Merge branch 'master' into doc

.daq_pm/configs/all_v7

@@ -0,0 +1,5 @@
+# It is a configuration file for [project_manager.vim](https://github.com/daquexian/project_manager.vim)
+name binary-nn
+type cpp
+build_dir build_v7
+cmake_options -DCMAKE_TOOLCHAIN_FILE=~/Android/Sdk/ndk-bundle/build/cmake/android.toolchain.cmake -DANDROID_PLATFORM=android-21 -DANDROID_ABI="armeabi-v7a with NEON" -DCMAKE_EXPORT_COMPILE_COMMANDS=ON -DCMAKE_BUILD_TYPE=Release -GNinja

.daq_pm/configs/bconv_test_x86

@@ -0,0 +1,7 @@
+# It is a configuration file for [project_manager.vim](https://github.com/daquexian/project_manager.vim)
+name binary-nn
+type cpp
+build_dir build_test_x86
+target bconv_test
+cmake_options  -DCMAKE_EXPORT_COMPILE_COMMANDS=ON -DBNN_BUILD_MAIN_LIB=ON -GNinja
+binary tests/bconv_test

.daq_pm/configs/bgemm_test

@@ -1,7 +1,7 @@
 # It is a configuration file for [project_manager.vim](https://github.com/daquexian/project_manager.vim)
 name binary-nn
 type cpp
-build_dir build_bgemm_test
+build_dir build_test
 target bgemm_test
-cmake_options -DCMAKE_TOOLCHAIN_FILE=~/Android/Sdk/ndk-bundle/build/cmake/android.toolchain.cmake -DANDROID_PLATFORM=android-28 -DANDROID_ABI=arm64-v8a -DCMAKE_EXPORT_COMPILE_COMMANDS=ON -DCMAKE_BUILD_TYPE=Release
+cmake_options -DCMAKE_TOOLCHAIN_FILE=~/Android/Sdk/ndk-bundle/build/cmake/android.toolchain.cmake -DANDROID_PLATFORM=android-28 -DANDROID_ABI=arm64-v8a -DCMAKE_EXPORT_COMPILE_COMMANDS=ON -DCMAKE_BUILD_TYPE=Release -GNinja
 binary ~/adb_push_and_run.sh tests/bgemm_test

.daq_pm/configs/bgemm_test_v7

@@ -0,0 +1,7 @@
+# It is a configuration file for [project_manager.vim](https://github.com/daquexian/project_manager.vim)
+name binary-nn
+type cpp
+build_dir build_test_v7
+target bgemm_test
+cmake_options -DCMAKE_TOOLCHAIN_FILE=~/Android/Sdk/ndk-bundle/build/cmake/android.toolchain.cmake -DANDROID_PLATFORM=android-25 -DANDROID_ABI="armeabi-v7a with NEON" -DCMAKE_EXPORT_COMPILE_COMMANDS=ON -DCMAKE_BUILD_TYPE=Release -GNinja
+binary ~/adb_push_and_run.sh tests/bgemm_test

.daq_pm/configs/bgemm_test_x86

@@ -0,0 +1,7 @@
+# It is a configuration file for [project_manager.vim](https://github.com/daquexian/project_manager.vim)
+name binary-nn
+type cpp
+build_dir build_test_x86
+target bgemm_test
+cmake_options  -DCMAKE_EXPORT_COMPILE_COMMANDS=ON -DBNN_BUILD_MAIN_LIB=ON -GNinja
+binary tests/bgemm_test

.daq_pm/configs/bitpack_test

@@ -1,7 +1,7 @@
 # It is a configuration file for [project_manager.vim](https://github.com/daquexian/project_manager.vim)
 name binary-nn
 type cpp
-build_dir build
+build_dir build_test
 target bitpack_test
-cmake_options -DCMAKE_TOOLCHAIN_FILE=~/Android/Sdk/ndk-bundle/build/cmake/android.toolchain.cmake -DANDROID_PLATFORM=android-28 -DANDROID_ABI=arm64-v8a -DCMAKE_EXPORT_COMPILE_COMMANDS=ON -DCMAKE_BUILD_TYPE=Release
+cmake_options -DCMAKE_TOOLCHAIN_FILE=~/Android/Sdk/ndk-bundle/build/cmake/android.toolchain.cmake -DANDROID_PLATFORM=android-28 -DANDROID_ABI=arm64-v8a -DCMAKE_EXPORT_COMPILE_COMMANDS=ON -DCMAKE_BUILD_TYPE=Release -GNinja
 binary ~/adb_push_and_run.sh tests/bitpack_test

