Skip to content

BM3D denoising filter for VapourSynth, implemented in CUDA, AVX2, HIP and SYCL

License

Notifications You must be signed in to change notification settings

WolframRhodium/VapourSynth-BM3DCUDA

Repository files navigation

VapourSynth-BM3DCUDA

Copyright© 2021 WolframRhodium

BM3D denoising filter for VapourSynth, implemented in CUDA.

Description

  • Please check VapourSynth-BM3D.

  • The _rtc version compiles GPU code at runtime, which might runs faster than standard version at the cost of a slight overhead.

  • The cpu version is implemented in AVX and AVX2 intrinsics, serves as a reference implementation on CPU. However, bitwise identical outputs are not guaranteed across CPU and CUDA implementations.

Requirements

  • CPU with AVX support.

  • CUDA-enabled GPU(s) of compute capability 5.0 or higher (Maxwell+).

  • GPU driver 450 or newer.

The minimum requirement on compute capability is 3.5, which requires manual compilation (specifying nvcc flag -gencode arch=compute_35,code=sm_35).

The cpu version does not require any external libraries but requires AVX2 support on CPU in addition.

Parameters

{bm3dcuda, bm3dcuda_rtc, bm3dcpu}.BM3D(clip clip[, clip ref=None, float[] sigma=3.0, int[] block_step=8, int[] bm_range=9, int radius=0, int[] ps_num=2, int[] ps_range=4, bint chroma=False, int device_id=0, bool fast=True, int extractor_exp=0])
  • clip:

    The input clip. Must be of 32 bit float format. Each plane is denoised separately if chroma is set to False. Data of unprocessed planes is undefined. Frame properties of the output clip are copied from it.

  • ref:

    The reference clip. Must be of the same format, width, height, number of frames as clip.

    Used in block-matching and as the reference in empirical Wiener filtering, i.e. bm3d.Final / bm3d.VFinal:

    basic = core.{bm3dcpu, bm3dcuda, bm3dcuda_rtc}.BM3D(src, radius=0)
    final = core.{bm3d...}.BM3D(src, ref=basic, radius=0)
    
    vbasic = core.{bm3d...}.BM3D(src, radius=radius_nonzero).bm3d.VAggregate(radius=radius_nonzero)
    vfinal = core.{bm3d...}.BM3D(src, ref=vbasic, radius=r).bm3d.VAggregate(radius=r)
    
    # alternatively, using the v2 interface
    basic_or_vbasic = core.{bm3dcpu, bm3dcuda, bm3dcuda_rtc}.BM3Dv2(src, radius=r)
    final_or_vfinal = core.{bm3d...}.BM3Dv2(src, ref=basic_or_vbasic, radius=r)

    corresponds to the followings (ignoring color space handling and other differences in implementation), respectively

    basic = core.bm3d.Basic(clip)
    final = core.bm3d.Final(basic, ref=src)
    
    vbasic = core.bm3d.VBasic(src, radius=r).bm3d.VAggregate(radius=r, sample=1)
    vfinal = core.bm3d.VFinal(src, ref=vbasic, radius=r).bm3d.VAggregate(radius=r)
  • sigma: The strength of denoising for each plane.

    The strength is similar (but not strictly equal) as VapourSynth-BM3D due to differences in implementation. (coefficient normalization is not implemented, for example)

    Default [3,3,3].

  • block_step, bm_range, radius, ps_num, ps_range:

    Same as those in VapourSynth-BM3D.

    If chroma is set to True, only the first value is in effect.

    Otherwise an array of values may be specified for each plane (except radius).

    Note: It is generally not recommended to take a large value of ps_num as current implementations do not take duplicate block-matching candidates into account during temporary searching, which may leads to regression in denoising quality. This issue is not present in VapourSynth-BM3D.

    Note2: Lowering the value of "block_step" will be useful in reducing blocking artifacts at the cost of slower processing.

  • chroma:

    CBM3D algorithm. clip must be of YUV444PS format.

    Y channel is used in block-matching of chroma channels.

    Default False.

  • device_id:

    Set GPU to be used.

    Default 0.

  • fast:

    Multi-threaded copy between CPU and GPU at the expense of 4x memory consumption.

    Default True.

  • extractor_exp:

    Used for deterministic (bitwise) output. This parameter is not present in the cpu version since the implementation always produces deterministic output.

    Pre-rounding is employed for associative floating-point summation.

    The value should be a positive integer not less than 3, and may need to be higher depending on the source video and filter parameters.

    Default 0. (non-determinism)

Notes

  • bm3d.VAggregate should be called after temporal filtering, as in VapourSynth-BM3D. Alternatively, you may use the BM3Dv2() interface for both spatial and temporal denoising in one step.

  • The _rtc version has three additional experimental parameters:

    • bm_error_s: (string)

      Specify cost for block similarity measurement.

      Currently implemented costs: SSD (Sum of Squared Differences), SAD (Sum of Absolute Differences), ZSSD (Zero-mean SSD), ZSAD (Zero-mean SAD), SSD/NORM.

      Default SSD.

    • transform_2d_s/transform_1d_s: (string)

      Specify type of transform.

      Currently implemented transforms: DCT (Discrete Cosine Transform), Haar (Haar Transform), WHT (Walsh–Hadamard Transform), Bior1.5 (transform based on a bi-orthogonal spline wavelet).

      Default DCT.

    These features are not implemented in the standard version due to performance and binary size concerns.

Statistics

GPU memory consumptions:

(ref ? 4 : 3) * (chroma ? 3 : 1) * (fast ? 4 : 1) * (2 * radius + 1) * size_of_a_single_frame

Compilation

  • The CMake configuration of BM3DCUDA_RTC links to NVRTC static library by default, which requires CUDA 11.5 or later.
cmake -S . -B build -D CMAKE_BUILD_TYPE=Release -D CMAKE_CUDA_FLAGS="--threads 0 --use_fast_math -Wno-deprecated-gpu-targets" -D CMAKE_CUDA_ARCHITECTURES="50;61-real;75-real;86"

cmake --build build --config Release