Fast GPU Histogram Analysis for Scene Post-Processing - PowerPoint Presentation

Slide1
Slide2

Fast GPU Histogram Analysis for Scene Post-Processing

Andy Luedke

Halo Development Team

Microsoft Game StudiosSlide3

Why do Histogram Analysis?

Dynamically adjust post-processing settings based on rendered scene content

Drive tone adjustments by discovering intensity levels and adjusting

tonemapper

settings

Make environments feel consistent with a wide range of illumination

Mimic eye’s natural adaptation to exposure and focal rangesSlide4

Existing Techniques

Average Scene Luminance

Varies significantly with small perceived changes in HDR scenes

Luminance Histogram

Provides more useful exposure data

Limited by fixed number of bins

CPU generated from locked texture

Adjustable granularity, poor performance

GPU queries to update histogram bins

Low granularity, delayed scene responseSlide5

Luminance Histogram

Used to find interesting exposure control points

Median luminance (50

th

percentile)

Bright point (90

th

– 95

th

percentile)

Search histogram for each point

Only contains luminance data from previously rendered frames

Expensive to generate and search

Histograms are not great for exposure controlSlide6

Sorted Luminance Buffer

Sorting the luminance fixes many problems with histogram method

Expensive to sort on the CPU

Sort on the GPU instead

Parallel sorts are quite fast on GPUs

Works on current frame’s data

Easy to find percentiles in a sorted luminance buffer

Sample center of buffer for median value, or at X*N/100 for

X

th

percentileSlide7

GPU Sorting

Avoids histogram range clamping and bin granularity problems

Works on current frame’s values

Sorts multiple channels at once

Sort luminance and depth in a two channel buffer, or more in 4 channels

Sorted buffer remains on GPU

CPU processing of exposure control can be moved to the GPU exclusivelySlide8

GPU Sorting (continued)

Bitonic

sort works well on the GPU

Well suited for shader implementation

Exactly ½*(log

2

n * (log

2

n+1)) passes

Scale to slower hardware by reducing size of sorting buffer

Exposure control point lookups are still direct, but have less resolution

Bitonic

sort works best on power of 2 textures, but can be tweaked to work on other sizesSlide9

Bitonic Sort Demo

Red = Average luminance

Green = Maximum luminanceSlide10

GPU Exposure Processing

Shader samples sorted luminance buffer and outputs updated exposure control values

Use GPU to sample many points and do complex adjustments (curves, etc)

Blend new exposure control values with previous values over time

Another shader generates further

tonemapping

settings

Bloom settings, saturation, tone, etc.Slide11

Local Exposure Control

Use one channel of sort buffer as a key for another channel’s sort

Sort regions of the screen in addition to the full frame’s values

Still direct access as long as each region has a known number of pixels

RGBA=[

Lum

, Depth, Local

Lum

, Key]

Allows you to divide the screen into multiple exposure zones and mix local and global adjustmentsSlide12

Keyed Luminance Sort

Red/Green =

Avg

/Max Luminance

Blue = Regional

Avg

LuminanceSlide13

Local Exposure Control

Use different region masks to customize to your game’s needs

Must know how many pixels in each region for direct value accessSlide14

Focal Range Control

Sorted depth gives useful information for DOF ranges

Detect changes in depth range, adjusting DOF settings to simulate eye’s adjustment of focal range

Extracting depth has sampling cost, but no additional sorting

cost

May not need to filter on

downsampleSlide15

CPU Exposure PipelineSlide16

GPU Exposure PipelineSlide17

Questions?

aluedke@microsoft.com

Please fill out your surveys

References

GPU Gems

, Chapter 37

GPU Gems 2, Chapter 46

UberFlow

: A

GPU-based particle engine [

Kipfer

, et al.]

Wikipedia, Sorting

Algoritms