服务器

版本:25.09

ANNC 使用手册

1 ANNC 介绍

ANNC(Accelerated Neural Network Compiler)是专注于加速神经网络计算的编译器,聚焦于通过计算图优化,高性能融合算子生成和对接技术以及高效代码生成和优化能力,加速推荐和大模型的推理性能,支持主流开源推理框架接入。

2 ANNC 的安装构建

2.1 直接安装ANNC(通过eur获取)

bash
wget https://eur.openeuler.openatom.cn/results/lesleyzheng1103/ANNC/openeuler-22.03_LTS_SP3-aarch64/00109829-ANNC/ANNC-0.0.2-1.aarch64.rpm

#  安装到 / 目录
rpm -ivh ANNC-0.0.2-1.aarch64.rpm

2.2 RPM包构建安装流程(推荐)

  1. 使用 root 权限,安装 rpmbuild、rpmdevtools,具体命令如下:

    bash
    # 安装 rpmbuild
    yum install dnf-plugins-core rpm-build
    # 安装 rpmdevtools
    yum install rpmdevtools
  2. 在主目录/root下生成 rpmbuild 文件夹:

    bash
    rpmdev-setuptree
    # 检查自动生成的目录结构
    ls ~/rpmbuild/
    BUILD  BUILDROOT  RPMS  SOURCES  SPECS  SRPMS
  3. 使用git clone -b master https://gitee.com/src-openeuler/ANNC.git,从目标仓库的 master 分支拉取代码,并把目标文件放入 rpmbuild 的相应文件夹下:

    shell
    cp ANNC/*.tar.gz* ~/rpmbuild/SOURCES
    cp ANNC/*.patch ~/rpmbuild/SOURCES/
    cp ANNC/ANNC.spec ~/rpmbuild/SPECS/
  4. 用户可通过以下步骤生成 ANNC 的 RPM 包:

    bash
    # 安装 ANNC 所需依赖
    yum-builddep ~/rpmbuild/SPECS/ANNC.spec
    # 构建 ANNC 依赖包
    # 若出现 check-rpaths 相关报错,则需要在 rpmbuild 前添加 QA_RPATHS=0x0002,例如
    # QA_RPATHS=0x0002 rpmbuild -ba ~/rpmbuild/SPECS/ANNC.spec
    rpmbuild -ba ~/rpmbuild/SPECS/ANNC.spec
    # 安装 RPM 包
    cd ~/rpmbuild/RPMS/<arch>
    rpm -ivh ANNC-<version>-<release>.<arch>.rpm

    注意事项:若系统因存有旧版本的 RPM 安装包而导致文件冲突,可以通过以下方式解决:

    bash
    # 解决方案一:强制安装新版本
    rpm -ivh ANNC-<version>-<release>.<arch>.rpm --force
    # 解决方案二:更新安装包
    rpm -Uvh ANNC-<version>-<release>.<arch>.rpm

2.3 源码构建安装流程

ANNC 的源码地址:https://gitee.com/openeuler/ANNC

保证以下依赖包已安装:

shell
yum install -y gcc gcc-c++ bzip2 python3-devel python3-numpy python3-setuptools python3-wheel libstdc++-static java-11-openjdk java-11-openjdk-devel make

安装bazel, 从该地址 https://releases.bazel.build/6.5.0/release/bazel-6.5.0-dist.zip 获取bazel-6.5.0包

bash
unzip bazel-6.5.0-dist.zip -d bazel-6.5.0
cd bazel-6.5.0
env EXTRA_BAZEL_ARGS="--tool_java_runtime_version=local_jdk" bash ./compile.sh

export PATH=/path/to/bazel-6.5.0/output:$PATH
bazel --version

准备XNNPACK

bash
git clone https://gitee.com/openeuler/ANNC.git
export ANNC="your_path_to_ANNC"

cd $ANNC/annc/service/cpu/xla/libs
bash xnnpack.sh

cd $ANNC/annc/service/cpu/xla/libs/XNNPACK/build
cp libXNNPACK.so /usr/lib64
export XNNPACK_BASE="$ANNC/annc/service/cpu/xla/libs"
export XNNPACK_DIR="$XNNPACK_BASE/XNNPACK"

CPLUS_INCLUDE_PATH+="$ANNC/annc/service/cpu/xla/:"
CPLUS_INCLUDE_PATH+="$ANNC/annc/service/:"
CPLUS_INCLUDE_PATH+="$XNNPACK_DIR/:"
CPLUS_INCLUDE_PATH+="$XNNPACK_DIR/include/:"
CPLUS_INCLUDE_PATH+="$XNNPACK_DIR/src/:"
CPLUS_INCLUDE_PATH+="$XNNPACK_DIR/build/pthreadpool-source/include/:"
export CPLUS_INCLUDE_PATH

安装ANNC, 从源码地址下载ANNC源码包。

bash
cd $ANNC

bash build.sh

cp bazel-bin/annc/service/cpu/libannc.so /usr/lib64
mkdir -p /usr/include/annc
cp annc/service/cpu/kdnn_rewriter.h /usr/include/annc
cd python
python3 setup.py bdist_wheel
python3 -m pip install dist/*.whl

