上一篇中,我们介绍了在 PC 如何使用 C++ 加载我们保存的模型并测试。这一篇,我们介绍在 PC 上交叉编译 aarch64 平台的 tensorflow 源码过程,这个难度比我想象的要难太多了。(耗时10天不止,一把心酸一把泪),首先看一下官方在文档介绍:
这里,我选择了 tensorflow 官方测试过支持 gcc 的最后版本 2.12.0。然后介绍下 PC 的配置:
然后看下 python3 和 bazel 的版本:
▸ python3 --version
Python 3.11.9
▸ bazel-5.3.0-linux-x86_64 --version
bazel 5.3.0
▸ aarch64-linux-gcc -v
Using built-in specs.
COLLECT_GCC=/home/red/Samba/arm-gnu-toolchain-13.3.rel1-x86_64-aarch64-none-linux-gnu/bin/aarch64-linux-gcc
COLLECT_LTO_WRAPPER=/home/red/Samba/arm-gnu-toolchain-13.3.rel1-x86_64-aarch64-none-linux-gnu/bin/../libexec/gcc/aarch64-none-linux-gnu/13.3.1/lto-wrapper
Target: aarch64-none-linux-gnu
Configured with: /data/jenkins/workspace/GNU-toolchain/arm-13/src/gcc/configure --target=aarch64-none-linux-gnu --prefix= --with-sysroot=/aarch64-none-linux-gnu/libc --with-build-sysroot=/data/jenkins/workspace/GNU-toolchain/arm-13/build-aarch64-none-linux-gnu/install//aarch64-none-linux-gnu/libc --with-bugurl=https://bugs.linaro.org/ --enable-gnu-indirect-function --enable-shared --disable-libssp --disable-libmudflap --enable-checking=release --enable-languages=c,c++,fortran --with-gmp=/data/jenkins/workspace/GNU-toolchain/arm-13/build-aarch64-none-linux-gnu/host-tools --with-mpfr=/data/jenkins/workspace/GNU-toolchain/arm-13/build-aarch64-none-linux-gnu/host-tools --with-mpc=/data/jenkins/workspace/GNU-toolchain/arm-13/build-aarch64-none-linux-gnu/host-tools --with-isl=/data/jenkins/workspace/GNU-toolchain/arm-13/build-aarch64-none-linux-gnu/host-tools --enable-fix-cortex-a53-843419 --with-pkgversion='Arm GNU Toolchain 13.3.Rel1 (Build arm-13.24)'
Thread model: posix
Supported LTO compression algorithms: zlib
gcc version 13.3.1 20240614 (Arm GNU Toolchain 13.3.Rel1 (Build arm-13.24))
特别要说明的是,交叉编译工具联的版本也要选择合适的,否则会出现莫名奇妙的问题(现在这个版本都还存在问题:在 tensorflow 的 logging.h 这里,呜呜呜)。
基础性的介绍完了之后,我们开始正文。
交叉编译的目的是编译出对应的库和引用头文件,为什么选择交叉编译是因为 PC 性能强啊(实测在这台 PC 上一切顺利编译也要3h+,如果是本地编译可想而知那要多久啊)。首先要说明一下 tensorflow 使用 Bazel 构建。而如何在 bazel 环境下选择自己的交叉编译工具链呢?不像 makefile 或者 cmake 直接指定就行了,bazel 有点类似 GN,要配置好多东西,但是个人感觉比 GN 还要复杂,想哭。
刚开始,我走了好多弯路,一通晚上乱搜,最后搜的文章水平参差不齐导致徒劳浪费了不少时间,还对自己的能力产生了怀疑。最后还是根据官方文档,一点一点查漏补缺搭建好的交叉编译工具链的配置文件。首先,附上官方的指导文档连接 [Bazel Tutorial: Configure C++ Toolchains](Configuring C++ toolchains - Bazel 5.3.0)。
最后创建的涉及到交叉编译工具链的文档有两个:
这里要简单介绍下 bazel 构建工具的一些基本概念:
@myrepo//my/app/main:app_binary
这个标签,前半部分 @myrepo
表示仓库的名字是 myrepo 仓库,如果本来就是在这个仓库下,那么可以进一步简写为//
,另外标签的第二部分 my/app/main
表示这个包的名字,如果本来就在 @myrepo//my/app/main
包下,可以简写为 :app_binary
,进一步地,如果是文件类型的可以间写为 app_binary
但是规则类型的还是要保留这个符号 :
。至此,简单概括了 bazel 中涉及到的关键概念,下面结合配置交叉编译工具链添加的文件,我们可以知道在根仓库下新建了一个 toolchain 的包。看下这个包的 BUILD 文件:
