GSoC project
Implement AVX instruction set to mask generation to accelerate painting performance on Circular Gauss, Circular soft, Rectangular Gauss, Rectangular Soft, Stamp brush masks.
| Brush | Ported to processors | Implemented Vc version | Merged into master |
| Default Circle | x | x | x |
| Default Rect | x | x | x |
| Soft Circle | x | x | x |
| Soft Rect | x | x | x |
| Gaussian Circle | x | x | x |
| Gaussian Rect | x | x | x |
| Stamp | |||
| Status | Assigned | Task | ||
|---|---|---|---|---|
| Resolved | vanyossi | T8580 Krita AVX optimization for brush mask generation: Gauss, Soft, Stamp, Rectangular Gauss, Rectangular Soft | ||
| Open | vanyossi | T8581 Base unit test kis_brush_mask_equality | ||
| Resolved | vanyossi | T8734 Implement Circular Gauss Mask Optim AVX | ||
| Resolved | vanyossi | T8868 Implement Circular Soft mask Vector Vc optimization | ||
| Resolved | vanyossi | T9010 Optimize Rectangular Gauss Mask | ||
| Resolved | vanyossi | T9133 Optimize Rectangular Soft with Vc code | ||
| Resolved | vanyossi | T9344 Optimize Rectangular Default Brush | ||
| Open | vanyossi | T9346 Optimize predefined Mask generator |
Made a benchfark for the generation of the mask, I let both scalar and vectorized coexsist to check how much different they are between them (although this was an after thought, af first only vector mask were in).
Preversion of KisMaskGeneratorBenchmark
Some preliminary results in my computer:
./KisMaskGeneratorBenchmark -iterations 300
Config: Using QtTest library 5.10.0, Qt 5.10.0 (x86_64-little_endian-lp64 shared (dynamic) release build; by Clang 9.0.0 (clang-900.0.39.2) (Apple))
PASS : KisMaskGeneratorBenchmark::initTestCase()
QDEBUG : KisMaskGeneratorBenchmark::testDefaultScalarMask() Invalid profile : "/Library/ColorSync/Profiles//WebSafeColors.icc" "Web Safe Colors"
PASS : KisMaskGeneratorBenchmark::testDefaultScalarMask()
RESULT : KisMaskGeneratorBenchmark::testDefaultScalarMask():
24.27 msecs per iteration (total: 7,282, iterations: 300)
PASS : KisMaskGeneratorBenchmark::testDefaultVectorMask()
RESULT : KisMaskGeneratorBenchmark::testDefaultVectorMask():
1.52 msecs per iteration (total: 458, iterations: 300)
PASS : KisMaskGeneratorBenchmark::testCircularGaussScalarMask()
RESULT : KisMaskGeneratorBenchmark::testCircularGaussScalarMask():
68.450 msecs per iteration (total: 20,535, iterations: 300)
PASS : KisMaskGeneratorBenchmark::testCircularGaussVectorMask()
RESULT : KisMaskGeneratorBenchmark::testCircularGaussVectorMask():
6.253 msecs per iteration (total: 1,876, iterations: 300)
PASS : KisMaskGeneratorBenchmark::cleanupTestCase()
Totals: 6 passed, 0 failed, 0 skipped, 0 blacklisted, 62320ms
Hi, @vanyossi!
The benchmark results look good. It is a bit suspicious is that for Default tip speed benefit is 16x and for Gaussian tip it is only 10x. It might be a natural difference, but I would probably try to spend a small bit of time (not more than a couple of hours) to check why this difference happens. Again, it might be perfectly okay that this difference exists, but I would check it, just out of curiosity :)