DataCopyPad(ISASI)

Applicability

Function Usage

Performs unaligned data transfer from the global memory to the local memory. During the transfer, developers can customize how padding data is filled as required.

Prototype

• dataCopyParams is of the DataCopyExtParams type. It supports a larger value range for parameters such as the operand stride compared to the DataCopyParams type.
  • Transfer path: global memory -> local memory

    1 2	template <typename T, PaddingMode mode = PaddingMode::Normal> __aicore__ inline void DataCopyPad(const LocalTensor<T>& dst, const GlobalTensor<T>& src, const DataCopyExtParams& dataCopyParams, const DataCopyPadExtParams<T>& padParams)

    1 2	template <typename T> __aicore__ inline void DataCopyPad(const LocalTensor<T>& dst, const GlobalTensor<T>& src, const DataCopyExtParams& dataCopyParams, const DataCopyPadExtParams<T>& padParams)

  • Transfer path: local memory -> global memory

    1 2	template <typename T, PaddingMode mode = PaddingMode::Normal> __aicore__ inline void DataCopyPad(const GlobalTensor<T>& dst, const LocalTensor<T>& src, const DataCopyExtParams& dataCopyParams)

    1 2	template <typename T> __aicore__ inline void DataCopyPad(const GlobalTensor<T>& dst, const LocalTensor<T>& src, const DataCopyExtParams& dataCopyParams)

  • Transfer path: local memory -> local memory. The actual transfer process goes through: VECIN or VECOUT -> global memory -> TSCM.

    1 2	template <typename T> __aicore__ inline void DataCopyPad(const LocalTensor<T>& dst, const LocalTensor<T>& src, const DataCopyExtParams& dataCopyParams, const Nd2NzParams& nd2nzParams)

• dataCopyParams is of the DataCopyParams type.
  • Transfer path: global memory -> local memory

    1 2	template<typename T, PaddingMode mode = PaddingMode::Normal> __aicore__ inline void DataCopyPad(const LocalTensor<T>& dst, const GlobalTensor<T>& src, const DataCopyParams& dataCopyParams, const DataCopyPadParams& padParams)

    1 2	template<typename T> __aicore__ inline void DataCopyPad(const LocalTensor<T>& dst, const GlobalTensor<T>& src, const DataCopyParams& dataCopyParams, const DataCopyPadParams& padParams)

  • Transfer path: local memory -> global memory

    1 2	template<typename T, PaddingMode mode = PaddingMode::Normal> __aicore__ inline void DataCopyPad(const GlobalTensor<T>& dst, const LocalTensor<T>& src,const DataCopyParams& dataCopyParams)

    1 2	template<typename T> __aicore__ inline void DataCopyPad(const GlobalTensor<T>& dst, const LocalTensor<T>& src,const DataCopyParams& dataCopyParams)

  • Transfer path: local memory -> local memory. The actual transfer process goes through: VECIN or VECOUT -> global memory -> TSCM.

    1 2	template<typename T> __aicore__ inline void DataCopyPad(const LocalTensor<T>& dst, const LocalTensor<T>& src, const DataCopyParams& dataCopyParams, const Nd2NzParams& nd2nzParams)

Different product models support different function prototypes. Refer to the following table to select a function prototype supported by your product model.

Parameters

1
2
3
4

enum class PaddingMode : uint8_t {
    Normal = 0,  // Default mode, which keeps the same data transfer format as the original. Data is padded to 32 bytes during each data transfer.
    Compact,     // Compact mode, which allows unaligned data during a single transfer. The entire data block is padded at the end to achieve 32-byte alignment.
};

For details about the structure parameters listed in the following table, see ${INSTALL_DIR}/include/ascendc/basic_api/interface/kernel_struct_data_copy.h. Replace ${INSTALL_DIR} with the actual CANN component directory.

