|
| 1 | +/*========================================================================= |
| 2 | + * |
| 3 | + * Copyright NumFOCUS |
| 4 | + * |
| 5 | + * Licensed under the Apache License, Version 2.0 (the "License"); |
| 6 | + * you may not use this file except in compliance with the License. |
| 7 | + * You may obtain a copy of the License at |
| 8 | + * |
| 9 | + * http://www.apache.org/licenses/LICENSE-2.0.txt |
| 10 | + * |
| 11 | + * Unless required by applicable law or agreed to in writing, software |
| 12 | + * distributed under the License is distributed on an "AS IS" BASIS, |
| 13 | + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 14 | + * See the License for the specific language governing permissions and |
| 15 | + * limitations under the License. |
| 16 | + * |
| 17 | + *=========================================================================*/ |
| 18 | +#ifndef itkVkMultiResolutionPyramidImageFilter_h |
| 19 | +#define itkVkMultiResolutionPyramidImageFilter_h |
| 20 | + |
| 21 | +#include "itkMultiResolutionPyramidImageFilter.h" |
| 22 | + |
| 23 | +#include "itkDiscreteGaussianImageFilter.h" |
| 24 | +#include "itkFFTDiscreteGaussianImageFilter.h" |
| 25 | +#include "itkVector.h" |
| 26 | +#include "itkMacro.h" |
| 27 | +#include "VkFFTBackendExport.h" |
| 28 | + |
| 29 | +#include <string> |
| 30 | + |
| 31 | +namespace itk |
| 32 | +{ |
| 33 | + |
| 34 | +/** \class VkMultiResolutionPyramidImageFilter |
| 35 | + * \brief Creates a multi-resolution pyramid with FFT acceleration |
| 36 | + * |
| 37 | + * VkMultiResolutionPyramidImageFilter re-implements a framework |
| 38 | + * for creating an image pyramid as laid out in |
| 39 | + * MultiResolutionPyramidImageFilter. Conditional logic is added |
| 40 | + * to preemptively select the optimal image smoothing pipeline |
| 41 | + * that is expected to give the best performance for different |
| 42 | + * pyramid levels. |
| 43 | + * |
| 44 | + * Separable spatial convolution with DiscreteGaussianImageFilter |
| 45 | + * runs quickly for small kernel sizes but scales poorly with |
| 46 | + * increasing kernel size. By contrast ITK FFT convolution accelerated |
| 47 | + * with a VkFFT GPU backend scales slowly with increasing kernel size |
| 48 | + * but is typically outperformed by spatial convolution filters |
| 49 | + * for small kernel sizes. |
| 50 | + * |
| 51 | + * VkMultiResolutionPyramidImageFilter allows the user to fix the |
| 52 | + * metric threshold at which a performance tradeoff is expected |
| 53 | + * between spatial and FFT convolution. The exact threshold depends |
| 54 | + * on user hardware and can be estimated through benchmarking with |
| 55 | + * scripts in the ITKVkFFTBackend repository. |
| 56 | + * |
| 57 | + * By mitigating blurring times on levels with large kernel sizes |
| 58 | + * VkMultiResolutionPyramidImageFilter has been observed to run in |
| 59 | + * as little as 50% of the time of its base class. |
| 60 | + * |
| 61 | + * See documentation of MultiResolutionPyramidImageFilter |
| 62 | + * for information on how to specify a multi-resolution schedule. |
| 63 | + * |
| 64 | + * \sa MultiResolutionPyramidImageFilter |
| 65 | + * \sa DiscreteGaussianImageFilter |
| 66 | + * \sa FFTDiscreteGaussianImageFilter |
| 67 | + * \sa ShrinkImageFilter |
| 68 | + * |
| 69 | + * \ingroup VkFFTBackend |
| 70 | + * \ingroup PyramidImageFilter |
| 71 | + * \ingroup ITKRegistrationCommon |
| 72 | + */ |
| 73 | +template <typename TInputImage, typename TOutputImage> |
| 74 | +class ITK_TEMPLATE_EXPORT VkMultiResolutionPyramidImageFilter |
| 75 | + : public MultiResolutionPyramidImageFilter<TInputImage, TOutputImage> |
| 76 | +{ |
| 77 | +public: |
| 78 | + ITK_DISALLOW_COPY_AND_MOVE(VkMultiResolutionPyramidImageFilter); |
| 79 | + |
| 80 | + /** Standard class type aliases. */ |
| 81 | + using Self = VkMultiResolutionPyramidImageFilter; |
| 82 | + using Superclass = MultiResolutionPyramidImageFilter<TInputImage, TOutputImage>; |
| 83 | + using Pointer = SmartPointer<Self>; |
| 84 | + using ConstPointer = SmartPointer<const Self>; |
| 85 | + |
| 86 | + /** Method for creation through the object factory. */ |
| 87 | + itkNewMacro(Self); |
| 88 | + |
| 89 | + /** Run-time type information (and related methods). */ |
| 90 | + itkTypeMacro(VkMultiResolutionPyramidImageFilter, MultiResolutionPyramidImageFilter); |
| 91 | + |
| 92 | + /** ImageDimension enumeration. */ |
| 93 | + static constexpr unsigned int ImageDimension = TInputImage::ImageDimension; |
| 94 | + |
| 95 | + /** Inherit types from Superclass. */ |
| 96 | + using typename Superclass::InputImageType; |
| 97 | + using typename Superclass::OutputImageType; |
| 98 | + using typename Superclass::InputImagePointer; |
| 99 | + using typename Superclass::OutputImagePointer; |
