Add per-example gfx arch checks

j-stephan · j-stephan · commit bc4819e080d7 · 2025-11-10T08:49:05.000-05:00
Signed-off-by: Jan Stephan &lt;jan.stephan@amd.com&gt;
diff --git a/Libraries/ComposableKernel/CMakeLists.txt b/Libraries/ComposableKernel/CMakeLists.txt
@@ -31,14 +31,6 @@ endif()
 file(RELATIVE_PATH folder_bin ${CMAKE_SOURCE_DIR} ${CMAKE_CURRENT_SOURCE_DIR})
 set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin/${folder_bin})
 
-set(CK_SUPPORTED_ARCH gfx90a gfx942 gfx950)
-
-foreach(ARCH ${CMAKE_HIP_ARCHITECTURES})
-    if(NOT ARCH IN_LIST CK_SUPPORTED_ARCH)
-        message(STATUS "CK Tile examples are not supported for ${ARCH}. Not building CK Tile examples")
-        return()
-    endif()
-endforeach()
 
 include(CTest)
 
diff --git a/Libraries/ComposableKernel/attention/fmha/CMakeLists.txt b/Libraries/ComposableKernel/attention/fmha/CMakeLists.txt
@@ -41,6 +41,14 @@ select_gpu_language()
 enable_language(${ROCM_EXAMPLES_GPU_LANGUAGE})
 select_hip_platform()
 
+set(EXAMPLE_SUPPORTED_ARCH gfx942)
+foreach(ARCH ${CMAKE_HIP_ARCHITECTURES})
+    if(NOT ARCH IN_LIST EXAMPLE_SUPPORTED_ARCH)
+        message(STATUS "CK Tile FMHA examples are not supported for ${ARCH}. Not building CK Tile FMHA examples")
+        return()
+    endif()
+endforeach()
+
 if(ROCM_EXAMPLES_GPU_LANGUAGE STREQUAL "CUDA")
     message(STATUS "Composable Kernel examples are not supported for CUDA.")
     return()
diff --git a/Libraries/ComposableKernel/attention/fmha/Makefile b/Libraries/ComposableKernel/attention/fmha/Makefile
@@ -44,7 +44,7 @@ ICPPFLAGS += $(CPPFLAGS)
 ILDFLAGS  += $(LDFLAGS)
 ILDLIBS   += $(LDLIBS)
 
-FWD_PP_DEFS := -DCK_TILE_FMHA_FWD_FAST_EXP2=1 -DCK_TILE_FMHA_FWD_SPLITKV_API=1 -DCK_TILE_FMHA_FWD_APPENDKV_API=1 -DCK_TILE_FMHA_FWD_PAGEDKV_API=1
+FWD_PP_DEFS := -DCK_TILE_FMHA_FWD_FAST_EXP2=1
 FWD_SRC_DIR := fwd_sources
 FWD_OBJ_DIR := fwd_objects
 BWD_SRC_DIR := bwd_sources
@@ -64,11 +64,11 @@ $(FWD_SRC_DIR) $(FWD_OBJ_DIR) $(BWD_SRC_DIR) $(BWD_OBJ_DIR):
 generate: $(FWD_GENERATION_DONE) $(BWD_GENERATION_DONE)
 
 $(FWD_GENERATION_DONE): $(EXAMPLE_GENERATOR) | $(FWD_SRC_DIR)
-	python3 $(EXAMPLE_GENERATOR) --api fwd,fwd_splitkv,fwd_appendkv --optdim 32,64,128,256 --output_dir $(FWD_SRC_DIR)
+	python3 $(EXAMPLE_GENERATOR) --api fwd --receipt 2 --optdim 64 --filter "*fp16_group*" --output_dir $(FWD_SRC_DIR)
 	@touch $@
 
 $(BWD_GENERATION_DONE): $(EXAMPLE_GENERATOR) | $(BWD_SRC_DIR)
-	python3 $(EXAMPLE_GENERATOR) --api bwd --receipt 3 --optdim 32,64,128,256 --output_dir $(BWD_SRC_DIR)
+	python3 $(EXAMPLE_GENERATOR) --api bwd --receipt 3 --optdim 64 --filter "*fp16_group*" --output_dir $(BWD_SRC_DIR)
 	@touch $@
 
 FWD_SRCS := fmha_fwd.cpp
diff --git a/Libraries/ComposableKernel/attention/fmha/README.md b/Libraries/ComposableKernel/attention/fmha/README.md
@@ -7,14 +7,7 @@ with the `_fwd` suffix contain the algorithm's forward pass, while the `_bwd` fi
 
 ### Supported architectures
 
-The forward pass example is supported for the following GPU architectures:
-
-* `gfx908`
-* `gfx90a`
-* `gfx942`
-* `gfx950`
-
-The backward pass is only supported for the `gfx942` architecture.
+The example is only supported for the `gfx942` architecture.
 
