Extend XDL kernel to Support RDNA3/4 - Part 5 (#2725)

* Enable xdl in gfx11 & gfx12 * update cmake file * fix all instance build (cmake) * fix batched_gemm_gemm(cmake) * rebase cmake files * fix cmake build error * remve CK_ENABLE_DYNAMIC_WARP_SIZE * update cmake build error2 * fix gfx11 build CK_USE_XDL is enabled on gfx11 and gfx12 * fix gfx10 build * fix gfx11 error --------- Co-authored-by: Lin, Qun <Quentin.Lin+amdeng@amd.com>
2026-04-20 06:49:15 +00:00 · 2025-09-16 01:59:25 +08:00
parent 03b59f8c76
commit f22740df82
33 changed files with 243 additions and 397 deletions
--- a/example/01_gemm/CMakeLists.txt
+++ b/example/01_gemm/CMakeLists.txt
@@ -44,8 +44,7 @@ list(APPEND GEMM_OPTIONS "SHELL: -mllvm -greedy-reverse-local-assignment=1 -mllv
 example_compile_options(example_gemm_xdl_fp8_v3 PRIVATE ${GEMM_OPTIONS})
 example_compile_options(example_gemm_xdl_bf16_v3 PRIVATE ${GEMM_OPTIONS})

-
-list(APPEND gpu_list gfx942 gfx950)
+list(APPEND gpu_list gfx942 gfx950 gfx1200 gfx1201 gfx12-generic)
 set(target 0)
 foreach(gpu IN LISTS GPU_TARGETS)
    if(gpu IN_LIST gpu_list AND target EQUAL 0)
@@ -89,7 +88,14 @@ foreach(gpu IN LISTS GPU_TARGETS)

        add_example_executable(example_gemm_xdl_lds_direct_load_fp16 gemm_xdl_lds_direct_load_fp16.cpp)
        add_example_dependencies(example_gemm_xdl example_gemm_xdl_lds_direct_load_fp16)
+        set(target 1)
+    endif()
+endforeach()

+list(APPEND gpu_list gfx90a gfx942 gfx950 gfx1200 gfx1201 gfx12-generic)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+    if(gpu IN_LIST gpu_list AND target EQUAL 0)
        add_example_executable(example_gemm_xdl_bf16_streamk_v3 gemm_xdl_bf16_streamk_v3.cpp)
        add_example_dependencies(example_gemm_xdl example_gemm_xdl_bf16_streamk_v3)

--- a/example/18_batched_gemm_reduce/CMakeLists.txt
+++ b/example/18_batched_gemm_reduce/CMakeLists.txt
@@ -1,8 +1 @@
-list(APPEND gpu_list gfx908 gfx90a gfx942 gfx950)
-set(target 0)
-foreach(gpu IN LISTS GPU_TARGETS)
- if(gpu IN_LIST gpu_list AND target EQUAL 0)
-   add_example_executable(example_batched_gemm_reduce_xdl_fp16 batched_gemm_reduce_xdl_fp16.cpp)
-   set(target 1)
- endif()
-endforeach()
+add_example_executable(example_batched_gemm_reduce_xdl_fp16 batched_gemm_reduce_xdl_fp16.cpp)
--- a/example/28_grouped_gemm_bias_e_permute/grouped_gemm_bias_e_permute_xdl_fp16.cpp
+++ b/example/28_grouped_gemm_bias_e_permute/grouped_gemm_bias_e_permute_xdl_fp16.cpp
@@ -53,7 +53,7 @@ using DeviceOpInstanceKKNN = ck::tensor_operation::device::
        //############################################|        |        |        |  Type|  Type|    Type| DataType|       Type|  Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Spacialization| Spacialization| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
        //############################################|        |        |        |      |      |        |         |           |      |    Operation|   Operation|    Operation|               |               |               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
        //############################################|        |        |        |      |      |        |         |           |      |             |            |             |               |               |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
-        DeviceGroupedContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK,   F16,   F16,     F32,      F16, DsDataType,   F16,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,         ABSpec,         ABSpec,         DESpec,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1, 32, 1, 4>,               8>;
+        DeviceGroupedContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK,   F16,   F16,     F32,      F16, DsDataType,   F16,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,         ABSpec,         ABSpec,         DESpec,        1,   256,   256,   128,    32,   8,   8,   16,   16,    8,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1, 32, 1, 4>,               4>;
