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/29_batched_gemm_bias_e_permute/CMakeLists.txt

-add_example_executable(example_batched_gemm_bias_e_permute_xdl_fp16 batched_gemm_bias_e_permute_xdl_fp16.cpp)
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_batched_gemm_bias_e_permute_xdl_fp16 batched_gemm_bias_e_permute_xdl_fp16.cpp)
 
-if(GPU_TARGETS MATCHES "gfx1100" OR GPU_TARGETS MATCHES "gfx1101" OR GPU_TARGETS MATCHES "gfx1102")
-    add_example_executable(example_batched_gemm_bias_e_permute_wmma_fp16 batched_gemm_bias_e_permute_wmma_fp16.cpp)
+    if(GPU_TARGETS MATCHES "gfx1100" OR GPU_TARGETS MATCHES "gfx1101" OR GPU_TARGETS MATCHES "gfx1102")
+        add_example_executable(example_batched_gemm_bias_e_permute_wmma_fp16 batched_gemm_bias_e_permute_wmma_fp16.cpp)
+    endif()
 endif()

example/30_grouped_conv_fwd_multiple_d/CMakeLists.txt

@@ -5,23 +5,29 @@ set(target 0)
 foreach(gpu IN LISTS GPU_TARGETS)
  if(gpu IN_LIST gpu_list1 AND target EQUAL 0)
    add_custom_target(example_grouped_conv_fwd_multiple_d)
-
-   add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_fp16 grouped_conv_fwd_bias_relu_add_xdl_fp16.cpp)
-   add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_fp32 grouped_conv_fwd_bias_relu_add_xdl_fp32.cpp)
-   add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_bf16 grouped_conv_fwd_bias_relu_add_xdl_bf16.cpp)
-   add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_int8 grouped_conv_fwd_bias_relu_add_xdl_int8.cpp)
-
-   add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_fp16)
-   add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_fp32)
-   add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_bf16)
-   add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_int8)
-
+   if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+      add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_fp16 grouped_conv_fwd_bias_relu_add_xdl_fp16.cpp)
+      add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_fp16)
+      add_example_executable(example_grouped_conv_fwd_xdl_fp16 grouped_conv_fwd_xdl_fp16.cpp)
+      add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_xdl_fp16)
+   endif()
+   if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+      add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_fp32 grouped_conv_fwd_bias_relu_add_xdl_fp32.cpp)
+      add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_fp32)
+   endif()
+   if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+      add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_bf16 grouped_conv_fwd_bias_relu_add_xdl_bf16.cpp)
+      add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_bf16)
+   endif()
+   if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+      add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_int8 grouped_conv_fwd_bias_relu_add_xdl_int8.cpp)
+      add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_int8)
+   endif()
    if(USE_BITINT_EXTENSION_INT4)
       add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_int4 grouped_conv_fwd_bias_relu_add_xdl_int4.cpp)
       add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_int4)
    endif() # USE_BITINT_EXTENSION_INT4
-   add_example_executable(example_grouped_conv_fwd_xdl_fp16 grouped_conv_fwd_xdl_fp16.cpp)
-   add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_xdl_fp16)
+
    set(target 1)
  endif()
 endforeach()
@@ -29,8 +35,12 @@ endforeach()
 set(target 0)
 foreach(gpu IN LISTS GPU_TARGETS)
  if(gpu IN_LIST gpu_list2 AND target EQUAL 0)
-   add_example_executable(example_grouped_conv_fwd_bias_relu_add_wmma_fp16 grouped_conv_fwd_bias_relu_add_wmma_fp16.cpp)
-   add_example_executable(example_grouped_conv_fwd_bias_relu_add_wmma_int8 grouped_conv_fwd_bias_relu_add_wmma_int8.cpp)
+   if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+      add_example_executable(example_grouped_conv_fwd_bias_relu_add_wmma_fp16 grouped_conv_fwd_bias_relu_add_wmma_fp16.cpp)
+   endif()
+   if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+      add_example_executable(example_grouped_conv_fwd_bias_relu_add_wmma_int8 grouped_conv_fwd_bias_relu_add_wmma_int8.cpp)
+   endif()
    set(target 1)
  endif()
 endforeach()

