Scaling, colorspace conversion, and dithering library

[Stephen R. Savage]

Some folks have expressed interest in having a program library to perform the common tasks of manipulating resolution, colorspace, and depth. The "z" library (name subject to change) provides a simple interface to perform those tasks. An open source Vapoursynth plugin is provided, demonstrating the usage of the API.

NAME:
    z - z.lib for VapourSynth

SYNOPSIS:
    z.Format(clip clip,
             int "width",
             int "height",
             int "format",
             enum "matrix",
             enum "transfer",
             enum "primaries",
             enum "range",
             enum "chromaloc",
             enum "matrix_in",
             enum "transfer_in",
             enum "primaries_in",
             enum "range_in",
             enum "chromaloc_in",
             string "resample_filter",
             float "filter_param_a",
             float "filter_param_b",
             string "resample_filter_uv",
             float "filter_param_a_uv",
             float "filter_param_b_uv",
             string "dither_type")

    z.Subresize(clip clip,
                int width,
                int height,
                float "shift_w",
                float "shift_h",
                float "subwidth",
                float "subheight",
                string "resample_filter",
                float "filter_param_a",
                float "filter_param_b",
                string "dither_type")

DESCRIPTION:
    z.Format is a drop-in replacement for the built-in VapourSynth resize
    functions. It converts a clip of known or unknown format to another clip
    of known or unknown format, changing only the parameters specified by the
    user. z.Subresize provides advanced resampling capabilities intended for
    use by script writers.

    Arguments denoted as type "enum" may be specified by numerical index
    (see ITU-T H.265 Annex E.3) or by name. Enums specified by name have their
    argument name suffixed with "_s".

    clip:                   input clip
        The input may be of COMPAT color family (requires VS R28).

    width,
    height:                 output image dimensions

    format:                 output format preset id
        The output may be of COMPAT color family (requires VS R28).

    matrix,
    transfer,
    primaries:              output colorspace specification
        If not provided, the corresponding attribute from the input clip will
        be selected, except for YCoCg and RGB color families, where the
        corresponding matrix is set by default.

    range:                  output pixel range
        For integer formats, this allows selection of the legal code values.
        Even when set, out of range values (BTB/WTW) may be generated. If the
        input format is of a different color family, the default range is
        studio/limited for YUV and full-range for RGB.

    chromaloc:              output chroma location
        For subsampled formats, specifies the chroma location. If the input
        format is 4:4:4 or RGB and the output is subsampled, the default
        location is left-aligned, as per MPEG.

    matrix_in,
    transfer_in,
    primaries_in,
    range_in,
    chromaloc_in:           input colorspace/format specification
        If the corresponding frame property is set to a value other than
        unspecified, the frame property is used instead of this parameter.
        Default values are set for certain color families.

    resample_filter,
    filter_param_a,
    filter_param_b:         scaling method for RGB and Y-channel
        For the bicubic filter, filter_param_a/b represent the "b" and "c"
        parameters. For the lanczos filter, filter_param_a represents the
        number of taps.

    resample_filter_uv,
    resample_filter_uv_a,
    resample_filter_uv_b:   scaling method for UV channels

    dither_type:            dithering method
        Dithering is used only for conversions resulting in an integer format.

    shift_w,
    shift_h:                offset of image top-left corner
        The top-left image corner is assumed to be at coordinate (0, 0) and
        the first sample centered at coordinate (0.5, 0.5). An offset may be
        applied to the assumed image origin to "shift" the image.

    subwidth,
    subheight:              fractional dimensions of input image
        The input image is assumed to span from its origin a distance equal to
        its dimensions in pixels. An alternative image resolution may be
        specified.

