asc_copy_gm2l1【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√功能说明将矩阵数据从Global Memory搬运到L1 Buffer中支持以下两种搬运模式填充模式搬运过程中可以对数据进行填充或移除填充值。2D搬运模式支持2D矩阵数据搬运支持压缩能力。NPU_ARCH 220x版本不支持.函数原型填充模式__aicore__ inline void asc_copy_gm2l1(__cbuf__ void* dst, __gm__ void* src, uint32_t n_burst, uint32_t len_burst, uint8_t pad_func_mode, uint64_t src_stride, uint32_t dst_stride)填充模式同步__aicore__ inline void asc_copy_gm2l1_sync(__cbuf__ void* dst, __gm__ void* src, uint32_t n_burst, uint32_t len_burst, uint8_t pad_func_mode, uint64_t src_stride, uint32_t dst_stride)2D搬运模式__aicore__ inline void asc_copy_gm2l1(__cbuf__ bfloat16_t *dst, __gm__ bfloat16_t *src, uint32_t m_start_position, uint32_t k_start_position, uint16_t dst_stride, uint16_t m_step, uint16_t k_step, uint8_t decomp_mode, uint8_t l2_cache_ctl) __aicore__ inline void asc_copy_gm2l1(__cbuf__ float *dst, __gm__ float *src, uint32_t m_start_position, uint32_t k_start_position, uint16_t dst_stride, uint16_t m_step, uint16_t k_step, uint8_t decomp_mode, uint8_t l2_cache_ctl) __aicore__ inline void asc_copy_gm2l1(__cbuf__ fp8_e4m3fn_t *dst, __gm__ fp8_e4m3fn_t *src, uint32_t m_start_position, uint32_t k_start_position, uint16_t dst_stride, uint16_t m_step, uint16_t k_step, uint8_t decomp_mode, uint8_t l2_cache_ctl) __aicore__ inline void asc_copy_gm2l1(__cbuf__ fp8_e5m2_t *dst, __gm__ fp8_e5m2_t *src, uint32_t m_start_position, uint32_t k_start_position, uint16_t dst_stride, uint16_t m_step, uint16_t k_step, uint8_t decomp_mode, uint8_t l2_cache_ctl) __aicore__ inline void asc_copy_gm2l1(__cbuf__ half *dst, __gm__ half *src, uint32_t m_start_position, uint32_t k_start_position, uint16_t dst_stride, uint16_t m_step, uint16_t k_step, uint8_t decomp_mode, uint8_t l2_cache_ctl) __aicore__ inline void asc_copy_gm2l1(__cbuf__ hifloat8_t *dst, __gm__ hifloat8_t *src, uint32_t m_start_position, uint32_t k_start_position, uint16_t dst_stride, uint16_t m_step, uint16_t k_step, uint8_t decomp_mode, uint8_t l2_cache_ctl) __aicore__ inline void asc_copy_gm2l1(__cbuf__ int16_t *dst, __gm__ int16_t *src, uint32_t m_start_position, uint32_t k_start_position, uint16_t dst_stride, uint16_t m_step, uint16_t k_step, uint8_t decomp_mode, uint8_t l2_cache_ctl) __aicore__ inline void asc_copy_gm2l1(__cbuf__ int32_t *dst, __gm__ int32_t *src, uint32_t m_start_position, uint32_t k_start_position, uint16_t dst_stride, uint16_t m_step, uint16_t k_step, uint8_t decomp_mode, uint8_t l2_cache_ctl) __aicore__ inline void asc_copy_gm2l1(__cbuf__ int8_t *dst, __gm__ int8_t *src, uint32_t m_start_position, uint32_t k_start_position, uint16_t dst_stride, uint16_t m_step, uint16_t k_step, uint8_t decomp_mode, uint8_t l2_cache_ctl) __aicore__ inline void asc_copy_gm2l1(__cbuf__ uint16_t *dst, __gm__ uint16_t *src, uint32_t m_start_position, uint32_t k_start_position, uint16_t dst_stride, uint16_t m_step, uint16_t k_step, uint8_t decomp_mode, uint8_t l2_cache_ctl) __aicore__ inline void asc_copy_gm2l1(__cbuf__ uint32_t *dst, __gm__ uint32_t *src, uint32_t m_start_position, uint32_t k_start_position, uint16_t dst_stride, uint16_t m_step, uint16_t k_step, uint8_t