onnxruntime/js/web/lib/wasm/jsep/webgpu/ops/layer-norm.ts

// Copyright (c) Microsoft Corporation. All rights reserved.
// Licensed under the MIT License.

import {DataType} from '../../../wasm-common';
import {TensorView} from '../../tensor-view';
import {ShapeUtil} from '../../util';
import {AttributeWithCacheKey, createAttributeWithCacheKey} from '../attribute-with-cache-key';
import {ComputeContext, ProgramInfo} from '../types';

import {castToF32, fillVector, getMaxComponents, inputVariable, outputVariable, ShaderHelper, sumVector, tensorTypeToWsglStorageType,} from './common';

export interface LayerNormAttributes extends AttributeWithCacheKey {
  axis: number;
  epsilon: number;
}

const validateInputs = (inputs: readonly TensorView[]): void => {
  if (!inputs || inputs.length < 2) {
    throw new Error('layerNorm requires at least 2 inputs.');
  }
};

const createLayerNormProgramInfo =
    (inputs: readonly TensorView[], attributes: LayerNormAttributes, outputCount: number): ProgramInfo => {
      const xShape = inputs[0].dims;
      const scale = inputs[1];
      const bias = inputs[2];

      const outputShape = xShape;
      const axis = ShapeUtil.normalizeAxis(attributes.axis, xShape.length);
      const normCount = ShapeUtil.sizeToDimension(xShape, axis);
      const normSize = ShapeUtil.sizeFromDimension(xShape, axis);

      const scaleSize = ShapeUtil.size(scale.dims);
      const biasSize = bias ? ShapeUtil.size(bias.dims) : 0;
      if (scaleSize !== normSize || (bias && biasSize !== normSize)) {
        throw new Error(`Size of X.shape()[axis:] == ${normSize}.
       Size of scale and bias (if provided) must match this.
       Got scale size of ${scaleSize} and bias size of ${biasSize}`);
      }

      const meanInvStdDevDim = [];
      for (let i = 0; i < xShape.length; ++i) {
        if (i < axis) {
          meanInvStdDevDim.push(xShape[i]);
        } else {
          meanInvStdDevDim.push(1);
        }
      }

      const components = getMaxComponents(normSize);
      const dataType = tensorTypeToWsglStorageType(inputs[0].dataType);
      const variables = [
        inputVariable('x', inputs[0].dataType, inputs[0].dims, components),
        inputVariable('scale', scale.dataType, scale.dims, components),
      ];
      if (bias) {
        variables.push(inputVariable('bias', bias.dataType, bias.dims, components));
      }
      variables.push(outputVariable('output', inputs[0].dataType, outputShape, components));

      const hasMeanDataOutput = outputCount > 1;
      const hasInvStdOutput = outputCount > 2;

      if (hasMeanDataOutput) {
        variables.push(outputVariable('meanDataOutput', DataType.float, meanInvStdDevDim));
      }
      if (hasInvStdOutput) {
        variables.push(outputVariable('invStdOutput', DataType.float, meanInvStdDevDim));
      }

      const getShaderSource = (shaderHelper: ShaderHelper) => `
  const normSize: f32 = ${normSize};
  const normSizeVectorized: u32 = ${normSize / components};
  const epsilon: f32 = ${attributes.epsilon};

  ${shaderHelper.declareVariables(...variables)}
  ${shaderHelper.mainStart()}
    ${shaderHelper.guardAgainstOutOfBoundsWorkgroupSizes(normCount)}
    let offset = global_idx * normSizeVectorized;
    var meanVector = ${fillVector('f32', components)};
    var meanSquareVector = ${fillVector('f32', components)};

    for (var h: u32 = 0u; h < normSizeVectorized; h++) {
      let value = ${castToF32(dataType, components, 'x[h + offset]')};
      meanVector += value;
      meanSquareVector += value * value;
    }
    let mean = ${sumVector('meanVector', components)} / normSize;
    let meanSquare = sqrt(${sumVector('meanSquareVector', components)} 
      / normSize - mean * mean + epsilon);

    for (var j: u32 = 0; j < normSizeVectorized; j++) {
      let f32input = ${castToF32(dataType, components, 'x[j + offset]')};
      let f32scale = ${castToF32(dataType, components, 'scale[j]')};
      output[j + offset] = ${variables[0].type.value}((f32input - mean) / meanSquare * f32scale
        ${bias ? `+ ${castToF32(dataType, components, 'bias[j]')}` : ''}
      );
    }

    ${hasMeanDataOutput ? 'meanDataOutput[global_idx] = mean' : ''};
    ${hasInvStdOutput ? 'invStdOutput[global_idx] = 1 / meanSquare' : ''};
  }`;
      const outputs = [{dims: outputShape, dataType: inputs[0].dataType}];
      if (hasMeanDataOutput) {
        outputs.push({dims: meanInvStdDevDim, dataType: DataType.float});
      }
      if (hasInvStdOutput) {
        outputs.push({dims: meanInvStdDevDim, dataType: DataType.float});
      }

      return {
        name: 'LayerNormalization',
        shaderCache: {hint: `${attributes.cacheKey}|${outputCount}|${inputs.length}`},
        getRunData: () => ({outputs, dispatchGroup: {x: Math.ceil(normCount / 64 /* workgroup size */)}}),
        getShaderSource,
      };
    };

export const parseLayerNormAttributes = (attributes: LayerNormAttributes): LayerNormAttributes =>
    createAttributeWithCacheKey({axis: attributes.axis, epsilon: attributes.epsilon});

export const layerNorm = (context: ComputeContext, attributes: LayerNormAttributes): void => {
  validateInputs(context.inputs);
  context.compute(createLayerNormProgramInfo(context.inputs, attributes, context.outputCount));
};