conv_backprop_gpu_depthwise.ts

/**
 * @license
 * Copyright 2018 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} from '@tensorflow/tfjs-core';
import {GPGPUProgram} from './gpgpu_math';

export class DepthwiseConv2DDerFilterProgram implements GPGPUProgram {
  variableNames = ['x', 'dy'];
  outputShape: number[];
  userCode: string;

  constructor(convInfo: backend_util.Conv2DInfo) {
    this.outputShape = convInfo.filterShape;

    const strideHeight = convInfo.strideHeight;
    const strideWidth = convInfo.strideWidth;
    const padTop = convInfo.padInfo.top;
    const padLeft = convInfo.padInfo.left;
    const channelMul = convInfo.outChannels / convInfo.inChannels;

    this.userCode = `
      void main() {
        ivec4 coords = getOutputCoords();
        int wR = coords.x;
        int wC = coords.y;
        int d1 = coords.z;
        int dm = coords.w;
        int d2 = d1 * ${channelMul} + dm;

        float dotProd = 0.0;

        // TO DO: Vec4 over the batch size
        for (int b = 0; b < ${convInfo.batchSize}; b++) {
          for (int yR = 0; yR < ${convInfo.outHeight}; yR++) {
            int xR = wR + yR * ${strideHeight} - ${padTop};

            if (xR < 0 || xR >= ${convInfo.inHeight}) {
              continue;
            }

            for (int yC = 0; yC < ${convInfo.outWidth}; yC++) {
              int xC = wC + yC * ${strideWidth} - ${padLeft};

              if (xC < 0 || xC >= ${convInfo.inWidth}) {
                continue;
              }

              float dyValue = getDy(b, yR, yC, d2);
              float xValue = getX(b, xR, xC, d1);
              dotProd += (xValue * dyValue);
            }
          }
        }
        setOutput(dotProd);
      }
    `;
  }
}

export class DepthwiseConv2DDerInputProgram implements GPGPUProgram {
  variableNames = ['dy', 'W'];
  outputShape: number[];
  userCode: string;

  constructor(convInfo: backend_util.Conv2DInfo) {
    this.outputShape = convInfo.inShape;

    const filterHeight = convInfo.filterHeight;
    const filterWidth = convInfo.filterWidth;
    const strideHeight = convInfo.strideHeight;
    const strideWidth = convInfo.strideWidth;

    const padTop = filterHeight - 1 - convInfo.padInfo.top;
    const padLeft = filterWidth - 1 - convInfo.padInfo.left;
    const channelMul = convInfo.outChannels / convInfo.inChannels;

    this.userCode = `
      const ivec2 pads = ivec2(${padTop}, ${padLeft});

      void main() {
        ivec4 coords = getOutputCoords();
        int batch = coords[0];
        int d1 = coords[3];
        ivec2 dyCorner = coords.yz - pads;
        int dyRCorner = dyCorner.x;
        int dyCCorner = dyCorner.y;

        float dotProd = 0.0;

        for (int wR = 0; wR < ${filterHeight}; wR++) {
          float dyR = float(dyRCorner + wR) / ${strideHeight}.0;

          if (dyR < 0.0 || dyR >= ${convInfo.outHeight}.0 || fract(dyR) > 0.0) {
            continue;
          }
          int idyR = int(dyR);

          int wRPerm = ${filterHeight} - 1 - wR;

          for (int wC = 0; wC < ${filterWidth}; wC++) {
            float dyC = float(dyCCorner + wC) / ${strideWidth}.0;

            if (dyC < 0.0 || dyC >= ${convInfo.outWidth}.0 ||
                fract(dyC) > 0.0) {
              continue;
            }
            int idyC = int(dyC);

            int wCPerm = ${filterWidth} - 1 - wC;

            // TO DO: Vec4 over the channelMul
            for (int dm = 0; dm < ${channelMul}; dm++) {
              int d2 = d1 * ${channelMul} + dm;
              float xValue = getDy(batch, idyR, idyC, d2);
              float wValue = getW(wRPerm, wCPerm, d1, dm);
              dotProd += xValue * wValue;
            }
          }
        }
        setOutput(dotProd);
      }
    `;
  }
}