3 使用流程

注意事项:

  • ANNC使用者需提前部署好tf-serving,通过编译选项和代码补丁的方式接入ANNC编译优化扩展套件

3.1 图融合结合手动大算子

下载基线模型

bash
git clone https://gitee.com/openeuler/sra_benchmark.git

从基线模型库中获取以下目标推荐模型 DeepFM、DFFM、DLRM、W&D

命令行实现图融合。

bash
# 安装依赖库

python3 -m pip install tensorflow==2.15.1

# 运行模型转换,以DeepFM模型为例

annc-opt -I /path/to/model_DeepFM/1730800001/1 -O deepfm_new/1 dnn_sparse linear_sparse
cp -r /path/to/model_DeepFM/1730800001/1/variables deepfm_new/1

输出目录deepfm_new/1下应生成新的模型文件saved_model.pbtxt,搜索KPFusedSparseEmbedding, 确认图融合算子正确生成

然后将ANNC提供的开源算子库注册到tf-serving:

bash
# 进入tf-serving目录,创建自定义算子文件夹

cd /path/to/serving
mkdir tensorflow_serving/custom_ops

# 将ANNC算子拷贝到该目录下

cp /usr/include/annc/fused*.cc tensorflow_serving/custom_ops/

创建算子编译文件tensorflow_serving/custom_ops/BUILD, 并在该文件中写入以下内容:

ini
package(
   default_visibility = [
        "//visibility:public",
       ],
       licenses = ["notice"],
)

cc_library(
    name = 'recom_embedding_ops',
    srcs = [
      "fused_sparse_embedding.cc",
      "fused_linear_embedding_with_hash_bucket.cc",
      "fused_dnn_embedding_with_hash_bucket.cc"
     ],
     alwayslink = 1,
     deps = [
       "@org_tensorflow//tensorflow/core:framework",
     ]
)
bash
# 打开 tensorflow_serving/model_servers/BUILD, 搜索SUPPORTED_TENSORFLOW_OPS, 添加以下内容注册我们的算子:

"//tensorflow_serving/custom_ops:recom_embedding_ops"

完成算子注册后,使用以下命令重新编译tf-serving,编译成功即表示算子成功注册:

bash
bazel --output_user_root=./output build -c opt --distdir=./proxy \
   --define tflite_with_xnnpack=false \
   tensorflow_serving/model_servers:tensorflow_model_server

3.2 使能算子优化和图优化

在构建好的server的tensorflow的xla路径下,通过补丁脚本使能以下补丁:

bash
export TF_PATH="$HOME/serving/output/XXX/external/org_tensorflow"
export XLA_PATH="$HOME/serving/output/XXX/external/org_tensorflow/third_party/xla"

# 通过方式一安装的ANNC:
cd /usr/include/annc/tfserver/xla

# 修改xla2.sh前两行为:
TF_PATCH_PATH="$ANNC" 
PATH_OF_PATCHES="$ANNC/xla"
export ANNC_PATH=/usr/include/annc
bash xla2.sh

# 通过方式二安装的ANNC:
cd $ANNC/install/tfserver/xla
export ANNC_PATH=$ANNC
bash xla2.sh

# 重新编译
bazel --output_user_root=./output build -c opt --distdir=./proxy \
   --define tflite_with_xnnpack=false \
   tensorflow_serving/model_servers:tensorflow_model_server

3.3 图优化

设置环境变量,开启优化特性。

bash
export 'TF_XLA_FLAGS=--tf_xla_auto_jit=2 --tf_xla_cpu_global_jit --tf_xla_min_cluster_size=16'
export OMP_NUM_THREADS=1
export PORT=7004  # 端口号
ANNC_FLAGS="--graph-opt" ENABLE_BISHENG_GRAPH_OPT="" ./bazel-bin/tensorflow_serving/model_servers/tensorflow_model_server 
--port=$PORT --rest_api_port=7005 
--model_base_path=/path/to/model_Boss/ 
--model_name=deepfm 
--tensorflow_intra_op_parallelism=1 --tensorflow_inter_op_parallelism=-1 
--xla_cpu_compilation_enabled=true

3.4 算子优化

配置环境变量ANNC_FLAGS, 开启matmul下发和对接openblas优化选项,启动TF-Serving,指定目标模型。

bash
export 'TF_XLA_FLAGS=--tf_xla_auto_jit=2 --tf_xla_cpu_global_jit --tf_xla_min_cluster_size=16'
export OMP_NUM_THREADS=1
export PORT=7004  # 端口号
ANNC_FLAGS="--gemm-opt"  XLA_FLAGS="--xla_cpu_enable_xnnpack=true" ./bazel-bin/tensorflow_serving/model_servers/tensorflow_model_server \
    --port=$PORT --rest_api_port=7005 \
    --model_base_path=/path/to/model_DeepFM/1730800001/ \
    --model_name=deepfm \
    --tensorflow_intra_op_parallelism=1 --tensorflow_inter_op_parallelism=-1 \
    --xla_cpu_compilation_enabled=true