CANN/Ascend C SIMD数据搬运API

asc_loadalign【免费下载链接】asc-devkit本项目是CANN 推出的昇腾AI处理器专用的算子程序开发语言原生支持C和C标准规范主要由类库和语言扩展层构成提供多层级API满足多维场景算子开发诉求。项目地址: https://gitcode.com/cann/asc-devkit产品支持情况产品是否支持Ascend 950PR/Ascend 950DT√功能说明对齐数据搬运接口从UB连续对齐搬入目的操作数支持多种搬入模式接口通过int32_t传入偏移用户可以选择更新偏移或者更新源操作数的地址。NORM搬入模式正常模式搬运VL数据。BRC搬入模式搬运一个b8/b16/b32类型的数据并Broadcast到所有元素位置。UP2X搬入模式数据2倍上采样加载VL/2个数据每个输入元素重复两次数据类型为b8/b16。DOWN2X搬入模式数据2倍下采样加载2倍VL的数据数据每隔一个保留数据类型为b8/b16。UNPACK搬入模式解压缩模式按无符号整型u8/u16/u32加载VL/2长度数据unpack到VL长度u16/u32/u64类型中间位置补0。UNPACKV2搬入模式解压缩模式按无符号整型u8加载VL/4长度数据unpack到VL长度u32类型中间位置补0。BRCV2搬入模式读取一个DataBlock32B并广播到VL。BRCV3搬入模式加载VL/DataBlockB的数据并将每个元素16bit/32bit广播到一个DataBlock32B中。DEINTLV搬入模式双搬入模式基于元素的交错搬运从src中读取2*VL长度数据将偶数索引的元素存入dst0将奇数索引的元素存入dst1数据类型为b8/b16/b32。函数原型// NORM搬入模式 __simd_callee__ inline void asc_loadalign(vector_int8_t dst, __ubuf__ int8_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign(vector_uint8_t dst, __ubuf__ uint8_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign(vector_fp4x2_e2m1_t dst, __ubuf__ fp4x2_e2m1_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign(vector_fp4x2_e1m2_t dst, __ubuf__ fp4x2_e1m2_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign(vector_fp8_e8m0_t dst, __ubuf__ fp8_e8m0_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign(vector_fp8_e5m2_t dst, __ubuf__ fp8_e5m2_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign(vector_fp8_e4m3fn_t dst, __ubuf__ fp8_e4m3fn_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign(vector_hifloat8_t dst, __ubuf__ hifloat8_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign(vector_int16_t dst, __ubuf__ int16_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign(vector_uint16_t dst, __ubuf__ uint16_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign(vector_half dst, __ubuf__ half* src, int32_t offset) __simd_callee__ inline void asc_loadalign(vector_bfloat16_t dst, __ubuf__ bfloat16_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign(vector_int32_t dst, __ubuf__ int32_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign(vector_uint32_t dst, __ubuf__ uint32_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign(vector_float dst, __ubuf__ float* src, int32_t offset) __simd_callee__ inline void asc_loadalign(vector_int64_t dst, __ubuf__ int64_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign(vector_uint64_t dst, __ubuf__ uint64_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign(vector_int4x2_t dst, __ubuf__ int4b_t* src, int32_t offset) // BRC搬入模式 __simd_callee__ inline void asc_loadalign_brc(vector_int8_t dst, __ubuf__ int8_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_brc(vector_uint8_t dst, __ubuf__ uint8_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_brc(vector_fp4x2_e2m1_t dst, __ubuf__ fp4x2_e2m1_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_brc(vector_fp4x2_e1m2_t dst, __ubuf__ fp4x2_e1m2_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_brc(vector_fp8_e8m0_t dst, __ubuf__ fp8_e8m0_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_brc(vector_fp8_e5m2_t dst, __ubuf__ fp8_e5m2_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_brc(vector_fp8_e4m3fn_t dst, __ubuf__ fp8_e4m3fn_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_brc(vector_hifloat8_t dst, __ubuf__ hifloat8_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_brc(vector_int16_t dst, __ubuf__ int16_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_brc(vector_uint16_t dst, __ubuf__ uint16_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_brc(vector_half dst, __ubuf__ half* src, int32_t offset) __simd_callee__ inline void asc_loadalign_brc(vector_bfloat16_t dst, __ubuf__ bfloat16_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_brc(vector_int32_t dst, __ubuf__ int32_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_brc(vector_uint32_t dst, __ubuf__ uint32_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_brc(vector_float dst, __ubuf__ float* src, int32_t offset) __simd_callee__ inline void asc_loadalign_brc(vector_int4x2_t dst, __ubuf__ int4b_t* src, int32_t offset) // UP2X搬入模式 __simd_callee__ inline void asc_loadalign_upsample(vector_int8_t dst, __ubuf__ int8_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_upsample(vector_uint8_t dst, __ubuf__ uint8_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_upsample(vector_fp4x2_e2m1_t dst, __ubuf__ fp4x2_e2m1_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_upsample(vector_fp4x2_e1m2_t dst, __ubuf__ fp4x2_e1m2_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_upsample(vector_fp8_e8m0_t dst, __ubuf__ fp8_e8m0_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_upsample(vector_fp8_e5m2_t dst, __ubuf__ fp8_e5m2_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_upsample(vector_fp8_e4m3fn_t