asc_copy_l0c2ub【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支持Atlas A3 训练系列产品/Atlas A3 推理系列产品不支持Atlas A2 训练系列产品/Atlas A2 推理系列产品不支持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Unified Buffer中。量化参数共有2个quant_pre和quant_post分别对应预处理和后处理阶段。quant_pre可选量化模式分别为NoQuant不开启量化功能。F322BF16float量化成bfloat16_t。量化结果不支持INF_NAN模式。F322F16float量化成half。量化结果支持INF_NAN模式。DEQF16int32_t量化成half。量化结果不支持INF_NAN模式。VDEQF16int32_t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。VQF322B8_PREfloat量化成uint8_t/int8_t。矢量量化。REQ8int32_t量化成uint8_t/int8_t。scalar量化。VREQ8int32_t量化成uint8_t/int8_t。矢量量化。QF322FP8_PREfloat量化成fp8_e4m3fn_tscalar量化。VQF322FP8_PREfloat量化成fp8_e4m3fn_t矢量量化。QF322HIF8_PREfloat量化成hifloat8_tHalf to Away Roundscalar量化。VQF322HIF8_PREfloat量化成hifloat8_tHalf to Away Round矢量量化。QF322HIF8_PRE_HYBRIDfloat量化成hifloat8_tHybrid Roundscalar量化。VQF322HIF8_PRE_HYBRIDfloat量化成hifloat8_tHybrid Round矢量量化。QS322BF16_PREint32_t量化成bfloat16_tscalar量化。VQS322BF16_PREint32_t量化成bfloat16_t矢量量化。QF322F16_PREfloat量化成halfscalar量化。VQF322F16_PREfloat量化成half矢量量化。QF322BF16_PREfloat量化成bfloat16_tscalar量化。VQF322BF16_PREfloat量化成bfloat16_t矢量量化。QF322F32_PREfloat量化成floatscalar量化。该量化模式精度无法达到双万分之一可以达到双千分之一。VQF322F32_PREfloat量化成float矢量量化。该量化模式精度无法达到双万分之一可以达到双千分之一。quant_post可选量化模式分别为NoConv不开启量化功能。QS162B8_POSTint16_t量化成uint8_t/int8_tscalar量化。VQS162B8_POSTint16_t量化成uint8_t/int8_t矢量量化。QF162B8_POSThalf量化成uint8_t/int8_tscalar量化。VQF162B8_POSThalf量化成uint8_t/int8_t矢量量化。QS162S4_POSTint16_t量化成int4_tscalar量化。VQS162S4_POSTint16_t量化成int4_t矢量量化。QF162S4_POSThalf量化成int4_t类型scalar量化。VQF162S4_POSThalf量化成int4_t类型矢量量化。QS162S16_POSTint16_t量化成int16_tscalar量化。VQS162S16_POSTint16_t量化成int16_t矢量量化。QF162S16_POSThalf量化成int16_tscalar量化。VQF162S16_POSThalf量化成int16_t矢量量化。函数原型常规搬运__aicore__ inline void asc_copy_l0c2ub(__ubuf__ bfloat16_t *dst_addr, __cc__ float *src_addr, uint16_t n_size, uint16_t m_size, uint32_t loop_dst_stride, uint16_t loop_src_stride, uint8_t dual_dst_ctl, bool sub_blockid, uint8_t clip_relu_pre, uint8_t unit_flag_ctl, uint64_t quant_pre, uint8_t relu_pre, bool split_en, bool nz2nd_en, uint64_t quant_post, uint8_t relu_post, bool clip_relu_post, uint8_t eltwise_op, bool eltwise_antq_en, bool c0_pad_en, bool broadcast_en, bool nz2dn_en) __aicore__ inline void asc_copy_l0c2ub(__ubuf__ half *dst_addr, __cc__ float *src_addr, uint16_t n_size, uint16_t m_size, uint32_t loop_dst_stride, uint16_t loop_src_stride, uint8_t dual_dst_ctl, bool sub_blockid, uint8_t clip_relu_pre, uint8_t unit_flag_ctl, uint64_t quant_pre, uint8_t relu_pre, bool split_en, bool nz2nd_en, uint64_t quant_post, uint8_t relu_post, bool clip_relu_post, uint8_t eltwise_op, bool eltwise_antq_en, bool c0_pad_en, bool broadcast_en, bool nz2dn_en) __aicore__ inline void asc_copy_l0c2ub(__ubuf__ fp8_e4m3fn_t *dst_addr, __cc__ float *src_addr, uint16_t n_size, uint16_t m_size, uint32_t loop_dst_stride, uint16_t loop_src_stride, uint8_t dual_dst_ctl, bool sub_blockid, uint8_t clip_relu_pre, uint8_t unit_flag_ctl, uint64_t quant_pre, uint8_t relu_pre, bool split_en, bool nz2nd_en, uint64_t