JincResize-r7.2

• Add parameters quant_x, quant_y for controlling sub-pixel quantization.

• Rename parameters, keeping consistent with AviSynth plugins.

  • crop_left -> src_left
  • crop_top -> src_top
  • crop_width -> src_width
  • crop_height -> src_height

• Give up using AVX2 on calculating convolution for 8~16bit input.

JincResize-r7.1

• Update msvc for AVX2

• The macro USE_AVX2 is defined in \include\Helper.hpp, you can modify it to use AVX2 or not.

• Todo list: Add a parameter opt to select the type of cpu optimizations.

JincResize-r7

• Fix bug

  • Fix #7: Abnormal black points on white words when 16bit or 8bit input
  • Fix bug for blur when downscaling

• Add AVX2 for convolution

• Use -O3 instead of -O2 when compilation

• Other changes in details

• Note

  • JincResize-r7: no optimization for convolution
  • JincResize-r7-avx2: use AVX2 for convolution

JincResize-r7-RC1

• Rebuild

  • Port AviSynth JincResize plugin, mainly for elliptical weighted averaging (EWA) resample

  Test 720p->1080p (100 frames) with Intel i5-4260U: core.jinc.JincResize(src, 1920, 1280)

  Version	Speed
  r7-RC1	10.839 fps
  r6.1	1.468 fps
  r6	1.704 fps

  (About 536% faster than r6, 638% faster than r6.1)

• Note

  • There is bug when there are texts on screen for 8bit and 16 bit input, while it's normal for 32bit input and 9-15bit input.

JincResize-r6.1

• Fix bug (#6)
• Update for compiler about c++17
• Adjust code structure

JincResize-r6

• Speed up

  • Add AVX optimization for LUT (look up table)
  • Some optimization on code details
  • Adjust the calculation methord of Jinc Function when tap=8 (#4)

  Test 720p->1080p (100 frames) with Intel i5-4260U: core.jinc.JincResize(src, 1920, 1280)

  Version	Speed
  r5	1.269 fps
  r6	1.594 fps

  (About 25% faster)

• Add meson for build

• Adjust code structure

• TODO List for r7

  • port AviSynth version about EWA resample

• Extra note

  • R38-R46: Adapted to VapourSynth 64bit R38-R46
  • R47-R48: Adapted to VapourSynth 64bit R47-R48

JincResize-r5

• Speed up

  • Use Taylor series for Jinc function when radius is not large. Thanks to @luglio.

    Test 720p->1080p (1200 frames) with AMD 2700x: core.jinc.JincResize(src, 1920, 1080)

    Version	Speed
    r4	9.514 fps
    r5	9.543 fps

• Other changes in details

• Extra note

  • R38-R46: Adapted to VapourSynth 64bit R38-R46
  • R47-R48: Adapted to VapourSynth 64bit R47-R48

JincResize-r4

• Add 32bit support

JincResize-r3

• Change function name: core.ewa.Lanczos -> core.jinc.JincResize
• Change parameter: float radius -> int tap
  • The original meaning of this parameter is the zero points of Jinc function
  • In old versions, users need to enter the value of the zero point by themselves with radius. Of course, users can enter any float value, but deviating from the original meaning.
  • In the current version, we provide 16 zero points. Users can choose suited zero point with tap. With the increasing of tap value, the value of zero points increases.
  • In default settings, the old and current versions are the same.
• Modifying code, maybe speed up a little.

JincResize-r2

• Speed up

Compiled by g++ instead of VS, the running speed of dynamic link library has increased significantly.

Zooming from 1280x720 to 1920x1080.
JincResize-r1 needs about 9 seconds per frame in VapourSynthEditor.
JincResize-r2 needs about 1 to 2 seconds per frame in VapourSynthEditor.

Of course, the speed of r2 version is still slow. But after all, it’s not as unacceptable as r1 version.