Jetson Nano安装MNN深度学习框架

sugardisk

已于 2024-03-18 18:00:50 修改

阅读量187

点赞数 2

文章标签： mnn 人工智能深度学习 Orin

于 2023-11-06 16:47:58 首次发布

本文链接：https://blog.csdn.net/weixin_52303102/article/details/134249742

版权

文章讲述了在NVIDIAJetsonNano设备上安装阿里巴巴MNN深度学习框架时遇到的编译错误，涉及`Vec.hpp`文件中的类型转换问题，并提供了可能的解决方法，如检查并启用`-flax-vector-conversions`编译选项。

摘要由CSDN通过智能技术生成

Jetson Orin Nano安装MNN深度学习框架

安装阿里巴巴的MNN框架

请参阅这里的文档: 在Jetson Nano上安装MNN深度学习框架

MNN 编译错误

1. question

In file included from /root/mnn-llm/MNN/source/backend/cpu/CPUBinary.cpp:18:
/root/mnn-llm/MNN/source/math/Vec.hpp: In member function ‘MNN::Math::Vec<float, 4>::VecType MNN::Math::Vec<float, 4>::operator<(const VecType&) const’:
/root/mnn-llm/MNN/source/math/Vec.hpp:508:51: error: could not convert ‘{vbslq_s32(res, ((int32x4_t)((const MNN::Math::Vec<float, 4>*)this)->MNN::Math::Vec<float, 4>::one), ((int32x4_t)((const MNN::Math::Vec<float, 4>*)this)->MNN::Math::Vec<float, 4>::zero))}’ from ‘<brace-enclosed initializer list>’ to ‘MNN::Math::Vec<float, 4>::VecType’ {aka ‘MNN::Math::Vec<float, 4>’}
  508 |         VecType dst = { vbslq_s32(res, one, zero) };
      |                                                   ^
      |                                                   |
      |                                                   <brace-enclosed initializer list>
/root/mnn-llm/MNN/source/math/Vec.hpp: In member function ‘MNN::Math::Vec<float, 4>::VecType MNN::Math::Vec<float, 4>::operator>(const VecType&) const’:
/root/mnn-llm/MNN/source/math/Vec.hpp:513:51: error: could not convert ‘{vbslq_s32(res, ((int32x4_t)((const MNN::Math::Vec<float, 4>*)this)->MNN::Math::Vec<float, 4>::one), ((int32x4_t)((const MNN::Math::Vec<float, 4>*)this)->MNN::Math::Vec<float, 4>::zero))}’ from ‘<brace-enclosed initializer list>’ to ‘MNN::Math::Vec<float, 4>::VecType’ {aka ‘MNN::Math::Vec<float, 4>’}
  513 |         VecType dst = { vbslq_s32(res, one, zero) };
      |                                                   ^
      |                                                   |
      |                                                   <brace-enclosed initializer list>
/root/mnn-llm/MNN/source/math/Vec.hpp: In member function ‘MNN::Math::Vec<float, 4>::VecType MNN::Math::Vec<float, 4>::operator<=(const VecType&) const’:
/root/mnn-llm/MNN/source/math/Vec.hpp:518:51: error: could not convert ‘{vbslq_s32(res, ((int32x4_t)((const MNN::Math::Vec<float, 4>*)this)->MNN::Math::Vec<float, 4>::one), ((int32x4_t)((const MNN::Math::Vec<float, 4>*)this)->MNN::Math::Vec<float, 4>::zero))}’ from ‘<brace-enclosed initializer list>’ to ‘MNN::Math::Vec<float, 4>::VecType’ {aka ‘MNN::Math::Vec<float, 4>’}
  518 |         VecType dst = { vbslq_s32(res, one, zero) };
      |                                                   ^
      |                                                   |
      |                                                   <brace-enclosed initializer list>
/root/mnn-llm/MNN/source/math/Vec.hpp: In member function ‘MNN::Math::Vec<float, 4>::VecType MNN::Math::Vec<float, 4>::operator>=(const VecType&) const’:
/root/mnn-llm/MNN/source/math/Vec.hpp:523:51: error: could not convert ‘{vbslq_s32(res, ((int32x4_t)((const MNN::Math::Vec<float, 4>*)this)->MNN::Math::Vec<float, 4>::one), ((int32x4_t)((const MNN::Math::Vec<float, 4>*)this)->MNN::Math::Vec<float, 4>::zero))}’ from ‘<brace-enclosed initializer list>’ to ‘MNN::Math::Vec<float, 4>::VecType’ {aka ‘MNN::Math::Vec<float, 4>’}
  523 |         VecType dst = { vbslq_s32(res, one, zero) };
      |                                                   ^
      |                                                   |
      |                                                   <brace-enclosed initializer list>
/root/mnn-llm/MNN/source/math/Vec.hpp: In member function ‘MNN::Math::Vec<float, 4>::VecType MNN::Math::Vec<float, 4>::operator==(const VecType&) const’:
/root/mnn-llm/MNN/source/math/Vec.hpp:528:51: error: could not convert ‘{vbslq_s32(res, ((int32x4_t)((const MNN::Math::Vec<float, 4>*)this)->MNN::Math::Vec<float, 4>::one), ((int32x4_t)((const MNN::Math::Vec<float, 4>*)this)->MNN::Math::Vec<float, 4>::zero))}’ from ‘<brace-enclosed initializer list>’ to ‘MNN::Math::Vec<float, 4>::VecType’ {aka ‘MNN::Math::Vec<float, 4>’}
  528 |         VecType dst = { vbslq_s32(res, one, zero) };
      |                                                   ^
      |                                                   |
      |                                                   <brace-enclosed initializer list>

