asc_copy_gm2l1_dn2nz

【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"。

将矩阵数据从Global Memory搬运到L1 Buffer，在此过程中执行DN->NZ/NCHW->NC1HWC0/NCHW->C1HWNC0操作。若D未对齐C0的大小，将会在L1 Buffer的最内层维度填充0值。GM中的数据以DN类型存储；L1中的数据以NZ形式存储。

DN->NZ的搬运形式如下图：

函数原型

• 常规搬运

  __aicore__ inline void asc_copy_gm2l1_dn2nz(__cbuf__ int8_t* dst, __gm__ int8_t* src, uint64_t loop1_src_stride, uint8_t l2_cache_ctl, uint16_t n_value, uint32_t d_value, uint64_t loop4_src_stride, bool smallc0_en) __aicore__ inline void asc_copy_gm2l1_dn2nz(__cbuf__ uint8_t* dst, __gm__ uint8_t* src, uint64_t loop1_src_stride, uint8_t l2_cache_ctl, uint16_t n_value, uint32_t d_value, uint64_t loop4_src_stride, bool smallc0_en) __aicore__ inline void asc_copy_gm2l1_dn2nz(__cbuf__ hifloat8_t* dst, __gm__ hifloat8_t* src, uint64_t loop1_src_stride, uint8_t l2_cache_ctl, uint16_t n_value, uint32_t d_value, uint64_t loop4_src_stride, bool smallc0_en) __aicore__ inline void asc_copy_gm2l1_dn2nz(__cbuf__ fp8_e5m2_t* dst, __gm__ fp8_e5m2_t* src, uint64_t loop1_src_stride, uint8_t l2_cache_ctl, uint16_t n_value, uint32_t d_value, uint64_t loop4_src_stride, bool smallc0_en) __aicore__ inline void asc_copy_gm2l1_dn2nz(__cbuf__ fp8_e4m3fn_t* dst, __gm__ fp8_e4m3fn_t* src, uint64_t loop1_src_stride, uint8_t l2_cache_ctl, uint16_t n_value, uint32_t d_value, uint64_t loop4_src_stride, bool smallc0_en) __aicore__ inline void asc_copy_gm2l1_dn2nz(__cbuf__ int16_t* dst, __gm__ int16_t* src, uint64_t loop1_src_stride, uint8_t l2_cache_ctl, uint16_t n_value, uint32_t d_value, uint64_t loop4_src_stride, bool smallc0_en) __aicore__ inline void asc_copy_gm2l1_dn2nz(__cbuf__ uint16_t* dst, __gm__ uint16_t* src, uint64_t loop1_src_stride, uint8_t l2_cache_ctl, uint16_t n_value, uint32_t d_value, uint64_t loop4_src_stride, bool smallc0_en) __aicore__ inline void asc_copy_gm2l1_dn2nz(__cbuf__ half* dst, __gm__ half* src, uint64_t loop1_src_stride, uint8_t l2_cache_ctl, uint16_t n_value, uint32_t d_value, uint64_t loop4_src_stride, bool smallc0_en) __aicore__ inline void asc_copy_gm2l1_dn2nz(__cbuf__ bfloat16_t* dst, __gm__ bfloat16_t* src, uint64_t loop1_src_stride, uint8_t l2_cache_ctl, uint16_t n_value, uint32_t d_value, uint64_t loop4_src_stride, bool smallc0_en) __aicore__ inline void asc_copy_gm2l1_dn2nz(__cbuf__ int32_t* dst, __gm__ int32_t* src, uint64_t loop1_src_stride, uint8_t l2_cache_ctl, uint16_t n_value, uint32_t d_value, uint64_t loop4_src_stride, bool smallc0_en) __aicore__ inline void asc_copy_gm2l1_dn2nz(__cbuf__ uint32_t* dst, __gm__ uint32_t* src, uint64_t loop1_src_stride, uint8_t l2_cache_ctl, uint16_t n_value, uint32_t d_value, uint64_t loop4_src_stride, bool smallc0_en) __aicore__ inline void asc_copy_gm2l1_dn2nz(__cbuf__ float* dst, __gm__ float* src, uint64_t loop1_src_stride, uint8_t l2_cache_ctl, uint16_t n_value, uint32_t d_value, uint64_t loop4_src_stride, bool smallc0_en)

