C-API-Reg矢量计算接口理论性能汇总

本节汇总介绍了C-API Reg矢量计算接口的理论性能，本节内容仅针对如下型号生效：

Ascend 950PR/Ascend 950DT

本节汇总的理论性能均为单指令理论性能，软仿实现的接口和数据类型不提供理论性能。

表1 基础算术

接口	源操作数数据类型	目的操作数数据类型	理论并行度（element/cycle）
asc_abs	int8_t	int8_t	128
asc_abs	int16_t	int16_t	64
asc_abs	half	half	128
asc_abs	int32_t	int32_t	32
asc_abs	float	float	64
asc_add_scalar	int8_t	int8_t	128
asc_add_scalar	uint8_t	uint8_t	128
asc_add_scalar	int16_t	int16_t	64
asc_add_scalar	uint16_t	uint16_t	64
asc_add_scalar	half	half	64
asc_add_scalar	bfloat16_t	bfloat16_t	64
asc_add_scalar	int32_t	int32_t	32
asc_add_scalar	uint32_t	uint32_t	32
asc_add_scalar	float	float	32
asc_ln	half	half	16
asc_ln	float	float	16
asc_max_scalar	int8_t	int8_t	256
asc_max_scalar	uint8_t	uint8_t	256
asc_max_scalar	int16_t	int16_t	128
asc_max_scalar	uint16_t	uint16_t	128
asc_max_scalar	half	half	128
asc_max_scalar	bfloat16_t	bfloat16_t	128
asc_max_scalar	int32_t	int32_t	64
asc_max_scalar	uint32_t	uint32_t	64
asc_max_scalar	float	float	64
asc_min_scalar	int8_t	int8_t	256
asc_min_scalar	uint8_t	uint8_t	256
asc_min_scalar	int16_t	int16_t	128
asc_min_scalar	uint16_t	uint16_t	128
asc_min_scalar	half	half	128
asc_min_scalar	int32_t	int32_t	64
asc_min_scalar	uint32_t	uint32_t	64
asc_min_scalar	float	float	64
asc_mul_scalar	int16_t	int16_t	64
asc_mul_scalar	uint16_t	uint16_t	64
asc_mul_scalar	half	half	128
asc_mul_scalar	int32_t	int32_t	32
asc_mul_scalar	uint32_t	uint32_t	32
asc_mul_scalar	float	float	64
asc_neg	int8_t	int8_t	128
asc_neg	int16_t	int16_t	64
asc_neg	half	half	128
asc_neg	int32_t	int32_t	32
asc_neg	float	float	64
asc_relu	half	half	128
asc_relu	int32_t	int32_t	64
asc_relu	float	float	64

表2 逻辑算术

接口	源操作数数据类型	目的操作数数据类型	理论并行度（element/cycle）
asc_and	bool	bool	128
asc_and	int8_t	int8_t	256
asc_and	uint8_t	uint8_t	256
asc_and	fp8_e5m2_t	fp8_e5m2_t	256
asc_and	fp8_e4m3fn_t	fp8_e4m3fn_t	256
asc_and	int16_t	int16_t	128
asc_and	uint16_t	uint16_t	128
asc_and	half	half	128
asc_and	bfloat16_t	bfloat16_t	128
asc_and	int32_t	int32_t	64
asc_and	uint32_t	uint32_t	64
asc_and	float	float	64
asc_or	bool	bool	128
asc_or	int8_t	int8_t	256
asc_or	uint8_t	uint8_t	256
asc_or	int16_t	int16_t	128
asc_or	uint16_t	uint16_t	128
asc_or	half	half	128
asc_or	bfloat16_t	bfloat16_t	128
asc_or	int32_t	int32_t	64
asc_or	uint32_t	uint32_t	64
asc_or	float	float	64

表3 复合计算

接口	源操作数数据类型	目的操作数数据类型	理论并行度（element/cycle）
asc_abs_sub	half	half	128
asc_abs_sub	float	float	64
asc_axpy	half	half	128
asc_axpy	float	float	64