 // clang-format on

 // hardcoded for NumDimM == NumDimN == NumDimK == 2
--- a/example/31_batched_gemm_gemm/CMakeLists.txt
+++ b/example/31_batched_gemm_gemm/CMakeLists.txt
@@ -11,6 +11,6 @@ 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)

-if(NOT GPU_TARGETS MATCHES "gfx94" AND NOT GPU_TARGETS MATCHES "gfx95" AND NOT GPU_TARGETS MATCHES "gfx1")
+if(NOT GPU_TARGETS MATCHES "gfx94" AND NOT GPU_TARGETS MATCHES "gfx95")
   add_example_executable(example_batched_gemm_gemm_xdl_int8 batched_gemm_gemm_xdl_int8.cpp)
 endif()
--- a/example/62_convnd_activ/binary/CMakeLists.txt
+++ b/example/62_convnd_activ/binary/CMakeLists.txt
@@ -1,15 +1,9 @@
-list(APPEND gpu_list gfx908 gfx90a gfx942 gfx950)
-set(target 0)
-foreach(gpu IN LISTS GPU_TARGETS)
- if(gpu IN_LIST gpu_list AND target EQUAL 0)
-      add_custom_target(example_convnd_activ_binary_xdl)
-      # Bilinear residual
-      add_example_executable(example_convnd_fwd_xdl_bilinear_residual_fp16 convnd_fwd_xdl_bilinear_residual_fp16.cpp)
-      add_example_dependencies(example_convnd_activ_binary_xdl example_convnd_fwd_xdl_bilinear_residual_fp16)
-      add_example_executable(example_convnd_bwd_data_xdl_bilinear_residual_fp16 convnd_bwd_data_xdl_bilinear_residual_fp16.cpp)
-      add_example_dependencies(example_convnd_activ_binary_xdl example_convnd_bwd_data_xdl_bilinear_residual_fp16)
-      add_example_executable(example_convnd_bwd_weight_xdl_bilinear_residual_fp16 convnd_bwd_weight_xdl_bilinear_residual_fp16.cpp)
-      add_example_dependencies(example_convnd_activ_binary_xdl example_convnd_bwd_weight_xdl_bilinear_residual_fp16)
-      set(target 1)
- endif()
-endforeach()
+add_custom_target(example_convnd_activ_binary_xdl)
+# Bilinear residual
+add_example_executable(example_convnd_fwd_xdl_bilinear_residual_fp16 convnd_fwd_xdl_bilinear_residual_fp16.cpp)
+add_example_dependencies(example_convnd_activ_binary_xdl example_convnd_fwd_xdl_bilinear_residual_fp16)
+add_example_executable(example_convnd_bwd_data_xdl_bilinear_residual_fp16 convnd_bwd_data_xdl_bilinear_residual_fp16.cpp)
+add_example_dependencies(example_convnd_activ_binary_xdl example_convnd_bwd_data_xdl_bilinear_residual_fp16)
+add_example_executable(example_convnd_bwd_weight_xdl_bilinear_residual_fp16 convnd_bwd_weight_xdl_bilinear_residual_fp16.cpp)
+add_example_dependencies(example_convnd_activ_binary_xdl example_convnd_bwd_weight_xdl_bilinear_residual_fp16)
+
--- a/example/62_convnd_activ/convinvscale/CMakeLists.txt
+++ b/example/62_convnd_activ/convinvscale/CMakeLists.txt
@@ -1,10 +1,5 @@
-list(APPEND gpu_list gfx908 gfx90a gfx942 gfx950)
-set(target 0)
-foreach(gpu IN LISTS GPU_TARGETS)
- if(gpu IN_LIST gpu_list AND target EQUAL 0)
-      add_custom_target(example_convnd_activ_xdl_convinvscale)
-      add_example_executable(example_convnd_fwd_xdl_convinvscale_fp8 convnd_fwd_xdl_convinvscale_fp8.cpp)
-      add_example_dependencies(example_convnd_activ_xdl_convinvscale example_convnd_fwd_xdl_convinvscale_fp8)
-      set(target 1)
- endif()
-endforeach()
+if (NOT GPU_TARGETS MATCHES "gfx11")
+    add_custom_target(example_convnd_activ_xdl_convinvscale)
+    add_example_executable(example_convnd_fwd_xdl_convinvscale_fp8 convnd_fwd_xdl_convinvscale_fp8.cpp)
+    add_example_dependencies(example_convnd_activ_xdl_convinvscale example_convnd_fwd_xdl_convinvscale_fp8)
+endif()
--- a/example/62_convnd_activ/convscale/CMakeLists.txt
+++ b/example/62_convnd_activ/convscale/CMakeLists.txt
@@ -1,20 +1,14 @@
-list(APPEND gpu_list gfx908 gfx90a gfx942 gfx950)
-set(target 0)
-foreach(gpu IN LISTS GPU_TARGETS)
- if(gpu IN_LIST gpu_list AND target EQUAL 0)
-     add_custom_target(example_convnd_activ_xdl_convscale)
-     add_example_executable(example_convnd_fwd_xdl_convscale_fp8 convnd_fwd_xdl_convscale_fp8.cpp)
-     add_example_dependencies(example_convnd_activ_xdl_convscale example_convnd_fwd_xdl_convscale_fp8 )
+if (NOT GPU_TARGETS MATCHES "gfx11")
