support`-flto` #310

headupinclouds · 2017-02-18T16:43:28Z

UPDATE 2/21/17: clang toolchains seems to work for iOS and OS X builds. Android NDK toolchains (both clang and gcc hit various roadblocks. See discussion below

See: http://stackoverflow.com/a/25649861 : -flto -O3 may optimize further but increase size, while -flto -Os should reduce size. Also, -flto -Os (or whatever optimization setting is used) must be used consistently for compilation and linking within the project and all external (Hunter) dependencies. Thus, -flto belongs in the toolchain, and we must specify a CMAKE_BUILD_TYPE of MinSizeRel, which can be achieved for the internal project and hunter dependencies using the following polly flags respectively:

--config=MinSizeRel
--fwd HUNTER_CONFIGURATION_TYPES=MinSizeRel

Relates: https://github.com/ruslo/hunter/issues/22

Note: Polly toolchain requires HUNTER_TOOLCHAIN_UNDETECTABLE_ID flag:

# There is no macro to detect this flags on toolchain calculation so we must
# mark this toolchain explicitly.
list(APPEND HUNTER_TOOLCHAIN_UNDETECTABLE_ID "lto")

Experiment 1: `-O3 -ffunction-sections -fdata-sections`

$ du -sh _install/xcode-hid-sections/{bin,lib}/*

5.8M	_install/xcode-hid-sections/bin/drishti-acf
8.0M	_install/xcode-hid-sections/bin/drishti-eye
9.0M	_install/xcode-hid-sections/bin/drishti-face
4.7M	_install/xcode-hid-sections/bin/opencv_size
6.6M	_install/xcode-hid-sections/lib/libdrishti.0.8.0.dylib
4.0K	_install/xcode-hid-sections/lib/libdrishti.0.dylib
4.0K	_install/xcode-hid-sections/lib/libdrishti.dylib
4.1M	_install/xcode-hid-sections/lib/libdrishti_c.dylib

Experiment 2: `-flto -O3 -ffunction-sections -fdata-sections`

du -sh _install/xcode-hid-sections-lto/{bin,lib}/*

4.3M	_install/xcode-hid-sections-lto/bin/drishti-acf
5.3M	_install/xcode-hid-sections-lto/bin/drishti-eye
7.2M	_install/xcode-hid-sections-lto/bin/drishti-face
2.8M	_install/xcode-hid-sections-lto/bin/opencv_size
6.9M	_install/xcode-hid-sections-lto/lib/libdrishti.0.8.0.dylib
4.0K	_install/xcode-hid-sections-lto/lib/libdrishti.0.dylib
4.0K	_install/xcode-hid-sections-lto/lib/libdrishti.dylib
3.8M	_install/xcode-hid-sections-lto/lib/libdrishti_c.dylib

Experiment 3: `-flto -Os -ffunction-sections -fdata-sections`

Note: needs Eigen patch for MinSizeRel

3.7M	_install/xcode-hid-sections-lto/bin/drishti-acf
4.9M	_install/xcode-hid-sections-lto/bin/drishti-eye
6.6M	_install/xcode-hid-sections-lto/bin/drishti-face
2.5M	_install/xcode-hid-sections-lto/bin/opencv_size
5.5M	_install/xcode-hid-sections-lto/lib/libdrishti-MinSizeRel.0.8.0.dylib
4.0K	_install/xcode-hid-sections-lto/lib/libdrishti-MinSizeRel.0.dylib
4.0K	_install/xcode-hid-sections-lto/lib/libdrishti-MinSizeRel.dylib
2.9M	_install/xcode-hid-sections-lto/lib/libdrishti_c-MinSizeRel.dylib

 gzip _install/xcode-hid-sections-lto/lib/libdrishti_c-MinSizeRel.dylib
du -sh _install/xcode-hid-sections-lto/lib/libdrishti_c-MinSizeRel.dylib.gz 
1.1M	_install/xcode-hid-sections-lto/lib/libdrishti_c-MinSizeRel.dylib.gz

Experiment 4: `-flto -Os -ffunction-sections -fdata-sections` w/ cereal

DRISHTI_SERIALIZE_WITH_BOOST=OFF
DRISHTI_SERIALIZE_WITH_CEREAL=ON
DRISHTI_SERIALIZE_WITH_CVMATIO=OFF

2.7M	_install/xcode-hid-sections-lto/bin/drishti-acf
3.2M	_install/xcode-hid-sections-lto/bin/drishti-eye
4.1M	_install/xcode-hid-sections-lto/bin/drishti-face
2.5M	_install/xcode-hid-sections-lto/bin/opencv_size
5.1M	_install/xcode-hid-sections-lto/lib/libdrishti-MinSizeRel.0.8.0.dylib
4.0K	_install/xcode-hid-sections-lto/lib/libdrishti-MinSizeRel.0.dylib
4.0K	_install/xcode-hid-sections-lto/lib/libdrishti-MinSizeRel.dylib
2.5M	_install/xcode-hid-sections-lto/lib/libdrishti_c-MinSizeRel.dylib

gzip _install/xcode-hid-sections-lto/lib/libdrishti_c-MinSizeRel.dylib
du -sh _install/xcode-hid-sections-lto/lib/libdrishti_c-MinSizeRel.dylib.gz 
1004K	_install/xcode-hid-sections-lto/lib/libdrishti_c-MinSizeRel.dylib.gz