    The following tables list values of selected colorspace enumerations and
    their abbreviated names. For all possible values, see ITU-T H.265.
        Matrix coefficients (ITU-T H.265 Table E.5):
        rgb         Identity
                    The identity matrix.
                    Typically used for GBR (often referred to as RGB);
                    however, may also be used for YZX (often referred to as
                    XYZ);
        709         KR = 0.2126; KB = 0.0722
                    ITU-R Rec. BT.709-5
        unspec      Unspecified
                    Image characteristics are unknown or are determined by the
                    application.
        470bg       KR = 0.299; KB = 0.114
                    ITU-R Rec. BT.470-6 System B, G (historical)
                    (functionally the same as the value 6 (170m))
        170m        KR = 0.299; KB = 0.114
                    SMPTE 170M (2004)
                    (functionally the same as the value 5 (470bg))
        ycgco       YCgCo
        2020ncl     KR = 0.2627; KB = 0.0593
                    Rec. ITU-R BT.2020 non-constant luminance system
        2020cl      KR = 0.2627; KB = 0.0593
                    Rec. ITU-R BT.2020 constant luminance system

        Transfer characteristics (ITU-T H.265 Table E.4):
        709         V = a * Lc0.45 - ( a - 1 ) for 1 >= Lc >= b
                    V = 4.500 * Lc for b > Lc >= 0
                    Rec. ITU-R BT.709-5
                    (functionally the same as the values 6 (601),
                    14 (2020_10) and 15 (2020_12))
        unspec      Unspecified
                    Image characteristics are unknown or are determined by the
                    application.
        601         V = a * Lc0.45 - ( a - 1 ) for 1 >= Lc >= b
                    V = 4.500 * Lc for b > Lc >= 0
                    Rec. ITU-R BT.601-6 525 or 625
                    (functionally the same as the values 1 (709),
                    14 (2020_10) and 15 (2020_12))
        linear      V = Lc for all values of Lc
                    Linear transfer characteristics
        2020_10     V = a * Lc0.45 - ( a - 1 ) for 1 >= Lc >= b
                    V = 4.500 * Lc for b > Lc >= 0
                    Rec. ITU-R BT.2020
                    (functionally the same as the values 1 (709),
                    6 (601) and 15 (2020_12))
        2020_12     V = a * Lc0.45 - ( a - 1 ) for 1 >= Lc >= b
                    V = 4.500 * Lc for b > Lc >= 0
                    Rec. ITU-R BT.2020
                    (functionally the same as the values 1 (709),
                    6 (601) and 14 (2020_10))

        Color primaries (ITU-T H.265 Table E.3):
        709         primary x y
                    green 0.300 0.600
                    blue 0.150 0.060
                    red 0.640 0.330
                    white D65 0.3127 0.3290
                    Rec. ITU-R BT.709-5
        unspec      Unspecified
                    Image characteristics are unknown or are determined by the
                    application.
        170m        primary x y
                    green 0.310 0.595
                    blue 0.155 0.070
                    red 0.630 0.340
                    white D65 0.3127 0.3290
                    SMPTE 170M (2004)
                    (functionally the same as the value 7 (240m))
        240m        primary x y
                    green 0.310 0.595
                    blue 0.155 0.070
                    red 0.630 0.340
                    white D65 0.3127 0.3290
                    SMPTE 240M (1999)
                    (functionally the same as the value 6 (170m))
        2020        primary x y
                    green 0.170 0.797
                    blue 0.131 0.046
                    red 0.708 0.292
                    white D65 0.3127 0.3290
                    Rec. ITU-R BT.2020

        Pixel range (ITU-T H.265 Eq E-4 to E-15):
        limited     Y = Clip1Y( Round( ( 1 << ( BitDepthY - 8 ) ) *
                                              ( 219 * E'Y + 16 ) ) )
                    Cb = Clip1C( Round( ( 1 << ( BitDepthC - 8 ) ) *
                                               ( 224 * E'PB + 128 ) ) )
                    Cr = Clip1C( Round( ( 1 << ( BitDepthC - 8 ) ) *
                                               ( 224 * E'PR + 128 ) ) )

