0. 引言
通过源码方式安装，并进行一定程度的解读，有助于理解tensorflow源码，本文主要基于tensorflow v1.8源码，并借鉴于如何阅读TensorFlow源码.
首先，自然是需要去bazel官网了解下必备知识，如(1)什么是bazel; (2)bazel如何对cpp项目进行构建的; (3)bazel构建时候的函数大全。然后就是bazel官网的一些其他更细节部分了。下文中会给出超链接。
ps: 找了很久，基本可以确定bazel除了官网是没有如书籍等资料出现的，所以只有官网和别人博客这2个途径进行学习了解
因为bazel官网很多链接不在左边的导航中，所以推荐直接将整个网站镜像下来
wget -m -c -x -np -k -E -p https://docs.bazel.build/versions/master/bazel-overview.html
1. 从源码编译tensorflow
如下图所示：
图1.1 github上tensorflow v1.8源码目录
1.1 先配置
源代码树的根目录中包含了一个名为 configure 的 bash 脚本。此脚本会要求您确定所有相关 TensorFlow 依赖项的路径名，并指定其他构建配置选项，例如编译器标记。您必须先运行此脚本，然后才能创建 pip 软件包并安装 TensorFlow
然后是运行该configure
./configure
$ cd tensorflow # cd to the top-level directory created
$ ./configure
Please specify the location of python. [Default is /usr/bin/python]: /usr/bin/python2.7 # python解释器路径
Found possible Python library paths:
/usr/local/lib/python2.7/dist-packages
/usr/lib/python2.7/dist-packages
Please input the desired Python library path to use. Default is [/usr/lib/python2.7/dist-packages] # python 库路径
Using python library path: /usr/local/lib/python2.7/dist-packages
Please specify optimization flags to use during compilation when bazel option "--config=opt" is specified [Default is -march=native]: # 是否在编译期间启用优化
Do you wish to use jemalloc as the malloc implementation? [Y/n] # 是否将 jemalloc 作为malloc的实现
jemalloc enabled
Do you wish to build TensorFlow with Google Cloud Platform support? [y/N] # 是否开启google云平台支持
No Google Cloud Platform support will be enabled for TensorFlow
Do you wish to build TensorFlow with Hadoop File System support? [y/N] # 是否开启hdfs的支持
No Hadoop File System support will be enabled for TensorFlow
Do you wish to build TensorFlow with the XLA just-in-time compiler (experimental)? [y/N] # 是否启用尚在实验性质的XLA jit编译
No XLA support will be enabled for TensorFlow
Do you wish to build TensorFlow with VERBS support? [y/N] # 是否开启VERBS支持
No VERBS support will be enabled for TensorFlow
Do you wish to build TensorFlow with OpenCL support? [y/N] # 是否开启OpenCL支持
No OpenCL support will be enabled for TensorFlow
Do you wish to build TensorFlow with CUDA support? [y/N] Y # 是否开启CUDA支持
CUDA support will be enabled for TensorFlow
Do you want to use clang as CUDA compiler? [y/N] # 是否将clang作为CUDA的编译器
nvcc will be used as CUDA compiler
Please specify the CUDA SDK version you want to use, e.g. 7.0. [Leave empty to default to CUDA 9.0]: 9.0 # 选择cuda版本
Please specify the location where CUDA 9.0 toolkit is installed. Refer to README.md for more details. [Default is /usr/local/cuda]: # 告知cuda的安装路径
Please specify which gcc should be used by nvcc as the host compiler. [Default is /usr/bin/gcc]: # 指定host侧的 编译器
Please specify the cuDNN version you want to use. [Leave empty to default to cuDNN 7.0]: 7 # cuDNN版本
Please specify the location where cuDNN 7 library is installed. Refer to README.md for more details. [Default is /usr/local/cuda]: # 告知cuDNN 的安装路径
Please specify a list of comma-separated CUDA compute capabilities you want to build with. # 告知当前机器上GPU的计算力
You can find the compute capability of your device at: https://developer.nvidia.com/cuda-gpus.
Please note that each additional compute capability significantly increases your build time and binary size.