The text was updated successfully, but these errors were encountered:

headupinclouds · 2017-02-18T17:05:51Z

Eigen won't support MinSizeRel:

Build/Eigen/args.cmake -DCMAKE_MINSIZEREL_POSTFIX=-MinSizeRel -DCMAKE_BUILD_TYPE=MinSizeRel -DCMAKE_CONFIGURATION_TYPES=MinSizeRel -DCMAKE_INSTALL_PREFIX=/Users/dhirvonen/.hunter/_Base/1b9e720/61a0300/3e2bb40/Build/Eigen/Install -DCMAKE_TOOLCHAIN_FILE=/Users/dhirvonen/.hunter/_Base/1b9e720/61a0300/3e2bb40/Build/Eigen/toolchain.cmake -GXcode /Users/dhirvonen/.hunter/_Base/1b9e720/61a0300/3e2bb40/Build/Eigen/Source
loading initial cache file /Users/dhirvonen/.hunter/_Base/1b9e720/61a0300/3e2bb40/cache.cmake
loading initial cache file /Users/dhirvonen/.hunter/_Base/1b9e720/61a0300/3e2bb40/Build/Eigen/args.cmake
-- [polly] Used toolchain: Xcode / clang / LLVM Standard C++ Library (libc++) / c++11 support / hidden / data-sections / function-sections / LTO
-- The C compiler identification is AppleClang 8.0.0.8000042
-- The CXX compiler identification is AppleClang 8.0.0.8000042
-- Check for working C compiler: /Applications/develop/ide/xcode/8.1/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/bin/clang
-- Check for working C compiler: /Applications/develop/ide/xcode/8.1/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/bin/clang -- works
-- Detecting C compile features
-- Detecting C compile features - done
-- Check for working CXX compiler: /Applications/develop/ide/xcode/8.1/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/bin/clang++
-- Check for working CXX compiler: /Applications/develop/ide/xcode/8.1/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/bin/clang++ -- works
-- Detecting CXX compile features
-- Detecting CXX compile features - done
CMake Error at CMakeLists.txt:26 (message):
  Unknown build type "MinSizeRel".  Allowed values are Debug, Release,
  RelWithDebInfo (case-insensitive).

headupinclouds · 2017-02-20T22:36:37Z

All NDK toolchains shown below report "plugin needed to handle lto object". Sorting through various posts to understand this.

TOOLCHAINS=
(
    libcxx-hid-sections${POLLY_LTO_EXT}
    ios-10-1-dep-8-0-libcxx-hid-sections${POLLY_LTO_EXT}
    android-ndk-r10e-api-19-armeabi-v7a-neon-hid-sections${POLLY_LTO_EXT}  # plugin needed to handle lto object
    android-ndk-r10e-api-21-arm64-v8a-gcc-49-hid-sections${POLLY_LTO_EXT} #  plugin needed to handle lto object
    android-ndk-r10e-api-16-x86-hid-sections${POLLY_LTO_EXT} #  plugin needed to handle lto object
    android-ndk-r10e-api-21-x86-64-hid-sections${POLLY_LTO_EXT} #  plugin needed to handle lto object
)

See: android/ndk#108

headupinclouds · 2017-02-21T19:12:56Z

Reading android/ndk#108, and android/ndk#137

As a starting point, it would be useful to create any working build of drishti w/ LTO using a polly android toolchain. Ideally, this would be for all toolchains in bin/toolchains.sh.

Both gcc and clang seem to turn up a few issues, which are described in the GitHub issue linked above.

`android-ndk-r10e-api-19-armeabi-v7a-neon-hid-sections-lto`

Create new toolchain from *-hid-sections.cmake:

cp ${POLLY_ROOT}/android-ndk-r10e-api-19-armeabi-v7a-neon-hid-sections.cmake ${POLLY_ROOT}/android-ndk-r10e-api-19-armeabi-v7a-neon-hid-sections-lto.cmake`

Update include guard, polly_init, add lto.cmake flag, etc:

Add ANDROID toolchain:

include("${CMAKE_CURRENT_LIST_DIR}/utilities/polly_common.cmake")
include("${CMAKE_CURRENT_LIST_DIR}/flags/cxx11.cmake") # before toolchain!
include("${CMAKE_CURRENT_LIST_DIR}/flags/function-sections.cmake")
include("${CMAKE_CURRENT_LIST_DIR}/flags/data-sections.cmake")
include("${CMAKE_CURRENT_LIST_DIR}/flags/hidden.cmake")
include("${CMAKE_CURRENT_LIST_DIR}/flags/lto.cmake") #### ADD ME
include("${CMAKE_CURRENT_LIST_DIR}/os/android.cmake")

${POLLY_ROOT}/flags/lto.cmake

# Copyright (c) 2014-2017, Ruslan Baratov
# Copyright (c) 2017, David Hirvonen
# All rights reserved.

if(DEFINED POLLY_FLAGS_LTO_CMAKE_)
  return()
else()
  set(POLLY_FLAGS_LTO_CMAKE_ 1)
endif()

include(polly_add_cache_flag)

string(COMPARE EQUAL "${ANDROID_NDK_VERSION}" "" _not_android)

# TODO: test other platfroms, CMAKE_CXX_FLAGS_INIT should work for all
if(_not_android)
  polly_add_cache_flag(CMAKE_CXX_FLAGS "-flto")
  polly_add_cache_flag(CMAKE_C_FLAGS "-flto")
else()
  polly_add_cache_flag(CMAKE_CXX_FLAGS_INIT "-flto")
  polly_add_cache_flag(CMAKE_C_FLAGS_INIT "-flto")

  # SECTION A
  #polly_add_cache_flag(CMAKE_EXE_LINKER_FLAGS_INIT "-fuse-ld=gold")
  #polly_add_cache_flag(CMAKE_SHARED_LINKER_FLAGS_INIT "-fuse-ld=gold")
  #polly_add_cache_flag(CMAKE_EXE_LINKER_FLAGS_INIT "-flto")
  #polly_add_cache_flag(CMAKE_SHARED_LINKER_FLAGS_INIT "-flto")

  # SECTION B
  #polly_add_cache_flag(CMAKE_EXE_LINKER_FLAGS_INIT "-Wl,-flto")
  #polly_add_cache_flag(CMAKE_SHARED_LINKER_FLAGS_INIT "-Wl,-flto")
  #polly_add_cache_flag(CMAKE_EXE_LINKER_FLAGS_INIT "-Wl,-fuse-ld=gold")
  #polly_add_cache_flag(CMAKE_SHARED_LINKER_FLAGS_INIT "-Wl,-fuse-ld=gold")
  
endif()

# There is no macro to detect this flags on toolchain calculation so we must
# mark this toolchain explicitly.
list(APPEND HUNTER_TOOLCHAIN_UNDETECTABLE_ID "lto")

When we run this toolchain as shown above (SECTIONA and SECTIONB commented out), but with -flto passed to CMAKE_{C,CXX}_FLAGS, we end up with a build error (plugin needed to handle lto objec) that first shows up here:

/usr/local/Cellar/cmake/3.7.2/bin/cmake -E cmake_link_script CMakeFiles/zlib.dir/link.txt --verbose=1
/Users/dhirvonen/pkg/android/android-ndk-r10e/toolchains/arm-linux-androideabi-4.9/prebuilt/darwin-x86_64/bin/arm-linux-androideabi-ar qc libz.a  CMakeFiles/zlib.dir/adler32.c.o CMakeFiles/zlib.dir/compress.c.o CMakeFiles/zlib.dir/crc32.c.o CMakeFiles/zlib.dir/deflate.c.o CMakeFiles/zlib.dir/gzclose.c.o CMakeFiles/zlib.dir/gzlib.c.o CMakeFiles/zlib.dir/gzread.c.o CMakeFiles/zlib.dir/gzwrite.c.o CMakeFiles/zlib.dir/inflate.c.o CMakeFiles/zlib.dir/infback.c.o CMakeFiles/zlib.dir/inftrees.c.o CMakeFiles/zlib.dir/inffast.c.o CMakeFiles/zlib.dir/trees.c.o CMakeFiles/zlib.dir/uncompr.c.o CMakeFiles/zlib.dir/zutil.c.o
BFD: CMakeFiles/zlib.dir/adler32.c.o: plugin needed to handle lto object

This seems to be due to a mismatched ar. We can tell CMake to use *gcc-ar by adding this to the toolchain file at the bottom:

set(CMAKE_AR "${ANDROID_NDK}/toolchains/arm-linux-androideabi-4.9/prebuilt/darwin-x86_64/bin/arm-linux-androideabi-gcc-ar")

But we still end up with

/Users/dhirvonen/pkg/android/android-ndk-r10e/toolchains/arm-linux-androideabi-4.9/prebuilt/darwin-x86_64/bin/arm-linux-androideabi-gcc-ar qc libz.a  CMakeFiles/zlib.dir/adler32.c.o CMakeFiles/zlib.dir/compress.c.o CMakeFiles/zlib.dir/crc32.c.o CMakeFiles/zlib.dir/deflate.c.o CMakeFiles/zlib.dir/gzclose.c.o CMakeFiles/zlib.dir/gzlib.c.o CMakeFiles/zlib.dir/gzread.c.o CMakeFiles/zlib.dir/gzwrite.c.o CMakeFiles/zlib.dir/inflate.c.o CMakeFiles/zlib.dir/infback.c.o CMakeFiles/zlib.dir/inftrees.c.o CMakeFiles/zlib.dir/inffast.c.o CMakeFiles/zlib.dir/trees.c.o CMakeFiles/zlib.dir/uncompr.c.o CMakeFiles/zlib.dir/zutil.c.o
/Users/dhirvonen/pkg/android/android-ndk-r10e/toolchains/arm-linux-androideabi-4.9/prebuilt/darwin-x86_64/bin/arm-linux-androideabi-ranlib libz.a
BFD: adler32.c.o: plugin needed to handle lto object

We can enable SECTIONB in the lto.cmake flag shown above and build again... but this causes a cmake build check to fail:

  /Users/dhirvonen/pkg/android/android-ndk-r10e/toolchains/arm-linux-androideabi-4.9/prebuilt/darwin-x86_64/bin/arm-linux-androideabi-gcc
  --sysroot=/Users/dhirvonen/pkg/android/android-ndk-r10e/platforms/android-19/arch-arm
  -flto -fvisibility=hidden -fdata-sections -ffunction-sections
-- Configuring incomplete, errors occurred!
  -march=armv7-a -marm -mfpu=neon -mfloat-abi=softfp -funwind-tables
  -no-canonical-prefixes -fexceptions -g -Wl,-fuse-ld=gold -Wl,-flto
  -Wl,--fix-cortex-a8 -fPIE -pie -Wl,--gc-sections -Wl,-z,nocopyreloc
See also "/Users/dhirvonen/devel/elucideye/drishti/_builds/android-ndk-r10e-api-19-armeabi-v7a-neon-hid-sections-lto-Release/CMakeFiles/CMakeOutput.log".
  CMakeFiles/cmTC_f83e0.dir/testCCompiler.c.o -o cmTC_f83e0


See also "/Users/dhirvonen/devel/elucideye/drishti/_builds/android-ndk-r10e-api-19-armeabi-v7a-neon-hid-sections-lto-Release/CMakeFiles/CMakeError.log".
  /Users/dhirvonen/pkg/android/android-ndk-r10e/toolchains/arm-linux-androideabi-4.9/prebuilt/darwin-x86_64/bin/../lib/gcc/arm-linux-androideabi/4.9/../../../../arm-linux-androideabi/bin/ld:
  fatal error: -f/--auxiliary may not be used without -shared

So it seems we are getting the correct linker, but hitting and error where gold is confused by the option that is used to call itself:-f/--auxiliary may not be used without -shared.

Note: With clang, according to this comment android/ndk#137 (comment), we pass -flto directly to the linker flags (not -Wl,-flto. Passing this directly to gcc` seems to have no effect.

headupinclouds · 2017-02-21T22:58:58Z

Note: Errors now from ranlib step. May be useful: http://stackoverflow.com/a/39256013

headupinclouds · 2017-02-21T23:20:54Z

Possible solution here: https://bugs.archlinux.org/task/43367

The "fix" is to use 'gcc-ar' in place of 'ar' (which invokes 'ar' with --plugin=/usr/lib/gcc/x86_64-unknown-linux-gnu/4.9.2/liblto_plugin.so or similar).

headupinclouds · 2017-02-21T23:54:29Z

Testing this now. Build is about 1/2 way through. Seems to resolve some of the basic setup issues for gcc + ndk. Could probably be done cleaner.

include("${CMAKE_CURRENT_LIST_DIR}/os/android.cmake")

set(CMAKE_AR "${ANDROID_NDK}/toolchains/arm-linux-androideabi-4.9/prebuilt/darwin-x86_64/bin/arm-linux-androideabi-gcc-ar")
set(CMAKE_RANLIB "/usr/bin/true") # noop for gcc

message("CMAKE_AR: ${CMAKE_AR}")
  
# set(CMAKE_C_ARCHIVE_CREATE "<CMAKE_AR> qcs <TARGET> <LINK_FLAGS> <OBJECTS>")
# set(CMAKE_C_ARCHIVE_FINISH "true") # Or any other no-op command

# set(CMAKE_CXX_ARCHIVE_CREATE "<CMAKE_AR> qcs <TARGET> <LINK_FLAGS> <OBJECTS>")
# set(CMAKE_CXX_ARCHIVE_FINISH "true") # Or any other no-op command

# Need to use ${CMAKE_AR} here else opencv build breaks
set(CMAKE_C_ARCHIVE_CREATE "${CMAKE_AR} qcs <TARGET> <LINK_FLAGS> <OBJECTS>")
set(CMAKE_C_ARCHIVE_FINISH "true") # Or any other no-op command

set(CMAKE_CXX_ARCHIVE_CREATE "${CMAKE_AR} qcs <TARGET> <LINK_FLAGS> <OBJECTS>")
set(CMAKE_CXX_ARCHIVE_FINISH "true") # Or any other no-op command

