add "unpack_bits_from_int8" function

10 files changed, 361 insertions, 1 deletions

diff --git a/config-model/src/test/derived/tensor/attributes.cfg b/config-model/src/test/derived/tensor/attributes.cfg
index 752e236bb19..d050d553416 100644
--- a/config-model/src/test/derived/tensor/attributes.cfg
+++ b/config-model/src/test/derived/tensor/attributes.cfg
@@ -143,3 +143,32 @@ attribute[].index.hnsw.enabled false
 attribute[].index.hnsw.maxlinkspernode 16
 attribute[].index.hnsw.neighborstoexploreatinsert 200
 attribute[].index.hnsw.multithreadedindexing true
+attribute[].name "f7"
+attribute[].datatype TENSOR
+attribute[].collectiontype SINGLE
+attribute[].dictionary.type BTREE
+attribute[].dictionary.match UNCASED
+attribute[].match UNCASED
+attribute[].removeifzero false
+attribute[].createifnonexistent false
+attribute[].fastsearch false
+attribute[].paged false
+attribute[].ismutable false
+attribute[].sortascending true
+attribute[].sortfunction UCA
+attribute[].sortstrength PRIMARY
+attribute[].sortlocale ""
+attribute[].enableonlybitvector false
+attribute[].fastaccess false
+attribute[].arity 8
+attribute[].lowerbound -9223372036854775808
+attribute[].upperbound 9223372036854775807
+attribute[].densepostinglistthreshold 0.4
+attribute[].tensortype "tensor<int8>(p{},x[5])"
+attribute[].imported false
+attribute[].maxuncommittedmemory 77777
+attribute[].distancemetric EUCLIDEAN
+attribute[].index.hnsw.enabled false
+attribute[].index.hnsw.maxlinkspernode 16
+attribute[].index.hnsw.neighborstoexploreatinsert 200
+attribute[].index.hnsw.multithreadedindexing true
diff --git a/config-model/src/test/derived/tensor/documentmanager.cfg b/config-model/src/test/derived/tensor/documentmanager.cfg
index bae2db34040..5676f60ef46 100644
--- a/config-model/src/test/derived/tensor/documentmanager.cfg
+++ b/config-model/src/test/derived/tensor/documentmanager.cfg
@@ -49,6 +49,7 @@ doctype[].fieldsets{[document]}.fields[] "f3"
 doctype[].fieldsets{[document]}.fields[] "f4"
 doctype[].fieldsets{[document]}.fields[] "f5"
 doctype[].fieldsets{[document]}.fields[] "f6"
+doctype[].fieldsets{[document]}.fields[] "f7"
 doctype[].tensortype[].idx 10017
 doctype[].tensortype[].detailedtype "tensor(x[3])"
 doctype[].tensortype[].idx 10018
@@ -59,6 +60,8 @@ doctype[].tensortype[].idx 10020
 doctype[].tensortype[].detailedtype "tensor(x[10],y[10])"
 doctype[].tensortype[].idx 10021
 doctype[].tensortype[].detailedtype "tensor<float>(x[10])"
+doctype[].tensortype[].idx 10022
+doctype[].tensortype[].detailedtype "tensor<int8>(p{},x[5])"
 doctype[].structtype[].idx 10016
 doctype[].structtype[].name "tensor.header"
 doctype[].structtype[].field[].name "f1"
@@ -79,3 +82,6 @@ doctype[].structtype[].field[].type 10021
 doctype[].structtype[].field[].name "f6"
 doctype[].structtype[].field[].internalid 596352344
 doctype[].structtype[].field[].type 10005
+doctype[].structtype[].field[].name "f7"
+doctype[].structtype[].field[].internalid 981728328
+doctype[].structtype[].field[].type 10022
diff --git a/config-model/src/test/derived/tensor/documenttypes.cfg b/config-model/src/test/derived/tensor/documenttypes.cfg
index d10ecd37c8f..d069b3764ba 100644
--- a/config-model/src/test/derived/tensor/documenttypes.cfg
+++ b/config-model/src/test/derived/tensor/documenttypes.cfg
@@ -54,6 +54,7 @@ doctype[].fieldsets{[document]}.fields[] "f3"
 doctype[].fieldsets{[document]}.fields[] "f4"
 doctype[].fieldsets{[document]}.fields[] "f5"
 doctype[].fieldsets{[document]}.fields[] "f6"
+doctype[].fieldsets{[document]}.fields[] "f7"
 doctype[].tensortype[].idx 10017
 doctype[].tensortype[].detailedtype "tensor(x[3])"
 doctype[].tensortype[].idx 10018
@@ -64,6 +65,8 @@ doctype[].tensortype[].idx 10020
 doctype[].tensortype[].detailedtype "tensor(x[10],y[10])"
 doctype[].tensortype[].idx 10021
 doctype[].tensortype[].detailedtype "tensor<float>(x[10])"
