auto">In this release I'm going to work on performance of most common operations. While Pillow-SIMD is the fastest library for image convolutional resize on CPU, there is still a room for optimizations without using SIMD.

Fast image allocation #2655 Fast image allocation affects almost every operation. Most of them do not require expensive computations and limited only by memory bandwidth, especially with multithreading
New Image.getchannel method #2661 New Image.getchannel() method returns single band from image, which is common operation, especially for alpha channel
Use RGBX rawmode for RGB JPEG images #1989 Using RGBX rawmode for JPEG image loading and saving eliminates unnecessary color conversion for most common image format
Fast split() #2676 Fast merge() #2677 Methods image.split() and Image.merge() use several calls of low-level functions im.getband() and im.putband() which is not effective enough. We need specific low-level functions for this operations which will be 2-2.5 times faster
Fast filters #2679 All filters from ImageFilter.py module are single channel. They are applied to multichannel images by splitting original image to bands, applying filter to each band and merging bands again. Applying filters in one pass should be much more effective. This also unlocks SIMD optimizations
Fast unpacking and partially packing #2693 Many functions in Pack.c and Unpack.c operate byte-by-byte (for example ImagingUnpackRGB). This should be more effective to read 4 bytes at once, apply transformation and write back 4 bytes. The same should be much much more effective using 128-bit SIMD registers

I also worried about VIPS benchmarks. Pillow doesn't show good result, because ≈70% of execution time is spending in filter applying (which is one of the most ineffective operation in Pillow). Funny thing is according to my tests Pillow is already faster than Python VIPS in concurrent mode on 4-cores, even with such ineffective filters. My aim to next release is to make Pillow-SIMD on single core as fast as Python VIPS on all 6 cores of Xeon E5-1650 :-> I also want to add VIPS benchmarks to pillow-perf despite the fact that Python VIPS is very hard to install.

Uh oh!

Release Pillow 4.3.0 on October 1, 2017 #2664

Description

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions