fix(cmake/midout): fix cmake midout

GitOrigin-RevId: 0fb93deb191ba235cbe312f7e7deaa979e810c9f

CMakeLists.txt

@@ -50,7 +50,7 @@ option(MGE_WITH_JIT "Build MegEngine with JIT." ON)
 option(MGE_WITH_JIT_MLIR "Build MegEngine with MLIR JIT." OFF)
 option(MGE_WITH_HALIDE "Build MegEngine with Halide JIT" OFF)
 option(MGE_WITH_MIDOUT_PROFILE "Build MegEngine with Midout profile." OFF)
-option(MGE_WITH_MINIMUM_SIZE "Swith off MGE_ENABLE_RTTI、MGE_ENABLE_EXCEPTIONS、MGE_ENABLE_LOGGING and switch on MGE_INFERENCE_ONLY so that compile minimum load_and_run. Take effect only when MGE_BIN_REDUCE was set" OFF)
+option(MGE_WITH_MINIMUM_SIZE "Swith off MGE_ENABLE_RTTI、MGE_ENABLE_EXCEPTIONS、MGE_ENABLE_LOGGING and switch on MGE_INFERENCE_ONLY so that compile minimum load_and_run." OFF)
 option(MGE_ARMV8_2_FEATURE_FP16 "Enable armv8.2-a+fp16 support" OFF)
 option(MGE_DISABLE_FLOAT16 "Disable MegEngine float16 support." OFF)
 option(MGE_WITH_CUDA "Enable MegEngine CUDA support." ON)
@@ -143,16 +143,32 @@ else()
     set(MGE_WITH_ANY_CUDA_STUB OFF)
 endif()
 
-if(NOT ${MGE_BIN_REDUCE} STREQUAL "")
-    message(STATUS "build with BIN REDUCE")
-    if(MGE_WITH_MINIMUM_SIZE)
-        set(MGE_ENABLE_RTTI OFF)
-        set(MGE_ENABLE_LOGGING OFF)
-        set(MGE_ENABLE_EXCEPTIONS OFF)
-        set(MGE_INFERENCE_ONLY ON)
+if(MGE_WITH_MIDOUT_PROFILE)
+    message(STATUS "build with MIDOUT PROFILE and force set MGE_WITH_MINIMUM_SIZE off and force rtti ON")
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DMIDOUT_PROFILING")
+    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DMIDOUT_PROFILING")
+    set(MGE_WITH_MINIMUM_SIZE OFF)
+    set(MGE_ENABLE_RTTI ON)
+    if(WIN32)
+        message(FATAL_ERROR "do not support midout at WIN32")
     endif()
-    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -include ${MGE_BIN_REDUCE}")
-    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -include ${MGE_BIN_REDUCE}")
+endif()
+
+set(BIN_REDUCE ${PROJECT_SOURCE_DIR}/src/bin_reduce_cmake.h)
+if(MGE_WITH_MINIMUM_SIZE)
+    message(STATUS "build with MGE_WITH_MINIMUM_SIZE bin_reduce header is: ${BIN_REDUCE}")
+    set(MGE_ENABLE_RTTI OFF)
+    set(MGE_ENABLE_LOGGING OFF)
+    set(MGE_ENABLE_EXCEPTIONS OFF)
+    set(MGE_INFERENCE_ONLY ON)
+    # MGE_WITH_MINIMUM_SIZE will triger unused-parameter
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-unused-parameter")
+    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -Wno-unused-parameter")
+endif()
+
+if(NOT MGE_WITH_MIDOUT_PROFILE AND NOT WIN32)
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -include ${BIN_REDUCE}")
+    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -include ${BIN_REDUCE}")
 endif()
 
 if (NOT APPLE)
@@ -176,12 +192,6 @@ else()
     message(STATUS "lto is not supported in this compiler")
 endif()
 
-if(MGE_WITH_MIDOUT_PROFILE)
-    message(STATUS "build with MIDOUT PROFILE")
-    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DMIDOUT_PROFILING")
-    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DMIDOUT_PROFILING")
-endif()
-
 if (APPLE)
     set (BUILD_SHARED_LIBS OFF)
     message(STATUS "build static for xcode framework require")
@@ -253,9 +263,9 @@ if(NOT CMAKE_CONFIGURATION_TYPES AND NOT CMAKE_BUILD_TYPE)
     set(CMAKE_BUILD_TYPE RelWithDebInfo)
 endif()
 
