Skip to content

Latest commit

 

History

History
239 lines (157 loc) · 6.95 KB

building-from-source.md

File metadata and controls

239 lines (157 loc) · 6.95 KB
Raw

Prerequisites

  • Linux or Windows running on x86_64 or arm64 CPUs
  • GCC >= 4.9 or LLVM/Clang >= 6.0, or Visual Studio >= 2015
  • CMake >= 3.14
  • Git >= 2.7.0
  • CUDA Toolkit >= 9.0 (for CUDA)
  • Python >= 3.5 (for CUDA and Python API support)
  • Lua >= 5.2.0 (optional, for Lua API support)

Cloning Source Code

git clone https://github.com/openppl-public/ppl.nn.git

Building X86_64 Engine

Linux

./build.sh -DPPLNN_USE_X86_64=ON

Headers and libraries are installed in pplnn-build/install.

If you want to enable openmp, please specify PPLNN_USE_OPENMP as following:

./build.sh -DPPLNN_USE_X86_64=ON -DPPLNN_USE_OPENMP=ON

Windows

Using vs2015 for example:

build.bat -G "Visual Studio 14 2015 Win64" -DPPLNN_USE_X86_64=ON

Headers and libraries are installed in pplnn-build/install.

Building CUDA Engine

Linux

./build.sh -DPPLNN_USE_CUDA=ON

Note that if you want to build X86 engine along with CUDA engine, you should specify -DPPLNN_USE_X86_64=ON explicitly like this:

./build.sh -DPPLNN_USE_X86_64=ON -DPPLNN_USE_CUDA=ON

Headers and libraries are installed in pplnn-build/install.

If you want to use specified CUDA toolkit version, please specify CUDA_TOOLKIT_ROOT_DIR as following:

./build.sh -DPPLNN_USE_CUDA=ON -DCUDA_TOOLKIT_ROOT_DIR=/path/to/cuda-toolkit-root-dir

Cross Compiling for Arm on X86

Using the following command:

CUDA_TOOLKIT_ROOT=/path/to/cuda/toolkit/root/dir ./build.sh -DPPLNN_USE_CUDA=ON -DPPLNN_TOOLCHAIN_DIR=/path/to/arm/toolchain/dir -DCMAKE_TOOLCHAIN_FILE=cmake/toolchains/aarch64-linux-gnu.cmake

Note that the CUDA_TOOLKIT_ROOT environment variable is required.

You can also specify CUDA_TOOLKIT_ROOT_DIR without setting CUDA_TOOLKIT_ROOT, which will be set to CUDA_TOOLKIT_ROOT_DIR by ppl.nn:

./build.sh -DPPLNN_USE_CUDA=ON -DPPLNN_TOOLCHAIN_DIR=/path/to/arm/toolchain/dir -DCMAKE_TOOLCHAIN_FILE=cmake/toolchains/aarch64-linux-gnu.cmake -DCUDA_TOOLKIT_ROOT_DIR=/path/to/cuda/toolkit/root/dir

Windows

Using vs2015 for example:

build.bat -G "Visual Studio 14 2015 Win64" -DPPLNN_USE_CUDA=ON

Headers and libraries are installed in pplnn-build/install.

Other useful options

We use runtime-compiling version by default. If you want to use static version (build all kernels in advance), please specify PPLNN_ENABLE_CUDA_JIT as following:

./build.sh -DPPLNN_USE_CUDA=ON -DPPLNN_ENABLE_CUDA_JIT=OFF

If you want to run debug model, please specify CMAKE_BUILD_TYPE as following:

./build.sh -DPPLNN_USE_CUDA=ON -DCMAKE_BUILD_TYPE=Debug

If you want to profile running time for each kernel, please specify PPLNN_ENABLE_KERNEL_PROFILING as following and add arg --enable-profiling during executing pplnn.