+    add_custom_target(example_convnd_activ_xdl_convscale)
+    add_example_executable(example_convnd_fwd_xdl_convscale_fp8 convnd_fwd_xdl_convscale_fp8.cpp)
+    add_example_dependencies(example_convnd_activ_xdl_convscale example_convnd_fwd_xdl_convscale_fp8 )

-     add_example_executable(example_convnd_fwd_xdl_convscale_bf8 convnd_fwd_xdl_convscale_bf8.cpp)
-     add_example_dependencies(example_convnd_activ_xdl_convscale example_convnd_fwd_xdl_convscale_bf8)
+    add_example_executable(example_convnd_fwd_xdl_convscale_bf8 convnd_fwd_xdl_convscale_bf8.cpp)
+    add_example_dependencies(example_convnd_activ_xdl_convscale example_convnd_fwd_xdl_convscale_bf8)

-     add_example_executable(example_convnd_fwd_xdl_convscale_fp8_bf8 convnd_fwd_xdl_convscale_fp8_bf8.cpp)
-     add_example_dependencies(example_convnd_activ_xdl_convscale example_convnd_fwd_xdl_convscale_fp8_bf8)
+    add_example_executable(example_convnd_fwd_xdl_convscale_fp8_bf8 convnd_fwd_xdl_convscale_fp8_bf8.cpp)
+    add_example_dependencies(example_convnd_activ_xdl_convscale example_convnd_fwd_xdl_convscale_fp8_bf8)

-     add_example_executable(example_convnd_fwd_xdl_convscale_bf8_fp8 convnd_fwd_xdl_convscale_bf8_fp8.cpp)
-     add_example_dependencies(example_convnd_activ_xdl_convscale example_convnd_fwd_xdl_convscale_bf8_fp8)
-
-     set(target 1)
- endif()
-endforeach()
+    add_example_executable(example_convnd_fwd_xdl_convscale_bf8_fp8 convnd_fwd_xdl_convscale_bf8_fp8.cpp)
+    add_example_dependencies(example_convnd_activ_xdl_convscale example_convnd_fwd_xdl_convscale_bf8_fp8)
+endif()
--- a/example/62_convnd_activ/convscale_add/CMakeLists.txt
+++ b/example/62_convnd_activ/convscale_add/CMakeLists.txt
@@ -1,11 +1,5 @@
-list(APPEND gpu_list gfx908 gfx90a gfx942 gfx950)
-set(target 0)
-foreach(gpu IN LISTS GPU_TARGETS)
- if(gpu IN_LIST gpu_list AND target EQUAL 0)
-     add_custom_target(example_convnd_activ_xdl_convscale_add)
-     add_example_executable(example_convnd_fwd_xdl_convscale_add_fp8 convnd_fwd_xdl_convscale_add_fp8.cpp)
-     add_example_dependencies(example_convnd_activ_xdl_convscale_add example_convnd_fwd_xdl_convscale_add_fp8 )
-
-     set(target 1)
- endif()
-endforeach()
+if (NOT GPU_TARGETS MATCHES "gfx11")
+    add_custom_target(example_convnd_activ_xdl_convscale_add)
+    add_example_executable(example_convnd_fwd_xdl_convscale_add_fp8 convnd_fwd_xdl_convscale_add_fp8.cpp)
+    add_example_dependencies(example_convnd_activ_xdl_convscale_add example_convnd_fwd_xdl_convscale_add_fp8)
+endif()
--- a/example/62_convnd_activ/convscale_reduce/CMakeLists.txt
+++ b/example/62_convnd_activ/convscale_reduce/CMakeLists.txt
@@ -1,14 +1,8 @@
-list(APPEND gpu_list gfx908 gfx90a gfx942 gfx950)
-set(target 0)
-foreach(gpu IN LISTS GPU_TARGETS)
- if(gpu IN_LIST gpu_list AND target EQUAL 0)
-     add_custom_target(example_convnd_activ_xdl_convscale_reduce)
-     add_example_executable(example_convnd_fwd_xdl_convscale_relu_amax_fp8 convnd_fwd_xdl_convscale_relu_amax_fp8.cpp)
-     add_example_dependencies(example_convnd_activ_xdl_convscale_reduce example_convnd_fwd_xdl_convscale_relu_amax_fp8)
+if (NOT GPU_TARGETS MATCHES "gfx11")
+    add_custom_target(example_convnd_activ_xdl_convscale_reduce)
+    add_example_executable(example_convnd_fwd_xdl_convscale_relu_amax_fp8 convnd_fwd_xdl_convscale_relu_amax_fp8.cpp)
+    add_example_dependencies(example_convnd_activ_xdl_convscale_reduce example_convnd_fwd_xdl_convscale_relu_amax_fp8)

-     add_example_executable(example_convnd_fwd_xdl_convscale_amax_fp8 convnd_fwd_xdl_convscale_amax_fp8.cpp)
-     add_example_dependencies(example_convnd_activ_xdl_convscale_reduce example_convnd_fwd_xdl_convscale_amax_fp8)
-
-     set(target 1)
- endif()
-endforeach()
+    add_example_executable(example_convnd_fwd_xdl_convscale_amax_fp8 convnd_fwd_xdl_convscale_amax_fp8.cpp)
+    add_example_dependencies(example_convnd_activ_xdl_convscale_reduce example_convnd_fwd_xdl_convscale_amax_fp8)