The build is much closer but does hit a few errors.

ruslo · 2017-02-22T00:26:57Z

Eigen won't support MinSizeRel

#314

ruslo · 2017-02-22T00:53:59Z

WIP toolchain:

https://gist.github.com/ruslo/4bf32a1b71e63172b29da7137ae7bd10

ruslo · 2017-02-23T14:57:54Z

From GCC documentation:

To create static libraries suitable for LTO, use gcc-ar and gcc-ranlib instead of ar and ranlib

So instead of '/usr/bin/true' it should be arm-linux-androideabi-gcc-ranlib:

ruslo/polly@bffb365

ruslo · 2017-02-24T06:44:26Z

android-ndk-r10e-api-19-armeabi-v7a-neon-hid-sections${POLLY_LTO_EXT}
android-ndk-r10e-api-21-arm64-v8a-gcc-49-hid-sections${POLLY_LTO_EXT}
android-ndk-r10e-api-16-x86-hid-sections${POLLY_LTO_EXT}
android-ndk-r10e-api-21-x86-64-hid-sections${POLLY_LTO_EXT}

Non of those is Android + Clang. What toolchain have you tried?

headupinclouds · 2017-02-24T12:58:38Z

I tried a few versions. These could be a starting point:

android-ndk-r10e-api-16-armeabi-v7a-neon-clang-35-hid-sections-lto.cmake
android-ndk-r11c-api-21-arm64-v8a-clang-hid-sections-lto.cmake

headupinclouds · 2017-02-24T13:02:43Z

In some cases, I believe toolchains that maintain a specific clang version (i.e., "clang3.5") have been updated to newer NDKs where that clang version is not available.

headupinclouds · 2017-02-28T03:02:22Z

Note: Updated experiment after cleaning the hunter cache w/ shared OpenCV + lto

Experiment 8: TOOLCHAIN=android-ndk-r10e-api-19-armeabi-v7a-neon-hid-sections-lto (i.e., gcc) w/ MinSizerel and opencv CMAKE_ARGS BUILD_SHARED_LIBS=ON + OpenCV LTO and `-flto -Os -ffunction-sections -fdata-sections`

du -sh _install/android-ndk-r10e-api-19-armeabi-v7a-neon-hid-sections-lto/lib/*
932K	_install/android-ndk-r10e-api-19-armeabi-v7a-neon-hid-sections-lto/lib/libdrishti.so
836K	_install/android-ndk-r10e-api-19-armeabi-v7a-neon-hid-sections-lto/lib/libdrishti_c.so

gzip -9 *.so
du -sh *
396K	libdrishti.so.gz
352K	libdrishti_c.so.gz

🎆 👏 👍

DRISHTI_SERIALIZE_WITH_BOOST=OFF
DRISHTI_SERIALIZE_WITH_CEREAL=ON
DRISHTI_SERIALIZE_WITH_CVMATIO=OFF

"cmake" "-H." "-B/Users/dhirvonen/devel/elucideye/drishti/_builds/android-ndk-r10e-api-19-armeabi-v7a-neon-hid-sections-lto-MinSizeRel" "-DCMAKE_BUILD_TYPE=MinSizeRel" "-GUnix Makefiles" "-DCMAKE_TOOLCHAIN_FILE=/Users/dhirvonen/devel/elucideye/drishti/src/3rdparty/polly/android-ndk-r10e-api-19-armeabi-v7a-neon-hid-sections-lto.cmake" "-DCMAKE_VERBOSE_MAKEFILE=ON" "-DPOLLY_STATUS_DEBUG=ON" "-DHUNTER_STATUS_DEBUG=ON" "-DCMAKE_INSTALL_PREFIX=/Users/dhirvonen/devel/elucideye/drishti/_install/android-ndk-r10e-api-19-armeabi-v7a-neon-hid-sections-lto" "-DANDROID=TRUE" "-DHUNTER_CONFIGURATION_TYPES=Release" "-DDRISHTI_BUILD_EXAMPLES=ON" "-DDRISHTI_BUILD_TESTS=OFF" "-DDRISHTI_BUILD_FACE=OFF" "-DDRISHTI_BUILD_ACF=OFF" "-DDRISHTI_BUILD_HCI=OFF" "-DDRISHTI_BUILD_OGLES_GPGPU=OFF" "-DDRISHTI_BUILD_REGRESSION_FIXED_POINT=ON" "-DDRISHTI_BUILD_REGRESSION_SIMD=ON" "-DDRISHTI_SERIALIZE_WITH_BOOST=OFF" "-DDRISHTI_SERIALIZE_WITH_CEREAL=ON" "-DDRISHTI_SERIALIZE_WITH_CVMATIO=OFF" "-DDRISHTI_DISABLE_DSYM=OFF" "-DDRISHTI_BUILD_C_INTERFACE=ON" "-DCMAKE_VISIBILITY_INLINES_HIDDEN=ON" "-DCMAKE_CXX_VISIBILITY_PRESET=hidden" "-DCMAKE_XCODE_ATTRIBUTE_GCC_INLINES_ARE_PRIVATE_EXTERN=YES" "-DCMAKE_XCODE_ATTRIBUTE_GCC_SYMBOLS_PRIVATE_EXTERN=YES" "-DDRISHTI_BUILD_MIN_SIZE=ON"

