GeGLU

Applicability

Function Usage

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 by substituting the GELU approximate formula:

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

Prototype

• Pass 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 or 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 temporary 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 size of the temporary space (BufferSize) to be reserved, use the API provided in GetGeGLUMaxMinTmpSize.

Parameters

Returns

None

Constraints

• 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 cannot overlap that of the source or destination operand.

Examples

For a complete operator sample, see geglu operator sample.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14

// dstLocal: tensor for storing the GeGLU computation result
// src0Local: tensor for storing the GeGLU computation input
// src1Local: tensor for storing the GeGLU computation input
// sharedTmpBuffer: tensor for storing the temporary buffer during GeGLU computation

// Allocate the temporary space through the API framework, all of which is used for computation.
AscendC::GeGLU<srcType, false>(dstLocal, src0Local, src1Local);
// Allocate the temporary space through the API framework, part of which is used for computation, with the number of elements involved in the computation being 512.
AscendC::GeGLU<srcType, false>(dstLocal, src0Local, src1Local, 512);

// Pass the temporary space through the sharedTmpBuffer input parameter, all of which is used for computation.
AscendC::GeGLU<srcType, false>(dstLocal, src0Local, src1Local, sharedTmpBuffer);
// Pass the temporary space through the sharedTmpBuffer input parameter, part of which is used for computation, with the number of elements involved in computation being 512.
AscendC::GeGLU<srcType, false>(dstLocal, src0Local, src1Local, sharedTmpBuffer, 512);

1
2
3
4
5
6

Input (src0Local):
[-2.56 -2.55 -2.54 ... 2.53  2.54  2.55]
Input (src1Local):
[2. 2. 2. ... 2. 2. 2.]
Output (dstLocal):
[-5.004   -4.984   -4.96 ... 4.94     4.96     4.984]