Merge pull request #30000 from vespa-engine/balder/add-method-returning-possible-object-to-avoid-double-lookupv8.290.51

Balder/add method returning possible object to avoid double lookup

8 files changed, 72 insertions, 33 deletions

diff --git a/vespajlib/abi-spec.json b/vespajlib/abi-spec.json
index 4a65b00a6a4..6eb0aaf340e 100644
--- a/vespajlib/abi-spec.json
+++ b/vespajlib/abi-spec.json
@@ -912,6 +912,7 @@
       "public com.yahoo.tensor.DirectIndexedAddress directAddress()",
       "public varargs float getFloat(long[])",
       "public double get(com.yahoo.tensor.TensorAddress)",
+      "public java.lang.Double getAsDouble(com.yahoo.tensor.TensorAddress)",
       "public boolean has(com.yahoo.tensor.TensorAddress)",
       "public abstract double get(long)",
       "public abstract float getFloat(long)",
@@ -968,6 +969,7 @@
       "public int sizeAsInt()",
       "public double get(com.yahoo.tensor.TensorAddress)",
       "public boolean has(com.yahoo.tensor.TensorAddress)",
+      "public java.lang.Double getAsDouble(com.yahoo.tensor.TensorAddress)",
       "public java.util.Iterator cellIterator()",
       "public java.util.Iterator valueIterator()",
       "public java.util.Map cells()",
@@ -1048,6 +1050,7 @@
       "public com.yahoo.tensor.TensorType type()",
       "public long size()",
       "public double get(com.yahoo.tensor.TensorAddress)",
+      "public java.lang.Double getAsDouble(com.yahoo.tensor.TensorAddress)",
       "public boolean has(com.yahoo.tensor.TensorAddress)",
       "public java.util.Iterator cellIterator()",
       "public java.util.Iterator valueIterator()",
@@ -1174,6 +1177,7 @@
       "public int sizeAsInt()",
       "public abstract double get(com.yahoo.tensor.TensorAddress)",
       "public abstract boolean has(com.yahoo.tensor.TensorAddress)",
+      "public abstract java.lang.Double getAsDouble(com.yahoo.tensor.TensorAddress)",
       "public abstract java.util.Iterator cellIterator()",
       "public abstract java.util.Iterator valueIterator()",
       "public abstract java.util.Map cells()",
diff --git a/vespajlib/src/main/java/com/yahoo/tensor/IndexedTensor.java b/vespajlib/src/main/java/com/yahoo/tensor/IndexedTensor.java
index 5d384e0329b..f26174d9576 100644
--- a/vespajlib/src/main/java/com/yahoo/tensor/IndexedTensor.java
+++ b/vespajlib/src/main/java/com/yahoo/tensor/IndexedTensor.java
@@ -120,6 +120,17 @@ public abstract class IndexedTensor implements Tensor {
     }
 
     @Override
+    public Double getAsDouble(TensorAddress address) {
+        try {
+            long index = toValueIndex(address, dimensionSizes, type);
+            if (index < 0 || size() <= index) return null;
+            return get(index);
+        } catch (IllegalArgumentException e) {
+            return null;
+        }
+    }
+
+    @Override
     public boolean has(TensorAddress address) {
         try {
             long index = toValueIndex(address, dimensionSizes, type);
diff --git a/vespajlib/src/main/java/com/yahoo/tensor/MappedTensor.java b/vespajlib/src/main/java/com/yahoo/tensor/MappedTensor.java
index 5471ea65b97..3e0df5f2261 100644
--- a/vespajlib/src/main/java/com/yahoo/tensor/MappedTensor.java
+++ b/vespajlib/src/main/java/com/yahoo/tensor/MappedTensor.java
@@ -41,6 +41,9 @@ public class MappedTensor implements Tensor {
     public boolean has(TensorAddress address) { return cells.containsKey(address); }
 
     @Override
+    public Double getAsDouble(TensorAddress address) { return cells.get(address); }
+
+    @Override
     public Iterator<Cell> cellIterator() { return new CellIteratorAdaptor(cells.entrySet().iterator()); }
 