• 同步搬运

  __aicore__ inline void asc_copy_gm2l1_dn2nz_sync(__cbuf__ int8_t* dst, __gm__ int8_t* src, uint64_t loop1_src_stride, uint8_t l2_cache_ctl, uint16_t n_value, uint32_t d_value, uint64_t loop4_src_stride, bool smallc0_en) __aicore__ inline void asc_copy_gm2l1_dn2nz_sync(__cbuf__ uint8_t* dst, __gm__ uint8_t* src, uint64_t loop1_src_stride, uint8_t l2_cache_ctl, uint16_t n_value, uint32_t d_value, uint64_t loop4_src_stride, bool smallc0_en) __aicore__ inline void asc_copy_gm2l1_dn2nz_sync(__cbuf__ hifloat8_t* dst, __gm__ hifloat8_t* src, uint64_t loop1_src_stride, uint8_t l2_cache_ctl, uint16_t n_value, uint32_t d_value, uint64_t loop4_src_stride, bool smallc0_en) __aicore__ inline void asc_copy_gm2l1_dn2nz_sync(__cbuf__ fp8_e5m2_t* dst, __gm__ fp8_e5m2_t* src, uint64_t loop1_src_stride, uint8_t l2_cache_ctl, uint16_t n_value, uint32_t d_value, uint64_t loop4_src_stride, bool smallc0_en) __aicore__ inline void asc_copy_gm2l1_dn2nz_sync(__cbuf__ fp8_e4m3fn_t* dst, __gm__ fp8_e4m3fn_t* src, uint64_t loop1_src_stride, uint8_t l2_cache_ctl, uint16_t n_value, uint32_t d_value, uint64_t loop4_src_stride, bool smallc0_en) __aicore__ inline void asc_copy_gm2l1_dn2nz_sync(__cbuf__ int16_t* dst, __gm__ int16_t* src, uint64_t loop1_src_stride, uint8_t l2_cache_ctl, uint16_t n_value, uint32_t d_value, uint64_t loop4_src_stride, bool smallc0_en) __aicore__ inline void asc_copy_gm2l1_dn2nz_sync(__cbuf__ uint16_t* dst, __gm__ uint16_t* src, uint64_t loop1_src_stride, uint8_t l2_cache_ctl, uint16_t n_value, uint32_t d_value, uint64_t loop4_src_stride, bool smallc0_en) __aicore__ inline void asc_copy_gm2l1_dn2nz_sync(__cbuf__ half* dst, __gm__ half* src, uint64_t loop1_src_stride, uint8_t l2_cache_ctl, uint16_t n_value, uint32_t d_value, uint64_t loop4_src_stride, bool smallc0_en) __aicore__ inline void asc_copy_gm2l1_dn2nz_sync(__cbuf__ bfloat16_t* dst, __gm__ bfloat16_t* src, uint64_t loop1_src_stride, uint8_t l2_cache_ctl, uint16_t n_value, uint32_t d_value, uint64_t loop4_src_stride, bool smallc0_en) __aicore__ inline void asc_copy_gm2l1_dn2nz_sync(__cbuf__ int32_t* dst, __gm__ int32_t* src, uint64_t loop1_src_stride, uint8_t l2_cache_ctl, uint16_t n_value, uint32_t d_value, uint64_t loop4_src_stride, bool smallc0_en) __aicore__ inline void asc_copy_gm2l1_dn2nz_sync(__cbuf__ uint32_t* dst, __gm__ uint32_t* src, uint64_t loop1_src_stride, uint8_t l2_cache_ctl, uint16_t n_value, uint32_t d_value, uint64_t loop4_src_stride, bool smallc0_en) __aicore__ inline void asc_copy_gm2l1_dn2nz_sync(__cbuf__ float* dst, __gm__ float* src, uint64_t loop1_src_stride, uint8_t l2_cache_ctl, uint16_t n_value, uint32_t d_value, uint64_t loop4_src_stride, bool smallc0_en)

参数说明

表1参数说明

表2l2_cache_ctl取值说明

返回值说明

无

流水类型

PIPE_MTE2

约束说明

• 注意当且仅当D <= 4时，smallc0_en才可以被启用，否则它会失效。

调用示例

//搬运关于最内层循环的步长为4个字节 constexpr uint64_t loop1_src_stride = 4; //L2 Cache采用NORMAL模式 constexpr uint8_t l2_cache_ctl = 1; //N值为3 constexpr uint16_t n_value = 3; //D值为3 constexpr uint32_t d_value = 3; //搬运关于最外层循环的步长为4个字节 constexpr uint64_t loop4_src_stride = 2; //每一个C0数据块都会被pad至32个字节 constexpr bool smallc0_en = true; __gm__ half src[256]; __cbuf__ half dst[256]; asc_copy_gm2l1_dn2nz(dst, src, loop1_src_stride, l2_cache_ctl, n_value, d_value, loop4_src_stride, smallc0_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