example/31_batched_gemm_gemm/CMakeLists.txt

@@ -3,10 +3,15 @@ list(APPEND gpu_list2 gfx908 gfx90a)
 set(target 0)
 foreach(gpu IN LISTS GPU_TARGETS)
  if(gpu IN_LIST gpu_list1 AND target EQUAL 0)
-   add_example_executable(example_batched_gemm_gemm_xdl_fp32 batched_gemm_gemm_xdl_fp32.cpp)
-   add_example_executable(example_batched_gemm_gemm_xdl_fp16 batched_gemm_gemm_xdl_fp16.cpp)
-   add_example_executable(example_batched_gemm_gemm_xdl_bf16 batched_gemm_gemm_xdl_bf16.cpp)
-
+   if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+      add_example_executable(example_batched_gemm_gemm_xdl_fp32 batched_gemm_gemm_xdl_fp32.cpp)
+   endif()
+   if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+      add_example_executable(example_batched_gemm_gemm_xdl_fp16 batched_gemm_gemm_xdl_fp16.cpp)
+   endif()
+   if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+      add_example_executable(example_batched_gemm_gemm_xdl_bf16 batched_gemm_gemm_xdl_bf16.cpp)
+   endif()
    if(USE_BITINT_EXTENSION_INT4)
       add_example_executable(example_batched_gemm_gemm_xdl_int4 batched_gemm_gemm_xdl_int4.cpp)
    endif(USE_BITINT_EXTENSION_INT4)
@@ -15,5 +20,7 @@ foreach(gpu IN LISTS GPU_TARGETS)
 endforeach()
 
 if(NOT GPU_TARGETS MATCHES "gfx94" AND NOT GPU_TARGETS MATCHES "gfx1")
-   add_example_executable(example_batched_gemm_gemm_xdl_int8 batched_gemm_gemm_xdl_int8.cpp)
+   if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+      add_example_executable(example_batched_gemm_gemm_xdl_int8 batched_gemm_gemm_xdl_int8.cpp)
+   endif()
 endif()

example/32_batched_gemm_scale_softmax_gemm/CMakeLists.txt

-add_example_executable(example_batched_gemm_scale_softmax_gemm_xdl_fp16 batched_gemm_scale_softmax_gemm_xdl_fp16.cpp)
-add_example_executable(example_batched_gemm_scale_softmax_gemm_xdl_bf16 batched_gemm_scale_softmax_gemm_xdl_bf16.cpp)
-add_example_executable(example_batched_gemm_scale_softmax_gemm_permute_xdl_fp16 batched_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp)
-add_example_executable(example_batched_gemm_scale_softmax_gemm_permute_xdl_bf16 batched_gemm_scale_softmax_gemm_permute_xdl_bf16.cpp)
-add_example_executable(example_grouped_gemm_scale_softmax_gemm_permute_xdl_fp16 grouped_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp)
-add_example_executable(example_batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16 batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp)
-add_example_executable(example_grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16 grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp)
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_batched_gemm_scale_softmax_gemm_xdl_fp16 batched_gemm_scale_softmax_gemm_xdl_fp16.cpp)
+    add_example_executable(example_batched_gemm_scale_softmax_gemm_permute_xdl_fp16 batched_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp)
+    add_example_executable(example_grouped_gemm_scale_softmax_gemm_permute_xdl_fp16 grouped_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp)
+    add_example_executable(example_batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16 batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp)
+    add_example_executable(example_grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16 grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp)
+endif()
+if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_batched_gemm_scale_softmax_gemm_xdl_bf16 batched_gemm_scale_softmax_gemm_xdl_bf16.cpp)
+    add_example_executable(example_batched_gemm_scale_softmax_gemm_permute_xdl_bf16 batched_gemm_scale_softmax_gemm_permute_xdl_bf16.cpp)
+endif()
 
 add_custom_target(example_gemm_scale_softmax_gemm)
-add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_xdl_fp16)
-add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_xdl_bf16)
-add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_permute_xdl_fp16)
-add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_permute_xdl_bf16)
-add_dependencies(example_gemm_scale_softmax_gemm example_grouped_gemm_scale_softmax_gemm_permute_xdl_fp16)
-add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16)
-add_dependencies(example_gemm_scale_softmax_gemm example_grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16)
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_xdl_fp16)
+    add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_permute_xdl_fp16)
+    add_dependencies(example_gemm_scale_softmax_gemm example_grouped_gemm_scale_softmax_gemm_permute_xdl_fp16)
+    add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16)
+    add_dependencies(example_gemm_scale_softmax_gemm example_grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16)
+endif()
+if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+    add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_xdl_bf16)
+    add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_permute_xdl_bf16)
+endif()