 ### Application flow
 
diff --git a/Libraries/ComposableKernel/basic/batched_transpose/CMakeLists.txt b/Libraries/ComposableKernel/basic/batched_transpose/CMakeLists.txt
@@ -36,6 +36,14 @@ select_gpu_language()
 enable_language(${ROCM_EXAMPLES_GPU_LANGUAGE})
 select_hip_platform()
 
+set(EXAMPLE_SUPPORTED_ARCH gfx908 gfx90a gfx942 gfx950)
+foreach(ARCH ${CMAKE_HIP_ARCHITECTURES})
+    if(NOT ARCH IN_LIST EXAMPLE_SUPPORTED_ARCH)
+        message(STATUS "CK Tile batched transpose example is not supported for ${ARCH}. Not building CK Tile batched transpose example")
+        return()
+    endif()
+endforeach()
+
 if(ROCM_EXAMPLES_GPU_LANGUAGE STREQUAL "CUDA")
     message(STATUS "Composable Kernel examples are not supported for CUDA.")
     return()
diff --git a/Libraries/ComposableKernel/basic/permute/CMakeLists.txt b/Libraries/ComposableKernel/basic/permute/CMakeLists.txt
@@ -37,6 +37,14 @@ select_gpu_language()
 enable_language(${ROCM_EXAMPLES_GPU_LANGUAGE})
 select_hip_platform()
 
+set(EXAMPLE_SUPPORTED_ARCH gfx908 gfx90a gfx942 gfx950)
+foreach(ARCH ${CMAKE_HIP_ARCHITECTURES})
+    if(NOT ARCH IN_LIST EXAMPLE_SUPPORTED_ARCH)
+        message(STATUS "CK Tile permute example is not supported for ${ARCH}. Not building CK Tile permute example")
+        return()
+    endif()
+endforeach()
+
 if(ROCM_EXAMPLES_GPU_LANGUAGE STREQUAL "CUDA")
     message(STATUS "Composable Kernel examples are not supported for CUDA.")
     return()
diff --git a/Libraries/ComposableKernel/convolution/grouped_convolution/CMakeLists.txt b/Libraries/ComposableKernel/convolution/grouped_convolution/CMakeLists.txt
@@ -37,6 +37,14 @@ select_gpu_language()
 enable_language(${ROCM_EXAMPLES_GPU_LANGUAGE})
 select_hip_platform()
 
+set(EXAMPLE_SUPPORTED_ARCH gfx908 gfx90a gfx942 gfx950)
+foreach(ARCH ${CMAKE_HIP_ARCHITECTURES})
+    if(NOT ARCH IN_LIST EXAMPLE_SUPPORTED_ARCH)
+        message(STATUS "CK Tile grouped convolution example is not supported for ${ARCH}. Not building CK Tile grouped convolution example")
+        return()
+    endif()
+endforeach()
+
 if(ROCM_EXAMPLES_GPU_LANGUAGE STREQUAL "CUDA")
     message(STATUS "Composable Kernel examples are not supported for CUDA.")
     return()
diff --git a/Libraries/ComposableKernel/gemm/batched_gemm/CMakeLists.txt b/Libraries/ComposableKernel/gemm/batched_gemm/CMakeLists.txt
@@ -35,6 +35,14 @@ select_gpu_language()
 enable_language(${ROCM_EXAMPLES_GPU_LANGUAGE})
 select_hip_platform()
 
+set(EXAMPLE_SUPPORTED_ARCH gfx908 gfx90a gfx942 gfx950)
+foreach(ARCH ${CMAKE_HIP_ARCHITECTURES})
+    if(NOT ARCH IN_LIST EXAMPLE_SUPPORTED_ARCH)
+        message(STATUS "CK Tile batched GEMM example is not supported for ${ARCH}. Not building CK Tile batched GEMM example")
+        return()
+    endif()
+endforeach()
+
 if(ROCM_EXAMPLES_GPU_LANGUAGE STREQUAL "CUDA")
     message(STATUS "Composable Kernel examples are not supported for CUDA.")
     return()
diff --git a/Libraries/ComposableKernel/gemm/block_scale_gemm/CMakeLists.txt b/Libraries/ComposableKernel/gemm/block_scale_gemm/CMakeLists.txt
@@ -35,6 +35,14 @@ select_gpu_language()
 enable_language(${ROCM_EXAMPLES_GPU_LANGUAGE})
 select_hip_platform()
 