-if(${CMAKE_BUILD_TYPE} STREQUAL "Release" AND NOT MGE_WITH_TEST AND NOT ${MGE_ARCH} STREQUAL "x86_64")
+if(${CMAKE_BUILD_TYPE} STREQUAL "Release" AND NOT MGE_WITH_TEST AND NOT ${MGE_ARCH} STREQUAL "x86_64" AND NOT MGE_WITH_MIDOUT_PROFILE)
     set(MGE_ENABLE_RTTI OFF)
-    message(STATUS "disable MGE_ENABLE_RTTI when Release/NON-x86_64 mode!!")
+    message(STATUS "disable MGE_ENABLE_RTTI when Release/NON-x86_64/NON-MGE_WITH_MIDOUT_PROFILE mode!!")
 endif()
 
 if(MSVC OR WIN32)

lite/src/mge/common.cpp

@@ -83,9 +83,11 @@ LTensorLayout lite::to_impl_layout(const Layout& layout) {
         case LiteDataType::LITE_FLOAT:
             mge_layout.dtype = mgb::dtype::Float32();
             break;
+#if !MEGDNN_DISABLE_FLOAT16
         case LiteDataType::LITE_HALF:
             mge_layout.dtype = mgb::dtype::Float16();
             break;
+#endif
         case LiteDataType::LITE_INT:
             mge_layout.dtype = mgb::dtype::Int32();
             break;
@@ -120,9 +122,11 @@ Layout lite::to_lite_layout(const LTensorLayout& mge_layout) {
         case mgb::DTypeEnum::Float32:
             layout.data_type = LiteDataType::LITE_FLOAT;
             break;
+#if !MEGDNN_DISABLE_FLOAT16
         case mgb::DTypeEnum::Float16:
             layout.data_type = LiteDataType::LITE_HALF;
             break;
+#endif
         case mgb::DTypeEnum::Int32:
             layout.data_type = LiteDataType::LITE_INT;
             break;

src/bin_reduce_cmake.h

+/**
+ * \file src/bin_reduce_cmake.h
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ */
+
+/*
+ * always empty, override by ../tools/gen_header_for_bin_reduce.py
+ */

src/megbrain_build_config.h.in

 /**
- * \file src/core/include/megbrain_build_config.h
+ * \file src/megbrain_build_config.h.in
  * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
  *
  * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.

tools/gen_header_for_bin_reduce.py

@@ -53,7 +53,16 @@ class HeaderGen:
                cls._megvii3_root_cache = str(wd)
                return cls._megvii3_root_cache
            wd = wd.parent
-        raise RuntimeError('This script is supposed to run in megvii3.')
+        return None
+
+    _megengine_root_cache = None
+    @classmethod
+    def get_megengine_root(cls):
+        if cls._megengine_root_cache is not None:
+            return cls._megengine_root_cache
+        wd = Path(__file__).resolve().parent.parent
+        cls._megengine_root_cache = str(wd)
+        return cls._megengine_root_cache
 
     def extend_netinfo(self, data):
         self._has_netinfo = True
@@ -125,11 +134,11 @@ class HeaderGen:
         with tempfile.NamedTemporaryFile() as ftmp:
             fpath = os.path.realpath(ftmp.name)
             subprocess.check_call(
-                ['./brain/megbrain/dnn/scripts/gen_param_defs.py',
+                ['./dnn/scripts/gen_param_defs.py',
                  '--write-enum-items', 'Elemwise:Mode',
-                 './brain/megbrain/dnn/scripts/opr_param_defs.py',
+                 './dnn/scripts/opr_param_defs.py',
                  fpath],
-                cwd=self.get_megvii3_root()
+                cwd=self.get_megengine_root()
             )
 
             with open(fpath) as fin:
@@ -178,8 +187,10 @@ class HeaderGen:
         if not self._midout_files:
             return
 
-        gen = os.path.join(self.get_megvii3_root(), 'brain', 'midout',
-                           'gen_header.py')
+        gen = os.path.join(self.get_megengine_root(), 'third_party', 'midout', 'gen_header.py')
+        if self.get_megvii3_root():
+            gen = os.path.join(self.get_megvii3_root(), 'brain', 'midout', 'gen_header.py')
+        print('use {} to gen bin_reduce header'.format(gen))
         cvt = subprocess.run(
             [gen] + self._midout_files,
             stdout=subprocess.PIPE, check=True,
@@ -212,10 +223,13 @@ def main():
         '  1. json files generated by '
         'megbrain.serialize_comp_graph_to_file() in python; '
         '  2. trace files generated by midout library')
-    parser.add_argument('-o', '--output', help='output file',
-                        default=os.path.join(HeaderGen.get_megvii3_root(),
-                                             'utils', 'bin_reduce.h'))
+    default_file=os.path.join(HeaderGen.get_megengine_root(), 'src', 'bin_reduce_cmake.h')
+    is_megvii3 = HeaderGen.get_megvii3_root()
+    if is_megvii3:
+        default_file=os.path.join(HeaderGen.get_megvii3_root(), 'utils', 'bin_reduce.h')
+    parser.add_argument('-o', '--output', help='output file', default=default_file)
     args = parser.parse_args()
+    print('config output file: {}'.format(args.output))
 
     gen = HeaderGen()
     for i in args.inputs: