算能RISC-V通用云编译飞桨paddlepaddle@openKylin留档

skywalk8163

已于 2024-03-19 07:14:51 修改

阅读量1.9k

点赞数 11

于 2024-02-24 00:57:44 首次发布

本文链接：https://blog.csdn.net/skywalk8163/article/details/136264198

版权

项目实践同时被 2 个专栏收录

122 篇文章 0 订阅

订阅专栏

人工智能

101 篇文章 4 订阅

订阅专栏

尝试在算能云riscv环境里编译飞桨。

先总结操作步骤：

飞桨编译安装简洁步骤

下载飞桨代码GitHub - PaddlePaddle/Paddle: PArallel Distributed Deep LEarning: Machine Learning Framework from Industrial Practice （『飞桨』核心框架，深度学习&机器学习高性能单机、分布式训练和跨平台部署）：git clone https://github.com/paddlepaddle/paddle

参照prAdding a compile option to Paddle for Risc-V · PaddlePaddle/Paddle@d3db383 · GitHub修改代码

然后编译

cmake ../ -DWITH_GPU=OFF -DWITH_RISCV=ON
make -j 128 TARGET=RISCV64_GENERIC

（注意以上两句最好分开执行，以便发现cmake是否报错，尤其是修改了cmake 配置文件之后。否则可能没生效而重复编译）

编译好后安装：

pip install paddlepaddle-0.0.0-cp38-cp38-linux_riscv64.whl -i https://mirror.baidu.com/pypi/simple

最后一步：注册libpaddle.so ，使用命令：patchelf --add-needed libatomic.so.1 /usr/local/lib/python3.8/dist-packages/paddle/base/libpaddle.so

这样大家就可以愉快的使用飞桨在RISCV环境下进行AI机器学习拉！

详细的飞桨编译安装过程

先git下载源码

git clone https://github.com/paddlepaddle/padle

配置环境

pip3 install protobuf

还有一些库等，具体可以参考编译pytorch的文档：算能RISC-V通用云开发空间编译pytorch @openKylin留档-CSDN博客

然后进入paddle目录编译

mkdir build

cd build

cmake ../

我的天，除了protobuf，竟然一下子编译完成了！

Automatic code generation for paddle/fluid/primitive succeed.
Automatic code generation for decomp interface succeed.
WITH_DLNNE:
-- Configuring done
-- Generating done
-- Build files have been written to: /root/github/paddle/build

是我肤浅了，后面还需要编译呢root@863c89a419ec:~/github/paddle/build# make

make -j 16

make -j 8 TARGET=RISCV64_GENERIC -dw

如果碰到报错，根据报错进行处理，比如哪个第三方库编译失败，就删除那个目录，然后在third_party目录执行：git submodule update --init --recursive

然后再make 即可。

困了，让它编译去吧。

报错了

通过官网issue，发现cmake指令为：

cmake ../ -DWITH_GPU=OFF -WITH_RISCV=ON

make 指令为;

make -j 16 TARGET=RISCV64_GENERIC

发现cmake指令有问题

拼写错误，应该是：

cmake ../ -DWITH_GPU=OFF -DWITH_RISCV=ON

然后再make

看MakeFile文件里面没有RISCV选项，不知道是不是cmake编译器那边支持，先执行看看。（后来知道，人家这个RISCV是对应了专门的pr的，目前这个pr还没有合并，所以需要手工改代码，具体修改方法如下）

cmake结束之后提示WITH_RISCV没有生效

手工添加飞桨对riscv的支持

修改CMakeFile文件，在WITH_ARM的后面加上下面内容：

if(WITH_RISCV)
  set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -fPIC")
  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fPIC")
  set(WITH_XBYAK
      OFF
      CACHE STRING "Disable XBYAK when compiling WITH_RISCV=ON." FORCE)
  set(WITH_MKL
      OFF
      CACHE STRING "Disable MKL when compiling WITH_RISCV=ON." FORCE)
  set(WITH_AVX
      OFF
      CACHE STRING "Disable AVX when compiling WITH_AVX=OFF." FORCE)
  add_definitions(-DPADDLE_WITH_RISCV)
endif()

