BilinearInterpolation (ISASI)

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	x

Function Usage

Functions are classified into horizontal iteration and vertical iteration. In each horizontal iteration, eight offset values are read from src0Offset in sequence, indicating the offset of src0. Each offset value points to the start address of a data block in src0. If repeatMode is set to false, a value is obtained from src1 and multiplied by each value in eight data blocks in src0. If repeatMode is set to true, eight values are obtained from src1 and multiplied by the values in the eight data blocks in src0 in sequence. The dst result of the current iteration and the previous dst result are accumulated by data block and stored in the destination address, the dst address remains unchanged in the same horizontal iteration. Then, vertical iteration is performed. The dst start address of vertical iteration is the dst start address of the previous vertical iteration plus vROffset. The dst space occupied by this round of vertical iteration is the eight blocks after the dst start address. In each round of vertical iteration, hRepeat horizontal iterations are performed.

Prototype

Bitwise mask mode:

        
             template <typename T>
__aicore__ inline void BilinearInterpolation(const LocalTensor<T>& dst, const LocalTensor<T>& src0, const LocalTensor<uint32_t>& src0Offset, const LocalTensor<T>& src1, uint64_t mask[], uint8_t hRepeat, bool repeatMode, uint16_t dstBlkStride, uint16_t vROffset, uint8_t vRepeat, const LocalTensor<uint8_t> &sharedTmpBuffer)

Contiguous mask mode:

        
             template <typename T>
__aicore__ inline void BilinearInterpolation(const LocalTensor<T>& dst, const LocalTensor<T>& src0, const LocalTensor<uint32_t>& src0Offset, const LocalTensor<T>& src1, uint64_t mask, uint8_t hRepeat, bool repeatMode, uint16_t dstBlkStride, uint16_t vROffset, uint8_t vRepeat, const LocalTensor<uint8_t> &sharedTmpBuffer)

Parameters

**Table 1** Template parameters
Parameter	Description
T	Operand data type. For the Atlas 350 Accelerator Card, the supported data type is half. For the Atlas A2 training product / Atlas A2 inference product , the supported data type is half. For the Atlas A3 training product / Atlas A3 inference product , the supported data type is half. For the Atlas inference product AI Core, the supported data type is half.

**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.
src0 and 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. Both source operands must have the same data type as the destination operand.
src0Offset	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.
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].
hRepeat	Input	Number of horizontal iterations. The value range is [1, 255].
repeatMode	Input	Repeat mode. false: Each value in the eight data blocks read by src0 in each iteration is multiplied by a single value in src1. true: Each data block of src0 is multiplied by a single element of src1 every iteration. A total of eight blocks and eight elements are consumed.
dstBlkStride	Input	Address stride of the destination operand between different data blocks in a single repeat, in the unit of 32 bytes.
vROffset	Input	Address offset of the destination operand between vertical iterations, in the unit of elements. The value range is [128, 65535). vROffset × sizeof(T) must be 32-byte aligned.
vRepeat	Input	Number of vertical iterations. The value range is [1, 255].
sharedTmpBuffer	Input	Temporary space. For the Atlas 350 Accelerator Card, no temporary space needs to be allocated. For the Atlas A2 training product / Atlas A2 inference product , ensure that at least a space of src0.GetSize() × 32 bytes + src1.GetSize() × 32 bytes is allocated. For the Atlas A3 training product / Atlas A3 inference product , ensure that at least a space of src0.GetSize() × 32 bytes + src1.GetSize() × 32 bytes is allocated. For the Atlas inference product AI Core, ensure that at least a space of src0OffsetLocal.GetSize() × sizeof(uint32_t) bytes is allocated.

Returns

None

Restrictions

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

The addresses of src0, src1, and src0Offset cannot overlap. In addition, the destination addresses of two vertical iterations cannot overlap.

Examples

API example - contiguous mask mode

        
             AscendC::LocalTensor<half> dstLocal, src0Local, src1Local;
AscendC::LocalTensor<uint32_t> src0OffsetLocal;
AscendC::LocalTensor<uint8_t> tmpLocal;
uint64_t mask = 128;        // Continuous mask mode
uint8_t hRepeat = 2;        // Two horizontal iterations
bool repeatMode = false;    // Iteration mode
uint16_t dstBlkStride = 1;  // Data is continuously written in a single iteration.
uint16_t vROffset = 128;    // Data is continuously written between adjacent iterations.
uint8_t vRepeat = 2;        // Two vertical iterations

AscendC::BilinearInterpolation(dstLocal, src0Local, src0OffsetLocal, src1Local, mask, hRepeat, repeatMode,
            dstBlkStride, vROffset, vRepeat, tmpLocal);

API example - bitwise mask mode

        
             AscendC::LocalTensor<half> dstLocal, src0Local, src1Local;
AscendC::LocalTensor<uint32_t> src0OffsetLocal;
AscendC::LocalTensor<uint8_t> tmpLocal;
uint64_t mask[2] = { UINT64_MAX, UINT64_MAX};  // Bitwise mask mode
uint8_t hRepeat = 2;        // Two horizontal iterations
bool repeatMode = false;    // Iteration mode
uint16_t dstBlkStride = 1;  // Data is continuously written in a single iteration.
uint16_t vROffset = 128;    // Data is continuously written between adjacent iterations.
uint8_t vRepeat = 2;        // Two vertical iterations

AscendC::BilinearInterpolation(dstLocal, src0Local, src0OffsetLocal, src1Local, mask, hRepeat, repeatMode,
            dstBlkStride, vROffset, vRepeat, tmpLocal);

Result example:

Input (src0Local,half): [1, 2, 3, ..., 512]
Input (src1Local,half): [2, 3, 4, ..., 17]
Input (src0OffsetLocal,uint32_t): [0, 32, 64, ..., 992]
Output (dstLocal,half): [389, 394, 399, 404, ..., 4096]

Parent topic: Basic Arithmetic