Optimized Gaussian blur function with configurable sigma and radius.
- Play around with sigma and radius values to find the optimal parameters.
- Replace the sigma and radius parameters with literal values to let the Unity shader compiler produce efficient code.
#define GAUSSIAN_BLUR_UNROLL 1before including this file.
Include this shader file (path may vary):
// #define GAUSSIAN_BLUR_UNROLL 1
#include "Assets/Shaders/GaussianBlur.hlsl"It is highly recommended to use this variation if you have access to a temporary texture to ping-pong buffer.
The number of texture samples needed for each pass is radius + 1.
For example, a 5x5 kernel has a radius of 2, which means 3 texture samples are needed for each pass.
First pass:
return GaussianBlurHorizontal(_SourceTex, sampler_SourceTex, _SourceTex_TexelSize, i.uv, 1, 3);Second pass:
return GaussianBlurVertical(_SourceTex, sampler_SourceTex, _SourceTex_TexelSize, i.uv, 1, 3);This variant is used to perform Gaussian blur in a single pass. This is useful when it is not possible or is too difficult to prepare a temporary texture for the 2 pass approach.
The number of texture samples needed is (radius + 1)^2.
For example, a 5x5 kernel has a radius of 2, which means 9 texture samples are needed in this case.
This is significantly smaller than the naive implementation, which requires (radius * 2 + 1)^2, but still is an O(n^2) algorithm.
return GaussianBlurSingle(_SourceTex, sampler_SourceTex, _SourceTex_TexelSize, i.uv, 1, 2);I could no longer be bothered writing kernel calculation code on CPU, and pass parameters to the shader for experimenting with sigma and radius, so I wrote this to let the Unity shader compiler generate the code I want.
I also saw various implementations of unoptimized Gaussian blur in many places, which led me to write this.
- The function minimizes the texture fetches by bilinear filtering.
- The produced code will be more efficient if
GAUSSIAN_BLUR_UNROLLis set. However, it won't compile if the radius parameter is not a literal value. - UV calculations can be moved to vertex shader if further optimizations are required.
- Loop unroll will succeed even if the sigma parameter is dynamic. However, the generated code is not very efficient. It is desirable to move these weight calculations to the CPU.
- Vertical pass
- Sigma = 3
- Radius = 6
u_xlat0.x = 0.0;
u_xlat0.y = _SourceTex_TexelSize.y;
u_xlat1 = vec4(-5.35180569, -5.35180569, -3.40398479, -3.40398479) * u_xlat0.xyxy + vs_TEXCOORD0.xyxy;
u_xlat16_2 = texture(_SourceTex, u_xlat1.zw);
u_xlat16_1 = texture(_SourceTex, u_xlat1.xy);
u_xlat2 = u_xlat16_2 * vec4(0.139440298, 0.139440298, 0.139440298, 0.139440298);
u_xlat1 = u_xlat16_1 * vec4(0.0527109653, 0.0527109653, 0.0527109653, 0.0527109653) + u_xlat2;
u_xlat2 = vec4(-1.45842957, -1.45842957, 1.45842957, 1.45842957) * u_xlat0.xyxy + vs_TEXCOORD0.xyxy;
u_xlat0 = vec4(3.40398479, 3.40398479, 5.35180569, 5.35180569) * u_xlat0.xyxy + vs_TEXCOORD0.xyxy;
u_xlat16_3 = texture(_SourceTex, u_xlat2.xy);
u_xlat16_2 = texture(_SourceTex, u_xlat2.zw);
u_xlat1 = u_xlat16_3 * vec4(0.239337295, 0.239337295, 0.239337295, 0.239337295) + u_xlat1;
u_xlat16_3 = texture(_SourceTex, vs_TEXCOORD0.xy);
u_xlat1 = u_xlat16_3 * vec4(0.137022808, 0.137022808, 0.137022808, 0.137022808) + u_xlat1;
u_xlat1 = u_xlat16_2 * vec4(0.239337295, 0.239337295, 0.239337295, 0.239337295) + u_xlat1;
u_xlat16_2 = texture(_SourceTex, u_xlat0.xy);
u_xlat16_0 = texture(_SourceTex, u_xlat0.zw);
u_xlat1 = u_xlat16_2 * vec4(0.139440298, 0.139440298, 0.139440298, 0.139440298) + u_xlat1;
u_xlat0 = u_xlat16_0 * vec4(0.0527109653, 0.0527109653, 0.0527109653, 0.0527109653) + u_xlat1;
SV_Target0 = u_xlat0;- Sigma = 1
- Radius = 2
u_xlat0 = _SourceTex_TexelSize.xyxy * vec4(-1.1824255, -1.1824255, 0.0, -1.1824255) + vs_TEXCOORD0.xyxy;
u_xlat16_1 = texture(_SourceTex, u_xlat0.zw);
u_xlat16_0 = texture(_SourceTex, u_xlat0.xy);
u_xlat1 = u_xlat16_1 * vec4(0.12025857, 0.12025857, 0.12025857, 0.12025857);
u_xlat0 = u_xlat16_0 * vec4(0.0892157406, 0.0892157406, 0.0892157406, 0.0892157406) + u_xlat1;
u_xlat1 = _SourceTex_TexelSize.xyxy * vec4(1.1824255, -1.1824255, -1.1824255, 0.0) + vs_TEXCOORD0.xyxy;
u_xlat16_2 = texture(_SourceTex, u_xlat1.xy);
u_xlat16_1 = texture(_SourceTex, u_xlat1.zw);
u_xlat0 = u_xlat16_2 * vec4(0.0892157406, 0.0892157406, 0.0892157406, 0.0892157406) + u_xlat0;
u_xlat0 = u_xlat16_1 * vec4(0.12025857, 0.12025857, 0.12025857, 0.12025857) + u_xlat0;
u_xlat16_1 = texture(_SourceTex, vs_TEXCOORD0.xy);
u_xlat0 = u_xlat16_1 * vec4(0.162102833, 0.162102833, 0.162102833, 0.162102833) + u_xlat0;
u_xlat1 = _SourceTex_TexelSize.xyxy * vec4(1.1824255, 0.0, -1.1824255, 1.1824255) + vs_TEXCOORD0.xyxy;
u_xlat16_2 = texture(_SourceTex, u_xlat1.xy);
u_xlat16_1 = texture(_SourceTex, u_xlat1.zw);
u_xlat0 = u_xlat16_2 * vec4(0.12025857, 0.12025857, 0.12025857, 0.12025857) + u_xlat0;
u_xlat0 = u_xlat16_1 * vec4(0.0892157406, 0.0892157406, 0.0892157406, 0.0892157406) + u_xlat0;
u_xlat1 = _SourceTex_TexelSize.xyxy * vec4(0.0, 1.1824255, 1.1824255, 1.1824255) + vs_TEXCOORD0.xyxy;
u_xlat16_2 = texture(_SourceTex, u_xlat1.xy);
u_xlat16_1 = texture(_SourceTex, u_xlat1.zw);
u_xlat0 = u_xlat16_2 * vec4(0.12025857, 0.12025857, 0.12025857, 0.12025857) + u_xlat0;
u_xlat0 = u_xlat16_1 * vec4(0.0892157406, 0.0892157406, 0.0892157406, 0.0892157406) + u_xlat0;
SV_Target0 = u_xlat0;