Implementation:Tensorflow Tfjs CPU Dilation2D Kernel

Knowledge Sources	Tensorflow_Tfjs
Domains	Deep_Learning, CPU_Backend
Last Updated	2026-02-10 06:00 GMT

Overview

The Dilation2D kernel implements 2D morphological dilation on the CPU backend. For each spatial output position and channel, it computes the maximum of input + filter over the filter window. This operation is analogous to max-pooling but adds the structuring element (filter) values before taking the maximum. The implementation follows the TensorFlow C++ reference kernel.

Code Reference

Source Location

tfjs-backend-cpu/src/kernels/Dilation2D.ts (GitHub)

Signature

The kernel function is defined inline as an anonymous function within the KernelConfig object:

export const dilation2DConfig: KernelConfig = {
  kernelName: Dilation2D,
  backendName: 'cpu',
  kernelFunc: ({inputs, backend, attrs}) => { ... }
};

Imports

import {backend_util, Dilation2D, Dilation2DAttrs, Dilation2DInputs, KernelConfig, TypedArray, util} from '@tensorflow/tfjs-core';
import {MathBackendCPU} from '../backend_cpu';

I/O Contract

Parameter	Type	Description
inputs.x	TensorInfo	4D input tensor of shape [batch, inHeight, inWidth, inChannels] (NHWC format)
inputs.filter	TensorInfo	3D structuring element of shape [filterHeight, filterWidth, depth]
attrs.strides	[number, number]	Stride for each spatial dimension
attrs.pad	'same'	Padding mode
attrs.dilations	[number, number]	Dilation rate for the filter
Return	TensorInfo	Output tensor of shape [batch, outHeight, outWidth, inChannels]

Implementation Details

• Computes dilation metadata via backend_util.computeDilation2DInfo with NHWC data format.
• Allocates an output array of the computed output size using util.getArrayFromDType.
• Iterates over batches, output spatial positions, and channels.
• For each output element, initializes curVal to Number.MIN_SAFE_INTEGER and scans the filter window.
• At each filter position, computes the input coordinate factoring in dilation (hIn = hBeg + h * dilationHeight).
• Skips out-of-bounds positions. For valid positions, computes val = xVals[xIndex] + filterVals[filterIndex] and keeps the maximum.
• Uses util.locToIndex with precomputed strides for flat array indexing.
• Writes the final output via cpuBackend.write.

Usage Examples

// Dilation2D is registered via dilation2DConfig and invoked through tf.dilation2d()
import * as tf from '@tensorflow/tfjs-core';

const result = tf.dilation2d(inputTensor, filterTensor, strides, pad, dilations);

Related Pages

• Principle:Tensorflow_Tfjs_CPU_Kernel_Registration