Function maxPoolWithArgmax

tfjs-backend-webgpu/src/kernels/MaxPoolWithArgmax.ts:24–64

(args: {
  inputs: MaxPoolWithArgmaxInputs,
  attrs: MaxPoolWithArgmaxAttrs,
  backend: WebGPUBackend
})

import {Pool2DProgram} from '../pool_webgpu';

export function maxPoolWithArgmax(args: {
  inputs: MaxPoolWithArgmaxInputs,
  attrs: MaxPoolWithArgmaxAttrs,
  backend: WebGPUBackend
}): TensorInfo[] {
  const {inputs, backend, attrs} = args;
  const {filterSize, strides, pad, includeBatchInIndex} = attrs;
  const {x} = inputs;

  util.assert(
      x.shape.length === 4,
      () => `Error in maxPool: input must be rank 4 but got rank ${
          x.shape.length}.`);
  const dilations: [number, number] = [1, 1];
  util.assert(
      backend_util.eitherStridesOrDilationsAreOne(strides, dilations),
      () => 'Error in maxPool: Either strides or dilations must be 1. ' +
          `Got strides ${strides} and dilations '${dilations}'`);

  const convInfo = backend_util.computePool2DInfo(
      x.shape as [number, number, number, number], filterSize, strides,
      dilations, pad);

  const uniformData = [
    {type: 'int32', data: [convInfo.strideHeight, convInfo.strideWidth]},
    {type: 'int32', data: [convInfo.padInfo.top, convInfo.padInfo.left]},
    {type: 'int32', data: [convInfo.dilationHeight, convInfo.dilationWidth]},
    {type: 'int32', data: [convInfo.inHeight, convInfo.inWidth]}, {
      type: 'int32',
      data: [convInfo.effectiveFilterHeight, convInfo.effectiveFilterWidth]
    }
  ];
  let program = new Pool2DProgram(convInfo, 'max', false);
  const poolOutput =
      backend.runWebGPUProgram(program, [x], x.dtype, uniformData);

  program = new Pool2DProgram(convInfo, 'max', true, true, includeBatchInIndex);
  const indexOutput =
      backend.runWebGPUProgram(program, [x], 'int32', uniformData);
  return [poolOutput, indexOutput];
}

export const maxPoolWithArgmaxConfig: KernelConfig = {
  kernelName: MaxPoolWithArgmax,