm_max_xdl_bf16.cpp)
-
-   add_example_executable(example_gemm_add_add_mean_meansquare_xdl_fp16 gemm_add_add_mean_meansquare_xdl_fp16.cpp)
-
-   add_example_executable(example_gemm_mean_meansquare_xdl_fp16 gemm_mean_meansquare_xdl_fp16.cpp)
-   add_example_executable(example_gemm_mean_meansquare_xdl_fp32 gemm_mean_meansquare_xdl_fp32.cpp)
-   add_example_executable(example_gemm_mean_meansquare_xdl_bf16 gemm_mean_meansquare_xdl_bf16.cpp)
-   add_example_executable(example_gemm_add_addsquare_xdl_int8 gemm_add_addsquare_xdl_int8.cpp)
-
-   add_dependencies(example_gemm_reduce_xdl_max
-                 example_gemm_max_xdl_bf16
-                 example_gemm_max_xdl_fp16
-                 example_gemm_max_xdl_fp32
-                 example_gemm_max_xdl_int8)
-
-   add_dependencies(example_gemm_reduce_xdl_mean_meansquare
-                 example_gemm_mean_meansquare_xdl_fp16
-                 example_gemm_mean_meansquare_xdl_fp32
-                 example_gemm_mean_meansquare_xdl_bf16
-                 example_gemm_add_addsquare_xdl_int8)
-
-   add_dependencies(example_gemm_add_add_mean_meansquare_xdl example_gemm_add_add_mean_meansquare_xdl_fp16)
-
+   if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_gemm_max_xdl_fp16 gemm_max_xdl_fp16.cpp)
+    add_example_executable(example_gemm_add_add_mean_meansquare_xdl_fp16 gemm_add_add_mean_meansquare_xdl_fp16.cpp)
+    add_example_executable(example_gemm_mean_meansquare_xdl_fp16 gemm_mean_meansquare_xdl_fp16.cpp)
+    add_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_fp16)
+    add_dependencies(example_gemm_add_add_mean_meansquare_xdl example_gemm_add_add_mean_meansquare_xdl_fp16)
+    add_dependencies(example_gemm_reduce_xdl_mean_meansquare example_gemm_mean_meansquare_xdl_fp16)
+   endif()
+   if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+    add_example_executable(example_gemm_max_xdl_int8 gemm_max_xdl_int8.cpp)
+    add_example_executable(example_gemm_add_addsquare_xdl_int8 gemm_add_addsquare_xdl_int8.cpp)
+    add_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_int8)
+    add_dependencies(example_gemm_reduce_xdl_mean_meansquare example_gemm_add_addsquare_xdl_int8)
+   endif()
+   if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+    add_example_executable(example_gemm_max_xdl_fp32 gemm_max_xdl_fp32.cpp)
+    add_example_executable(example_gemm_mean_meansquare_xdl_fp32 gemm_mean_meansquare_xdl_fp32.cpp)
+    add_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_fp32)
+    add_dependencies(example_gemm_reduce_xdl_mean_meansquare example_gemm_mean_meansquare_xdl_fp32)
+   endif()
+   if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_gemm_max_xdl_bf16 gemm_max_xdl_bf16.cpp)
+    add_example_executable(example_gemm_mean_meansquare_xdl_bf16 gemm_mean_meansquare_xdl_bf16.cpp)
+    add_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_bf16)
+    add_dependencies(example_gemm_reduce_xdl_mean_meansquare example_gemm_mean_meansquare_xdl_bf16)
+   endif()
+   
    add_dependencies(example_gemm_reduce_xdl
                  example_gemm_reduce_xdl_mean_meansquare
                  example_gemm_reduce_xdl_max

example/17_convnd_bwd_data/CMakeLists.txt

+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
 list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
 set(target 0)
 foreach(gpu IN LISTS GPU_TARGETS)
@@ -7,5 +8,8 @@ foreach(gpu IN LISTS GPU_TARGETS)
    set(target 1)
  endif()
 endforeach()
-add_example_executable(example_convnd_bwd_data_dl_fp16 convnd_bwd_data_dl_fp16.cpp)
-target_link_libraries(example_convnd_bwd_data_dl_fp16 PRIVATE utility)
+  if(DL_KERNELS)
+    add_example_executable(example_convnd_bwd_data_dl_fp16 convnd_bwd_data_dl_fp16.cpp)
+    target_link_libraries(example_convnd_bwd_data_dl_fp16 PRIVATE utility)
+  endif()
+endif()

