Why DownsampleBox13Tap in post-processing stack do so much sample?

Unity Engine
,
1

This function is a part of fragment shader in Sampling.hlsl used in bloom.Each edge of destination texture is 1/2 of the source texture.

// . A . B . C .
// . . D . E . .
// . F . G . H .
// . . I . J . .
// . K . L . M .

half4 DownsampleBox13Tap(TEXTURE2D_ARGS(tex, samplerTex), float2 uv, float2 texelSize)
{
half4 A = SAMPLE_TEXTURE2D(tex, samplerTex, UnityStereoTransformScreenSpaceTex(uv + texelSize * float2(-1.0, -1.0)));
half4 B = SAMPLE_TEXTURE2D(tex, samplerTex, UnityStereoTransformScreenSpaceTex(uv + texelSize * float2( 0.0, -1.0)));
half4 C = SAMPLE_TEXTURE2D(tex, samplerTex, UnityStereoTransformScreenSpaceTex(uv + texelSize * float2( 1.0, -1.0)));
half4 D = SAMPLE_TEXTURE2D(tex, samplerTex, UnityStereoTransformScreenSpaceTex(uv + texelSize * float2(-0.5, -0.5)));
half4 E = SAMPLE_TEXTURE2D(tex, samplerTex, UnityStereoTransformScreenSpaceTex(uv + texelSize * float2( 0.5, -0.5)));
half4 F = SAMPLE_TEXTURE2D(tex, samplerTex, UnityStereoTransformScreenSpaceTex(uv + texelSize * float2(-1.0, 0.0)));
half4 G = SAMPLE_TEXTURE2D(tex, samplerTex, UnityStereoTransformScreenSpaceTex(uv ));
half4 H = SAMPLE_TEXTURE2D(tex, samplerTex, UnityStereoTransformScreenSpaceTex(uv + texelSize * float2( 1.0, 0.0)));
half4 I = SAMPLE_TEXTURE2D(tex, samplerTex, UnityStereoTransformScreenSpaceTex(uv + texelSize * float2(-0.5, 0.5)));
half4 J = SAMPLE_TEXTURE2D(tex, samplerTex, UnityStereoTransformScreenSpaceTex(uv + texelSize * float2( 0.5, 0.5)));
half4 K = SAMPLE_TEXTURE2D(tex, samplerTex, UnityStereoTransformScreenSpaceTex(uv + texelSize * float2(-1.0, 1.0)));
half4 L = SAMPLE_TEXTURE2D(tex, samplerTex, UnityStereoTransformScreenSpaceTex(uv + texelSize * float2( 0.0, 1.0)));
half4 M = SAMPLE_TEXTURE2D(tex, samplerTex, UnityStereoTransformScreenSpaceTex(uv + texelSize * float2( 1.0, 1.0)));

half2 div = (1.0 / 4.0) * half2(0.5, 0.125);

half4 o = (D + E + I + J) * div.x;
o += (A + B + G + F) * div.y;
o += (B + C + H + G) * div.y;
o += (F + G + L + K) * div.y;
o += (G + H + M + L) * div.y;

return o;
}

I think “G” is equal to “(D + E + I + J)*0.25”.
So it just samples G is enough,why does it sample D,E,I,J?It looks like not necessary.There must be some reason to add so mush extra sampling.

2

This shader is based on a talk Jorge Jimenez – Next Generation Post Processing in Call of Duty: Advanced Warfare
If you look at the slides it becomes clear that all 13 samples are supposed to be bilinear fetches across a 6 by 6 texel grid, but I think the shader code done by Unity Tech. might be implementing it incorrectly. Below is the screenshot of the slide in question:

4571578--425350--upload_2019-5-23_14-26-34.png
4571578--425350--upload_2019-5-23_14-26-34.png500×568 23.8 KB

In the original slides G(center-most sample) would not be equal (D + E + I + J) * 0.25 (red samples) because all the components are already bilinear interpolations, not texels.

As the DownsampleBox13Tap function is now, it looks like samples D, E, I and J are interpolated(since they use half-texel offsets) and the rest seem to be uninterpolated. Have you investigated this further? Any interesting discoveries?

Disclaimer: I might be completely wrong.

3

That kernel assumes the starting UV is at the corner between texels. When downsampling, the texture UV of the lower res texture will already be centered between the 4 texels of the previous texture.