PERF: Speed up full, grid, sparse mask samplers by a local sampleVector

Did speed up `ImageFullSampler`, `ImageGridSampler`, and `ImageRandomSamplerSparseMask` by doing the `push_back`'s of samples to a local `std::vector` of samples (instead of directly to the sample container), and then eventually swapping the vector "into" the sample container. When using `ImageFullSampler` on a 4096 x 4096 input image, a `sampler.Update()` call originally took more than 0.38 seconds, but it went down to almost 0.20 seconds with this commit, as observed using Visual C++ 2019 (Release configuration).
SuperElastix · Sep 13, 2023 · 4d028f7 · 4d028f7
1 parent 5e054ee
commit 4d028f7
Show file tree

Hide file tree

Showing 3 changed files with 38 additions and 16 deletions.
diff --git a/Common/ImageSamplers/itkImageFullSampler.hxx b/Common/ImageSamplers/itkImageFullSampler.hxx
@@ -45,8 +45,10 @@ ImageFullSampler<TInputImage>::GenerateData()
   typename ImageSampleContainerType::Pointer sampleContainer = this->GetOutput();
   typename MaskType::ConstPointer            mask = this->GetMask();
 
-  /** Clear the container. */
-  sampleContainer->clear();
+  // Take capacity from the output container, and clear it.
+  std::vector<ImageSampleType> sampleVector;
+  sampleContainer->swap(sampleVector);
+  sampleVector.clear();
 
   const auto croppedInputImageRegion = this->GetCroppedInputImageRegion();
 
@@ -61,7 +63,7 @@ ImageFullSampler<TInputImage>::GenerateData()
      */
     try
     {
-      sampleContainer->reserve(croppedInputImageRegion.GetNumberOfPixels());
+      sampleVector.reserve(croppedInputImageRegion.GetNumberOfPixels());
     }
     catch (const std::exception & excp)
     {
@@ -90,7 +92,7 @@ ImageFullSampler<TInputImage>::GenerateData()
       tempSample.m_ImageValue = iter.Get();
 
       /** Store in container */
-      sampleContainer->push_back(tempSample);
+      sampleVector.push_back(tempSample);
 
     } // end for
   }   // end if no mask
@@ -115,12 +117,15 @@ ImageFullSampler<TInputImage>::GenerateData()
         tempSample.m_ImageValue = iter.Get();
 
         /** Store in container. */
-        sampleContainer->push_back(tempSample);
+        sampleVector.push_back(tempSample);
 
       } // end if
     }   // end for
   }     // end else (if mask exists)
 
+  // Move the samples from the vector into the output container.
+  sampleContainer->swap(sampleVector);
+
 } // end GenerateData()
 
 
@@ -139,6 +144,11 @@ ImageFullSampler<TInputImage>::ThreadedGenerateData(const InputImageRegionType &
   ImageSampleContainerPointer &   sampleContainerThisThread // & ???
     = this->m_ThreaderSampleContainer[threadId];
 
+  // Take capacity from the container of this thread, and clear it.
+  std::vector<ImageSampleType> sampleVector;
+  sampleContainerThisThread->swap(sampleVector);
+  sampleVector.clear();
+
   /** Set up a region iterator within the user specified image region. */
   using InputImageIterator = ImageRegionConstIteratorWithIndex<InputImageType>;
   // InputImageIterator iter( inputImage, this->GetCroppedInputImageRegion() );
@@ -152,8 +162,7 @@ ImageFullSampler<TInputImage>::ThreadedGenerateData(const InputImageRegionType &
      */
     try
     {
-      sampleContainerThisThread->clear();
-      sampleContainerThisThread->reserve(chunkSize);
+      sampleVector.reserve(chunkSize);
     }
     catch (const std::exception & excp)
     {
@@ -182,7 +191,7 @@ ImageFullSampler<TInputImage>::ThreadedGenerateData(const InputImageRegionType &
       tempSample.m_ImageValue = iter.Get();
 