Do you wish to build TensorFlow with MPI support? [y/N] # 是否开启MPI支持
MPI support will not be enabled for TensorFlow
Configuration finished
我们先来看看configure到底做了什么事情，
#!/usr/bin/env bash
set -e
set -o pipefail
if [ -z "$PYTHON_BIN_PATH" ]; then
PYTHON_BIN_PATH=$(which python || which python3 || true)
fi
# Set all env variables
CONFIGURE_DIR=$(dirname "$0")
"$PYTHON_BIN_PATH" "${CONFIGURE_DIR}/configure.py" "$@" # 这行表明该configure文件是通过调用 对应的configure.py来完成配置过程的
echo "Configuration finished"
从configure.py的第1491行开始，发现如上述运行代码中展示的配置过程
set_build_var(environ_cp, 'TF_NEED_JEMALLOC', 'jemalloc as malloc',
'with_jemalloc', True)
set_build_var(environ_cp, 'TF_NEED_GCP', 'Google Cloud Platform',
'with_gcp_support', True, 'gcp')
set_build_var(environ_cp, 'TF_NEED_HDFS', 'Hadoop File System',
'with_hdfs_support', True, 'hdfs')
set_build_var(environ_cp, 'TF_NEED_AWS', 'Amazon AWS Platform',
'with_aws_support', True, 'aws')
set_build_var(environ_cp, 'TF_NEED_KAFKA', 'Apache Kafka Platform',
'with_kafka_support', True, 'kafka')
set_build_var(environ_cp, 'TF_ENABLE_XLA', 'XLA JIT', 'with_xla_support',
False, 'xla')
set_build_var(environ_cp, 'TF_NEED_GDR', 'GDR', 'with_gdr_support',
False, 'gdr')
set_build_var(environ_cp, 'TF_NEED_VERBS', 'VERBS', 'with_verbs_support',
False, 'verbs')
所以配置过程可以简单的理解，就是各种参数的收集，最后会有3个文件的时间信息更新（即生成或者修改的）：
其中.bazelrc内容如下:
import /mnt/d/tensorflow/tensorflow-master/.tf_configure.bazelrc
即导入的是在当前文件夹下新生成的文件.tf_configure.bazelrc，而该文件就纪录了配置
build --action_env PYTHON_BIN_PATH="/home/shouhuxianjian/anaconda3/bin/python"
build --action_env PYTHON_LIB_PATH="/home/shouhuxianjian/anaconda3/lib/python3.6/site-packages"
build --python_path="/home/shouhuxianjian/anaconda3/bin/python"
build --define with_jemalloc=true
build:gcp --define with_gcp_support=true
build:hdfs --define with_hdfs_support=true
build:aws --define with_aws_support=true
build:kafka --define with_kafka_support=true
build:xla --define with_xla_support=true
build:gdr --define with_gdr_support=true
build:verbs --define with_verbs_support=true
build --action_env TF_NEED_OPENCL_SYCL="0"
build --action_env TF_NEED_CUDA="0"
build --action_env TF_DOWNLOAD_CLANG="0"
build --define grpc_no_ares=true
build:opt --copt=-march=native
build:opt --host_copt=-march=native
build:opt --define with_default_optimizations=true
build --strip=always
其中的build:hdfs等形式等效于build --config=hdfs ，见这里的--config
上述在hdfs,gcp,aws,kafka选择时点击了N，如果点击Y则会变换成如下形式:
build --define with_gcp_support=true
build --define with_hdfs_support=true
build --define with_aws_support=true
build --define with_kafka_support=true
可以发现和
build --define with_jemalloc=true
一样了。而对于bazel而言，如果build:package形式，则编译时候会忽略该包(hdfs包中BUILD内容为：
# 文档在 tensorflow-master/third_party/hadoop/BUILD
package(default_visibility = ["//visibility:public"])
licenses(["notice"]) # Apache 2.0
exports_files(["LICENSE.txt"])
cc_library(
name = "hdfs",
hdrs = ["hdfs.h"],
)
所以下面真的调用bazel进行编译的时候，需要显示采用--config=opt来告知bazel，不要忽略opt这个package（这里是为了使用command:name中group这个特性）。
1.2 再bazel编译
如果只编译支持cpu的，敲如下代码
$ bazel build --config=opt //tensorflow/tools/pip_package:build_pip_package
如果需要gpu支持的，敲如下代码：
$ bazel build --config=opt --config=cuda //tensorflow/tools/pip_package:build_pip_package
1.2.1 BUILD文件结构格式推荐
在解读tensorflow-master/tensorflow/tools/pip_package/BUILD的时候，需要温习bazel构建时候的函数大全，还有官方推荐的BUILD文件结构格式File structure. 如下形式：
Package description (a comment)
All load() statements
The package() function.
