Minimum

对两个输入数据执行逐元素取最小值操作。对于每一对对应的输入元素,比较其大小,并将较小者存入输出地址。

\[output_i = \min(input0_i, input1_i)\]
输入:
  • input0 - 第一个输入数据地址。

  • input1 - 第二个输入数据地址。

  • param - 参数数组(7个元素),包含以下内容:
    • param[0]: input0_dims 地址(根据维度数量,分配相应空间)

    • param[1]: input1_dims 地址(根据维度数量,分配相应空间)

    • param[2]: output_dims 地址(根据维度数量,分配相应空间)

    • param[3]: strides0 地址(需分配 num_dims * sizeof(int) 空间)

    • param[4]: strides1 地址(需分配 num_dims * sizeof(int) 空间)

    • param[5]: strides_output 地址(需分配 num_dims * sizeof(int) 空间)

    • param[6]: num_dims(维度数量)

  • core_mask(int, 可选) - 核掩码(仅适用于共享存储版本)。

输出:
  • output - 计算结果地址。

支持平台:

FT78NE MT7004

备注

  • FT78NE 支持:int8 (i8), int16 (i16), int32 (i32), fp32 (fp), fp64 (dp)

  • MT7004 支持:int16 (i16), int32 (i32), fp16 (hp), fp32 (fp)

  • 两个输入数据与输出数据的数据类型必须一致。

共享存储版本:

void i8_minimum_s(int8_t *input0, int8_t *input1, int8_t *output, unsigned long long *param, int core_mask)
void i16_minimum_s(int16_t *input0, int16_t *input1, int16_t *output, unsigned long long *param, int core_mask)
void i32_minimum_s(int32_t *input0, int32_t *input1, int32_t *output, unsigned long long *param, int core_mask)
void hp_minimum_s(half *input0, half *input1, half *output, unsigned long long *param, int core_mask)
void fp_minimum_s(float *input0, float *input1, float *output, unsigned long long *param, int core_mask)
void dp_minimum_s(double *input0, double *input1, double *output, unsigned long long *param, int core_mask)

C调用示例:

 1// MT7004 示例(共享存储多核并行)
 2#include <stdio.h>
 3#include <minimum.h>
 4
 5int main(int argc, char* argv[]) {
 6    float *in0 = (float *)0x88000000;   // 输入0在共享存储空间
 7    float *in1 = (float *)0x82000000;   // 输入1在共享存储空间
 8    float *out = (float *)0x98000000;   // 输出在共享存储空间
 9
10    int length = 960001;
11    int dims[1] = {length};
12    int strides[1] = {1};
13    int num_dims = 1;
14    unsigned long long param[9];
15    param[0] = (unsigned long long)dims;
16    param[1] = (unsigned long long)dims;
17    param[2] = (unsigned long long)dims;
18    param[3] = (unsigned long long)strides;
19    param[4] = (unsigned long long)strides;
20    param[5] = (unsigned long long)strides;
21    param[6] = (unsigned long long)num_dims;
22
23    int core_mask = 0xFF;                 // 使用所有核心并行计算
24    fp_minimum_s(in0, in1, out, param, core_mask);
25    return 0;
26}

私有存储版本:

void i8_minimum_p(int8_t *input0, int8_t *input1, int8_t *output, unsigned long long *param)
void i16_minimum_p(int16_t *input0, int16_t *input1, int16_t *output, unsigned long long *param)
void i32_minimum_p(int32_t *input0, int32_t *input1, int32_t *output, unsigned long long *param)
void hp_minimum_p(half *input0, half *input1, half *output, unsigned long long *param)
void fp_minimum_p(float *input0, float *input1, float *output, unsigned long long *param)
void dp_minimum_p(double *input0, double *input1, double *output, unsigned long long *param)

C调用示例:

 1// MT7004 示例(私有存储单核)
 2#include <stdio.h>
 3#include <minimum.h>
 4
 5int main(int argc, char* argv[]) {
 6    // 输入与输出均位于芯片私有存储空间
 7    float *in0 = (float *)0x10000000;
 8    float *in1 = (float *)0x10010000;
 9    float *out = (float *)0x10020000;
10
11    int length = 1024;
12    int dims[1] = {length};
13    int strides[1] = {1};
14    int num_dims = 1;
15    unsigned long long param[9];
16    param[0] = (unsigned long long)dims;
17    param[1] = (unsigned long long)dims;
18    param[2] = (unsigned long long)dims;
19    param[3] = (unsigned long long)strides;
20    param[4] = (unsigned long long)strides;
21    param[5] = (unsigned long long)strides;
22    param[6] = (unsigned long long)num_dims;
23
24    fp_minimum_p(in0, in1, out, param);
25    return 0;
26}