example/35_splitK_gemm/CMakeLists.txt

@@ -3,17 +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_splitK_gemm_xdl)
-   add_example_executable(example_splitK_gemm_xdl_fp32 splitK_gemm_xdl_fp32.cpp)
-   add_example_executable(example_splitK_gemm_xdl_fp16 splitK_gemm_xdl_fp16.cpp)
-   add_example_executable(example_splitK_gemm_xdl_bfp16 splitK_gemm_xdl_bfp16.cpp)
-   add_example_executable(example_splitK_gemm_xdl_int8 splitK_gemm_xdl_int8.cpp)
-
-   add_dependencies(example_splitK_gemm_xdl
-                 example_splitK_gemm_xdl_fp32
-                 example_splitK_gemm_xdl_fp16
-                 example_splitK_gemm_xdl_bfp16
-                 example_splitK_gemm_xdl_int8)
-
+   if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+    add_example_executable(example_splitK_gemm_xdl_fp32 splitK_gemm_xdl_fp32.cpp)
+    add_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_fp32)
+   endif()
+   if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+    add_example_executable(example_splitK_gemm_xdl_fp16 splitK_gemm_xdl_fp16.cpp)
+    add_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_fp16)
+   endif()
+   if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+    add_example_executable(example_splitK_gemm_xdl_bfp16 splitK_gemm_xdl_bfp16.cpp)
+    add_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_bfp16)
+   endif()
+   if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+    add_example_executable(example_splitK_gemm_xdl_int8 splitK_gemm_xdl_int8.cpp)
+    add_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_int8)
+   endif()
    if(USE_BITINT_EXTENSION_INT4)
       add_example_executable(example_splitK_gemm_xdl_int4 splitK_gemm_xdl_int4.cpp)
       add_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_int4)

example/37_batched_gemm_add_add_relu_gemm_add/CMakeLists.txt

-add_example_executable(example_batched_gemm_add_add_relu_gemm_add_xdl_fp16 batched_gemm_add_add_relu_gemm_add_xdl_fp16.cpp)
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_batched_gemm_add_add_relu_gemm_add_xdl_fp16 batched_gemm_add_add_relu_gemm_add_xdl_fp16.cpp)
+endif()

example/38_grouped_conv_bwd_data_multiple_d/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)
@@ -10,4 +11,5 @@ foreach(gpu IN LISTS GPU_TARGETS)
    add_dependencies(example_grouped_conv_bwd_data example_grouped_conv_bwd_data_bias_relu_fp16)
    set(target 1)
  endif()
-endforeach()
\ No newline at end of file
+endforeach()
+endif()

example/39_permute/CMakeLists.txt

-add_custom_target(example_permute)
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_custom_target(example_permute)
 
-add_example_executable(example_permute_1xHxW_fp16 permute_1xHxW_fp16.cpp)
-add_example_executable(example_permute_NxHxW_fp16 permute_NxHxW_fp16.cpp)
-add_example_executable(example_permute_HxWx4_fp16 permute_HxWx4_fp16.cpp)
+    add_example_executable(example_permute_1xHxW_fp16 permute_1xHxW_fp16.cpp)
+    add_example_executable(example_permute_NxHxW_fp16 permute_NxHxW_fp16.cpp)
+    add_example_executable(example_permute_HxWx4_fp16 permute_HxWx4_fp16.cpp)
 
-add_dependencies(example_permute example_permute_1xHxW_fp16)
-add_dependencies(example_permute example_permute_NxHxW_fp16)
-add_dependencies(example_permute example_permute_HxWx4_fp16)
+    add_dependencies(example_permute example_permute_1xHxW_fp16)
+    add_dependencies(example_permute example_permute_NxHxW_fp16)
+    add_dependencies(example_permute example_permute_HxWx4_fp16)
+endif()

example/40_conv2d_fwd_quantization/CMakeLists.txt

@@ -10,21 +10,19 @@ foreach(gpu IN LISTS GPU_TARGETS)
    set(target 1)
  endif()
 endforeach()
-# Conv perlayer quantization
-add_example_executable(example_conv2d_fwd_dl_perlayer_quantization_int8 conv2d_fwd_dl_perlayer_quantization_int8.cpp)
 
