Adaptive Deferred Shading

An adaptive deferred shading implementation based on the paper Deferred Adaptive Compute Shading

Algorithm Overview

⬜: shading pixel 🟩: shaded pixel ⬛: unshaded pixel

Workflow

Pass 0

⬜⬛⬛⬛
⬛⬛⬛⬛
⬛⬛⬛⬛
⬛⬛⬛⬛
Dispatch Dimension: (Screen Width / 4) * (Screen Height / 4) * 1

Access Pattern:

1. Compute Block Idx
   
2. Apply Pixel Offset
   
3. Shade Target Pixel

Pass 1

🟩⬛⬛⬛
⬛⬛⬛⬛
⬛⬛⬜⬛
⬛⬛⬛⬛
Dispatch Dimension: (Screen Width / 4) * (Screen Height / 4) * 1

Access Pattern:

1. Compute Block Idx
   
2. Apply Pixel Offset
   
3. Derive Statistics Property
   
4. Interpolate Or Shade The Target Pixel Based On Stats Info.

Interpolate between neighbors if (i + 2, j + 2), (i - 2, j + 2), (i + 2, j - 2), (i - 2, j - 2) are similar.

Pass 2

🟩⬛⬜⬛
⬛⬛⬛⬛
⬜⬛🟩⬛
⬛⬛⬛⬛
Dispatch Dimension: (Screen Width / 4) * (Screen Height / 4) * 2

Interpolate between neighbors if (i + 2, j), (i - 2, j), (i, j - 2), (i, j + 2) are similar.

Pass 3

🟩⬛🟩⬛
⬛⬜⬛⬜
🟩⬛🟩⬛
⬛⬜⬛⬜
Dispatch Dimension: (Screen Width / 4) * (Screen Height / 4) * 4

Interpolate between neighbors if (i + 1, j + 1), (i - 1, j - 1), (i + 1, j - 1), (i - 1, j + 1) are similar.

Pass 4

🟩⬜🟩⬜
⬜🟩⬜🟩
🟩⬜🟩⬜
⬜🟩⬜🟩
Dispatch Dimension: (Screen Width / 4) * (Screen Height / 4) * 8

Interpolate between neighbors if (i + 1, j), (i - 1, j), (i, j - 1), (i, j + 1) are similar.

Pixel Similarity Definition

Given 4 pixels, they are similar if the variance is lower than the threshold.

Wave Sorting

⬜: shading 🟩: interpolation

Assume Warp Size = 16

⬜⬜⬜⬜
⬜🟩⬜🟩
🟩⬜|⬜|⬜
⬜⬜⬜⬜

Total Shading Count WaveActiveCountBits(shading) = 13,

Given thread 10, srcLaneIdx = 10, WavePrefixCountBits(shading) = 7 (not including itself),

interpolation before thread 10 is srcLaneIdx - WavePrefixCountBits(shading) = 3

dstLaneIdx = srcLaneIdx is shading? WavePrefixCountBits(shading) = 7: (WaveActiveCountBits(shading) + (srcLaneIdx - WavePrefixCountBits(shading)) = 16) = 7

Finally, srcLaneIdx = 10, dstLaneIdx = 7

srcLaneIdx = 0, WavePrefixCountBits(shading) = 0, dstLaneIdx = 0
srcLaneIdx = 1, WavePrefixCountBits(shading) = 1, dstLaneIdx = 1
srcLaneIdx = 2, WavePrefixCountBits(shading) = 2, dstLaneIdx = 2
srcLaneIdx = 3, WavePrefixCountBits(shading) = 3, dstLaneIdx = 3

srcLaneIdx = 4, WavePrefixCountBits(shading) = 4, dstLaneIdx = 4
srcLaneIdx = 5, WavePrefixCountBits(shading) = 5, dstLaneIdx = 13
srcLaneIdx = 6, WavePrefixCountBits(shading) = 5, dstLaneIdx = 5
srcLaneIdx = 7, WavePrefixCountBits(shading) = 6, dstLaneIdx = 14

srcLaneIdx = 8, WavePrefixCountBits(shading) = 6, dstLaneIdx = 15
srcLaneIdx = 9, WavePrefixCountBits(shading) = 6, dstLaneIdx = 6
srcLaneIdx = 10, WavePrefixCountBits(shading) = 7, dstLaneIdx = 7
srcLaneIdx = 11, WavePrefixCountBits(shading) = 8, dstLaneIdx = 8

srcLaneIdx = 12, WavePrefixCountBits(shading) = 9, dstLaneIdx = 9
srcLaneIdx = 13, WavePrefixCountBits(shading) = 10, dstLaneIdx = 10
srcLaneIdx = 14, WavePrefixCountBits(shading) = 11, dstLaneIdx = 11
srcLaneIdx = 15, WavePrefixCountBits(shading) = 12, dstLaneIdx = 12

Multi-Wave Sorting

⬜: shading 🟩: interpolation

Assume Warp Size = 16

Wave 0

⬜⬜⬜⬜
⬜🟩⬜🟩
🟩⬜⬜⬜
⬜⬜⬜⬜

Wave 1

⬜🟩🟩🟩
⬜🟩⬜🟩
🟩⬜🟩⬜
🟩⬜🟩🟩