quant_post, uint8_t relu_post, bool clip_relu_post, uint8_t eltwise_op, bool eltwise_antq_en, bool c0_pad_en, bool broadcast_en, bool nz2dn_en) __aicore__ inline void asc_copy_l0c2ub(__ubuf__ fp8_e5m2_t *dst_addr, __cc__ float *src_addr, uint16_t n_size, uint16_t m_size, uint32_t loop_dst_stride, uint16_t loop_src_stride, uint8_t dual_dst_ctl, bool sub_blockid, uint8_t clip_relu_pre, uint8_t unit_flag_ctl, uint64_t quant_pre, uint8_t relu_pre, bool split_en, bool nz2nd_en, uint64_t quant_post, uint8_t relu_post, bool clip_relu_post, uint8_t eltwise_op, bool eltwise_antq_en, bool c0_pad_en, bool broadcast_en, bool nz2dn_en) __aicore__ inline void asc_copy_l0c2ub(__ubuf__ hifloat8_t *dst_addr, __cc__ float *src_addr, uint16_t n_size, uint16_t m_size, uint32_t loop_dst_stride, uint16_t loop_src_stride, uint8_t dual_dst_ctl, bool sub_blockid, uint8_t clip_relu_pre, uint8_t unit_flag_ctl, uint64_t quant_pre, uint8_t relu_pre, bool split_en, bool nz2nd_en, uint64_t quant_post, uint8_t relu_post, bool clip_relu_post, uint8_t eltwise_op, bool eltwise_antq_en, bool c0_pad_en, bool broadcast_en, bool nz2dn_en) __aicore__ inline void asc_copy_l0c2ub(__ubuf__ int8_t *dst_addr, __cc__ float *src_addr, uint16_t n_size, uint16_t m_size, uint32_t loop_dst_stride, uint16_t loop_src_stride, uint8_t dual_dst_ctl, bool sub_blockid, uint8_t clip_relu_pre, uint8_t unit_flag_ctl, uint64_t quant_pre, uint8_t relu_pre, bool split_en, bool nz2nd_en, uint64_t quant_post, uint8_t relu_post, bool clip_relu_post, uint8_t eltwise_op, bool eltwise_antq_en, bool c0_pad_en, bool broadcast_en, bool nz2dn_en) __aicore__ inline void asc_copy_l0c2ub(__ubuf__ uint8_t *dst_addr, __cc__ float *src_addr, uint16_t n_size, uint16_t m_size, uint32_t loop_dst_stride, uint16_t loop_src_stride, uint8_t dual_dst_ctl, bool sub_blockid, uint8_t clip_relu_pre, uint8_t unit_flag_ctl, uint64_t quant_pre, uint8_t relu_pre, bool split_en, bool nz2nd_en, uint64_t quant_post, uint8_t relu_post, bool clip_relu_post, uint8_t eltwise_op, bool eltwise_antq_en, bool c0_pad_en, bool broadcast_en, bool nz2dn_en) __aicore__ inline void asc_copy_l0c2ub(__ubuf__ float *dst_addr, __cc__ float *src_addr, uint16_t n_size, uint16_t m_size, uint32_t loop_dst_stride, uint16_t loop_src_stride, uint8_t dual_dst_ctl, bool sub_blockid, uint8_t clip_relu_pre, uint8_t unit_flag_ctl, uint64_t quant_pre, uint8_t relu_pre, bool split_en, bool nz2nd_en, uint64_t quant_post, uint8_t relu_post, bool clip_relu_post, uint8_t eltwise_op, bool eltwise_antq_en, bool c0_pad_en, bool broadcast_en, bool nz2dn_en) __aicore__ inline void asc_copy_l0c2ub(__ubuf__ bfloat16_t *dst_addr, __cc__ int32_t *src_addr, uint16_t n_size, uint16_t m_size, uint32_t loop_dst_stride, uint16_t loop_src_stride, uint8_t dual_dst_ctl, bool sub_blockid, uint8_t clip_relu_pre, uint8_t unit_flag_ctl, uint64_t quant_pre, uint8_t relu_pre, bool split_en, bool nz2nd_en, uint64_t quant_post, uint8_t relu_post, bool clip_relu_post, uint8_t eltwise_op, bool eltwise_antq_en, bool c0_pad_en, bool