| 100 | + using typename Superclass::InputImageConstPointer; |
| 101 | + using InputSizeType = typename InputImageType::SizeType; |
| 102 | + using OutputPixelType = typename OutputImageType::PixelType; |
| 103 | + using OutputSizeType = typename OutputImageType::SizeType; |
| 104 | + using typename Superclass::ScheduleType; |
| 105 | + |
| 106 | + using VarianceType = itk::Vector<double, ImageDimension>; |
| 107 | + using KernelSizeType = OutputSizeType; |
| 108 | + |
| 109 | + /** Types for acceleration. |
| 110 | + * Assumes and does not verify that FFT backend is accelerated. */ |
| 111 | + using BaseSmootherType = DiscreteGaussianImageFilter<OutputImageType, OutputImageType>; |
| 112 | + using SpatialSmootherType = DiscreteGaussianImageFilter<OutputImageType, OutputImageType>; |
| 113 | + using FFTSmootherType = FFTDiscreteGaussianImageFilter<OutputImageType, OutputImageType>; |
| 114 | + |
| 115 | + /** Set the metric threshold to decide between |
| 116 | + * accelerated methods such as CPU-based separable smoothing |
| 117 | + * versus GPU-based FFT smoothing. |
| 118 | + * We can predictively compare spatial and FFT smoothing |
| 119 | + * performance using the following metric: |
| 120 | + * |
| 121 | + * f(i,j,k,x,y,z) = log((i + j + k) * x * y * z) |
| 122 | + * |
| 123 | + * where i,j,k are the dimensions of the kernel for a given |
| 124 | + * pyramid level and x,y,z are the dimensions of the |
| 125 | + * output image region. |
| 126 | + * |
| 127 | + * The equation above approximates the difference in runtime complexity |
| 128 | + * between separable spatial Gaussian smoothing and FFT Gaussian smoothing. |
| 129 | + * Under separable smoothing each pixel [xi,yi,zi] is used in computation |
| 130 | + * approximately (i + j + k) times. FFT smoothing meanwhile has significant |
| 131 | + * overhead in setup but scales much more slowly with kernel and image sizes. |
| 132 | + * As a result there is an approximate threshold where GPU-accelerated |
| 133 | + * smoothing outperforms spatial smoothing for a given pyramid level. |
| 134 | + * |
| 135 | + * The default threshold value 8.0 has been empirically determined as |
| 136 | + * a reasonable approximation such that f(...) < 8.0 indicates that |
| 137 | + * spatial convolution will run faster while f(...) > 8.0 indicates that |
| 138 | + * FFT convolution will run faster. The threshold value is not universal |
| 139 | + * and may need to be adjusted to better match benchmarking results for |
| 140 | + * particular hardware and expected image sizes so that nuances such as |
| 141 | + * multithreading and GPU performance may be taken into account. |
| 142 | + */ |
| 143 | + itkSetMacro(MetricThreshold, float); |
| 144 | + itkGetMacro(MetricThreshold, float); |
| 145 | + |
| 146 | + /** Set the metric threshold from a certain parameter set describing the input size |
| 147 | + * and kernel radius threshold that is expected to be equally fast with separable |
| 148 | + * spatial smoothing and FFT smoothing */ |
| 149 | + void |
| 150 | + SetMetricThreshold(const InputSizeType & inputSize, const KernelSizeType & kernelRadius) |
| 151 | + { |
| 152 | + this->SetMetricThreshold(ComputeMetricValue(inputSize, kernelRadius)); |
| 153 | + } |
| 154 | + |
| 155 | + float |
| 156 | + ComputeMetricValue(const InputSizeType & inputSize, const KernelSizeType & kernelRadius) const; |
| 157 | + |
| 158 | + /** Estimate the kernel radius from ilevel settings */ |
| 159 | + KernelSizeType |
| 160 | + GetKernelRadius(unsigned int ilevel) const; |
| 161 | + |
| 162 | + /** Get the kernel variance for the given pyramid level |
| 163 | + * based on the current schedule */ |
| 164 | + VarianceType |
| 165 | + GetVariance(unsigned int ilevel) const; |
| 166 | + |
| 167 | + /** Get whether FFT smoothing will be used for the given |
| 168 | + * pyramid level */ |
| 169 | + bool |
| 170 | + GetUseFFT(const KernelSizeType & kernelRadius) const; |
| 171 | + |
| 172 | +protected: |
| 173 | + VkMultiResolutionPyramidImageFilter() = default; |
| 174 | + ~VkMultiResolutionPyramidImageFilter() override = default; |
| 175 | + |
| 176 | + /** Generate the output data. */ |
| 177 | + void |
| 178 | + GenerateData() override; |
| 179 | + |
| 180 | + void |
| 181 | + PrintSelf(std::ostream & os, Indent indent) const override; |
| 182 | + |
| 183 | +private: |
| 184 | + float m_MetricThreshold = 8.0f; |
| 185 | + typename SpatialSmootherType::Pointer spatialSmoother = SpatialSmootherType::New(); |
| 186 | + typename FFTSmootherType::Pointer fftSmoother = FFTSmootherType::New(); |
| 187 | + |
| 188 | +}; |
| 189 | +} // namespace itk |
| 190 | + |
| 191 | +#ifndef ITK_MANUAL_INSTANTIATION |
| 192 | +# include "itkVkMultiResolutionPyramidImageFilter.hxx" |
| 193 | +#endif |
| 194 | + |
| 195 | +#endif |
0 commit comments