Calls to rules and macros
1.2.2 tensorflow/tools/pip_package/BUILD文件解读
在下面的tensorflow/tools/pip_package/BUILD文件中，你可以看到package描述，load函数，transitive_hdrs，生成python的pb_binary，内部变量COMMON_PIP_DEPS，filegroup，生成shell的sh_binary和genrule等等。
# Description:
# Tools for building the TensorFlow pip package.
# 原型：package(default_deprecation, default_testonly, default_visibility, features)
# 此函数声明适用于包中每个后续规则的元数据。 它最多只能在一个包（BUILD文件）中使用一次。
# 此函数应该出现文件顶部，在所有load（）语句之后，任何规则之前的范围内，调用package（）函数。
# [package](https://docs.bazel.build/versions/master/be/functions.html#package)
# private表示后续的规则默认情况下只能在当前包内可见 https://docs.bazel.build/versions/master/be/common-definitions.html#common-attributes
package(default_visibility = ["//visibility:private"])
# Bazel的扩展是以.bzl结尾的文件。 通过使用load语句从可以从bazel的扩展文件中导入对应符号到当前BUILD中使用。
# [load](https://docs.bazel.build/versions/master/skylark/concepts.html)
load(
"//tensorflow:tensorflow.bzl",
"if_not_windows",
"if_windows",
"transitive_hdrs",
)
load("//third_party/mkl:build_defs.bzl", "if_mkl")
load("//tensorflow:tensorflow.bzl", "if_cuda")
load("@local_config_tensorrt//:build_defs.bzl", "if_tensorrt")
load("//tensorflow/core:platform/default/build_config_root.bzl", "tf_additional_license_deps")
# This returns a list of headers of all public header libraries (e.g.,
# framework, lib), and all of the transitive dependencies of those
# public headers. Not all of the headers returned by the filegroup
# are public (e.g., internal headers that are included by public
# headers), but the internal headers need to be packaged in the
# pip_package for the public headers to be properly included.
#
# Public headers are therefore defined by those that are both:
#
# 1) "publicly visible" as defined by bazel
# 2) Have documentation.
#
# This matches the policy of "public" for our python API.
transitive_hdrs(
name = "included_headers",
deps = [
"//tensorflow/core:core_cpu",
"//tensorflow/core:framework",
"//tensorflow/core:lib",
"//tensorflow/core:protos_all_cc",
"//tensorflow/core:stream_executor",
"//third_party/eigen3",
] + if_cuda([
"@local_config_cuda//cuda:cuda_headers",
]),
)
py_binary(
name = "simple_console",
srcs = ["simple_console.py"],
srcs_version = "PY2AND3",
deps = ["//tensorflow:tensorflow_py"],
)
COMMON_PIP_DEPS = [
":licenses",
"MANIFEST.in",
"README",
"setup.py",
":included_headers",
"//tensorflow:tensorflow_py",
"//tensorflow/contrib/autograph:autograph",
"//tensorflow/contrib/autograph/converters:converters",
"//tensorflow/contrib/autograph/converters:test_lib",
"//tensorflow/contrib/autograph/impl:impl",
"//tensorflow/contrib/autograph/pyct:pyct",
"//tensorflow/contrib/autograph/pyct/static_analysis:static_analysis",
"//tensorflow/contrib/boosted_trees:boosted_trees_pip",
"//tensorflow/contrib/cluster_resolver:cluster_resolver_pip",
"//tensorflow/contrib/data/python/kernel_tests:dataset_serialization_test",
"//tensorflow/contrib/data/python/ops:contrib_op_loader",
"//tensorflow/contrib/eager/python/examples:examples_pip",
"//tensorflow/contrib/eager/python:checkpointable_utils",
"//tensorflow/contrib/eager/python:evaluator",
"//tensorflow/contrib/gan:gan",
"//tensorflow/contrib/graph_editor:graph_editor_pip",
"//tensorflow/contrib/keras:keras",
"//tensorflow/contrib/labeled_tensor:labeled_tensor_pip",
"//tensorflow/contrib/nn:nn_py",
"//tensorflow/contrib/predictor:predictor_pip",
"//tensorflow/contrib/proto:proto_pip",
"//tensorflow/contrib/receptive_field:receptive_field_pip",
"//tensorflow/contrib/rpc:rpc_pip",
"//tensorflow/contrib/session_bundle:session_bundle_pip",
"//tensorflow/contrib/signal:signal_py",
"//tensorflow/contrib/signal:test_util",
"//tensorflow/contrib/slim:slim",
"//tensorflow/contrib/slim/python/slim/data:data_pip",
"//tensorflow/contrib/slim/python/slim/nets:nets_pip",
"//tensorflow/contrib/specs:specs",
"//tensorflow/contrib/summary:summary_test_util",
"//tensorflow/contrib/tensor_forest:init_py",
"//tensorflow/contrib/tensor_forest/hybrid:hybrid_pip",
"//tensorflow/contrib/timeseries:timeseries_pip",
"//tensorflow/contrib/tpu",
"//tensorflow/examples/tutorials/mnist:package",
"//tensorflow/python:distributed_framework_test_lib",
"//tensorflow/python:meta_graph_testdata",
"//tensorflow/python:spectral_ops_test_util",
"//tensorflow/python:util_example_parser_configuration",
"//tensorflow/python/debug:debug_pip",
"//tensorflow/python/eager:eager_pip",
"//tensorflow/python/kernel_tests/testdata:self_adjoint_eig_op_test_files",
"//tensorflow/python/saved_model:saved_model",
"//tensorflow/python/tools:tools_pip",
"//tensorflow/python:test_ops",
"//tensorflow/tools/dist_test/server:grpc_tensorflow_server",
]
