EEVEE-Next: Add tile deferred lighting #115820

Merged

Clément Foucault merged 14 commits from fclem/blender:eevee-next-deferred-lighting-tile into main

2023-12-07 17:05:06 +01:00

Clément Foucault commented

2023-12-05 21:45:46 +01:00

Member

This adds a tile classification pass to the gbuffer.
This is then compact into streams of tiles for each
complexity level of lighting evaluation.

The benefit over a simpler approach of using a per
object stencil value is that we can have a per
tile granularity of the lighting complexity.

To avoid quad overshading, we use a prepass that
tags different stencil values for each complexity
level. This allows to still use a fullscreen quad
for the light evaluation pass and remove the
diagonal overshading cost.

This doesn't use compute shader at all to leverage
render pass merging and in-tile memory loads.

Using atomicOr for adding together the eClosureBits
revealed to be too slow. Using multiple non-atomic
writes to many data values is faster and not much
memory hungry.

Performance

The whole tile scheduling process takes ~70µs for
a half covered 3800x790 framebuffer and doesn't
get much more slower than this.

Using simpler lighting shader helps reduce the cost
of the lighting pass by half in most common cases.
SSS materials stay the most costly.

TODO

Make lighting shader variations.
Merge tile mask generation to the gbuffer pass and use subpass input.
Test Metal and remove workaround.

This adds a tile classification pass to the gbuffer. This is then compact into streams of tiles for each complexity level of lighting evaluation. The benefit over a simpler approach of using a per object stencil value is that we can have a per tile granularity of the lighting complexity. To avoid quad overshading, we use a prepass that tags different stencil values for each complexity level. This allows to still use a fullscreen quad for the light evaluation pass and remove the diagonal overshading cost. This doesn't use compute shader at all to leverage render pass merging and in-tile memory loads. Using `atomicOr` for adding together the `eClosureBits` revealed to be too slow. Using multiple non-atomic writes to many data values is faster and not much memory hungry. ### Performance The whole tile scheduling process takes ~70µs for a half covered 3800x790 framebuffer and doesn't get much more slower than this. Using simpler lighting shader helps reduce the cost of the lighting pass by half in most common cases. SSS materials stay the most costly. ### TODO - [x] Make lighting shader variations. - [x] Merge tile mask generation to the gbuffer pass and use subpass input. - [x] Test Metal and remove workaround.

Clément Foucault added this to the 4.1 milestone 2023-12-05 21:45:46 +01:00

Clément Foucault added the

label 2023-12-05 21:45:46 +01:00

Clément Foucault added 5 commits 2023-12-05 21:45:59 +01:00

b3cd8a745d EEVEE-Next: Add tile deferred lighting

26c6a2d54f WIP

6906c07200 Use tile prepass to set stencil values.

3d844b3413 Add tile compaction stage

fbed71af0b Adjust tile size going over a certain number of tile

Clément Foucault added 1 commit 2023-12-05 22:00:21 +01:00

d12d7e378b Revert unwanted changes

Clément Foucault added 1 commit 2023-12-05 22:02:32 +01:00

4d2c5366fc Remove unused file

Clément Foucault added 2 commits 2023-12-06 17:54:53 +01:00

cc24346ab9 Add light evaluation shader variations

b0c16811aa Add ROG, Fix wrong stencil and missing bind

Clément Foucault force-pushed eevee-next-deferred-lighting-tile from 20f3a24ea2 to 811354d166

2023-12-06 21:48:25 +01:00

Compare

Clément Foucault added 1 commit 2023-12-06 21:48:49 +01:00

buildbot/vexp-code-patch-coordinator Build done. Details

d8cf40e83b Merge branch 'main' into eevee-next-deferred-lighting-tile

Clément Foucault changed title from ~~WIP: EEVEE-Next: Add tile deferred lighting~~ to EEVEE-Next: Add tile deferred lighting

2023-12-06 21:48:59 +01:00

Clément Foucault commented

2023-12-06 21:49:09 +01:00

Author

Member

@blender-bot build

@blender-bot build

Jacob Merrill commented

2023-12-06 23:35:57 +01:00

First-time contributor

"The whole tile scheduling process takes ~70µs for
a half covered 3800x790 framebuffer and doesn't
scale much more than this."

this is not realtime :(

Is there room here to fix its performance another way?

"The whole tile scheduling process takes ~70µs for a half covered 3800x790 framebuffer and doesn't scale much more than this." this is not realtime :( Is there room here to fix its performance another way?