                    R = Clip1Y( ( 1 << ( BitDepthY - 8 ) ) *
                                       ( 219 * E'R + 16 ) )
                    G = Clip1Y( ( 1 << ( BitDepthY - 8 ) ) *
                                       ( 219 * E'G + 16 ) )
                    B = Clip1Y( ( 1 << ( BitDepthY - 8 ) ) *
                                       ( 219 * E'B + 16 ) )
        full        Y = Clip1Y( Round( ( ( 1 << BitDepthY ) - 1 ) * E'Y ) )
                    Cb = Clip1C( Round( ( ( 1 << BitDepthC ) - 1 ) * E'PB +
                                          ( 1 << ( BitDepthC - 1 ) ) ) )
                    Cr = Clip1C( Round( ( ( 1 << BitDepthC ) - 1 ) * E'PR +
                                          ( 1 << ( BitDepthC - 1 ) ) ) )

                    R = Clip1Y( ( ( 1 << BitDepthY ) - 1 ) * E'R )
                    G = Clip1Y( ( ( 1 << BitDepthY ) - 1 ) * E'G )
                    B = Clip1Y( ( ( 1 << BitDepthY ) - 1 ) * E'B )

        Chroma location (ITU-T H.265 Figure E.1):
        left
        center
        top_left
        top
        bottom_left
        bottom

    The following scaling methods are available:
        point, bilinear, bicubic, spline16, spline36, lanczos
    The following dithering methods are available:
        none, ordered, random, error_diffusion

Release 2.0.2: Download link

[kolak]

How this is different than fmtconv, which I found very good?

[kolak]

Is library fully free? Is it cross-platform?

What error diffusion method is implemented? Floyd-Steinberg?

[kolak]

Can it be implemented to ffmpeg?
Ffmpeg current 'conversions' are not the best.

[Stephen R. Savage]

Originally Posted by kolak

Can it be implemented to ffmpeg?
Ffmpeg current 'conversions' are not the best.

The API header is included in the ZIP package, so surely someone more experienced with ffmpeg than me could determine if this is the case.

[Daemon404]

Originally Posted by Stephen R. Savage

The API header is included in the ZIP package, so surely someone more experienced with ffmpeg than me could determine if this is the case.

Would be pretty simple to hook up as a filter in libavfilter (which itself is pretty gross), and with and with the right filter negotiation could perhaps take the place of vf_scale. There are a bazillion places where swscale is used though that would be slightly uglier.

An alternative filter to vf_scale might be accepted upstream, since it has a C API (although they might be all NIH and butthurt about it), but swapping the actual use in the ffmpeg cli fully is pretty much a "never going to ever be accepted" scenario.

[jackoneill]

Since no one posted such a thing yet, here are some speed comparisons.

CPU is a mobile Core 2 Duo T5470, 1.6 GHz, no hyper-threading.
Due to a lack of AVX2, F16C, and FMA, all the tests use zimg's SSE2 paths.

Input is 700×480 YUV420P8, h264, 1000 frames, decoded with ffms2.

Command used:

vspipe test.py /dev/null --end 999

with an additional "--requests 1" for the 1 thread tests.

zimg version is d2e712dc54fadf45a2c55169f5a49dd74e86d62e.
fmtconv version is r8.
swscale is from ffmpeg 2.4.3.

Note that swscale never processes more than one frame at a time, because
it doesn't like multithreading (great library design). Only the input
frames are maybe fetched in parallel in the 2 thread tests.

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Upscaling by 2 using lanczos (700×480 -> 1400×960), 8 bit input:
    1 thread:
        fmtconv:    31.88 fps
        zimg:       32.11 fps
        swscale:    28.93 fps

    2 threads:
        fmtconv:    46.33 fps
        zimg:       45.19 fps
        swscale:    30.33 fps