./build.sh -DPPLNN_USE_CUDA=ON -DPPLNN_ENABLE_KERNEL_PROFILING=ON

Building RISCV Engine

AllWinner D1

You need to download c906 toolchain package from https://occ.t-head.cn/community/download?id=3913221581316624384.

tar -xf riscv64-linux-x86_64-20210512.tar.gz
export RISCV_ROOT_PATH=/path/to/riscv64-linux-x86_64-20210512

Build pplnn:

./build.sh -DPPLNN_TOOLCHAIN_DIR=$RISCV_ROOT_PATH -DCMAKE_TOOLCHAIN_FILE=cmake/toolchains/riscv64-linux-gnu.cmake -DPPLNN_USE_RISCV64=ON -DPPLNN_ENABLE_KERNEL_PROFILING=ON -DPPLNN_ENABLE_PYTHON_API=OFF -DPPLNN_ENABLE_LUA_API=OFF -DCMAKE_INSTALL_PREFIX=pplnn-build/install

Headers and libraries are installed in pplnn-build/install.

Building ARM Engine

Linux

./build.sh -DPPLNN_USE_AARCH64=ON

Headers and libraries are installed in pplnn-build/install.

If you want to enable openmp, please specify PPLNN_USE_OPENMP as following:

./build.sh -DPPLNN_USE_AARCH64=ON -DPPLNN_USE_OPENMP=ON

If you want to enable FP16 inference, please specify PPLNN_USE_ARMV8_2 (your compiler must have armv8.2-a ISA support):

./build.sh -DPPLNN_USE_AARCH64=ON -DPPLNN_USE_ARMV8_2=ON

If your system has multiple NUMA nodes, it is recommended to build with PPLNN_USE_NUMA (please make sure libnuma has been installed in your system):

./build.sh -DPPLNN_USE_AARCH64=ON -DPPLNN_USE_NUMA=ON

If you want to run on mobile platforms, please use the Android NDK package:

./build.sh -DPPLNN_USE_AARCH64=ON -DANDROID_PLATFORM=android-22 -DANDROID_ABI=arm64-v8a -DANDROID_ARM_NEON=ON -DCMAKE_TOOLCHAIN_FILE=<path_to_android_ndk_package>/android-ndk-r22b/build/cmake/android.toolchain.cmake

Buliding Python API support

add -DPPLNN_ENABLE_PYTHON_API=ON to the build command if you want to use PPLNN in python:

./build.sh -DPPLNN_ENABLE_PYTHON_API=ON

If you want to use a specified version of python, you can pass PYTHON3_INCLUDE_DIRS to build.sh:

./build.sh -DPPLNN_ENABLE_PYTHON_API=ON -DPYTHON3_INCLUDE_DIRS=/path/to/your/python/include/dir [other options]

Run the python demo with the following command:

PYTHONPATH=./pplnn-build/install/lib python3 ./tools/pplnn.py [--use-x86 | --use-cuda] --onnx-model tests/testdata/conv.onnx

or use both engines:

cd ppl.nn
PYTHONPATH=./pplnn-build/install/lib python3 ./tools/pplnn.py --use-x86 --use-cuda --onnx-model tests/testdata/conv.onnx

There is a python packaging configuration in python/package. You can build a .whl package:

./build.sh

and then install this package with pip:

cd /tmp/pyppl-package/dist
pip3 install pyppl*.whl

After installation, you can use from pyppl import nn directly without setting the PYTHONPATH env.

Buliding Lua API support

add -DPPLNN_ENABLE_LUA_API=ON to the build command if you want to use PPLNN in lua:

./build.sh -DPPLNN_ENABLE_LUA_API=ON

If you want to use a specified version of lua, you can pass LUA_SRC_DIR to build.sh:

./build.sh -DPPLNN_ENABLE_LUA_API=ON -DLUA_SRC_DIR=/path/to/lua/src [other options]

or you already have a pre-compiled version, you can pass LUA_INCLUDE_DIR and LUA_LIBRARIES to build.sh:

./build.sh -DPPLNN_ENABLE_LUA_API=ON -DLUA_INCLUDE_DIR=/path/to/your/lua/include/dir -DLUA_LIBRARIES=/path/to/your/lua/lib [other options]

Run the lua demo with the following commands:

cd ppl.nn
LUAPATH=./pplnn-build/install/lib /path/to/your/lua-interpreter ./tools/pplnn.lua

Note that your lua interpreter should be compiled with options MYCFLAGS="-DLUA_USE_DLOPEN -fPIC" MYLIBS=-ldl to enable loading .so plugins.

Testing

There is a test tool named pplnn generated from tools/pplnn.cc. You can run pplnn using the following command:

./pplnn-build/tools/pplnn [--use-x86 | --use-cuda] --onnx-model tests/testdata/conv.onnx