鲲鹏平台，arrch64操作系统，ncnn推理结果不正确

[wangyangneu]

前馈输出与x86的结果对不上，请问有什么解决办法么 @nihui

[nihui]

ctest 能 100% 跑完嘛？

[wangyangneu]

ctest 能 100% 跑完嘛？

ctest有三个failed

[nihui]

ctest --output-on-failure 输出下具体错误

[wangyangneu]

ctest --output-on-failure 输出下具体错误

  Start  5: test_squeezenet

5/66 Test #5: test_squeezenet ..................***Failed 0.19 sec
top 0 index not match expect 532 but got 791
test_squeezenet cpu failed use_packing_layout=0 use_fp16_packed=0 use_fp16_storage=0 use_shader_pack8=0 use_bf16_storage=0 use_image_storage=0
CMake Error at /root/work/code/wangyang/ncnn/cmake/run_test.cmake:4 (message):
Test failed with return value '255'

  Start 14: test_convolution

14/66 Test #14: test_convolution .................***Failed 3.26 sec
value not match at c:0 h:0 w:0 expect 3.338143 but got 5.268054
test_layer_cpu failed
test_layer Convolution failed use_packing_layout=0 use_fp16_packed=0 use_fp16_storage=0 use_fp16_arithmetic=0 use_shader_pack8=0 use_bf16_storage=0 use_image_storage=0
test_convolution failed w=9 h=7 c=16 outch=16 kernel=3 dilation=1 stride=1 pad=1 bias=0 act=5 actparams=[-0.013927,0.071340]
CMake Error at /root/work/code/wangyang/ncnn/cmake/run_test.cmake:4 (message):
Test failed with return value '1'

  Start 31: test_innerproduct

31/66 Test #31: test_innerproduct ................***Failed 0.16 sec
CMake Error at /root/work/code/wangyang/ncnn/cmake/run_test.cmake:4 (message):
Test failed with return value 'Segmentation fault'

[tpoisonooo]

复现了。有一点历史渊源...华为 arm 那套服务器，默认用自己的欧拉。
机器上默认的 g++ 编译器应该是这个：

$ /usr/bin/c++ --version
c++ (GCC) 4.8.5 20150623 (Red Hat 4.8.5-44)
Copyright (C) 2015 Free Software Foundation, Inc. 

这个编译器里除了__aarch64__以外，没有__ARM_NEON及其他可能的所有 neon 条件编译。也就是说下面的代码会很奇怪地正常编译过：

#if defined(__ARM_NEON__)
#error "__ARM_NEON__"
#elif defined(_M_ARM)
#error "_M_ARM"
#elif defined(_M_ARM64)
#error "_M_ARM64"
#elif defined(_M_HYBRID_X86_ARM64)
#error "hybrid"
#elif defined(__aarch64__)
// #error "__aarch64__"
#endif

这会导致一些 __ARM_NEON 包的代码没有执行，例如这句

#if __ARM_NEON
    if (opt.use_packing_layout || opt.use_int8_inference)
    {
        flatten = ncnn::create_layer(ncnn::LayerType::Flatten);

        ncnn::ParamDict pd;

        flatten->load_param(pd);

        flatten->create_pipeline(opt);
    }
#endif // __ARM_NEON

flatten handle 是 0x0，触发后面的 crash。因此最简单的修复方法是：

安装 g++-7

以 centos 为例（我印象里华为 stable 的 OS 只有两个版本，也没啥好选的）

$ yum install -y centos-release-scl
$ yum install -y devtoolset-7-gcc devtoolset-7-gcc-c++
$ source /opt/rh/devtoolset-7/enable 

改掉默认的 CMAKE_CXX_COMPILER

在 toolchain 里面抄一个新的 host-c.gcc.toolchain.cmake，内容就三行

$ cat ../toolchains/host-c.gcc.toolchain.cmake 
SET ( CMAKE_C_COMPILER "gcc" )
SET ( CMAKE_CXX_COMPILER "g++" )
SET ( CMAKE_CXX_FLAGS "-std=c++11 ${CMAKE_CXX_FLAGS}" )

创建个新的 build 目录，编译运行

$ mkdir build && build && cmake3 -DCMAKE_TOOLCHAIN_FILE=../toolchains/host-c.gcc.toolchain.cmake  ..
$ make -j
$ ./tests/test_innerproduct

我这边就正常了。我是幸福，祝福你狗子...

[nihui]

应该修好啦，有问题可以 reopen issue