package(default_visibility = ["//visibility:public"])
# 定义了一个名字是 cross_gcc_suite 的 target,这个 target 是 C++ 工具链的集合
cc_toolchain_suite(
# name 和 toolchains 是必须项
name = "cross_gcc_suite",
# 声明 aarch64 平台对应的工具链名字是当前 package 下的 aarch64_toolchain 目标
# 根据 --cpu 和 --compiler 选项选择工具链
toolchains = {
"aarch64": ":aarch64_toolchain",
},
)
# 定义一个名字为 empty 的空 target
filegroup(name = "empty")
# 定义了一个名字是 aarch64_toolchain 的 c++ 工具链 target
cc_toolchain(
name = "aarch64_toolchain",
toolchain_identifier = "aarch64-toolchain",
# 工具联的配置项
toolchain_config = ":aarch64_toolchain_config",
all_files = ":empty",
compiler_files = ":empty",
dwp_files = ":empty",
linker_files = ":empty",
objcopy_files = ":empty",
strip_files = ":empty",
supports_param_files = 0,
)
# 在当前 package 下的 cc_toolchain_config.bzl 中,导入 cc_toolchain_config 函数
load(":cc_toolchain_config.bzl", "cc_toolchain_config")
# 通过 cc_toolchain_config 函数,定义一个名为 aarch64_toolchain_config 的 target
cc_toolchain_config(name = "aarch64_toolchain_config")
下面我们看下 toolchain 包下的 cc_toolchain_config.bzl 文件内容:
# toolchain/cc_toolchain_config.bzl:
# 加载一些函数
load("@bazel_tools//tools/build_defs/cc:action_names.bzl", "ACTION_NAMES")
load("@bazel_tools//tools/cpp:cc_toolchain_config_lib.bzl", "feature", "flag_group", "flag_set", "tool_path")
# 定义变量
all_link_actions = [ # NEW
ACTION_NAMES.cpp_link_executable,
ACTION_NAMES.cpp_link_dynamic_library,
ACTION_NAMES.cpp_link_nodeps_dynamic_library,
]
all_compile_actions = [
ACTION_NAMES.assemble,
ACTION_NAMES.c_compile,
ACTION_NAMES.clif_match,
ACTION_NAMES.cpp_compile,
ACTION_NAMES.cpp_header_parsing,
ACTION_NAMES.cpp_module_codegen,
ACTION_NAMES.cpp_module_compile,
ACTION_NAMES.linkstamp_compile,
ACTION_NAMES.lto_backend,
ACTION_NAMES.preprocess_assemble,
]
all_cpp_compile_actions = [
ACTION_NAMES.cpp_compile,
ACTION_NAMES.cpp_header_parsing,
ACTION_NAMES.cpp_module_codegen,
ACTION_NAMES.cpp_module_compile,
]
# 定义函数
def _impl(ctx):
tool_paths = [ # NEW
tool_path(
name = "gcc",
# 如下替换为指定的工具链的路径
path = "/home/red/Samba/arm-gnu-toolchain-13.3.rel1-x86_64-aarch64-none-linux-gnu/bin/aarch64-linux-gcc",
),
tool_path(
name = "g++",
path = "/home/red/Samba/arm-gnu-toolchain-13.3.rel1-x86_64-aarch64-none-linux-gnu/bin/aarch64-linux-g++",
),
tool_path(
name = "ld",
path = "/home/red/Samba/arm-gnu-toolchain-13.3.rel1-x86_64-aarch64-none-linux-gnu/bin/aarch64-linux-ld",
),
tool_path(
name = "ar",
path = "/home/red/Samba/arm-gnu-toolchain-13.3.rel1-x86_64-aarch64-none-linux-gnu/bin/aarch64-linux-ar",
),
tool_path(
name = "cpp",