.daq_pm/configs/net_test

@@ -3,5 +3,5 @@ name binary-nn
 type cpp
 build_dir build_net_test
 target net_test
-cmake_options -DCMAKE_TOOLCHAIN_FILE=~/Android/Sdk/ndk-bundle/build/cmake/android.toolchain.cmake -DANDROID_PLATFORM=android-28 -DANDROID_ABI=arm64-v8a -DCMAKE_EXPORT_COMPILE_COMMANDS=ON -DCMAKE_BUILD_TYPE=Release
+cmake_options -DCMAKE_TOOLCHAIN_FILE=~/Android/Sdk/ndk-bundle/build/cmake/android.toolchain.cmake -DANDROID_PLATFORM=android-21 -DANDROID_ABI=arm64-v8a -DCMAKE_EXPORT_COMPILE_COMMANDS=ON -DCMAKE_BUILD_TYPE=Release -GNinja
 binary ~/adb_push_and_run.sh tests/net_test

.daq_pm/configs/net_test_v7

@@ -0,0 +1,7 @@
+# It is a configuration file for [project_manager.vim](https://github.com/daquexian/project_manager.vim)
+name binary-nn
+type cpp
+build_dir build_test_v7
+target net_test
+cmake_options -DCMAKE_TOOLCHAIN_FILE=~/Android/Sdk/ndk-bundle/build/cmake/android.toolchain.cmake -DANDROID_PLATFORM=android-21 -DANDROID_ABI="armeabi-v7a with NEON" -DCMAKE_EXPORT_COMPILE_COMMANDS=ON -DCMAKE_BUILD_TYPE=Release -GNinja
+binary ~/adb_push_and_run.sh tests/net_test

.daq_pm/configs/net_test_x86

@@ -0,0 +1,7 @@
+# It is a configuration file for [project_manager.vim](https://github.com/daquexian/project_manager.vim)
+name binary-nn
+type cpp
+build_dir build_net_test_x86
+target net_test
+cmake_options -DCMAKE_EXPORT_COMPILE_COMMANDS=ON -DCMAKE_BUILD_TYPE=Release -DBNN_BUILD_MAIN_LIB=ON -GNinja 
+binary tests/net_test

CMakeLists.txt

@@ -22,11 +22,16 @@ option(BNN_SYSTEM_PROTOBUF "Use system protobuf to build onnx2bnn" ON)
 option(BNN_BUILD_PYTHON "Build onnx2bnn python interface" OFF)
 option(BNN_USE_MSVC_STATIC_RUNTIME "Link onnx2bnn to msvc static runtime" ON)
 
-if (${CMAKE_SYSTEM_NAME} STREQUAL "Android")
-    set(BNN_BUILD_ANDROID ON)
-else()
-    set(BNN_BUILD_ANDROID OFF)
+message(STATUS "Target architecture: ${CMAKE_SYSTEM_PROCESSOR}")
+if (NOT DEFINED BNN_BUILD_MAIN_LIB)
+    if (CMAKE_SYSTEM_PROCESSOR STREQUAL "aarch64" OR 
+            CMAKE_SYSTEM_PROCESSOR STREQUAL "armv7-a")
+        set(BNN_BUILD_MAIN_LIB ON)
+    else()
+        set(BNN_BUILD_MAIN_LIB OFF)
+    endif()
 endif()
+message(STATUS "BNN_BUILD_MAIN_LIB: ${BNN_BUILD_MAIN_LIB}")
 
 include(cmake/utils.cmake)
 bnn_add_msvc_runtime_flag()
@@ -42,7 +47,7 @@ if (${BNN_NET_BENCHMARK})
     add_compile_options("-DBNN_BENCHMARK")
 endif()
 
-if (${BNN_BUILD_ANDROID})
+if (BNN_BUILD_MAIN_LIB)
     set(CMAKE_CXX_STANDARD 17)
 
     set(CMAKE_CXX_FLAGS_RELEASE "-O3 -DNDEBUG")
@@ -64,7 +69,6 @@ if (${BNN_BUILD_ANDROID})
         configure_benchmark()
         add_subdirectory(benchmark)
     endif()
-
 else()
     set(CMAKE_CXX_STANDARD 11)
 

README.md

@@ -16,7 +16,7 @@ Join chat at [Gitter (English)](https://gitter.im/dabnn/dabnn) or QQ Group (Chin
 
 Binary neural networks (BNNs) have great potential on edge devices since they replace float operations by efficient bit-wise operations. However, to leverage the efficiency of bit-wise operations, the reimplmentation of convolution layer and also other layers is needed. 
 
-To our best knowledge, dabnn is the first highly-optimized binary neural networks inference framework for mobile platform. We implemented binary convolutions with armv8 assembly. On Google Pixel 1, our dabnn is as **800%~2400% faster** as [BMXNet](https://github.com/hpi-xnor/BMXNet) (the only one open-sourced BNN inference framework except dabnn to our best knowledge) on a single binary convolution, and as about **700% faster** as it on binarized ResNet-18.
+To our best knowledge, dabnn is the first highly-optimized binary neural networks inference framework for mobile platform. We implemented binary convolutions with ARM assembly. On Google Pixel 1, our dabnn is as **800%~2400% faster** as [BMXNet](https://github.com/hpi-xnor/BMXNet) (the only one open-sourced BNN inference framework except dabnn to our best knowledge) on a single binary convolution, and as about **700% faster** as it on binarized ResNet-18.
 
 ## Benchmark and Comparison
 
@@ -46,11 +46,11 @@ The following is the comparison between our dabnn and [Caffe](http://caffe.berke
 
 ## Convert ONNX Model
 
-We provide a conversion tool, named onnx2bnn, to convert an ONNX model to a dabnn model. To get the conversion tool, just build the project using the native toolchain (instead of arm cross-compiling toolchain). For Linux users, we provide pre-built onnx2bnn AppImage. Linux users can download it from [GitHub Releases](https://github.com/JDAI-CV/dabnn/releases). For the usage and other information about AppImage, please check out https://appimage.org .
+We provide a conversion tool, named onnx2bnn, to convert an ONNX model to a dabnn model. We provide onnx2bnn pre-built binaries for all platforms in [GitHub Releases](https://github.com/JDAI-CV/dabnn/releases). For Linux users, the onnx2bnn pre-built binary is [AppImage](https://appimage.org) format, see https://appimage.org for details.
 
-Note: Binary convolution is a custom operator, so whether the ONNX model is dabnn-comptabile heavily depends on the implementation of the binary convolution in the training code. We will soon provide an dabnn-comptabile PyTorch implementation of binary convolution.
+Note: Binary convolution is a custom operator, so whether the ONNX model is dabnn-comptabile heavily depends on the implementation of the binary convolution in the training code. Please check out [our wiki](https://github.com/JDAI-CV/dabnn/wiki/Train,-export-and-convert-a-dabnn-model) for the further information.
 
-After conversion, the generated dabnn model can be deployed on armv8 devices. For Android developer, we have provided Android AAR package and published it on [jcenter](https://bintray.com/daquexian566/maven/dabnn/_latestVersion), for the usage please check out [example project](https://github.com/JDAI-CV/dabnn-example).
+After conversion, the generated dabnn model can be deployed on ARM devices (e.g., mobile phones and embedded devices). For Android developer, we have provided Android AAR package and published it on [jcenter](https://bintray.com/daquexian566/maven/dabnn/_latestVersion), for the usage please check out [example project](https://github.com/JDAI-CV/dabnn-example).
 