# On Windows, python binary is a zip file of runfiles tree.
# Add everything to its data dependency for generating a runfiles tree
# for building the pip package on Windows.
py_binary(
name = "simple_console_for_windows",
srcs = ["simple_console_for_windows.py"],
data = COMMON_PIP_DEPS,
srcs_version = "PY2AND3",
deps = ["//tensorflow:tensorflow_py"],
)
filegroup(
name = "licenses",
data = [
"//third_party/eigen3:LICENSE",
"//third_party/fft2d:LICENSE",
"//third_party/hadoop:LICENSE.txt",
"@absl_py//absl/flags:LICENSE",
"@arm_neon_2_x86_sse//:LICENSE",
"@astor_archive//:LICENSE",
"@aws//:LICENSE",
"@boringssl//:LICENSE",
"@com_google_absl//:LICENSE",
"@com_googlesource_code_re2//:LICENSE",
"@cub_archive//:LICENSE.TXT",
"@curl//:COPYING",
"@eigen_archive//:COPYING.MPL2",
"@farmhash_archive//:COPYING",
"@fft2d//:fft/readme.txt",
"@flatbuffers//:LICENSE.txt",
"@gast_archive//:PKG-INFO",
"@gemmlowp//:LICENSE",
"@gif_archive//:COPYING",
"@grpc//:LICENSE",
"@highwayhash//:LICENSE",
"@jemalloc//:COPYING",
"@jpeg//:LICENSE.md",
"@kafka//:LICENSE",
"@libxsmm_archive//:LICENSE",
"@lmdb//:LICENSE",
"@local_config_nccl//:LICENSE",
"@local_config_sycl//sycl:LICENSE.text",
"@grpc//third_party/nanopb:LICENSE.txt",
"@grpc//third_party/address_sorting:LICENSE",
"@nasm//:LICENSE",
"@nsync//:LICENSE",
"@pcre//:LICENCE",
"@png_archive//:LICENSE",
"@protobuf_archive//:LICENSE",
"@six_archive//:LICENSE",
"@snappy//:COPYING",
"@swig//:LICENSE",
"@termcolor_archive//:COPYING.txt",
"@zlib_archive//:zlib.h",
"@org_python_pypi_backports_weakref//:LICENSE",
] + if_mkl([
"//third_party/mkl:LICENSE",
]) + tf_additional_license_deps(),
)
# 对应的shell二进制规则，其中涉及到了select(主要用来做平台依赖选择)，在bazel的编译命令中，并未显式的指定build_pip_package的属性，所以这里采用了默认的条件
# [select](https://docs.bazel.build/versions/master/skylark/lib/globals.html#select)
# [select](https://docs.bazel.build/versions/master/be/functions.html#select)
sh_binary(
name = "build_pip_package",
srcs = ["build_pip_package.sh"],
data = select({
"//tensorflow:windows": [":simple_console_for_windows"],
"//tensorflow:windows_msvc": [":simple_console_for_windows"],
"//conditions:default": COMMON_PIP_DEPS + [
":simple_console",
"//tensorflow/contrib/lite/python:interpreter_test_data",
"//tensorflow/contrib/lite/python:tf_lite_py_pip",
"//tensorflow/contrib/lite/toco:toco",
"//tensorflow/contrib/lite/toco/python:toco_wrapper",
"//tensorflow/contrib/lite/toco/python:toco_from_protos",
],
}) + if_mkl(["//third_party/mkl:intel_binary_blob"]) + if_tensorrt([
"//tensorflow/contrib/tensorrt:init_py",
]),
)
