InitBufPool

Applicability

Product	Supported
Atlas 350 Accelerator Card	√
Atlas A3 training product/Atlas A3 inference product	√
Atlas A2 training product/Atlas A2 inference product	√
Atlas 200I/500 A2 inference product	x
Atlas inference product AI Core	√
Atlas inference product Vector Core	x
Atlas training product	√

Function Usage

TPipe::InitBufPool obtains a TbufPool resource block, which can be further divided into small resource blocks by TBufPool::InitBufPool.

Prototype

Non-sharing mode

template <class T>
__aicore__ inline bool InitBufPool(T& bufPool, uint32_t len)

Sharing mode

template <class T, class U>
__aicore__ inline bool InitBufPool(T& bufPool, uint32_t len, U& shareBuf)

Parameters

**Table 1** Template parameters
Parameter	Description
T	Type of the bufPool parameter.
U	shareBuf type.

**Table 2** Parameters in InitBufPool(T& bufPool, uint32_t len) prototype definition
Parameter	Input/Output	Meaning
bufPool	Input	New resource pool. The type is TBufPool.
len	Input	Length of the new resource pool, in bytes. If the length is not 32-byte aligned, it is automatically padded up to 32-byte aligned.

**Table 3** Parameters in InitBufPool(T& bufPool, uint32_t len, U& shareBuf) prototype definition
Parameter	Input/Output	Meaning
bufPool	Input	New resource pool. The type is TBufPool.
len	Input	Length of the new resource pool, in bytes. If the length is not 32-byte aligned, it is automatically padded up to 32-byte aligned.
shareBuf	Input	Reused resource pool of the TBufPool type. The new resource pool shares the start address and length with the reused resource pool.

Constraints

The physical memory of the new resource pool must be the same as that of the reused resource pool. The two resource pools share the start address and length.
The entered length must be less than or equal to the length of the reused resource pool.
For details about other restrictions, see TBufPool.

Returns

None

Example

When the data volume is large and the memory is limited, data movement cannot be completed at a time. In this case, the data movement needs to be split into multiple phases. Each phase uses a part of the data. The TBufPool resource pool can be used for memory address reuse. In this example, the resource pool TBufPool0 is allocated from the TPipe. After TBufPool0 allocates space to src0Gm, TBufPool1 is allocated. TBufPool1 and TBufPool2 are reused and used for the first and second rounds of computation, respectively. In this case, TBufPool1 and TBufPool2 share the start address and length.

AscendC::TPipe pipe;
AscendC::TBufPool<AscendC::TPosition::VECCALC> tbufPool0, tbufPool1, tbufPool2;
// srcQue0, srcQue1, and srcQue2 are TQue instances on VECIN, and dstQue0 and dstQue1 are TQue instances on VECOUT.
// Divide tbufPool0 from Tpipe.
pipe.InitBufPool(tbufPool0, 131072);
// Allocate space to srcQue0.
tbufPool0.InitBuffer(srcQue0, 1, 65536); // Total src0
// Allocate space to tbufPool1 through tbufPool0, and specify that tbufPool1 and tbufPool2 share the same space.
tbufPool0.InitBufPool(tbufPool1, 65536);
tbufPool0.InitBufPool(tbufPool2, 65536, tbufPool1);
// Allocate space to srcQue1 and dstQue0 through tbufPool1.
tbufPool1.InitBuffer(srcQue1, 1, 32768);
tbufPool1.InitBuffer(dstQue0, 1, 32768);
// Switch the resource pool to tbufPool2 and allocate space to srcQue2 and dstQue1 through tbufPool2.
tbufPool1.Reset();
tbufPool2.InitBuffer(srcQue2, 1, 32768);
tbufPool2.InitBuffer(dstQue1, 1, 32768);
tbufPool2.Reset();
tbufPool0.Reset();
pipe.Reset();

Parent topic: TBufPool