dst, __ubuf__ fp8_e4m3fn_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_upsample(vector_hifloat8_t dst, __ubuf__ hifloat8_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_upsample(vector_int16_t dst, __ubuf__ int16_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_upsample(vector_uint16_t dst, __ubuf__ uint16_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_upsample(vector_half dst, __ubuf__ half* src, int32_t offset) __simd_callee__ inline void asc_loadalign_upsample(vector_bfloat16_t dst, __ubuf__ bfloat16_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_upsample(vector_int4x2_t dst, __ubuf__ int4b_t* src, int32_t offset) // DOWN2X搬入模式 __simd_callee__ inline void asc_loadalign_downsample(vector_int8_t dst, __ubuf__ int8_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_downsample(vector_uint8_t dst, __ubuf__ uint8_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_downsample(vector_fp4x2_e2m1_t dst, __ubuf__ fp4x2_e2m1_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_downsample(vector_fp4x2_e1m2_t dst, __ubuf__ fp4x2_e1m2_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_downsample(vector_fp8_e8m0_t dst, __ubuf__ fp8_e8m0_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_downsample(vector_fp8_e5m2_t dst, __ubuf__ fp8_e5m2_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_downsample(vector_fp8_e4m3fn_t dst, __ubuf__ fp8_e4m3fn_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_downsample(vector_hifloat8_t dst, __ubuf__ hifloat8_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_downsample(vector_int16_t dst, __ubuf__ int16_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_downsample(vector_uint16_t dst, __ubuf__ uint16_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_downsample(vector_half dst, __ubuf__ half* src, int32_t offset) __simd_callee__ inline void asc_loadalign_downsample(vector_bfloat16_t dst, __ubuf__ bfloat16_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_downsample(vector_int4x2_t dst, __ubuf__ int4b_t* src, int32_t offset) // UNPACK搬入模式 __simd_callee__ inline void asc_loadalign_unpack(vector_int8_t dst, __ubuf__ int8_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_unpack(vector_uint8_t dst, __ubuf__ uint8_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_unpack(vector_fp4x2_e2m1_t dst, __ubuf__ fp4x2_e2m1_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_unpack(vector_fp4x2_e1m2_t dst, __ubuf__ fp4x2_e1m2_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_unpack(vector_fp8_e8m0_t dst, __ubuf__ fp8_e8m0_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_unpack(vector_fp8_e5m2_t dst, __ubuf__ fp8_e5m2_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_unpack(vector_fp8_e4m3fn_t dst, __ubuf__ fp8_e4m3fn_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_unpack(vector_hifloat8_t dst, __ubuf__ hifloat8_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_unpack(vector_int16_t dst, __ubuf__ int16_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_unpack(vector_uint16_t dst, __ubuf__ uint16_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_unpack(vector_half dst, __ubuf__ half* src, int32_t offset) __simd_callee__ inline void asc_loadalign_unpack(vector_bfloat16_t dst, __ubuf__ bfloat16_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_unpack(vector_int32_t dst, __ubuf__ int32_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_unpack(vector_uint32_t dst, __ubuf__ uint32_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_unpack(vector_float dst, __ubuf__ float* src, int32_t offset) __simd_callee__ inline void asc_loadalign_unpack(vector_int4x2_t dst, __ubuf__ int4b_t* src, int32_t offset) // UNPACKV2搬入模式 __simd_callee__ inline void asc_loadalign_unpack_v2(vector_int8_t dst, __ubuf__ int8_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_unpack_v2(vector_uint8_t dst, __ubuf__ uint8_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_unpack_v2(vector_fp4x2_e2m1_t dst, __ubuf__ fp4x2_e2m1_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_unpack_v2(vector_fp4x2_e1m2_t dst, __ubuf__ fp4x2_e1m2_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_unpack_v2(vector_fp8_e8m0_t dst, __ubuf__ fp8_e8m0_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_unpack_v2(vector_fp8_e5m2_t dst, __ubuf__ fp8_e5m2_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_unpack_v2(vector_fp8_e4m3fn_t dst, __ubuf__ fp8_e4m3fn_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_unpack_v2(vector_hifloat8_t dst, __ubuf__ hifloat8_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_unpack_v2(vector_int4x2_t dst, __ubuf__ int4b_t* src, int32_t offset) // BRCV2搬入模式 __simd_callee__ inline void