LeakyRelu

Applicability

Product	Supported/Unsupported
Atlas A3 training products / Atlas A3 inference products	√
Atlas A2 training products / Atlas A2 inference products	√
Atlas 200I/500 A2 inference products	√
Atlas inference product 's AI Core	√
Atlas inference product 's Vector Core	x
Atlas training products	x

Function Usage

Performs the Leaky Rectified Linear Unit (Leaky ReLU) operation by element. The calculation formula is as follows:

$\text{[math]}$

Leaky ReLU is a common activation function in artificial neural networks. Its mathematical expression is as follows:

The difference between Leaky ReLU and ReLU is that ReLU sets all negative values to zero, while Leaky ReLU assigns a slope to all negative values. The following figure shows the difference between ReLU and Leaky ReLU.

Prototype

Computation of the first n pieces of data of a tensor

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

Contiguous mask mode

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

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 LeakyRelu(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 LeakyRelu(const LocalTensor<T>& dst, const LocalTensor<T>& src, const U& scalarValue, uint64_t mask[], const uint8_t repeatTime, const UnaryRepeatParams& repeatParams)

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 LeakyRelu(const LocalTensor<T>& dst, const LocalTensor<T>& src, const U& scalarValue, 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 's AI Core, the supported data types are half and float. For the Atlas A2 training products / Atlas A2 inference products , the supported data types are half and float. For the Atlas A3 training products / Atlas A3 inference products , the supported data types are half and float. For the Atlas 200I/500 A2 inference products , the supported data types are half and float.
U	Data type of scalarValue. For the Atlas inference product 's AI Core, the supported data types are half and float. For the Atlas A2 training products / Atlas A2 inference products , the supported data types are half and float. For the Atlas A3 training products / Atlas A3 inference products , the supported data types are half and float. For the Atlas 200I/500 A2 inference products , the supported data types are half and float.
isSetMask	Whether to set the mask mode and mask value inside the API. true: indicates that the settings are performed inside the API. The tensor high-dimensional tiling computation API or the API for computing the first n elements of a tensor uses the Normal/Counter mode of the mask. In general, you can retain the default value of isSetMask, which indicates that the mask mode and mask value are set inside the API based on the mask/count parameter passed by the developer. false: indicates that the settings are performed outside the API. For the tensor high-dimensional tiling computation API, in some scenarios with high performance requirements, you need to use SetMaskNorm or SetMaskCount to set the mask mode and use SetVectorMask 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 data elements in a tensor, in scenarios that require high performance, you need to use SetMaskCount to set the mask mode to counter mode and use the SetVectorMask API to set the mask value. The count parameter in the input parameter of this API does not take effect. You are advised to set it to 1. For the following models, the isSetMask parameter in the API for calculating the first n pieces of data in a tensor does not take effect. Retain the default value. Atlas 200I/500 A2 inference products

**Table 2** Parameters
Parameter	Input/Output	Meaning
dst	Output	Destination operand. The type is LocalTensor, and the supported TPosition is VECIN, VECCALC, or VECOUT. The start address of the LocalTensor must be 32-byte aligned.
src	Input	Source operand. The type is LocalTensor, and the supported TPosition is VECIN, VECCALC, or VECOUT. The start address of the LocalTensor must be 32-byte aligned. Its data type must be the same as that of the destination operand.
scalarValue	Input	Source operand, and its data type must be the same as the element type of the tensor in the destination operand.
count	Input	Number of elements involved in the computation.
mask/mask[]	Input	The mask parameter is used to control the elements involved 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 in the computation. The mask is in array form. The array length and the value range of the array elements are related to the data type of the operand. When the operand is 16-bit, the array length is 2. In this case, mask[0] and mask[1] must be in the range of [0, 2⁶⁴ – 1] and cannot be 0 at the same time. When the operand is 32-bit, the array length is 1. In this case, mask[0] must be in the range of (0, 2⁶⁴ – 1]. When the operand is 64-bit, the array length is 1. In this case, mask[0] must be in the range of (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 repeat 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 repeats. For details about this parameter, see High-dimensional Sharding APIs.
repeatParams	Input	Control structure information of element operations. For details, see UnaryRepeatParams.

Returns

None

Restrictions

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

Examples

For more examples, see here.

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

        
         
           
           
             #include "kernel_operator.h"
class KernelBinaryScalar {
public:
    __aicore__ inline KernelBinaryScalar() {}
    __aicore__ inline void Init(__gm__ uint8_t* src, __gm__ uint8_t* dstGm)
    {
        srcGlobal.SetGlobalBuffer((__gm__ half*)src);
        dstGlobal.SetGlobalBuffer((__gm__ half*)dstGm);
        pipe.InitBuffer(inQueueSrc, 1, 512 * sizeof(half));
        pipe.InitBuffer(outQueueDst, 1, 512 * sizeof(half));
    }
    __aicore__ inline void Process()
    {
        CopyIn();
        Compute();
        CopyOut();
    }
private:
    __aicore__ inline void CopyIn()
    {
        AscendC::LocalTensor<half> srcLocal = inQueueSrc.AllocTensor<half>();
        AscendC::DataCopy(srcLocal, srcGlobal, 512);
        inQueueSrc.EnQue(srcLocal);
    }
    __aicore__ inline void Compute()
    {
        AscendC::LocalTensor<half> srcLocal = inQueueSrc.DeQue<half>();
        AscendC::LocalTensor<half> dstLocal = outQueueDst.AllocTensor<half>();

       uint64_t mask = 128;
        half scalar = 2;
        // 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::LeakyRelu(dstLocal, srcLocal, scalar, mask, 4, {1, 1, 8, 8});

        outQueueDst.EnQue<half>(dstLocal);
        inQueueSrc.FreeTensor(srcLocal);
    }
    __aicore__ inline void CopyOut()
    {
        AscendC::LocalTensor<half> dstLocal = outQueueDst.DeQue<half>();
        AscendC::DataCopy(dstGlobal, dstLocal, 512);
        outQueueDst.FreeTensor(dstLocal);
    }
private:
    AscendC::TPipe pipe;
    AscendC::TQue<AscendC::TPosition::VECIN, 1> inQueueSrc;
    AscendC::TQue<AscendC::TPosition::VECOUT, 1> outQueueDst;
    AscendC::GlobalTensor<half> srcGlobal, dstGlobal;
};
extern "C" __global__ __aicore__ void binary_scalar_simple_kernel(__gm__ uint8_t* src, __gm__ uint8_t* dstGm)
{
    KernelBinaryScalar op;
    op.Init(src, dstGm);
    op.Process();
}

            

          

        
       

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

        
             uint64_t mask[2] = { UINT64_MAX, UINT64_MAX };
half scalar = 2;
// repeatTime = 4. 128 elements are processed in a single iteration, and four iterations are required to compute 512 elements.
// dstBlkStride, srcBlkStride = 1. The interval between src0 data addresses involved in calculation 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::LeakyRelu(dstLocal, srcLocal, scalar, mask, 4, {1, 1, 8, 8});

Example of computing the first n data elements of a tensor

        
             half scalar = 2;
AscendC::LeakyRelu(dstLocal, srcLocal, scalar, 512);

Result example:

Input (src0Local): [1 2 3 ... 512]
Input (scalar): 2
Output (dstLocal): [1. 2. 3. ... 512.]

Parent topic: Basic arithmetic operations