+endif()
--- a/example/62_convnd_activ/convscale_relu/CMakeLists.txt
+++ b/example/62_convnd_activ/convscale_relu/CMakeLists.txt
@@ -1,11 +1,5 @@
-list(APPEND gpu_list gfx908 gfx90a gfx942 gfx950)
-set(target 0)
-foreach(gpu IN LISTS GPU_TARGETS)
- if(gpu IN_LIST gpu_list AND target EQUAL 0)
-     add_custom_target(example_convnd_activ_xdl_convscale_relu)
-     add_example_executable(example_convnd_fwd_xdl_convscale_relu_fp8 convnd_fwd_xdl_convscale_relu_fp8.cpp)
-     add_example_dependencies(example_convnd_activ_xdl_convscale_relu example_convnd_fwd_xdl_convscale_relu_fp8 )
-
-     set(target 1)
- endif()
-endforeach()
+if (NOT GPU_TARGETS MATCHES "gfx11")
+    add_custom_target(example_convnd_activ_xdl_convscale_relu)
+    add_example_executable(example_convnd_fwd_xdl_convscale_relu_fp8 convnd_fwd_xdl_convscale_relu_fp8.cpp)
+    add_example_dependencies(example_convnd_activ_xdl_convscale_relu example_convnd_fwd_xdl_convscale_relu_fp8)
+endif()
--- a/example/62_convnd_activ/dynamic_unary/CMakeLists.txt
+++ b/example/62_convnd_activ/dynamic_unary/CMakeLists.txt
@@ -1,45 +1,37 @@
-list(APPEND gpu_list gfx908 gfx90a gfx942 gfx950)
-set(target 0)
-foreach(gpu IN LISTS GPU_TARGETS)
- if(gpu IN_LIST gpu_list AND target EQUAL 0)
-      add_custom_target(example_convnd_activ_dynamic_unary_xdl)
-      # Sigmoid
-      add_example_executable(example_convnd_fwd_xdl_dynamic_sigmoid_fp16 convnd_fwd_xdl_dynamic_sigmoid_fp16.cpp)
-      add_example_dependencies(example_convnd_activ_dynamic_unary_xdl example_convnd_fwd_xdl_dynamic_sigmoid_fp16)
-      # Tanh
-      add_example_executable(example_convnd_fwd_xdl_dynamic_tanh_fp16 convnd_fwd_xdl_dynamic_tanh_fp16.cpp)
-      add_example_dependencies(example_convnd_activ_dynamic_unary_xdl example_convnd_fwd_xdl_dynamic_tanh_fp16)
-      # Relu
-      add_example_executable(example_convnd_fwd_xdl_dynamic_relu_fp16 convnd_fwd_xdl_dynamic_relu_fp16.cpp)
-      add_example_dependencies(example_convnd_activ_dynamic_unary_xdl example_convnd_fwd_xdl_dynamic_relu_fp16)
-      # SoftRelu
-      add_example_executable(example_convnd_fwd_xdl_dynamic_softrelu_fp16 convnd_fwd_xdl_dynamic_softrelu_fp16.cpp)
-      add_example_dependencies(example_convnd_activ_dynamic_unary_xdl example_convnd_fwd_xdl_dynamic_softrelu_fp16)
-      # Abs
-      add_example_executable(example_convnd_fwd_xdl_dynamic_abs_fp16 convnd_fwd_xdl_dynamic_abs_fp16.cpp)
-      add_example_dependencies(example_convnd_activ_dynamic_unary_xdl example_convnd_fwd_xdl_dynamic_abs_fp16)
-      # Pow
-      add_example_executable(example_convnd_fwd_xdl_dynamic_pow_fp16 convnd_fwd_xdl_dynamic_pow_fp16.cpp)
-      add_example_dependencies(example_convnd_activ_dynamic_unary_xdl example_convnd_fwd_xdl_dynamic_pow_fp16)
-      # Clipped Relu
-      add_example_executable(example_convnd_fwd_xdl_dynamic_clippedrelu_fp16 convnd_fwd_xdl_dynamic_clippedrelu_fp16.cpp)
-      add_example_dependencies(example_convnd_activ_dynamic_unary_xdl example_convnd_fwd_xdl_dynamic_clippedrelu_fp16)
-      # Leaky Relu
-      add_example_executable(example_convnd_fwd_xdl_dynamic_leakyrelu_fp16 convnd_fwd_xdl_dynamic_leakyrelu_fp16.cpp)
-      add_example_dependencies(example_convnd_activ_dynamic_unary_xdl example_convnd_fwd_xdl_dynamic_leakyrelu_fp16)
-      # Elu
-      add_example_executable(example_convnd_fwd_xdl_dynamic_elu_fp16 convnd_fwd_xdl_dynamic_elu_fp16.cpp)
-      add_example_dependencies(example_convnd_activ_dynamic_unary_xdl example_convnd_fwd_xdl_dynamic_elu_fp16)
-      # Swish
-      add_example_executable(example_convnd_fwd_xdl_dynamic_swish_fp16 convnd_fwd_xdl_dynamic_swish_fp16.cpp)
-      add_example_dependencies(example_convnd_activ_dynamic_unary_xdl example_convnd_fwd_xdl_dynamic_swish_fp16)
-      # PassThrough
-      add_example_executable(example_convnd_fwd_xdl_dynamic_passthrough_fp16 convnd_fwd_xdl_dynamic_passthrough_fp16.cpp)
-      add_example_dependencies(example_convnd_activ_dynamic_unary_xdl example_convnd_fwd_xdl_dynamic_passthrough_fp16)
