Changeset 17320 for trunk/psLib/src/fft/psImageFFT.h

Timestamp:

Apr 4, 2008, 12:44:56 PM (18 years ago)

Author:

Paul Price

Message:

Optimising psImageConvolveFFT. Made a couple of optimisations: 1. Moved code into psImageFFT.c, so I can use FFTW directly. FFTW, through its 'advanced interface' allows two FFTs to be performed at the same time and 'the resulting plans can often be faster than calling FFTW multiple times for the individual transforms' (FFTW 3.1.2 manual, p30). 2. The convolved image needs to be normalised because FFTW doesn't take out the sqrt(N) factors in the FFT. Instead of multiplying the entire convolved image, we can multiply the much smaller number of pixels in the kernel. Tested this code with tap_psImageConvolve2 and it seems to work. Running convolutionBench unoptimised on mithrandir produces times that are comparable to what I got a while back running optimised. Running it optimized on alala produces speed-ups ranging from 25 to 350%.

File:

• trunk/psLib/src/fft/psImageFFT.h (modified) (3 diffs)

trunk/psLib/src/fft/psImageFFT.h

@@ -5 +5 @@
 /// @author Robert DeSonia, MHPCC
 ///
-/// @version $Revision: 1.9 $ $Name: not supported by cvs2svn $
-/// @date $Date: 2007-02-08 04:23:57 $
+/// @version $Revision: 1.10 $ $Name: not supported by cvs2svn $
+/// @date $Date: 2008-04-04 22:44:56 $
 /// Copyright 2004-2005 Maui High Performance Computing Center, University of Hawaii
 ///
@@ -14 +14 @@
 
 #include "psImage.h"
+#include "psImageConvolve.h"
 
 /// @addtogroup MathOps Mathematical Operations
@@ -61 +62 @@
     );
 
+/// Convolve an image with a kernel, using the FFT
+///
+/// This is appropriate for larger kernels, where the direct convolution is slow.  The input image and kernel
+/// are suitably padded to avoid wrap-around effects.
+psImage *psImageConvolveFFT(
+    psImage *out,                       ///< Output image, or NULL
+    const psImage *in,                  ///< Image to convolve
+    const psImage *mask,                ///< Corresponding mask
+    psMaskType maskVal,                 ///< Value to mask
+    const psKernel *kernel              ///< kernel to colvolve with
+);
+
 /// @}
 #endif // #ifndef PS_IMAGE_FFT_H