Add more copyExternalImageToTexture tests as Chrome/Safari/Firefox are failing to agree how to load an image #4545

Activity

1. greggman commented on Dec 23, 2025

   

   greggman

   on Dec 23, 2025

   MemberAuthor

   More actions

   Trying with Image instead of createImageBitmap

   https://jsgist.org/?src=2359aec968f72ec4b24f69c712753d59

   • Chrome: Top-left RGBA: 130 101 73 1
   • Safari: Top-left RGBA: 1 0 0 1
   • Firefox: Top-left RGBA: 130 101 73 1
2. kainino0x commented on Jan 2, 2026

   

   kainino0x

   on Jan 2, 2026

   Last edited by kainino0x

   Collaborator

   More actions

   PNG is unpremultiplied, so if you copyExternalImageToTexture without premultipliedAlpha: true and copy that directly into the canvas you'll end up with undefined results on the display. That is not allowed to affect the readback result, but it does affect how it appears on screen.

   Here are my notes from playing with this sample on Mac:

   • ImageMagick seems to confirm 130 101 73 1 (magick FurBaseColorAlpha.png -format '%[fx:255*u.p{0,0}.r] %[fx:255*u.p{0,0}.g] %[fx:255*u.p{0,0}.b] %[fx:255*u.p{0,0}.a]' info:) - noting it's not fully transparent so there shouldn't be any issues with transparent pixels having arbitrary color values.
   • Firefox is initially missing their texture-to-canvas-conversion step - probably specific to filling a canvas using copyTextureToTexture - causing the BGRA-RGBA conversion step and alphaMode: 'opaque' not to apply. It fixes after switching away from the tab for a few seconds and back.
   • EDIT: continued in next post
3. kainino0x commented on Jan 2, 2026

   

   kainino0x

   on Jan 2, 2026

   Collaborator

   More actions

   Oops, posted before I was done.

   • In all browsers I think the createImageBitmap colorSpaceConversion option is behaving correctly. It should have no effect because the image is already sRGB.
   • In Safari and Firefox, I think the createImageBitmap premultiplyAlpha option is behaving incorrectly. It always behaves like premultiplyAlpha: 'premultiply' which results in the R/G/B values getting "quantized" into the range of A (so either 0 or 1). Safari quantizes to 1,0,0, and Firefox quantizes to 1,1,1. Then they unpremultiply to 255,0,0 and 255,255,255, respectively.
   • In Safari something is going badly wrong with the Image path, I'm not sure what, but the color is very wrong (regardless of -srgb). At pixel 100,100, I'm getting RGBA 128,99,99,252 instead of the correct 130,100,71,252. Also, it doesn't do the red-blue flip when doing copyExternalImageToTexture into BGRA. No such issues in the ImageBitmap path.
4. kainino0x commented on Jan 2, 2026

   

   kainino0x

   on Jan 2, 2026

   Collaborator

   More actions

   We do have tests that are supposed to test some of these things, but I don't think we try very hard to test createImageBitmap itself. In particular, we don't have Blob-to-ImageBitmap or Image-to-ImageBitmap cases.
   https://github.com/gpuweb/cts/tree/main/src/webgpu/web_platform/copyToTexture

   It would be useful to see if createImageBitmap is broken even for WebGL.
5. kainino0x commented on Jan 2, 2026

   

   kainino0x

   on Jan 2, 2026

   Collaborator

   More actions

   https://gpuweb.github.io/cts/standalone/?q=webgpu:web_platform,copyToTexture,image:* does not pass in Safari at all - every test case fails. It passes in Chrome and Firefox (except for new formats Firefox doesn't implement yet).
6. greggman commented on Jan 5, 2026

   

   greggman

   on Jan 5, 2026

   MemberAuthor

   More actions

   I think the thing not being tested that's failing on firefox, there is no webgpu:web_platform,copyToTexture,ImageBitmap:from_ImageData:from_Image there is only from_ImageData, from_canvas