文档

LayerNorm Tiling

Function Usage

Ascend C provides a group of LayerNorm Tiling APIs for users to obtain the tiling parameter required for LayerNorm kernel computation.

To obtain the tiling parameter, perform the following two steps:

  1. Use the GetLayerNormMaxMinTmpSize API to obtain the maximum and minimum temporary space sizes required for LayerNorm computation. This is to ensure the reasonable allocation of computing space.
    To perform LayerNorm computation in the kernel, developers need to reserve or allocate the temporary space. This API is used to obtain the maximum and minimum sizes of the temporary space to be reserved or allocated on the host. Developers can select a proper size within this range as the tiling parameter and pass it to the kernel.
    • To ensure correct functions, the temporary space to be reserved or allocated cannot be less than the minimum temporary space.
    • Within the range between the minimum and maximum, as the temporary space increases, the API computing performance in the kernel can be optimized to some extent. To achieve better performance, reserve or allocate the space based on the actual buffer usage.
  2. Use the GetLayerNormNDTillingInfo API to obtain the tiling parameter required by the LayerNorm kernel API, and pass the input shape, remaining space for LayerNorm computation, and computation data types.

    The definition of the LayerNormTiling structure is as follows. You do not need to pay attention to the specific information of this tiling structure, but only need to pass it to the kernel and directly use it through LayerNorm high-level APIs.

     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
    struct LayerNormTiling {
    uint32_t bLength = 0;
    uint32_t sLength = 0;
    uint32_t hLength = 0;
    uint32_t originalHLength = 0;
    uint32_t inputXSize = 0;
    uint32_t meanVarSize = 0;
    uint32_t numberOfTmpBuf = 0;
    uint32_t meanTmpTensorPos = 0;
    uint32_t meanTmpTensorSize = 0;
    uint32_t varianceTmpTensorPos = 0;
    uint32_t varianceTmpTensorSize = 0;
    uint32_t tmpBufSize = 0;
    uint32_t oneTmpSize = 0;
    uint32_t firstTmpStartPos = 0;
    uint32_t secondTmpStartPos = 0;
    uint32_t thirdTmpStartPos = 0;
    uint32_t loopRound = 0;
    uint32_t inputRoundSize = 0;
    uint32_t inputTailSize = 0;
    uint32_t inputTailPos = 0;
    uint32_t meanVarRoundSize = 0;
    uint32_t meanVarTailSize = 0;
    uint32_t meanVarTailPos = 0;
    uint32_t bshCurLength = 0;
    uint32_t bsCurLength = 0;
    float lastDimValueBack = 0.0;
};

Prototype

  • GetLayerNormMaxMinTmpSize API
    • Temporary space required by the LayerNorm API for outputting the normalization result, mean, and variance
      1
      void GetLayerNormMaxMinTmpSize(const ge::Shape& srcShape, const uint32_t typeSize, const bool isReuseSource, uint32_t& maxValue, uint32_t& minValue)
  • GetLayerNormNDTillingInfo API
    • Tiling parameters required by the LayerNorm API for outputting the normalization result, mean, and variance
      1
      void GetLayerNormNDTillingInfo(const ge::Shape& srcShape, const uint32_t stackBufferSize, const uint32_t typeSize, const bool isReuseSource, optiling::LayerNormTiling& tilling)

Parameters

Table 1 GetLayerNormMaxMinTmpSize API parameters

API

Input/Output

Function

srcShape

Input
  • LayerNorm API for outputting the normalization result, mean, and variance:

    Shape {B, S, storageHLength, originHLength} of input data inputX, including the shape information of the current inputX and the original shape information before address alignment (if the H axis padding operation is performed).

    In scenarios where the API is supported, the values of storageHLength and originHLength must be the same.

typeSize

Input

Data type size of the input data inputX. The unit is byte. For example, if the input data type of the operator is half, set this parameter to 2.

isReuseSource

Input

Whether to reuse the buffer space of the source operand, which must be the same as that of the LayerNorm API.

maxValue

Output

Tiling information (maximum temporary space size) required by the LayerNorm API.

Maximum size of the temporary space required by LayerNorm computation. Any space exceeding this value will not be utilized by the API. Within the range between the minimum and maximum, as the temporary space increases, the API computing performance in the kernel can be optimized to some extent. To achieve better performance, reserve or allocate the space based on the actual buffer usage.

NOTE:

maxValue is for reference only and may be larger than the remaining space of the Unified Buffer. In this case, select a proper temporary space size based on the remaining space of the Unified Buffer.

minValue

Output

Tiling information (minimum temporary space size) required by the LayerNorm API.

