Shadow buffers #68470

New Issue

Clément Foucault · 2019-08-09T16:39:01+02:00

Clément Foucault commented

2019-08-09 16:39:01 +02:00

In order to reduce Shader register pressure during shading phase, we need to transfer the shadow sampling of opaque surfaces to a fullscreen pass.
We would store 1 or 2 bits per lights (limiting number of shadowed lights per pixels to 32 or 16) and use noise to dither the shadow shape.
The noise would disappear with accumulation.
This idea was explored in this presentation link.
The issue is that this technique will not work for alpha-blended surfaces that will still need the standard shadow code.
Contact shadow and cascade shadows cost could be reduced. This could be the opportunity to remove support for contact shadows for alpha blend surfaces.
This also will permit to have raytraced shadows in the future.

Estimate : 2 weeks

In order to reduce Shader register pressure during shading phase, we need to transfer the shadow sampling of opaque surfaces to a fullscreen pass. We would store 1 or 2 bits per lights (limiting number of shadowed lights per pixels to 32 or 16) and use noise to dither the shadow shape. The noise would disappear with accumulation. This idea was explored in this presentation [link](https://drive.google.com/file/d/19EHLJAfqQECjRRShHyS3mHrg4XFxpCOx/view). The issue is that this technique will not work for alpha-blended surfaces that will still need the standard shadow code. Contact shadow and cascade shadows cost could be reduced. This could be the opportunity to remove support for contact shadows for alpha blend surfaces. This also will permit to have raytraced shadows in the future. Estimate : 2 weeks

Clément Foucault commented

2019-08-09 16:39:01 +02:00

Added subscriber: @fclem

the Elder commented

2019-11-12 05:27:53 +01:00

Added subscriber: @theElder

the Elder commented

2019-11-12 05:27:59 +01:00

Removed subscriber: @theElder

Kenzie commented

2020-08-25 11:53:49 +02:00

Added subscriber: @KenzieMac130

Mindinsomnia commented

2020-10-27 02:12:37 +01:00

Added subscriber: @Grady

Mindinsomnia commented

2020-10-27 02:12:37 +01:00

In the linked presentation I particularly like page 16 which discusses the benefits of having two shadow maps, one for dynamic objects and one for static objects. In most scenes, most objects are static, most of the time.

Imagine a scene being rendered in Eevee that consists of a detailed environment of a city street, and only two characters animated inside of it standing next to a car, having a conversation from a few different camera angles. For each frame, Eevee would be re-rendering shadow maps that consist of almost identical data every frame for no benefit. If Eevee could retain shadow maps of static objects and dynamic objects separately, even if it requires a bit more memory, the performance boost would be substantial I imagine. An optimisation like that could go a long way to speeding up render times.

Unless Eevee is already doing this and no one has mentioned it to me yet?

But anywho, if that paper is roughly the idea that's being proposed here, perhaps that benefit of that system could also be adopted at the same time.

In the linked presentation I particularly like page 16 which discusses the benefits of having two shadow maps, one for dynamic objects and one for static objects. In most scenes, most objects are static, most of the time. Imagine a scene being rendered in Eevee that consists of a detailed environment of a city street, and only two characters animated inside of it standing next to a car, having a conversation from a few different camera angles. For each frame, Eevee would be re-rendering shadow maps that consist of almost identical data every frame for no benefit. If Eevee could retain shadow maps of static objects and dynamic objects separately, even if it requires a bit more memory, the performance boost would be substantial I imagine. An optimisation like that could go a long way to speeding up render times. Unless Eevee is already doing this and no one has mentioned it to me yet? But anywho, if that paper is roughly the idea that's being proposed here, perhaps that benefit of that system could also be adopted at the same time.

Jon Henry M. Faltis 9077149 commented

2020-12-22 03:55:39 +01:00

Added subscriber: @geocentric_wage

Mikhail Matrosov commented

2020-12-22 09:42:56 +01:00

Added subscriber: @ktdfly

Emi Martinez commented

2021-04-06 02:43:36 +02:00

Added subscriber: @Emi_Martinez

Philipp Oeser removed the

Download

What's New

Blender Studio

Manual

Developers Blog

Documentation

Benchmark

Blender Conference

Development Fund

One-time Donations

Shadow buffers #68470