WIP: do the 1/2 pixel translation only in RealizeOnDomainOperation #124841

Bill-Spitzak · 2024-07-17T06:34:59+02:00

Bill-Spitzak commented

2024-07-17 06:34:59 +02:00

If the input and output domains don't match even/odd parity of sizes, perform a 1/2 pixel shift so that there is no unexpected filtering.

This patch includes my fix for #124133

If the input and output domains don't match even/odd parity of sizes, perform a 1/2 pixel shift so that there is no unexpected filtering. This patch includes my fix for #124133

Bill-Spitzak added 5 commits 2024-07-17 06:35:16 +02:00

Fix #124133 : Transforms shift due to incorrect floor() 2da2dd133a

Previously there would always be a 1/2 pixel shift if the input and output
domains did not match even/odd sizes, this shift was always in the same
direction no matter which one was odd. All this made transforms somewhat
inaccurate and made viewport scaling look incorrect.

Compute full transform including corner offsets outside glsl 8ec8430a92

This changes the glsl realize_on_domain code to take a matrix that goes
directly between spaces where 0,0 is in the lower-left corner of the output
and input domains. Simplifies the math and allows more analysis to be done
of the transform.

Use the "nearest" filter if possible ee0d542e15

Minor optimization. As the current filters are interpolating, if the samples
are all in the middle of the pixel and the derivative between samples is 1
then the Nearest filter can be used.

Fix sampling calls in CPU compositor 6c2b76e768

These are adjusted to match the locations used by the GPU compositor,
which I believe is correct. Mostly this means adding .5 to the output
point before the transform, and subtracting .5 to get the input point.

I also fixed the Scale Absolute bounding box to produce the requested
size. Rounding errors were causing it to produce larger sizes.

COM_Translate is unchanged as the position is correct. However it is
putting transparent rows of pixel on the bottom and left edges of the
output rather than the correct value, I was unable to figure out the
cause of this.

Put the .5 shift into RealizeOnDomainOperation b3b2014f14

This is the transform that is automatically triggered when Domains
don't match. In this case the user probably is not expecting any
filtering due to pixel centers not aligning.

NYI for the CPU compositor.

Bill-Spitzak requested review from Omar Emara 2024-07-17 06:35:52 +02:00

Bill-Spitzak referenced this pull request

2024-07-17 06:36:56 +02:00

Fix #124133: Transforms shift due to incorrect floor() #124381

Bill-Spitzak referenced this pull request

2024-07-19 00:32:58 +02:00

WIP: halfpixel shift for odd domains #124760

Bill-Spitzak commented

2024-07-19 06:03:58 +02:00

I have been using it like this for a while and it seems to be working well.

Bill-Spitzak commented

2024-07-19 22:48:58 +02:00

I have merged this change into the main patch #124381 as that results in the minimal change of behavior.

I have merged this change into the main patch https://projects.blender.org/blender/blender/pulls/124381 as that results in the minimal change of behavior.

Bill-Spitzak closed this pull request

2024-07-19 22:49:00 +02:00

Pull request closed

This pull request cannot be reopened because the branch was deleted.

Sign in to join this conversation.

No reviewers

No Label

Download

What's New

Blender Studio

Manual

Developers Blog

Documentation

Benchmark

Blender Conference

Development Fund

One-time Donations

WIP: do the 1/2 pixel translation only in RealizeOnDomainOperation #124841

Pull request closed