[8.x] Add support for bitwise inner-product in painless (elastic#116082…

…) (elastic#116285)

* Add support for bitwise inner-product in painless (elastic#116082)

This adds bitwise inner product to painless. 

The idea here is:

 - For two bit arrays, which we determine to be a byte array whose dimensions match `dense_vector.dim/8`, we simply return bitwise `&`
 - For a stored bit array (remember, with `dense_vector.dim/8` bytes), sum up the provided byte or float array using the bit array as a mask.

This is effectively supporting asynchronous quantization. A prime
example of how this works is:
https://github.com/cohere-ai/BinaryVectorDB

Basically, you do your initial search against the binary space and then
rerank with a differently quantized vector allowing for more information
without additional storage space. 

closes:  elastic#111232

* removing unnecessary task adjustment

---------

Co-authored-by: Elastic Machine <[email protected]>

docs/changelog/116082.yaml

@@ -0,0 +1,5 @@
+pr: 116082
+summary: Add support for bitwise inner-product in painless
+area: Vector Search
+type: enhancement
+issues: []

docs/reference/vectors/vector-functions.asciidoc

@@ -16,7 +16,7 @@ This is the list of available vector functions and vector access methods:
 6. <<vector-functions-accessing-vectors,`doc[<field>].vectorValue`>> – returns a vector's value as an array of floats
 7. <<vector-functions-accessing-vectors,`doc[<field>].magnitude`>> – returns a vector's magnitude
 
-NOTE: The `cosineSimilarity` and `dotProduct` functions are not supported for `bit` vectors.
+NOTE: The `cosineSimilarity` function is not supported for `bit` vectors.
 
 NOTE: The recommended way to access dense vectors is through the
 `cosineSimilarity`, `dotProduct`, `l1norm` or `l2norm` functions. Please note
@@ -332,6 +332,92 @@ When using `bit` vectors, not all the vector functions are available. The suppor
 * <<vector-functions-hamming,`hamming`>> – calculates Hamming distance, the sum of the bitwise XOR of the two vectors
 * <<vector-functions-l1,`l1norm`>> – calculates L^1^ distance, this is simply the `hamming` distance
 * <<vector-functions-l2,`l2norm`>> - calculates L^2^ distance, this is the square root of the `hamming` distance
+* <<vector-functions-dot-product,`dotProduct`>> – calculates dot product. When comparing two `bit` vectors,
+this is the sum of the bitwise AND of the two vectors. If providing `float[]` or `byte[]`, who has `dims` number of elements, as a query vector, the `dotProduct` is
+the sum of the floating point values using the stored `bit` vector as a mask.
 
-Currently, the `cosineSimilarity` and `dotProduct` functions are not supported for `bit` vectors.
+Here is an example of using dot-product with bit vectors.
+
+[source,console]
+--------------------------------------------------
+PUT my-index-bit-vectors
+{
+  "mappings": {
+    "properties": {
+      "my_dense_vector": {
+        "type": "dense_vector",
+        "index": false,
+        "element_type": "bit",
+        "dims": 40 <1>
+      }
+    }
+  }
+}
+
+PUT my-index-bit-vectors/_doc/1
+{
+  "my_dense_vector": [8, 5, -15, 1, -7] <2>
+}
+
+PUT my-index-bit-vectors/_doc/2
+{
+  "my_dense_vector": [-1, 115, -3, 4, -128]
+}
+
+PUT my-index-bit-vectors/_doc/3
+{
+  "my_dense_vector": [2, 18, -5, 0, -124]
+}
+
+POST my-index-bit-vectors/_refresh
+--------------------------------------------------
+// TEST[continued]
+<1> The number of dimensions or bits for the `bit` vector.
+<2> This vector represents 5 bytes, or `5 * 8 = 40` bits, which equals the configured dimensions
+
+[source,console]
+--------------------------------------------------
+GET my-index-bit-vectors/_search
+{
+  "query": {
+    "script_score": {
+      "query" : {
+        "match_all": {}
+      },
+      "script": {
+        "source": "dotProduct(params.query_vector, 'my_dense_vector')",
+        "params": {
+          "query_vector": [8, 5, -15, 1, -7] <1>
+        }
+      }
+    }
+  }
+}
+--------------------------------------------------
+// TEST[continued]
+<1> This vector is 40 bits, and thus will compute a bitwise `&` operation with the stored vectors.
+
+[source,console]
+--------------------------------------------------
+GET my-index-bit-vectors/_search
+{
+  "query": {
+    "script_score": {
+      "query" : {
+        "match_all": {}
+      },
+      "script": {
+        "source": "dotProduct(params.query_vector, 'my_dense_vector')",
+        "params": {
+          "query_vector": [0.23, 1.45, 3.67, 4.89, -0.56, 2.34, 3.21, 1.78, -2.45, 0.98, -0.12, 3.45, 4.56, 2.78, 1.23, 0.67, 3.89, 4.12, -2.34, 1.56, 0.78, 3.21, 4.12, 2.45, -1.67, 0.34, -3.45, 4.56, -2.78, 1.23, -0.67, 3.89, -4.34, 2.12, -1.56, 0.78, -3.21, 4.45, 2.12, 1.67] <1>
+        }
+      }
+    }
+  }
+}
+--------------------------------------------------
+// TEST[continued]
+<1> This vector is 40 individual dimensions, and thus will sum the floating point values using the stored `bit` vector as a mask.
+
+Currently, the `cosineSimilarity` function is not supported for `bit` vectors.
 