-# Conv perchannel quantization
-add_example_executable(example_conv2d_fwd_dl_perchannel_quantization_int8 conv2d_fwd_dl_perchannel_quantization_int8.cpp)
-
-# Conv + bias + relu perlayer quantization
-add_example_executable(example_conv2d_fwd_dl_bias_relu_perlayer_quantization_int8 conv2d_fwd_dl_bias_relu_perlayer_quantization_int8.cpp)
-
-# Conv + bias + relu perchannel quantization
-add_example_executable(example_conv2d_fwd_dl_bias_relu_perchannel_quantization_int8 conv2d_fwd_dl_bias_relu_perchannel_quantization_int8.cpp)
-
-# Conv + bias + tanh perlayer quantization
-add_example_executable(example_conv2d_fwd_dl_bias_tanh_perlayer_quantization_int8 conv2d_fwd_dl_bias_tanh_perlayer_quantization_int8.cpp)
-
-# Conv + bias + tanh perchannel quantization
-add_example_executable(example_conv2d_fwd_dl_bias_tanh_perchannel_quantization_int8 conv2d_fwd_dl_bias_tanh_perchannel_quantization_int8.cpp)
+  if(DL_KERNELS)
+  # Conv perlayer quantization
+  add_example_executable(example_conv2d_fwd_dl_perlayer_quantization_int8 conv2d_fwd_dl_perlayer_quantization_int8.cpp)
+  # Conv perchannel quantization
+  add_example_executable(example_conv2d_fwd_dl_perchannel_quantization_int8 conv2d_fwd_dl_perchannel_quantization_int8.cpp)
+  # Conv + bias + relu perlayer quantization
+  add_example_executable(example_conv2d_fwd_dl_bias_relu_perlayer_quantization_int8 conv2d_fwd_dl_bias_relu_perlayer_quantization_int8.cpp)
+  # Conv + bias + relu perchannel quantization
+  add_example_executable(example_conv2d_fwd_dl_bias_relu_perchannel_quantization_int8 conv2d_fwd_dl_bias_relu_perchannel_quantization_int8.cpp)
+  # Conv + bias + tanh perlayer quantization
+  add_example_executable(example_conv2d_fwd_dl_bias_tanh_perlayer_quantization_int8 conv2d_fwd_dl_bias_tanh_perlayer_quantization_int8.cpp)
+  # Conv + bias + tanh perchannel quantization
+  add_example_executable(example_conv2d_fwd_dl_bias_tanh_perchannel_quantization_int8 conv2d_fwd_dl_bias_tanh_perchannel_quantization_int8.cpp)
+  endif()
 endif()
\ No newline at end of file

example/41_grouped_conv_conv_fwd/CMakeLists.txt

@@ -3,9 +3,15 @@ list(APPEND gpu_list2 gfx908 gfx90a)
 set(target 0)
 foreach(gpu IN LISTS GPU_TARGETS)
  if(gpu IN_LIST gpu_list1 AND target EQUAL 0)
-   add_example_executable(example_grouped_conv_conv_fwd_xdl_fp32 grouped_conv_conv_fwd_xdl_fp32.cpp)
-   add_example_executable(example_grouped_conv_conv_fwd_xdl_fp16 grouped_conv_conv_fwd_xdl_fp16.cpp)
-   add_example_executable(example_grouped_conv_conv_fwd_xdl_bf16 grouped_conv_conv_fwd_xdl_bf16.cpp)
+   if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+      add_example_executable(example_grouped_conv_conv_fwd_xdl_fp32 grouped_conv_conv_fwd_xdl_fp32.cpp)
+   endif()
+   if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+      add_example_executable(example_grouped_conv_conv_fwd_xdl_fp16 grouped_conv_conv_fwd_xdl_fp16.cpp)
+   endif()
+   if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+      add_example_executable(example_grouped_conv_conv_fwd_xdl_bf16 grouped_conv_conv_fwd_xdl_bf16.cpp)
+   endif()
    if(USE_BITINT_EXTENSION_INT4)
       add_example_executable(example_grouped_conv_conv_fwd_xdl_int4 grouped_conv_conv_fwd_xdl_int4.cpp)
    endif(USE_BITINT_EXTENSION_INT4)
@@ -14,5 +20,7 @@ foreach(gpu IN LISTS GPU_TARGETS)
 endforeach()
 
