目录

GlobalTensor

Function Description

Stores the global data of the global memory (external storage).

Prototype

template <typename T> class GlobalTensor : public BaseGlobalTensor<T> {
public:
    using PrimType = PrimT<T>;
    __aicore__ inline GlobalTensor<T>() {}
    __aicore__ inline void SetGlobalBuffer(__gm__ PrimType* buffer, uint64_t bufferSize); // Pass in the pointer to the global data, set a buffer size, and initialize the GlobalTensor.
    __aicore__ inline void SetGlobalBuffer(__gm__ PrimType* buffer); // Pass in the pointer to the global data, and initialize the GlobalTensor. The buffer size does not necessarily need to be passed in, but the length obtained by using GetSize is a random value.
    __aicore__ inline const __gm__ PrimType* GetPhyAddr() const; // Return the address of the global data.
    __aicore__ inline __gm__ PrimType* GetPhyAddr(const uint64_t offset) const; // Return the address (with *offset* elements) of the global data.
    __aicore__ inline __inout_pipe__(S) PrimType GetValue(const uint64_t offset) const; // Obtain the value of the corresponding offset position of a tensor.
    __aicore__ inline __inout_pipe__(S) __gm__ PrimType& operator()(const uint64_t offset) const; // Return the reference of an element numbered as *index*.
    __aicore__ inline void SetValue(const uint64_t offset, PrimType value); // Set the value of the corresponding offset position of a tensor.
    __aicore__ inline uint64_t GetSize() const; // Return the number of elements in a tensor.
    __aicore__ inline GlobalTensor operator[](const uint64_t offset) const; // Specify the offset and returns a global tensor. The unit of offset is element.
    __aicore__ inline void SetShapeInfo(const ShapeInfo& shapeInfo);
    __aicore__ inline ShapeInfo GetShapeInfo() const;
    template<CacheRwMode rwMode = CacheRwMode::RW> 
    __aicore__ inline void SetL2CacheHint(CacheMode mode); // Set the mode (allowed/forbidden) for writing tensors to the L2 cache.
    // ...
};

Function Usage

Type T supports basic data types, and supports the data types of the instructions that use the GlobalTensor.

Table 1 Function description

Function Name Input Parameter Description
GetValue offset: offset value, in elements Obtains a value in GlobalTensor. The immediate of the T type is returned.
SetGlobalBuffer

buffer: global data pointer passed from the host

bufferSize: number of T data elements. Ensure that the value does not exceed the actual data length.

 Sets the storage position of the GlobalTensor: buffer points to the start address of the external storage, and bufferSize indicates the size of the external storage occupied by the tensor. For example, if the pointed external storage has 256 consecutive int32_t data elements, bufferSize is 256.
GetPhyAddr - Returns the address of the GlobalTensor. If offset is passed in, offset elements are offset.
GetSize - Returns the number of elements in the GlobalTensor.
operator[] offset: offset position specified by the developer Returns a new tensor based on the input offset. The unit of offset is the number of elements.
operator() offset: subscript index Obtains the reference of the indexth variable of GlobalTensor. It is similar to operator() of LocalTensor.
SetShapeInfo shapeInfo: ShapeInfo structure Sets shapeInfo of the GlobalTensor.
GetShapeInfo -

Obtains shapeInfo of the GlobalTensor.

Note: The shape information does not have a default value. This API can be called to obtain the correct shape information only after the shape information is set by using SetShapeInfo.

Calling Example

void Init(__gm__ uint8_t *src_gm, __gm__ uint8_t *dst_gm)
{
    uint64_t dataSize = 256; // Set the size of input_global to 256.
 
    AscendC::GlobalTensor<int32_t> inputGlobal; // The type is int32_t.
    inputGlobal.SetGlobalBuffer(reinterpret_cast<__gm__ int32_t *>(src_gm), dataSize); // Set the start address of the source operand in the global memory to src_gm and the size of the external storage occupied by the source operand to 256 int32_t data elements.
 
    AscendC::LocalTensor<int32_t> inputLocal = inQueueX.AllocTensor<int32_t>();
    AscendC::DataCopy(inputLocal, inputGlobal, dataSize); // Copy inputGlobal from the global memory to inputLocal of the local memory.
    // ...
}