AddReluCast

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	√
Atlas inference product Vector Core	x
Atlas training product	x

Function Usage

Performs element-wise addition, followed by a ReLU computation (comparing the result with 0 and taking the larger value), and converts the precision based on the data types of the source and destination operand tensors. The formula is as follows, where dstType indicates the data type of the destination operand:

$\text{[math]}$

Prototype

Computation of the first n pieces of data of a tensor

        
             template <typename T, typename U>
__aicore__ inline void AddReluCast(const LocalTensor<T>& dst, const LocalTensor<U>& src0, const LocalTensor<U>& src1, const uint32_t count)

High-dimensional tensor sharding computation

Bitwise mask mode

          
               template <typename T, typename U, bool isSetMask = true>
__aicore__ inline void AddReluCast(const LocalTensor<T>& dst, const LocalTensor<U>& src0, const LocalTensor<U>& src1, uint64_t mask[], const uint8_t repeatTime, const BinaryRepeatParams& repeatParams)

Contiguous mask mode

          
               template <typename T, typename U, bool isSetMask = true>
__aicore__ inline void AddReluCast(const LocalTensor<T>& dst, const LocalTensor<U>& src0, const LocalTensor<U>& src1, uint64_t mask, const uint8_t repeatTime, const BinaryRepeatParams& repeatParams)

Parameters

**Table 1** Template parameters
Parameter	Description
T	Data type of the destination operand. For details about precision conversion rules for different data types, see Table 3. For the Atlas 350 Accelerator Card, the supported data types are int8_t, half, int32_t, and float. For the Atlas A3 training product / Atlas A3 inference product , the supported data types are int8_t and half. For the Atlas A2 training product / Atlas A2 inference product , the supported data types are int8_t and half. For the Atlas 200I/500 A2 inference product , the supported data types are int8_t and half. For the Atlas inference product AI Core, the supported data types are int8_t and half.
U	Data type of the source operand. For the Atlas 350 Accelerator Card, the supported data types are int16_t, half, float, and int64_t. For the Atlas A3 training product / Atlas A3 inference product , the supported data types are int16_t, half, and float. For the Atlas A2 training product / Atlas A2 inference product , the supported data types are int16_t, half, and float. For the Atlas 200I/500 A2 inference product , the supported data types are half and float. For the Atlas inference product AI Core, the supported data types are int16_t, half, 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.
src0/src1	Input	Source operands. The type is LocalTensor, and TPosition can be VECIN, VECCALC, or VECOUT. The start address of LocalTensor must be 32-byte aligned.
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]. When the number of bits of the source operand is different from that of the destination operand, the data type with more bytes is used for the computation. For example, if the source operand is of the half type and the destination operand is of the int8_t type, half is used to compute the mask.
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. For details about this parameter, see High-dimensional Sharding APIs.
repeatParams	Input	Parameters that control the operand address strides. They are of the BinaryRepeatParams type, and contain such parameters as those that specify the address stride of the operand for the same data block between adjacent iterations and 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.

**Table 3** Precision conversion rules
Source Operand	Destination Operand	Type Conversion Mode
float	half	The source operand is cast to the values representable by the half format in CAST_NONE mode, and then stored in the destination operand in half format (with overflow handled by saturation by default).
half	int8_t	The source operand is rounded in CAST_NONE mode, and then stored in the destination operand in int8_t format (with overflow handled by saturation by default).
int16_t	int8_t	The source operand is cast to the values representable by the int8_t format in CAST_NONE mode, and then stored in the destination operand in int8_t format (with overflow handled by saturation by default).
int64_t	float	The source operand is rounded in CAST_NONE mode, and then stored in the destination operand in float format (with overflow handled by saturation by default).
int64_t	int32_t	The source operand is cast to the values representable by the int32_t format in CAST_NONE mode, and then stored in the destination operand in int32_t format (with overflow handled by saturation by default).

Returns

None

Restrictions

For the Atlas 350 Accelerator Card, int64_t supports only the APIs that compute the first n pieces of data in a tensor.

Examples

In the examples, srcLocal is of half type and dstLocal is of int8_t type. mask is computed based on the half type.

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

        
             uint64_t mask = 256 / sizeof(half); // 128
// repeatTime = 4. 128 elements are computed in one iteration, and 512 elements are computed in total.
// dstBlkStride, src0BlkStride, src1BlkStride = 1. Data is continuously read and written in a single iteration.
// dstRepStride = 4, src0RepStride, src1RepStride = 8. Data is continuously read and written between adjacent iterations.
AscendC::AddReluCast(dstLocal, src0Local, src1Local, mask, 4, { 1, 1, 1, 4, 8, 8 });

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

        
             uint64_t mask[2] = { UINT64_MAX, UINT64_MAX };
// repeatTime = 4. 128 elements are computed in one iteration, and 512 elements are computed in total.
// dstBlkStride, src0BlkStride, src1BlkStride = 1. Data is continuously read and written in a single iteration.
// dstRepStride = 4, src0RepStride, src1RepStride = 8. Data is continuously read and written between adjacent iterations.
AscendC::AddReluCast(dstLocal, src0Local, src1Local, mask, 4, { 1, 1, 1, 4, 8, 8 });

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

        
             AscendC::AddReluCast(dstLocal, src0Local, src1Local, 512);

Result example:

Input (src0Local): [1 1 3 ... 512]
Input (src1Local): [0 0.5 -4 ... -513]
Output (dstLocal): [1 2 0 ... 0]

Parent topic: Compound Computation