path = "/home/red/Samba/arm-gnu-toolchain-13.3.rel1-x86_64-aarch64-none-linux-gnu/bin/aarch64-linux-cpp",
),
tool_path(
name = "gcov",
path = "/home/red/Samba/arm-gnu-toolchain-13.3.rel1-x86_64-aarch64-none-linux-gnu/bin/aarch64-linux-gcov",
),
tool_path(
name = "nm",
path = "/home/red/Samba/arm-gnu-toolchain-13.3.rel1-x86_64-aarch64-none-linux-gnu/bin/aarch64-linux-nm",
),
tool_path(
name = "objdump",
path = "/home/red/Samba/arm-gnu-toolchain-13.3.rel1-x86_64-aarch64-none-linux-gnu/bin/aarch64-linux-objdump",
),
tool_path(
name = "strip",
path = "/home/red/Samba/arm-gnu-toolchain-13.3.rel1-x86_64-aarch64-none-linux-gnu/bin/aarch64-linux-strip",
),
]
features= [ # NEW
feature(
name = "default_linker_flags",
enabled = True,
flag_sets = [
flag_set(
actions = all_link_actions,
flag_groups = ([
flag_group(
flags = [
"-static", #建议交叉编译还是带上这个参数,要不然无法在 PC 简单地正常执行 aarch64 elf 格式的工具
"-lstdc++",
],
),
]),
),
],
),
feature(
name = "default_compiler_flags",
enabled = True,
flag_sets = [
flag_set(
actions = all_cpp_compile_actions,
flag_groups = ([
flag_group(
flags = [
"-fpermissive",
],
),
]),
),
],
),
]
return cc_common.create_cc_toolchain_config_info(
ctx = ctx,
features = features,
cxx_builtin_include_directories = [ # NEW
# 替换自己相关的头文件,这部分可以先空起来,在编译过程中提示缺少头文件的时候会告诉我们缺少哪些,到时候再追加也可以
"/home/red/Samba/arm-gnu-toolchain-13.3.rel1-x86_64-aarch64-none-linux-gnu/aarch64-none-linux-gnu/include/c++/13.3.1/bits",
"/home/red/Samba/arm-gnu-toolchain-13.3.rel1-x86_64-aarch64-none-linux-gnu/aarch64-none-linux-gnu/include/c++/13.3.1",
"/home/red/Samba/arm-gnu-toolchain-13.3.rel1-x86_64-aarch64-none-linux-gnu/aarch64-none-linux-gnu/libc/usr/include",
"/home/red/Samba/arm-gnu-toolchain-13.3.rel1-x86_64-aarch64-none-linux-gnu/lib/gcc/aarch64-none-linux-gnu/13.3.1/include",
"/home/red/Samba/arm-gnu-toolchain-13.3.rel1-x86_64-aarch64-none-linux-gnu/lib/gcc/aarch64-none-linux-gnu/13.3.1/include-fixed",
],
toolchain_identifier = "aarch64-linux",
host_system_name = "x86_64",
target_system_name = "aarch64",
target_cpu = "aarch64",
target_libc = "unknown",
compiler = "g++",
abi_version = "unknown",
abi_libc_version = "unknown",
tool_paths = tool_paths, # NEW
)
# 定义了一个新的规则
cc_toolchain_config = rule(
# 规则的实现函数
implementation = _impl,
attrs = {},
provides = [CcToolchainConfigInfo],
)
有了这两个文件,交叉编译工具链就配置好了。下面为了方便使用,修改 .bazelrc 文件,追加如下两行,可以看到使用 :
build:elinux_aarch64 --crosstool_top=//toolchain:cross_gcc_suite
build:elinux_aarch64 --host_crosstool_top=@bazel_tools//tools/cpp:toolchain
这里使用的 --crosstool_top 和 --host_crosstool_top 是 Bazel 中用于指定交叉编译工具和主机编译工具链配置的重要参数,这里要注意必须要使用 --host_crosstool_top 选项指定一个默认的 PC(或者 k8)平台的工具链,要不然会报错的。