 ## Pretrained Models
 

README_CN.md

@@ -18,7 +18,7 @@ QQ 群：1021964010, 入群答案: nndab
 
 然而，想发挥出位运算的高效率，就需要用位运算对卷积进行重新实现。一直以来，始终没有人针对二值网络推出一个高度优化的 inference 框架，这让二值网络令人遗憾的失去了可以部署在现有通用设备上这个巨大的优势。
 
-为了填补这个巨大的空白，我们推出了 dabnn，一个用 armv8 汇编重写了卷积，高度优化的二值网络 inference 框架。实验显示 dabnn 相对于现有的二值网络和浮点网络 inference 框架均能带来极大的加速。在运行单个二值卷积时，我们的速度是 [BMXNet](https://github.com/hpi-xnor/BMXNet) （除 dabnn 外唯一一个二值网络 inference 框架） 的 **800%~2400%**，在运行二值化的 ResNet-18 时，我们的速度是 BMXNet 的约 **700%**。我们希望 dabnn 的开源可以成为二值网络部署在终端设备的关键一步，也可以有助于使这一领域的爱好者了解二值网络在真实设备上的表现。
+为了填补这个巨大的空白，我们推出了 dabnn，一个用 ARM 汇编重写了卷积，高度优化的二值网络 inference 框架。实验显示 dabnn 相对于现有的二值网络和浮点网络 inference 框架均能带来极大的加速。在运行单个二值卷积时，我们的速度是 [BMXNet](https://github.com/hpi-xnor/BMXNet) （除 dabnn 外唯一一个二值网络 inference 框架） 的 **800%~2400%**，在运行二值化的 ResNet-18 时，我们的速度是 BMXNet 的约 **700%**。我们希望 dabnn 的开源可以成为二值网络部署在终端设备的关键一步，也可以有助于使这一领域的爱好者了解二值网络在真实设备上的表现。
 
 ## 速度
 
@@ -48,11 +48,11 @@ dabnn_bireal18_imagenet_stem   43279353 ns   41533009 ns         14       <--- 
 
 ## 如何转换 ONNX 模型
 
-我们提供模型转换工具 onnx2bnn 将 ONNX 模型转换为 dabnn 格式的模型。用本地编译工具链（而不是 arm 交叉编译工具链）编译这个项目就可以编译出 onnx2dnn。对 Linux 用户我们提供可以在 Linux 下无需编译直接运行的 AppImage，从 [GitHub Releases](https://github.com/JDAI-CV/dabnn/releases) 下载即可。AppImage 的使用方法和其它相关信息请参考 https://appimage.org/。
+我们提供模型转换工具 onnx2bnn 将 ONNX 模型转换为 dabnn 格式的模型。在 [GitHub Releases](https://github.com/JDAI-CV/dabnn/releases) 里有各个平台的 onnx2bnn 预编译二进制文件，可以直接下载运行。Linux 用户我们提供的是 AppImage 格式的二进制文件，AppImage 的使用方法和其它相关信息请参考 https://appimage.org/。
 
-注意：因为二值卷积是一种自定义操作，所以 ONNX 模型是否与 dabnn 兼容极大程度上依赖于训练代码中二值卷积的实现。我们很快会提供一个与 dabnn 兼容的二值卷积 PyTorch 实现。
+注意：因为二值卷积是一种自定义操作，所以 ONNX 模型是否与 dabnn 兼容极大程度上依赖于训练代码中二值卷积的实现，在 [wiki](https://github.com/JDAI-CV/dabnn/wiki/Train,-export-and-convert-a-dabnn-model) 中有详细的进一步描述。
 
-转换完成后得到的 dabnn 模型就可以在 armv8 设备上使用。对 Android 开发者我们已经把 Android AAR 包上传到了 [jcenter](https://bintray.com/daquexian566/maven/dabnn/_latestVersion)，使用方法请看[示例工程](https://github.com/JDAI-CV/dabnn-example)。
+转换完成后得到的 dabnn 模型就可以在 ARM 设备（例如手机和嵌入式设备）上使用。对 Android 开发者我们已经把 Android AAR 包上传到了 [jcenter](https://bintray.com/daquexian566/maven/dabnn/_latestVersion)，使用方法请看[示例工程](https://github.com/JDAI-CV/dabnn-example)。
 
 ## 预训练模型
 

benchmark/benchmark.cpp

@@ -21,13 +21,15 @@ static void BM_pack_mat_64_small(benchmark::State &state) {
     }
 }
 