 if(NOT GPU_TARGETS MATCHES "gfx94" AND NOT GPU_TARGETS MATCHES "gfx1")
-   add_example_executable(example_grouped_conv_conv_fwd_xdl_int8 grouped_conv_conv_fwd_xdl_int8.cpp)
+   if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+      add_example_executable(example_grouped_conv_conv_fwd_xdl_int8 grouped_conv_conv_fwd_xdl_int8.cpp)
+   endif()
 endif()

example/42_groupnorm/CMakeLists.txt

-add_example_executable(example_groupnorm_sigmoid_mul_fp16 groupnorm_sigmoid_mul_fp16.cpp)
-add_example_executable(example_groupnorm_splitk_fp16 groupnorm_splitk_fp16.cpp)
-add_example_executable(example_groupnorm_swish_fp16 groupnorm_swish_fp16.cpp)
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_groupnorm_sigmoid_mul_fp16 groupnorm_sigmoid_mul_fp16.cpp)
+    add_example_executable(example_groupnorm_splitk_fp16 groupnorm_splitk_fp16.cpp)
+    add_example_executable(example_groupnorm_swish_fp16 groupnorm_swish_fp16.cpp)
+endif()

example/43_splitk_gemm_bias_e_permute/CMakeLists.txt

-add_example_executable(example_splitk_gemm_bias_e_permute_xdl_fp16 splitk_gemm_bias_e_permute_xdl_fp16.cpp)
-add_example_executable(example_splitk_gemm_bias_e_permute_xdl_fp32 splitk_gemm_bias_e_permute_xdl_fp32.cpp)
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_splitk_gemm_bias_e_permute_xdl_fp16 splitk_gemm_bias_e_permute_xdl_fp16.cpp)
+endif()
+if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+    add_example_executable(example_splitk_gemm_bias_e_permute_xdl_fp32 splitk_gemm_bias_e_permute_xdl_fp32.cpp)
+endif()

example/44_elementwise_permute/CMakeLists.txt

-add_example_executable(example_elementwise_permute_4D_fp16 elementwise_permute_4D_fp16.cpp)
-add_example_executable(example_elementwise_permute_4D_fp16_2d elementwise_permute_4D_fp16_2d.cpp)
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_elementwise_permute_4D_fp16 elementwise_permute_4D_fp16.cpp)
+    add_example_executable(example_elementwise_permute_4D_fp16_2d elementwise_permute_4D_fp16_2d.cpp)
+endif()

example/46_gemm_add_multiply/CMakeLists.txt

-add_example_executable(example_gemm_add_multiply_dl_fp16 gemm_add_multiply_dl_fp16.cpp)
-add_example_executable(example_gemm_add_multiply_xdl_fp16 gemm_add_multiply_xdl_fp16.cpp)
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    if(DL_KERNELS)
+        add_example_executable(example_gemm_add_multiply_dl_fp16 gemm_add_multiply_dl_fp16.cpp)
+    endif()
+    add_example_executable(example_gemm_add_multiply_xdl_fp16 gemm_add_multiply_xdl_fp16.cpp)
+endif()

example/48_pool3d_fwd/CMakeLists.txt

-add_example_executable(example_pool3d_fwd_fp16 pool3d_fwd_fp16.cpp)
-
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_pool3d_fwd_fp16 pool3d_fwd_fp16.cpp)
+endif()

example/49_maxpool2d_bwd/CMakeLists.txt

-add_example_executable(example_maxpool2d_bwd_bf16 maxpool2d_bwd_bf16.cpp)
-add_example_executable(example_maxpool2d_bwd_fp16 maxpool2d_bwd_fp16.cpp)
-add_example_executable(example_maxpool2d_bwd_fp32 maxpool2d_bwd_fp32.cpp)
+if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_maxpool2d_bwd_bf16 maxpool2d_bwd_bf16.cpp)
+endif()
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_maxpool2d_bwd_fp16 maxpool2d_bwd_fp16.cpp)
+endif()
+if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+    add_example_executable(example_maxpool2d_bwd_fp32 maxpool2d_bwd_fp32.cpp)
+endif()

example/50_put_element/CMakeLists.txt

-add_example_executable(example_put_element_fp16 put_element_fp16.cpp)
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_put_element_fp16 put_element_fp16.cpp)
+endif()

