Star638
910
代码
代码
Issues122
Pull Requests164
Actions
项目讨论
Wiki
分析
Star638
910
cann-robotcann-robot优化矩阵搬出类接口入参命名，优化资料文档
48d3eb53创建于 1 天前历史提交

asc_copy_l0c2ub

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

该接口用于将矩阵计算的结果从L0C Buffer搬运至Unified Buffer（UB）中，并且在搬运过程中支持与其它接口组合使用，配置多种随路能力。

下图展示了随路量化、随路ReLU、随路格式转换、随路通道拆分以及随路通道合并的有效组合、中间数据类型和数据路径。下图中的F32->F16与F32->BF16为Cast，其余为随路scalar/tensor量化模式。

图1 asc_copy_l0c2ub随路功能组合

quant_pre_mode量化模式参数支持的枚举值如下：

  • NoQuant：不开启量化功能
  • DEQF16：int32_t量化成half， scalar量化
  • VDEQF16：int32_t量化成half，tensor量化
  • REQ4：int32_t量化成int4b_t，scalar量化
  • VREQ4：int32_t量化成int4b_t，tensor量化
  • REQ8：int32_t量化成int8_t/uint8_t，scalar量化
  • VREQ8：int32_t量化成int8_t/uint8_t，tensor量化
  • QS322BF16_PRE：int32_t量化成bfloat16_t，scalar量化
  • VQS322BF16_PRE：int32_t量化成bfloat16_t，tensor量化
  • QF322F16_PRE：float量化成half，scalar量化
  • VQF322F16_PRE：float量化成half，tensor量化
  • QF322BF16_PRE：float量化成bfloat16_t，scalar量化
  • VQF322BF16_PRE：float量化成bfloat16_t，tensor量化
  • F322F16：float cast成half，cast mode为CAST_RINT模式
  • F322BF16：float cast成bfloat16_t，cast mode为CAST_RINT模式
  • QF322S4_PRE：float量化成int4b_t，scalar量化
  • VQF322S4_PRE：float量化成int4b_t，tensor量化
  • QF322B8_PRE：float量化成int8_t/uint8_t，scalar量化
  • VQF322B8_PRE：float量化成int8_t/uint8_t，tensor量化
  • QF322FP8_PRE：float量化成fp8_e4m3fn_t，scalar量化
  • VQF322FP8_PRE：float量化成fp8_e4m3fn_t，tensor量化
  • QF322HIF8_PRE：float量化成hifloat8_t(Half to Away Round)，scalar量化
  • VQF322HIF8_PRE：float量化成hifloat8_t(Half to Away Round)，tensor量化
  • QF322HIF8_PRE_HYBRID：float量化成hifloat8_t(Hybrid Round)，scalar量化
  • VQF322HIF8_PRE_HYBRID：float量化成hifloat8_t(Hybrid Round)，tensor量化
  • QF322F32_PRE：float量化成float，scalar量化，精度可以达到双千分之一，无法达到双万分之一。如果有双万分之一的精度要求，建议使用AscendDeQuant高阶API。
  • VQF322F32_PRE：float量化成float，tensor量化，精度可以达到双千分之一，无法达到双万分之一。如果有双万分之一的精度要求，建议使用AscendDeQuant高阶API。

