Optimize land lookup on blendmaps sampling

Minimize number of getLand calls by grouping samples per cell.

apps/openmw_test_suite/esmterrain/testgridsampling.cpp

@@ -430,8 +430,8 @@ namespace ESMTerrain
             EXPECT_THAT(samples,
                 ElementsAre( //
                     CellSample{ .mCellX = -1, .mCellY = 0, .mSrcRow = 7, .mSrcCol = 0, .mDstRow = 0, .mDstCol = 0 },
-                    CellSample{ .mCellX = 0, .mCellY = 0, .mSrcRow = 0, .mSrcCol = 0, .mDstRow = 1, .mDstCol = 0 },
                     CellSample{ .mCellX = -1, .mCellY = 0, .mSrcRow = 7, .mSrcCol = 1, .mDstRow = 0, .mDstCol = 1 },
+                    CellSample{ .mCellX = 0, .mCellY = 0, .mSrcRow = 0, .mSrcCol = 0, .mDstRow = 1, .mDstCol = 0 },
                     CellSample{ .mCellX = 0, .mCellY = 0, .mSrcRow = 0, .mSrcCol = 1, .mDstRow = 1, .mDstCol = 1 }));
         }
 
@@ -468,23 +468,23 @@ namespace ESMTerrain
                     CellSample{ .mCellX = 0, .mCellY = -2, .mSrcRow = 0, .mSrcCol = 1, .mDstRow = 4, .mDstCol = 0 },
                     CellSample{ .mCellX = -2, .mCellY = -1, .mSrcRow = 0, .mSrcCol = 0, .mDstRow = 0, .mDstCol = 1 },
                     CellSample{ .mCellX = -2, .mCellY = -1, .mSrcRow = 1, .mSrcCol = 0, .mDstRow = 1, .mDstCol = 1 },
-                    CellSample{ .mCellX = -1, .mCellY = -1, .mSrcRow = 0, .mSrcCol = 0, .mDstRow = 2, .mDstCol = 1 },
-                    CellSample{ .mCellX = -1, .mCellY = -1, .mSrcRow = 1, .mSrcCol = 0, .mDstRow = 3, .mDstCol = 1 },
-                    CellSample{ .mCellX = 0, .mCellY = -1, .mSrcRow = 0, .mSrcCol = 0, .mDstRow = 4, .mDstCol = 1 },
                     CellSample{ .mCellX = -2, .mCellY = -1, .mSrcRow = 0, .mSrcCol = 1, .mDstRow = 0, .mDstCol = 2 },
                     CellSample{ .mCellX = -2, .mCellY = -1, .mSrcRow = 1, .mSrcCol = 1, .mDstRow = 1, .mDstCol = 2 },
+                    CellSample{ .mCellX = -1, .mCellY = -1, .mSrcRow = 0, .mSrcCol = 0, .mDstRow = 2, .mDstCol = 1 },
+                    CellSample{ .mCellX = -1, .mCellY = -1, .mSrcRow = 1, .mSrcCol = 0, .mDstRow = 3, .mDstCol = 1 },
                     CellSample{ .mCellX = -1, .mCellY = -1, .mSrcRow = 0, .mSrcCol = 1, .mDstRow = 2, .mDstCol = 2 },
                     CellSample{ .mCellX = -1, .mCellY = -1, .mSrcRow = 1, .mSrcCol = 1, .mDstRow = 3, .mDstCol = 2 },
+                    CellSample{ .mCellX = 0, .mCellY = -1, .mSrcRow = 0, .mSrcCol = 0, .mDstRow = 4, .mDstCol = 1 },
                     CellSample{ .mCellX = 0, .mCellY = -1, .mSrcRow = 0, .mSrcCol = 1, .mDstRow = 4, .mDstCol = 2 },
                     CellSample{ .mCellX = -2, .mCellY = 0, .mSrcRow = 0, .mSrcCol = 0, .mDstRow = 0, .mDstCol = 3 },
                     CellSample{ .mCellX = -2, .mCellY = 0, .mSrcRow = 1, .mSrcCol = 0, .mDstRow = 1, .mDstCol = 3 },
-                    CellSample{ .mCellX = -1, .mCellY = 0, .mSrcRow = 0, .mSrcCol = 0, .mDstRow = 2, .mDstCol = 3 },
-                    CellSample{ .mCellX = -1, .mCellY = 0, .mSrcRow = 1, .mSrcCol = 0, .mDstRow = 3, .mDstCol = 3 },
-                    CellSample{ .mCellX = 0, .mCellY = 0, .mSrcRow = 0, .mSrcCol = 0, .mDstRow = 4, .mDstCol = 3 },
                     CellSample{ .mCellX = -2, .mCellY = 0, .mSrcRow = 0, .mSrcCol = 1, .mDstRow = 0, .mDstCol = 4 },
                     CellSample{ .mCellX = -2, .mCellY = 0, .mSrcRow = 1, .mSrcCol = 1, .mDstRow = 1, .mDstCol = 4 },
+                    CellSample{ .mCellX = -1, .mCellY = 0, .mSrcRow = 0, .mSrcCol = 0, .mDstRow = 2, .mDstCol = 3 },
+                    CellSample{ .mCellX = -1, .mCellY = 0, .mSrcRow = 1, .mSrcCol = 0, .mDstRow = 3, .mDstCol = 3 },
                     CellSample{ .mCellX = -1, .mCellY = 0, .mSrcRow = 0, .mSrcCol = 1, .mDstRow = 2, .mDstCol = 4 },
                     CellSample{ .mCellX = -1, .mCellY = 0, .mSrcRow = 1, .mSrcCol = 1, .mDstRow = 3, .mDstCol = 4 },
+                    CellSample{ .mCellX = 0, .mCellY = 0, .mSrcRow = 0, .mSrcCol = 0, .mDstRow = 4, .mDstCol = 3 },
                     CellSample{ .mCellX = 0, .mCellY = 0, .mSrcRow = 0, .mSrcCol = 1, .mDstRow = 4, .mDstCol = 4 }));
         }
     }