# A genrule for generating a marker file for the pip package on Windows
#
# This only works on Windows, because :simple_console_for_windows is a
# python zip file containing everything we need for building the pip package.
# However, on other platforms, due to https://github.com/bazelbuild/bazel/issues/4223,
# when C++ extensions change, this generule doesn't rebuild.
genrule(
name = "win_pip_package_marker",
srcs = if_windows([
":build_pip_package",
":simple_console_for_windows",
]),
outs = ["win_pip_package_marker_file"],
cmd = select({
"//conditions:default": "touch $@",
"//tensorflow:windows": "md5sum $(locations :build_pip_package) $(locations :simple_console_for_windows) > $@",
}),
visibility = ["//visibility:public"],
)
1.2.3 编译build_pip_package的过程
因编译命令显式的编译build_pip_package，对应上述文件中的sh_binary。sh_binary中主要负责依赖的data的生成，其中基于平台依赖选用了select函数，且bazel命令行中并未对当前build_pip_package做显式的选择，所以读取默认配置，
COMMON_PIP_DEPS + [
":simple_console",
"//tensorflow/contrib/lite/python:interpreter_test_data",
"//tensorflow/contrib/lite/python:tf_lite_py_pip",
"//tensorflow/contrib/lite/toco:toco",
"//tensorflow/contrib/lite/toco/python:toco_wrapper",
"//tensorflow/contrib/lite/toco/python:toco_from_protos",
]
那么现在焦点就转移到COMMON_PIP_DEPS 部分了。该变量中，一开始的三个文件MANIFEST.in、README、setup.py是直接存在的，因此不会有什么操作。然后我们看下一行的
:included_headers
这里表示当前范围内的target，所以是对应的
# This matches the policy of "public" for our python API.
transitive_hdrs(
name = "included_headers",
deps = [
"//tensorflow/core:core_cpu",
"//tensorflow/core:framework",
"//tensorflow/core:lib",
"//tensorflow/core:protos_all_cc",
"//tensorflow/core:stream_executor",
"//third_party/eigen3",
] + if_cuda([
"@local_config_cuda//cuda:cuda_headers",
]),
)
而transitive_hdrs 并不是关键字类型的函数，是由上面的load导入的
load(
"//tensorflow:tensorflow.bzl",
"if_not_windows",
"if_windows",
"transitive_hdrs",
)
transitive_hdrs在tensorflow:tensorflow.bzl中的实现为
# Bazel rule for collecting the header files that a target depends on.
def _transitive_hdrs_impl(ctx):
outputs = depset()
for dep in ctx.attr.deps:
outputs += dep.cc.transitive_headers
return struct(files=outputs)
# 这里调用了对应的rule函数
# [rule](https://docs.bazel.build/versions/master/skylark/lib/globals.html#rule)
_transitive_hdrs = rule(
attrs = {
"deps": attr.label_list(
allow_files = True,
providers = ["cc"],
),
},
implementation = _transitive_hdrs_impl,
)
# transitive_hdrs所在的位置，其通过内部的_transitive_hdrs规则，而该规则是通过_transitive_hdrs_impl 实现的
def transitive_hdrs(name, deps=[], **kwargs):
_transitive_hdrs(name=name + "_gather", deps=deps)
native.filegroup(name=name, srcs=[":" + name + "_gather"])
这部分先放下，我们接着找和cpp交互的部分。我们先关注下，接下来的
"//tensorflow:tensorflow_py",
当前WORKSPACE所在的位置为根位置//，后面的tensorflow表示对应的tensorflow文件夹，后面的tensorflow_py可以定位到文件tensorflow/BUILD中
# 当前文件为tensorflow/BUILD的539-548行
py_library(
name = "tensorflow_py",
srcs = ["__init__.py"],
srcs_version = "PY2AND3",
visibility = ["//visibility:public"],
deps = [
"//tensorflow/python",
"//tensorflow/tools/api/generator:python_api",
],
)
这里依赖于//tensorflow/python这个包，这个包依赖于tensorflow/python/BUILD进行生成，其内部
# 当前文件为tensorflow/python/BUILD
py_library(
name = "python",
srcs = ["__init__.py"],
srcs_version = "PY2AND3",
visibility = [
"//tensorflow:__pkg__",
"//tensorflow/compiler/aot/tests:__pkg__", # TODO(b/34059704): remove when fixed
"//tensorflow/contrib/learn:__pkg__", # TODO(b/34059704): remove when fixed
"//tensorflow/contrib/learn/python/learn/datasets:__pkg__", # TODO(b/34059704): remove when fixed
"//tensorflow/contrib/lite/toco/python:__pkg__", # TODO(b/34059704): remove when fixed
"//tensorflow/python/debug:__pkg__", # TODO(b/34059704): remove when fixed
"//tensorflow/python/tools:__pkg__", # TODO(b/34059704): remove when fixed
"//tensorflow/tools/api/generator:__pkg__",
"//tensorflow/tools/quantization:__pkg__", # TODO(b/34059704): remove when fixed
],
deps = [
":no_contrib",
"//tensorflow/contrib:contrib_py",
],
)
这里依赖于:no_contrib 这个target，那么我们关注下
# 当前文件为tensorflow/python/BUILD
py_library(
name = "no_contrib",
srcs = ["__init__.py"],
srcs_version = "PY2AND3",
visibility = [
"//tensorflow:__pkg__",
],
deps = [
":array_ops",
":bitwise_ops",
":boosted_trees_ops",
":check_ops",
":client",
":client_testlib",
":confusion_matrix",
":control_flow_ops",
":cudnn_rnn_ops_gen",
":errors",
":framework",
":framework_for_generated_wrappers",
":functional_ops",
":gradient_checker",
":graph_util",
":histogram_ops",
":image_ops",
":initializers_ns",
":io_ops",
":layers",
":lib",
":list_ops",
":manip_ops",
":math_ops",
":metrics",
":nn",
":ops",
":platform",
":pywrap_tensorflow",
":saver_test_utils",
":script_ops",
":session_ops",
":sets",
":sparse_ops",
":spectral_ops",
":spectral_ops_test_util",
":standard_ops",
":state_ops",
":string_ops",
":subscribe",
":summary",
":tensor_array_ops",
":test_ops", # TODO: Break testing code out into separate rule.