       /** Store in container. */
-      sampleContainerThisThread->push_back(tempSample);
+      sampleVector.push_back(tempSample);
 
     } // end for
   }   // end if no mask
@@ -207,12 +216,15 @@ ImageFullSampler<TInputImage>::ThreadedGenerateData(const InputImageRegionType &
         tempSample.m_ImageValue = iter.Get();
 
         /**  Store in container. */
-        sampleContainerThisThread->push_back(tempSample);
+        sampleVector.push_back(tempSample);
 
       } // end if
     }   // end for
   }     // end else (if mask exists)
 
+  // Move the samples from the vector into the container for this thread.
+  sampleContainerThisThread->swap(sampleVector);
+
 } // end ThreadedGenerateData()
 
 

diff --git a/Common/ImageSamplers/itkImageGridSampler.hxx b/Common/ImageSamplers/itkImageGridSampler.hxx
@@ -55,8 +55,10 @@ ImageGridSampler<TInputImage>::GenerateData()
   typename ImageSampleContainerType::Pointer sampleContainer = this->GetOutput();
   typename MaskType::ConstPointer            mask = this->GetMask();
 
-  /** Clear the container. */
-  sampleContainer->Initialize();
+  // Take capacity from the output container, and clear it.
+  std::vector<ImageSampleType> sampleVector;
+  sampleContainer->swap(sampleVector);
+  sampleVector.clear();
 
   /** Take into account the possibility of a smaller bounding box around the mask */
   this->SetNumberOfSamples(this->m_RequestedNumberOfSamples);
@@ -120,7 +122,7 @@ ImageGridSampler<TInputImage>::GenerateData()
             index[0] += this->m_SampleGridSpacing[0];
 
             // Store sample in container.
-            sampleContainer->push_back(tempSample);
+            sampleVector.push_back(tempSample);
 
           } // end x
           index[0] = sampleGridIndex[0];
@@ -165,7 +167,7 @@ ImageGridSampler<TInputImage>::GenerateData()
               tempSample.m_ImageValue = inputImage->GetPixel(index);
 
               // Store sample in container.
-              sampleContainer->push_back(tempSample);
+              sampleVector.push_back(tempSample);
 
             } // end if in mask
               // Jump to next position on grid
@@ -190,6 +192,9 @@ ImageGridSampler<TInputImage>::GenerateData()
     } // end t
   }   // else (if mask exists)
 
+  // Move the samples from the vector into the output container.
+  sampleContainer->swap(sampleVector);
+
 } // end GenerateData()
 
 

diff --git a/Common/ImageSamplers/itkImageRandomSamplerSparseMask.hxx b/Common/ImageSamplers/itkImageRandomSamplerSparseMask.hxx
@@ -44,8 +44,10 @@ ImageRandomSamplerSparseMask<TInputImage>::GenerateData()
   InputImageConstPointer      inputImage = this->GetInput();
   ImageSampleContainerPointer sampleContainer = this->GetOutput();
 
-  /** Clear the container. */
-  sampleContainer->Initialize();
+  // Take capacity from the output container, and clear it.
+  std::vector<ImageSampleType> sampleVector;
+  sampleContainer->swap(sampleVector);
+  sampleVector.clear();
 
   /** Make sure the internal full sampler is up-to-date. */
   this->m_InternalFullSampler->SetInput(inputImage);
@@ -92,9 +94,12 @@ ImageRandomSamplerSparseMask<TInputImage>::GenerateData()
   for (unsigned int i = 0; i < this->GetNumberOfSamples(); ++i)
   {
     unsigned long randomIndex = this->m_RandomGenerator->GetIntegerVariate(numberOfValidSamples - 1);
-    sampleContainer->push_back(allValidSamples->ElementAt(randomIndex));
+    sampleVector.push_back(allValidSamples->ElementAt(randomIndex));
   }
 
+  // Move the samples from the vector into the output container.
+  sampleContainer->swap(sampleVector);
+
 } // end GenerateData()