VectorPadding (ISASI)

Applicability

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

Function Usage

Performs the padding operation on the source operand by the data block based on padMode and padSide.

Suppose that a data block of the source operand has 16 numbers: data block[0:15] = a to p.

padSide==false: Pads from the left of the data block along its starting direction (a->p).

padSide==true: Pads from the right of the data block along its ending direction (p->a).
padMode==0: uses the adjacent number as the padding value, for example, aaa|abc (padSide=false) and nop|ppp (padSide=true).
padMode==1: uses the adjacent data block value for symmetric padding, for example, cba|abc (padSide=false) and nop|pon (padSide=true).
padMode==2: uses the adjacent data block value that is offset by a number for symmetric padding. For example:
- In padSide=false: xcb|abc, where xcb is padded as follows: If a is discarded, x is padded with 0 symmetrically.
- In padSide=true: nop|onx, where onx is padded as follows: If p is discarded, x is padded with 0 symmetrically.

Prototype

Computation of the first n data elements of a tensor

        
             template <typename T>
__aicore__ inline void VectorPadding(const LocalTensor<T>& dst, const LocalTensor<T>& src, const uint8_t padMode, const bool padSide, const uint32_t count)

High-dimensional tensor sharding computation

Bitwise mask mode

          
               template <typename T, bool isSetMask = true>
__aicore__ inline void VectorPadding(const LocalTensor<T>& dst, const LocalTensor<T>& src, const uint8_t padMode, const bool padSide, const uint64_t mask[], const uint8_t repeatTime, const UnaryRepeatParams& repeatParams)

Contiguous mask mode

          
               template <typename T, bool isSetMask = true>
__aicore__ inline void VectorPadding(const LocalTensor<T>& dst, const LocalTensor<T>& src, const uint8_t padMode, const bool padSide, const uint64_t mask, const uint8_t repeatTime, const UnaryRepeatParams& repeatParams)

Parameters

**Table 1** Parameters in the template
Parameter	Description
T	Operand data type. For the Atlas inference product AI Core, the supported data types are int16_t, uint16_t, half, int32_t, uint32_t, and float.
isSetMask	Indicates whether to set mask inside the API. true: sets mask inside the API. false: sets mask outside the API. Developers need to use the SetVectorMask API to set the mask value. In this mode, the mask value in the input parameter of this API must be set to the placeholder MASK_PLACEHOLDER.

**Table 2** Parameters
Parameter	Input/Output	Description
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.
src	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.
padMode	Input	Padding mode. The type is uint8_t. The value range is [0, 2]. 0: The adjacent number is used as the padding value. 1: The adjacent data block value is used for symmetric padding. 2: The adjacent data block value is offset by a number for symmetric padding.
padSide	Input	Padding direction. The value is of the bool type. false: left true: right
count	Input	Number of elements involved in the computation.
mask[]/mask	Input	mask controls the elements that participate in computation in each iteration. Bitwise mode: controls the elements that participate in computation by bit. If a bit is set to 1, the corresponding element participates in the computation. If a bit is set to 0, the corresponding element is masked from the computation. The mask value is an array. The array length and the value range of the array elements are related to the operand data type. When the operand is 16-bit, the array length is 2, with mask[0] and mask[1] each in the range [0, 2⁶⁴ – 1], and they cannot both be 0 at the same time. When the operand is 32-bit, the array length is 1, with mask[0] in the range (0, 2⁶⁴ – 1]. When the operand is 64-bit, the array length is 1, with mask[0] in the range (0, 2³² – 1]. For example, if mask = [0, 8] and 8 = 0b1000, only the fourth element participates in computation. Contiguous mode: indicates the number of contiguous elements that participate in computation. The value range is related to the operand data type. The maximum number of elements that can be processed in each iteration varies according to the data type. When the operand is 16-bit, mask ∈ [1, 128]. When the operand is 32-bit, mask ∈ [1, 64]. When the operand is 64-bit, mask ∈ [1, 32].
repeatTime	Input	Number of iteration repeats. The vector compute unit reads 256 bytes of contiguous data for computation each time. To process the input data, the data needs to be read and computed over multiple repeats. repeatTime indicates the number of repeats. For details about this parameter, see High-dimensional Sharding APIs.
repeatParams	Input	Parameters that control the operand address strides. They are of the UnaryRepeatParams type (see UnaryRepeatParams), and contain parameters such as the address stride of the operand for the same Data Block between adjacent iterations and the address stride of the operand between different Data Blocks in a single iteration. For details about the address stride of the operand between adjacent iterations, see repeatStride. For details about the address stride of the operand between different data blocks in a single iteration, see dataBlockStride.

Returns

None

Restrictions

For details about the operand address alignment requirements, see General Address Alignment Restrictions.

mask controls only the write operation on the destination operand. It is irrelevant to the read operation on the source operand.
count indicates the total number of elements written to the destination operand. It is irrelevant to the read operation on the source operand.

Examples

In this example, srcLocal and dstLocal are of the half type.

For more examples, see LINK.

Example of high-dimensional tensor sharding computation (contiguous mask mode)

        
             uint64_t mask = 256 / sizeof(half);
uint8_t padMode = 0;
bool padSide = false;
// repeatTime = 4, 128 elements one repeat, 512 elements total
// dstBlkStride, srcBlkStride = 1, no gap between blocks in one repeat
// dstRepStride, srcRepStride = 8, no gap between repeats
AscendC::VectorPadding(dstLocal, srcLocal, padMode, padSide, mask, 4, { 1, 1, 8, 8 });

Example of high-dimensional tensor sharding computation (bitwise mask mode)

        
             uint64_t mask[2] = { UINT64_MAX, UINT64_MAX };
uint8_t padMode = 0;
bool padSide = false;
// repeatTime = 4, 128 elements one repeat, 512 elements total
// dstBlkStride, srcBlkStride = 1, no gap between blocks in one repeat
// dstRepStride, srcRepStride = 8, no gap between repeats
AscendC::VectorPadding(dstLocal, srcLocal, padMode, padSide, mask, 4, { 1, 1, 8, 8 });

Example of computing the first n data elements of a tensor

        
             uint8_t padMode = 0;
bool padSide = false;
AscendC::VectorPadding(dstLocal, srcLocal, padMode, padSide, 512);

Result example:

// In srcLocal, there are 16 numbers in a data block.
Input (srcLocal): [6.938 -8.86 -0.2263 ... 1.971 1.778]
Output (dstLocal):
[6.938 6.938 6.938 ... 6.938 6.938]

Parent topic: Data Padding