     @Override
diff --git a/vespajlib/src/main/java/com/yahoo/tensor/MixedTensor.java b/vespajlib/src/main/java/com/yahoo/tensor/MixedTensor.java
index cf6e737bf27..aab4a4e1297 100644
--- a/vespajlib/src/main/java/com/yahoo/tensor/MixedTensor.java
+++ b/vespajlib/src/main/java/com/yahoo/tensor/MixedTensor.java
@@ -83,6 +83,16 @@ public class MixedTensor implements Tensor {
     }
 
     @Override
+    public Double getAsDouble(TensorAddress address) {
+        var block = index.blockOf(address);
+        int denseOffset = index.denseOffsetOf(address);
+        if (block == null || denseOffset < 0 || denseOffset >= block.cells.length) {
+            return null;
+        }
+        return block.cells[denseOffset];
+    }
+
+    @Override
     public boolean has(TensorAddress address) {
         var block = index.blockOf(address);
         int denseOffset = index.denseOffsetOf(address);
diff --git a/vespajlib/src/main/java/com/yahoo/tensor/Tensor.java b/vespajlib/src/main/java/com/yahoo/tensor/Tensor.java
index cff17fdfd7c..d034ac551f8 100644
--- a/vespajlib/src/main/java/com/yahoo/tensor/Tensor.java
+++ b/vespajlib/src/main/java/com/yahoo/tensor/Tensor.java
@@ -90,6 +90,8 @@ public interface Tensor {
 
     /** Returns true if this cell exists */
     boolean has(TensorAddress address);
+    /** null = no value present. More efficient that if (t.has(key)) t.get(key) */
+    Double getAsDouble(TensorAddress address);
 
     /**
      * Returns the cell of this in some undefined order.
diff --git a/vespajlib/src/main/java/com/yahoo/tensor/functions/Join.java b/vespajlib/src/main/java/com/yahoo/tensor/functions/Join.java
index 7a336233de0..712e5528fc6 100644
--- a/vespajlib/src/main/java/com/yahoo/tensor/functions/Join.java
+++ b/vespajlib/src/main/java/com/yahoo/tensor/functions/Join.java
@@ -129,8 +129,9 @@ public class Join<NAMETYPE extends Name> extends PrimitiveTensorFunction<NAMETYP
         for (Iterator<Tensor.Cell> i = a.cellIterator(); i.hasNext(); ) {
             Map.Entry<TensorAddress, Double> aCell = i.next();
             var key = aCell.getKey();
-            if (b.has(key)) {
-                builder.cell(key, combinator.applyAsDouble(aCell.getValue(), b.get(key)));
+            Double bVal = b.getAsDouble(key);
+            if (bVal != null) {
+                builder.cell(key, combinator.applyAsDouble(aCell.getValue(), bVal));
             }
         }
         return builder.build();
@@ -206,11 +207,12 @@ public class Join<NAMETYPE extends Name> extends PrimitiveTensorFunction<NAMETYP
         for (Iterator<Tensor.Cell> i = superspace.cellIterator(); i.hasNext(); ) {
             Map.Entry<TensorAddress, Double> supercell = i.next();
             TensorAddress subaddress = mapAddressToSubspace(supercell.getKey(), subspaceIndexes);
-            if (subspace.has(subaddress)) {
-                double subspaceValue = subspace.get(subaddress);
+            Double subspaceValue = subspace.getAsDouble(subaddress);
+            if (subspaceValue != null) {
                 builder.cell(supercell.getKey(),
-                        reversedArgumentOrder ? combinator.applyAsDouble(supercell.getValue(), subspaceValue)
-                             : combinator.applyAsDouble(subspaceValue, supercell.getValue()));
+                        reversedArgumentOrder
+                                ? combinator.applyAsDouble(supercell.getValue(), subspaceValue)
+                                : combinator.applyAsDouble(subspaceValue, supercell.getValue()));
             }
         }
         return builder.build();
diff --git a/vespajlib/src/main/java/com/yahoo/tensor/functions/Merge.java b/vespajlib/src/main/java/com/yahoo/tensor/functions/Merge.java
index 59394785382..ddad91dc060 100644
--- a/vespajlib/src/main/java/com/yahoo/tensor/functions/Merge.java
+++ b/vespajlib/src/main/java/com/yahoo/tensor/functions/Merge.java
@@ -121,10 +121,11 @@ public class Merge<NAMETYPE extends Name> extends PrimitiveTensorFunction<NAMETY
         for (Iterator<Tensor.Cell> i = a.cellIterator(); i.hasNext(); ) {
             Map.Entry<TensorAddress, Double> aCell = i.next();
             var key = aCell.getKey();
-            if (! b.has(key)) {
+            Double bVal = b.getAsDouble(key);
+            if (bVal == null) {
                 builder.cell(key, aCell.getValue());
             } else if (combinator != null) {
-                builder.cell(key, combinator.applyAsDouble(aCell.getValue(), b.get(key)));
+                builder.cell(key, combinator.applyAsDouble(aCell.getValue(), bVal));
             }
         }
     }
diff --git a/vespajlib/src/test/java/com/yahoo/tensor/TensorFunctionBenchmark.java b/vespajlib/src/test/java/com/yahoo/tensor/TensorFunctionBenchmark.java
index 5c4d5f1ffcf..37c0fb87be0 100644
--- a/vespajlib/src/test/java/com/yahoo/tensor/TensorFunctionBenchmark.java
+++ b/vespajlib/src/test/java/com/yahoo/tensor/TensorFunctionBenchmark.java
@@ -107,45 +107,51 @@ public class TensorFunctionBenchmark {
     public static void main(String[] args) {
         double time = 0;
 
