Speedup MedianFilter::apply() when calling repeatedly

Speedup MedianFilter::apply(), avoid to newly create the two GpuMat every time calling filter->apply(), which is very time consuming.

modules/cudafilters/src/filtering.cpp

@@ -1068,6 +1068,8 @@ namespace
     private:
         int windowSize;
         int partitions;
+        GpuMat devHist;
+        GpuMat devCoarseHist;
     };
 
     MedianFilter::MedianFilter(int srcType, int _windowSize, int _partitions) :
@@ -1098,10 +1100,13 @@ namespace
 
         // Note - these are hardcoded in the actual GPU kernel. Do not change these values.
         int histSize=256, histCoarseSize=8;
-
-        BufferPool pool(_stream);
-        GpuMat devHist = pool.getBuffer(1, src.cols*histSize*partitions,CV_32SC1);
-        GpuMat devCoarseHist = pool.getBuffer(1,src.cols*histCoarseSize*partitions,CV_32SC1);
+        int devHistCols = src.cols*histSize*partitions, devCoarseHistCols = src.cols*histCoarseSize*partitions;
+        if(devHist.empty() || devCoarseHist.empty() || devHist.cols != devHistCols || devCoarseHist.cols != devCoarseHistCols)
+        {
+          BufferPool pool(_stream);
+          devHist = pool.getBuffer(1, devHistCols, CV_32SC1);
+          devCoarseHist = pool.getBuffer(1, devCoarseHistCols, CV_32SC1);
+        }
 
         devHist.setTo(0, _stream);
         devCoarseHist.setTo(0, _stream);