headupinclouds · 2017-02-28T03:09:25Z

Experiment 9: TOOLCHAIN=android-ndk-r10e-api-19-armeabi-v7a-neon-hid-sections-lto (i.e., gcc) w/ MinSizerel and opencv CMAKE_ARGS BUILD_SHARED_LIBS=OFF + OpenCV LTO and -flto -Os -ffunction-sections -fdata-sections

Same as above but with static opencv:

2.4M	_install/android-ndk-r10e-api-19-armeabi-v7a-neon-hid-sections-lto/lib/libdrishti.so
1.4M	_install/android-ndk-r10e-api-19-armeabi-v7a-neon-hid-sections-lto/lib/libdrishti_c.so

$ gzip -9 *.so
1.1M	libdrishti.so.gz
624K	libdrishti_c.so.gz

headupinclouds · 2017-02-28T15:30:57Z

For NDK lto builds I notice these warnings. I believe they are innocuous, but worth a review:

/Users/dhirvonen/pkg/ndk/android-ndk-r10e/sources/cxx-stl/gnu-libstdc++/4.9/include/iostream:62:0: warning: type of 'cerr' does not match original declaration
   extern ostream cerr;  /// Linked to standard error (unbuffered)
 ^
/Users/dhirvonen/pkg/ndk/android-ndk-r10e/sources/cxx-stl/gnu-libstdc++/4.9/include/iostream:62:18: warning: type of 'cerr' does not match original declaration
   extern ostream cerr;  /// Linked to standard error (unbuffered)
                  ^
/Users/dhirvonen/pkg/ndk/android-ndk-r10e/sources/cxx-stl/gnu-libstdc++/4.9/include/iostream:62:18: note: previously declared here
   extern ostream cerr;  /// Linked to standard error (unbuffered)
                  ^
/Users/dhirvonen/pkg/ndk/android-ndk-r10e/sources/cxx-stl/gnu-libstdc++/4.9/include/istream:58:11: warning: type 'struct basic_istream' violates one definition rule
     class basic_istream : virtual public basic_ios<_CharT, _Traits>
           ^
/Users/dhirvonen/pkg/ndk/android-ndk-r10e/sources/cxx-stl/gnu-libstdc++/4.9/include/istream:58:11: note: a type with the same name but different layout is defined in another translation unit
     class basic_istream : virtual public basic_ios<_CharT, _Traits>
           ^
/Users/dhirvonen/pkg/ndk/android-ndk-r10e/sources/cxx-stl/gnu-libstdc++/4.9/include/bits/basic_ios.h:66:11: warning: type 'struct basic_ios' violates one definition rule
     class basic_ios : public ios_base
           ^
/Users/dhirvonen/pkg/ndk/android-ndk-r10e/sources/cxx-stl/gnu-libstdc++/4.9/include/bits/basic_ios.h:66:11: note: a type with the same name but different layout is defined in another translation unit
     class basic_ios : public ios_base
           ^
/Users/dhirvonen/pkg/ndk/android-ndk-r10e/sources/cxx-stl/gnu-libstdc++/4.9/include/bits/locale_facets.h:674:11: warning: type 'struct ctype' violates one definition rule
     class ctype<char> : public locale::facet, public ctype_base
           ^
/Users/dhirvonen/pkg/ndk/android-ndk-r10e/sources/cxx-stl/gnu-libstdc++/4.9/include/bits/locale_facets.h:674:11: note: a type with the same name but different layout is defined in another translation unit
     class ctype<char> : public locale::facet, public ctype_base
           ^
In member function '_ZNKSt5ctypeIcE5widenEc.part.18.constprop':
/Users/dhirvonen/pkg/ndk/android-ndk-r10e/sources/cxx-stl/gnu-libstdc++/4.9/include/bits/locale_facets.h:674:11: warning: type 'struct ctype' violates one definition rule
     class ctype<char> : public locale::facet, public ctype_base
           ^
/Users/dhirvonen/pkg/ndk/android-ndk-r10e/sources/cxx-stl/gnu-libstdc++/4.9/include/bits/locale_facets.h:674:11: note: a type with the same name but different layout is defined in another translation unit
     class ctype<char> : public locale::facet, public ctype_base

headupinclouds · 2017-03-05T18:56:33Z

Experiment 10: TOOLCHAIN=android-ndk-r10e-api-19-armeabi-v7a-neon-hid-sections-lto (i.e., gcc) w/ MinSizerel and opencv CMAKE_ARGS BUILD_SHARED_LIBS=ON + OpenCV LTO and -flto -Os -ffunction-sections -fdata-sections