-        // ---------------- Mapped with extra space (sidesteps current special-case optimizations):
-        // 7.8 ms
-        time = new TensorFunctionBenchmark().benchmark(1000, vectors(100, 300, TensorType.Dimension.Type.mapped), TensorType.Dimension.Type.mapped, true);
-        System.out.printf("Mapped vectors, x space  time per join: %1$8.3f ms\n", time);
-        // 7.7 ms
-        time = new TensorFunctionBenchmark().benchmark(1000, matrix(100, 300, TensorType.Dimension.Type.mapped), TensorType.Dimension.Type.mapped, true);
-        System.out.printf("Mapped matrix, x space   time per join: %1$8.3f ms\n", time);
+        // ---------------- Indexed unbound:
+        time = new TensorFunctionBenchmark().benchmark(50000, vectors(100, 300, TensorType.Dimension.Type.indexedUnbound), TensorType.Dimension.Type.indexedUnbound, false);
+        System.out.printf("Indexed unbound vectors, time per join: %1$8.3f ms\n", time);
+        time = new TensorFunctionBenchmark().benchmark(50000, matrix(100, 300, TensorType.Dimension.Type.indexedUnbound), TensorType.Dimension.Type.indexedUnbound, false);
+        System.out.printf("Indexed unbound matrix,  time per join: %1$8.3f ms\n", time);
+
+        // ---------------- Indexed bound:
+        time = new TensorFunctionBenchmark().benchmark(50000, vectors(100, 300, TensorType.Dimension.Type.indexedBound), TensorType.Dimension.Type.indexedBound, false);
+        System.out.printf("Indexed bound vectors,   time per join: %1$8.3f ms\n", time);
+
+        time = new TensorFunctionBenchmark().benchmark(50000, matrix(100, 300, TensorType.Dimension.Type.indexedBound), TensorType.Dimension.Type.indexedBound, false);
+        System.out.printf("Indexed bound matrix,    time per join: %1$8.3f ms\n", time);
 
