CMake: merge two condition blocks (#273)

Author: danieldk

build2cmake/src/templates/preamble.cmake

@@ -78,7 +78,6 @@ else()
   set(CUDA_DEFAULT_KERNEL_ARCHS "7.0;7.2;7.5;8.0;8.6;8.7;8.9;9.0+PTX")
 endif()
 
-
 # Basic checks for each GPU language.
 if(GPU_LANG STREQUAL "CUDA")
   if(NOT CUDA_FOUND)
@@ -99,6 +98,31 @@ if(GPU_LANG STREQUAL "CUDA")
     endif()
   {% endif %}
 
+  # This clears out -gencode arguments from `CMAKE_CUDA_FLAGS`, which we need
+  # to set our own set of capabilities.
+  clear_gencode_flags()
+
+  # Get the capabilities without +PTX suffixes, so that we can use them as
+  # the target archs in the loose intersection with a kernel's capabilities.
+  cuda_remove_ptx_suffixes(CUDA_ARCHS "${CUDA_DEFAULT_KERNEL_ARCHS}")
+  message(STATUS "CUDA supported base architectures: ${CUDA_ARCHS}")
+
+  if(BUILD_ALL_SUPPORTED_ARCHS)
+    set(CUDA_KERNEL_ARCHS "${CUDA_DEFAULT_KERNEL_ARCHS}")
+  else()
+    try_run_python(CUDA_KERNEL_ARCHS SUCCESS "import torch; cc=torch.cuda.get_device_capability(); print(f\"{cc[0]}.{cc[1]}\")" "Failed to get CUDA capability")
+    if(NOT SUCCESS)
+      message(WARNING "Failed to detect CUDA capability, using default capabilities.")
+      set(CUDA_KERNEL_ARCHS "${CUDA_DEFAULT_KERNEL_ARCHS}")
+    endif()
+  endif()
+
+  message(STATUS "CUDA supported kernel architectures: ${CUDA_KERNEL_ARCHS}")
+
+  if(NVCC_THREADS AND GPU_LANG STREQUAL "CUDA")
+    list(APPEND GPU_FLAGS "--threads=${NVCC_THREADS}")
+  endif()
+
   # TODO: deprecate one of these settings.
   add_compile_definitions(USE_CUDA=1)
   add_compile_definitions(CUDA_KERNEL)
@@ -112,6 +136,10 @@ elseif(GPU_LANG STREQUAL "HIP")
   # .hip extension automatically, HIP must be enabled explicitly.
   enable_language(HIP)
 
+  override_gpu_arches(GPU_ARCHES HIP ${HIP_SUPPORTED_ARCHS})
+  set(ROCM_ARCHS ${GPU_ARCHES})
+  message(STATUS "ROCM supported target architectures: ${ROCM_ARCHS}")
+
   # TODO: deprecate one of these settings.
   add_compile_definitions(USE_ROCM=1)
   add_compile_definitions(ROCM_KERNEL)
@@ -132,44 +160,6 @@ elseif(GPU_LANG STREQUAL "SYCL")
     message(FATAL_ERROR "Intel SYCL C++ compiler (icpx) and/or C compiler (icx) not found. Please install Intel oneAPI toolkit.")
   endif()
 
-  add_compile_definitions(XPU_KERNEL)
-  add_compile_definitions(USE_XPU)
-else()
-  message(FATAL_ERROR "Unsupported GPU language: ${GPU_LANG}")
-endif()
-
-# CUDA build options.
-if(GPU_LANG STREQUAL "CUDA")
-  # This clears out -gencode arguments from `CMAKE_CUDA_FLAGS`, which we need
-  # to set our own set of capabilities.
-  clear_gencode_flags()
-
-  # Get the capabilities without +PTX suffixes, so that we can use them as
-  # the target archs in the loose intersection with a kernel's capabilities.
-  cuda_remove_ptx_suffixes(CUDA_ARCHS "${CUDA_DEFAULT_KERNEL_ARCHS}")
-  message(STATUS "CUDA supported base architectures: ${CUDA_ARCHS}")
-
-  if(BUILD_ALL_SUPPORTED_ARCHS)
-    set(CUDA_KERNEL_ARCHS "${CUDA_DEFAULT_KERNEL_ARCHS}")
-  else()
-    try_run_python(CUDA_KERNEL_ARCHS SUCCESS "import torch; cc=torch.cuda.get_device_capability(); print(f\"{cc[0]}.{cc[1]}\")" "Failed to get CUDA capability")
-    if(NOT SUCCESS)
-      message(WARNING "Failed to detect CUDA capability, using default capabilities.")
-      set(CUDA_KERNEL_ARCHS "${CUDA_DEFAULT_KERNEL_ARCHS}")
-    endif()
-  endif()
-
-  message(STATUS "CUDA supported kernel architectures: ${CUDA_KERNEL_ARCHS}")
-
-  if(NVCC_THREADS AND GPU_LANG STREQUAL "CUDA")
-    list(APPEND GPU_FLAGS "--threads=${NVCC_THREADS}")
-  endif()
-
-elseif(GPU_LANG STREQUAL "HIP")
-  override_gpu_arches(GPU_ARCHES HIP ${HIP_SUPPORTED_ARCHS})
-  set(ROCM_ARCHS ${GPU_ARCHES})
-  message(STATUS "ROCM supported target architectures: ${ROCM_ARCHS}")
-elseif(GPU_LANG STREQUAL "SYCL")
   execute_process(
     COMMAND ${ICPX_COMPILER} --version
     OUTPUT_VARIABLE ICPX_VERSION_OUTPUT
@@ -191,10 +181,12 @@ elseif(GPU_LANG STREQUAL "SYCL")
   set(sycl_flags "-fsycl;-fhonor-nans;-fhonor-infinities;-fno-associative-math;-fno-approx-func;-fno-sycl-instrument-device-code;--offload-compress;-fsycl-targets=spir64_gen,spir64;")
   set(GPU_FLAGS "${sycl_flags}")
   set(GPU_ARCHES "")
+
+
+  add_compile_definitions(XPU_KERNEL)
+  add_compile_definitions(USE_XPU)
 else()
-  override_gpu_arches(GPU_ARCHES
-    ${GPU_LANG}
-    "${${GPU_LANG}_SUPPORTED_ARCHS}")
+  message(FATAL_ERROR "Unsupported GPU language: ${GPU_LANG}")
 endif()
 
 # Initialize SRC list for kernel and binding sources