Script used:

import vapoursynth as vs

c = vs.get_core(threads=2) # or threads=1 

src = c.ffms2.Source("700x480 YUV420P8 h264.mkv")


def resize_zimg(clip):
    src = clip
    src = c.z.Depth(src, depth=16)
    src = c.z.Resize(src, width=2*src.width, height=2*src.height, filter="lanczos")
    src = c.z.Depth(src, depth=8, dither="ordered")
    return src

def resize_fmtconv(clip):
    src = clip
    src = c.fmtc.resample(src, w=2*src.width, h=2*src.height, kernel="lanczos")
    src = c.fmtc.bitdepth(src, bits=8, dmode=0)
    return src

def resize_swscale(clip):
    src = clip
    src = c.resize.Lanczos(src, width=2*src.width, height=2*src.height)
    return src


src = resize_zimg(src)
#src = resize_swscale(src)
#src = resize_fmtconv(src)

src.set_output()

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Upscaling by 2 using lanczos (700×480 -> 1400×960), 16 bit input:
    1 thread:
        fmtconv:    40.66 fps
        zimg:       36.54 fps
        swscale:    22.89 fps

    2 threads:
        fmtconv:    55.60 fps
        zimg:       50.99 fps
        swscale:    24.66 fps

Script used:

import vapoursynth as vs

c = vs.get_core(threads=2)

src = c.ffms2.Source("700x480 YUV420P8 h264.mkv")
src = c.fmtc.bitdepth(src, bits=16)


def resize_zimg(clip):
    src = clip
    src = c.z.Resize(src, width=2*src.width, height=2*src.height, filter="lanczos")
    return src

def resize_fmtconv(clip):
    src = clip
    src = c.fmtc.resample(src, w=2*src.width, h=2*src.height, kernel="lanczos")
    return src

def resize_swscale(clip):
    src = clip
    src = c.resize.Lanczos(src, width=2*src.width, height=2*src.height)
    return src


src = resize_zimg(src)
#src = resize_swscale(src)
#src = resize_fmtconv(src)

src.set_output()

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Conversion from YUV420P8 to RGB24:
    1 thread:
        fmtconv:    60.58 fps
        zimg:       54.88 fps
        swscale:    59.05 fps

    2 threads:
        fmtconv:    73.32 fps
        zimg:       60.79 fps
        swscale:    64.14 fps

Script used:

import vapoursynth as vs

c = vs.get_core(threads=2)

src = c.ffms2.Source("700x480 YUV420P8 h264.mkv")


def test_zimg(clip):
    src = clip
    src = c.z.Depth(src, sample=1, depth=32)
    src = c.z.Resize(src, width=src.width, height=src.height, filter_uv="lanczos", subsample_w=0, subsample_h=0)
    src = c.z.Colorspace(src, 6, 6, 6, 0)
    src = c.z.Depth(src, sample=0, depth=8, dither="ordered")
    return src

def test_swscale(clip):
    src = clip
    src = c.resize.Lanczos(src, format=vs.RGB24)
    return src

def test_fmtconv(clip):
    src = clip
    src = c.fmtc.resample(src, kernel="lanczos", css="444")
    src = c.fmtc.matrix(src, mat="601", col_fam=vs.RGB)
    src = c.fmtc.bitdepth(src, bits=8, dmode=0)
    return src


src = test_zimg(src)
#src = test_swscale(src)
#src = test_fmtconv(src)

src.set_output()

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Conversion from YUV420P10 to RGB24:
    1 thread:
        fmtconv:    56.96 fps
        zimg:       53.05 fps
        swscale:    56.43 fps

    2 threads:
        fmtconv:    70.60 fps
        zimg:       59.14 fps
        swscale:    60.84 fps