components/esmterrain/gridsampling.hpp

@@ -132,26 +132,16 @@ namespace ESMTerrain
         return static_cast<int>(textureSize * size) + 1;
     }
 
-    inline void adjustTextureCoordinates(int textureSize, int& cellX, int& cellY, int& x, int& y)
+    inline std::pair<std::size_t, std::size_t> getBlendmapLocalRange(
+        int cell, int minCell, int textureSize, int min, int max)
     {
-        --x;
+        const int cellMin = (cell - minCell) * textureSize;
-        if (x < 0)
+        const int cellMax = cellMin + textureSize - 1;
-        {
+        const std::size_t begin = min > cellMin ? min - cellMin : 0;
-            --cellX;
+        const std::size_t end = cellMax < max ? textureSize : max - cellMin + 1;
-            x += textureSize;
+        assert(begin < end);
-        }
+        assert(static_cast<int>(end) <= textureSize);
-
+        return { begin, end };
-        while (x >= textureSize)
-        {
-            ++cellX;
-            x -= textureSize;
-        }
-
-        while (y >= textureSize)
-        {
-            ++cellY;
-            y -= textureSize;
-        }
     }
 
     template <class F>
@@ -163,32 +153,52 @@ namespace ESMTerrain
         if (textureSize <= 0)
             throw std::invalid_argument("Invalid texture size for blendmap sampling: " + std::to_string(textureSize));
 
-        const int beginCellX = static_cast<int>(std::floor(minX));
+        int minCellX = static_cast<int>(std::floor(minX));
-        const int beginCellY = static_cast<int>(std::floor(minY));
+        const int minCellY = static_cast<int>(std::floor(minY));
-        const int beginRow = static_cast<int>((minX - beginCellX) * (textureSize + 1));
+        int minRow = static_cast<int>((minX - minCellX) * (textureSize + 1)) - 1;
-        const int beginCol = static_cast<int>((minY - beginCellY) * (textureSize + 1));
+        const int minCol = static_cast<int>((minY - minCellY) * (textureSize + 1));
-        const int blendmapSize = getBlendmapSize(size, textureSize);
 
-        for (int y = 0; y < blendmapSize; y++)
+        if (minRow < 0)
         {
-            for (int x = 0; x < blendmapSize; x++)
+            --minCellX;
+            minRow += textureSize;
+        }
+
+        const int maxRow = minRow + static_cast<int>(textureSize * size);
+        const int maxCol = minCol + static_cast<int>(textureSize * size);
+        const int maxCellX = minCellX + maxRow / textureSize;
+        const int maxCellY = minCellY + maxCol / textureSize;
+
+        std::size_t baseDstCol = 0;
+        for (int cellY = minCellY; cellY <= maxCellY; ++cellY)
         {
-                int cellX = beginCellX;
+            std::size_t baseDstRow = 0;
-                int cellY = beginCellY;
-                int srcX = x + beginRow;
-                int srcY = y + beginCol;
 
-                adjustTextureCoordinates(textureSize, cellX, cellY, srcX, srcY);
+            const auto [localBeginCol, localEndCol]
+                = getBlendmapLocalRange(cellY, minCellY, textureSize, minCol, maxCol);
+            const std::size_t colCount = localEndCol - localBeginCol;
 
+            for (int cellX = minCellX; cellX <= maxCellX; ++cellX)
+            {
+                const auto [localBeginRow, localEndRow]
+                    = getBlendmapLocalRange(cellX, minCellX, textureSize, minRow, maxRow);
+                const std::size_t rowCount = localEndRow - localBeginRow;
+
+                for (std::size_t col = 0; col < colCount; ++col)
+                    for (std::size_t row = 0; row < rowCount; ++row)
                         f(CellSample{
                             .mCellX = cellX,
                             .mCellY = cellY,
-                    .mSrcRow = static_cast<std::size_t>(srcX),
+                            .mSrcRow = localBeginRow + row,
-                    .mSrcCol = static_cast<std::size_t>(srcY),
+                            .mSrcCol = localBeginCol + col,
-                    .mDstRow = static_cast<std::size_t>(x),
+                            .mDstRow = baseDstRow + row,
-                    .mDstCol = static_cast<std::size_t>(y),
+                            .mDstCol = baseDstCol + col,
                         });
+
+                baseDstRow += rowCount;
             }
+
+            baseDstCol += colCount;
         }
     }
 }