perf(mgb/compile): improve compile time for megbrain

GitOrigin-RevId: 12d7467c8bccc07731646ff307dba71ea983d88f

CMakeLists.txt

@@ -97,6 +97,11 @@ option(MGE_WITH_LARGE_ARCHIVE "Enable big archive link support" OFF)
 option(MGE_BUILD_WITH_ASAN "Enable build with ASAN, need compiler support" OFF)
 option(MGE_WITH_CUSTOM_OP "Build with Custom op" OFF)
 option(MGE_SYNC_THIRD_PARTY "help sync third_party submodule" OFF)
+option(MGE_PROFILE_COMPILE_TIME "help profile compile time per file" OFF)
+
+if(MGE_PROFILE_COMPILE_TIME)
+  set_property(GLOBAL PROPERTY RULE_LAUNCH_COMPILE "cmake -E time")
+endif()
 
 # TODO: add windows support
 cmake_dependent_option(MGE_WITH_CUPTI "Build with CUPTI" ON

dnn/src/CMakeLists.txt

@@ -188,8 +188,21 @@ if(MGE_WITH_CUDA)
   gen_cutlass_kimpl(dwconv2d_dgrad tensorop884 CUTLASS_SOURCES)
   gen_cutlass_kimpl(dwconv2d_wgrad simt CUTLASS_SOURCES)
   gen_cutlass_kimpl(dwconv2d_wgrad tensorop884 CUTLASS_SOURCES)
-  list(APPEND SOURCES ${CUTLASS_SOURCES})
-  list(APPEND SOURCES ${CUSOURCES})
+  # Compile the following opr first
+  file(
+    GLOB_RECURSE
+    PRIORITY_DIRS
+    cuda/type_cvt/*cu
+    cuda/conv_bias/*cu
+    cuda/reduce/*cu
+    cuda/relayout/*cu
+    cuda/relayout_format/*cu
+    cuda/powc/*cu
+    cuda/warp_perspective/*cu)
+  list(PREPEND PRIORITY_DIRS ${CUTLASS_SOURCES})
+  list(PREPEND CUSOURCES ${PRIORITY_DIRS})
+  list(REMOVE_DUPLICATES CUSOURCES)
+  list(PREPEND SOURCES ${CUSOURCES})
 endif()
 
 if(MGE_WITH_ATLAS)