Implement optimized version of alpha darken composite op for RGBF32
AbandonedPublic
Actions

Authored by dkazakov on Feb 29 2016, 2:52 PM.

Details

Reviewers

Group Reviewers

Krita: Stable
Krita

Summary

Added optimized version for alpha darken composite op for RGBF32 colorspace.
Added tests to test performance and results of new implementation against legacy.

The diff needed in the test compare is do to the fact that the compiler calculates 1.0/255.0 and multiplying the result instead if dividing by 255.0 for the mask.

Here are the results of the benchmark on my Intel i5-2520M CPU

QDEBUG : KisCompositionBenchmark::testRgbF32CompositeAlphaDarkenLegacy() krita.general: Testing Composite Op: "alphadarken" ( "Legacy" )
QDEBUG : KisCompositionBenchmark::testRgbF32CompositeAlphaDarkenLegacy() krita.general: "Aligned Mask SrcRand DstRand" RESULT: 67 msec
QDEBUG : KisCompositionBenchmark::testRgbF32CompositeAlphaDarkenLegacy() krita.general: "DstUnalig Mask SrcRand DstRand" RESULT: 68 msec
QDEBUG : KisCompositionBenchmark::testRgbF32CompositeAlphaDarkenLegacy() krita.general: "SrcUnalig Mask SrcRand DstRand" RESULT: 69 msec
QDEBUG : KisCompositionBenchmark::testRgbF32CompositeAlphaDarkenLegacy() krita.general: "Unaligned Mask SrcRand DstRand" RESULT: 66 msec
QDEBUG : KisCompositionBenchmark::testRgbF32CompositeAlphaDarkenLegacy() krita.general: "Aligned NoMask SrcRand DstRand" RESULT: 33 msec
QDEBUG : KisCompositionBenchmark::testRgbF32CompositeAlphaDarkenLegacy() krita.general: "Aligned NoMask SrcZero DstRand" RESULT: 32 msec
QDEBUG : KisCompositionBenchmark::testRgbF32CompositeAlphaDarkenLegacy() krita.general: "Aligned NoMask SrcUnit DstRand" RESULT: 32 msec
QDEBUG : KisCompositionBenchmark::testRgbF32CompositeAlphaDarkenLegacy() krita.general: "Aligned NoMask SrcRand DstZero" RESULT: 31 msec
QDEBUG : KisCompositionBenchmark::testRgbF32CompositeAlphaDarkenLegacy() krita.general: "Aligned NoMask SrcZero DstZero" RESULT: 28 msec
QDEBUG : KisCompositionBenchmark::testRgbF32CompositeAlphaDarkenLegacy() krita.general: "Aligned NoMask SrcUnit DstZero" RESULT: 31 msec
QDEBUG : KisCompositionBenchmark::testRgbF32CompositeAlphaDarkenLegacy() krita.general: "Aligned NoMask SrcRand DstUnit" RESULT: 32 msec
QDEBUG : KisCompositionBenchmark::testRgbF32CompositeAlphaDarkenLegacy() krita.general: "Aligned NoMask SrcZero DstUnit" RESULT: 32 msec
QDEBUG : KisCompositionBenchmark::testRgbF32CompositeAlphaDarkenLegacy() krita.general: "Aligned NoMask SrcUnit DstUnit" RESULT: 33 msec

QDEBUG : KisCompositionBenchmark::testRgbF32CompositeAlphaDarkenOptimized() krita.general: "Aligned Mask SrcRand DstRand" RESULT: 12 msec
QDEBUG : KisCompositionBenchmark::testRgbF32CompositeAlphaDarkenOptimized() krita.general: "DstUnalig Mask SrcRand DstRand" RESULT: 12 msec
QDEBUG : KisCompositionBenchmark::testRgbF32CompositeAlphaDarkenOptimized() krita.general: "SrcUnalig Mask SrcRand DstRand" RESULT: 16 msec
QDEBUG : KisCompositionBenchmark::testRgbF32CompositeAlphaDarkenOptimized() krita.general: "Unaligned Mask SrcRand DstRand" RESULT: 16 msec
QDEBUG : KisCompositionBenchmark::testRgbF32CompositeAlphaDarkenOptimized() krita.general: "Aligned NoMask SrcRand DstRand" RESULT: 9 msec
QDEBUG : KisCompositionBenchmark::testRgbF32CompositeAlphaDarkenOptimized() krita.general: "Aligned NoMask SrcZero DstRand" RESULT: 9 msec
QDEBUG : KisCompositionBenchmark::testRgbF32CompositeAlphaDarkenOptimized() krita.general: "Aligned NoMask SrcUnit DstRand" RESULT: 14 msec
QDEBUG : KisCompositionBenchmark::testRgbF32CompositeAlphaDarkenOptimized() krita.general: "Aligned NoMask SrcRand DstZero" RESULT: 9 msec
QDEBUG : KisCompositionBenchmark::testRgbF32CompositeAlphaDarkenOptimized() krita.general: "Aligned NoMask SrcZero DstZero" RESULT: 9 msec
QDEBUG : KisCompositionBenchmark::testRgbF32CompositeAlphaDarkenOptimized() krita.general: "Aligned NoMask SrcUnit DstZero" RESULT: 9 msec
QDEBUG : KisCompositionBenchmark::testRgbF32CompositeAlphaDarkenOptimized() krita.general: "Aligned NoMask SrcRand DstUnit" RESULT: 9 msec
QDEBUG : KisCompositionBenchmark::testRgbF32CompositeAlphaDarkenOptimized() krita.general: "Aligned NoMask SrcZero DstUnit" RESULT: 10 msec
QDEBUG : KisCompositionBenchmark::testRgbF32CompositeAlphaDarkenOptimized() krita.general: "Aligned NoMask SrcUnit DstUnit" RESULT: 9 msec

This is a speedup of factor 3 to 6.

Diff Detail

Repository

R37 Krita

Lint

Lint Skipped

Unit

Unit Tests Skipped

zachmann updated this revision to Diff 2495.Feb 29 2016, 2:52 PM

zachmann retitled this revision from to Implement optimized version of alpha darken composite op for RGBF32.

zachmann updated this object.

zachmann edited the test plan for this revision. (Show Details)

zachmann added reviewers: Krita: Stable, Krita, dkazakov.

zachmann set the repository for this revision to R37 Krita.

zachmann added projects: Krita, Krita: Stable.

Restricted Application added a subscriber: woltherav. · View Herald TranscriptFeb 29 2016, 2:52 PM

Hi, Thorsten!

Thank you very much for the patch! It looks perfect! And it compiles and runs perfect as well! Feel free to commit it to master :)

This revision is now accepted and ready to land.Mar 1 2016, 12:11 PM

The revision is pushed to master long ago. So I'm commandeering it to close as Submitted.

This revision now requires review to proceed.Aug 1 2016, 6:18 PM

dkazakov abandoned this revision.Aug 1 2016, 6:18 PM

Revision Contents
Changeset List

		Path
M		benchmarks/kis_composition_benchmark.h (9 lines)
M		benchmarks/kis_composition_benchmark.cpp (65 lines)
A	M	libs/pigment/compositeops/KoOptimizedCompositeOpAlphaDarken128.h (221 lines)
M		libs/pigment/compositeops/KoOptimizedCompositeOpFactory.h (1 line)
M		libs/pigment/compositeops/KoOptimizedCompositeOpFactory.cpp (5 lines)
M		libs/pigment/compositeops/KoOptimizedCompositeOpFactoryPerArch.h (3 lines)
M		libs/pigment/compositeops/KoOptimizedCompositeOpFactoryPerArch.cpp (9 lines)