asc_loadalign_brc_v2(vector_int8_t dst, __ubuf__ int8_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_brc_v2(vector_uint8_t dst, __ubuf__ uint8_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_brc_v2(vector_fp4x2_e2m1_t dst, __ubuf__ fp4x2_e2m1_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_brc_v2(vector_fp4x2_e1m2_t dst, __ubuf__ fp4x2_e1m2_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_brc_v2(vector_fp8_e8m0_t dst, __ubuf__ fp8_e8m0_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_brc_v2(vector_fp8_e5m2_t dst, __ubuf__ fp8_e5m2_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_brc_v2(vector_fp8_e4m3fn_t dst, __ubuf__ fp8_e4m3fn_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_brc_v2(vector_hifloat8_t dst, __ubuf__ hifloat8_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_brc_v2(vector_int16_t dst, __ubuf__ int16_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_brc_v2(vector_uint16_t dst, __ubuf__ uint16_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_brc_v2(vector_half dst, __ubuf__ half* src, int32_t offset) __simd_callee__ inline void asc_loadalign_brc_v2(vector_bfloat16_t dst, __ubuf__ bfloat16_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_brc_v2(vector_int32_t dst, __ubuf__ int32_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_brc_v2(vector_uint32_t dst, __ubuf__ uint32_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_brc_v2(vector_float dst, __ubuf__ float* src, int32_t offset) __simd_callee__ inline void asc_loadalign_brc_v2(vector_int4x2_t dst, __ubuf__ int4b_t* src, int32_t offset) // BRCV3搬入模式 __simd_callee__ inline void asc_loadalign_brc_v3(vector_int16_t dst, __ubuf__ int16_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_brc_v3(vector_uint16_t dst, __ubuf__ uint16_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_brc_v3(vector_half dst, __ubuf__ half* src, int32_t offset) __simd_callee__ inline void asc_loadalign_brc_v3(vector_bfloat16_t dst, __ubuf__ bfloat16_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_brc_v3(vector_int32_t dst, __ubuf__ int32_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_brc_v3(vector_uint32_t dst, __ubuf__ uint32_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_brc_v3(vector_float dst, __ubuf__ float* src, int32_t offset) // DEINTLV搬入模式 __simd_callee__ inline void asc_loadalign_deintlv(vector_int8_t dst0, vector_int8_t dst1, __ubuf__ int8_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_deintlv(vector_uint8_t dst0, vector_uint8_t dst1, __ubuf__ uint8_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_deintlv(vector_fp4x2_e2m1_t dst0, vector_fp4x2_e2m1_t dst1, __ubuf__ fp4x2_e2m1_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_deintlv(vector_fp4x2_e1m2_t dst0, vector_fp4x2_e1m2_t dst1, __ubuf__ fp4x2_e1m2_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_deintlv(vector_fp8_e8m0_t dst0, vector_fp8_e8m0_t dst1, __ubuf__ fp8_e8m0_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_deintlv(vector_fp8_e5m2_t dst0, vector_fp8_e5m2_t dst1, __ubuf__ fp8_e5m2_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_deintlv(vector_fp8_e4m3fn_t dst0, vector_fp8_e4m3fn_t dst1, __ubuf__ fp8_e4m3fn_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_deintlv(vector_hifloat8_t dst0, vector_hifloat8_t dst1, __ubuf__ hifloat8_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_deintlv(vector_int16_t dst0, vector_int16_t dst1, __ubuf__ int16_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_deintlv(vector_uint16_t dst0, vector_uint16_t dst1, __ubuf__ uint16_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_deintlv(vector_half dst0, vector_half dst1, __ubuf__ half* src, int32_t offset) __simd_callee__ inline void asc_loadalign_deintlv(vector_bfloat16_t dst0, vector_bfloat16_t dst1, __ubuf__ bfloat16_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_deintlv(vector_int32_t dst0, vector_int32_t dst1, __ubuf__ int32_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_deintlv(vector_uint32_t dst0, vector_uint32_t dst1, __ubuf__ uint32_t* src, int32_t offset) __simd_callee__ inline void asc_loadalign_deintlv(vector_float dst0, vector_float dst1, __ubuf__ float* src, int32_t offset) __simd_callee__ inline void asc_loadalign_deintlv(vector_int4x2_t dst0, vector_int4x2_t dst1, __ubuf__ int4b_t* src, int32_t offset)参数说明参数名输入/输出描述dst输出目的操作数矢量数据寄存器。dst0输出目的操作数矢量数据寄存器。dst1输出目的操作数矢量数据寄存器。src输入源操作数矢量的起始地址。offset输入地址偏移量。矢量数据寄存器的详细说明请参见reg数据类型定义.md。返回值说明无流水类型PIPE_V约束说明无调用示例vector_half dst; __ubuf__ half* src; int32_t offset; asc_loadalign(dst, src, offset);【免费下载链接】asc-devkit本项目是CANN 推出的昇腾AI处理器专用的算子程序开发语言原生支持C和C标准规范主要由类库和语言扩展层构成提供多层级API满足多维场景算子开发诉求。项目地址: https://gitcode.com/cann/asc-devkit创作声明：本文部分内容由AI辅助生成（AIGC），仅供参考