修改文件：cmake/flags.cmake

在ARM后面加上：AND NOT WITH_RISCV

同时找到了-m64的位置。因为加了这句，就不用手工删除-m64这句。

修改文件：paddle/fluid/operators/search_compute.h

一共需要修改四处，在四处加上defined(PADDLE_WITH_RISCV)：

#pragma once

#if !defined(PADDLE_WITH_ARM) && !defined(PADDLE_WITH_SW) && \
    !defined(PADDLE_WITH_MIPS) && !defined(PADDLE_WITH_LOONGARCH) && \
	!defined(PADDLE_WITH_RISCV)
#include <immintrin.h>
#endif
#include <cfloat>
@@ -103,7 +104,8 @@ void call_gemm_batched(const framework::ExecutionContext& ctx,
}

#if !defined(PADDLE_WITH_ARM) && !defined(PADDLE_WITH_SW) && \
    !defined(PADDLE_WITH_MIPS) && !defined(PADDLE_WITH_LOONGARCH) && \
	!defined(PADDLE_WITH_RISCV)

#define __m256x __m256

@@ -144,7 +146,8 @@ inline void axpy(const T* x, T* y, size_t len, const T alpha) {
                      _mm256_mul_px(mm_alpha, _mm256_load_px(x + jjj))));
  }
#elif defined(PADDLE_WITH_ARM) || defined(PADDLE_WITH_SW) || \
    defined(PADDLE_WITH_MIPS) || defined(PADDLE_WITH_LOONGARCH) || \
	defined(PADDLE_WITH_RISCV)
  PADDLE_THROW(platform::errors::Unimplemented("axpy is not supported"));
#else
  lll = len & ~SSE_CUT_LEN_MASK;
@@ -174,7 +177,8 @@ inline void axpy_noadd(const T* x, T* y, size_t len, const T alpha) {
    _mm256_store_px(y + jjj, _mm256_mul_px(mm_alpha, _mm256_load_px(x + jjj)));
  }
#elif defined(PADDLE_WITH_ARM) || defined(PADDLE_WITH_SW) || \
    defined(PADDLE_WITH_MIPS) || defined(PADDLE_WITH_LOONGARCH) || \
	defined(PADDLE_WITH_RISCV)
  PADDLE_THROW(platform::errors::Unimplemented("axpy_noadd is not supported"));
#else
  lll = len & ~SSE_CUT_LEN_MASK;

修改文件：paddle/fluid/platform/denormal.cc

33行if那句后面加上RISCV，变成这样：

#if !defined(GCC_WITHOUT_INTRINSICS) && !defined(PADDLE_WITH_ARM) && \
    !defined(PADDLE_WITH_SW) && !defined(PADDLE_WITH_MIPS) &&        \
    !defined(_WIN32) && !defined(PADDLE_WITH_LOONGARCH) &&           \
        !defined(PADDLE_WITH_RISCV)
#define DENORM_USE_INTRINSICS

修改文件paddle/phi/backends/cpu/cpu_info.cc

  } else {
#if !defined(WITH_NV_JETSON) && !defined(PADDLE_WITH_ARM) &&  \
    !defined(PADDLE_WITH_SW) && !defined(PADDLE_WITH_MIPS) && \
    !defined(PADDLE_WITH_LOONGARCH) && !defined(PADDLE_WITH_RISCV)
    std::array<int, 4> reg;
    cpuid(reg.data(), 0);
    int nIds = reg[0];

修改文件paddle/phi/backends/cpu/cpu_info.h

#define cpuid(reg, x) __cpuidex(reg, x, 0)
#else
#if !defined(WITH_NV_JETSON) && !defined(PADDLE_WITH_ARM) &&  \
    !defined(PADDLE_WITH_SW) && !defined(PADDLE_WITH_MIPS) && \
    !defined(PADDLE_WITH_LOONGARCH) && !defined(PADDLE_WITH_RISCV)
#include <cpuid.h>
inline void cpuid(int reg[4], int x) {

感觉离曙光很近了。

重新cmake和make，希望不会碰到“pmmintrin.h”文件找不到的错误。

发现少修改一个文件

修改文件cmake/third_party.cmake

if(WITH_POCKETFFT)
  include(external/pocketfft)
  list(APPEND third_party_deps extern_pocketfft)
  add_definitions(-DPADDLE_WITH_POCKETFFT)
if(WITH_RISCV)
  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-maybe-uninitialized") # Warnings in pocketfft_hdronly.h
endif()
endif()