-      # Logistic
-      add_example_executable(example_convnd_fwd_xdl_dynamic_logistic_fp16 convnd_fwd_xdl_dynamic_logistic_fp16.cpp)
-      add_example_dependencies(example_convnd_activ_dynamic_unary_xdl example_convnd_fwd_xdl_dynamic_logistic_fp16)
-   
-   set(target 1)
- endif()
-endforeach()
+add_custom_target(example_convnd_activ_dynamic_unary_xdl)
+# Sigmoid
+add_example_executable(example_convnd_fwd_xdl_dynamic_sigmoid_fp16 convnd_fwd_xdl_dynamic_sigmoid_fp16.cpp)
+add_example_dependencies(example_convnd_activ_dynamic_unary_xdl example_convnd_fwd_xdl_dynamic_sigmoid_fp16)
+# Tanh
+add_example_executable(example_convnd_fwd_xdl_dynamic_tanh_fp16 convnd_fwd_xdl_dynamic_tanh_fp16.cpp)
+add_example_dependencies(example_convnd_activ_dynamic_unary_xdl example_convnd_fwd_xdl_dynamic_tanh_fp16)
+# Relu
+add_example_executable(example_convnd_fwd_xdl_dynamic_relu_fp16 convnd_fwd_xdl_dynamic_relu_fp16.cpp)
+add_example_dependencies(example_convnd_activ_dynamic_unary_xdl example_convnd_fwd_xdl_dynamic_relu_fp16)
+# SoftRelu
+add_example_executable(example_convnd_fwd_xdl_dynamic_softrelu_fp16 convnd_fwd_xdl_dynamic_softrelu_fp16.cpp)
+add_example_dependencies(example_convnd_activ_dynamic_unary_xdl example_convnd_fwd_xdl_dynamic_softrelu_fp16)
+# Abs
+add_example_executable(example_convnd_fwd_xdl_dynamic_abs_fp16 convnd_fwd_xdl_dynamic_abs_fp16.cpp)
+add_example_dependencies(example_convnd_activ_dynamic_unary_xdl example_convnd_fwd_xdl_dynamic_abs_fp16)
+# Pow
+add_example_executable(example_convnd_fwd_xdl_dynamic_pow_fp16 convnd_fwd_xdl_dynamic_pow_fp16.cpp)
+add_example_dependencies(example_convnd_activ_dynamic_unary_xdl example_convnd_fwd_xdl_dynamic_pow_fp16)
+# Clipped Relu
+add_example_executable(example_convnd_fwd_xdl_dynamic_clippedrelu_fp16 convnd_fwd_xdl_dynamic_clippedrelu_fp16.cpp)
+add_example_dependencies(example_convnd_activ_dynamic_unary_xdl example_convnd_fwd_xdl_dynamic_clippedrelu_fp16)
+# Leaky Relu
+add_example_executable(example_convnd_fwd_xdl_dynamic_leakyrelu_fp16 convnd_fwd_xdl_dynamic_leakyrelu_fp16.cpp)
+add_example_dependencies(example_convnd_activ_dynamic_unary_xdl example_convnd_fwd_xdl_dynamic_leakyrelu_fp16)
+# Elu
+add_example_executable(example_convnd_fwd_xdl_dynamic_elu_fp16 convnd_fwd_xdl_dynamic_elu_fp16.cpp)
+add_example_dependencies(example_convnd_activ_dynamic_unary_xdl example_convnd_fwd_xdl_dynamic_elu_fp16)
+# Swish
+add_example_executable(example_convnd_fwd_xdl_dynamic_swish_fp16 convnd_fwd_xdl_dynamic_swish_fp16.cpp)
+add_example_dependencies(example_convnd_activ_dynamic_unary_xdl example_convnd_fwd_xdl_dynamic_swish_fp16)
+# PassThrough
+add_example_executable(example_convnd_fwd_xdl_dynamic_passthrough_fp16 convnd_fwd_xdl_dynamic_passthrough_fp16.cpp)
+add_example_dependencies(example_convnd_activ_dynamic_unary_xdl example_convnd_fwd_xdl_dynamic_passthrough_fp16)
+# Logistic
+add_example_executable(example_convnd_fwd_xdl_dynamic_logistic_fp16 convnd_fwd_xdl_dynamic_logistic_fp16.cpp)
+add_example_dependencies(example_convnd_activ_dynamic_unary_xdl example_convnd_fwd_xdl_dynamic_logistic_fp16)
--- a/example/62_convnd_activ/multi_AB/CMakeLists.txt
+++ b/example/62_convnd_activ/multi_AB/CMakeLists.txt
@@ -1,17 +1,10 @@
-list(APPEND gpu_list gfx908 gfx90a gfx942 gfx950)
-set(target 0)
-foreach(gpu IN LISTS GPU_TARGETS)
- if(gpu IN_LIST gpu_list AND target EQUAL 0)
-      add_custom_target(example_convnd_activ_multi_ab_xdl)
-      # ScaleAdd on A and B
-      add_example_executable(example_conv_fwd_xdl_scaleadd_ab_fp16 conv_fwd_xdl_scaleadd_ab_fp16.cpp)
-      add_example_dependencies(example_convnd_activ_multi_ab_xdl example_conv_fwd_xdl_scaleadd_ab_fp16)
-      add_example_executable(example_conv_fwd_xdl_scaleadd_ab_fp32 conv_fwd_xdl_scaleadd_ab_fp32.cpp)
-      add_example_dependencies(example_convnd_activ_multi_ab_xdl example_conv_fwd_xdl_scaleadd_ab_fp32)
-      add_example_executable(example_conv_fwd_xdl_scaleadd_ab_bf16 conv_fwd_xdl_scaleadd_ab_bf16.cpp)