+set(EXAMPLE_SUPPORTED_ARCH gfx942 gfx950)
+foreach(ARCH ${CMAKE_HIP_ARCHITECTURES})
+    if(NOT ARCH IN_LIST EXAMPLE_SUPPORTED_ARCH)
+        message(STATUS "CK Tile block-scale GEMM example is not supported for ${ARCH}. Not building CK Tile block-scale GEMM example")
+        return()
+    endif()
+endforeach()
+
 if(ROCM_EXAMPLES_GPU_LANGUAGE STREQUAL "CUDA")
     message(STATUS "Composable Kernel examples are not supported for CUDA.")
     return()
diff --git a/Libraries/ComposableKernel/gemm/flatmm/CMakeLists.txt b/Libraries/ComposableKernel/gemm/flatmm/CMakeLists.txt
@@ -36,6 +36,14 @@ select_gpu_language()
 enable_language(${ROCM_EXAMPLES_GPU_LANGUAGE})
 select_hip_platform()
 
+set(EXAMPLE_SUPPORTED_ARCH gfx908 gfx90a gfx942 gfx950)
+foreach(ARCH ${CMAKE_HIP_ARCHITECTURES})
+    if(NOT ARCH IN_LIST EXAMPLE_SUPPORTED_ARCH)
+        message(STATUS "CK Tile FLATMM example is not supported for ${ARCH}. Not building CK Tile FLATMM example")
+        return()
+    endif()
+endforeach()
+
 if(ROCM_EXAMPLES_GPU_LANGUAGE STREQUAL "CUDA")
     message(STATUS "Composable Kernel examples are not supported for CUDA.")
     return()
diff --git a/Libraries/ComposableKernel/gemm/gemm/CMakeLists.txt b/Libraries/ComposableKernel/gemm/gemm/CMakeLists.txt
@@ -40,6 +40,14 @@ select_gpu_language()
 enable_language(${ROCM_EXAMPLES_GPU_LANGUAGE})
 select_hip_platform()
 
+set(EXAMPLE_SUPPORTED_ARCH gfx908 gfx90a gfx942 gfx950)
+foreach(ARCH ${CMAKE_HIP_ARCHITECTURES})
+    if(NOT ARCH IN_LIST EXAMPLE_SUPPORTED_ARCH)
+        message(STATUS "CK Tile GEMM example is not supported for ${ARCH}. Not building CK Tile GEMM example")
+        return()
+    endif()
+endforeach()
+
 if(ROCM_EXAMPLES_GPU_LANGUAGE STREQUAL "CUDA")
     message(STATUS "Composable Kernel examples are not supported for CUDA.")
     return()
diff --git a/Libraries/ComposableKernel/gemm/gemm_multi_d/CMakeLists.txt b/Libraries/ComposableKernel/gemm/gemm_multi_d/CMakeLists.txt
@@ -35,6 +35,14 @@ select_gpu_language()
 enable_language(${ROCM_EXAMPLES_GPU_LANGUAGE})
 select_hip_platform()
 
+set(EXAMPLE_SUPPORTED_ARCH gfx908 gfx90a gfx942 gfx950)
+foreach(ARCH ${CMAKE_HIP_ARCHITECTURES})
+    if(NOT ARCH IN_LIST EXAMPLE_SUPPORTED_ARCH)
+        message(STATUS "CK Tile batched multi-D GEMM is not supported for ${ARCH}. Not building CK Tile multi-D GEMM example")
+        return()
+    endif()
+endforeach()
+
 if(ROCM_EXAMPLES_GPU_LANGUAGE STREQUAL "CUDA")
     message(STATUS "Composable Kernel examples are not supported for CUDA.")
     return()
diff --git a/Libraries/ComposableKernel/gemm/grouped_gemm/CMakeLists.txt b/Libraries/ComposableKernel/gemm/grouped_gemm/CMakeLists.txt
@@ -36,6 +36,14 @@ select_gpu_language()
 enable_language(${ROCM_EXAMPLES_GPU_LANGUAGE})
 select_hip_platform()
 
