improve convolve speed of skpdiff using direct pointers

experimental/skpdiff/SkPMetric.cpp

Index: experimental/skpdiff/SkPMetric.cpp
diff --git a/experimental/skpdiff/SkPMetric.cpp b/experimental/skpdiff/SkPMetric.cpp
index a9628539b556584a888d2cbe1443692261aeba4c..1bf448fd0bc9c0e6af2e566a0d1110155a57c780 100644
--- a/experimental/skpdiff/SkPMetric.cpp
+++ b/experimental/skpdiff/SkPMetric.cpp
@@ -38,6 +38,10 @@ struct Image2D {
         *pixel = image[y * width + x];
     }
 
+    T * getRow(int y) const {

[djsollen, 2013/07/12 18:07:10]

T* no space

+        return &image[y * width];
+    }
+
     void writePixel(int x, int y, const T& pixel) {
         SkASSERT(x >= 0);
         SkASSERT(y >= 0);
@@ -196,10 +200,19 @@ static void convolve(const ImageL* imageL,
                      ImageL* outImageL) {
     SkASSERT(imageL->width == outImageL->width);
     SkASSERT(imageL->height == outImageL->height);
+
+    // Keep track of what rows are being operated on for quick access.
+    int filterCount = radius * 2 + 1;
+    float * rowPtrs[8];

[djsollen, 2013/07/12 18:07:10]

float* here and on 210

+    for (int y = radius; y < filterCount; y++) {
+        rowPtrs[y] = imageL->getRow(y - radius);

[djsollen, 2013/07/12 18:07:10]

how can you ensure that y is never greater than 7?

[Zach Reizner, 2013/07/12 18:51:43]

convolve is static and only ever called from pmetric. The filter always has a
size of 5. Thinking about it now, I may as well keep the filter here and avoid
future misuse of convolve.

+    }
+    float * writeRow = outImageL->getRow(0);
+
+
     for (int y = 0; y < imageL->height; y++) {
         for (int x = 0; x < imageL->width; x++) {
             float lSum = 0.0f;
-            float l;
             for (int xx = -radius; xx <= radius; xx++) {
                 int nx = x;
                 int ny = y;
@@ -224,12 +237,18 @@ static void convolve(const ImageL* imageL,
                     }
                 }
 
-                imageL->readPixel(nx, ny, &l);
                 float weight = matrix[xx + radius];
-                lSum += l * weight;
+                lSum += rowPtrs[ny - y + radius][nx] * weight;
             }
-            outImageL->writePixel(x, y, lSum);
+            writeRow[x] = lSum;
+        }
+        // As we move down, scroll the row pointers down with us
+        for (int y = 0; y < filterCount - 1; y++)
+        {
+            rowPtrs[y] = rowPtrs[y + 1];
         }
+        rowPtrs[filterCount - 1] += imageL->width;
+        writeRow += imageL->width;
     }
 }