参阅这里: NVIDIA Jetson编译报错
安装 MNN 2.7.1 编译不会报错
版本号 2.7.1

2. question

In file included from /home/orangepi/MNN/source/backend/cpu/CPUBinary.cpp:18:
/home/orangepi/MNN/source/math/Vec.hpp: In static member function ‘static void MNN::Math::Vec<int, 4>::transpose4(MNN::Math::Vec<int, 4>::VecType&, MNN::Math::Vec<int, 4>::VecType&, MNN::Math::Vec<int, 4>::VecType&, MNN::Math::Vec<int, 4>::VecType&)’:
/home/orangepi/MNN/source/math/Vec.hpp:300:66: note: use ‘-flax-vector-conversions’ to permit conversions between vectors with differing element types or numbers of subparts
300 | vec0.value = vsetq_lane_s64(vgetq_lane_s64(m2m3.val[0], 0), vec0.value, 1);
| ^
/home/orangepi/MNN/source/math/Vec.hpp:300:62: error: cannot convert ‘int32x4_t’ {aka ‘__vector(4) int’} to ‘int64x2_t’ {aka ‘__vector(2) long long int’}
300 | vec0.value = vsetq_lane_s64(vgetq_lane_s64(m2m3.val[0], 0), vec0.value, 1);
| ~~~~~~~~~~^
| |
| int32x4_t {aka __vector(4) int}
In file included from /home/orangepi/MNN/source/math/Vec.hpp:16,
from /home/orangepi/MNN/source/backend/cpu/CPUBinary.cpp:18:
/usr/lib/gcc/arm-linux-gnueabihf/9/include/arm_neon.h:6307:27: note: initializing argument 1 of ‘int64_t vgetq_lane_s64(int64x2_t, int)’
6307 | vgetq_lane_s64 (int64x2_t __a, const int __b)
| ~~~~~~~~~~^~~
In file included from /home/orangepi/MNN/source/backend/cpu/CPUBinary.cpp:18:
/home/orangepi/MNN/source/math/Vec.hpp:301:62: error: cannot convert ‘int32x4_t’ {aka ‘__vector(4) int’} to ‘int64x2_t’ {aka ‘__vector(2) long long int’}
301 | vec1.value = vsetq_lane_s64(vgetq_lane_s64(m2m3.val[1], 0), vec1.value, 1);
| ~~~~~~~~~~^
| |
| int32x4_t {aka __vector(4) int}
In file included from /home/orangepi/MNN/source/math/Vec.hpp:16,
from /home/orangepi/MNN/source/backend/cpu/CPUBinary.cpp:18:
/usr/lib/gcc/arm-linux-gnueabihf/9/include/arm_neon.h:6307:27: note: initializing argument 1 of ‘int64_t vgetq_lane_s64(int64x2_t, int)’
6307 | vgetq_lane_s64 (int64x2_t __a, const int __b)
| ~~~~~~~~~~^~~
In file included from /home/orangepi/MNN/source/backend/cpu/CPUBinary.cpp:18:
/home/orangepi/MNN/source/math/Vec.hpp:302:62: error: cannot convert ‘int32x4_t’ {aka ‘__vector(4) int’} to ‘int64x2_t’ {aka ‘__vector(2) long long int’}
302 | vec2.value = vsetq_lane_s64(vgetq_lane_s64(m0m1.val[0], 1), vec2.value, 0);
| ~~~~~~~~~~^
| |
| int32x4_t {aka __vector(4) int}