函数原型

  • 常规搬运

    __aicore__ inline void asc_copy_l0c2ub(
    __ubuf__ bfloat16_t* dst, __cc__ float* src, uint16_t n_size, uint16_t m_size, uint32_t dst_stride,
    uint16_t src_stride, uint8_t dual_dst_ctrl, bool sub_blockid, uint8_t enable_clip_relu_pre,
    uint8_t unit_flag_mode, uint64_t quant_pre_mode, uint8_t relu_pre_mode, bool enable_channel_split,
    bool enable_nz2nd, 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 enable_nz2dn)
__aicore__ inline void asc_copy_l0c2ub(
    __ubuf__ half* dst, __cc__ float* src, uint16_t n_size, uint16_t m_size, uint32_t dst_stride,
    uint16_t src_stride, uint8_t dual_dst_ctrl, bool sub_blockid, uint8_t enable_clip_relu_pre,
    uint8_t unit_flag_mode, uint64_t quant_pre_mode, uint8_t relu_pre_mode, bool enable_channel_split,
    bool enable_nz2nd, 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 enable_nz2dn)
__aicore__ inline void asc_copy_l0c2ub(
    __ubuf__ fp8_e4m3fn_t* dst, __cc__ float* src, uint16_t n_size, uint16_t m_size, uint32_t dst_stride,
    uint16_t src_stride, uint8_t dual_dst_ctrl, bool sub_blockid, uint8_t enable_clip_relu_pre,
    uint8_t unit_flag_mode, uint64_t quant_pre_mode, uint8_t relu_pre_mode, bool enable_channel_split,
    bool enable_nz2nd, 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 enable_nz2dn)
__aicore__ inline void asc_copy_l0c2ub(
    __ubuf__ hifloat8_t* dst, __cc__ float* src, uint16_t n_size, uint16_t m_size, uint32_t dst_stride,
    uint16_t src_stride, uint8_t dual_dst_ctrl, bool sub_blockid, uint8_t enable_clip_relu_pre,
    uint8_t unit_flag_mode, uint64_t quant_pre_mode, uint8_t relu_pre_mode, bool enable_channel_split,
    bool enable_nz2nd, 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 enable_nz2dn)
__aicore__ inline void asc_copy_l0c2ub(
    __ubuf__ int8_t* dst, __cc__ float* src, uint16_t n_size, uint16_t m_size, uint32_t dst_stride,
    uint16_t src_stride, uint8_t dual_dst_ctrl, bool sub_blockid, uint8_t enable_clip_relu_pre,
    uint8_t unit_flag_mode, uint64_t quant_pre_mode, uint8_t relu_pre_mode, bool enable_channel_split,
    bool enable_nz2nd, 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 enable_nz2dn)
__aicore__ inline void asc_copy_l0c2ub(
    __ubuf__ uint8_t* dst, __cc__ float* src, uint16_t n_size, uint16_t m_size, uint32_t dst_stride,
    uint16_t src_stride, uint8_t dual_dst_ctrl, bool sub_blockid, uint8_t enable_clip_relu_pre,
    uint8_t unit_flag_mode, uint64_t quant_pre_mode, uint8_t relu_pre_mode, bool enable_channel_split,
    bool enable_nz2nd, 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 enable_nz2dn)
__aicore__ inline void asc_copy_l0c2ub(
    __ubuf__ float* dst, __cc__ float* src, uint16_t n_size, uint16_t m_size, uint32_t dst_stride,
    uint16_t src_stride, uint8_t dual_dst_ctrl, bool sub_blockid, uint8_t enable_clip_relu_pre,
    uint8_t unit_flag_mode, uint64_t quant_pre_mode, uint8_t relu_pre_mode, bool enable_channel_split,
    bool enable_nz2nd, 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 enable_nz2dn)
__aicore__ inline void asc_copy_l0c2ub(
    __ubuf__ bfloat16_t* dst, __cc__ int32_t* src, uint16_t n_size, uint16_t m_size, uint32_t dst_stride,
    uint16_t src_stride, uint8_t dual_dst_ctrl, bool sub_blockid, uint8_t enable_clip_relu_pre,
    uint8_t unit_flag_mode, uint64_t quant_pre_mode, uint8_t relu_pre_mode, bool enable_channel_split,
    bool enable_nz2nd, 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 enable_nz2dn)
__aicore__ inline void asc_copy_l0c2ub(
    __ubuf__ half* dst, __cc__ int32_t* src, uint16_t n_size, uint16_t m_size, uint32_t dst_stride,
    uint16_t src_stride, uint8_t dual_dst_ctrl, bool sub_blockid, uint8_t enable_clip_relu_pre,
    uint8_t unit_flag_mode, uint64_t quant_pre_mode, uint8_t relu_pre_mode, bool enable_channel_split,
    bool enable_nz2nd, 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 enable_nz2dn)
__aicore__ inline void asc_copy_l0c2ub(
    __ubuf__ 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 dual_dst_ctrl, bool sub_blockid, uint8_t enable_clip_relu_pre,
    uint8_t unit_flag_mode, uint64_t quant_pre_mode, uint8_t relu_pre_mode, bool enable_channel_split,
    bool enable_nz2nd, 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 enable_nz2dn)
__aicore__ inline void asc_copy_l0c2ub(
    __ubuf__ 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 dual_dst_ctrl, bool sub_blockid, uint8_t enable_clip_relu_pre,
    uint8_t unit_flag_mode, uint64_t quant_pre_mode, uint8_t relu_pre_mode, bool enable_channel_split,
    bool enable_nz2nd, 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 enable_nz2dn)
__aicore__ inline void asc_copy_l0c2ub(
    __ubuf__ int32_t* dst, __cc__ int32_t* src, uint16_t n_size, uint16_t m_size, uint32_t dst_stride,
    uint16_t src_stride, uint8_t dual_dst_ctrl, bool sub_blockid, uint8_t enable_clip_relu_pre,
    uint8_t unit_flag_mode, uint64_t quant_pre_mode, uint8_t relu_pre_mode, bool enable_channel_split,
    bool enable_nz2nd, 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 enable_nz2dn)
__aicore__ inline void asc_copy_l0c2ub(
    __ubuf__ int4b_t* dst, __cc__ float* src, uint16_t n_size, uint16_t m_size, uint32_t dst_stride,
    uint16_t src_stride, uint8_t dual_dst_ctrl, bool sub_blockid, uint8_t enable_clip_relu_pre,
    uint8_t unit_flag_mode, uint64_t quant_pre_mode, uint8_t relu_pre_mode, bool enable_channel_split,
    bool enable_nz2nd, 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 enable_nz2dn)
__aicore__ inline void asc_copy_l0c2ub(
    __ubuf__ int4b_t* dst, __cc__ int32_t* src, uint16_t n_size, uint16_t m_size, uint32_t dst_stride,
    uint16_t src_stride, uint8_t dual_dst_ctrl, bool sub_blockid, uint8_t enable_clip_relu_pre,
    uint8_t unit_flag_mode, uint64_t quant_pre_mode, uint8_t relu_pre_mode, bool enable_channel_split,
    bool enable_nz2nd, 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 enable_nz2dn)

