文档

GeGLU

Applicability

Product

Supported

Atlas A3 training products/Atlas A3 inference products

√

Atlas A2 training products/Atlas A2 inference products

√

Atlas 200I/500 A2 inference products

x

Atlas inference product's AI Core

√

Atlas inference product's Vector Core

x

Atlas training products

x

Function

GeGLU is a GLU variant that uses GELU as the activation function. Below is the formula.

The formula for calculating the GELU activation function is as follows:

erf in the foregoing formula is an error function:

The error function does not have an analytical expression and utilizes the tanh approximate expression, which is widely accepted in the industry:

The GeGLU expression can be derived if the GELU approximate formula is substituted:

a equals –0.0713548162726 and b equals 2.2363860002236e1. x1 and x0 are elements in srcTensor1 and srcTensor0.

Prototype

  • Pass to the temporary space through the sharedTmpBuffer input parameter.
    • All or part of the source operand tensors are involved in computation.
      1
2
      template <typename T, bool isReuseSource = false>
__aicore__ inline void GeGLU(const LocalTensor<T>& dstTensor, const LocalTensor<T>& srcTensor0, const LocalTensor<T>& srcTensor1, const LocalTensor<uint8_t>& sharedTmpBuffer, uint32_t calCount)
    • All source operand tensors are involved in computation.
      1
2
      template <typename T, bool isReuseSource = false>
__aicore__ inline void GeGLU(const LocalTensor<T>& dstTensor, const LocalTensor<T>& srcTensor0, const LocalTensor<T>& srcTensor1, const LocalTensor<uint8_t>& sharedTmpBuffer)
  • Allocate the temporary space through the API framework.
    • All or part of the source operand tensors are involved in computation.
      1
2
      template <typename T, bool isReuseSource = false>
__aicore__ inline void GeGLU(const LocalTensor<T>& dstTensor, const LocalTensor<T>& srcTensor0, const LocalTensor<T>& srcTensor1, uint32_t calCount)
    • All source operand tensors are involved in computation.
      1
2
      template <typename T, bool isReuseSource = false>
__aicore__ inline void GeGLU(const LocalTensor<T>& dstTensor, const LocalTensor<T>& srcTensor0, const LocalTensor<T>& srcTensor1)

Due to the complex mathematical computation involved in the internal implementation of this API, extra temporary space is required to store intermediate variables generated during computation. The temporary space can be passed by developers through the sharedTmpBuffer input parameter or allocated through the API framework.

  • When the sharedTmpBuffer input parameter is used for passing the temporary space, the tensor serves as the temporary space. In this case, the API framework is not required for temporary space allocation. This enables developers to manage the sharedTmpBuffer space and reuse the buffer after calling the API, so that the buffer is not repeatedly allocated and deallocated, improving the flexibility and buffer utilization.
  • When the API framework is used for temporary space allocation, developers do not need to allocate the space, but must reserve the required size for the space.

If sharedTmpBuffer is used, developers must allocate space for the tensor. If the API framework is used, developers must reserve the temporary space. To obtain the temporary space (BufferSize) to be reserved, use the API provided in GetGeGLUMaxMinTmpSize.

Parameters

Table 1 Template parameters

Parameter

Description

T

Data type of the operand.

For the Atlas A3 training products/Atlas A3 inference products, 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 inference product's AI Core, the supported data types are half and float.

isReuseSource

Whether the source operand can be modified. This parameter is reserved. Pass the default value false.
Table 2 API parameters

Parameter

Input/Output

Description

dstTensor

Output

Destination operand.

The type is LocalTensor, and the supported TPosition is VECIN, VECCALC, or VECOUT.

srcTensor0/

srcTensor1

Input

Source operand.

The source operand must have the same data type as the destination operand.

The type is LocalTensor, and the supported TPosition is VECIN, VECCALC, or VECOUT.

sharedTmpBuffer

Input

Temporary buffer.

The type is LocalTensor, and the supported TPosition is VECIN, VECCALC, or VECOUT.

This parameter is used to store intermediate variables during complex computation in GeGLU and is provided by developers.

For details about how to obtain the temporary space size (BufferSize), see GetGeGLUMaxMinTmpSize.

calCount

Input

Number of elements involved in the computation.

Returns

None

Restrictions

  • For details about the operand address alignment requirements, see General Address Alignment Restrictions.
  • The source operand address must not overlap the destination operand address.
  • Currently, only the ND format is supported.
  • The address of sharedTmpBuffer must not overlap that of the source or destination operand.