Note: Full build w/ real-time + face api:

HUNTER_CONFIGURATION_TYPES=MinSizeRel
DRISHTI_BUILD_CONFIG=MinSizeRel

DRISHTI_SERIALIZE_WITH_BOOST=OFF
DRISHTI_SERIALIZE_WITH_CEREAL=ON
DRISHTI_SERIALIZE_WITH_CVMATIO=OFF
DRISHTI_BUILD_C_INTERFACE=ON
DRISHTI_BUILD_ACF=ON
DRISHTI_BUILD_FACE=ON
DRISHTI_BUILD_HCI=ON
DRISHTI_BUILD_OGLES_GPGPU=ON

 du -sh _install/android-ndk-r10e-api-19-armeabi-v7a-neon-hid-sections-lto/lib/*
1.4M	_install/android-ndk-r10e-api-19-armeabi-v7a-neon-hid-sections-lto/lib/libdrishti-MinSizeRel.so
1.3M	_install/android-ndk-r10e-api-19-armeabi-v7a-neon-hid-sections-lto/lib/libdrishti_c-MinSizeRel.so

$ gzip -9 *
$ du -sh *
612K	libdrishti-MinSizeRel.so.gz
568K	libdrishti_c-MinSizeRel.so.gz

headupinclouds · 2017-03-05T19:13:07Z

Experiment 11: TOOLCHAIN=android-ndk-r10e-api-19-armeabi-v7a-neon-hid-sections-lto (i.e., gcc) w/ MinSizerel and opencv CMAKE_ARGS BUILD_SHARED_LIBS=OFF + OpenCV LTO and -flto -Os -ffunction-sections -fdata-sections

Note: Full build w/ real-time + face api (same as 10 above but w/ static OpenCV):

HUNTER_CONFIGURATION_TYPES=MinSizeRel
DRISHTI_BUILD_CONFIG=MinSizeRel

DRISHTI_SERIALIZE_WITH_BOOST=OFF
DRISHTI_SERIALIZE_WITH_CEREAL=ON
DRISHTI_SERIALIZE_WITH_CVMATIO=OFF
DRISHTI_BUILD_C_INTERFACE=ON
DRISHTI_BUILD_ACF=ON
DRISHTI_BUILD_FACE=ON
DRISHTI_BUILD_HCI=ON
DRISHTI_BUILD_OGLES_GPGPU=ON

$ du -sh *
3.6M	libdrishti-MinSizeRel.so
2.1M	libdrishti_c-MinSizeRel.so

$ gzip -9 *.so
$ du -sh *.gz
1.6M	libdrishti-MinSizeRel.so.gz
944K	libdrishti_c-MinSizeRel.so.gz

headupinclouds · 2017-04-12T14:54:35Z

I think this main issue can be closed w/ the addition of the polly lto flag and toolchains. This should be sufficient for current needs, and remaining issues seem to be mostly upstream. (CMake related issues will remain open.)

Caveats:

old ndk clang stuff doesn't seem to work with -lto funcionality
new ndk clang works for -lto w/ -O[123] but -Os functionality is waiting on an LLVM fix + merge to NDK
old ndk gcc toolchains (deprecated) seem to work for -Os and -O123
Apple Clang -lto seems to work fine
MSVC (not tested, and not required)... still curious

ruslo · 2017-04-12T15:10:55Z

MSVC (not tested, and not required)... still curious

They have it as an optimization choice as far as I remember, like no-optimization, agressive-optimization, size-optimization, lto-optimization.

headupinclouds · 2017-04-20T19:51:32Z

Closing this main issue since the core feature is implemented.

