Resample vertical pass #1933
Resample vertical pass #1933
Conversation
|
Changes Unknown when pulling f09067e on uploadcare:resample-vertical-pass into * on python-pillow:master*. |
| @@ -40,6 +40,7 @@ def getmode(mode): | |||
| _modes[m] = ModeDescriptor(m, bands, basemode, basetype) | |||
| # extra experimental modes | |||
| _modes["LA"] = ModeDescriptor("LA", ("L", "A"), "L", "L") | |||
| _modes["La"] = ModeDescriptor("La", ("L", "a"), "L", "L") | |||
There was a problem hiding this comment.
This required for Image.merge('La', ...). But there is no 'RGBa', so I used 'RGBX', which is. Modes support in Pillow is a mess ((
There was a problem hiding this comment.
@wiredfool If you don't mind about adding La in _modes, I'd like to merge it. I don't see any other controversial points in this RP.
There was a problem hiding this comment.
I went through the review, hit this, started wondering why we didn't have RGBa as a mode here, and then determined that modes are a mess.
There was a problem hiding this comment.
I can add RGBa here as a temporary solution.
|
As a bonus: It will be easier for me to maintain Pillow-SIMD, because Pillow-SIMD does two separate passes from the start. |
|
Changes Unknown when pulling 3b7923c on uploadcare:resample-vertical-pass into * on python-pillow:master*. |
|
Changes Unknown when pulling c826266 on uploadcare:resample-vertical-pass into * on python-pillow:master*. |
|
Changes Unknown when pulling 12c8cf9 on uploadcare:resample-vertical-pass into * on python-pillow:master*. |
|
Are you ready for this to be merged? |
|
Yes |
This PR implements separate vertical pass of resampling instead of two transposes.
Originally I've replaced two passes with transpose for the following reasons:
Since then there are many tests for resampling were written and many optimizations were implemented. Now it is time to return upscaling performance.
Result on aws ec2 c4.large with 1920×1280 RGB image. Megapixels of source image per second.
I also include results of stable version and fixed-point separetely to show what FPI doesn't affect upscaling too much, while vertical pass does. Vertical pass faster for upscaling in average by 40% and also makes downscaling to reasonable sizes 10% faster in average.
This last two improvements together makes all resampling operations about 60% faster.