然后和 PC 端那样,首先是 ./configure
然后就是触发构建了,具体命令如下:
bazel build --config=elinux_aarch64 --copt="-fPIC" --cxxopt="-fPIC" --verbose_failures //tensorflow:libtensorflow.so //tensorflow:install_headers
在编译过程中,我使用的工具链需要修改如下地方:
diff --git a/tensorflow/tsl/platform/default/logging.h b/tensorflow/tsl/platform/default/logging.h
index 3578bedf0f1..24c74607a96 100644
--- a/tensorflow/tsl/platform/default/logging.h
+++ b/tensorflow/tsl/platform/default/logging.h
@@ -310,7 +310,7 @@ inline uint64 GetReferenceableValue(uint64 t) { return t; }
// it uses the definition for operator< < , with a few special cases below.
template < typename T >
inline void MakeCheckOpValueString(std::ostream* os, const T& v) {
- // (*os) < < v;
+ //(*os) < < v;
}
// Overrides for char types provide readable values for unprintable
这部分后续应该可以不用这么修改(暂时还没有解决)。
然后最重要的是在后期,静态连接 libtensorflow_framework.so.2.12.0 的时候会提示错误:
这时候需要强制动态链接才行,怎么做呢? diff 文件是:
diff --git a/tensorflow/BUILD b/tensorflow/BUILD
index 0d27a8294f5..adcdef8f8a9 100644
--- a/tensorflow/BUILD
+++ b/tensorflow/BUILD
@@ -1100,6 +1100,8 @@ tf_cc_shared_library(
],
"//conditions:default": [
"-Wl,--version-script,$(location //tensorflow:tf_framework_version_script.lds)",
+ "-Wl,-Bdynamic",
],
}),
linkstatic = 1,
下面构建的时候又出现新的错误,动态链接交叉编译出来的工具无法正常生成一些链接 libtensorflow 库的文件,这时候我通过 sudo chrpath -r
进行处理一下,记得要提前将需要的库放在一个指定的目录,还是很复杂的。
具体集合到 tensorflow 的编译环境的 diff 文件是:
diff --git a/tensorflow/tensorflow.bzl b/tensorflow/tensorflow.bzl
index 115ff76b414..2fb733de643 100644
--- a/tensorflow/tensorflow.bzl
+++ b/tensorflow/tensorflow.bzl
@@ -1113,12 +1113,13 @@ def tf_gen_op_wrapper_cc(
],
srcs = srcs,
tools = [":" + tool] + tf_binary_additional_srcs(),
- cmd = ("$(location :" + tool + ") $(location :" + out_ops_file + ".h) " +
+ cmd = ("echo "xxx" | sudo -S chrpath -r /home/red/Rcc/lib64 " + "$(location :" + tool + ")" + ";" + "$(location :" + tool + ") $(location :" + out_ops_file + ".h) " +
"$(location :" + out_ops_file + ".cc) " +
str(include_internal_ops) + " " + api_def_args_str),
compatible_with = compatible_with,
至此,关键的修改就是这些了,其他编译过程中需要的修改,根据提示改一下就好了。
看一下最后编译成功的截图:
至此就有了开发 aarch64 平台 tensorflow 开发相关的库和头文件了:
这个库文件真够大的,哈哈。然后我们测试下,例程还是用上一篇的一个加载光照度模型并打印预测值的C++代码:
#include < tensorflow/cc/saved_model/loader.h >
using namespace tensorflow;
using namespace std;
int main() {
SessionOptions options;
RunOptions run_options;
SavedModelBundle bundle;
Status status = LoadSavedModel(options, run_options, "/home/red/Downloads/fivek_dataset/test_mark_illuminance_level/illu_v03", {"serve"}, &bundle);