headupinclouds added this to the 0.4 milestone Feb 21, 2017

headupinclouds changed the title ~~review-flto~~ support-flto Feb 21, 2017

ruslo self-assigned this Feb 22, 2017

ruslo mentioned this issue Feb 22, 2017

Eigen + MinSizeRel #314

Closed

ruslo mentioned this issue Feb 22, 2017

Android LTO: Clang error with default 'ar' (?) #315

Closed

ruslo mentioned this issue Feb 24, 2017

Android LTO: clang toolchain #316

Closed

headupinclouds added the polly label Mar 5, 2017

headupinclouds closed this as completed Apr 20, 2017

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

support`-flto` #310

support`-flto` #310

headupinclouds commented Feb 18, 2017 •

edited

Loading

headupinclouds commented Feb 18, 2017

headupinclouds commented Feb 20, 2017 •

edited

Loading

headupinclouds commented Feb 21, 2017 •

edited

Loading

headupinclouds commented Feb 21, 2017

headupinclouds commented Feb 21, 2017

headupinclouds commented Feb 21, 2017 •

edited

Loading

ruslo commented Feb 22, 2017

ruslo commented Feb 22, 2017

ruslo commented Feb 23, 2017

ruslo commented Feb 24, 2017

headupinclouds commented Feb 24, 2017

headupinclouds commented Feb 24, 2017

headupinclouds commented Feb 28, 2017 •

edited

Loading

headupinclouds commented Feb 28, 2017 •

edited

Loading

headupinclouds commented Feb 28, 2017 •

edited

Loading

headupinclouds commented Mar 5, 2017 •

edited

Loading

headupinclouds commented Mar 5, 2017

headupinclouds commented Apr 12, 2017

ruslo commented Apr 12, 2017

headupinclouds commented Apr 20, 2017

support-flto #310

support-flto #310

Comments

headupinclouds commented Feb 18, 2017 • edited Loading

Experiment 1: -O3 -ffunction-sections -fdata-sections

Experiment 2: -flto -O3 -ffunction-sections -fdata-sections

Experiment 3: -flto -Os -ffunction-sections -fdata-sections

Experiment 4: -flto -Os -ffunction-sections -fdata-sections w/ cereal

headupinclouds commented Feb 18, 2017

headupinclouds commented Feb 20, 2017 • edited Loading

headupinclouds commented Feb 21, 2017 • edited Loading

android-ndk-r10e-api-19-armeabi-v7a-neon-hid-sections-lto

headupinclouds commented Feb 21, 2017

headupinclouds commented Feb 21, 2017

headupinclouds commented Feb 21, 2017 • edited Loading

ruslo commented Feb 22, 2017

ruslo commented Feb 22, 2017

ruslo commented Feb 23, 2017

ruslo commented Feb 24, 2017

headupinclouds commented Feb 24, 2017

headupinclouds commented Feb 24, 2017

headupinclouds commented Feb 28, 2017 • edited Loading

Experiment 8: TOOLCHAIN=android-ndk-r10e-api-19-armeabi-v7a-neon-hid-sections-lto (i.e., gcc) w/ MinSizerel and opencv CMAKE_ARGS BUILD_SHARED_LIBS=ON + OpenCV LTO and -flto -Os -ffunction-sections -fdata-sections

headupinclouds commented Feb 28, 2017 • edited Loading

Experiment 9: TOOLCHAIN=android-ndk-r10e-api-19-armeabi-v7a-neon-hid-sections-lto (i.e., gcc) w/ MinSizerel and opencv CMAKE_ARGS BUILD_SHARED_LIBS=OFF + OpenCV LTO and -flto -Os -ffunction-sections -fdata-sections

headupinclouds commented Feb 28, 2017 • edited Loading

headupinclouds commented Mar 5, 2017 • edited Loading

Experiment 10: TOOLCHAIN=android-ndk-r10e-api-19-armeabi-v7a-neon-hid-sections-lto (i.e., gcc) w/ MinSizerel and opencv CMAKE_ARGS BUILD_SHARED_LIBS=ON + OpenCV LTO and -flto -Os -ffunction-sections -fdata-sections

headupinclouds commented Mar 5, 2017

Experiment 11: TOOLCHAIN=android-ndk-r10e-api-19-armeabi-v7a-neon-hid-sections-lto (i.e., gcc) w/ MinSizerel and opencv CMAKE_ARGS BUILD_SHARED_LIBS=OFF + OpenCV LTO and -flto -Os -ffunction-sections -fdata-sections

headupinclouds commented Apr 12, 2017

ruslo commented Apr 12, 2017

headupinclouds commented Apr 20, 2017

support`-flto` #310

support`-flto` #310

headupinclouds commented Feb 18, 2017 •

edited

Loading

Experiment 1: `-O3 -ffunction-sections -fdata-sections`

Experiment 2: `-flto -O3 -ffunction-sections -fdata-sections`

Experiment 3: `-flto -Os -ffunction-sections -fdata-sections`

Experiment 4: `-flto -Os -ffunction-sections -fdata-sections` w/ cereal

headupinclouds commented Feb 20, 2017 •

edited

Loading

headupinclouds commented Feb 21, 2017 •

edited

Loading

`android-ndk-r10e-api-19-armeabi-v7a-neon-hid-sections-lto`

headupinclouds commented Feb 21, 2017 •

edited

Loading

headupinclouds commented Feb 28, 2017 •

edited

Loading

Experiment 8: TOOLCHAIN=android-ndk-r10e-api-19-armeabi-v7a-neon-hid-sections-lto (i.e., gcc) w/ MinSizerel and opencv CMAKE_ARGS BUILD_SHARED_LIBS=ON + OpenCV LTO and `-flto -Os -ffunction-sections -fdata-sections`

headupinclouds commented Feb 28, 2017 •

edited

Loading

headupinclouds commented Feb 28, 2017 •

edited

Loading

headupinclouds commented Mar 5, 2017 •

edited

Loading