broadcast_en, bool nz2dn_en) __aicore__ inline void asc_copy_l0c2ub(__ubuf__ half *dst_addr, __cc__ int32_t *src_addr, uint16_t n_size, uint16_t m_size, uint32_t loop_dst_stride, uint16_t loop_src_stride, uint8_t dual_dst_ctl, bool sub_blockid, uint8_t clip_relu_pre, uint8_t unit_flag_ctl, uint64_t quant_pre, uint8_t relu_pre, bool split_en, bool nz2nd_en, uint64_t quant_post, uint8_t relu_post, bool clip_relu_post, uint8_t eltwise_op, bool eltwise_antq_en, bool c0_pad_en, bool broadcast_en, bool nz2dn_en) __aicore__ inline void asc_copy_l0c2ub(__ubuf__ fp8_e4m3fn_t *dst_addr, __cc__ int32_t *src_addr, uint16_t n_size, uint16_t m_size, uint32_t loop_dst_stride, uint16_t loop_src_stride, uint8_t dual_dst_ctl, bool sub_blockid, uint8_t clip_relu_pre, uint8_t unit_flag_ctl, uint64_t quant_pre, uint8_t relu_pre, bool split_en, bool nz2nd_en, uint64_t quant_post, uint8_t relu_post, bool clip_relu_post, uint8_t eltwise_op, bool eltwise_antq_en, bool c0_pad_en, bool broadcast_en, bool nz2dn_en) __aicore__ inline void asc_copy_l0c2ub(__ubuf__ fp8_e5m2_t *dst_addr, __cc__ int32_t *src_addr, uint16_t n_size, uint16_t m_size, uint32_t loop_dst_stride, uint16_t loop_src_stride, uint8_t dual_dst_ctl, bool sub_blockid, uint8_t clip_relu_pre, uint8_t unit_flag_ctl, uint64_t quant_pre, uint8_t relu_pre, bool split_en, bool nz2nd_en, uint64_t quant_post, uint8_t relu_post, bool clip_relu_post, uint8_t eltwise_op, bool eltwise_antq_en, bool c0_pad_en, bool broadcast_en, bool nz2dn_en) __aicore__ inline void asc_copy_l0c2ub(__ubuf__ hifloat8_t *dst_addr, __cc__ int32_t *src_addr, uint16_t n_size, uint16_t m_size, uint32_t loop_dst_stride, uint16_t loop_src_stride, uint8_t dual_dst_ctl, bool sub_blockid, uint8_t clip_relu_pre, uint8_t unit_flag_ctl, uint64_t quant_pre, uint8_t relu_pre, bool split_en, bool nz2nd_en, uint64_t quant_post, uint8_t relu_post, bool clip_relu_post, uint8_t eltwise_op, bool eltwise_antq_en, bool c0_pad_en, bool broadcast_en, bool nz2dn_en) __aicore__ inline void asc_copy_l0c2ub(__ubuf__ int8_t *dst_addr, __cc__ int32_t *src_addr, uint16_t n_size, uint16_t m_size, uint32_t loop_dst_stride, uint16_t loop_src_stride, uint8_t dual_dst_ctl, bool sub_blockid, uint8_t clip_relu_pre, uint8_t unit_flag_ctl, uint64_t quant_pre, uint8_t relu_pre, bool split_en, bool nz2nd_en, uint64_t quant_post, uint8_t relu_post, bool clip_relu_post, uint8_t eltwise_op, bool eltwise_antq_en, bool c0_pad_en, bool broadcast_en, bool nz2dn_en) __aicore__ inline void asc_copy_l0c2ub(__ubuf__ uint8_t *dst_addr, __cc__ int32_t *src_addr, uint16_t n_size, uint16_t m_size, uint32_t loop_dst_stride, uint16_t loop_src_stride, uint8_t dual_dst_ctl, bool sub_blockid, uint8_t clip_relu_pre, uint8_t unit_flag_ctl, uint64_t quant_pre, uint8_t relu_pre, bool split_en, bool nz2nd_en, uint64_t quant_post, uint8_t relu_post, bool clip_relu_post, uint8_t eltwise_op, bool eltwise_antq_en, bool c0_pad_en, bool broadcast_en, bool nz2dn_en) __aicore__ inline void asc_copy_l0c2ub(__ubuf__ int32_t *dst_addr, __cc__ int32_t *src_addr, uint16_t n_size, uint16_t m_size, uint32_t loop_dst_stride, uint16_t loop_src_stride, uint8_t dual_dst_ctl, bool sub_blockid, uint8_t clip_relu_pre, uint8_t unit_flag_ctl, uint64_t quant_pre, uint8_t relu_pre, bool split_en, bool nz2nd_en, uint64_t quant_post, uint8_t relu_post, bool clip_relu_post, uint8_t eltwise_op, bool eltwise_antq_en, bool c0_pad_en, bool broadcast_en, bool nz2dn_en) __aicore__ inline void asc_copy_l0c2ub(__ubuf__ void *dst_addr, __cc__ float *src_addr, uint16_t n_size, uint16_t m_size, uint32_t loop_dst_stride, uint16_t loop_src_stride, uint8_t dual_dst_ctl, bool sub_blockid, uint8_t clip_relu_pre, uint8_t unit_flag_ctl, uint64_t quant_pre, uint8_t relu_pre, bool split_en, bool nz2nd_en, uint64_t quant_post, uint8_t relu_post, bool clip_relu_post, uint8_t eltwise_op, bool eltwise_antq_en, bool c0_pad_en, bool broadcast_en, bool nz2dn_en) __aicore__ inline void asc_copy_l0c2ub(__ubuf__ void *dst_addr, __cc__ int32_t *src_addr, uint16_t n_size, uint16_t m_size, uint32_t loop_dst_stride, uint16_t loop_src_stride, uint8_t dual_dst_ctl, bool sub_blockid, uint8_t clip_relu_pre, uint8_t unit_flag_ctl, uint64_t quant_pre, uint8_t relu_pre, bool split_en, bool nz2nd_en, uint64_t quant_post, uint8_t relu_post, bool clip_relu_post, uint8_t eltwise_op, bool eltwise_antq_en, bool c0_pad_en, bool broadcast_en, bool nz2dn_en)同步搬运__aicore__ inline void asc_copy_l0c2ub_sync(__ubuf__ bfloat16_t* dst_addr, __cc__ float* src_addr, uint16_t n_size, uint16_t m_size, uint32_t loop_dst_stride, uint16_t loop_src_stride, uint8_t dual_dst_ctl, bool sub_blockid, uint8_t clip_relu_pre, uint8_t unit_flag_ctl, uint64_t quant_pre, uint8_t relu_pre, bool split_en, bool nz2nd_en, uint64_t quant_post, uint8_t relu_post, bool clip_relu_post, uint8_t eltwise_op, bool eltwise_antq_en, bool c0_pad_en, bool broadcast_en, bool nz2dn_en) __aicore__ inline void asc_copy_l0c2ub_sync(__ubuf__ half* dst_addr, __cc__ float* src_addr, uint16_t n_size, uint16_t m_size, uint32_t loop_dst_stride, uint16_t loop_src_stride, uint8_t dual_dst_ctl, bool sub_blockid, uint8_t clip_relu_pre, uint8_t unit_flag_ctl, uint64_t quant_pre, uint8_t relu_pre, bool split_en, bool nz2nd_en, uint64_t quant_post, uint8_t relu_post, bool clip_relu_post, uint8_t eltwise_op, bool eltwise_antq_en, bool c0_pad_en, bool broadcast_en, bool nz2dn_en) __aicore__ inline void asc_copy_l0c2ub_sync(__ubuf__ fp8_e4m3fn_t* dst_addr, __cc__ float* src_addr, uint16_t n_size, uint16_t m_size, uint32_t loop_dst_stride, uint16_t loop_src_stride, uint8_t dual_dst_ctl, bool sub_blockid, uint8_t clip_relu_pre, uint8_t unit_flag_ctl, uint64_t quant_pre, uint8_t relu_pre, bool split_en, bool nz2nd_en, uint64_t quant_post, uint8_t relu_post, bool clip_relu_post, uint8_t eltwise_op, bool eltwise_antq_en, bool c0_pad_en, bool broadcast_en, bool nz2dn_en) __aicore__ inline void asc_copy_l0c2ub_sync(__ubuf__ fp8_e5m2_t* dst_addr, __cc__ float* src_addr, uint16_t n_size, uint16_t m_size, uint32_t loop_dst_stride, uint16_t loop_src_stride, uint8_t