if (!status.ok()) {
std::cerr < < "Error loading model: " < < status.ToString() < < std::endl;
return 1;
}
// Access the session
Session* session = bundle.session.get();
// Create input tensor
Tensor input_tensor(DT_FLOAT, TensorShape({1, 255, 255, 3}));
// Fill input tensor with data
auto input_tensor_flat = input_tensor.flat< float >();
std::cout < < "size of input tensor is " < < input_tensor_flat.size() < < std::endl;
for (int i = 0; i < input_tensor_flat.size(); ++i) {
input_tensor_flat(i) = 255.0;
}
// Run inference
std::vector< Tensor > outputs;
Status run_status = session- >Run({{"serving_default_rescaling_input", input_tensor}}, {"StatefulPartitionedCall"}, {}, &outputs);
if (!run_status.ok()) {
std::cerr < < "Error running model: " < < run_status.ToString() < < std::endl;
return 1;
}
const Eigen::TensorMap< Eigen::Tensor< float, 1, Eigen::RowMajor >, Eigen::Aligned >& prediction = outputs[0].flat< float >();
const long count = prediction.size();
for (int i = 0; i < count; ++i) {
const float value = prediction(i);
// value是该张量以一维数组表示时在索引i处的值。
std::cout < < "hey hey " < < value < < std::endl;
}
// Process output tensor
Tensor ans = outputs[0];
// auto ans_value = ans.tensor< float, 1 >();
auto ans_value = ans.tensor< float, 2 >();
std::cout < < ans_value(0,0) < < std::endl;
return 0;
}
对应的 Makefile 文件要改一下:
CROSS_COMPILE:=/home/red/Samba/arm-gnu-toolchain-13.3.rel1-x86_64-aarch64-none-linux-gnu/bin/aarch64-linux-
TARGET=tfcpp
CFLAGS:=-I/home/red/.cache/bazel/_bazel_red/81f6b3978d226a63c6d017ab1c0efa9f/execroot/org_tensorflow/bazel-out/aarch64-opt/bin/tensorflow/include/
CFLAGS+=-I/home/red/.cache/bazel/_bazel_red/81f6b3978d226a63c6d017ab1c0efa9f/execroot/org_tensorflow/bazel-out/aarch64-opt/bin/tensorflow/include/src
CFLAGS+=-I/home/red/.cache/bazel/_bazel_red/81f6b3978d226a63c6d017ab1c0efa9f/execroot/org_tensorflow/bazel-out/aarch64-opt/bin/tensorflow/include/_virtual_includes/float8/
CFLAGS+=-I/home/red/.cache/bazel/_bazel_red/81f6b3978d226a63c6d017ab1c0efa9f/execroot/org_tensorflow/bazel-out/aarch64-opt/bin/tensorflow/include/_virtual_includes/int4/
LDFLAGS:=-L/home/red/.cache/bazel/_bazel_red/81f6b3978d226a63c6d017ab1c0efa9f/execroot/org_tensorflow/bazel-out/aarch64-opt/bin/tensorflow -ltensorflow_framework
LDFLAGS+=-L/home/red/.cache/bazel/_bazel_red/81f6b3978d226a63c6d017ab1c0efa9f/execroot/org_tensorflow/bazel-out/aarch64-opt/bin/tensorflow -ltensorflow
$(TARGET):$(TARGET).cpp
$(CROSS_COMPILE)g++ $(CFLAGS) $(LDFLAGS) $^ -o $@
clean:
rm -frv $(TARGET)
编译生成测试程序并测试对一个全白色图片的测试结果
可以看到和上一篇 PC 端对比的结果还是很一致的。
全部0条评论
快来发表一下你的评论吧 !