表4 比较与选择

接口	源操作数数据类型	目的操作数数据类型	理论并行度（element/cycle）
asc_eq	int8_t	bool	256
asc_eq	uint8_t	bool	256
asc_eq	int16_t	bool	128
asc_eq	uint16_t	bool	128
asc_eq	half	bool	128
asc_eq	bfloat16_t	bool	128
asc_eq	int32_t	bool	64
asc_eq	uint32_t	bool	64
asc_eq	float	bool	64
asc_ge	int8_t	bool	256
asc_ge	uint8_t	bool	256
asc_ge	int16_t	bool	128
asc_ge	uint16_t	bool	128
asc_ge	half	bool	128
asc_ge	bfloat16_t	bool	128
asc_ge	int32_t	bool	64
asc_ge	uint32_t	bool	64
asc_ge	float	bool	64
asc_ge_scalar	int8_t	bool	256
asc_ge_scalar	uint8_t	bool	256
asc_ge_scalar	int16_t	bool	128
asc_ge_scalar	uint16_t	bool	128
asc_ge_scalar	half	bool	128
asc_ge_scalar	bfloat16_t	bool	128
asc_ge_scalar	int32_t	bool	64
asc_ge_scalar	uint32_t	bool	64
asc_ge_scalar	float	bool	64
asc_gt	int8_t	bool	256
asc_gt	uint8_t	bool	256
asc_gt	int16_t	bool	128
asc_gt	uint16_t	bool	128
asc_gt	half	bool	128
asc_gt	bfloat16_t	bool	128
asc_gt	int32_t	bool	64
asc_gt	uint32_t	bool	64
asc_gt	float	bool	64
asc_gt_scalar	int8_t	bool	256
asc_gt_scalar	uint8_t	bool	256
asc_gt_scalar	int16_t	bool	128
asc_gt_scalar	uint16_t	bool	128
asc_gt_scalar	half	bool	128
asc_gt_scalar	bfloat16_t	bool	128
asc_gt_scalar	int32_t	bool	64
asc_gt_scalar	uint32_t	bool	64
asc_gt_scalar	float	bool	64
asc_lt	int8_t	bool	256
asc_lt	uint8_t	bool	256
asc_lt	int16_t	bool	128
asc_lt	uint16_t	bool	128
asc_lt	half	bool	128
asc_lt	bfloat16_t	bool	128
asc_lt	int32_t	bool	64
asc_lt	uint32_t	bool	64
asc_lt	float	bool	64
asc_ne	int8_t	bool	256
asc_ne	uint8_t	bool	256
asc_ne	int16_t	bool	128
asc_ne	uint16_t	bool	128
asc_ne	half	bool	128
asc_ne	bfloat16_t	bool	128
asc_ne	int32_t	bool	64
asc_ne	uint32_t	bool	64
asc_ne	float	bool	64
asc_ne_scalar	int8_t	bool	256
asc_ne_scalar	uint8_t	bool	256
asc_ne_scalar	int16_t	bool	128
asc_ne_scalar	uint16_t	bool	128
asc_ne_scalar	half	bool	128
asc_ne_scalar	bfloat16_t	bool	128
asc_ne_scalar	int32_t	bool	64
asc_ne_scalar	uint32_t	bool	64
asc_ne_scalar	float	bool	64
asc_select	bool	bool	256
asc_select	int8_t	int8_t	256
asc_select	uint8_t	uint8_t	256
asc_select	hifloat8_t	hifloat8_t	256
asc_select	fp8_e5m2_t	fp8_e5m2_t	256
asc_select	fp8_e4m3fn_t	fp8_e4m3fn_t	256
asc_select	int16_t	int16_t	128
asc_select	uint16_t	uint16_t	128
asc_select	half	half	128
asc_select	bfloat16_t	bfloat16_t	128
asc_select	int32_t	int32_t	64
asc_select	uint32_t	uint32_t	64
asc_select	float	float	64

表5 类型转换

接口	源操作数数据类型	目的操作数数据类型	理论并行度（element/cycle）
asc_int322float	int32_t	float	32
asc_int322int16	int32_t	int16_t	64
asc_int322int64	int32_t	int64_t	32
asc_int322uint16	int32_t	uint16_t	64

表6 归约计算

接口	源操作数数据类型	目的操作数数据类型	理论并行度（element/cycle）
asc_reduce_max	int16_t	int16_t	128
asc_reduce_max	uint16_t	uint16_t	128
asc_reduce_max	half	half	128
asc_reduce_max	int32_t	int32_t	64
asc_reduce_max	uint32_t	uint32_t	64
asc_reduce_max	float	float	64
asc_reduce_max_datablock	int16_t	int16_t	128
asc_reduce_max_datablock	uint16_t	uint16_t	128
asc_reduce_max_datablock	half	half	128
asc_reduce_max_datablock	int32_t	int32_t	64
asc_reduce_max_datablock	uint32_t	uint32_t	64
asc_reduce_max_datablock	float	float	64
asc_reduce_min	int16_t	int16_t	128
asc_reduce_min	uint16_t	uint16_t	128
asc_reduce_min	half	half	128
asc_reduce_min	int32_t	int32_t	64
asc_reduce_min	uint32_t	uint32_t	64
asc_reduce_min	float	float	64
asc_reduce_min_datablock	int16_t	int16_t	128
asc_reduce_min_datablock	uint16_t	uint16_t	128
asc_reduce_min_datablock	half	half	128
asc_reduce_min_datablock	int32_t	int32_t	64
asc_reduce_min_datablock	uint32_t	uint32_t	64
asc_reduce_min_datablock	float	float	64
asc_reduce_sum	int16_t	int32_t	64
asc_reduce_sum	uint16_t	uint32_t	64
asc_reduce_sum	half	half	128
asc_reduce_sum	int32_t	int32_t	64
asc_reduce_sum	uint32_t	uint32_t	64
asc_reduce_sum	float	float	64
asc_reduce_sum_datablock	int16_t	int32_t	64
asc_reduce_sum_datablock	uint16_t	uint32_t	64
asc_reduce_sum_datablock	half	half	128
asc_reduce_sum_datablock	int32_t	int32_t	64
asc_reduce_sum_datablock	uint32_t	uint32_t	64
asc_reduce_sum_datablock	float	float	64

表7 直方图计算

接口	源操作数数据类型	目的操作数数据类型	理论并行度（element/cycle）
asc_cumulative_histogram_bin0	uint8_t	uint16_t	64
asc_cumulative_histogram_bin1	uint8_t	uint16_t	64