":tf_cluster",
":tf_item",
":tf_optimizer",
":training",
":util",
":weights_broadcast_ops",
"//tensorflow/core:protos_all_py",
"//tensorflow/python/data",
"//tensorflow/python/estimator:estimator_py",
"//tensorflow/python/feature_column:feature_column_py",
"//tensorflow/python/keras",
"//tensorflow/python/ops/distributions",
"//tensorflow/python/ops/linalg",
"//tensorflow/python/ops/losses",
"//tensorflow/python/profiler",
"//tensorflow/python/saved_model",
"//third_party/py/numpy",
],
)
我们也跟随.如何阅读TensorFlow源码去找pywrap_tensorflow这个部分,其中pywrap_tensorflow target依赖于pywrap_tensorflow_internal这个target的，而pywrap_tensorflow_internal就是通过swig从cc文件生成对应的python接口文件部分了
# 当前文件为tensorflow/python/BUILD 3421行
py_library(
name = "pywrap_tensorflow",
srcs = [
"pywrap_tensorflow.py",
] + if_static(
["pywrap_dlopen_global_flags.py"],
# Import will fail, indicating no global dlopen flags
otherwise = [],
),
srcs_version = "PY2AND3",
deps = [":pywrap_tensorflow_internal"],
)
tf_py_wrap_cc(
name = "pywrap_tensorflow_internal",
srcs = ["tensorflow.i"],
swig_includes = [
"client/device_lib.i",
"client/events_writer.i",
"client/tf_session.i",
"client/tf_sessionrun_wrapper.i",
"framework/cpp_shape_inference.i",
"framework/python_op_gen.i",
"grappler/cluster.i",
"grappler/cost_analyzer.i",
"grappler/item.i",
"grappler/model_analyzer.i",
"grappler/tf_optimizer.i",
"lib/core/bfloat16.i",
"lib/core/py_exception_registry.i",
"lib/core/py_func.i",
"lib/core/strings.i",
"lib/io/file_io.i",
"lib/io/py_record_reader.i",
"lib/io/py_record_writer.i",
"platform/base.i",
"platform/stacktrace_handler.i",
"pywrap_tfe.i",
"training/quantize_training.i",
"training/server_lib.i",
"util/kernel_registry.i",
"util/port.i",
"util/py_checkpoint_reader.i",
"util/stat_summarizer.i",
"util/tfprof.i",
"util/transform_graph.i",
"util/util.i",
],
win_def_file = select({
"//tensorflow:windows": ":pywrap_tensorflow_filtered_def_file",
"//conditions:default": None,
}),
deps = [
":bfloat16_lib",
":cost_analyzer_lib",
":model_analyzer_lib",
":cpp_python_util",
":cpp_shape_inference",
":kernel_registry",
":numpy_lib",
":safe_ptr",
":py_exception_registry",
":py_func_lib",
":py_record_reader_lib",
":py_record_writer_lib",
":python_op_gen",
":tf_session_helper",
"//tensorflow/c:c_api",
"//tensorflow/c:checkpoint_reader",
"//tensorflow/c:python_api",
"//tensorflow/c:tf_status_helper",
"//tensorflow/c/eager:c_api",
"//tensorflow/core/distributed_runtime/rpc:grpc_rpc_factory_registration",
"//tensorflow/core/distributed_runtime/rpc:grpc_server_lib",
"//tensorflow/core/distributed_runtime/rpc:grpc_session",
"//tensorflow/core/grappler:grappler_item",
"//tensorflow/core/grappler:grappler_item_builder",
"//tensorflow/core/grappler/clusters:cluster",
"//tensorflow/core/grappler/clusters:single_machine",
"//tensorflow/core/grappler/clusters:virtual_cluster",
"//tensorflow/core/grappler/costs:graph_memory",
"//tensorflow/core/grappler/optimizers:meta_optimizer",
"//tensorflow/core:lib",
"//tensorflow/core:reader_base",
"//tensorflow/core/debug",
"//tensorflow/core/distributed_runtime:server_lib",
"//tensorflow/core/profiler/internal:print_model_analysis",
"//tensorflow/tools/graph_transforms:transform_graph_lib",
"//tensorflow/python/eager:pywrap_tfe_lib",
"//tensorflow/python/eager:python_eager_op_gen",
"//util/python:python_headers",
] + (tf_additional_lib_deps() +
tf_additional_plugin_deps() +
tf_additional_verbs_deps() +
tf_additional_mpi_deps() +
tf_additional_gdr_deps()),
)
而tf_py_wrap_cc不是在bazel内置的规则中，所以是tensorflow自定义的一个规则，通过
load("//tensorflow:tensorflow.bzl", "tf_py_wrap_cc")
找到其实现为
# 此文件为tensorflow/tensorflow.bzl 1404行
def tf_py_wrap_cc(name,
srcs,
swig_includes=[],
deps=[],
copts=[],
**kwargs):
module_name = name.split("/")[-1]
