Argmin
=================


沿指定轴查找最小 `topk` 个值的索引。当 `topk=1` 时，该算子等价于 `ArgMin`。

.. math::

    Y_i = \underset{k}{\operatorname{argmin}} (X_{slice_i})

其中 :math:`X_{slice_i}` 是输入张量中沿指定轴的一个切片，函数返回该切片中最小值的索引 :math:`k`。

输入：
    - **input** - 输入数据地址。
    - **params** - 其他参数打包成数组。
    - **core_mask** - 核掩码。

输出：
    - **output** - 存储索引的输出张量。
    - **output_value** - 如果 `return_values` 为 true，则此处存储找到的值。

支持平台：
    ``FT78NE``
    ``MT7004``

.. note::
    - FT78NE 支持fp32
    - MT7004 支持fp16, fp32


**参数数组结构：**

.. code-block:: c
    :linenos:

    long long params[12];
    params[0] = (long long)in_shape;输入张量的维度信息数组。
    params[1] = (long long)in_strides;输入张量的步长信息数组。
    params[2] = (long long)out_strides;输出张量的步长信息数组。
    params[3] = (long long)arg_elements;用于存放候选值的临时工作空间地址。
    params[4] = (long long)index;用于存放候选索引的临时工作空间地址。
    params[5] = (long long)topk;需要查找的最小值的数量。
    params[6] = (long long)out_value;是否返回数值的标志。若为非0，则 `output_value` 必须提供有效地址。
    params[7] = (long long)in_shape_size;输入张量的维度数 (即 `in_shape` 数组的长度)。
    params[8] = (long long)axis;执行查找操作的轴。


**共享存储版本:**

.. c:function:: void fp_arg_min_s(float* input, void* output, float* output_value, long long* params, int core_mask)
.. c:function:: void hp_arg_min_s(half* input, void* output, half* output_value, long long* params, int core_mask)

   
**C调用示例：**

.. code-block:: c
    :linenos:
    :emphasize-lines: 21-22

    //FT78NE示例
    #include <stdio.h>
    #include <argmin.h>
    int main(int argc, char* argv[]) {
         float* input = (float*)0x81000000; //需要初始化
        void* output = (void*)0x82000000; //不需要初始化
        float* output_value = (float*)0x83000000; //可选
        float* arg_elements = (float*)0x84000000;//不需要初始化
        int* index = (int*)0x85000000; //不需要初始化

        int core_mask = 0b1111;

        int *in_strides = (int*)0x86000000;
        int *out_strides = (int*)0x86000200;

        int in_shape_size = 4;//最多只考虑4维
        int in_shape[4] = {4, 8, 16, 8};
        
        int axis = 1; //要操作的维度,不能超过3
        int topk = 3; //不超过in_shape[axis]
    
        int out_value = 0; //是否输出值，1表示输出值，0表示输出索引
        float *outputfp32 = (float *)output;
        int *outputint = (int *)output;
        
        srand(time(0));
        // 初始化测试数据，包含各种情况
        int i, j;
        //tensor 1 int32
        int in_total_elements = in_shape[0] * in_shape[1] * in_shape[2] * in_shape[3];
        for(i = 0; i < in_total_elements; i ++) {
            input[i] = (float)(rand()%100);
        }
        
        long long params[9];
        params[0] = (long long)in_shape;
        params[1] = (long long)in_strides;
        params[2] = (long long)out_strides;
        params[3] = (long long)arg_elements;
        params[4] = (long long)index;
        params[5] = (long long)topk;
        params[6] = (long long)out_value;
        params[7] = (long long)in_shape_size;
        params[8] = (long long)axis;
        
        fp_arg_min_s(input, output, output_value, params, core_mask);
        return 0;
    }


**私有存储版本:**

.. c:function:: void fp_arg_min_p(float *input, void *output, float *output_value, long long *params)
.. c:function:: void hp_arg_min_p(half *input, void *output, half *output_value, long long *params)

   
**C调用示例：**

.. code-block:: c
    :linenos:
    :emphasize-lines: 20-21

    //FT78NE示例
    #include <stdio.h>
    #include <argmin.h>
    int main(int argc, char* argv[]) {
        float* input = (float*)0x10010000; //需要初始化
        void* output = (void*)0x10020000; //不需要初始化
        float* output_value = (float*)0x10030000; //可选
        float* arg_elements = (float*)0x10040000;//不需要初始化
        int* index = (int*)0x10050000; //不需要初始化

        int *in_strides = (int*)0x1004E000;
        int *out_strides = (int*)0x1004E200;

        int in_shape_size = 4;//最多只考虑4维
        int in_shape[4] = {4, 8, 16, 8};
        
        int axis = 1; //要操作的维度,不能超过3
        int topk = 3; //不超过in_shape[axis]
    
        int out_value = 0; //是否输出值，1表示输出值，0表示输出索引
        float *outputfp32 = (float *)output;
        int *outputint = (int *)output;
        
        srand(time(0));
        // 初始化测试数据，包含各种情况
        int i, j;
        //tensor 1 int32
        int in_total_elements = in_shape[0] * in_shape[1] * in_shape[2] * in_shape[3];
        for(i = 0; i < in_total_elements; i ++) {
            input[i] = (float)(rand()%100);
        }
        
        long long params[9];
        params[0] = (long long)in_shape;
        params[1] = (long long)in_strides;
        params[2] = (long long)out_strides;
        params[3] = (long long)arg_elements;
        params[4] = (long long)index;
        params[5] = (long long)topk;
        params[6] = (long long)out_value;
        params[7] = (long long)in_shape_size;
        params[8] = (long long)axis;
        
        fp_arg_min_p(input, output, output_value, params);
        return 0;
    }