Script used:

import vapoursynth as vs

c = vs.get_core(threads=2)

src = c.ffms2.Source("700x480 YUV420P8 h264.mkv")
src = c.fmtc.bitdepth(src, bits=10)


def test_zimg(clip):
    src = clip
    src = c.z.Depth(src, sample=1, depth=32)
    src = c.z.Resize(src, width=src.width, height=src.height, filter_uv="lanczos", subsample_w=0, subsample_h=0)
    src = c.z.Colorspace(src, 6, 6, 6, 0)
    src = c.z.Depth(src, sample=0, depth=8, dither="ordered")
    return src

def test_swscale(clip):
    src = clip
    src = c.resize.Lanczos(src, format=vs.RGB24)
    return src

def test_fmtconv(clip):
    src = clip
    src = c.fmtc.resample(src, kernel="lanczos", css="444")
    src = c.fmtc.matrix(src, mat="601", col_fam=vs.RGB)
    src = c.fmtc.bitdepth(src, bits=8, dmode=0)
    return src


src = test_zimg(src)
#src = test_swscale(src)
#src = test_fmtconv(src)

src.set_output()

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Bit depth conversion from 16 to 8 bits:
    1 thread:
        No dithering:
            fmtconv:    127.38 fps
            zimg:       138.32 fps

        Ordered dithering:
            fmtconv:    126.02 fps
            zimg:       139.20 fps

        Floyd-Steinberg error diffusion:
            fmtconv:    99.35 fps
            zimg:       56.43 fps

    2 threads:
        No dithering:
            fmtconv:    131.94 fps
            zimg:       134.10 fps

        Ordered dithering:
            fmtconv:    123.25 fps
            zimg:       128.98 fps

        Floyd-Steinberg error diffusion:
            fmtconv:    105.70 fps
            zimg:        69.97 fps

I have no clue what sort of dithering swscale uses, if any.
The VapourSynth filter doesn't have any parameters for it.

    1 thread:
        swscale:    142.85 fps

    2 threads:
        swscale:    142.04 fps

For these tests I used 2000 frames instead of 1000.

Script used:

import vapoursynth as vs

c = vs.get_core(threads=2)

src = c.ffms2.Source("700x480 YUV420P8 h264.mkv")
src = c.fmtc.bitdepth(src, bits=16)


def bits_zimg(clip):
    src = clip
    src = c.z.Depth(src, depth=8, dither="none") # or "ordered", or "error_diffusion"
    return src

def bits_fmtconv(clip):
    src = clip
    src = c.fmtc.bitdepth(src, bits=8, dmode=1) # or 0 for ordered, or 6 for Floyd-Steinberg error diffusion
    return src

def bits_swscale(clip):
    src = clip
    src = c.resize.Lanczos(src, format=vs.YUV420P8)
    return src


src = bits_zimg(src)
#src = bits_fmtconv(src)
#src = bits_swscale(src)

src.set_output()

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Bit depth conversion from 8 to 16 bits:
    1 thread:
        fmtconv:    159.20 fps
        zimg:       145.33 fps
        swscale:    150.64 fps

    2 threads:
        fmtconv:    148.23 fps
        zimg:       155.85 fps
        swscale:    161.81 fps

Script used:

import vapoursynth as vs

c = vs.get_core(threads=2)

src = c.ffms2.Source("700x480 YUV420P8 h264.mkv")


def bits_zimg(clip):
    src = clip
    src = c.z.Depth(src, depth=16)
    return src

def bits_fmtconv(clip):
    src = clip
    src = c.fmtc.bitdepth(src, bits=16)
    return src

def bits_swscale(clip):
    src = clip
    src = c.resize.Lanczos(src, format=vs.YUV420P16)
    return src


src = bits_zimg(src)
#src = bits_fmtconv(src)
#src = bits_swscale(src)

src.set_output()

[kolak]

What is the pure decoding speed? Does decoding use CPU or GPU?
For such a test it's better to use uncompressed source stored on fast raid or ram disk.

Thanks a lot for your time.

[jackoneill]

Originally Posted by kolak

What is the pure decoding speed? Does decoding use CPU or GPU?
For such a test it's better to use uncompressed source stored on fast raid or ram disk.

Thanks a lot for your time.

Just decoding the file is 200 fps. It uses the CPU.

[kolak]

In this case your results are affected by the decoding speed and it's possible than some may be 'incorrect'.