+#ifdef __aarch64__
 static void BM_pack_mat_128_small(benchmark::State &state) {
     const bnn::Mat a(1, 32, 32, 128, bnn::DataType::Float, 0);
     bnn::Mat b(1, 32, 32, 128, bnn::DataType::Bit, 0);
     for (auto _ : state) {
         pack_mat_128(a, b);
     }
 }
+#endif  // __aarch64__
 
 static void BM_pack_mat_64(benchmark::State &state) {
     const bnn::Mat a(1, 64, 64, 128, bnn::DataType::Float);
@@ -37,13 +39,15 @@ static void BM_pack_mat_64(benchmark::State &state) {
     }
 }
 
+#ifdef __aarch64__
 static void BM_pack_mat_128(benchmark::State &state) {
     const bnn::Mat a(1, 64, 64, 128, bnn::DataType::Float);
     bnn::Mat b(1, 64, 64, 128, bnn::DataType::Bit);
     for (auto _ : state) {
         pack_mat_128(a, b);
     }
 }
+#endif  // __aarch64__
 
 #define SETUP_BCONV_FLOAT(size_a, size_b, num_output)                         \
     const size_t AHEIGHT = size_a;                                            \
@@ -73,6 +77,7 @@ static void BM_pack_mat_128(benchmark::State &state) {
                                                                               \
     bnn::Mat c(CHEIGHT, CWIDTH, NUM_OUTPUT, bnn::DataType::Float);
 
+#ifdef __aarch64__
 static void BM_bconv_float_3x3_128(benchmark::State &state) {
     SETUP_BCONV_FLOAT(30, 3, 128);
     for (auto _ : state) {
@@ -88,6 +93,7 @@ static void BM_bconv_float_1x1_128(benchmark::State &state) {
         bnn::bconv_1x1_128(a_binary, b, c);
     }
 }
+#endif  // __aarch64__
 
 #undef SETUP_BCONV_FLOAT
 
@@ -135,6 +141,7 @@ static void BM_bnn_bconv_1x1_naive_128(benchmark::State &state) {
     }
 }
 
+#ifdef __aarch64__
 static void BM_bnn_bconv_1x1_64(benchmark::State &state) {
     SETUP_BCONV(56, 1, 64, 1);
     for (auto _ : state) {
@@ -162,6 +169,7 @@ static void BM_bnn_bconv_1x1_512(benchmark::State &state) {
         bnn::bconv_1x1_512(a, b, c);
     }
 }
+#endif  // __aarch64__
 
 static void BM_bnn_bconv_3x3_64(benchmark::State &state) {
     SETUP_BCONV(58, 3, 64, 1);

ci/build_dabnn_v7.sh

@@ -0,0 +1,11 @@
+#! /usr/bin/env bash
+set -e
+
+echo "y" | $ANDROID_HOME/tools/bin/sdkmanager --install 'ndk-bundle'
+nproc=$(ci/get_cores.sh)
+
+mkdir build_dabnn && cd build_dabnn
+cmake -DCMAKE_TOOLCHAIN_FILE=$ANDROID_HOME/ndk-bundle/build/cmake/android.toolchain.cmake -DANDROID_PLATFORM=android-21 -DANDROID_ABI="armeabi-v7a with NEON" -DCMAKE_EXPORT_COMPILE_COMMANDS=ON -DCMAKE_BUILD_TYPE=Release ..
+cmake --build . -- -j$nproc
+cd -
+