-      add_example_dependencies(example_convnd_activ_multi_ab_xdl example_conv_fwd_xdl_scaleadd_ab_bf16)
-      add_example_executable(example_conv_fwd_xdl_scaleadd_ab_int8 conv_fwd_xdl_scaleadd_ab_int8.cpp)
-      add_example_dependencies(example_convnd_activ_multi_ab_xdl example_conv_fwd_xdl_scaleadd_ab_int8)
-   set(target 1)
- endif()
-endforeach()
+add_custom_target(example_convnd_activ_multi_ab_xdl)
+# ScaleAdd on A and B
+add_example_executable(example_conv_fwd_xdl_scaleadd_ab_fp16 conv_fwd_xdl_scaleadd_ab_fp16.cpp)
+add_example_dependencies(example_convnd_activ_multi_ab_xdl example_conv_fwd_xdl_scaleadd_ab_fp16)
+add_example_executable(example_conv_fwd_xdl_scaleadd_ab_fp32 conv_fwd_xdl_scaleadd_ab_fp32.cpp)
+add_example_dependencies(example_convnd_activ_multi_ab_xdl example_conv_fwd_xdl_scaleadd_ab_fp32)
+add_example_executable(example_conv_fwd_xdl_scaleadd_ab_bf16 conv_fwd_xdl_scaleadd_ab_bf16.cpp)
+add_example_dependencies(example_convnd_activ_multi_ab_xdl example_conv_fwd_xdl_scaleadd_ab_bf16)
+add_example_executable(example_conv_fwd_xdl_scaleadd_ab_int8 conv_fwd_xdl_scaleadd_ab_int8.cpp)
+add_example_dependencies(example_convnd_activ_multi_ab_xdl example_conv_fwd_xdl_scaleadd_ab_int8)
--- a/example/62_convnd_activ/unary/CMakeLists.txt
+++ b/example/62_convnd_activ/unary/CMakeLists.txt
@@ -1,45 +1,37 @@
-list(APPEND gpu_list gfx908 gfx90a gfx942 gfx950)
-set(target 0)
-foreach(gpu IN LISTS GPU_TARGETS)
- if(gpu IN_LIST gpu_list AND target EQUAL 0)
-      add_custom_target(example_convnd_activ_unary_xdl)
-      # Sigmoid
-      add_example_executable(example_convnd_fwd_xdl_sigmoid_fp16 convnd_fwd_xdl_sigmoid_fp16.cpp)
-      add_example_dependencies(example_convnd_activ_unary_xdl example_convnd_fwd_xdl_sigmoid_fp16)
-      # Tanh
-      add_example_executable(example_convnd_fwd_xdl_tanh_fp16 convnd_fwd_xdl_tanh_fp16.cpp)
-      add_example_dependencies(example_convnd_activ_unary_xdl example_convnd_fwd_xdl_tanh_fp16)
-      # Relu
-      add_example_executable(example_convnd_fwd_xdl_relu_fp16 convnd_fwd_xdl_relu_fp16.cpp)
-      add_example_dependencies(example_convnd_activ_unary_xdl example_convnd_fwd_xdl_relu_fp16)
-      # SoftRelu
-      add_example_executable(example_convnd_fwd_xdl_softrelu_fp16 convnd_fwd_xdl_softrelu_fp16.cpp)
-      add_example_dependencies(example_convnd_activ_unary_xdl example_convnd_fwd_xdl_softrelu_fp16)
-      # Abs
-      add_example_executable(example_convnd_fwd_xdl_abs_fp16 convnd_fwd_xdl_abs_fp16.cpp)
-      add_example_dependencies(example_convnd_activ_unary_xdl example_convnd_fwd_xdl_abs_fp16)
-      # Pow
-      add_example_executable(example_convnd_fwd_xdl_pow_fp16 convnd_fwd_xdl_pow_fp16.cpp)
-      add_example_dependencies(example_convnd_activ_unary_xdl example_convnd_fwd_xdl_pow_fp16)
-      # Clipped Relu
-      add_example_executable(example_convnd_fwd_xdl_clippedrelu_fp16 convnd_fwd_xdl_clippedrelu_fp16.cpp)
-      add_example_dependencies(example_convnd_activ_unary_xdl example_convnd_fwd_xdl_clippedrelu_fp16)
-      # Leaky Relu
-      add_example_executable(example_convnd_fwd_xdl_leakyrelu_fp16 convnd_fwd_xdl_leakyrelu_fp16.cpp)
-      add_example_dependencies(example_convnd_activ_unary_xdl example_convnd_fwd_xdl_leakyrelu_fp16)
-      # Elu
-      add_example_executable(example_convnd_fwd_xdl_elu_fp16 convnd_fwd_xdl_elu_fp16.cpp)
-      add_example_dependencies(example_convnd_activ_unary_xdl example_convnd_fwd_xdl_elu_fp16)
-      # Swish
-      add_example_executable(example_convnd_fwd_xdl_swish_fp16 convnd_fwd_xdl_swish_fp16.cpp)
-      add_example_dependencies(example_convnd_activ_unary_xdl example_convnd_fwd_xdl_swish_fp16)
-      # PassThrough
-      add_example_executable(example_convnd_fwd_xdl_passthrough_fp16 convnd_fwd_xdl_passthrough_fp16.cpp)
-      add_example_dependencies(example_convnd_activ_unary_xdl example_convnd_fwd_xdl_passthrough_fp16)
-      # Logistic
