A working solution

integral_image.cpp

@@ -0,0 +1,67 @@
+#include "integral_image.h"
+
+#include <omp.h>
+
+
+namespace integral_image {
+
+
+Mat integral_image_serial(const Mat &image)
+{
+    if (image.cols == 0 || image.rows == 0)
+        return Mat();
+
+    Mat result = image.clone();
+
+    for (size_t row = 1; row < result.rows; ++row)
+        result[row][0] += result[row - 1][0];
+
+    for (size_t col = 1; col < result.cols; ++col)
+        result[0][col] += result[0][col - 1];
+
+    for (size_t row = 1; row < result.rows; ++row)
+        for (size_t col = 1; col < result.cols; ++col)
+            result[row][col] += result[row - 1][col] + result[row][col - 1] - result[row - 1][col - 1];
+
+    return result;
+}
+
+Mat integral_image_openmp(const Mat &image, int thread_number)
+{
+    if (image.cols == 0 || image.rows == 0)
+        return Mat();
+
+    if (0 != thread_number) {
+
+        omp_set_dynamic(0);
+        omp_set_num_threads(thread_number);
+    }
+
+    Mat result = image.clone();
+
+    #pragma omp parallel for
+    for (int row = 0; row < result.rows; ++row) {
+
+        for (int col = 1; col < result.cols; ++col) {
+
+            result[row][col] += result[row][col - 1];
+        }
+    }
+
+    //This loop is likely to have lots of cache misses that can probably be avoided by transposing data, processing data
+    //in a way similar to the previous loop, and than transposing data again.
+    //TODO: benchmark
+
+    #pragma omp parallel for
+    for (int col = 0; col < result.cols; ++col) {
+
+        for (int row = 1; row < result.rows; ++row) {
+
+            result[row][col] += result[row - 1][col];
+        }
+    }
+
+    return result;
+}
+
+}