The following provides configuration examples for different scenarios:

• GM -> VECIN or VECOUT
• VECIN or VECOUT -> GM
• VECIN or VECOUT -> TSCM
• Configuration example of the transfer mode
• GM -> VECIN or VECOUT
  • Description
    • When blockLen + leftPadding + rightPadding is 32-byte aligned and isPad is false, the values used for left and right padding will default to random values; otherwise, they will be set to paddingValue.
    • When blockLen + leftPadding + rightPadding is not 32-byte aligned, the framework pads some dummy data to ensure the 32-byte alignment. When both leftPadding and rightPadding are 0, the dummy value fills with the first element of the data chunk being transferred by default. When leftPadding or rightPadding is not 0, if isPad is set to false, the data values padded on the left and right sides and dummy values are random numbers; if not, the data values padded on the left and right sides and dummy values are configured padding values.
  • Configuration example 1

    • blockLen = 64 indicates that each data chunk contains 64 bytes.

      srcStride = 1 indicates that the stride between adjacent data chunks of the source operand is 1 byte. The logical position of the source operand is GM, and the unit of srcStride is byte.

      dstStride = 1 indicates that the stride between adjacent data chunks of the destination operand is one data block. The logical position of the destination operand is VECIN or VECOUT, and the unit of dstStride is the number of data blocks (32 bytes per data block).

    • When blockLen + leftPadding + rightPadding is 32-byte aligned, if isPad is set to false, the data values padded on the left and right sides are random numbers by default; if not, the data values padded on the left and right sides are configured padding values. In this example, both leftPadding and rightPadding are 0, indicating that padding is not performed.
    • When blockLen + leftPadding + rightPadding is not 32-byte aligned, the framework pads some dummy data to ensure the 32-byte alignment. When leftPadding or rightPadding is not 0, if isPad is set to false, the data values padded on the left and right sides and dummy values are random numbers; if not, the data values padded on the left and right sides are configured padding values.

    

  • Configuration example 2

    • blockLen = 47 indicates that each data chunk contains 47 bytes.

      srcStride = 1 indicates that the stride between adjacent data chunks of the source operand is 1 byte.

      dstStride = 1 indicates that the stride between adjacent data chunks of the destination operand is one data block.

    • When blockLen + leftPadding + rightPadding is not 32-byte aligned, and both leftPadding and rightPadding are 0, the dummy value fills with the first element of the data chunk being transferred by default.
    • When blockLen + leftPadding + rightPadding is not 32-byte aligned, and leftPadding or rightPadding is not 0, if isPad is set to false, the data values padded on the left and right sides and dummy values are random numbers; if not, the data values padded on the left and right sides and dummy values are configured padding values.

    

  • Configuration example 3

    • blockLen = 48 indicates that each data chunk contains 48 bytes.

      srcStride = –48 indicates that the stride between adjacent data chunks of the source operand is –48 bytes. This means that the data chunks transferred each time are the same.

      dstStride = 1 indicates that the stride between adjacent data chunks of the destination operand is one data block.

    • When blockLen + leftPadding + rightPadding is not 32-byte aligned, and both leftPadding and rightPadding are 0, the dummy value fills with the first element of the data chunk being transferred by default.
    • When blockLen + leftPadding + rightPadding is not 32-byte aligned, and leftPadding or rightPadding is not 0, if isPad is set to false, the data values padded on the left and right sides and dummy values are random numbers; if not, the data values padded on the left and right sides and dummy values are configured padding values.

    

Returns

None

Restrictions

• For both leftPadding and rightPadding, the total number of bytes occupied must not exceed 32 bytes.
• For the Atlas 350 Accelerator Card, the following data types are supported for data transfer from VECIN or VECOUT to TSCM: half, bfloat16_t, int16_t, uint16_t, float, int32_t, uint32_t, int8_t, uint8_t, int64_t, uint64_t, and double.

Examples

1
2
3
4

AscendC::DataCopyExtParams copyParams{1, 20 * sizeof(half), 0, 0, 0}; 
AscendC::DataCopyPadExtParams<half> padParams{true, 0, 2, 0};
// Transfer 40 bytes from GM to VECIN.
AscendC::DataCopyPad(dstGlobal, src0Global, copyParams, padParams); 

Input (src0Global): [1 2 3 ... 32]
Output (dstGlobal): [1 2 3 ... 20]