Real Time Soft Shadows
on GeForce class hardware

Requirements

To display soft shadows, you'll need a 3D card with:

• A Stencil Buffer
• A 8 bit Alpha buffer

Introduction

The basic idea is to display many times the shadows while shaking the light, and to blur them together. As GeForce hardware is not fast enough to display so many volumes at the same time, we will use time coherency to fake it. During one frame, we will display the current volume, and blur it with the volumes of the previous frames. This idea is based on Motion Blur.

Implementation

High level view

We will store the volumes in the Alpha buffer. The different steps of the Soft Shadow Stencil Volumes algorithm are:

• Shake the light
• Compute the volume in the Stencil Buffer as usual (see Carmack's reversed method, or NVidia Robust method)
• Display the current volume in the Alpha Buffer
• Fade out the Alpha Buffer, to blur the current volumes and to make too old volume disappear
• Display shadows with Blending switched on in the RGB Buffer

Shaking the light

To shake the light, I add one of these eight movement to the light position:

As you can see, I select 4 corners of 2 cubes around the light, so that there are as few positions in the same plane as possible. As a result, each volume will not be totally included in the seven others.

Fading the Alpha

Fading is done by displaying a quad with the following blending mode and color:

glColorMask( GL_FALSE,GL_FALSE,GL_FALSE,GL_TRUE );
glBlendFunc(GL_ZERO,GL_SRC_ALPHA);
glColor4f(0,0,0,0.95f);

Dynamic lighting

When the light is moving, you can chose between either shaking of not the light. Both method works fine and gives similar results.

Dynamic camera

The problem with dynamic camera is that the previous volumes are not placed where they should be.
At high frame rates, the error in the positioning of the previous volumes is not significant.
At lower frame rates, the error gets too important, and we have to reduce the number of displayed volumes, by fading them out faster.
I usually display 8 volumes when the camera is static, and 2 volumes when the camera is moving.

More realistic on faster hardware

On faster hardware, you can display all the volumes at the same time during the same frame, and you will not need shadow motion blurring anymore.
The steps are now:
For volume = 0 To MaxVolume

• Shake the light
• Compute the shadow volume in the Stencil Buffer as usual
• Add the volume in the Alpha Buffer
• Display shadows with Blending switched on in the RGB Buffer

Results

           

The author