asc_copy_l12l0a【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√功能说明用于搬运存放在L1 Buffer里的512B大小的矩阵到LOA Buffer里包含2D格式搬运、2D格式转置搬运、3D格式搬运。函数原型常规搬运 2D格式__aicore__ inline void asc_copy_l12l0a(__ca__ bfloat16_t* dst, __cbuf__ bfloat16_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a(__ca__ fp8_e4m3fn_t* dst, __cbuf__ fp8_e4m3fn_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a(__ca__ fp8_e5m2_t* dst, __cbuf__ fp8_e5m2_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a(__ca__ half* dst, __cbuf__ half* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a(__ca__ float* dst, __cbuf__ float* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a(__ca__ hifloat8_t* dst, __cbuf__ hifloat8_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a(__ca__ int16_t* dst, __cbuf__ int16_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a(__ca__ int32_t* dst, __cbuf__ int32_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a(__ca__ int8_t* dst, __cbuf__ int8_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a(__ca__ uint16_t* dst, __cbuf__ uint16_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a(__ca__ uint32_t* dst, __cbuf__ uint32_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a(__ca__ uint8_t* dst, __cbuf__ uint8_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a(__ca__ int4b_t* dst, __cbuf__ int4b_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a(__ca__ fp4x2_e2m1_t* dst, __cbuf__ fp4x2_e2m1_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a(__ca__ fp4x2_e1m2_t* dst, __cbuf__ fp4x2_e1m2_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride)同步常规搬运 2D格式__aicore__ inline void asc_copy_l12l0a_sync(__ca__ bfloat16_t* dst, __cbuf__ bfloat16_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a_sync(__ca__ fp8_e4m3fn_t* dst, __cbuf__ fp8_e4m3fn_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a_sync(__ca__ fp8_e5m2_t* dst, __cbuf__ fp8_e5m2_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a_sync(__ca__ half* dst, __cbuf__ half* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a_sync(__ca__ float* dst, __cbuf__ float* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a_sync(__ca__ hifloat8_t* dst, __cbuf__ hifloat8_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a_sync(__ca__ int16_t* dst, __cbuf__ int16_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a_sync(__ca__ int32_t* dst, __cbuf__ int32_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a_sync(__ca__ int8_t* dst, __cbuf__ int8_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a_sync(__ca__ uint16_t* dst, __cbuf__ uint16_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a_sync(__ca__ uint32_t* dst, __cbuf__ uint32_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a_sync(__ca__ uint8_t* dst, __cbuf__ uint8_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a_sync(__ca__ int4b_t* dst, __cbuf__ int4b_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a_sync(__ca__ fp4x2_e2m1_t* dst, __cbuf__ fp4x2_e2m1_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a_sync(__ca__ fp4x2_e1m2_t* dst, __cbuf__ fp4x2_e1m2_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride)转置搬运 2D格式__aicore__ inline void asc_copy_l12l0a_transpose(__ca__ bfloat16_t* dst, __cbuf__ bfloat16_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a_transpose(__ca__ fp8_e4m3fn_t* dst, __cbuf__ fp8_e4m3fn_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a_transpose(__ca__ fp8_e5m2_t* dst, __cbuf__ fp8_e5m2_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a_transpose(__ca__ half* dst, __cbuf__ half* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a_transpose(__ca__ float* dst, __cbuf__ float* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a_transpose(__ca__ hifloat8_t* dst, __cbuf__ hifloat8_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a_transpose(__ca__ int16_t* dst, __cbuf__ int16_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a_transpose(__ca__ int32_t* dst, __cbuf__ int32_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a_transpose(__ca__ int8_t* dst, __cbuf__ int8_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a_transpose(__ca__ uint16_t* dst, __cbuf__ uint16_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a_transpose(__ca__ uint32_t* dst, __cbuf__ uint32_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a_transpose(__ca__ uint8_t* dst, __cbuf__ uint8_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a_transpose(__ca__ int4b_t* dst, __cbuf__ int4b_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a_transpose(__ca__ fp4x2_e2m1_t* dst, __cbuf__ fp4x2_e2m1_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a_transpose(__ca__ fp4x2_e1m2_t* dst, __cbuf__ fp4x2_e1m2_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride)同步转置搬运 2D格式__aicore__ inline void asc_copy_l12l0a_transpose_sync(__ca__ bfloat16_t* dst, __cbuf__ bfloat16_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a_transpose_sync(__ca__ fp8_e4m3fn_t* dst, __cbuf__ fp8_e4m3fn_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a_transpose_sync(__ca__ fp8_e5m2_t* dst, __cbuf__ fp8_e5m2_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a_transpose_sync(__ca__ half* dst, __cbuf__ half* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a_transpose_sync(__ca__ float* dst, __cbuf__ float* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a_transpose_sync(__ca__ hifloat8_t* dst, __cbuf__ hifloat8_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a_transpose_sync(__ca__ int16_t* dst, __cbuf__ int16_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a_transpose_sync(__ca__ int32_t* dst, __cbuf__ int32_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a_transpose_sync(__ca__ int8_t* dst, __cbuf__ int8_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a_transpose_sync(__ca__ uint16_t* dst, __cbuf__ uint16_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a_transpose_sync(__ca__ uint32_t* dst, __cbuf__ uint32_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a_transpose_sync(__ca__ uint8_t* dst, __cbuf__ uint8_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a_transpose_sync(__ca__ int4b_t* dst, __cbuf__ int4b_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a_transpose_sync(__ca__ fp4x2_e2m1_t* dst, __cbuf__ fp4x2_e2m1_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride) __aicore__ inline void asc_copy_l12l0a_transpose_sync(__ca__ fp4x2_e1m2_t* dst, __cbuf__ fp4x2_e1m2_t* src, uint16_t m_start_position, uint16_t k_start_position, uint8_t m_step, uint8_t k_step, int16_t src_stride, uint16_t dst_stride)高维切分搬运 3D格式__aicore__ inline void