Hi Depeng

从你的知乎文章 移动端arm cpu优化学习笔记第4弹--内联汇编入门 看过来的。

我的NDK版本是r21b，代码不修改的情况下，内联汇编最后一部分有编译报错，原因是同一个变量，既作为 输入 也作为 输出。我找到的修改方法是，只在 output 部分传入，并且将=r修改为+r。

例如rgb转gray的：

        :[src]        "=r"(src),
         [dst]        "=r"(dst),
         [neonLen]    "=r"(neonLen)
        :[src]        "0"(src),
         [dst]        "1"(dst),
         [neonLen]    "2"(neonLen)
        :"cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13"

改为如下：

        :[src]        "+r"(src),
         [dst]        "+r"(dst),
         [neonLen]    "+r"(neonLen)
        :
        :"cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13"

My run produces (left: input; middle: my implementation; right: this implementation; radius=10):

Can you try this input? It's a 201x201 FP32 square, stored with little-endian:
noisy-square.tar.gz

I use this routine to read the file:

template<typename T>
std::vector<T> read_raw(const std::string &path, bool little_endian = false)
{
    std::ifstream file(path, std::ifstream::binary);
    std::streamsize buffer_size = 1 << 20;
    char buffer[buffer_size];
    std::vector<T> raw;
    while(!file.eof())
    {
        file.read(buffer, buffer_size);
        for (int i = 0; i < std::min(buffer_size, file.gcount()); i += sizeof(T))
        {
            char *begin = &buffer[i];
            if (little_endian)
                std::reverse(begin, begin + sizeof(T));
            raw.push_back(*(T *)begin);
        }
    }
    file.close();
    return raw;
}

Thank you.

请问大佬， 有办法对进行arm neon代码进行单步调试吗？或者开发的时候 如何去debug，arm neon汇编或者intrinsic

CMake Error at CMakeLists.txt:35 (add_subdirectory):
add_subdirectory given source "third_party/googletest/googletest" which is
not an existing directory.

Radius 5
In 3.4Ghz Cpu
i use avx2 to improve add histogram and sub histogram
but the time speed greater than 200ms
image size 1280*1024

求问为啥后面的neon指令看上去并没有加速很多？这里面的原因是为啥？