Clément Foucault commented

2023-12-07 00:20:20 +01:00

Author

Member

"The whole tile scheduling process takes ~70µs for
a half covered 3800x790 framebuffer and doesn't
scale much more than this."

this is not realtime :(

Is there room here to fix its performance another way?

You clearly misunderstood what I wrote. I updated the PR description to avoid any misunderstanding.

~70µs is 0.4% of a 60Hz frame, and if it helps save 5% it's still a good. And that saving does scale up with resolution, whereas the preprocess doesn't cost much more.

> "The whole tile scheduling process takes ~70µs for > a half covered 3800x790 framebuffer and doesn't > scale much more than this." > > this is not realtime :( > > Is there room here to fix its performance another way? You clearly misunderstood what I wrote. I updated the PR description to avoid any misunderstanding. ~70µs is 0.4% of a 60Hz frame, and if it helps save 5% it's still a good. And that saving does scale up with resolution, whereas the preprocess doesn't cost much more.

❤️ 1 🎉 1

Clément Foucault added 1 commit 2023-12-07 00:31:40 +01:00

3a2b0a54fd Fix issues on Metal

Jacob Merrill commented

2023-12-07 00:31:54 +01:00

First-time contributor

"The whole tile scheduling process takes ~70µs for
a half covered 3800x790 framebuffer and doesn't
scale much more than this."

this is not realtime :(

Is there room here to fix its performance another way?

You clearly misunderstood what I wrote. I updated the PR description to avoid any misunderstanding.

~70µs is 0.4% of a 60Hz frame, and if it helps save 5% it's still a good. And that saving does scale up with resolution, whereas the preprocess doesn't cost much more.

~70µs is 0.4% - for some reason I did a quick google says its 70 milliseconds - not .07 ms - sorry for the confusion
A microsecond != Milliseconds *

> > "The whole tile scheduling process takes ~70µs for > > a half covered 3800x790 framebuffer and doesn't > > scale much more than this." > > > > this is not realtime :( > > > > Is there room here to fix its performance another way? > > You clearly misunderstood what I wrote. I updated the PR description to avoid any misunderstanding. > > ~70µs is 0.4% of a 60Hz frame, and if it helps save 5% it's still a good. And that saving does scale up with resolution, whereas the preprocess doesn't cost much more. ~70µs is 0.4% - for some reason I did a quick google says its 70 milliseconds - not .07 ms - sorry for the confusion A microsecond != Milliseconds *

Clément Foucault referenced this issue from a commit

2023-12-07 17:05:06 +01:00

EEVEE-Next: Add tile deferred lighting

Clément Foucault merged commit 2cf64b9100 into main

2023-12-07 17:05:06 +01:00

Clément Foucault deleted branch eevee-next-deferred-lighting-tile

2023-12-07 17:05:08 +01:00

Jacob Merrill commented

2023-12-12 02:41:46 +01:00

First-time contributor

how does one set the pipe to forward+ vs tiled deferred ?

how does one set the pipe to forward+ vs tiled deferred ?

Clément Foucault commented

2023-12-12 13:30:55 +01:00

Author

Member

@JacobMerrill-1 this is what the Dithered/Blended option is for.

But this isn't the place for asking this sort of thing. This is better asked on devtalk or blender.chat .

@JacobMerrill-1 this is what the Dithered/Blended option is for. But this isn't the place for asking this sort of thing. This is better asked on devtalk or blender.chat .

Brian Savery (AMD) referenced this issue from a commit

2023-12-28 20:07:43 +01:00

EEVEE-Next: Add tile deferred lighting

Sign in to join this conversation.

No reviewers

No Label

Animation & Rigging

Asset Browser Project Overview

Automated Testing

Blender Asset Bundle

Dependency Graph

Development Management

EEVEE & Viewport

Images & Movies

Motion Tracking

Nodes & Physics

Pipeline, Assets & IO

Platforms, Builds & Tests

Render & Cycles

Render Pipeline

Sculpt, Paint & Texture

Video Sequencer

Virtual Reality

Blender 2.8 Project

Milestone 1: Basic, Local Asset Browser

Good First Issue

Animation & Rigging

Development Management

EEVEE & Viewport

Nodes & Physics

Pipeline, Assets & IO

Platforms, Builds & Tests

Render & Cycles

Sculpt, Paint & Texture

Needs Info from Developers

Needs Information from User

No Milestone

No project

No Assignees

2 Participants

Notifications

Due Date

The due date is invalid or out of range. Please use the format 'yyyy-mm-dd'.

No due date set.

Dependencies

No dependencies set.

Reference: blender/blender#115820

No description provided.