asc_copy_l12l0a(__ca__ int8_t* dst, __cbuf__ int8_t* src, uint16_t k_extension, uint16_t m_extension, uint16_t k_start_pt, uint16_t m_start_pt, uint8_t stride_w, uint8_t stride_h, uint8_t filter_w, uint8_t filter_h, uint8_t dilation_filter_w, uint8_t dilation_filter_h, bool filter_size_w, bool filter_size_h, bool transpose, bool f_matrix_ctrl, uint16_t channel_size) __aicore__ inline void asc_copy_l12l0a(__ca__ uint8_t* dst, __cbuf__ uint8_t* src, uint16_t k_extension, uint16_t m_extension, uint16_t k_start_pt, uint16_t m_start_pt, uint8_t stride_w, uint8_t stride_h, uint8_t filter_w, uint8_t filter_h, uint8_t dilation_filter_w, uint8_t dilation_filter_h, bool filter_size_w, bool filter_size_h, bool transpose, bool f_matrix_ctrl, uint16_t channel_size) __aicore__ inline void asc_copy_l12l0a(__ca__ fp8_e4m3fn_t* dst, __cbuf__ fp8_e4m3fn_t* src, uint16_t k_extension, uint16_t m_extension, uint16_t k_start_pt, uint16_t m_start_pt, uint8_t stride_w, uint8_t stride_h, uint8_t filter_w, uint8_t filter_h, uint8_t dilation_filter_w, uint8_t dilation_filter_h, bool filter_size_w, bool filter_size_h, bool transpose, bool f_matrix_ctrl, uint16_t channel_size) __aicore__ inline void asc_copy_l12l0a(__ca__ fp8_e5m2_t* dst, __cbuf__ fp8_e5m2_t* src, uint16_t k_extension, uint16_t m_extension, uint16_t k_start_pt, uint16_t m_start_pt, uint8_t stride_w, uint8_t stride_h, uint8_t filter_w, uint8_t filter_h, uint8_t dilation_filter_w, uint8_t dilation_filter_h, bool filter_size_w, bool filter_size_h, bool transpose, bool f_matrix_ctrl, uint16_t channel_size) __aicore__ inline void asc_copy_l12l0a(__ca__ hifloat8_t* dst, __cbuf__ hifloat8_t* src, uint16_t k_extension, uint16_t m_extension, uint16_t k_start_pt, uint16_t m_start_pt, uint8_t stride_w, uint8_t stride_h, uint8_t filter_w, uint8_t filter_h, uint8_t dilation_filter_w, uint8_t dilation_filter_h, bool filter_size_w, bool filter_size_h, bool transpose, bool f_matrix_ctrl, uint16_t channel_size) __aicore__ inline void asc_copy_l12l0a(__ca__ int16_t* dst, __cbuf__ int16_t* src, uint16_t k_extension, uint16_t m_extension, uint16_t k_start_pt, uint16_t m_start_pt, uint8_t stride_w, uint8_t stride_h, uint8_t filter_w, uint8_t filter_h, uint8_t dilation_filter_w, uint8_t dilation_filter_h, bool filter_size_w, bool filter_size_h, bool transpose, bool f_matrix_ctrl, uint16_t channel_size) __aicore__ inline void asc_copy_l12l0a(__ca__ uint16_t* dst, __cbuf__ uint16_t* src, uint16_t k_extension, uint16_t m_extension, uint16_t k_start_pt, uint16_t m_start_pt, uint8_t stride_w, uint8_t stride_h, uint8_t filter_w, uint8_t filter_h, uint8_t dilation_filter_w, uint8_t dilation_filter_h, bool filter_size_w, bool filter_size_h, bool transpose, bool f_matrix_ctrl, uint16_t channel_size) __aicore__ inline void asc_copy_l12l0a(__ca__ bfloat16_t* dst, __cbuf__ bfloat16_t* src, uint16_t k_extension, uint16_t m_extension, uint16_t k_start_pt, uint16_t m_start_pt, uint8_t stride_w, uint8_t stride_h, uint8_t filter_w, uint8_t filter_h, uint8_t dilation_filter_w, uint8_t dilation_filter_h, bool filter_size_w, bool filter_size_h, bool transpose, bool f_matrix_ctrl, uint16_t channel_size) __aicore__ inline void asc_copy_l12l0a(__ca__ half* dst, __cbuf__ half* src, uint16_t k_extension, uint16_t m_extension, uint16_t k_start_pt, uint16_t