depthwise_conv2d_vec4_webgpu.ts

/**
 * @license
 * Copyright 2021 Google LLC. All Rights Reserved.
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 * http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 * =============================================================================
 */

import {backend_util, util} from '@tensorflow/tfjs-core';
import {activationFnSnippet, biasActivationSnippet} from './activation_util';
import {getMainHeaderString as main, WebGPUProgram} from './webgpu_program';
import {computeDispatch, flatDispatchLayout} from './webgpu_util';

export class DepthwiseConv2DVec4Program implements WebGPUProgram {
  outputShape: number[];
  shaderKey: string;
  dispatchLayout: {x: number[]};
  dispatch: [number, number, number];
  variableNames = ['x', 'W'];
  uniforms = 'pads : vec2<i32>, inDims : vec2<i32>, virtualWidth : i32,';
  workgroupSize: [number, number, number] = [64, 1, 1];
  workPerThread = 4;
  convInfo: backend_util.Conv2DInfo;
  addBias: boolean;
  activation: backend_util.Activation;
  hasPreluActivation: boolean;
  outputComponent = 4;
  virtualWidth: number;

  constructor(
      convInfo: backend_util.Conv2DInfo, addBias = false,
      activation: backend_util.Activation = null, hasPreluActivation = false) {
    this.outputShape = convInfo.outShape;
    this.virtualWidth = Math.ceil(this.outputShape[2] / this.workPerThread) *
        this.workPerThread;
    const virtualOutputShape = [
      this.outputShape[0], this.outputShape[1], this.virtualWidth,
      this.outputShape[3]
    ];
    this.dispatchLayout = flatDispatchLayout(virtualOutputShape);

    this.dispatch = computeDispatch(
        this.dispatchLayout, virtualOutputShape, this.workgroupSize,
        [this.outputComponent * this.workPerThread, 1, 1]);

    util.assert(
        convInfo.dataFormat === 'channelsLast',
        () => 'TODO: NCHW is unimplemented');

    if (addBias) {
      this.variableNames.push('bias');
    }
    if (hasPreluActivation) {
      this.variableNames.push('preluActivationWeights');
    }

    this.convInfo = convInfo;
    this.addBias = addBias;
    this.activation = activation;
    this.hasPreluActivation = hasPreluActivation;

    this.shaderKey =
        `depthwiseVec4_${activation}_${this.convInfo.filterHeight}_${
            this.convInfo.filterWidth}_${this.convInfo.strideHeight}_${
            this.convInfo.strideWidth}_${this.workPerThread}`;
  }

  getUserCode(): string {
    const xNumber = (this.workPerThread - 1) * this.convInfo.strideWidth +
        this.convInfo.filterWidth;
    const strideHeight = this.convInfo.strideHeight;
    const strideWidth = this.convInfo.strideWidth;

    const userCode = `
      ${activationFnSnippet(this.activation, this.hasPreluActivation, true, 4)}
      fn readX(batch : i32, row : i32, col : i32, channel : i32) -> vec4<f32> {
        var value = vec4<f32>(0.0);
        if (col >=0 && col < uniforms.inDims[1]) {
          value = getX(batch, row, col, channel);
        }
        return value;
      }

      ${main('index')} {
        let width0 = uniforms.outShape[3] / ${this.outputComponent};
        let d1 = (index % width0) * ${this.outputComponent};
        var index1 = index / width0;
        let width1 = uniforms.virtualWidth / ${this.workPerThread};
        let c = (index1 % width1) * ${this.workPerThread};
        index1 = index1 / width1;
        let r = index1 % uniforms.outShape[1];
        let batch = index1 / uniforms.outShape[1];

        let xRCCorner = vec2<i32>(r, c) * vec2<i32>(${strideHeight}, ${
        strideWidth}) - uniforms.pads;

        let xRCorner = xRCCorner.x;
        let xCCorner = xRCCorner.y;
        var xVals : array<vec4<f32>, ${xNumber}>;
        var dotProd : array<vec4<f32>, ${this.workPerThread}>;
        for (var i = 0; i < ${this.workPerThread}; i++) {
          dotProd[i] = vec4<f32>(0.0);
        }

        // Use constant instead of uniform can give better performance.
        for (var wR = 0; wR < ${this.convInfo.filterHeight}; wR = wR + 1) {
          let xR = xRCorner + wR;
          if (xR >=0 && xR < uniforms.inDims[0]) {
            for (var i = 0; i < ${xNumber}; i++) {
              xVals[i] = readX(batch, xR, xCCorner + i, d1);
            }
            for (var wC = 0; wC < ${this.convInfo.filterWidth}; wC = wC + 1) {
              let wValue = getW(wR, wC, d1, 0);
              for (var i = 0; i < ${this.workPerThread}; i++) {
                dotProd[i] = fma(xVals[i * ${
        strideWidth} + wC], wValue, dotProd[i]);
              }
            }
          }
        }

        for (var i = 0; i < ${this.workPerThread}; i = i + 1) {
          let coords = vec4<i32>(batch, r, c + i, d1);
          if (coordsInBounds4D(coords, uniforms.outShape)) {
            var value = dotProd[i];
            ${biasActivationSnippet(this.addBias, this.activation)}
            setOutputAtCoords(coords[0], coords[1], coords[2], coords[3], value);
          }
        }
      }
    `;
    return userCode;
  }
}