In addition to the existing alpha blending rules (24 and 34), I propose 3 new additional rules.
Check https://en.wikipedia.org/wiki/Alpha_compositing
Existing rule 34 is a correct implementation of Alpha Blending for scaled forms, i.e. if premultiplied alpha is used. But rule 24 is a correct implementation of non-scaled forms, only for the case where destination is opaque (i.e. it's alpha is 1.0 in every pixel). If destination includes translucency (or it is completely transparent), the result is not correct. In order to fix it, the RGB of each pixel need to be divided by the pixel alpha.
The proposed blendUnscaled is the basic (for non-scaled forms) alpha blending described in Wikipedia. Note that users knowing that destination background is opaque might call the faster rule 24 instead.
The other two additional proposed rules are for converting to and from premultiplied alpha.
EXISTING blend 24 alphaBlend resultAlpha = srcAlpha + destAlpha*(1-srcAlpha) resultRGB = srcAlpha*source + (1-srcAlpha)*dest
EXISTING blendAlphaScaled 34 alphaBlendScaled resultRGBA = source + (1-srcAlpha)*dest
NEW PROPOSED multiplyRGBByAlpha Non premultiplied alpha -> premultiplied alpha. Only uses destination. Alpha unmodified. For each RGB component, resultRGB = dest*destAlpha
NEW PROPOSED divideRGBByAlpha Premultiplied alpha -> non premultiplied alpha. Only uses destination. Alpha unmodified. For each RGB component, resultRGB = dest/destAlpha
NEW PROPOSED blendUnscaled Equivalent to blend, and then divideRGBByAlpha resultAlpha = srcAlpha + destAlpha*(1-srcAlpha) resultRGB = (srcAlpha*source + (1-srcAlpha)*dest) / resultAlpha
This would allow handling of scaled (premultiplied-alpha) forms, and blending of regular forms including translucency, without the need to call slower smalltalk code to fix the results.
I believe this proposal is motivated by the issue discussed here:
https://lists.cuis.st/mailman/archives/cuis-dev/2020-June/001817.html
Is that right?
Dave
On Tue, Jun 02, 2020 at 12:06:42PM -0700, Juan Vuletich wrote:
In addition to the existing alpha blending rules (24 and 34), I propose 3 new additional rules.
Check https://en.wikipedia.org/wiki/Alpha_compositing
Existing rule 34 is a correct implementation of Alpha Blending for scaled forms, i.e. if premultiplied alpha is used. But rule 24 is a correct implementation of non-scaled forms, only for the case where destination is opaque (i.e. it's alpha is 1.0 in every pixel). If destination includes translucency (or it is completely transparent), the result is not correct. In order to fix it, the RGB of each pixel need to be divided by the pixel alpha.
The proposed blendUnscaled is the basic (for non-scaled forms) alpha blending described in Wikipedia. Note that users knowing that destination background is opaque might call the faster rule 24 instead.
The other two additional proposed rules are for converting to and from premultiplied alpha.
EXISTING blend 24 alphaBlend resultAlpha = srcAlpha + destAlpha*(1-srcAlpha) resultRGB = srcAlpha*source + (1-srcAlpha)*dest
EXISTING blendAlphaScaled 34 alphaBlendScaled resultRGBA = source + (1-srcAlpha)*dest
NEW PROPOSED multiplyRGBByAlpha Non premultiplied alpha -> premultiplied alpha. Only uses destination. Alpha unmodified. For each RGB component, resultRGB = dest*destAlpha
NEW PROPOSED divideRGBByAlpha Premultiplied alpha -> non premultiplied alpha. Only uses destination. Alpha unmodified. For each RGB component, resultRGB = dest/destAlpha
NEW PROPOSED blendUnscaled Equivalent to blend, and then divideRGBByAlpha resultAlpha = srcAlpha + destAlpha*(1-srcAlpha) resultRGB = (srcAlpha*source + (1-srcAlpha)*dest) / resultAlpha
This would allow handling of scaled (premultiplied-alpha) forms, and blending of regular forms including translucency, without the need to call slower smalltalk code to fix the results.
-- You are receiving this because you are subscribed to this thread. Reply to this email directly or view it on GitHub: https://github.com/OpenSmalltalk/opensmalltalk-vm/issues/505
Yep. That's right.
Thanks Juan to report it.
Would be nice to also have the Smalltalk fall-back code for those new rules. :-)
WIP http://source.squeak.org/VMMaker/VMMaker.oscog-nice.2909.diff
On 2020-12-23, at 12:56 PM, Nicolas Cellier notifications@github.com wrote:
WIP http://source.squeak.org/VMMaker/VMMaker.oscog-nice.2909.diff
Without having actually looked into this I just want ot wave a flag to remind us all to be sure it integrates ok with the ARM special purpose bitlbt routines.
Also, we *might* just possibly be getting the gift of some Ben Avison time to make the 64bit ARM equivalents.
tim -- tim Rowledge; tim@rowledge.org; http://www.rowledge.org/tim Strange OpCodes: MBR: Move Bits Randomly
I won't provide any accelerated assembler for the new ops, not my cup of tea. But as I understand it, the PI will fall back to the code generated for the BitBlt plugin when no accelerated path is provided, so it should continue to work. What would be nice now would be to start writing image side code & unit tests in either Cuis or Squeak.
http://source.squeak.org/inbox/GraphicsTests-nice.57.diff http://source.squeak.org/inbox/Graphics-nice.442.diff
I have revised the implementation which was previously not correct. See https://source.squeak.org/VMMaker/VMMaker.oscog-nice.3170.diff and also the image-side tests See https://source.squeak.org/inbox/GraphicsTests-nice.58.diff the image side tests pass on locally generated VM
integrated in https://github.com/OpenSmalltalk/opensmalltalk-vm/commit/f77943115bbb8e55fc1...
Closed #505.
vm-dev@lists.squeakfoundation.org
-
David T. Lewis -
hilaire -
Juan Vuletich
-
Marcel Taeumel
-
Nicolas Cellier
-
Stéphane Rollandin -
tim Rowledge