Sort32

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	√
Atlas inference product AI Core	x
Atlas inference product Vector Core	x
Atlas training product	x

Function Usage

Sorts a maximum of 32 elements in each repeat. The data needs to be stored according to the following structure:

Scores and indexes are stored in src0 and src1 respectively. The sorting is based on scores in descending order. The sorted scores and their indexes are stored in dst in the (score, index) structure. No matter whether the scores are of the half or float type, the (score, index) structure in dst always occupies 8 bytes.

See the following examples:

When the score type is float and the index type is uint32_t, in the computation result, the indexes are stored in the upper 4 bytes and the scores are stored in the lower 4 bytes.
When the score type is half and the index type is uint32_t, in the computation result, the indexes are stored in the upper 4 bytes, the scores are stored in the lower 2 bytes, and the middle 2 bytes are reserved.

Prototype

      
           template <typename T>
__aicore__ inline void Sort32(const LocalTensor<T>& dst, const LocalTensor<T>& src0, const LocalTensor<uint32_t>& src1, const int32_t repeatTime)

Parameters

**Table 1** Parameters in the template
Parameter	Description
T	Operand data type. For the Atlas 350 Accelerator Card, the supported data types are half and float. For the Atlas A3 training product / Atlas A3 inference product , the supported data types are half and float. For the Atlas A2 training product / Atlas A2 inference product , the supported data types are half and float. For the Atlas 200I/500 A2 inference product , the supported data types are half and float.

**Table 2** Parameters
Parameter	Input/Output	Meaning
dst	Output	Destination operand. The type is LocalTensor, and TPosition can be VECIN, VECCALC, or VECOUT. The start address of LocalTensor must be 32-byte aligned.
src0	Input	Source operand. The type is LocalTensor, and TPosition can be VECIN, VECCALC, or VECOUT. The start address of LocalTensor must be 32-byte aligned. The source operand must have the same data type as the destination operand.
src1	Input	Source operand. The type is LocalTensor, and TPosition can be VECIN, VECCALC, or VECOUT. The start address of LocalTensor must be 32-byte aligned. This source operand is fixed at the uint32_t data type.
repeatTime	Input	Number of iteration repeats. The value is of the int32_t type. 32 elements are sorted in each iteration. In the next iteration, 32 elements are skipped in src0 and src1 respectively, and 32 x 8 bytes are skipped in dst. Value range: repeatTime ∈ [0,255]

Returns

None

Restrictions

If score [i] and score [j] are the same and i is greater than j, score [j] is selected first.
Data within each iteration is sorted, but data among different iterations is not sorted.
For details about the operand address alignment requirements, see General Address Alignment Restrictions.

Examples

       
            AscendC::LocalTensor<float> srcLocal0 = inQueueSrc0.DeQue<float>();
AscendC::LocalTensor<uint32_t> srcLocal1 = inQueueSrc1.DeQue<uint32_t>();
AscendC::LocalTensor<float> dstLocal = outQueueDst.AllocTensor<float>();
// repeatTime = 4. Divide 128 elements into four groups and sort the 32 elements in each group each time.
AscendC::Sort32<float>(dstLocal, srcLocal0, srcLocal1, 4);
outQueueDst.EnQue<float>(dstLocal);
inQueueSrc0.FreeTensor(srcLocal0);
inQueueSrc1.FreeTensor(srcLocal1);

Parent topic: Combination Sorting (ISASI)