m_start_pt, uint8_t stride_w, uint8_t stride_h, uint8_t filter_w, uint8_t filter_h, uint8_t dilation_filter_w, uint8_t dilation_filter_h, bool filter_size_w, bool filter_size_h, bool transpose, bool f_matrix_ctrl, uint16_t channel_size) __aicore__ inline void asc_copy_l12l0a(__ca__ int32_t* dst, __cbuf__ int32_t* src, uint16_t k_extension, uint16_t m_extension, uint16_t k_start_pt, uint16_t m_start_pt, uint8_t stride_w, uint8_t stride_h, uint8_t filter_w, uint8_t filter_h, uint8_t dilation_filter_w, uint8_t dilation_filter_h, bool filter_size_w, bool filter_size_h, bool transpose, bool f_matrix_ctrl, uint16_t channel_size) __aicore__ inline void asc_copy_l12l0a(__ca__ uint32_t* dst, __cbuf__ uint32_t* src, uint16_t k_extension, uint16_t m_extension, uint16_t k_start_pt, uint16_t m_start_pt, uint8_t stride_w, uint8_t stride_h, uint8_t filter_w, uint8_t filter_h, uint8_t dilation_filter_w, uint8_t dilation_filter_h, bool filter_size_w, bool filter_size_h, bool transpose, bool f_matrix_ctrl, uint16_t channel_size) __aicore__ inline void asc_copy_l12l0a(__ca__ float* dst, __cbuf__ float* src, uint16_t k_extension, uint16_t m_extension, uint16_t k_start_pt, uint16_t m_start_pt, uint8_t stride_w, uint8_t stride_h, uint8_t filter_w, uint8_t filter_h, uint8_t dilation_filter_w, uint8_t dilation_filter_h, bool filter_size_w, bool filter_size_h, bool transpose, bool f_matrix_ctrl, uint16_t channel_size)同步高维切分搬运 3D格式__aicore__ inline void asc_copy_l12l0a_sync(__ca__ int8_t* dst, __cbuf__ int8_t* src, uint16_t k_extension, uint16_t m_extension, uint16_t k_start_pt, uint16_t m_start_pt, uint8_t stride_w, uint8_t stride_h, uint8_t filter_w, uint8_t filter_h, uint8_t dilation_filter_w, uint8_t dilation_filter_h, bool filter_size_w, bool filter_size_h, bool transpose, bool f_matrix_ctrl, uint16_t channel_size) __aicore__ inline void asc_copy_l12l0a_sync(__ca__ uint8_t* dst, __cbuf__ uint8_t* src, uint16_t k_extension, uint16_t m_extension, uint16_t k_start_pt, uint16_t m_start_pt, uint8_t stride_w, uint8_t stride_h, uint8_t filter_w, uint8_t filter_h, uint8_t dilation_filter_w, uint8_t dilation_filter_h, bool filter_size_w, bool filter_size_h, bool transpose, bool f_matrix_ctrl, uint16_t channel_size) __aicore__ inline void asc_copy_l12l0a_sync(__ca__ fp8_e4m3fn_t* dst, __cbuf__ fp8_e4m3fn_t* src, uint16_t k_extension, uint16_t m_extension, uint16_t k_start_pt, uint16_t m_start_pt, uint8_t stride_w, uint8_t stride_h, uint8_t filter_w, uint8_t filter_h, uint8_t dilation_filter_w, uint8_t dilation_filter_h, bool filter_size_w, bool filter_size_h, bool transpose, bool f_matrix_ctrl, uint16_t channel_size) __aicore__ inline void asc_copy_l12l0a_sync(__ca__ fp8_e5m2_t* dst, __cbuf__ fp8_e5m2_t* src, uint16_t k_extension, uint16_t m_extension, uint16_t k_start_pt, uint16_t m_start_pt, uint8_t stride_w, uint8_t stride_h, uint8_t filter_w, uint8_t filter_h, uint8_t dilation_filter_w, uint8_t dilation_filter_h, bool filter_size_w, bool filter_size_h, bool transpose, bool f_matrix_ctrl, uint16_t channel_size) __aicore__ inline void asc_copy_l12l0a_sync(__ca__ hifloat8_t* dst, __cbuf__ hifloat8_t* src, uint16_t k_extension, uint16_t m_extension, uint16_t k_start_pt, uint16_t m_start_pt, uint8_t stride_w, uint8_t stride_h, uint8_t filter_w, uint8_t filter_h, uint8_t