libs/simdvec/src/main/java/org/elasticsearch/simdvec/ESVectorUtil.java

@@ -9,13 +9,36 @@
 
 package org.elasticsearch.simdvec;
 
+import org.apache.lucene.util.BitUtil;
+import org.apache.lucene.util.Constants;
 import org.elasticsearch.simdvec.internal.vectorization.ESVectorUtilSupport;
 import org.elasticsearch.simdvec.internal.vectorization.ESVectorizationProvider;
 
+import java.lang.invoke.MethodHandle;
+import java.lang.invoke.MethodHandles;
+import java.lang.invoke.MethodType;
+
 import static org.elasticsearch.simdvec.internal.vectorization.ESVectorUtilSupport.B_QUERY;
 
 public class ESVectorUtil {
 
+    private static final MethodHandle BIT_COUNT_MH;
+    static {
+        try {
+            // For xorBitCount we stride over the values as either 64-bits (long) or 32-bits (int) at a time.
+            // On ARM Long::bitCount is not vectorized, and therefore produces less than optimal code, when
+            // compared to Integer::bitCount. While Long::bitCount is optimal on x64. See
+            // https://bugs.openjdk.org/browse/JDK-8336000
+            BIT_COUNT_MH = Constants.OS_ARCH.equals("aarch64")
+                ? MethodHandles.lookup()
+                    .findStatic(ESVectorUtil.class, "andBitCountInt", MethodType.methodType(int.class, byte[].class, byte[].class))
+                : MethodHandles.lookup()
+                    .findStatic(ESVectorUtil.class, "andBitCountLong", MethodType.methodType(int.class, byte[].class, byte[].class));
+        } catch (NoSuchMethodException | IllegalAccessException e) {
+            throw new AssertionError(e);
+        }
+    }
+
     private static final ESVectorUtilSupport IMPL = ESVectorizationProvider.getInstance().getVectorUtilSupport();
 
     public static long ipByteBinByte(byte[] q, byte[] d) {
@@ -24,4 +47,103 @@ public static long ipByteBinByte(byte[] q, byte[] d) {
         }
         return IMPL.ipByteBinByte(q, d);
     }
+
+    /**
+     * Compute the inner product of two vectors, where the query vector is a byte vector and the document vector is a bit vector.
+     * This will return the sum of the query vector values using the document vector as a mask.
+     * @param q the query vector
+     * @param d the document vector
+     * @return the inner product of the two vectors
+     */
+    public static int ipByteBit(byte[] q, byte[] d) {
+        if (q.length != d.length * Byte.SIZE) {
+            throw new IllegalArgumentException("vector dimensions incompatible: " + q.length + "!= " + Byte.SIZE + " x " + d.length);
+        }
+        int result = 0;
+        // now combine the two vectors, summing the byte dimensions where the bit in d is `1`
+        for (int i = 0; i < d.length; i++) {
+            byte mask = d[i];
+            for (int j = 0; j < Byte.SIZE; j++) {
+                if ((mask & (1 << j)) != 0) {
+                    result += q[i * Byte.SIZE + j];
+                }
+            }
+        }
+        return result;
+    }
+
+    /**
+     * Compute the inner product of two vectors, where the query vector is a float vector and the document vector is a bit vector.
+     * This will return the sum of the query vector values using the document vector as a mask.
+     * @param q the query vector
+     * @param d the document vector
+     * @return the inner product of the two vectors
+     */
+    public static float ipFloatBit(float[] q, byte[] d) {
+        if (q.length != d.length * Byte.SIZE) {
+            throw new IllegalArgumentException("vector dimensions incompatible: " + q.length + "!= " + Byte.SIZE + " x " + d.length);
+        }
+        float result = 0;
+        for (int i = 0; i < d.length; i++) {
+            byte mask = d[i];
+            for (int j = 0; j < Byte.SIZE; j++) {
+                if ((mask & (1 << j)) != 0) {
+                    result += q[i * Byte.SIZE + j];
+                }
+            }
+        }
+        return result;
+    }
+
+    /**
+     * AND bit count computed over signed bytes.
+     * Copied from Lucene's XOR implementation
+     * @param a bytes containing a vector
+     * @param b bytes containing another vector, of the same dimension
+     * @return the value of the AND bit count of the two vectors
+     */
+    public static int andBitCount(byte[] a, byte[] b) {
+        if (a.length != b.length) {
+            throw new IllegalArgumentException("vector dimensions differ: " + a.length + "!=" + b.length);
+        }
+        try {
+            return (int) BIT_COUNT_MH.invokeExact(a, b);
+        } catch (Throwable e) {
+            if (e instanceof Error err) {
+                throw err;
+            } else if (e instanceof RuntimeException re) {
+                throw re;
+            } else {
+                throw new RuntimeException(e);
+            }
+        }
+    }
+
+    /** AND bit count striding over 4 bytes at a time. */
+    static int andBitCountInt(byte[] a, byte[] b) {
+        int distance = 0, i = 0;
+        // limit to number of int values in the array iterating by int byte views
+        for (final int upperBound = a.length & -Integer.BYTES; i < upperBound; i += Integer.BYTES) {
+            distance += Integer.bitCount((int) BitUtil.VH_NATIVE_INT.get(a, i) & (int) BitUtil.VH_NATIVE_INT.get(b, i));
+        }
+        // tail:
+        for (; i < a.length; i++) {
+            distance += Integer.bitCount((a[i] & b[i]) & 0xFF);
+        }
+        return distance;
+    }
+
+    /** AND bit count striding over 8 bytes at a time**/
+    static int andBitCountLong(byte[] a, byte[] b) {
+        int distance = 0, i = 0;
+        // limit to number of long values in the array iterating by long byte views
+        for (final int upperBound = a.length & -Long.BYTES; i < upperBound; i += Long.BYTES) {
+            distance += Long.bitCount((long) BitUtil.VH_NATIVE_LONG.get(a, i) & (long) BitUtil.VH_NATIVE_LONG.get(b, i));
+        }
+        // tail:
+        for (; i < a.length; i++) {
+            distance += Integer.bitCount((a[i] & b[i]) & 0xFF);
+        }
+        return distance;
+    }
 }