decomp_mode, uint8_t l2_cache_ctl) __aicore__ inline void asc_copy_gm2l1(__cbuf__ uint8_t *dst, __gm__ uint8_t *src, uint32_t m_start_position, uint32_t k_start_position, uint16_t dst_stride, uint16_t m_step, uint16_t k_step, uint8_t decomp_mode, uint8_t l2_cache_ctl) __aicore__ inline void asc_copy_gm2l1(__cbuf__ fp4x2_e1m2_t *dst, __gm__ fp4x2_e1m2_t *src, uint32_t m_start_position, uint32_t k_start_position, uint16_t dst_stride, uint16_t m_step, uint16_t k_step, uint8_t decomp_mode, uint8_t l2_cache_ctl) __aicore__ inline void asc_copy_gm2l1(__cbuf__ fp4x2_e2m1_t *dst, __gm__ fp4x2_e2m1_t *src, uint32_t m_start_position, uint32_t k_start_position, uint16_t dst_stride, uint16_t m_step, uint16_t k_step, uint8_t decomp_mode, uint8_t l2_cache_ctl) __aicore__ inline void asc_copy_gm2l1(__cbuf__ void *dst, __gm__ void *src, uint32_t m_start_position, uint32_t k_start_position, uint16_t dst_stride, uint16_t m_step, uint16_t k_step, uint8_t decomp_mode, uint8_t l2_cache_ctl)2D搬运模式同步__aicore__ inline void asc_copy_gm2l1_sync(__cbuf__ bfloat16_t *dst, __gm__ bfloat16_t *src, uint32_t m_start_position, uint32_t k_start_position, uint16_t dst_stride, uint16_t m_step, uint16_t k_step, uint8_t decomp_mode, uint8_t l2_cache_ctl) __aicore__ inline void asc_copy_gm2l1_sync(__cbuf__ float *dst, __gm__ float *src, uint32_t m_start_position, uint32_t k_start_position, uint16_t dst_stride, uint16_t m_step, uint16_t k_step, uint8_t decomp_mode, uint8_t l2_cache_ctl) __aicore__ inline void asc_copy_gm2l1_sync(__cbuf__ fp8_e4m3fn_t *dst, __gm__ fp8_e4m3fn_t *src, uint32_t m_start_position, uint32_t k_start_position, uint16_t dst_stride, uint16_t m_step, uint16_t k_step, uint8_t decomp_mode, uint8_t l2_cache_ctl) __aicore__ inline void asc_copy_gm2l1_sync(__cbuf__ fp8_e5m2_t *dst, __gm__ fp8_e5m2_t *src, uint32_t m_start_position, uint32_t k_start_position, uint16_t dst_stride, uint16_t m_step, uint16_t k_step, uint8_t decomp_mode, uint8_t l2_cache_ctl) __aicore__ inline void asc_copy_gm2l1_sync(__cbuf__ half *dst, __gm__ half *src, uint32_t m_start_position, uint32_t k_start_position, uint16_t dst_stride, uint16_t m_step, uint16_t k_step, uint8_t decomp_mode, uint8_t l2_cache_ctl) __aicore__ inline void asc_copy_gm2l1_sync(__cbuf__ hifloat8_t *dst, __gm__ hifloat8_t *src, uint32_t m_start_position, uint32_t k_start_position, uint16_t dst_stride, uint16_t m_step, uint16_t k_step, uint8_t decomp_mode, uint8_t l2_cache_ctl) __aicore__ inline void asc_copy_gm2l1_sync(__cbuf__ int16_t *dst, __gm__ int16_t *src, uint32_t m_start_position, uint32_t k_start_position, uint16_t dst_stride, uint16_t m_step, uint16_t k_step, uint8_t decomp_mode, uint8_t l2_cache_ctl) __aicore__ inline void asc_copy_gm2l1_sync(__cbuf__ int32_t *dst, __gm__ int32_t *src, uint32_t m_start_position, uint32_t k_start_position, uint16_t dst_stride, uint16_t m_step, uint16_t k_step, uint8_t decomp_mode, uint8_t l2_cache_ctl) __aicore__ inline void asc_copy_gm2l1_sync(__cbuf__ int8_t *dst, __gm__ int8_t *src, uint32_t m_start_position, uint32_t k_start_position, uint16_t dst_stride, uint16_t m_step, uint16_t k_step, uint8_t