-      add_example_executable(example_convnd_fwd_xdl_logistic_fp16 convnd_fwd_xdl_logistic_fp16.cpp)
-      add_example_dependencies(example_convnd_activ_unary_xdl example_convnd_fwd_xdl_logistic_fp16)
-   
-   set(target 1)
- endif()
-endforeach()
+add_custom_target(example_convnd_activ_unary_xdl)
+# Sigmoid
+add_example_executable(example_convnd_fwd_xdl_sigmoid_fp16 convnd_fwd_xdl_sigmoid_fp16.cpp)
+add_example_dependencies(example_convnd_activ_unary_xdl example_convnd_fwd_xdl_sigmoid_fp16)
+# Tanh
+add_example_executable(example_convnd_fwd_xdl_tanh_fp16 convnd_fwd_xdl_tanh_fp16.cpp)
+add_example_dependencies(example_convnd_activ_unary_xdl example_convnd_fwd_xdl_tanh_fp16)
+# Relu
+add_example_executable(example_convnd_fwd_xdl_relu_fp16 convnd_fwd_xdl_relu_fp16.cpp)
+add_example_dependencies(example_convnd_activ_unary_xdl example_convnd_fwd_xdl_relu_fp16)
+# SoftRelu
+add_example_executable(example_convnd_fwd_xdl_softrelu_fp16 convnd_fwd_xdl_softrelu_fp16.cpp)
+add_example_dependencies(example_convnd_activ_unary_xdl example_convnd_fwd_xdl_softrelu_fp16)
+# Abs
+add_example_executable(example_convnd_fwd_xdl_abs_fp16 convnd_fwd_xdl_abs_fp16.cpp)
+add_example_dependencies(example_convnd_activ_unary_xdl example_convnd_fwd_xdl_abs_fp16)
+# Pow
+add_example_executable(example_convnd_fwd_xdl_pow_fp16 convnd_fwd_xdl_pow_fp16.cpp)
+add_example_dependencies(example_convnd_activ_unary_xdl example_convnd_fwd_xdl_pow_fp16)
+# Clipped Relu
+add_example_executable(example_convnd_fwd_xdl_clippedrelu_fp16 convnd_fwd_xdl_clippedrelu_fp16.cpp)
+add_example_dependencies(example_convnd_activ_unary_xdl example_convnd_fwd_xdl_clippedrelu_fp16)
+# Leaky Relu
+add_example_executable(example_convnd_fwd_xdl_leakyrelu_fp16 convnd_fwd_xdl_leakyrelu_fp16.cpp)
+add_example_dependencies(example_convnd_activ_unary_xdl example_convnd_fwd_xdl_leakyrelu_fp16)
+# Elu
+add_example_executable(example_convnd_fwd_xdl_elu_fp16 convnd_fwd_xdl_elu_fp16.cpp)
+add_example_dependencies(example_convnd_activ_unary_xdl example_convnd_fwd_xdl_elu_fp16)
+# Swish
+add_example_executable(example_convnd_fwd_xdl_swish_fp16 convnd_fwd_xdl_swish_fp16.cpp)
+add_example_dependencies(example_convnd_activ_unary_xdl example_convnd_fwd_xdl_swish_fp16)
+# PassThrough
+add_example_executable(example_convnd_fwd_xdl_passthrough_fp16 convnd_fwd_xdl_passthrough_fp16.cpp)
+add_example_dependencies(example_convnd_activ_unary_xdl example_convnd_fwd_xdl_passthrough_fp16)
+# Logistic
+add_example_executable(example_convnd_fwd_xdl_logistic_fp16 convnd_fwd_xdl_logistic_fp16.cpp)
+add_example_dependencies(example_convnd_activ_unary_xdl example_convnd_fwd_xdl_logistic_fp16)
--- a/example/65_gemm_multiply_multiply/CMakeLists.txt
+++ b/example/65_gemm_multiply_multiply/CMakeLists.txt
@@ -16,7 +16,7 @@ add_example_executable(example_moe_gemm2_xdl_fp8 moe_gemm2_xdl_fp8.cpp)
 add_example_executable(example_moe_gemm2_xdl_fp8_blockscale moe_gemm2_xdl_fp8_blockscale.cpp)
 add_example_executable(example_moe_gemm1_xdl_fp8_blockscale moe_gemm1_xdl_fp8_blockscale.cpp)

-list(APPEND gpu_list gfx942 gfx950)
+list(APPEND gpu_list gfx942 gfx950 gfx1100 gfx1101 gfx1102 gfx1103 gfx1150 gfx1151 gfx1152 gfx1200 gfx1201 gfx11-generic gfx12-generic)
 set(target 0)
 foreach(gpu IN LISTS GPU_TARGETS)
    if(gpu IN_LIST gpu_list AND target EQUAL 0)
--- a/example/CMakeLists.txt
+++ b/example/CMakeLists.txt
@@ -69,7 +69,7 @@ function(add_example_executable EXAMPLE_NAME FILE_NAME)
            list(REMOVE_ITEM FILE_NAME "${source}")
        endif()
        #Do not build any XDL examples if gfx9 targets are not on the list
-        if(NOT EX_TARGETS MATCHES "gfx9" AND source_name MATCHES "_xdl")
+        if(NOT EX_TARGETS MATCHES "gfx9" AND NOT EX_TARGETS MATCHES "gfx11" AND NOT EX_TARGETS MATCHES "gfx12" AND source_name MATCHES "_xdl")
            message(DEBUG "removing xdl example ${source} ")
            list(REMOVE_ITEM FILE_NAME "${source}")
        endif()
@@ -93,8 +93,8 @@ function(add_example_executable EXAMPLE_NAME FILE_NAME)