# Convert a rule name such as foo/bar/baz to foo/bar/_baz.so
# and use that as the name for the rule producing the .so file.
cc_library_name = "/".join(name.split("/")[:-1] + ["_" + module_name + ".so"])
cc_library_pyd_name = "/".join(
name.split("/")[:-1] + ["_" + module_name + ".pyd"])
extra_deps = []
_py_wrap_cc(
name=name + "_py_wrap",
srcs=srcs,
swig_includes=swig_includes,
deps=deps + extra_deps,
toolchain_deps=["//tools/defaults:crosstool"],
module_name=module_name,
py_module_name=name)
vscriptname=name+"_versionscript"
_append_init_to_versionscript(
name=vscriptname,
module_name=module_name,
is_version_script=select({
"@local_config_cuda//cuda:darwin":False,
"//conditions:default":True,
}),
template_file=select({
"@local_config_cuda//cuda:darwin":clean_dep("//tensorflow:tf_exported_symbols.lds"),
"//conditions:default":clean_dep("//tensorflow:tf_version_script.lds")
})
)
extra_linkopts = select({
"@local_config_cuda//cuda:darwin": [
"-Wl,-exported_symbols_list",
"%s.lds"%vscriptname,
],
clean_dep("//tensorflow:windows"): [],
clean_dep("//tensorflow:windows_msvc"): [],
"//conditions:default": [
"-Wl,--version-script",
"%s.lds"%vscriptname,
]
})
extra_deps += select({
"@local_config_cuda//cuda:darwin": [
"%s.lds"%vscriptname,
],
clean_dep("//tensorflow:windows"): [],
clean_dep("//tensorflow:windows_msvc"): [],
"//conditions:default": [
"%s.lds"%vscriptname,
]
})
tf_cc_shared_object(
name=cc_library_name,
srcs=[module_name + ".cc"],
copts=(copts + if_not_windows([
"-Wno-self-assign", "-Wno-sign-compare", "-Wno-write-strings"
]) + tf_extension_copts()),
linkopts=tf_extension_linkopts() + extra_linkopts,
linkstatic=1,
deps=deps + extra_deps,
**kwargs)
native.genrule(
name="gen_" + cc_library_pyd_name,
srcs=[":" + cc_library_name],
outs=[cc_library_pyd_name],
cmd="cp $< $@",)
native.py_library(
name=name,
srcs=[":" + name + ".py"],
srcs_version="PY2AND3",
data=select({
clean_dep("//tensorflow:windows"): [":" + cc_library_pyd_name],
"//conditions:default": [":" + cc_library_name],
}))
因为swig是你编写好对应的example.c文件和example.i文件，然后通过调用swig命令生成example_wrap.c文件，通过gcc编译这2个c文件，就能生成对应的o文件，通过连接生成so文件，这时候就能够被python导入了。
上述自定义规则中
- tf_cc_shared_object 负责生成 so文件;
- 而native.py_library负责???