dual_dst_ctl, bool sub_blockid, uint8_t clip_relu_pre, uint8_t unit_flag_ctl, uint64_t quant_pre, uint8_t relu_pre, bool split_en, bool nz2nd_en, uint64_t quant_post, uint8_t relu_post, bool clip_relu_post, uint8_t eltwise_op, bool eltwise_antq_en, bool c0_pad_en, bool broadcast_en, bool nz2dn_en) __aicore__ inline void asc_copy_l0c2ub_sync(__ubuf__ hifloat8_t* dst_addr, __cc__ float* src_addr, uint16_t n_size, uint16_t m_size, uint32_t loop_dst_stride, uint16_t loop_src_stride, uint8_t dual_dst_ctl, bool sub_blockid, uint8_t clip_relu_pre, uint8_t unit_flag_ctl, uint64_t quant_pre, uint8_t relu_pre, bool split_en, bool nz2nd_en, uint64_t quant_post, uint8_t relu_post, bool clip_relu_post, uint8_t eltwise_op, bool eltwise_antq_en, bool c0_pad_en, bool broadcast_en, bool nz2dn_en) __aicore__ inline void asc_copy_l0c2ub_sync(__ubuf__ int8_t* dst_addr, __cc__ float* src_addr, uint16_t n_size, uint16_t m_size, uint32_t loop_dst_stride, uint16_t loop_src_stride, uint8_t dual_dst_ctl, bool sub_blockid, uint8_t clip_relu_pre, uint8_t unit_flag_ctl, uint64_t quant_pre, uint8_t relu_pre, bool split_en, bool nz2nd_en, uint64_t quant_post, uint8_t relu_post, bool clip_relu_post, uint8_t eltwise_op, bool eltwise_antq_en, bool c0_pad_en, bool broadcast_en, bool nz2dn_en) __aicore__ inline void asc_copy_l0c2ub_sync(__ubuf__ uint8_t* dst_addr, __cc__ float* src_addr, uint16_t n_size, uint16_t m_size, uint32_t loop_dst_stride, uint16_t loop_src_stride, uint8_t dual_dst_ctl, bool sub_blockid, uint8_t clip_relu_pre, uint8_t unit_flag_ctl, uint64_t quant_pre, uint8_t relu_pre, bool split_en, bool nz2nd_en, uint64_t quant_post, uint8_t relu_post, bool clip_relu_post, uint8_t eltwise_op, bool eltwise_antq_en, bool c0_pad_en, bool broadcast_en, bool nz2dn_en) __aicore__ inline void asc_copy_l0c2ub_sync(__ubuf__ float* dst_addr, __cc__ float* src_addr, uint16_t n_size, uint16_t m_size, uint32_t loop_dst_stride, uint16_t loop_src_stride, uint8_t dual_dst_ctl, bool sub_blockid, uint8_t clip_relu_pre, uint8_t unit_flag_ctl, uint64_t quant_pre, uint8_t relu_pre, bool split_en, bool nz2nd_en, uint64_t quant_post, uint8_t relu_post, bool clip_relu_post, uint8_t eltwise_op, bool eltwise_antq_en, bool c0_pad_en, bool broadcast_en, bool nz2dn_en) __aicore__ inline void asc_copy_l0c2ub_sync(__ubuf__ bfloat16_t* dst_addr, __cc__ int32_t* src_addr, uint16_t n_size, uint16_t m_size, uint32_t loop_dst_stride, uint16_t loop_src_stride, uint8_t dual_dst_ctl, bool sub_blockid, uint8_t clip_relu_pre, uint8_t unit_flag_ctl, uint64_t quant_pre, uint8_t relu_pre, bool split_en, bool nz2nd_en, uint64_t quant_post, uint8_t relu_post, bool clip_relu_post, uint8_t eltwise_op, bool eltwise_antq_en, bool c0_pad_en, bool broadcast_en, bool nz2dn_en) __aicore__ inline void asc_copy_l0c2ub_sync(__ubuf__ half* dst_addr, __cc__ int32_t* src_addr, uint16_t n_size, uint16_t m_size, uint32_t loop_dst_stride, uint16_t loop_src_stride, uint8_t dual_dst_ctl, bool sub_blockid, uint8_t clip_relu_pre, uint8_t unit_flag_ctl, uint64_t