ShiftRight

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

Performs a right shift on each element of the source operand. The number of bits to shift is determined by a scalar. A right shift operation is classified into the following types based on the data type of the source operand:

If the data type is unsigned, a logical right shift is performed. In this case, a right shift moves a binary number to the right by the specified number of bits, discarding the least significant bits and filling the most significant bits with zeros. For example, the binary number 1010101010101010 (uint16_t) after a logical right shift by 1 bit becomes 0101010101010101.
If the data type is signed, an arithmetic right shift is performed. In this case, a right shift moves the binary number to the right by a specified number of bits, discarding the least significant bits and replicating the sign bit into the most significant bits. For example, the binary number 1010101010101010 (int16_t) becomes 1101010101010101 after an arithmetic right shift by 1 bit, and becomes 1111010101010101 after an arithmetic right shift by 3 bits.

$\text{[math]}$

Prototype

Computation of the first n pieces of data of a tensor

        
             template <typename T, bool isSetMask = true>
__aicore__ inline void ShiftRight(const LocalTensor<T>& dst, const LocalTensor<T>& src, const T& scalarValue, const int32_t& count)

High-dimensional tensor sharding computation

Bitwise mask mode

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

Contiguous mask mode

          
               template <typename T, bool isSetMask = true>
__aicore__ inline void ShiftRight(const LocalTensor<T>& dst, const LocalTensor<T>& src, const T& scalarValue, uint64_t mask, const uint8_t repeatTime, const UnaryRepeatParams& repeatParams, bool roundEn = false)

If dst and src use the TensorTrait data structure, their data type (represented by LiteType in TensorTrait) may be different from the data type of scalarValue. So, a new template parameter U needs to be created to indicate the data type of scalarValue. std::enable_if is used to check whether LiteType extracted from T is the same as U. If they are the same, the API passes the compilation. Otherwise, the compilation fails. The API prototype is defined as follows:

Computation of the first n pieces of data of a tensor

        
             template <typename T, typename U, bool isSetMask = true, typename Std::enable_if<Std::is_same<PrimT<T>, U>::value, bool>::type = true>
__aicore__ inline void ShiftRight(const LocalTensor<T>& dst, const LocalTensor<T>& src, const U& scalarValue, const int32_t& count)

High-dimensional tensor sharding computation

Bitwise mask mode

          
               template <typename T, typename U, bool isSetMask = true, typename Std::enable_if<Std::is_same<PrimT<T>, U>::value, bool>::type = true>
__aicore__ inline void ShiftRight(const LocalTensor<T>& dst, const LocalTensor<T>& src, const U& scalarValue, uint64_t mask[], const uint8_t repeatTime, const UnaryRepeatParams& repeatParams, bool roundEn)

Contiguous mask mode

          
               template <typename T, typename U, bool isSetMask = true, typename Std::enable_if<Std::is_same<PrimT<T>, U>::value, bool>::type = true>
__aicore__ inline void ShiftRight(const LocalTensor<T>& dst, const LocalTensor<T>& src, const U& scalarValue, uint64_t mask, const uint8_t repeatTime, const UnaryRepeatParams& repeatParams, bool roundEn)

Parameters

**Table 1** Template parameters
Parameter	Description
T	Operand data type. For the Atlas A2 training product / Atlas A2 inference product , the supported data types are uint16_t, int16_t, uint32_t, and int32_t. For the Atlas A3 training product / Atlas A3 inference product , the supported data types are uint16_t, int16_t, uint32_t, and int32_t. For the Atlas 200I/500 A2 inference product , the supported data types are uint16_t, int16_t, uint32_t, and int32_t. For the Atlas 350 Accelerator Card, the supported data types are uint8_t, int8_t, uint16_t, int16_t, uint32_t, int32_t, uint64_t, and int64_t.
U	Data type of scalarValue. For the Atlas A2 training product / Atlas A2 inference product , the supported data types are uint16_t, int16_t, uint32_t, and int32_t. For the Atlas A3 training product / Atlas A3 inference product , the supported data types are uint16_t, int16_t, uint32_t, and int32_t. For the Atlas 200I/500 A2 inference product , the supported data types are uint16_t, int16_t, uint32_t, and int32_t. For the Atlas 350 Accelerator Card, the supported data types are uint8_t, int8_t, uint16_t, int16_t, uint32_t, int32_t, uint64_t, and int64_t.
isSetMask	Whether to set the mask mode and mask value inside the API. true: indicates that the settings are performed inside the API. The APIs for high-dimensional tensor sharding computation and for computing the first n pieces of data in a tensor use the Normal or Counter mode of the mask. Generally, retain the default value of isSetMask, indicating that the mask mode and mask value are set in the API based on the mask and count parameters passed by developers. false: indicates that the settings are performed outside the API. For high-dimensional tensor sharding computation APIs, in some scenarios that require high performance, developers need to use SetMaskNorm or SetMaskCount to set the mask mode and use the SetVectorMask API to set the mask value. The mask value in the input parameter of this API must be set to MASK_PLACEHOLDER. For the APIs that compute the first n pieces of data in a tensor, in some scenarios that require high performance, developers need to use SetMaskCount to set the mask mode to Counter and use the SetVectorMask API to set the mask value. The count value in the input parameter of this API does not take effect. You are advised to set it to 1. For the models below, isSetMask is invalid for the APIs that compute the first n pieces of data in a tensor. Retain the default value. Atlas 350 Accelerator Card Atlas 200I/500 A2 inference product