libs/simdvec/src/test/java/org/elasticsearch/simdvec/ESVectorUtilTests.java

@@ -21,6 +21,10 @@ public class ESVectorUtilTests extends BaseVectorizationTests {
     static final ESVectorizationProvider defaultedProvider = BaseVectorizationTests.defaultProvider();
     static final ESVectorizationProvider defOrPanamaProvider = BaseVectorizationTests.maybePanamaProvider();
 
+    public void testBitAndCount() {
+        testBasicBitAndImpl(ESVectorUtil::andBitCountLong);
+    }
+
     public void testIpByteBinInvariants() {
         int iterations = atLeast(10);
         for (int i = 0; i < iterations; i++) {
@@ -41,6 +45,23 @@ interface IpByteBin {
         long apply(byte[] q, byte[] d);
     }
 
+    interface BitOps {
+        long apply(byte[] q, byte[] d);
+    }
+
+    void testBasicBitAndImpl(BitOps bitAnd) {
+        assertEquals(0, bitAnd.apply(new byte[] { 0 }, new byte[] { 0 }));
+        assertEquals(0, bitAnd.apply(new byte[] { 1 }, new byte[] { 0 }));
+        assertEquals(0, bitAnd.apply(new byte[] { 0 }, new byte[] { 1 }));
+        assertEquals(1, bitAnd.apply(new byte[] { 1 }, new byte[] { 1 }));
+        byte[] a = new byte[31];
+        byte[] b = new byte[31];
+        random().nextBytes(a);
+        random().nextBytes(b);
+        int expected = scalarBitAnd(a, b);
+        assertEquals(expected, bitAnd.apply(a, b));
+    }
+
     void testBasicIpByteBinImpl(IpByteBin ipByteBinFunc) {
         assertEquals(15L, ipByteBinFunc.apply(new byte[] { 1, 1, 1, 1 }, new byte[] { 1 }));
         assertEquals(30L, ipByteBinFunc.apply(new byte[] { 1, 2, 1, 2, 1, 2, 1, 2 }, new byte[] { 1, 2 }));
@@ -115,6 +136,14 @@ static int scalarIpByteBin(byte[] q, byte[] d) {
         return res;
     }
 
+    static int scalarBitAnd(byte[] a, byte[] b) {
+        int res = 0;
+        for (int i = 0; i < a.length; i++) {
+            res += Integer.bitCount((a[i] & b[i]) & 0xFF);
+        }
+        return res;
+    }
+
     public static int popcount(byte[] a, int aOffset, byte[] b, int length) {
         int res = 0;
         for (int j = 0; j < length; j++) {

modules/lang-painless/build.gradle

@@ -53,7 +53,7 @@ tasks.named("dependencyLicenses").configure {
 restResources {
     restApi {
         include '_common', 'cluster', 'nodes', 'indices', 'index', 'search', 'get', 'bulk', 'update',
-                'scripts_painless_execute', 'put_script', 'delete_script'
+                'scripts_painless_execute', 'put_script', 'delete_script', 'capabilities'
     }
 }