decomp_mode, uint8_t l2_cache_ctl) __aicore__ inline void asc_copy_gm2l1_sync(__cbuf__ uint16_t *dst, __gm__ uint16_t *src, uint32_t m_start_position, uint32_t k_start_position, uint16_t dst_stride, uint16_t m_step, uint16_t k_step, uint8_t decomp_mode, uint8_t l2_cache_ctl) __aicore__ inline void asc_copy_gm2l1_sync(__cbuf__ uint32_t *dst, __gm__ uint32_t *src, uint32_t m_start_position, uint32_t k_start_position, uint16_t dst_stride, uint16_t m_step, uint16_t k_step, uint8_t decomp_mode, uint8_t l2_cache_ctl) __aicore__ inline void asc_copy_gm2l1_sync(__cbuf__ uint8_t *dst, __gm__ uint8_t *src, uint32_t m_start_position, uint32_t k_start_position, uint16_t dst_stride, uint16_t m_step, uint16_t k_step, uint8_t decomp_mode, uint8_t l2_cache_ctl) __aicore__ inline void asc_copy_gm2l1_sync(__cbuf__ fp4x2_e1m2_t *dst, __gm__ fp4x2_e1m2_t *src, uint32_t m_start_position, uint32_t k_start_position, uint16_t dst_stride, uint16_t m_step, uint16_t k_step, uint8_t decomp_mode, uint8_t l2_cache_ctl) __aicore__ inline void asc_copy_gm2l1_sync(__cbuf__ fp4x2_e2m1_t *dst, __gm__ fp4x2_e2m1_t *src, uint32_t m_start_position, uint32_t k_start_position, uint16_t dst_stride, uint16_t m_step, uint16_t k_step, uint8_t decomp_mode, uint8_t l2_cache_ctl) __aicore__ inline void asc_copy_gm2l1_sync(__cbuf__ void *dst, __gm__ void *src, uint32_t m_start_position, uint32_t k_start_position, uint16_t dst_stride, uint16_t m_step, uint16_t k_step, uint8_t decomp_mode, uint8_t l2_cache_ctl)参数说明填充模式参数参数名输入/输出描述dst输出目的操作数矢量的起始地址。src输入源操作数矢量的起始地址。n_burst输入待搬运的连续传输数据块个数。len_burst输入待搬运的每个连续传输数据块的长度单位为32个字节。pad_func_mode输入一个4Bit的值用于指示通道上的padding功能他会在32个字节的数据中填充数据或从每个32个字节的读取数据中移除padding值。• 0不做padding。• 1每有1个字节的数据插入31个padding值。• 2每有2个字节的数据插入15个padding值。• 3每有4个字节的数据插入14个padding值。• 4每有8个字节的数据插入12个padding值。• 5每有16个字节的数据插入8个padding值。• 6每有32个字节的数据移除28个最高有效位数据只保留4个最低有效位元素。• 7每有32个字节的数据移除24个最高有效位数据只保留8个最低有效位元素。• 8每有32个字节的数据移除16个最高有效位数据只保留16个最低有效位元素。src_stride输入输入数据中两个相邻连续数据块之间的距离。dst_stride输入输出数据中两个相邻连续数据块之间的距离。2D搬运模式参数参数名输入/输出描述dst输出目的操作数矢量的起始地址。src输入源操作数矢量的起始地址。m_start_position输入M方向上的起始位置。k_start_position输入K方向上的起始位置。dst_stride输入目的操作数步长。m_step输入M方向上的步长。k_step输入K方向上的步长。decomp_mode输入设置不同的压缩模式。l2_cache_ctl输入配置数据在L2 Cache中的管理策略。取值说明如下• 0DISABLE模式适用于仅需访问一次的数据。• 1NORMAL模式适用于重用模式未知或不极端的数据。• 2LAST模式适用于高频重复访问的数据。• 4PERSISTENT模式适用于需要长期驻留在缓存中的数据。返回值说明无流水类型PIPE_MTE2约束说明填充模式约束对于插入padding值的场景len_burst的单位为32个字节的指的是写入L1 Buffer的实际数据数它必须设置为1此时一次仅从GM里读取1/2/4/8/16个字节的数据。每一次burst都是从GM中连续读取因此不需要设置src_stride。dst_stride以32个字节为单位。对于移除padding值的场景源操作数中一次burst操作的数据为len_burst * 32个字节目的操作数中一次burst的实际数据为len_burst * 4/8/16个字节。每一次burst后都连续地往L1 Buffer中写入因此不需要设置dst_stride。src_stride以32个字节为单位。2D搬运模式约束dst、src的起始地址需要32字节对齐。为了兼容性需要设置m_step为1。调用示例//待搬运的连续传输数据块个数为1 constexpr uint32_t n_burst 1; //待搬运的每个连续传输数据块的长度为128个字节 constexpr uint32_t len_burst 4; //不对数据做padding constexpr uint8_t pad_func_mode 0; //输入输出数据中两个相邻连续数据块之间的距离为0 constexpr uint64_t src_stride 0; constexpr uint32_t dst_stride 0; __gm__ half src[256]; __cbuf__ half dst[256]; asc_copy_gm2l1(dst, src, n_burst, len_burst, pad_func_mode, 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