Example

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
#include "kernel_operator.h"

template <typename srcType>
class KernelGeGLU
{
public:
    __aicore__ inline KernelGeGLU() {}
    __aicore__ inline void Init(GM_ADDR src0Gm, GM_ADDR src1Gm, GM_ADDR dstGm, uint32_t inputSize,
                                uint32_t tmpBufSize)
    {
        dataSize = inputSize;
        uint32_t bufSize = 4 x tmpBufSize;
        src0Global.SetGlobalBuffer(reinterpret_cast<__gm__ srcType *>(src0Gm), dataSize);
        src1Global.SetGlobalBuffer(reinterpret_cast<__gm__ srcType *>(src1Gm), dataSize);
        dstGlobal.SetGlobalBuffer(reinterpret_cast<__gm__ srcType *>(dstGm), dataSize);
        pipe.InitBuffer(inQueue0, 1, dataSize * sizeof(srcType));
        pipe.InitBuffer(inQueue1, 1, dataSize * sizeof(srcType));
        pipe.InitBuffer(outQueue, 1, dataSize * sizeof(srcType));
        if ((sizeof(srcType) == sizeof(half)) && (tmpBufSize > 0))
        {
            pipe.InitBuffer(buf, bufSize * sizeof(srcType));
        }
    }
    __aicore__ inline void Process(uint32_t tmpBufSize, uint32_t calCount)
    {
        CopyIn();
        Compute(tmpBufSize, calCount);
        CopyOut();
    }

private:
    __aicore__ inline void CopyIn()
    {
        AscendC::LocalTensor<srcType> src0Local = inQueue0.AllocTensor<srcType>();
        AscendC::LocalTensor<srcType> src1Local = inQueue1.AllocTensor<srcType>();
        AscendC::DataCopy(src0Local, src0Global, dataSize);
        AscendC::DataCopy(src1Local, src1Global, dataSize);
        inQueue0.EnQue(src0Local);
        inQueue1.EnQue(src1Local);
    }
    __aicore__ inline void Compute(uint32_t tmpBufSize, uint32_t calCount)
    {
        AscendC::LocalTensor<srcType> dstLocal = outQueue.AllocTensor<srcType>();
        AscendC::LocalTensor<srcType> src0Local = inQueue0.DeQue<srcType>();
        AscendC::LocalTensor<srcType> src1Local = inQueue1.DeQue<srcType>();
        AscendC::LocalTensor<uint8_t> temp;
        if ((sizeof(srcType) == sizeof(half)) && (tmpBufSize > 0)) {
            temp = buf.Get<uint8_t>();
        }
        if ((tmpBufSize > 0) && (calCount > 0)) {
            AscendC::GeGLU<srcType, false>(dstLocal, src0Local, src1Local, temp, calCount);
        } else if (tmpBufSize > 0) {
            AscendC::GeGLU<srcType, false>(dstLocal, src0Local, src1Local, temp);
        } else if (calCount > 0) {
            AscendC::GeGLU<srcType, false>(dstLocal, src0Local, src1Local, calCount);
        } else {
            AscendC::GeGLU<srcType, false>(dstLocal, src0Local, src1Local);
        }
        outQueue.EnQue<srcType>(dstLocal);
        inQueue0.FreeTensor(src0Local);
        inQueue1.FreeTensor(src1Local);
    }
    __aicore__ inline void CopyOut()
    {
        AscendC::LocalTensor<srcType> dstLocal = outQueue.DeQue<srcType>();
        AscendC::DataCopy(dstGlobal, dstLocal, dataSize);
        outQueue.FreeTensor(dstLocal);
    }

private:
    AscendC::GlobalTensor<srcType> src0Global;
    AscendC::GlobalTensor<srcType> src1Global;
    AscendC::GlobalTensor<srcType> dstGlobal;
    AscendC::TPipe pipe;
    AscendC::TQue<AscendC::TPosition::VECIN, 1> inQueue0;
    AscendC::TQue<AscendC::TPosition::VECIN, 1> inQueue1;
    AscendC::TQue<AscendC::TPosition::VECOUT, 1> outQueue;
    AscendC::TBuf<AscendC::TPosition::VECCALC> buf;
    uint32_t dataSize = 0;
};
template <typename dataType>
__aicore__ void kernel_geglu_operator(GM_ADDR src0Gm, GM_ADDR src1Gm, GM_ADDR dstGm, uint32_t srcSize,
                                      uint32_t tmpBufSize, uint32_t calCount)
{
    KernelGeGLU<dataType> op;
    op.Init(src0Gm, src1Gm, dstGm, srcSize, tmpBufSize);
    op.Process(tmpBufSize, calCount);
}
Result example:
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
Input data (srcTensor0):
[ 1.6025391   3.4765625   3.4316406   3.7539062  -1.3330078   0.72314453
 -3.0078125   0.85498047 -1.3691406   2.6894531  -2.9101562  -3.6992188
 -2.2734375  -2.859375    2.5683594  -1.7802734 ]

Input data (srcTensor1)
[-0.6015625  1.9589844  1.9257812  3.8769531  0.5878906  2.9179688
 -1.8847656  3.2304688  2.8945312  2.4550781  1.3730469 -1.9248047
  0.7919922 -2.5332031 -2.1425781 -2.9433594]

Output data (dstLocal): [-0.263916015625000000 6.640625000000000000 6.429687500000000000
14.554687500000000000 -0.565429687500000000 2.107421875000000000 0.168579101562500000
2.759765625000000000 -3.957031250000000000 6.558593750000000000 -3.656250000000000000
0.192993164062500000 -1.415039062500000000 0.039642333984375000 -0.087890625000000000
0.007740020751953125]
Parent topic: GeGLU APIs
我们使用cookie来确保您的高速浏览体验。继续浏览本站，即表示您同意我们使用cookie。查看详情