Merge branch 'develop' into aosewski/gemm_tile_loop

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/20_grouped_conv_bwd_weight/run_grouped_conv_bwd_weight_example.inc

@@ -72,10 +72,11 @@ bool run_grouped_conv_bwd_weight(const ExecutionConfig& config,
     // init to 0
     wei_device_buf.SetZero();
 
-    std::array<ck::index_t, NDimSpatial> input_spatial_lengths{};
-    std::array<ck::index_t, NDimSpatial> filter_spatial_lengths{};
-    std::array<ck::index_t, NDimSpatial> output_spatial_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> input_lengths{};
     std::array<ck::index_t, NDimSpatial + 3> input_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> filter_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> weights_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> output_lengths{};
     std::array<ck::index_t, NDimSpatial + 3> output_strides{};
     std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
     std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
@@ -84,10 +85,11 @@ bool run_grouped_conv_bwd_weight(const ExecutionConfig& config,
 
     auto range_copy = [](const auto& from, auto to) { std::copy(begin(from), end(from), to); };
 
-    range_copy(conv_param.input_spatial_lengths_, begin(input_spatial_lengths));
-    range_copy(conv_param.filter_spatial_lengths_, begin(filter_spatial_lengths));
-    range_copy(conv_param.output_spatial_lengths_, begin(output_spatial_lengths));
+    range_copy(in_g_n_c_wis_desc.GetLengths(), begin(input_lengths));
     range_copy(in_g_n_c_wis_desc.GetStrides(), begin(input_strides));
+    range_copy(wei_g_k_c_xs_desc.GetLengths(), begin(filter_lengths));
+    range_copy(wei_g_k_c_xs_desc.GetStrides(), begin(weights_strides));
+    range_copy(out_g_n_k_wos_desc.GetLengths(), begin(output_lengths));
     range_copy(out_g_n_k_wos_desc.GetStrides(), begin(output_strides));
     range_copy(conv_param.conv_filter_strides_, begin(conv_filter_strides));
     range_copy(conv_param.conv_filter_dilations_, begin(conv_filter_dilations));
@@ -100,14 +102,11 @@ bool run_grouped_conv_bwd_weight(const ExecutionConfig& config,
     auto argument = conv.MakeArgument(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
                                       static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
                                       static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
-                                      conv_param.G_,
-                                      conv_param.N_,
-                                      conv_param.K_,
-                                      conv_param.C_,
-                                      input_spatial_lengths,
-                                      filter_spatial_lengths,
-                                      output_spatial_lengths,
+                                      input_lengths,
                                       input_strides,
+                                      filter_lengths,
+                                      weights_strides,
+                                      output_lengths,
                                       output_strides,
                                       conv_filter_strides,
                                       conv_filter_dilations,

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()

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()
+endforeach()
\ No newline at end of file
+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()