quant_pre, uint8_t relu_pre, bool split_en, bool nz2nd_en, uint64_t quant_post, uint8_t relu_post, bool clip_relu_post, uint8_t eltwise_op, bool eltwise_antq_en, bool c0_pad_en, bool broadcast_en, bool nz2dn_en) __aicore__ inline void asc_copy_l0c2ub_sync(__ubuf__ fp8_e4m3fn_t* dst_addr, __cc__ int32_t* src_addr, uint16_t n_size, uint16_t m_size, uint32_t loop_dst_stride, uint16_t loop_src_stride, uint8_t dual_dst_ctl, bool sub_blockid, uint8_t clip_relu_pre, uint8_t unit_flag_ctl, uint64_t quant_pre, uint8_t relu_pre, bool split_en, bool nz2nd_en, uint64_t quant_post, uint8_t relu_post, bool clip_relu_post, uint8_t eltwise_op, bool eltwise_antq_en, bool c0_pad_en, bool broadcast_en, bool nz2dn_en) __aicore__ inline void asc_copy_l0c2ub_sync(__ubuf__ fp8_e5m2_t* dst_addr, __cc__ int32_t* src_addr, uint16_t n_size, uint16_t m_size, uint32_t loop_dst_stride, uint16_t loop_src_stride, uint8_t dual_dst_ctl, bool sub_blockid, uint8_t clip_relu_pre, uint8_t unit_flag_ctl, uint64_t quant_pre, uint8_t relu_pre, bool split_en, bool nz2nd_en, uint64_t quant_post, uint8_t relu_post, bool clip_relu_post, uint8_t eltwise_op, bool eltwise_antq_en, bool c0_pad_en, bool broadcast_en, bool nz2dn_en) __aicore__ inline void asc_copy_l0c2ub_sync(__ubuf__ hifloat8_t* dst_addr, __cc__ int32_t* src_addr, uint16_t n_size, uint16_t m_size, uint32_t loop_dst_stride, uint16_t loop_src_stride, uint8_t dual_dst_ctl, bool sub_blockid, uint8_t clip_relu_pre, uint8_t unit_flag_ctl, uint64_t quant_pre, uint8_t relu_pre, bool split_en, bool nz2nd_en, uint64_t quant_post, uint8_t relu_post, bool clip_relu_post, uint8_t eltwise_op, bool eltwise_antq_en, bool c0_pad_en, bool broadcast_en, bool nz2dn_en) __aicore__ inline void asc_copy_l0c2ub_sync(__ubuf__ int8_t* dst_addr, __cc__ int32_t* src_addr, uint16_t n_size, uint16_t m_size, uint32_t loop_dst_stride, uint16_t loop_src_stride, uint8_t dual_dst_ctl, bool sub_blockid, uint8_t clip_relu_pre, uint8_t unit_flag_ctl, uint64_t quant_pre, uint8_t relu_pre, bool split_en, bool nz2nd_en, uint64_t quant_post, uint8_t relu_post, bool clip_relu_post, uint8_t eltwise_op, bool eltwise_antq_en, bool c0_pad_en, bool broadcast_en, bool nz2dn_en) __aicore__ inline void asc_copy_l0c2ub_sync(__ubuf__ uint8_t* dst_addr, __cc__ int32_t* src_addr, uint16_t n_size, uint16_t m_size, uint32_t loop_dst_stride, uint16_t loop_src_stride, uint8_t dual_dst_ctl, bool sub_blockid, uint8_t clip_relu_pre, uint8_t unit_flag_ctl, uint64_t quant_pre, uint8_t relu_pre, bool split_en, bool nz2nd_en, uint64_t quant_post, uint8_t relu_post, bool clip_relu_post, uint8_t eltwise_op, bool eltwise_antq_en, bool c0_pad_en, bool broadcast_en, bool nz2dn_en) __aicore__ inline void asc_copy_l0c2ub_sync(__ubuf__ int32_t* dst_addr, __cc__ int32_t* src_addr, uint16_t n_size, uint16_t m_size, uint32_t loop_dst_stride, uint16_t loop_src_stride, uint8_t dual_dst_ctl, bool sub_blockid, uint8_t clip_relu_pre, uint8_t unit_flag_ctl, uint64_t quant_pre, uint8_t relu_pre, bool split_en, bool nz2nd_en, uint64_t