  • 同步搬运

    __aicore__ inline void asc_copy_l0c2ub_sync(
    __ubuf__ bfloat16_t* dst, __cc__ float* src, uint16_t n_size, uint16_t m_size, uint32_t dst_stride,
    uint16_t src_stride, uint8_t dual_dst_ctrl, bool sub_blockid, uint8_t enable_clip_relu_pre,
    uint8_t unit_flag_mode, uint64_t quant_pre_mode, uint8_t relu_pre_mode, bool enable_channel_split,
    bool enable_nz2nd, 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 enable_nz2dn)
__aicore__ inline void asc_copy_l0c2ub_sync(
    __ubuf__ half* dst, __cc__ float* src, uint16_t n_size, uint16_t m_size, uint32_t dst_stride,
    uint16_t src_stride, uint8_t dual_dst_ctrl, bool sub_blockid, uint8_t enable_clip_relu_pre,
    uint8_t unit_flag_mode, uint64_t quant_pre_mode, uint8_t relu_pre_mode, bool enable_channel_split,
    bool enable_nz2nd, 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 enable_nz2dn)
__aicore__ inline void asc_copy_l0c2ub_sync(
    __ubuf__ fp8_e4m3fn_t* dst, __cc__ float* src, uint16_t n_size, uint16_t m_size, uint32_t dst_stride,
    uint16_t src_stride, uint8_t dual_dst_ctrl, bool sub_blockid, uint8_t enable_clip_relu_pre,
    uint8_t unit_flag_mode, uint64_t quant_pre_mode, uint8_t relu_pre_mode, bool enable_channel_split,
    bool enable_nz2nd, 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 enable_nz2dn)
__aicore__ inline void asc_copy_l0c2ub_sync(
    __ubuf__ hifloat8_t* dst, __cc__ float* src, uint16_t n_size, uint16_t m_size, uint32_t dst_stride,
    uint16_t src_stride, uint8_t dual_dst_ctrl, bool sub_blockid, uint8_t enable_clip_relu_pre,
    uint8_t unit_flag_mode, uint64_t quant_pre_mode, uint8_t relu_pre_mode, bool enable_channel_split,
    bool enable_nz2nd, 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 enable_nz2dn)
__aicore__ inline void asc_copy_l0c2ub_sync(
    __ubuf__ int8_t* dst, __cc__ float* src, uint16_t n_size, uint16_t m_size, uint32_t dst_stride,
    uint16_t src_stride, uint8_t dual_dst_ctrl, bool sub_blockid, uint8_t enable_clip_relu_pre,
    uint8_t unit_flag_mode, uint64_t quant_pre_mode, uint8_t relu_pre_mode, bool enable_channel_split,
    bool enable_nz2nd, 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 enable_nz2dn)
__aicore__ inline void asc_copy_l0c2ub_sync(
    __ubuf__ uint8_t* dst, __cc__ float* src, uint16_t n_size, uint16_t m_size, uint32_t dst_stride,
    uint16_t src_stride, uint8_t dual_dst_ctrl, bool sub_blockid, uint8_t enable_clip_relu_pre,
    uint8_t unit_flag_mode, uint64_t quant_pre_mode, uint8_t relu_pre_mode, bool enable_channel_split,
    bool enable_nz2nd, 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 enable_nz2dn)
__aicore__ inline void asc_copy_l0c2ub_sync(
    __ubuf__ float* dst, __cc__ float* src, uint16_t n_size, uint16_t m_size, uint32_t dst_stride,
    uint16_t src_stride, uint8_t dual_dst_ctrl, bool sub_blockid, uint8_t enable_clip_relu_pre,
    uint8_t unit_flag_mode, uint64_t quant_pre_mode, uint8_t relu_pre_mode, bool enable_channel_split,
