Sobel filter shader for Unity.

SobelFilter.shader

Shader "Mattatz/SobelFilter" {

	Properties {
		_MainTex ("Base (RGB)", 2D) = "white" {}
		_DeltaX ("Delta X", Float) = 0.01
		_DeltaY ("Delta Y", Float) = 0.01
	}
	
	SubShader {
		Tags { "RenderType"="Opaque" }
		LOD 200
		
		CGINCLUDE
		
		#include "UnityCG.cginc"
		
		sampler2D _MainTex;
		float _DeltaX;
		float _DeltaY;
		
		float sobel (sampler2D tex, float2 uv) {
			float2 delta = float2(_DeltaX, _DeltaY);
			
			float4 hr = float4(0, 0, 0, 0);
			float4 vt = float4(0, 0, 0, 0);
			
			hr += tex2D(tex, (uv + float2(-1.0, -1.0) * delta)) *  1.0;
			hr += tex2D(tex, (uv + float2( 0.0, -1.0) * delta)) *  0.0;
			hr += tex2D(tex, (uv + float2( 1.0, -1.0) * delta)) * -1.0;
			hr += tex2D(tex, (uv + float2(-1.0,  0.0) * delta)) *  2.0;
			hr += tex2D(tex, (uv + float2( 0.0,  0.0) * delta)) *  0.0;
			hr += tex2D(tex, (uv + float2( 1.0,  0.0) * delta)) * -2.0;
			hr += tex2D(tex, (uv + float2(-1.0,  1.0) * delta)) *  1.0;
			hr += tex2D(tex, (uv + float2( 0.0,  1.0) * delta)) *  0.0;
			hr += tex2D(tex, (uv + float2( 1.0,  1.0) * delta)) * -1.0;
			
			vt += tex2D(tex, (uv + float2(-1.0, -1.0) * delta)) *  1.0;
			vt += tex2D(tex, (uv + float2( 0.0, -1.0) * delta)) *  2.0;
			vt += tex2D(tex, (uv + float2( 1.0, -1.0) * delta)) *  1.0;
			vt += tex2D(tex, (uv + float2(-1.0,  0.0) * delta)) *  0.0;
			vt += tex2D(tex, (uv + float2( 0.0,  0.0) * delta)) *  0.0;
			vt += tex2D(tex, (uv + float2( 1.0,  0.0) * delta)) *  0.0;
			vt += tex2D(tex, (uv + float2(-1.0,  1.0) * delta)) * -1.0;
			vt += tex2D(tex, (uv + float2( 0.0,  1.0) * delta)) * -2.0;
			vt += tex2D(tex, (uv + float2( 1.0,  1.0) * delta)) * -1.0;
			
			return sqrt(hr * hr + vt * vt);
		}
		
		float4 frag (v2f_img IN) : COLOR {
			float s = sobel(_MainTex, IN.uv);
			return float4(s, s, s, 1);
		}
		
		ENDCG
		
		Pass {
			CGPROGRAM
			#pragma vertex vert_img
			#pragma fragment frag
			ENDCG
		}
		
	} 
	FallBack "Diffuse"
}

Neat!

Very nice thank you very much!

You DO realize that about half of the texture lookups in your sobel function are multiplied by zero, and thus do absolutely nothing? Anyway, this is pretty neat for reference.

how to implement it

You DO realize that about half of the texture lookups in your sobel function are multiplied by zero, and thus do absolutely nothing? Anyway, this is pretty neat for reference.

You DO realize that these multiplications are the weights needed for a Sobel Filter? Read up on Convolution Filters, you will understand what I mean.

You DO realize that these multiplications are the weights needed for a Sobel Filter? Read up on Convolution Filters, you will understand what I mean.

Yes of course, but if a sample is multiplied by a constant value of zero, it will have no influence on the sum, and therefore it would be better to omit this sample to avoid an expensive texture lookup. There is a difference between the mathematical description of a kernel and how it is appropriate to implement it.

Maybe my phrasing was unecessarily rude - I just ran the code through DXC Compiler Explorer, and I can see that it will optimize away those lines anyway, so it's not like it matters.