quant_post, uint8_t relu_post, bool clip_relu_post, uint8_t eltwise_op, bool eltwise_antq_en, bool c0_pad_en, bool broadcast_en, bool nz2dn_en) __aicore__ inline void asc_copy_l0c2ub_sync(__ubuf__ void* dst_addr, __cc__ float* src_addr, uint16_t n_size, uint16_t m_size, uint32_t loop_dst_stride, uint16_t loop_src_stride, uint8_t dual_dst_ctl, bool sub_blockid, uint8_t clip_relu_pre, uint8_t unit_flag_ctl, uint64_t quant_pre, uint8_t relu_pre, bool split_en, bool nz2nd_en, uint64_t quant_post, uint8_t relu_post, bool clip_relu_post, uint8_t eltwise_op, bool eltwise_antq_en, bool c0_pad_en, bool broadcast_en, bool nz2dn_en) __aicore__ inline void asc_copy_l0c2ub_sync(__ubuf__ void* dst_addr, __cc__ int32_t* src_addr, uint16_t n_size, uint16_t m_size, uint32_t loop_dst_stride, uint16_t loop_src_stride, uint8_t dual_dst_ctl, bool sub_blockid, uint8_t clip_relu_pre, uint8_t unit_flag_ctl, uint64_t quant_pre, uint8_t relu_pre, bool split_en, bool nz2nd_en, uint64_t quant_post, uint8_t relu_post, bool clip_relu_post, uint8_t eltwise_op, bool eltwise_antq_en, bool c0_pad_en, bool broadcast_en, bool nz2dn_en)参数说明参数名输入/输出描述dst_addr输出目的操作数矢量的起始地址。src_addr输入源操作数矢量的起始地址。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]。loop_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。loop_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_Size16*sizeof(T)T为src_addr的数据类型。dual_dst_ctl输入双目的控制参数。sub_blockid输入子块ID。clip_relu_pre输入预处理阶段开启clip_relu需搭配normal relu归一化的relu函数一起使用且需要开启量化功能。unit_flag_ctl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输入预处理阶段量化参数。取值见功能说明。relu_pre输入预处理阶段开启relu。split_en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_addr和dst_addr都为float时才能开启通道拆分且不能同时开启split_en和NZ2ND功能。nz2nd_en输入开启NZ2ND开关。•false不开启•true开启。quant_post输入后处理阶段量化参数。取值见功能说明。relu_post输入后处理阶段开启relu。clip_relu_post输入后处理阶段开启clip_relu需搭配normal relu一起使用且需要开启量化功能。eltwise_op输入定义数据从l0c搬运至ub时的目的操作数地址和通道步长。eltwise_antq_en输入按位开启元素的反量化操作。c0_pad_en输入开启为C0配置填充位C0是通道循环的目标步长。broadcast_en输入是否开启广播能力。•false不开启•true开启在数据搬运时沿M轴方向进行数据广播。nz2dn_en输入开启NZ2DN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_addr的起始地址要求按照对应数据类型所占字节数对齐。dst_addr的起始地址要求32字节对齐。调用示例__ubuf__ bfloat16_t dst[256]; __cc__ float src[256]; uint16_t n_size 1; uint16_t m_size 1; uint32_t loop_dst_stride 0; uint16_t loop_src_stride 0; uint8_t dual_dst_ctl 5; bool sub_blockid true; uint8_t clip_relu_pre 0; uint8_t unit_flag_ctl 0; uint64_t quant_pre QuantMode_t::F322BF16; uint8_t relu_pre 0; bool split_en false; bool nz2nd_en true; uint64_t quant_post QuantMode_post::NoConv; uint8_t relu_post 0; bool clip_relu_post false; uint8_t eltwise_op 0; bool eltwise_antq_en true; bool c0_pad_en true; bool broadcast_en false; bool nz2dn_en false; asc_copy_l0c2ub(dst, src, n_size, m_size, loop_dst_stride, loop_src_stride, dual_dst_ctl, sub_blockid, clip_relu_pre, unit_flag_ctl, quant_pre, relu_pre, split_en, nz2nd_en, quant_post, relu_post, clip_relu_post, eltwise_op, eltwise_antq_en, c0_pad_en, broadcast_en, nz2dn_en);【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