Minimum size of the temporary space required by LayerNorm computation. To ensure correct functions, the size of the temporary space to be reserved or allocated during API computation cannot be less than the value of this parameter.
Table 2 GetLayerNormNDTillingInfo parameters

Parameter

Input/Output

Meaning

srcShape

Input
  • LayerNorm API for outputting the normalization result, mean, and variance:

    Shape {B, S, storageHLength, originHLength} of input data inputX, including the shape information of the current inputX and the original shape information before address alignment (if the H axis padding operation is performed).

stackBufferSize

Input

Size of the space that can be used by the LayerNorm API. The unit is byte.

typeSize

Input

Data type size of operator inputs. The unit is byte. For example, if the input data type of the operator is half, set this parameter to 2.

isReuseSource

Input

Whether the buffer space of inputX can be reused.

tilling

Output

Tilling information of input data.

Returns

None

Example

The following example describes the process of obtaining the tiling parameter on the host and the method of using the parameter on the kernel when LayerNorm high-level API for output variance is used. In this example, the shape size of the input tensor is [2, 16, 64], and the input data type is half.

  1. Add the LayerNormTiling structure parameter to the TilingData structure to function as a field.
    1
2
3
4
5
6
    BEGIN_TILING_DATA_DEF(TilingData)               // Register a tiling class and uses the tiling name as the input parameter.
  TILING_DATA_FIELD_DEF(uint32_t, totalLength); // Add the tiling field to compute the total data volume.
  TILING_DATA_FIELD_DEF(uint32_t, tileNum);     // Add the tiling field that specifies the total number of data blocks to be computed on each core.
  ...                                           // Add other tiling fields.
  TILING_DATA_FIELD_DEF_STRUCT(LayerNormTiling, layernormTilingData); // Add the LayerNormTiling structure parameter to the TilingData structure.
END_TILING_DATA_DEF;
  2. The tiling implementation function first calls the GetLayerNormMaxMinTmpSize API to obtain the maximum and minimum temporary space sizes required by the LayerNorm API to complete computation, sets an appropriate space size based on this range and the actual buffer usage, and then calls the GetLayerNormNDTillingInfo API to obtain the tiling parameter required by the LayerNorm kernel API based on the input shape and remaining size of computing space.
     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
    namespace optiling {
const uint32_t BLOCK_DIM = 8;
const uint32_t TILE_NUM = 8;
static ge::graphStatus TilingFunc(gert::TilingContext* context)
{
    TilingData tiling;
    uint32_t totalLength = context->GetInputTensor(0)->GetShapeSize();
    context->SetBlockDim(BLOCK_DIM);
    tiling.set_totalLength(totalLength);
    tiling.set_tileNum(TILE_NUM);
    // Set other tiling parameters.
    ...
    // {B, S, storageHLength, originHLength}
    std::vector<int64_t> shapeVec = {2, 16, 64, 64};
    ge::Shape srcShape(shapeVec);
    // This example is for reference only. Use GetLayerNormMaxMinTmpSize to obtain the minimum value and pass it to ensure correct functionality. Developers can pass a proper space size as required.
    uint32_t max;
    uint32_t min;
    AscendC::GetLayerNormMaxMinTmpSize(srcShape, sizeof(half), false, max, min);
    // Obtain the LayerNorm tiling parameter.
    AscendC::GetLayerNormNDTillingInfo(srcShape, min, sizeof(half), false, tiling.layernormTilingData); 
     ... // Other logic
    tiling.SaveToBuffer(context->GetRawTilingData()->GetData(), context->GetRawTilingData()->GetCapacity());
    context->GetRawTilingData()->SetDataSize(tiling.GetDataSize());
    context->SetTilingKey(1);
    return ge::GRAPH_SUCCESS;
}
} // namespace optiling
  3. The kernel calls GET_TILING_DATA in the kernel function to obtain TilingData, and then passes the LayerNormTiling information in TilingData to the LayerNorm API for computation. For details about the complete example on the kernel, see LayerNorm.
    1
2
3
4
5
6
7
8
9
    extern "C" __global__ __aicore__ void func_custom(GM_ADDR x, GM_ADDR y, GM_ADDR z, GM_ADDR workspace, GM_ADDR tiling)
{
    GET_TILING_DATA(tilingData, tiling);
    KernelFunc op;
    op.Init(x, y, z, tilingData.totalLength, tilingData.tileNum,tilingData.layernormTilingData);
    if (TILING_KEY_IS(1)) {
        op.Process();
    }
}
Parent topic: LayerNorm Tiling