            message(DEBUG "removing bf8 example ${source} ")
            list(REMOVE_ITEM FILE_NAME "${source}")
        endif()
-        # Build fp8 gemm_multiply_multiply and moe only on gfx94/95
-        if(NOT EX_TARGETS MATCHES "gfx94" AND NOT EX_TARGETS MATCHES "gfx95")
+        # Build fp8 gemm_multiply_multiply and moe only on gfx94/95 and gfx12
+        if(NOT EX_TARGETS MATCHES "gfx94" AND NOT EX_TARGETS MATCHES "gfx95" AND NOT EX_TARGETS MATCHES "gfx12")
            if(source_name MATCHES "fp8" AND source_name MATCHES "(gemm_multiply_multiply|moe)")
                message(DEBUG "Skipping ${source} example for current target")
                list(REMOVE_ITEM FILE_NAME "${source}")
@@ -109,14 +109,14 @@ function(add_example_executable EXAMPLE_NAME FILE_NAME)
    endforeach()
    if(FILE_NAME)
        if(source_name_list MATCHES "_xdl" AND NOT source_name_list MATCHES "_pk_i4")
-            list(REMOVE_ITEM EX_TARGETS gfx900 gfx906 gfx906:xnack- gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1150 gfx1151 gfx1152 gfx1200 gfx1201 gfx10-3-generic gfx11-generic gfx12-generic)
+            list(REMOVE_ITEM EX_TARGETS gfx900 gfx906 gfx906:xnack- gfx1030 gfx10-3-generic)
        elseif(source_name_list MATCHES "_wmma")
            list(REMOVE_ITEM EX_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack+ gfx908:xnack- gfx90a:xnack+ gfx90a:xnack- gfx908 gfx90a gfx942 gfx1030 gfx950)
        elseif(source_name_list MATCHES "_mx") #only build mx example for gfx950
            list(REMOVE_ITEM EX_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack+ gfx908:xnack- gfx90a:xnack+ gfx90a:xnack- gfx908 gfx90a gfx942 gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1150 gfx1151 gfx1152 gfx1200 gfx1201 gfx10-3-generic gfx11-generic gfx12-generic)
-        elseif(source_name_list MATCHES "_pk_i4") #only build these examples for gfx942 and gfx950
+        elseif(source_name_list MATCHES "_pk_i4") #only build these examples for gfx942 gfx950 and rdna3/4
            message(DEBUG "trimming targets for ${FILE_NAME}")
-            list(REMOVE_ITEM EX_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack+ gfx908:xnack- gfx90a:xnack+ gfx90a:xnack- gfx908 gfx90a gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1150 gfx1151 gfx1152 gfx1200 gfx1201 gfx10-3-generic gfx11-generic gfx12-generic)
+            list(REMOVE_ITEM EX_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack+ gfx908:xnack- gfx90a:xnack+ gfx90a:xnack- gfx908 gfx90a gfx1030 gfx10-3-generic)
        endif()
        set_source_files_properties(${FILE_NAME} PROPERTIES LANGUAGE HIP)
        add_executable(${EXAMPLE_NAME} ${FILE_NAME})
@@ -192,7 +192,7 @@ function(add_example_executable_no_testing EXAMPLE_NAME FILE_NAME)
            list(REMOVE_ITEM FILE_NAME "${source}")
        endif()
        #Do not build any XDL examples if gfx9 targets are not on the list
-        if(NOT EX_TARGETS MATCHES "gfx9" AND source_name MATCHES "_xdl")
+        if(NOT EX_TARGETS MATCHES "gfx9" AND NOT EX_TARGETS MATCHES "gfx11" AND NOT EX_TARGETS MATCHES "gfx12" AND source_name MATCHES "_xdl")
            message(DEBUG "removing xdl example ${source} ")
            list(REMOVE_ITEM FILE_NAME "${source}")
        endif()
@@ -206,7 +206,7 @@ function(add_example_executable_no_testing EXAMPLE_NAME FILE_NAME)
    #only continue if there are some source files left on the list
    if(FILE_NAME)
        if(source_name_list MATCHES "_xdl")
-            list(REMOVE_ITEM EX_TARGETS gfx900 gfx906 gfx906:xnack- gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1150 gfx1151 gfx1152 gfx1200 gfx1201 gfx10-3-generic gfx11-generic gfx12-generic)
+            list(REMOVE_ITEM EX_TARGETS gfx900 gfx906 gfx906:xnack- gfx1030 gfx10-3-generic)
        elseif(source_name_list MATCHES "_wmma")
            list(REMOVE_ITEM EX_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack+ gfx908:xnack- gfx90a:xnack+ gfx90a:xnack- gfx908 gfx90a gfx942 gfx1030 gfx950)
        endif()