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

@@ -16,7 +16,7 @@ namespace ck {
 namespace tensor_operation {
 namespace device {
 namespace instance {
-
+#ifdef __bf16__
 void add_device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gkn_gmn_instances(
     std::vector<std::unique_ptr<
         DeviceBatchedGemm<Col, Row, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
@@ -36,7 +36,8 @@ void add_device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gnk_gmn_instances(
     std::vector<std::unique_ptr<
         DeviceBatchedGemm<Row, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
         instances);
-
+#endif
+#ifdef __fp16__
 void add_device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instances(
     std::vector<std::unique_ptr<
         DeviceBatchedGemm<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
@@ -56,7 +57,8 @@ void add_device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instances(
     std::vector<std::unique_ptr<
         DeviceBatchedGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
         instances);
-
+#endif
+#ifdef __fp32__
 void add_device_batched_gemm_xdl_f32_f32_f32_gkm_gkn_gmn_instances(
     std::vector<std::unique_ptr<
         DeviceBatchedGemm<Col, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
@@ -76,7 +78,8 @@ void add_device_batched_gemm_xdl_f32_f32_f32_gmk_gnk_gmn_instances(
     std::vector<std::unique_ptr<
         DeviceBatchedGemm<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
         instances);
-
+#endif
+#ifdef __int8__
 void add_device_batched_gemm_xdl_int8_int8_int8_gkm_gkn_gmn_instances(
     std::vector<std::unique_ptr<DeviceBatchedGemm<Col,
                                                   Row,
@@ -120,7 +123,7 @@ void add_device_batched_gemm_xdl_int8_int8_int8_gmk_gnk_gmn_instances(
                                                   PassThrough,
                                                   PassThrough,
                                                   PassThrough>>>& instances);
-
+#endif
 template <typename ALayout,
           typename BLayout,
           typename CLayout,
@@ -151,7 +154,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceBatche
     static auto GetInstances()
     {
         std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
-
+#ifdef __fp32__
         if constexpr(is_same_v<ADataType, float> && is_same_v<BDataType, float> &&
                      is_same_v<CDataType, float>)
         {
@@ -176,8 +179,10 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceBatche
                 add_device_batched_gemm_xdl_f32_f32_f32_gkm_gnk_gmn_instances(op_ptrs);
             }
         }
-        else if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
-                          is_same_v<CDataType, half_t>)
+#endif
+#ifdef __fp16__
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
+                     is_same_v<CDataType, half_t>)
         {
             if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
                          is_same_v<CLayout, Row>)
@@ -200,8 +205,10 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceBatche
                 add_device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instances(op_ptrs);
             }
         }
-        else if constexpr(is_same_v<ADataType, bhalf_t> && is_same_v<BDataType, bhalf_t> &&
-                          is_same_v<CDataType, bhalf_t>)
+#endif
+#ifdef __bf16__
+        if constexpr(is_same_v<ADataType, bhalf_t> && is_same_v<BDataType, bhalf_t> &&
+                     is_same_v<CDataType, bhalf_t>)
         {
             if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
                          is_same_v<CLayout, Row>)
@@ -224,8 +231,10 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceBatche
                 add_device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gnk_gmn_instances(op_ptrs);
             }
         }
-        else if constexpr(is_same_v<ADataType, int8_t> && is_same_v<BDataType, int8_t> &&
-                          is_same_v<CDataType, int8_t>)
+#endif
+#ifdef __int8__
+        if constexpr(is_same_v<ADataType, int8_t> && is_same_v<BDataType, int8_t> &&
+                     is_same_v<CDataType, int8_t>)
         {
             if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
                          is_same_v<CLayout, Row>)
@@ -248,7 +257,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceBatche
                 add_device_batched_gemm_xdl_int8_int8_int8_gkm_gnk_gmn_instances(op_ptrs);
             }
         }
-
+#endif
         return op_ptrs;
     }
 };

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

@@ -14,7 +14,7 @@
 
 using CDE0ElementOp = ck::tensor_operation::element_wise::AddRelu;
 using CDE1ElementOp = ck::tensor_operation::element_wise::Add;
-
+#ifdef __fp16__
 namespace ck {
 namespace tensor_operation {
 namespace device {
@@ -137,3 +137,4 @@ struct DeviceOperationInstanceFactory<
 } // namespace device
 } // namespace tensor_operation
 } // namespace ck
+#endif

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

@@ -13,7 +13,7 @@
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
 #include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
-
+#ifdef __fp16__
 namespace ck {
 namespace tensor_operation {
 namespace device {
@@ -91,3 +91,4 @@ struct DeviceOperationInstanceFactory<
 } // namespace device
 } // namespace tensor_operation
 } // namespace ck
+#endif