11点开始编译，计时开始！15：37分编译完成，耗时4小时37分钟！

哈哈，它静静的躺在那里：

root@863c89a419ec:~/github/paddle/build/python# ls dist/
paddlepaddle-0.0.0-cp38-cp38-linux_riscv64.whl
root@863c89a419ec:~/github/paddle/build/python# pip3 install dist/paddlepaddle-0.0.0-cp38-cp38-linux_riscv64.whl  -i https://mirror.baidu.com/pypi/simple
Processing ./dist/paddlepaddle-0.0.0-cp38-cp38-linux_riscv64.whl

安装：

pip install paddlepaddle-0.0.0-cp38-cp38-linux_riscv64.whl

安装后执行import paddle 报错：ImportError: /usr/local/lib/python3.8/dist-packages/paddle/base/libpaddle.so: undefined symbol: __atomic_exchange_1

按照这个issue 算能云RISCV环境编译后import paddle报错 · Issue #62037 · PaddlePaddle/Paddle · GitHub里面修改，还是报错

将build文件夹删除，重新建立，重新cmake 和make，并把-j改成128

问题依旧

解决import paddle报错

突然想到pytorch安装的时候也碰到这个报错，于是找到了解决方法：

只需要用patchelf注册一下就行了`patchelf --add-needed libatomic.so.1 /usr/local/lib/python3.8/dist-packages/paddle/base/libpaddle.so `

如果没有patchelf，apt install patchelf安装即可。

终于飞桨可以用了：

root@863c89a419ec:~# python3
Python 3.8.2 (default, Jan 18 2024, 07:05:37) 
[GCC 9.3.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import paddle
>>> paddle.utils.run_check()
Running verify PaddlePaddle program ... 
I0228 07:37:38.344522 279426 program_interpreter.cc:220] New Executor is Running.
I0228 07:37:38.475822 279426 interpreter_util.cc:652] Standalone Executor is Used.
PaddlePaddle works well on 1 CPU.
PaddlePaddle is installed successfully! Let's start deep learning with PaddlePaddle now.

总结：

目前已编译成功。

目前import paddle成功。run_check()成功。飞桨成功编译好了，可以开始使用了！

>>> import paddle
>>> paddle.utils.run_check()
Running verify PaddlePaddle program ... 
I0228 07:37:38.344522 279426 program_interpreter.cc:220] New Executor is Running.
I0228 07:37:38.475822 279426 interpreter_util.cc:652] Standalone Executor is Used.
PaddlePaddle works well on 1 CPU.
PaddlePaddle is installed successfully! Let's start deep learning with PaddlePaddle now.
>>> x = paddle.randn((2,3))
>>> y = paddle.randn((2,3))
>>> z = x+y 
>>> z
Tensor(shape=[2, 3], dtype=float32, place=Place(cpu), stop_gradient=True,
       [[ 0.70946050, -0.07831943, -0.60504013],
        [-2.55267453, -0.59097183,  2.08291411]])

测试飞桨单机实例，pass第五十二回戴宗二取公孙胜李逵独劈罗真人-飞桨AI框架安装和使用示例-CSDN博客

调试

报错：Could NOT find PY_google.protobuf (missing: PY_GOOGLE.PROTOBUF)

Could NOT find PY_google.protobuf (missing: PY_GOOGLE.PROTOBUF)

安装protobuf

pip3 install protobuf

编译时到100%报错

[ 97%] Building CXX object gloo/CMakeFiles/gloo.dir/transport/tcp/unbound_buffer.cc.o
[100%] Linking CXX static library libgloo.a
[100%] Built target gloo
[ 3%] Performing install step for 'extern_gloo'
[ 3%] Completed 'extern_gloo'
[ 3%] Built target extern_gloo
make: *** [Makefile:136: all] Error 2
再重新编译，到了protobuf这里报错：

-- Installing: /root/github/paddle/build/third_party/install/protobuf/lib/cmake/protobuf/protobuf-config.cmake
[ 4%] Completed 'extern_protobuf'
[ 4%] Built target extern_protobuf
make: *** [Makefile:136: all] Error 2