_py_wrap_cc则负责执行swig的命令，该自定义规则在同文件的1122行
# 此文件为tensorflow/tensorflow.bzl 1090行,下面的1122行就是_py_wrap_cc的位置
# Bazel rules for building swig files.
def _py_wrap_cc_impl(ctx):
srcs = ctx.files.srcs
if len(srcs) != 1:
fail("Exactly one SWIG source file label must be specified.", "srcs")
module_name = ctx.attr.module_name
src = ctx.files.srcs[0]
inputs = depset([src])
inputs += ctx.files.swig_includes
for dep in ctx.attr.deps:
inputs += dep.cc.transitive_headers
inputs += ctx.files._swiglib
inputs += ctx.files.toolchain_deps
swig_include_dirs = depset(_get_repository_roots(ctx, inputs))
swig_include_dirs += sorted([f.dirname for f in ctx.files._swiglib])
args = [
"-c++", "-python", "-module", module_name, "-o", ctx.outputs.cc_out.path,
"-outdir", ctx.outputs.py_out.dirname
]
args += ["-l" + f.path for f in ctx.files.swig_includes]
args += ["-I" + i for i in swig_include_dirs]
args += [src.path]
outputs = [ctx.outputs.cc_out, ctx.outputs.py_out]
ctx.action(
executable=ctx.executable._swig,
arguments=args,
inputs=list(inputs),
outputs=outputs,
mnemonic="PythonSwig",
progress_message="SWIGing " + src.path)
return struct(files=depset(outputs))
_py_wrap_cc = rule(
attrs = {
"srcs": attr.label_list(
mandatory = True,
allow_files = True,
),
"swig_includes": attr.label_list(
cfg = "data",
allow_files = True,
),
"deps": attr.label_list(
allow_files = True,
providers = ["cc"],
),
"toolchain_deps": attr.label_list(
allow_files = True,
),
"module_name": attr.string(mandatory = True),
"py_module_name": attr.string(mandatory = True),
"_swig": attr.label(
default = Label("@swig//:swig"),
executable = True,
cfg = "host",
),
"_swiglib": attr.label(
default = Label("@swig//:templates"),
allow_files = True,
),
},
outputs = {
"cc_out": "%{module_name}.cc",
"py_out": "%{py_module_name}.py",
},
implementation = _py_wrap_cc_impl,
)
上述中ctx.executable._swig 是为执行部分，其对应的
"_swig": attr.label(
default = Label("@swig//:swig"),
executable = True,
cfg = "host",
),
而swig就位于third_party/swig.BUILD中
licenses(["restricted"]) # GPLv3
exports_files(["LICENSE"])
cc_binary(
name = "swig",
srcs = [
"Source/CParse/cparse.h",
"Source/CParse/cscanner.c",
"Source/CParse/parser.c",
"Source/CParse/parser.h",
"Source/CParse/templ.c",
"Source/CParse/util.c",
"Source/DOH/base.c",
"Source/DOH/doh.h",
"Source/DOH/dohint.h",
"Source/DOH/file.c",
"Source/DOH/fio.c",
"Source/DOH/hash.c",
"Source/DOH/list.c",
"Source/DOH/memory.c",
"Source/DOH/string.c",
"Source/DOH/void.c",
"Source/Include/swigconfig.h",
"Source/Include/swigwarn.h",
"Source/Modules/allocate.cxx",
"Source/Modules/browser.cxx",
"Source/Modules/contract.cxx",
"Source/Modules/directors.cxx",
......
可以看成swig是需要先定义生成的一个target。这样基本一个流程就串起来了
- 先生成swig可执行文件
- 再通过对应i文件生成对应的wrap文件，并进行编译生成对应的so文件和py文件
- 就可以正常导入了
1.2.4 如何从python端找到对应的c源码文件
假设有python文件如下
import tensorflow as tf
import numpy as np
x_data = np.random.rand(100).astype(np.float32)
y_data = x_data * 0.1 + 0.3
W = tf.Variable(tf.random_uniform([1], -1.0, 1.0))
b = tf.Variable(tf.zeros([1]))
y = W * x_data + b
loss = tf.reduce_mean(tf.square(y - y_data))
optimizer = tf.train.GradientDescentOptimizer(0.5)
train = optimizer.minimize(loss)
init = tf.initialize_all_variables()
sess = tf.Session()
sess.run(init)
for step in range(0, 201):
sess.run(train)
if step % 20 == 0:
print(step, sess.run(W), sess.run(b))
假设想找到Session的位置,最后在/anaconda3/lib/python3.6/site-packages/tensorflow/python/client/session.py中找到对应的类
1.2.5 python和cpp函数名的对应
在底层cpp代码中，google采用的是驼峰形式去编写cpp的代码，如AbcDefGh，而前端语言python遵循的是小写下划线的方式，如abc_def_gh，所以这二者中，tensorflow内置了一个转换函数，在
/tensorflow/python/framework/python_op_gen.cc
string function_name;
python_op_gen_internal::GenerateLowerCaseOpName(op_def.name(),
&function_name);
在后续版本中，该函数的定义迁移到了/tensorflow/python/framework/python_op_gen_internal.cc
void GenerateLowerCaseOpName(const string& str, string* result) {
const char joiner = '_';
const int last_index = str.size() - 1;
for (int i = 0; i <= last_index; ++i) {
const char c = str[i];
// Emit a joiner only if a previous-lower-to-now-upper or a
// now-upper-to-next-lower transition happens.
if (isupper(c) && (i > 0)) {
if (islower(str[i - 1]) || ((i < last_index) && islower(str[i + 1]))) {
result->push_back(joiner);
}
}
result->push_back(tolower(c));
}
}
如我们要找tf.conv2d在cpp中的实现就是去找Conv2d。
参考资料：