dilation_filter_w, uint8_t dilation_filter_h, bool filter_size_w, bool filter_size_h, bool transpose, bool f_matrix_ctrl, uint16_t channel_size) __aicore__ inline void asc_copy_l12l0a_sync(__ca__ int16_t* dst, __cbuf__ int16_t* src, uint16_t k_extension, uint16_t m_extension, uint16_t k_start_pt, uint16_t m_start_pt, uint8_t stride_w, uint8_t stride_h, uint8_t filter_w, uint8_t filter_h, uint8_t dilation_filter_w, uint8_t dilation_filter_h, bool filter_size_w, bool filter_size_h, bool transpose, bool f_matrix_ctrl, uint16_t channel_size) __aicore__ inline void asc_copy_l12l0a_sync(__ca__ uint16_t* dst, __cbuf__ uint16_t* src, uint16_t k_extension, uint16_t m_extension, uint16_t k_start_pt, uint16_t m_start_pt, uint8_t stride_w, uint8_t stride_h, uint8_t filter_w, uint8_t filter_h, uint8_t dilation_filter_w, uint8_t dilation_filter_h, bool filter_size_w, bool filter_size_h, bool transpose, bool f_matrix_ctrl, uint16_t channel_size) __aicore__ inline void asc_copy_l12l0a_sync(__ca__ bfloat16_t* dst, __cbuf__ bfloat16_t* src, uint16_t k_extension, uint16_t m_extension, uint16_t k_start_pt, uint16_t m_start_pt, uint8_t stride_w, uint8_t stride_h, uint8_t filter_w, uint8_t filter_h, uint8_t dilation_filter_w, uint8_t dilation_filter_h, bool filter_size_w, bool filter_size_h, bool transpose, bool f_matrix_ctrl, uint16_t channel_size) __aicore__ inline void asc_copy_l12l0a_sync(__ca__ half* dst, __cbuf__ half* src, uint16_t k_extension, uint16_t m_extension, uint16_t k_start_pt, uint16_t m_start_pt, uint8_t stride_w, uint8_t stride_h, uint8_t filter_w, uint8_t filter_h, uint8_t dilation_filter_w, uint8_t dilation_filter_h, bool filter_size_w, bool filter_size_h, bool transpose, bool f_matrix_ctrl, uint16_t channel_size) __aicore__ inline void asc_copy_l12l0a_sync(__ca__ int32_t* dst, __cbuf__ int32_t* src, uint16_t k_extension, uint16_t m_extension, uint16_t k_start_pt, uint16_t m_start_pt, uint8_t stride_w, uint8_t stride_h, uint8_t filter_w, uint8_t filter_h, uint8_t dilation_filter_w, uint8_t dilation_filter_h, bool filter_size_w, bool filter_size_h, bool transpose, bool f_matrix_ctrl, uint16_t channel_size) __aicore__ inline void asc_copy_l12l0a_sync(__ca__ uint32_t* dst, __cbuf__ uint32_t* src, uint16_t k_extension, uint16_t m_extension, uint16_t k_start_pt, uint16_t m_start_pt, uint8_t stride_w, uint8_t stride_h, uint8_t filter_w, uint8_t filter_h, uint8_t dilation_filter_w, uint8_t dilation_filter_h, bool filter_size_w, bool filter_size_h, bool transpose, bool f_matrix_ctrl, uint16_t channel_size) __aicore__ inline void asc_copy_l12l0a_sync(__ca__ float* dst, __cbuf__ float* src, uint16_t k_extension, uint16_t m_extension, uint16_t k_start_pt, uint16_t m_start_pt, uint8_t stride_w, uint8_t stride_h, uint8_t filter_w, uint8_t filter_h, uint8_t dilation_filter_w, uint8_t dilation_filter_h, bool filter_size_w, bool filter_size_h, bool transpose, bool f_matrix_ctrl, uint16_t channel_size)参数说明表1 2D格式参数说明 | 参数名 | 输入/输出 | 描述 | |:-----------------| :--- |:-------------------------------------------------------------------------------------------------------------------------------------------------------------| | dst | 输出 | 目的L0A Buffer地址。 | | src | 输入 | 源L1 Buffer地址。 | | m_start_position | 输入 | 以MK矩阵为例源矩阵M轴方向的起始位置单位为16个元素。 | | k_start_position | 输入 | 以MK矩阵为例源矩阵K轴方向的起始位置单位为32B。 | | m_step | 输入 | 以MK矩阵为例源矩阵M轴方向搬运长度单位为16个元素。取值范围[0, 255]。当进行转置搬运时还需满足以下额外约束数据位宽为4时m_step必须为4的倍数数据位宽为8时m_step必须为2的倍数数据位宽为16时m_step必须为1的倍数数据位宽为32时m_step无额外约束。 | | k_step | 输入 | 以MK矩阵为例源矩阵M轴方向搬运长度单位为32B。取值范围[0. 255]。当进行转置搬运时还需满足以下额外约束数据位宽为4、8或16时k_step没有额外约束数据位宽为32时k_step必须是2的倍数。 | | src_stride | 输入 | 以MK矩阵为例源矩阵K方向前一个分形起始地址与后一个分形起始地址的间隔单位为512B。 | | dst_stride | 输入 | 以MK矩阵为例目标矩阵K方向前一个分形起始地址与后一个分形起始地址的间隔单位为512B。 |表2 3D格式参数说明 | 参数名 | 输入/输出 | 描述 | | :--- | :--- | :--- | | dst | 输出 | 目的L0A Buffer地址。 | | src | 输入 | 源L1 Buffer地址。 | | k_extension | 输入 | 该指令在目的操作数width维度的传输长度。如果不覆盖最右侧的分形对于half类型应为16的倍数对于int8_t/uint8_t类型应为32的倍数如果覆盖最右侧的分形则无倍数要求。取值范围[1, 65535]。 | | m_extension | 输入 | 该指令在目的操作数height维度的传输长度。如果不覆盖最下侧的分形对于half/int8_t/uint8_t类型应为16的倍数如果覆盖最下侧的分形则无倍数要求。取值范围[1, 65535]。 | | k_start_pt | 输入 | 该指令在目的操作数width维度的起点。对于half类型应为16的倍数对于int8_t/uint8_t类型应为32的倍数。取值范围[0, 65535] | | m_start_pt | 输入 | 该指令在目的操作数height维度的起点如果不覆盖最下侧的分形对于half/int8_t/uint8_t应为16的倍数如果覆盖最下侧的分形则无倍数要求。取值范围[0, 65535] 。| | stride_w | 输入 | 卷积核在源操作数width维度滑动的步长取值范围[1, 63]。 | | stride_h | 输入 | 卷积核在源操作数height维度滑动的步长取值范围[1, 63]。 | | filter_w | 输入 | 卷积核width取值范围[1, 255]。 | | filter_h | 输入 | 卷积核height取值范围[1, 255]。 | | dilation_filter_w | 输入 | 卷积核width膨胀系数取值范围[1, 255]。 | | dilation_filter_h | 输入 | 卷积核height膨胀系数取值范围[1, 255]。 | | filter_size_w | 输入 | 是否在filter_w的基础上将卷积核width增加256个元素。true表示增加false表示不增加。 | | filter_size_h | 输入 | 是否在filter_h的基础上将卷积核height增加256个元素。true表示增加false表示不增加。 | | transpose | 输入 | 是否启用转置功能对整个目标矩阵进行转置仅在源操作数为half类型时有效。true表示启用false表示不启用。 | | f_matrix_ctrl | 输入 | 表示asc_copy_l12l0a指令从左矩阵还是右矩阵获取FeatureMap的属性描述当前只支持设置为false。 | | channel_size | 输入 | 源操作数的通道数取值范围[1, 63]。对于uint32_t/int32_t/floatchannelSize可取值为4N * 8N * 8 4对于half/bfloat16channelSize可取值为48N * 16N * 16 4N * 16 8对于int8_t/uint8_tchannelSize可取值为4816 32 * NN * 32 4N * 32 8N * 32 16对于int4b_tChannelSize可取值为81632N * 64N * 64 8N * 64 16N * 64 32。N为正整数。|返回值说明无流水类型PIPE_MTE1约束说明dst的起始地址需要512字节对齐src的起始地址需要32字节对齐。操作数地址重叠约束请参考通用地址重叠约束。3D数据格式说明要求输入的feature map和filter的格式是 NC1HWC0其中 C0 是最低维度而且 C0 是固定值为 16对于u8/s8类型为32C1C/C0。为了简化场景以下场景假设输入的 feature map 的 channel 为4即 Ci4。输入 feature maps 在 A1 中的形状为 (Hi,Wi,Ci)经过 load3dv1 处理后在 A2 的数据形状为(WoHo, HkWk*Ci)。其中 Wo 和 Ho 是卷积后输出的shapeHk 和 Wk 是 filter 的 shape。直观的来看img2col 的过程就是 filter 在 feature map 上扫过将对应 feature map 的数据展开成输出数据的每一行的过程。filter 首先在W方向上滑动 Wo 步然后在 H 方向上走一步然后重复以上过程最终输出 Wo * Ho 行数据。下图中红色和黄色的数据分别代表第一行和第二行。数字表示原始输入数据filter 和输出数据三者之间的关联关系。可以看到load3dv1 首先在输入数据的 Ci 维度搬运对应于 00 的 4 个数然后搬运对应于 01 的四个数最终这一行的大小为 HkWkCi 即 33436 个数。对应的feature map格式如下图对应的 filter 的格式如下图其中 n 为 filter 的个数可以看出维度排布为 (Hk,Wk,Ci,n)但是需要注意的是下图的格式还需要根据Mmad中B矩阵的格式转换。实际操作中由于存储空间或者计算能力限制我们通常会将整个卷积计算分块一次只搬运并计算一小块数据。对于 A2 的 feature map 来说有两种方案水平分块和垂直分块。分别对应参数中 repeatMode 的 0 和 1。注下图中的分形矩阵大小为 4x4实际应该为 16x16 (对于 u8/s8 类型为 16x32)repeatMode 0 时每次 repeat 会改变在 filter 窗口中读取数据点的位置然后跳到下一个 C0 的位置。repeatMode 1 的时候 filter 窗口中读取数据的位置保持不变每个 repeat 在 feature map 中前进 C0 个元素。调用示例__ca__ bfloat16_t dst[256]; __cbuf__ bfloat16_t src[256]; uint16_t m_start_position 8; uint16_t k_start_position 2; uint8_t m_step 4; uint8_t k_step 4; int16_t src_stride 1; uint16_t dst_stride 1; asc_copy_l12l0a(dst, src, m_start_position, k_start_position, m_step, k_step, src_stride, dst_stride);【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