example/18_batched_gemm_reduce/CMakeLists.txt

+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
 list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
 set(target 0)
 foreach(gpu IN LISTS GPU_TARGETS)
@@ -6,3 +7,4 @@ foreach(gpu IN LISTS GPU_TARGETS)
    set(target 1)
  endif()
 endforeach()
+endif()

example/20_grouped_conv_bwd_weight/CMakeLists.txt

@@ -3,18 +3,22 @@ set(target 0)
 foreach(gpu IN LISTS GPU_TARGETS)
  if(gpu IN_LIST gpu_list AND target EQUAL 0)
    add_custom_target(example_grouped_conv_bwd_weight)
-
-   add_example_executable(example_grouped_conv_bwd_weight_xdl_fp16 grouped_conv_bwd_weight_xdl_fp16.cpp)
-   add_example_executable(example_grouped_conv_bwd_weight_xdl_bf16 grouped_conv_bwd_weight_xdl_bf16.cpp)
-
-   add_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_xdl_fp16
-                                                 example_grouped_conv_bwd_weight_xdl_bf16)
+   if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_grouped_conv_bwd_weight_xdl_fp16 grouped_conv_bwd_weight_xdl_fp16.cpp)
+    add_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_xdl_fp16)
+   endif()
+   if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_grouped_conv_bwd_weight_xdl_bf16 grouped_conv_bwd_weight_xdl_bf16.cpp)
+    add_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_xdl_bf16)
+   endif()
    set(target 1)
  endif()
 endforeach()
 
-add_custom_target(example_grouped_conv_bwd_weight_dl)
-
-add_example_executable(example_grouped_conv_bwd_weight_dl_fp16 grouped_conv_bwd_weight_dl_fp16.cpp)
-
-add_dependencies(example_grouped_conv_bwd_weight_dl example_grouped_conv_bwd_weight_dl_fp16)
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+  if(DL_KERNELS)
+    add_custom_target(example_grouped_conv_bwd_weight_dl)
+    add_example_executable(example_grouped_conv_bwd_weight_dl_fp16 grouped_conv_bwd_weight_dl_fp16.cpp)
+    add_dependencies(example_grouped_conv_bwd_weight_dl example_grouped_conv_bwd_weight_dl_fp16)
+  endif()
+endif()
\ No newline at end of file

example/21_gemm_layernorm/CMakeLists.txt

+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
 list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
 set(target 0)
 foreach(gpu IN LISTS GPU_TARGETS)
@@ -9,3 +10,4 @@ foreach(gpu IN LISTS GPU_TARGETS)
    set(target 1)
  endif()
 endforeach()
+endif()

example/22_cgemm/CMakeLists.txt

 add_custom_target(example_cgemm_xdl)
 
-add_example_executable(example_cgemm_xdl_bf16 cgemm_xdl_bf16.cpp)
-add_example_executable(example_cgemm_xdl_fp16 cgemm_xdl_fp16.cpp)
+if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+  add_example_executable(example_cgemm_xdl_bf16 cgemm_xdl_bf16.cpp)
+  add_dependencies(example_cgemm_xdl example_cgemm_xdl_bf16)
+endif()
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+  add_example_executable(example_cgemm_xdl_fp16 cgemm_xdl_fp16.cpp)
+  add_dependencies(example_cgemm_xdl example_cgemm_xdl_fp16)
+endif()
+if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
 add_example_executable(example_cgemm_xdl_fp32 cgemm_xdl_fp32.cpp)
-add_example_executable(example_cgemm_xdl_int8 cgemm_xdl_int8.cpp)
-
-add_dependencies(example_cgemm_xdl 
-                 example_cgemm_xdl_bf16
-                 example_cgemm_xdl_fp16
-                 example_cgemm_xdl_fp32
-                 example_cgemm_xdl_int8)
-
+add_dependencies(example_cgemm_xdl example_cgemm_xdl_fp32)
+endif()
+if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+  add_example_executable(example_cgemm_xdl_int8 cgemm_xdl_int8.cpp)
+  add_dependencies(example_cgemm_xdl example_cgemm_xdl_int8)
+endif()
 if(USE_BITINT_EXTENSION_INT4)
   add_example_executable(example_cgemm_xdl_int4 cgemm_xdl_int4.cpp)
   add_dependencies(example_cgemm_xdl example_cgemm_xdl_int4)