再重新编译，发现这里报错：

报错：cc1: error: '-march=loongson3a': ISA string must begin with rv32 or rv64

cc1: error: requested ABI requires '-march' to subsume the 'D' extension
cc1: error: ABI requires '-march=rv64'
make[4]: *** [Makefile:737: isamin.o] Error 1
cc1: error: '-march=loongson3a': ISA string must begin with rv32 or rv64
make[4]: *** [Makefile:611: sasum.o] Error 1
cc1: error: '-march=loongson3a': ISA string must begin with rv32 or rv64
cc1: error: '-march=loongson3a': ISA string must begin with rv32 or rv64
cc1: error: requested ABI requires '-march' to subsume the 'D' extension
cc1: error: requested ABI requires '-march' to subsume the 'D' extension
cc1: error: ABI requires '-march=rv64'
cc1: error: ABI requires '-march=rv64'
cc1: error: requested ABI requires '-march' to subsume the 'D' extension
cc1: error: ABI requires '-march=rv64'
make[4]: *** [Makefile:647: snrm2.o] Error 1
make[4]: *** [Makefile:629: ssum.o] Error 1
make[4]: *** [Makefile:701: smax.o] Error 1
make[4]: *** [Makefile:665: samax.o] Error 1
make[4]: *** [Makefile:755: ismax.o] Error 1
make[4]: *** [Makefile:773: sdsdot.o] Error 1
make[3]: *** [Makefile:164: libs] Error 1
make[2]: *** [CMakeFiles/extern_openblas.dir/build.make:86: third_party/openblas/src/extern_openblas-stamp/extern_openblas-build] Error 2
make[1]: *** [CMakeFiles/Makefile2:4001: CMakeFiles/extern_openblas.dir/all] Error 2
[ 5%] Built target eager_python_c_codegen
[ 5%] Built target op_map_codegen
[ 5%] Built target eager_codegen
make: *** [Makefile:136: all] Error 2

参考这里重新来过：

如何在RISC-V平台编译paddle - 知乎

编译报错：

Live child 0x2af978fd70 (paddle/fluid/eager/auto_code_generator/generator/CMakeFiles/eager_codegen) PID 115530
Reaping winning child 0x2af978fd70 PID 115530
Removing child 0x2af978fd70 PID 115530 from chain.
Considering target file 'paddle/fluid/eager/auto_code_generator/generator/CMakeFiles/eager_codegen.dir/build'.
File 'paddle/fluid/eager/auto_code_generator/generator/CMakeFiles/eager_codegen.dir/build' does not exist.
Considering target file 'eager_codegen'.
   File 'eager_codegen' does not exist.
    Considering target file 'paddle/fluid/eager/auto_code_generator/generator/CMakeFiles/eager_codegen'.
    File 'paddle/fluid/eager/auto_code_generator/generator/CMakeFiles/eager_codegen' was considered already.
    Considering target file 'paddle/fluid/eager/auto_code_generator/generator/CMakeFiles/eager_codegen.dir/build.make'.
    File 'paddle/fluid/eager/auto_code_generator/generator/CMakeFiles/eager_codegen.dir/build.make' was considered already.
   Finished prerequisites of target file 'eager_codegen'.
Must remake target 'eager_codegen'.
Successfully remade target file 'eager_codegen'.
Finished prerequisites of target file 'paddle/fluid/eager/auto_code_generator/generator/CMakeFiles/eager_codegen.dir/build'.
Must remake target 'paddle/fluid/eager/auto_code_generator/generator/CMakeFiles/eager_codegen.dir/build'.
Successfully remade target file 'paddle/fluid/eager/auto_code_generator/generator/CMakeFiles/eager_codegen.dir/build'.
make[2]: Leaving directory '/root/github/paddle/build'
Reaping winning child 0x2b00eb0510 PID 115266
Live child 0x2b00eb0510 (paddle/fluid/eager/auto_code_generator/generator/CMakeFiles/eager_codegen.dir/all) PID 115532
[ 4%] Built target eager_codegen
Reaping winning child 0x2b00eb0510 PID 115532
Removing child 0x2b00eb0510 PID 115532 from chain.
make[1]: Leaving directory '/root/github/paddle/build'
Reaping losing child 0x2ad05e7c50 PID 115118
make: *** [Makefile:136: all] Error 2
Removing child 0x2ad05e7c50 PID 115118 from chain.
make: Leaving directory '/root/github/paddle/build'

在paddle/build/third_party删除eigen3目录，然后执行：git submodule update --init --recursive

问题解决。

问题没有解决，还是4%这里报错：

[ 4%] Built target eager_codegen
Reaping winning child 0x2b167efa70 PID 117670
Removing child 0x2b167efa70 PID 117670 from chain.
make[1]: Leaving directory '/root/github/paddle/build'
Reaping losing child 0x2b0332dc20 PID 117206
make: *** [Makefile:136: all] Error 2
Removing child 0x2b0332dc20 PID 117206 from chain.
make: Leaving directory '/root/github/paddle/build'

