LayerNormGradBeta Tiling

Function Usage

LayerNormGradBeta Tiling provides the following functions:

Obtain the maximum and minimum temporary space sizes reserved or allocated on the host.
To perform LayerNormGradBeta computation in the kernel, developers need to reserve or allocate the temporary space. The GetLayerNormGradBetaMaxMinTmpSize 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.

Use the GetLayerNormGradBetaNDTilingInfo API to obtain the tiling parameter required by the LayerNormGradBeta kernel API, and pass the input shapes, remaining space for LayerNormGradBeta computation, and computation data types.

The definition of the LayerNormGradBetaTiling 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 LayerNormGradBeta high-level APIs.

struct LayerNormGradBetaTiling {
    uint32_t stackBufferSize = 0;
    uint32_t bLength = 0;
    uint32_t sLength = 0;
    uint32_t hLength = 0;
    uint32_t originalHLength = 0;
    uint32_t bshLength = 0;
    uint32_t bsLength = 0;
    uint32_t oneCalSize = 0;
    uint32_t numberOfTmpBuf = 0;
    uint32_t loopRound = 0;
    uint32_t inputTailSize = 0;
    uint32_t inputTailPos = 0;
    uint32_t bsTailSize = 0;
    uint32_t bshCurLength = 0;
    uint32_t bsCurLength = 0;
    uint32_t gammaTempTensorPos = 0;
    uint32_t betaTempTensorPos = 0;
    uint32_t inputDyTmpTensorPos = 0;
    uint32_t resForGammaTmpTensorPos = 0;
    uint32_t reserved = 0;
};

Prototype

void GetLayerNormGradBetaMaxMinTmpSize(const ge::Shape &srcShape, const uint32_t typeSize, const bool isReuseSource, uint32_t &maxValue, uint32_t &minValue)

void GetLayerNormGradBetaNDTilingInfo(const ge::Shape srcShape, const uint32_t stackBufferSize, const uint32_t typeSize, const bool isReuseSource, optiling::LayerNormGradBetaTiling &tiling)

Parameters

**Table 1** **GetLayerNormGradBetaMaxMinTmpSize** API parameters
Parameter	Input/Output	Meaning
srcShape	Input	Shape {B, S, storageHLength, originHLength} of input data inputDy, including the shape information of the current inputDy 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 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 to reuse the buffer space of the source operand, which must be the same as that of the LayerNorm API.
maxValue	Output	Maximum size of the temporary space required by LayerNormGradBeta 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. If the maximum space size is 0, no temporary space is required. 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	Minimum size of the temporary space required by LayerNormGradBeta 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. If the minimum space size is 0, no temporary space is required.

**Table 2** **GetLayerNormGradBetaNDTilingInfo** API parameters
Parameter	Input/Output	Meaning
srcShape	Input	Shape of input data inputDy, including the current input shape information and the original shape information before address alignment.
stackBufferSize	Input	Size of the space that can be used by the API. The unit is the number of elements.
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 inputDy 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 LayerNormGradBeta high-level APIs are used. In this example, the shape size of the input tensor is [2, 16, 64], and the input data type is half.

Add the LayerNormGradBetaTiling structure parameter to the TilingData structure to function as a field.

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(LayerNormGradBetaTiling, layernormGradBetaTilingData); // Add the LayerNormGradBetaTiling structure parameter to the TilingData structure.
END_TILING_DATA_DEF;

The tiling implementation function first calls the GetLayerNormGradBetaMaxMinTmpSize API to obtain the maximum and minimum temporary space sizes required by the LayerNormGradBeta API to complete computation, sets an appropriate space size based on this range and the actual buffer usage, and then calls the GetLayerNormGradBetaNDTillingInfo API to obtain the tiling parameter required by the LayerNormGradBeta kernel API based on the input shape and remaining size of computing space.

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 only for reference. Use GetLayerNormGradBetaMaxMinTmpSize 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::GetLayerNormGradBetaMaxMinTmpSize(srcShape, sizeof(half), false, max, min);
    // Obtain LayernormGradBeta tiling parameters.
    AscendC::GetLayerNormGradBetaNDTillingInfo(srcShape, min, sizeof(half), false, tiling.layernormGradBetaTilingData); 
     ... // Other logic
    tiling.SaveToBuffer(context->GetRawTilingData()->GetData(), context->GetRawTilingData()->GetCapacity());
    context->GetRawTilingData()->SetDataSize(tiling.GetDataSize());
    context->SetTilingKey(1);
    return ge::GRAPH_SUCCESS;
}
} // namespace optiling

The kernel calls GET_TILING_DATA in the kernel function to obtain TilingData, and then passes the LayerNormGradBetaTiling information in TilingData to the LayerNormGradBeta API for computation. For details about the complete example on the kernel, see LayerNormGradBeta.

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.layernormGradBetaTilingData);
    if (TILING_KEY_IS(1)) {
        op.Process();
    }
}

Parent topic: LayerNorm Tiling