+set(EXAMPLE_SUPPORTED_ARCH gfx908 gfx90a gfx942 gfx950)
+foreach(ARCH ${CMAKE_HIP_ARCHITECTURES})
+    if(NOT ARCH IN_LIST EXAMPLE_SUPPORTED_ARCH)
+        message(STATUS "CK Tile grouped GEMM example is not supported for ${ARCH}. Not building CK Tile grouped GEMM example")
+        return()
+    endif()
+endforeach()
+
 if(ROCM_EXAMPLES_GPU_LANGUAGE STREQUAL "CUDA")
     message(STATUS "Composable Kernel examples are not supported for CUDA.")
     return()
diff --git a/Libraries/ComposableKernel/moe/fused_moe/CMakeLists.txt b/Libraries/ComposableKernel/moe/fused_moe/CMakeLists.txt
@@ -35,6 +35,14 @@ select_gpu_language()
 enable_language(${ROCM_EXAMPLES_GPU_LANGUAGE})
 select_hip_platform()
 
+set(EXAMPLE_SUPPORTED_ARCH gfx942 gfx950)
+foreach(ARCH ${CMAKE_HIP_ARCHITECTURES})
+    if(NOT ARCH IN_LIST EXAMPLE_SUPPORTED_ARCH)
+        message(STATUS "CK Tile fused MoE example is not supported for ${ARCH}. Not building CK Tile fused MoE example")
+        return()
+    endif()
+endforeach()
+
 if(ROCM_EXAMPLES_GPU_LANGUAGE STREQUAL "CUDA")
     message(STATUS "Composable Kernel examples are not supported for CUDA.")
     return()
diff --git a/Libraries/ComposableKernel/moe/moe_smoothquant/CMakeLists.txt b/Libraries/ComposableKernel/moe/moe_smoothquant/CMakeLists.txt
@@ -36,6 +36,14 @@ select_gpu_language()
 enable_language(${ROCM_EXAMPLES_GPU_LANGUAGE})
 select_hip_platform()
 
+set(EXAMPLE_SUPPORTED_ARCH gfx908 gfx90a gfx942 gfx950)
+foreach(ARCH ${CMAKE_HIP_ARCHITECTURES})
+    if(NOT ARCH IN_LIST EXAMPLE_SUPPORTED_ARCH)
+        message(STATUS "CK Tile MoE-SmoothQuant example is not supported for ${ARCH}. Not building CK Tile MoE-SmoothQuant example")
+        return()
+    endif()
+endforeach()
+
 if(ROCM_EXAMPLES_GPU_LANGUAGE STREQUAL "CUDA")
     message(STATUS "Composable Kernel examples are not supported for CUDA.")
     return()
diff --git a/Libraries/ComposableKernel/moe/moe_sorting/CMakeLists.txt b/Libraries/ComposableKernel/moe/moe_sorting/CMakeLists.txt
@@ -36,6 +36,14 @@ select_gpu_language()
 enable_language(${ROCM_EXAMPLES_GPU_LANGUAGE})
 select_hip_platform()
 
+set(EXAMPLE_SUPPORTED_ARCH gfx90a gfx942 gfx950)
+foreach(ARCH ${CMAKE_HIP_ARCHITECTURES})
+    if(NOT ARCH IN_LIST EXAMPLE_SUPPORTED_ARCH)
+        message(STATUS "CK Tile MoE sorting example is not supported for ${ARCH}. Not building CK Tile MoE sorting example")
+        return()
+    endif()
+endforeach()
+
 if(ROCM_EXAMPLES_GPU_LANGUAGE STREQUAL "CUDA")
     message(STATUS "Composable Kernel examples are not supported for CUDA.")
     return()
diff --git a/Libraries/ComposableKernel/normalization/add_rmsnorm2d_rdquant/CMakeLists.txt b/Libraries/ComposableKernel/normalization/add_rmsnorm2d_rdquant/CMakeLists.txt
@@ -37,6 +37,14 @@ select_gpu_language()
 enable_language(${ROCM_EXAMPLES_GPU_LANGUAGE})
 select_hip_platform()
 
