cast的解读

cast本意是映射，学过c++的应该都知道各种的cast，其实就是数据类型转换。

我们还是直接上代码，构造函数：

Cast::Cast()
{
    one_blob_only = true;
    support_inplace = false;
    support_packing = true;
}

这次发现一个之前没遇到的support_packing，按照常识来理解应该是将数据打包后进行计算，这一点我们具体看代码再来验证。

• 单输入单输出
• 不支持就地运算
• 支持packing

再来看参数加载函数：

int Cast::load_param(const ParamDict& pd)
{
    type_from = pd.get(0, 0);
    type_to = pd.get(1, 0);

    return 0;
}

两个参数：

• type_from：原始类型
• type_to：新类型

up主在头文件里面给出了注释：

// element type
// 0 = auto
// 1 = float32
// 2 = float16
// 3 = int8
// 4 = bfloat16

推理函数，还是分析过程写在注释里：

int Cast::forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const
{
    //如果原始类型等于新类型则直接赋值
    if (type_from == type_to)
    {
        top_blob = bottom_blob;
        return 0;
    }

    int w = bottom_blob.w;
    int h = bottom_blob.h;
    int channels = bottom_blob.c;
    int dims = bottom_blob.dims;
    size_t elemsize = bottom_blob.elemsize;
    int elempack = bottom_blob.elempack;

    size_t out_elemsize = elemsize;
    //根据新类型来确定输出矩阵元素的size,下面程序显然是以字节为单位，字节数乘打包字节个数就等于输出矩阵元素的size
    if (type_to == 1)
    {
        // float32
        out_elemsize = 4 * elempack;
    }
    else if (type_to == 2)
    {
        // float16
        out_elemsize = 2 * elempack;
    }
    else if (type_to == 3)
    {
        // int8
        out_elemsize = elempack;
    }
    else if (type_to == 4)
    {
        // bfloat16
        out_elemsize = 2 * elempack;
    }
    //为输出矩阵分配内存空间
    if (dims == 1)
    {
        top_blob.create(w, out_elemsize, elempack, opt.blob_allocator);
    }
    else if (dims == 2)
    {
        top_blob.create(w, h, out_elemsize, elempack, opt.blob_allocator);
    }
    else if (dims == 3)
    {
        top_blob.create(w, h, channels, out_elemsize, elempack, opt.blob_allocator);
    }
    if (top_blob.empty())
        return -100;

    int size = w * h * elempack;
    //float32 -> float16
    if (type_from == 1 && type_to == 2)
    {
        //openmp指令，用于多线程
        #pragma omp parallel for num_threads(opt.num_threads)
        for (int q = 0; q < channels; q++)
        {
            const float* ptr = bottom_blob.channel(q);
            unsigned short* outptr = top_blob.channel(q);

            for (int i = 0; i < size; i++)
            {
                outptr[i] = float32_to_float16(ptr[i]);
            }
        }
    }
    //float16 -> float32
    if (type_from == 2 && type_to == 1)
    {
        //openmp指令，用于多线程
        #pragma omp parallel for num_threads(opt.num_threads)
        for (int q = 0; q < channels; q++)
        {
            const unsigned short* ptr = bottom_blob.channel(q);
            float* outptr = top_blob.channel(q);

            for (int i = 0; i < size; i++)
            {
                outptr[i] = float16_to_float32(ptr[i]);
            }
        }
    }
    //int8 -> float32
    if (type_from == 3 && type_to == 1)
    {
        //openmp指令，用于多线程
        #pragma omp parallel for num_threads(opt.num_threads)
        for (int q = 0; q < channels; q++)
        {
            const signed char* ptr = bottom_blob.channel(q);
            float* outptr = top_blob.channel(q);

            for (int i = 0; i < size; i++)
            {
                outptr[i] = (float)ptr[i];
            }
        }
    }
    //float32 -> bfloat16
    if (type_from == 1 && type_to == 4)
    {
        //openmp指令，用于多线程
        #pragma omp parallel for num_threads(opt.num_threads)
        for (int q = 0; q < channels; q++)
        {
            const float* ptr = bottom_blob.channel(q);
            unsigned short* outptr = top_blob.channel(q);

            for (int i = 0; i < size; i++)
            {
                outptr[i] = float32_to_bfloat16(ptr[i]);
            }
        }
    }
    //bfloat16 -> float32
    if (type_from == 4 && type_to == 1)
    {
        //openmp指令，用于多线程
        #pragma omp parallel for num_threads(opt.num_threads)
        for (int q = 0; q < channels; q++)
        {
            const unsigned short* ptr = bottom_blob.channel(q);
            float* outptr = top_blob.channel(q);

            for (int i = 0; i < size; i++)
            {
                outptr[i] = bfloat16_to_float32(ptr[i]);
            }
        }
    }

    // TODO more cast type

    return 0;
}

总体来说cast运算还是很简单的，就是个类型转换。当然这里的类型转换的实现细节并没有说，也不是这个系列文章的解读重点。

下面是pr内容：

cast

y = cast(x)

• one_blob_only
• support_packing

Element type:

• 0 = auto
• 1 = float32
• 2 = float16
• 3 = int8
• 4 = bfloat16