I'm not sure why many people use heavily compressed sources to measure speed of some filters. Quite often decoding speed is lower than filter speed, so you can't reliably compare different filters speed.
The best to use uncompressed source or encoded with very fast codecs.

[jackoneill]

It's a realistic scenario, at least for me. My sources are always compressed.

While the numbers aren't as high as they could be (with blankclip), all three plugins were tested with the same source file, so the differences between them should be correct.

[kolak]

They may be correct in this case, but you have to be careful.

If decoding speed would be 100fps than your results could suggest that all filters are as fast, which may be not the case at all.

[Myrsloik]

Originally Posted by kolak

They may be correct in this case, but you have to be careful.

If decoding speed would be 100fps than your results could suggest that all filters are as fast, which may be not the case at all.

Everything is done on the cpu so the decoding speed will scale as well. I say the test is correct.

[Stephen R. Savage]

After much (?) work, I am pleased to announce the release of "zlib" v1.0 FINAL. The library and VapourSynth example are linked in the frist post. Work is underway to improve the multithreaded scalability of the software, making it a better fit for playback scenarios

[kolak]

Originally Posted by Myrsloik

Everything is done on the cpu so the decoding speed will scale as well. I say the test is correct.

Example:

Filter A speed (from uncompressed source) = 200fps
Filter B speed (from uncompressed source) = 100fps

Decoding of compressed source= 100fps

What will be processing speed for A,B filters with compressed source?
Will filter A give faster processing or both cases give about the same speed? Even if A will give faster speed, it won't be possible to tell that A is 2x faster than B.

I understand that this was rather real scenario case and not speed banchmark for each filter itself.

[kolak]

Originally Posted by Stephen R. Savage

After much (?) work, I am pleased to announce the release of "zlib" v1.0 FINAL. The library and VapourSynth example are linked in the frist post. Work is underway to improve the multithreaded scalability of the software, making it a better fit for playback scenarios

Great, thank you for your work.

Could you add some noise generator?
I found that Floyd Steinberg dithering with a tiny amount of noise gives very good results. This small noise helps to cover contouring left after dithering.
Another thing is Filter Light algorithm, which is close to Floyd, but apparently waay faster.

[mandarinka]

Wouldn't that be a step back (removal of error-diffusion dithering)?

[kolak]

No, Floyd method is staying.
Filter Light would be an additional option. It's very close to Floyd, but can be waaay faster. At leas this is what I have read.

[YamashitaRen]

Question of the day :
What you guys have with cats ?

Answer of the day :
I tried the filter on my Odroid-U2 (unlike fmtconv, it doesn't throw me SSE2 errors) for resizing/cropping a 1920x1080 Blu-Ray to 1280x536.
Swscale is 2x as fast as zlib (8.79fps against 4.36fps). It's probably caused by the lack of NEON optimizations.

[jammupatu]

Hi,

I am very new to Vapoursynth. Just beginning to test it and its' potential in hopes of moving on from Avisynth.

I'm getting script evaluations ok using VapourSynth Editor but the preview and VirtualDub crash when I try to actually use the script. I'm just trying a simple downsize. Any number will crash, even divide by 2 (/ or // operators, actually using / the script does not evaluate). Here is my script bit:

import vapoursynth as vs
core = vs.get_core()


def downscale_zimg(clip):
    clip = core.z.Depth(clip, depth=16)
    clip = core.z.Resize(clip, width=clip.width//2, height=clip.height//2, filter="bicubic")
    clip = core.z.Depth(clip, depth=8, dither="ordered")
    return clip

source = core.ffms2.Source(source='c:/temp/test.ts', fpsnum=25, fpsden=1)
source = downscale_zimg(source)
source.set_output()

Upscaling works fine. I'm running Win 8.1.1 64b and all things 64bit; VapourSynth, Zimg, Vdub, the editor, Python 3.4.3. All components should be the latest; Zimg 1.1, VS 26 etc.

Am I missing something or is this a bug? Using the native resize does not crash on downscale.

BR,

-J