How to Sharing depth between passes?

[jellychen]

Description

How to Sharing depth between passes?

Solution

I am developing with WebGPU and have two scenes. For specific reasons, I need to split them into two separate passes and then merge them. However, I want the second pass to share the depth from the first pass. Is there a convenient way to handle this using TSL?

Alternatives

none

Additional context

No response

[Mugen87]

Crossposting: https://discourse.threejs.org/t/how-to-sharing-depth-between-passes/90036

[sunag]

Perhaps we can move forward with something like this, it will need to modify the renderer, however:

const basePass = pass( scene, camera );

const secondPass = pass( scene, camera );
secondPass.autoClearDepth = false;

secondPass.renderTarget.depthTexture = basePass.renderTarget.depthTexture;

Initially we could simplify by modifying only PassNode and using a internal copyTextureToTexture() before rendering with something like this

const basePass = pass( scene, camera );

const secondPass = pass( scene, camera, { depthTexture: basePass.depthTexture } );
secondPass.autoClearDepth = false;

Are we aware of the problem this will have with transparent objects?

Maybe this type of filter could also be achieved with MRT similar to what was done in #32570 (comment) ?

[Mugen87]

PassNode has currently no clear flags because it is designed to maintain its own internal render target. If we want to allow to share buffer data, we probably should allow to share the entire render target. So similar to RenderPass, two instances of PassNode can render into the same destination. It should also be possible to share just the depth renderbuffer or texture in our case. I like the idea of passing the render target or depth texture via the options parameter.

So I guess we need four clear flags: autoClear, autoClearColor, autoClearDepth and autoClearStencil. These flags should be set to the respective renderer properties in updateBefore().

And we need renderTarget and depthTexture parameters in the options object that are used to configure the internal render target.

Are we aware of the problem this will have with transparent objects?

Transparent objects won't be properly blended if they are divided into multiple render passes. So I think apps must be careful how they organize their scenes.

I understand if apps render the scene with one pass and e.g. helpers or UI with a second one, they need a custom blend right now. Having both render into the same destination is indeed better for performance.