**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.
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 data type must be the same as that of the destination operand.
scalarValue	Input	Number of bits to shift (shift amount). Its data type must be the same as the tensor elements in the destination operand. For the Atlas A2 training product / Atlas A2 inference product , if src is of uint16_t or int16_t type, the value range of scalarValue is [0, 16]; if src is of uint32_t or int32_t type, the value range of scalarValue is [0, 32]. For the Atlas A3 training product / Atlas A3 inference product , if src is of uint16_t or int16_t type, the value range of scalarValue is [0, 16]; if src is of uint32_t or int32_t type, the value range of scalarValue is [0, 32]. For the Atlas 200I/500 A2 inference product , if src is of uint16_t or int16_t type, the value range of scalarValue is [0, 16]; if src is of uint32_t or int32_t type, the value range of scalarValue is [0, 32]. For the Atlas 350 Accelerator Card, scalarValue must be greater than or equal to 0. For a logical right shift, if the number of bits to shift exceeds the bit width of the src data type, all elements in dst are assigned the value 0. For an arithmetic right shift, if the number of bits to shift exceeds the bit width of the src data type, a positive element in the source operand results in element 0 in the destination operand and a negative element in the source operand results in element –1 in the destination operand.
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 Unit reads 256 bytes of contiguous data for computation each time. To read the complete data for processing, the unit needs to read the input data in multiple repeats. repeatTime indicates the number of iterations. For details about this parameter, see High-dimensional Sharding APIs.
repeatParams	Input	Structure for controlling element-wise operations. For details, see UnaryRepeatParams.
roundEn	Input	Whether to enable rounding. The value is of the bool type. true indicates rounding is enabled, and false indicates rounding is disabled. This parameter is valid only when src is of type int16_t or int32_t. For example, with the rounding function enabled and src of type int16_t, the 5-bit arithmetic right shift of src is computed as follows: src_ele = 17 = 0b0000000000010001 (the fifth bit is 1) dst_ele = arithmetic_right_shift(src_ele, 5) + 1 = 0b0000000000000000 + 1 = 0b0000000000000001 For the Atlas 200I/500 A2 inference product , the rounding function cannot be enabled and the value can only be false. For the Atlas 350 Accelerator Card, the rounding function cannot be enabled and the value can only be false.

Returns

None

Restrictions

For details about the operand address alignment requirements, see General Address Alignment Restrictions.
For details about the constraints on operand address overlapping, see General Address Overlapping Restrictions.

For the Atlas 350 Accelerator Card, int8_t, uint8_t, uint64_t, and int64_t support only the APIs that compute the first n pieces of data in a tensor.

Examples

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

        
             // dstLocal: tensor for storing the ShiftLeft computation result
// srcLocal: tensor for storing the ShiftLeft computation input

uint64_t mask = 128;
int16_t scalar = 2; // a right shift by 2 bits
// repeatTime = 4. 128 elements are processed in a single iteration. To compute 512 elements, four iterations are required.
// dstBlkStride, srcBlkStride = 1. The interval between src0 data addresses involved in computation in each iteration is one data block, indicating that data is continuously read and written in a single iteration.
// dstRepStride, srcRepStride = 8. The interval between addresses of adjacent iterations is eight data blocks, indicating that data is continuously read and written between adjacent iterations.
AscendC::ShiftRight(dstLocal, srcLocal, scalar, mask, 4, { 1, 1, 8, 8 }, false);

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

        
             // dstLocal: tensor for storing the ShiftLeft computation result
// srcLocal: tensor for storing the ShiftLeft computation input

uint64_t mask[2] = { UINT64_MAX, UINT64_MAX };
int16_t scalar = 2; // a right shift by 2 bits
// repeatTime = 4. 128 elements are processed in a single iteration. To compute 512 elements, four iterations are required.
// dstBlkStride, srcBlkStride = 1. The interval between src0 data addresses involved in computation in each iteration is one data block, indicating that data is continuously read and written in a single iteration.
// dstRepStride, srcRepStride = 8. The interval between addresses of adjacent iterations is eight data blocks, indicating that data is continuously read and written between adjacent iterations.
AscendC::ShiftRight(dstLocal, srcLocal, scalar, mask, 4, {1, 1, 8, 8}, false);

Example of computing the first n pieces of data of a tensor

        
             int16_t scalar = 2; // a right shift by 2 bits
// The input data type of the operator is int16_t, and the number of elements involved in the computation is 512.
AscendC::ShiftRight(dstLocal, srcLocal, scalar, 512);

Result example:

Input (srcLocal): [1 2 3 ... 512]
Input (scalar) = 2
Output (dstLocal): [0 0 0 1 1 1 1 ... 128]

Parent topic: Logical Computation