Unify CPU and GPU compositors #116694

New Issue

Habib Gahbiche · 2024-01-01T19:29:20+01:00

Habib Gahbiche commented

2024-01-01 19:29:20 +01:00

Summary

This task summarizes all remaining tasks to replace the Tiled compositor with the Full Frame compositor. The requirement to do so is to unify the behavior of CPU and GPU compositors.

The must-haves are necessary to finish before enabling the Full Frame compositor as the default CPU compositor. The images in the comparison document Full Frame vs. GPU were produced by running the compositor unit tests while setting execution mode to Fullframe and GPU respectively. Only nodes with different results are shown in PDF document.

Must have

Behavior	Full Frame	GPU	Status
Size inference: constant input	Constant input results in output with	Constant input results in constant output `x`	Done
Size inference: size undo	Limit canvas size to 1.5x render size	Unlimited `x`	Done
Automatic realization	Apply transformations from left to right `x`	Realization time can be delayed	Done
plane track motion and plane track plane	Produced mask is binary `x`	Produced mask contains value from original alpha channel of input	Done
Node mask aspect ratio	Mask is scaled by image aspect ratio `x`	Mask ignores image aspect ratio	Done
Node Texture Color	`x`		Done
Node Sun Beans			Done
Double Edge Mask		`x`	Done
Inpaint	(difference is visually significant)	(difference is visually significant) `x`	Done
Erode feather	(difference is visually significant for negative values) `x`	(difference is visually significant for negative values)	Done
d(irectional) blur	alpha channel considered `x`	alpha channel is ignored	Done

Legend:

x indicates which behavior should be implemented for both compositors.
Done decision about behavior was made, implementation is done.
Undecided need to agree on behavior

Nice to have

Nodes that visually look the same but dividing the two images gives values slightly lower or higher than 1. See #118548 for full list

## Summary This task summarizes all remaining tasks to replace the Tiled compositor with the Full Frame compositor. The requirement to do so is to unify the behavior of CPU and GPU compositors. The must-haves are necessary to finish before enabling the Full Frame compositor as the default CPU compositor. The images in the comparison document `Full Frame vs. GPU` were produced by running the compositor unit tests while setting execution mode to Fullframe and GPU respectively. Only nodes with different results are shown in PDF document. ## Must have | Behavior | Full Frame | GPU | Status | | ----------- | ----------- | ----------- | ----------- | | Size inference: constant input | Constant input results in output with ` ` | Constant input results in constant output `x` | Done | | Size inference: size undo | Limit canvas size to 1.5x render size ` ` | Unlimited `x` | Done | | Automatic realization | Apply transformations from left to right `x` | Realization time can be delayed ` ` | Done | | plane track motion and plane track plane | Produced mask is binary `x` | Produced mask contains value from original alpha channel of input ` ` | Done | | Node mask aspect ratio | Mask is scaled by image aspect ratio ![image](/attachments/eab7b036-51c9-4098-9681-d3a3d857af3d) `x` | Mask ignores image aspect ratio ![image](/attachments/e83b84b0-7804-4249-b415-55028cd17868) ` ` | Done | | Node Texture Color | ![image](/attachments/90997a73-b3f5-4081-ad3a-841334e8ff33) `x` | ![image](/attachments/a572fc20-bde2-4621-b253-2e340f9f3748) ` ` | Done | | Node Sun Beans | ![image](/attachments/25f94061-7d6c-490a-81b4-cee812ad7152) ` ` | ![image](/attachments/c297c55b-0e12-472d-b3e0-1bdbb7da5ac5) ` ` | Done | | Double Edge Mask | ![image](/attachments/5fc4e4dd-ef59-4268-b491-95138b583c94) ` ` | ![image](/attachments/acf07df0-45ec-467e-a859-8370bf6e51ef) `x` | Done | | Inpaint | (difference is visually significant) ` ` | (difference is visually significant) `x` | Done | | Erode feather | (difference is visually significant for negative values) `x` | (difference is visually significant for negative values) ` ` | Done | | d(irectional) blur | alpha channel considered `x` | alpha channel is ignored ` ` | Done | Legend: - `x` indicates which behavior should be implemented for both compositors. - Done decision about behavior was made, implementation is done. - Undecided need to agree on behavior ## Nice to have Nodes that visually look the same but dividing the two images gives values slightly lower or higher than 1. See https://projects.blender.org/blender/blender/issues/118548 for full list

❤️ 4 🎉 1 🚀 1

Habib Gahbiche added the

Type

To Do

label 2024-01-01 19:29:20 +01:00

Habib Gahbiche added this to the Compositing project 2024-01-01 19:29:22 +01:00

Sergey Sharybin referenced this issue

2024-01-23 17:40:05 +01:00

Compositor: Short term planning #106533

Habib Gahbiche commented

2024-03-17 10:08:49 +01:00

Closing as done, follow-up issue is #118548.