重新设置make 后，cmake ../ -DWITH_GPU=OFF -WITH_RISCV=ON 编译报错

报错c++: error: unrecognized command line option '-m64'

按照文档，应该“查找makefile文件的生成逻辑，最终发现在Paddle.cmake文件中有这样一段逻辑”，将该文件中的-64去掉。

但是没有找到这个文件。找到了，在这里：

cmake/flags.cmake

if(NOT WITH_NV_JETSON
     AND NOT WITH_ARM
     AND NOT WITH_RISCV
     AND NOT WITH_SW
     AND NOT WITH_MIPS
     AND NOT WITH_LOONGARCH)
    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -m64")
endif()

后期通过修改源码，添加RISCV选项的方法，使代码支持RISCV，并同时使用-m64失效

在cmake的时候加上DWITH_RISCV提示没有生效

WITH_DLNNE:
-- Configuring done
-- Generating done
CMake Warning:
Manually-specified variables were not used by the project:

WITH_RISCV

按照飞桨官网pr，修改飞桨make相关文件。

编译报错没有 pmmintrin.h文件

[ 7%] Building CXX object paddle/fluid/platform/CMakeFiles/denormal.dir/denormal.cc.o
/root/github/paddle/paddle/fluid/platform/denormal.cc:38:10: fatal error: pmmintrin.h: No such file or directory
38 | #include <pmmintrin.h>
| ^~~~~~~~~~~~~
compilation terminated.
make[2]: *** [paddle/fluid/platform/CMakeFiles/denormal.dir/build.make:76: paddle/fluid/platform/CMakeFiles/denormal.dir/denormal.cc.o] Error 1
make[1]: *** [CMakeFiles/Makefile2:4787: paddle/fluid/platform/CMakeFiles/denormal.dir/all] Error 2
make[1]: *** Waiting for unfinished jobs....

这个对应的就是paddle/fluid/platform/denormal.cc 这个文件，已经修改。

编译安装好后，

import paddle报错

oot@863c89a419ec:~/github/paddle/build# python3
Python 3.8.2 (default, Jan 18 2024, 07:05:37)
[GCC 9.3.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import paddle
Error: Can not import paddle core while this file exists: /usr/local/lib/python3.8/dist-packages/paddle/base/libpaddle.so
Traceback (most recent call last):
File "/usr/local/lib/python3.8/dist-packages/paddle/base/core.py", line 267, in <module>
from . import libpaddle
ImportError: /usr/local/lib/python3.8/dist-packages/paddle/base/libpaddle.so: undefined symbol: __atomic_exchange_1

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/usr/local/lib/python3.8/dist-packages/paddle/__init__.py", line 30, in <module>
    from .base import core # noqa: F401
File "/usr/local/lib/python3.8/dist-packages/paddle/base/__init__.py", line 38, in <module>
    from . import ( # noqa: F401
File "/usr/local/lib/python3.8/dist-packages/paddle/base/backward.py", line 25, in <module>
    from . import core, framework, log_helper, unique_name
File "/usr/local/lib/python3.8/dist-packages/paddle/base/core.py", line 377, in <module>
    if not avx_supported() and libpaddle.is_compiled_with_avx():
NameError: name 'libpaddle' is not defined

看文档，有人说是patchelf版本低的缘故，升级patchelf，再重新编译，问题依旧

https://github.com/PaddlePaddle/Paddle/issues/51536

已提issue ：https://github.com/PaddlePaddle/Paddle/issues/62037

根据issue里的提示，修改文件

目前还未解决报错问题

解决undefined symbol: __atomic_exchange_1报错

只需要用patchelf注册一下就行了patchelf --add-needed libatomic.so.1 /usr/local/lib/python3.8/dist-packages/paddle/base/libpaddle.so

如果没有patchelf，apt install patchelf安装即可。

查看算能云核心数

方法一：使用 lscpu 命令
方法二：使用 cat /proc/cpuinfo 命令
方法三：使用 nproc 命令

发现算能云是128核，所以修改参数-j 128

关注

11
点赞
踩
18

收藏

觉得还不错? 一键收藏
0
评论
算能RISC-V通用云编译飞桨paddlepaddle@openKylin留档

飞桨paddlepaddle在算能云riscv环境下已编译成功。目前import paddle成功。run_check()成功。飞桨成功编译好了，可以开始使用了！
复制链接

扫一扫