昇腾C L0C到L1数据搬运

asc_copy_l0c2l1【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Ascend 950PR/Ascend 950DT不支持Atlas A3 训练系列产品/Atlas A3 推理系列产品支持Atlas A2 训练系列产品/Atlas A2 推理系列产品支持Atlas 200I/500 A2 推理产品不支持Atlas 推理系列产品AI Core不支持Atlas 推理系列产品Vector Core不支持Atlas 训练系列产品不支持功能说明头文件路径c_api/cube_datamove/cube_datamove.h。矩阵计算完成后对结果进行量化处理之后将处理结果搬运到L1 Buffer中。量化模式共分为9种分别为NoQuant不开启量化功能。DEQF16int32_t量化成half。量化结果不支持INF_NAN模式。F322BF16float量化成bfloat16_t。量化结果不支持INF_NAN模式。F322F16float量化成half。量化结果不支持INF_NAN模式。QF322B8_PREfloat量化成uint8_t/int8_t。scalar量化。REQ8int32_t量化成uint8_t/int8_t。scalar量化。VDEQF16int32_t量化成half。量化结果不支持INF_NAN模式。VQF322B8_PREfloat量化成uint8_t/int8_t。矢量量化。VREQ8int32_t量化成uint8_t/int8_t。矢量量化。此外该搬运指令还支持int4b_t类型的量化相关量化类型分为两种int32_t量化为int4b_tfloat量化为int4b_t。函数原型常规搬运__aicore__ inline void asc_copy_l0c2l1(__cbuf__ half* dst, __cc__ float* src, uint16_t n_size,uint16_t m_size, uint32_t dst_stride, uint16_t src_stride, uint8_t unit_flag_mode, uint64_t quant_pre, uint8_t relu_pre, bool enable_channel_split, bool enable_nz2nd) __aicore__ inline void asc_copy_l0c2l1(__cbuf__ bfloat16_t* dst, __cc__ float* src, uint16_t n_size,uint16_t m_size, uint32_t dst_stride, uint16_t src_stride, uint8_t unit_flag_mode, uint64_t quant_pre, uint8_t relu_pre, bool enable_channel_split, bool enable_nz2nd) __aicore__ inline void asc_copy_l0c2l1(__cbuf__ int8_t* dst, __cc__ float* src, uint16_t n_size,uint16_t m_size, uint32_t dst_stride, uint16_t src_stride, uint8_t unit_flag_mode, uint64_t quant_pre, uint8_t relu_pre, bool enable_channel_split, bool enable_nz2nd) __aicore__ inline void asc_copy_l0c2l1(__cbuf__ half* dst, __cc__ int32_t* src, uint16_t n_size,uint16_t m_size, uint32_t dst_stride, uint16_t src_stride, uint8_t unit_flag_mode, uint64_t quant_pre, uint8_t relu_pre, bool enable_channel_split, bool enable_nz2nd) __aicore__ inline void asc_copy_l0c2l1(__cbuf__ int16_t* dst, __cc__ int32_t* src, uint16_t n_size,uint16_t m_size, uint32_t dst_stride, uint16_t src_stride, uint8_t unit_flag_mode, uint64_t quant_pre, uint8_t relu_pre, bool enable_channel_split, bool enable_nz2nd) __aicore__ inline void asc_copy_l0c2l1(__cbuf__ int8_t* dst, __cc__ int32_t* src, uint16_t n_size,uint16_t m_size, uint32_t dst_stride, uint16_t src_stride, uint8_t unit_flag_mode, uint64_t quant_pre, uint8_t relu_pre, bool enable_channel_split, bool enable_nz2nd) __aicore__ inline void asc_copy_l0c2l1(__cbuf__ uint8_t* dst, __cc__ int32_t* src, uint16_t n_size,uint16_t m_size, uint32_t dst_stride, uint16_t src_stride, uint8_t unit_flag_mode, uint64_t quant_pre, uint8_t relu_pre, bool enable_channel_split, bool enable_nz2nd) __aicore__ inline void asc_copy_l0c2l1(__cbuf__ void* dst, __cc__ float* src, uint16_t n_size,uint16_t m_size, uint32_t dst_stride, uint16_t src_stride, uint8_t unit_flag_mode, uint64_t quant_pre, uint8_t relu_pre, bool enable_channel_split, bool enable_nz2nd) __aicore__ inline void asc_copy_l0c2l1(__cbuf__ void* dst, __cc__ int32_t* src, uint16_t n_size,uint16_t m_size, uint32_t dst_stride, uint16_t src_stride, uint8_t unit_flag_mode, uint64_t quant_pre, uint8_t relu_pre, bool enable_channel_split, bool enable_nz2nd)同步搬运__aicore__ inline void asc_copy_l0c2l1_sync(__cbuf__ half* dst, __cc__ float* src, uint16_t n_size,uint16_t m_size, uint32_t dst_stride, uint16_t src_stride, uint8_t unit_flag_mode, uint64_t quant_pre, uint8_t relu_pre, bool enable_channel_split, bool enable_nz2nd) __aicore__ inline void asc_copy_l0c2l1_sync(__cbuf__ bfloat16_t* dst, __cc__ float* src, uint16_t n_size,uint16_t m_size, uint32_t dst_stride, uint16_t src_stride, uint8_t unit_flag_mode, uint64_t quant_pre, uint8_t relu_pre, bool enable_channel_split, bool enable_nz2nd) __aicore__ inline void asc_copy_l0c2l1_sync(__cbuf__ int8_t* dst, __cc__ float* src, uint16_t n_size,uint16_t m_size, uint32_t dst_stride, uint16_t src_stride, uint8_t unit_flag_mode, uint64_t quant_pre, uint8_t relu_pre, bool enable_channel_split, bool enable_nz2nd) __aicore__ inline void asc_copy_l0c2l1_sync(__cbuf__ half* dst, __cc__ int32_t* src, uint16_t n_size,uint16_t m_size, uint32_t dst_stride, uint16_t src_stride, uint8_t unit_flag_mode, uint64_t quant_pre, uint8_t relu_pre, bool enable_channel_split, bool enable_nz2nd) __aicore__ inline void asc_copy_l0c2l1_sync(__cbuf__ int16_t* dst, __cc__ int32_t* src, uint16_t n_size,uint16_t m_size, uint32_t dst_stride, uint16_t src_stride, uint8_t unit_flag_mode, uint64_t quant_pre, uint8_t relu_pre, bool enable_channel_split, bool enable_nz2nd) __aicore__ inline void asc_copy_l0c2l1_sync(__cbuf__ int8_t* dst, __cc__ int32_t* src, uint16_t n_size,uint16_t m_size, uint32_t dst_stride, uint16_t src_stride, uint8_t unit_flag_mode, uint64_t quant_pre, uint8_t relu_pre, bool enable_channel_split, bool enable_nz2nd) __aicore__ inline void asc_copy_l0c2l1_sync(__cbuf__ uint8_t* dst, __cc__ int32_t* src, uint16_t n_size,uint16_t m_size, uint32_t dst_stride, uint16_t src_stride, uint8_t unit_flag_mode, uint64_t quant_pre, uint8_t relu_pre, bool enable_channel_split, bool enable_nz2nd) __aicore__ inline void asc_copy_l0c2l1_sync(__cbuf__ void* dst, __cc__ float* src, uint16_t n_size,uint16_t m_size, uint32_t dst_stride, uint16_t src_stride, uint8_t unit_flag_mode, uint64_t quant_pre, uint8_t relu_pre, bool enable_channel_split, bool enable_nz2nd) __aicore__ inline void asc_copy_l0c2l1_sync(__cbuf__ void* dst, __cc__ int32_t* src, uint16_t n_size,uint16_t m_size, uint32_t dst_stride, uint16_t src_stride, uint8_t unit_flag_mode, uint64_t quant_pre, uint8_t relu_pre, bool enable_channel_split, bool enable_nz2nd)参数说明参数名输入/输出描述dst输出目的操作数矢量的起始地址。src输入源操作数的起始地址。n_size输入源NZ矩阵在N方向上的大小。• 不开启NZ2ND功能取值范围[1, 4095]• 开启NZ2ND功能取值范围[1, 4095]。m_size输入源NZ矩阵在M方向上的大小。• 不开启NZ2ND功能取值范围[1, 65535]• 开启NZ2ND功能取值范围[1, 8192]。dst_stride输入• 不开启NZ2ND功能目的NZ矩阵中相邻Z排布的起始地址偏移取值不为0单位element。• 开启NZ2ND/NZ2DN功能目的ND矩阵每一行中的元素个数取值不为0 单位element。src_stride输入源NZ矩阵中相邻Z排布的起始地址偏移取值范围[0, 65535]单位C0_Size(16*sizeof(T), T为src的数据类型)。unit_flag_mode输入与unit_flag参数相关取值如下• 0保留值• 2 开启unit_flag硬件执行完指令之后不会设置寄存器• 3 开启unit_flag硬件执行完指令后会将unit_flag关闭。quant_pre输入量化参数。取值见功能说明。relu_pre输入开启relu。enable_channel_split输入是否开启通道拆分的功能默认false不开启该功能。仅在src和dst都为float时才能开启通道拆分且不能同时开启channel_split和NZ2ND功能。enable_nz2nd输入开启NZ2ND开关。• false不开启• true开启。返回值说明无流水类型PIPE_FIX约束说明src的起始地址要求按照对应数据类型所占字节数对齐。dst的起始地址要求32字节对齐。如果需要执行多条asc_copy_l0c2l1指令且asc_copy_l0c2l1指令的目的地址存在重叠需要插入同步指令保证多个asc_copy_l0c2l1指令的串行化防止出现异常数据。调用示例// dst src分别对应目的操作数的输出地址和源操作数的输入地址 constexpr uint64_t total_length 256; __cbuf__ int32_t dst[total_length]; __cc__ int32_t src[total_length]; // 其余入参均已默认数值传入 uint16_t n_size 16; uint16_t m_size 16; uint32_t dst_stride 8; uint16_t src_stride 8; uint8_t unit_flag_mode 0; uint64_t quant_pre QuantMode_t::NoQuant; uint8_t relu_pre 0; bool channel_split false; bool enable_nz2nd false; asc_copy_l0c2l1(dst, src, n_size, m_size, dst_stride, src_stride, unit_flag_mode, quant_pre, relu_pre, enable_channel_split, enable_nz2nd);【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