昇腾社区首页
中文
注册

预留接口

本章节列出的接口均为预留接口,后续有可能变更或废弃,不建议开发者使用,开发者无需关注。

Matmul

  • __aicore__ inline void SetSubBlockIdx(uint8_t subBlockIdx);
  • __aicore__ inline uint8_t GetSubBlockIdx();
  • __aicore__ inline void SetTensorAWithCopy(const GlobalTensor<SrcAT>& gm, const LocalTensor<SrcAT> &leftMatrix, bool isTransposeA = false);
  • __aicore__ inline void SetTensorBWithCopy(const GlobalTensor<SrcBT>& gm, const LocalTensor<SrcBT> &rightMatrix, bool isTransposeB = false);
  • __aicore__ constexpr MatmulConfig GetSpecialBasicConfig(const uint32_t basicM, const uint32_t basicN, const uint32_t basicK, const uint32_t singleCoreM, const uint32_t singleCoreN, const uint32_t singleCoreK, const uint32_t stepM, const uint32_t stepN, const bool intrinsicsLimit = false, const bool batchLoop = false, const BatchMode bmmMode = BatchMode::BATCH_LESS_THAN_L1);

ConfusionTranspose

void GetConfusionTransposeOnlyTilingInfo(const ge::Shape &srcShape, const uint32_t stackBufferSize, const uint32_t typeSize, optiling::ConfusionTransposeTiling &tiling);

内存管理和同步控制

  • TPipe
    • GetAbsAddr
  • __aicore__ constexpr Hardware GetPhyType(TPosition pos);
  • template <typename T, TPosition pos> __aicore__ inline bool PopStackBuffer(LocalTensor<T>& popLocal);

矩阵计算

  • template <typename DstT, typename SrcT, const FixpipeConfig& config = CFG_ROW_MAJOR> void Fixpipe(const LocalTensor<DstT>& dstLocal, const LocalTensor<SrcT>& srcLocal, const FixpipeParamsV220& intriParams);
  • template <typename DstT, typename SrcT, const FixpipeConfig& config = CFG_ROW_MAJOR> void Fixpipe(const LocalTensor<DstT>& dstLocal, const LocalTensor<SrcT>& srcLocal, const LocalTensor<uint64_t>& cbufWorkspace, const FixpipeParamsV220& intriParams);
  • template <typename T, typename U> __aicore__ inline __inout_pipe__(V) void BroadCastVecToMM(const LocalTensor<T> &dstLocal, const LocalTensor<U> &srcLocal, const int32_t blockCount, const uint8_t blockLen, const uint8_t srcGap, const uint8_t dstGap);

其他

  • template <typename T> __aicore__ inline __in_pipe__(V) __out_pipe__(MTE3) void InitOutput(GlobalTensor<T> gmWorkspaceAddr, uint32_t size, T value = 0);