example/24_batched_gemm/CMakeLists.txt

 add_custom_target(example_batched_gemm_xdl)
-
-add_example_executable(example_batched_gemm_xdl_fp32 batched_gemm_xdl_fp32.cpp)
-add_example_executable(example_batched_gemm_xdl_fp16 batched_gemm_xdl_fp16.cpp)
-add_example_executable(example_batched_gemm_xdl_bfp16 batched_gemm_xdl_bfp16.cpp)
-add_example_executable(example_batched_gemm_xdl_int8 batched_gemm_xdl_int8.cpp)
-
-add_dependencies(example_batched_gemm_xdl
-                 example_batched_gemm_xdl_fp32
-                 example_batched_gemm_xdl_fp16
-                 example_batched_gemm_xdl_bfp16
-                 example_batched_gemm_xdl_int8)
-
+if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+  add_example_executable(example_batched_gemm_xdl_fp32 batched_gemm_xdl_fp32.cpp)
+  add_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_fp32)
+endif()
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+  add_example_executable(example_batched_gemm_xdl_fp16 batched_gemm_xdl_fp16.cpp)
+  add_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_fp16)
+endif()
+if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+  add_example_executable(example_batched_gemm_xdl_bfp16 batched_gemm_xdl_bfp16.cpp)
+  add_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_bfp16)
+endif()
+if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+  add_example_executable(example_batched_gemm_xdl_int8 batched_gemm_xdl_int8.cpp)
+  add_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_int8)
+endif()
 if(USE_BITINT_EXTENSION_INT4)
   add_example_executable(example_batched_gemm_xdl_int4 batched_gemm_xdl_int4.cpp)
   add_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_int4)

example/25_gemm_bias_e_permute/CMakeLists.txt

-add_example_executable(example_gemm_bias_e_permute_g1m3n2k1_xdl_fp16 gemm_bias_e_permute_g1m3n2k1_xdl_fp16.cpp)
-add_example_executable(example_gemm_bias_e_permute_g1m2n3k1_xdl_fp16 gemm_bias_e_permute_g1m2n3k1_xdl_fp16.cpp)
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_gemm_bias_e_permute_g1m3n2k1_xdl_fp16 gemm_bias_e_permute_g1m3n2k1_xdl_fp16.cpp)
+    add_example_executable(example_gemm_bias_e_permute_g1m2n3k1_xdl_fp16 gemm_bias_e_permute_g1m2n3k1_xdl_fp16.cpp)
+endif()

example/26_contraction/CMakeLists.txt

-add_example_executable(example_contraction_bilinear_xdl_fp32 contraction_bilinear_xdl_fp32.cpp)
-add_example_executable(example_contraction_scale_xdl_fp32 contraction_scale_xdl_fp32.cpp)
-
-add_example_executable(example_contraction_bilinear_xdl_fp64 contraction_bilinear_xdl_fp64.cpp)
-add_example_executable(example_contraction_scale_xdl_fp64 contraction_scale_xdl_fp64.cpp)
+if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+    add_example_executable(example_contraction_bilinear_xdl_fp32 contraction_bilinear_xdl_fp32.cpp)
+    add_example_executable(example_contraction_scale_xdl_fp32 contraction_scale_xdl_fp32.cpp)
+endif()
+if(DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES)
+    add_example_executable(example_contraction_bilinear_xdl_fp64 contraction_bilinear_xdl_fp64.cpp)
+    add_example_executable(example_contraction_scale_xdl_fp64 contraction_scale_xdl_fp64.cpp)
+endif()

example/27_layernorm/CMakeLists.txt

-add_example_executable(example_layernorm_fp16 layernorm_fp16.cpp)
-add_example_executable(example_layernorm_splitk_fp16 layernorm_splitk_fp16.cpp)
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_layernorm_fp16 layernorm_fp16.cpp)
+    add_example_executable(example_layernorm_splitk_fp16 layernorm_splitk_fp16.cpp)
+endif()

example/28_grouped_gemm_bias_e_permute/CMakeLists.txt

-add_example_executable(example_grouped_gemm_bias_e_permute_xdl_fp16 grouped_gemm_bias_e_permute_xdl_fp16.cpp)
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_grouped_gemm_bias_e_permute_xdl_fp16 grouped_gemm_bias_e_permute_xdl_fp16.cpp)
+endif()