ci/dabnn_build_and_test.yml

@@ -22,26 +22,34 @@ pr:
     - README.md
     - docs/*
 
-pool:
-    vmImage: 'macOS-10.14'
-steps:
-- checkout: self
-  submodules: true
-- bash: brew install watch gnu-sed
-  displayName: Install watch and gnu-sed
-- bash: ci/build_dabnn.sh
-  displayName: Build
-- bash: ci/start_android_emulator.sh
-  displayName: Start Android Emulator
-- bash: ci/adb_push_and_run.sh build_dabnn/tests/bconv_test
-  displayName: Binary Conv Test
-- bash: ci/adb_push_and_run.sh build_dabnn/tests/bgemm_test
-  displayName: Binary Gemm Test
-- bash: ci/download_models.sh
-  displayName: Download Models
-- bash: ci/adb_push_and_run.sh build_dabnn/tests/net_test
-  displayName: Model Test
-- bash: ci/build_aar.sh
-  env: 
-    BINTRAY_KEY: $(bintrayKey)
-  displayName: Build and Publish AAR package
+jobs:
+- job: aarch64
+  pool:
+      vmImage: 'macOS-10.14'
+  steps:
+  - checkout: self
+    submodules: true
+  - bash: brew install watch gnu-sed
+    displayName: Install watch and gnu-sed
+  - bash: ci/build_dabnn.sh
+    displayName: Build
+  - bash: ci/start_android_emulator.sh
+    displayName: Start Android Emulator
+  - template: template_dabnn_run_test.yml
+  - bash: ci/build_aar.sh
+    env: 
+      BINTRAY_KEY: $(bintrayKey)
+    displayName: Build and Publish AAR package
+- job: armv7a_with_NEON
+  pool:
+      vmImage: 'macOS-10.14'
+  steps:
+  - checkout: self
+    submodules: true
+  - bash: brew install watch gnu-sed
+    displayName: Install watch and gnu-sed
+  - bash: ci/build_dabnn_v7.sh
+    displayName: Build
+  - bash: ci/start_android_emulator_v7.sh
+    displayName: Start Android Emulator
+  - template: template_dabnn_run_test.yml

ci/start_android_emulator_v7.sh

@@ -0,0 +1,25 @@
+#! /usr/bin/env bash
+set -e
+
+export TERM=xterm
+
+echo "y" | $ANDROID_HOME/tools/bin/sdkmanager --install 'system-images;android-25;google_apis;armeabi-v7a'
+
+echo "no" | $ANDROID_HOME/tools/bin/avdmanager create avd -n android_emulator -k 'system-images;android-25;google_apis;armeabi-v7a' --force
+
+echo "Starting emulator"
+
+# Start emulator in background
+nohup $ANDROID_HOME/emulator/emulator -avd android_emulator -no-snapshot -no-audio &
+
+# start server in advance, so that the result of watch will only change when device gets online
+$ANDROID_HOME/platform-tools/adb start-server
+
+watch -g -n 1 '$ANDROID_HOME/platform-tools/adb devices | grep -c device$'
+
+echo "Emulator is online"
+
+$ANDROID_HOME/platform-tools/adb devices
+
+echo "Emulator started"
+

ci/template_dabnn_run_test.yml

@@ -0,0 +1,9 @@
+steps:
+- bash: ci/adb_push_and_run.sh build_dabnn/tests/bconv_test
+  displayName: Binary Conv Test
+- bash: ci/adb_push_and_run.sh build_dabnn/tests/bgemm_test
+  displayName: Binary Gemm Test
+- bash: ci/download_models.sh
+  displayName: Download Models
+- bash: ci/adb_push_and_run.sh build_dabnn/tests/net_test
+  displayName: Model Test

cmake/system.cmake

@@ -1,6 +1,7 @@
 # Copyright 2019 JD.com Inc. JD AI
 
-if (${CMAKE_SYSTEM_NAME} STREQUAL "Android")
+if ((NOT CMAKE_SYSTEM_NAME STREQUAL CMAKE_HOST_SYSTEM_NAME) OR 
+    (NOT CMAKE_SYSTEM_PROCESSOR STREQUAL CMAKE_HOST_SYSTEM_PROCESSOR))
     set(CMAKE_FIND_ROOT_PATH_MODE_PROGRAM NEVER)
     set(CMAKE_FIND_ROOT_PATH_MODE_LIBRARY ONLY)
     set(CMAKE_FIND_ROOT_PATH_MODE_INCLUDE ONLY)

common/baseline.h

@@ -10,8 +10,12 @@
 #include <dabnn/mat.h>
 
 inline int bitcount(uint64_t x) {
+#ifdef __aarch64__
+    return __builtin_popcountl(x);
+#else
     std::bitset<64> bs(x);
     return bs.count();
+#endif
 }
 
 inline void baseline_pack_mat(const bnn::Mat &float_mat, bnn::Mat &binary_mat) {