         // ---------------- Mapped:
-        // 2.1 ms
         time = new TensorFunctionBenchmark().benchmark(5000, vectors(100, 300, TensorType.Dimension.Type.mapped), TensorType.Dimension.Type.mapped, false);
         System.out.printf("Mapped vectors,          time per join: %1$8.3f ms\n", time);
-        // 7.0 ms
+
         time = new TensorFunctionBenchmark().benchmark(1000, matrix(100, 300, TensorType.Dimension.Type.mapped), TensorType.Dimension.Type.mapped, false);
         System.out.printf("Mapped matrix,           time per join: %1$8.3f ms\n", time);
 
         // ---------------- Indexed (unbound) with extra space (sidesteps current special-case optimizations):
-        // 14.5 ms
         time = new TensorFunctionBenchmark().benchmark(500, vectors(100, 300, TensorType.Dimension.Type.indexedUnbound), TensorType.Dimension.Type.indexedUnbound, true);
         System.out.printf("Indexed vectors, x space time per join: %1$8.3f ms\n", time);
-        // 8.9 ms
         time = new TensorFunctionBenchmark().benchmark(500, matrix(100, 300, TensorType.Dimension.Type.indexedUnbound), TensorType.Dimension.Type.indexedUnbound, true);
         System.out.printf("Indexed matrix, x space  time per join: %1$8.3f ms\n", time);
 
-        // ---------------- Indexed unbound:
-        // 0.14 ms
-        time = new TensorFunctionBenchmark().benchmark(50000, vectors(100, 300, TensorType.Dimension.Type.indexedUnbound), TensorType.Dimension.Type.indexedUnbound, false);
-        System.out.printf("Indexed unbound vectors, time per join: %1$8.3f ms\n", time);
-        // 0.44 ms
-        time = new TensorFunctionBenchmark().benchmark(50000, matrix(100, 300, TensorType.Dimension.Type.indexedUnbound), TensorType.Dimension.Type.indexedUnbound, false);
-        System.out.printf("Indexed unbound matrix,  time per join: %1$8.3f ms\n", time);
+        // ---------------- Mapped with extra space (sidesteps current special-case optimizations):
+        time = new TensorFunctionBenchmark().benchmark(1000, vectors(100, 300, TensorType.Dimension.Type.mapped), TensorType.Dimension.Type.mapped, true);
+        System.out.printf("Mapped vectors, x space  time per join: %1$8.3f ms\n", time);
+        time = new TensorFunctionBenchmark().benchmark(1000, matrix(100, 300, TensorType.Dimension.Type.mapped), TensorType.Dimension.Type.mapped, true);
+        System.out.printf("Mapped matrix, x space   time per join: %1$8.3f ms\n", time);
+
+        /** 2.4Ghz Intel Core i9, Macbook Pro 2019
+         * Indexed unbound vectors, time per join:    0,067 ms
+         * Indexed unbound matrix,  time per join:    0,107 ms
+         * Indexed bound vectors,   time per join:    0,068 ms
+         * Indexed bound matrix,    time per join:    0,105 ms
+         * Mapped vectors,          time per join:    1,780 ms
+         * Mapped matrix,           time per join:    5,339 ms
+         * Indexed vectors, x space time per join:    6,398 ms
+         * Indexed matrix, x space  time per join:    3,220 ms
+         * Mapped vectors, x space  time per join:   13,026 ms
+         * Mapped matrix, x space   time per join:   28,259 ms
+         */
 
-        // ---------------- Indexed bound:
-        // 0.32 ms
-        time = new TensorFunctionBenchmark().benchmark(50000, vectors(100, 300, TensorType.Dimension.Type.indexedBound), TensorType.Dimension.Type.indexedBound, false);
-        System.out.printf("Indexed bound vectors,   time per join: %1$8.3f ms\n", time);
-        // 0.44 ms
-        time = new TensorFunctionBenchmark().benchmark(50000, matrix(100, 300, TensorType.Dimension.Type.indexedBound), TensorType.Dimension.Type.indexedBound, false);
-        System.out.printf("Indexed bound matrix,    time per join: %1$8.3f ms\n", time);
     }
 
 }