+set(EXAMPLE_SUPPORTED_ARCH gfx908 gfx90a gfx942 gfx950)
+foreach(ARCH ${CMAKE_HIP_ARCHITECTURES})
+    if(NOT ARCH IN_LIST EXAMPLE_SUPPORTED_ARCH)
+        message(STATUS "CK Tile Add + RMSNorm + row-wise dynamic quantization example is not supported for ${ARCH}. Not building CK Tile Add + RMSNorm + row-wise dynamic quantization example")
+        return()
+    endif()
+endforeach()
+
 if(ROCM_EXAMPLES_GPU_LANGUAGE STREQUAL "CUDA")
     message(STATUS "Composable Kernel examples are not supported for CUDA.")
     return()
diff --git a/Libraries/ComposableKernel/normalization/layernorm2d/CMakeLists.txt b/Libraries/ComposableKernel/normalization/layernorm2d/CMakeLists.txt
@@ -36,6 +36,14 @@ select_gpu_language()
 enable_language(${ROCM_EXAMPLES_GPU_LANGUAGE})
 select_hip_platform()
 
+set(EXAMPLE_SUPPORTED_ARCH gfx908 gfx90a gfx942 gfx950)
+foreach(ARCH ${CMAKE_HIP_ARCHITECTURES})
+    if(NOT ARCH IN_LIST EXAMPLE_SUPPORTED_ARCH)
+        message(STATUS "CK Tile Layernorm2D example is not supported for ${ARCH}. Not building CK Tile Layernorm2D example")
+        return()
+    endif()
+endforeach()
+
 if(ROCM_EXAMPLES_GPU_LANGUAGE STREQUAL "CUDA")
     message(STATUS "Composable Kernel examples are not supported for CUDA.")
     return()
diff --git a/Libraries/ComposableKernel/normalization/layernorm2d/README.md b/Libraries/ComposableKernel/normalization/layernorm2d/README.md
@@ -5,6 +5,15 @@
 This example demonstrates how to perform the forward pass of the Layernorm2D operation with the CK Tile programming
 model.
 
+### Supported architectures
+
+The example is supported for the following GPU architectures:
+
+* `gfx908`
+* `gfx90a`
+* `gfx942`
+* `gfx950`
+
 ### Application flow
 
 1. Command line arguments are parsed to configure matrix dimensions and execution parameters.
diff --git a/Libraries/ComposableKernel/normalization/rmsnorm2d/CMakeLists.txt b/Libraries/ComposableKernel/normalization/rmsnorm2d/CMakeLists.txt
@@ -37,6 +37,14 @@ select_gpu_language()
 enable_language(${ROCM_EXAMPLES_GPU_LANGUAGE})
 select_hip_platform()
 
+set(EXAMPLE_SUPPORTED_ARCH gfx908 gfx90a gfx942 gfx950)
+foreach(ARCH ${CMAKE_HIP_ARCHITECTURES})
+    if(NOT ARCH IN_LIST EXAMPLE_SUPPORTED_ARCH)
+        message(STATUS "CK Tile RMSNorm2D example is not supported for ${ARCH}. Not building CK Tile RMSNorm2D example")
+        return()
+    endif()
+endforeach()
+
 if(ROCM_EXAMPLES_GPU_LANGUAGE STREQUAL "CUDA")
     message(STATUS "Composable Kernel examples are not supported for CUDA.")
     return()
diff --git a/Libraries/ComposableKernel/quantization/smoothquant/CMakeLists.txt b/Libraries/ComposableKernel/quantization/smoothquant/CMakeLists.txt
@@ -37,6 +37,14 @@ select_gpu_language()
 enable_language(${ROCM_EXAMPLES_GPU_LANGUAGE})
 select_hip_platform()
 
+set(EXAMPLE_SUPPORTED_ARCH gfx908 gfx90a gfx942 gfx950)
+foreach(ARCH ${CMAKE_HIP_ARCHITECTURES})
+    if(NOT ARCH IN_LIST EXAMPLE_SUPPORTED_ARCH)
+        message(STATUS "CK Tile SmoothQuant example is not supported for ${ARCH}. Not building CK Tile SmoothQuant example")
+        return()
+    endif()
+endforeach()
+
 if(ROCM_EXAMPLES_GPU_LANGUAGE STREQUAL "CUDA")
     message(STATUS "Composable Kernel examples are not supported for CUDA.")
     return()
diff --git a/Libraries/ComposableKernel/quantization/smoothquant/README.md b/Libraries/ComposableKernel/quantization/smoothquant/README.md
@@ -5,6 +5,15 @@
 This example shows how to implement SmoothQuant using the CK Tile programming model. There are two variants: One basic
 example with a fixed parameter set (`example_`) and one example with various available parameters.
 
+### Supported architectures
+
+The example is supported for the following GPU architectures:
+
+* `gfx908`
+* `gfx90a`
+* `gfx942`
+* `gfx950`
+
 ### Application flow
 
 1. Command line arguments are parsed to configure matrix dimensions and execution parameters.