    bool enable_nz2nd, 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 enable_nz2dn)
__aicore__ inline void asc_copy_l0c2ub_sync(
    __ubuf__ bfloat16_t* dst, __cc__ int32_t* src, uint16_t n_size, uint16_t m_size, uint32_t dst_stride,
    uint16_t src_stride, uint8_t dual_dst_ctrl, bool sub_blockid, uint8_t enable_clip_relu_pre,
    uint8_t unit_flag_mode, uint64_t quant_pre_mode, uint8_t relu_pre_mode, bool enable_channel_split,
    bool enable_nz2nd, 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 enable_nz2dn)
__aicore__ inline void asc_copy_l0c2ub_sync(
    __ubuf__ half* dst, __cc__ int32_t* src, uint16_t n_size, uint16_t m_size, uint32_t dst_stride,
    uint16_t src_stride, uint8_t dual_dst_ctrl, bool sub_blockid, uint8_t enable_clip_relu_pre,
    uint8_t unit_flag_mode, uint64_t quant_pre_mode, uint8_t relu_pre_mode, bool enable_channel_split,
    bool enable_nz2nd, 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 enable_nz2dn)
__aicore__ inline void asc_copy_l0c2ub_sync(
    __ubuf__ 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 dual_dst_ctrl, bool sub_blockid, uint8_t enable_clip_relu_pre,
    uint8_t unit_flag_mode, uint64_t quant_pre_mode, uint8_t relu_pre_mode, bool enable_channel_split,
    bool enable_nz2nd, 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 enable_nz2dn)
__aicore__ inline void asc_copy_l0c2ub_sync(
    __ubuf__ 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 dual_dst_ctrl, bool sub_blockid, uint8_t enable_clip_relu_pre,
    uint8_t unit_flag_mode, uint64_t quant_pre_mode, uint8_t relu_pre_mode, bool enable_channel_split,
    bool enable_nz2nd, 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 enable_nz2dn)
__aicore__ inline void asc_copy_l0c2ub_sync(
    __ubuf__ int32_t* dst, __cc__ int32_t* src, uint16_t n_size, uint16_t m_size, uint32_t dst_stride,
    uint16_t src_stride, uint8_t dual_dst_ctrl, bool sub_blockid, uint8_t enable_clip_relu_pre,
    uint8_t unit_flag_mode, uint64_t quant_pre_mode, uint8_t relu_pre_mode, bool enable_channel_split,
    bool enable_nz2nd, 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 enable_nz2dn)
__aicore__ inline void asc_copy_l0c2ub_sync(
    __ubuf__ int4b_t* dst, __cc__ float* src, uint16_t n_size, uint16_t m_size, uint32_t dst_stride,
    uint16_t src_stride, uint8_t dual_dst_ctrl, bool sub_blockid, uint8_t enable_clip_relu_pre,
    uint8_t unit_flag_mode, uint64_t quant_pre_mode, uint8_t relu_pre_mode, bool enable_channel_split,
    bool enable_nz2nd, 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 enable_nz2dn)
__aicore__ inline void asc_copy_l0c2ub_sync(
    __ubuf__ int4b_t* dst, __cc__ int32_t* src, uint16_t n_size, uint16_t m_size, uint32_t dst_stride,
    uint16_t src_stride, uint8_t dual_dst_ctrl, bool sub_blockid, uint8_t enable_clip_relu_pre,
    uint8_t unit_flag_mode, uint64_t quant_pre_mode, uint8_t relu_pre_mode, bool enable_channel_split,
    bool enable_nz2nd, 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 enable_nz2dn)

