Merge branch 'padded-convolution' into 'master'

Add function for padded convolutions

See merge request !169

include/schaapcommon/math/paddedconvolution.h

+#include "convolution.h"
+
+#include <aocommon/image.h>
+
+#ifndef SCHAAPCOMMON_MATH_PADDED_CONVOLUTION_H_
+#define SCHAAPCOMMON_MATH_PADDED_CONVOLUTION_H_
+
+namespace schaapcommon::math {
+
+/**
+ * Performs an in-place padded image convolution. The convolution kernel (psf)
+ * is appropriately transformed so that its centre pixel is the centre of the
+ * kernel. This overload requires two scratch images that have been allocated
+ * with the padded dimensions.
+ *
+ * This operation is sometimes referred to as correlation, because the kernel is
+ * not flipped as is formally done when convolving. The operation performed by
+ * this function is typically used to convolve an interferometric model image
+ * with the PSF.
+ *
+ * This overload allows specifying the scratch images, to minimize allocations
+ * for use-cases that perform multiple padded convolutions. Both scratch images
+ * should be allocated with the padded size on input.
+ */
+inline void PaddedConvolution(aocommon::Image& image,
+                              const aocommon::Image& psf,
+                              aocommon::Image& scratch_a,
+                              aocommon::Image& scratch_b, size_t padded_width,
+                              size_t padded_height) {
+  assert(image.Width() == psf.Width());
+  assert(image.Height() == psf.Height());
+  assert(image.Size() > 0);
+  assert(scratch_a.Width() * scratch_a.Height() >=
+         padded_width * padded_height);
+  assert(scratch_b.Width() * scratch_b.Height() >=
+         padded_width * padded_height);
+  assert(padded_width >= image.Width());
+  assert(padded_height >= image.Height());
+  using aocommon::Image;
+  // scratch_a = padded psf
+  Image::Untrim(scratch_a.Data(), padded_width, padded_height, psf.Data(),
+                psf.Width(), psf.Height());
+  // scratch_b = prepared padded psf
+  PrepareConvolutionKernel(scratch_b.Data(), scratch_a.Data(), padded_width,
+                           padded_height);
+  // scratch_a = padded image
+  Image::Untrim(scratch_a.Data(), padded_width, padded_height, image.Data(),
+                image.Width(), image.Height());
+
+  // Convolve and store in scratch_a
+  Convolve(scratch_a.Data(), scratch_b.Data(), padded_width, padded_height);
+
+  Image::Trim(image.Data(), image.Width(), image.Height(), scratch_a.Data(),
+              padded_width, padded_height);
+}
+
+/**
+ * Convenience overload of the above PaddedConvolution() function that does
+ * not require scratch images. The required scratch images are allocated
+ * inside the function. See other overload for help.
+ */
+inline void PaddedConvolution(aocommon::Image& image,
+                              const aocommon::Image& psf, size_t padded_width,
+                              size_t padded_height) {
+  aocommon::Image scratch_a(padded_width, padded_height);
+  aocommon::Image scratch_b(padded_width, padded_height);
+  PaddedConvolution(image, psf, scratch_a, scratch_b, padded_width,
+                    padded_height);
+}
+
+}  // namespace schaapcommon::math
+
+#endif

src/math/test/CMakeLists.txt

@@ -3,5 +3,11 @@
 
 include(${SCHAAPCOMMON_SOURCE_DIR}/cmake/unittest.cmake)
 
-add_unittest(math runtests.cc tconvolution.cc tdrawgaussian.cc tresampler.cc
-             trestoreimage.cc)
+add_unittest(
+  math
+  runtests.cc
+  tconvolution.cc
+  tdrawgaussian.cc
+  tpaddedconvolution.cc
+  tresampler.cc
+  trestoreimage.cc)

src/math/test/tpaddedconvolution.cc

+#include <boost/test/unit_test.hpp>
+
+#include <aocommon/image.h>
+
+#include "paddedconvolution.h"
+
+using aocommon::Image;
+
+namespace schaapcommon::math {
+namespace {
+
+void Check(size_t width, size_t height) {
+  // Use a padding factor of 3, so that both padded even sizes remain even
+  // and padded odd sizes remain odd.
+  constexpr size_t kPadding = 3;
+
+  Image image(width, height, 0.0f);
+  const Image zero(width, height, 0.0f);
+
+  PaddedConvolution(image, zero, width * kPadding, height * kPadding);
+  BOOST_CHECK_LT(image.RMS(), 1e-5);
+
+  Image delta(width, height, 0.0f);
+  delta.Value(width / 2, height / 2) = 1.0f;
+
+  PaddedConvolution(image, delta, width * kPadding, height * kPadding);
+  BOOST_CHECK_LT(image.RMS(), 1e-5);
+
+  for (size_t i = 0; i != image.Size(); ++i) image[i] = i + 3;
+  PaddedConvolution(image, delta, width * kPadding, height * kPadding);
+  for (size_t i = 0; i != image.Size(); ++i)
+    BOOST_CHECK_CLOSE_FRACTION(image[i], i + 3, 1e-6);
+
+  delta *= 2.0f;
+  PaddedConvolution(image, delta, width * kPadding, height * kPadding);
+  for (size_t i = 0; i != image.Size(); ++i)
+    BOOST_CHECK_CLOSE_FRACTION(image[i], (i + 3) * 2, 1e-5);
+}
+
+}  // namespace
+
+BOOST_AUTO_TEST_SUITE(padded_convolution)
+
+BOOST_AUTO_TEST_CASE(simple_even) { Check(10, 8); }
+
+BOOST_AUTO_TEST_CASE(simple_odd) { Check(9, 11); }
+
+BOOST_AUTO_TEST_CASE(full_example) {
+  Image image(8, 6, 0.0f);
+  image.Value(1, 3) = 3.0;
+  image.Value(4, 5) = 1.5;
+  image.Value(6, 5) = -1.0;
+
+  Image psf(8, 6, 0.0f);
+  psf.Value(4, 3) = 0.5;
+  psf.Value(3, 3) = 3.0;
+  psf.Value(5, 3) = 7.0;
+  psf.Value(4, 2) = 10.5;
+  psf.Value(4, 4) = 37.0;
+
+  PaddedConvolution(image, psf, 20, 30);
+  const Image reference(
+      8, 6,
+      {
+          0.0f, 0.0f,   0.0f,  0.0f, 0.0f,   0.0f, 0.0f,   0.0f,  // row 0
+          0.0f, 0.0f,   0.0f,  0.0f, 0.0f,   0.0f, 0.0f,   0.0f,  // row 1
+          0.0f, 31.5f,  0.0f,  0.0f, 0.0f,   0.0f, 0.0f,   0.0f,  // row 2
+          9.0f, 1.5f,   21.0f, 0.0f, 0.0f,   0.0f, 0.0f,   0.0f,  // row 3
+          0.0f, 111.0f, 0.0f,  0.0f, 15.75f, 0.0f, -10.5f, 0.0f,  // row 4
+          0.0f, 0.0f,   0.0f,  4.5f, 0.75f,  7.5f, -0.5f,  -7.0f  // row 5
+      });
+  for (size_t i = 0; i != reference.Size(); ++i) {
+    if (reference[i] == 0.0f)
+      BOOST_CHECK_LT(std::fabs(image[i]), 1e-5);
+    else
+      BOOST_CHECK_CLOSE_FRACTION(reference[i], image[i], 1e-5);
+  }
+}
+
+BOOST_AUTO_TEST_SUITE_END()
+
+}  // namespace schaapcommon::math