mad

功能说明

矩阵乘运算，c = a * b + c。

接口原型

void mad(__cc__ float *c, __ca__ float *a, __cb__ float *b, uint16_t m, uint16_t k, uint16_t n, uint8_t unitFlag, bool kDirectionAlign, bool cmatrixSource, bool cmatrixInitVal);

void mad(__cc__ float *dst, __ca__ bfloat16_t *src0, __cb__ bfloat16_t *src1, uint16_t m, uint16_t k, uint16_t n, uint8_t featOffset, uint8_t smaskOffset, uint8_t unitFlag, bool kDirectionAlign, bool isWeightOffset, bool cmatrixSource, bool cmatrixInitVal);


void mad(__cc__ int32_t *c, __ca__ int8_t *a, __cb__ int8_t *b, uint16_t m, uint16_t k, uint16_t n, uint8_t unitFlag, bool kDirectionAlign, bool cmatrixSource, bool cmatrixInitVal);

void mad(__cc__ float *c, __ca__ bfloat16_t *a, __cb__ bfloat16_t *b, uint16_t m, uint16_t k, uint16_t n, uint8_t unitFlag, bool kDirectionAlign, bool cmatrixSource, bool cmatrixInitVal);

void mad(__cc__ float *c, __ca__ half *a, __cb__ half *b, uint16_t m, uint16_t k, uint16_t n, uint8_t unitFlag, bool kDirectionAlign, bool cmatrixSource, bool cmatrixInitVal);

void mad(__cc__ float *c, __ca__ half *a, __cb__ half *b, uint16_t m, uint16_t k, uint16_t n, bool init_val_controlC);

void mad(__cc__ int32_t *c, __ca__ int8_t *a, __cb__ int8_t *b, uint16_t m, uint16_t k, uint16_t n, bool init_val_controlC);

void mad(__cc__ int32_t *dst, __ca__ int8_t *src0, __cb__ int8_t *src1, uint16_t m, uint16_t k, uint16_t n, uint8_t featOffset, uint8_t smaskOffset, uint8_t unitFlag, bool kDirectionAlign, bool isWeightOffset, bool cmatrixSource, bool cmatrixInitVal);

void mad(__cc__ float *dst, __ca__ float *src0, __cb__ float *src1, uint16_t m, uint16_t k, uint16_t n, uint8_t featOffset, uint8_t smaskOffset, uint8_t unitFlag, bool kDirectionAlign, bool isWeightOffset, bool cmatrixSource, bool cmatrixInitVal);

void mad(__cc__ float *dst, __ca__ half *src0, __cb__ half *src1, uint16_t m, uint16_t k, uint16_t n, uint8_t featOffset, uint8_t smaskOffset, uint8_t unitFlag, bool kDirectionAlign, bool isWeightOffset, bool cmatrixSource, bool cmatrixInitVal);
 
void mad_s4(__cc__ int32_t *c, __ca__ void *a, __cb__ void *b, uint16_t m, uint16_t k, uint16_t n, bool init_val_controlC); 

void mad_s4(__cc__ int32_t *c, __ca__ void *a, __cb__ void *b, uint16_t m, uint16_t k, uint16_t n, uint8_t unitFlag, bool kDirectionAlign, bool cmatrixSource, bool cmatrixInitVal);

void mad_s4(__cc__ int32_t *dst, __ca__ void *src0, __cb__ void *src1, uint16_t m, uint16_t k, uint16_t n, uint8_t featOffset, uint8_t smaskOffset, uint8_t unitFlag, bool kDirectionAlign, bool isWeightOffset, bool cmatrixSource, bool cmatrixInitVal); 

void mad_sp(__cc__ int32_t *c, __ca__ int8_t *a, __cb__ int8_t *b, uint16_t m, uint16_t k, uint16_t n, uint8_t unitFlagMode, bool cmatrixSource, bool cmatrixInitVal);

参数说明

表1 mad参数说明
参数名	说明	取值范围	单位
c	目的地址，矩阵c地址，存放在L0C上	/	/
a	左矩阵源地址，存放在L0A上	/	/
b	右矩阵源地址，存放在L0B上	/	/
m	左矩阵的高	[0, 2^12-1]	elem
k	左矩阵的宽，右矩阵的高	[0, 2^12-1]	elem
n	右矩阵的宽	[0, 2^12-1]	elem
unitFlag	预留参数，设置为0即可	/	/
kDirectionAlign	预留参数，设置为False即可	[0, 1]	/
cmatrixSource	True表示初始c矩阵在bias table buffer上，False表示初始c矩阵在L0C上	[0, 1]	/
cmatrixInitVal	True表示c矩阵内容为0，False使用c矩阵中具体数据	[0, 1]	/
featOffset	特征图矩阵偏移。如果数据类型为s8，则偏移量被视为s8；如果数据类型为u8，则偏移量被视为u8。如果不需要偏移量，则设置0	[0, 2^8-1]	elem
smaskOffset	SMASK缓冲区地址以2字节为单位加载权重偏移量。设置为1 表示SMASK缓冲区读取地址为2B	[0, 2^7-1]	2B
isWeightOffset	权重矩阵偏移使能位	[0, 1]	elem

流水类型

PIPE_M

父主题： 矩阵计算接口