参数说明

表1 参数说明

参数名 输入/输出 描述
dst 输出 目的操作数的起始地址。数据格式为NZ、ND或DN格式，地址需要满足32字节对齐。
src 输入 源操作数的起始地址。数据格式为NZ格式，地址需要满足64字节对齐。
n_size 输入 源NZ矩阵在N方向上的大小，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   • 若不开启NZ2ND功能，dst_stride表示目的NZ矩阵中相邻Z排布的起始地址偏移，取值不为0，单位：element。
  • 若开启NZ2ND/NZ2DN功能，dst_stride表示目的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)。
dual_dst_ctrl 输入 双目标模式控制参数。当同一AI Core内有一个Cube Core和两个Vector Core时，如果启用双目标模式控制，L0C Buffer中的M * N矩阵将被分成两半，并同时分别写入两个Vector Core各自的UB中，其中前半部分写入SUB BLOCK0，后半部分写入SUB BLOCK1，参数取值说明如下：
  • 0：单目标模式，将整个矩阵写入通过subBlockId参数配置的目标UB。
  • 1：双目标模式，按M维度拆分成形状为M / 2 * N的两个矩阵，分别写入两个UB, M必须为2的倍数。
  • 2：双目标模式，按N维度拆分成形状为M * N / 2的两个矩阵，分别写入两个UB, N须为32的倍数。
dualDstCtrl仅支持在普通搬运模式（NZ2NZ）或NZ2ND搬运场景下使用，不支持其他随路功能场景。
sub_blockid 输入 启用单目标模式时用于指示目标UB的SUB BLOCK ID。取值为0或1，取值为0时写入SUB BLOCK0，为1时写入SUB BLOCK1。
enable_clip_relu_pre 输入 是否开启Clip ReLU，需搭配Normal ReLU一起使用，且需要开启量化功能，取值如下：
  • 0：不开启Clip ReLU；
  • 1：开启Clip ReLU（scalar 模式）。
unit_flag_mode 输入 unit_flag是mmad类指令和矩阵搬出类指令细粒度的并行功能，开启该功能后，硬件每计算完一个分形，计算结果就会被搬出。取值说明如下：
  • 0：不开启unit_flag；
  • 2：开启unit_flag，硬件执行完指令之后，不复位单元标记位；
  • 3：开启unitFlag，硬件执行完指令之后，复位单元标记位。
开启该功能时，须将mmad类指令和矩阵搬出类指令的unit_flag值设置为2或3。
quant_pre_mode 输入 预处理阶段量化参数。取值见功能说明。
relu_pre_mode 输入 预处理阶段ReLU模式控制，取值如下：
  • 0：不开启ReLU；
  • 1：开启Normal ReLU；
  • 2：开启Scalar ReLU；
  • 3：开启Vector ReLU。
enable_channel_split 输入 是否开启通道拆分的功能。仅在src和dst都为float时才能开启Channel Split，且不能同时开启Channel Split和NZ2ND功能。
  • false：不开启；
  • true：开启。
enable_nz2nd 输入 是否开启NZ2ND功能。
  • false：不开启；
  • true：开启。
quant_post 输入 无效参数，用户无需关注，传入0即可。
relu_post 输入 无效参数，用户无需关注，传入0即可。
clip_relu_post 输入 无效参数，用户无需关注，传入false即可。
eltwise_op 输入 无效参数，用户无需关注，传入0即可。
eltwise_antq_en 输入 无效参数，用户无需关注，传入false即可。
c0_pad_en 输入 无效参数，用户无需关注，传入false即可。
broadcast_en 输入 无效参数，用户无需关注，传入false即可。
enable_nz2dn 输入 是否开启NZ2DN功能。
  • false：不开启；
  • true：开启。

返回值说明

无

流水类型

PIPE_FIX

约束说明

调用示例

// dst和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 = 128 * 64;
__ubuf__ float dst[total_length];
__cc__ float src[total_length];
uint16_t n_size =   64;
uint16_t m_size = 128;
uint32_t dst_stride = 2048;
uint16_t src_stride = 128;
uint8_t dual_dst_ctrl = 0;
bool sub_blockid = false;
uint8_t enable_clip_relu_pre = 0;
uint8_t unit_flag_mode = 0;
uint64_t quant_pre_mode = QuantMode_t::NoQuant;
uint8_t relu_pre_mode = 0;
bool enable_channel_split = false;
bool enable_nz2nd = false;
bool enable_nz2dn = false;
uint64_t quant_post = 0;  // 无效参数
uint8_t relu_post = 0;  // 无效参数
bool clip_relu_post = false;  // 无效参数
uint8_t eltwise_op = 0;  // 无效参数
bool eltwise_antq_en = false;  // 无效参数
bool c0_pad_en = false;  // 无效参数
bool broadcast_en = false;  // 无效参数
asc_copy_l0c2ub(dst, src, n_size, m_size, dst_stride, src_stride, dual_dst_ctrl, sub_blockid,
    enable_clip_relu_pre, unit_flag_mode, quant_pre_mode, relu_pre_mode, enable_channel_split,
    enable_nz2nd, quant_post, relu_post, clip_relu_post, eltwise_op, eltwise_antq_en,
    c0_pad_en, broadcast_en, enable_nz2dn);