+doctype[].tensortype[].idx 10022
+doctype[].tensortype[].detailedtype "tensor<int8>(p{},x[5])"
 doctype[].structtype[].idx 10016
 doctype[].structtype[].name "tensor.header"
 doctype[].structtype[].field[].name "f1"
@@ -84,4 +87,7 @@ doctype[].structtype[].field[].type 10021
 doctype[].structtype[].field[].name "f6"
 doctype[].structtype[].field[].internalid 596352344
 doctype[].structtype[].field[].type 10005
+doctype[].structtype[].field[].name "f7"
+doctype[].structtype[].field[].internalid 981728328
+doctype[].structtype[].field[].type 10022
 doctype[].structtype[].internalid 2125927172
diff --git a/config-model/src/test/derived/tensor/index-info.cfg b/config-model/src/test/derived/tensor/index-info.cfg
index 4d8ce8d1372..c9ce2433e17 100644
--- a/config-model/src/test/derived/tensor/index-info.cfg
+++ b/config-model/src/test/derived/tensor/index-info.cfg
@@ -35,3 +35,9 @@ indexinfo[].command[].indexname "f6"
 indexinfo[].command[].command "numerical"
 indexinfo[].command[].indexname "f6"
 indexinfo[].command[].command "type float"
+indexinfo[].command[].indexname "f7"
+indexinfo[].command[].command "attribute"
+indexinfo[].command[].indexname "f7"
+indexinfo[].command[].command "type tensor<int8>(p{},x[5])"
+indexinfo[].command[].indexname "f7"
+indexinfo[].command[].command "word"
diff --git a/config-model/src/test/derived/tensor/rank-profiles.cfg b/config-model/src/test/derived/tensor/rank-profiles.cfg
index 1ec3d67cb47..cd8375cb68d 100644
--- a/config-model/src/test/derived/tensor/rank-profiles.cfg
+++ b/config-model/src/test/derived/tensor/rank-profiles.cfg
@@ -1,4 +1,6 @@
 rankprofile[].name "default"
+rankprofile[].fef.property[].name "vespa.type.attribute.f7"
+rankprofile[].fef.property[].value "tensor<int8>(p{},x[5])"
 rankprofile[].fef.property[].name "vespa.type.attribute.f2"
 rankprofile[].fef.property[].value "tensor<float>(x[2],y[1])"
 rankprofile[].fef.property[].name "vespa.type.attribute.f3"
@@ -16,6 +18,8 @@ rankprofile[].fef.property[].name "vespa.hitcollector.arraysize"
 rankprofile[].fef.property[].value "0"
 rankprofile[].fef.property[].name "vespa.dump.ignoredefaultfeatures"
 rankprofile[].fef.property[].value "true"
+rankprofile[].fef.property[].name "vespa.type.attribute.f7"
+rankprofile[].fef.property[].value "tensor<int8>(p{},x[5])"
 rankprofile[].fef.property[].name "vespa.type.attribute.f2"
 rankprofile[].fef.property[].value "tensor<float>(x[2],y[1])"
 rankprofile[].fef.property[].name "vespa.type.attribute.f3"
@@ -29,6 +33,8 @@ rankprofile[].fef.property[].name "vespa.rank.firstphase"
 rankprofile[].fef.property[].value "rankingExpression(firstphase)"
 rankprofile[].fef.property[].name "rankingExpression(firstphase).rankingScript"
 rankprofile[].fef.property[].value "reduce(map(attribute(f4), f(x)(x * x)) + reduce(tensor(x[2],y[3])(random), count) * rename(attribute(f4), (x, y), (y, x)), sum)"
+rankprofile[].fef.property[].name "vespa.type.attribute.f7"
+rankprofile[].fef.property[].value "tensor<int8>(p{},x[5])"
 rankprofile[].fef.property[].name "vespa.type.attribute.f2"
 rankprofile[].fef.property[].value "tensor<float>(x[2],y[1])"
 rankprofile[].fef.property[].name "vespa.type.attribute.f3"
@@ -42,6 +48,8 @@ rankprofile[].fef.property[].name "vespa.rank.firstphase"
 rankprofile[].fef.property[].value "rankingExpression(firstphase)"
 rankprofile[].fef.property[].name "rankingExpression(firstphase).rankingScript"
 rankprofile[].fef.property[].value "reduce(reduce(join(attribute(f4), tensor(x[10],y[10],z[3])((x==y)*(y==z)), f(a,b)(a * b)), sum, x), sum)"
+rankprofile[].fef.property[].name "vespa.type.attribute.f7"
+rankprofile[].fef.property[].value "tensor<int8>(p{},x[5])"
 rankprofile[].fef.property[].name "vespa.type.attribute.f2"
 rankprofile[].fef.property[].value "tensor<float>(x[2],y[1])"
 rankprofile[].fef.property[].name "vespa.type.attribute.f3"
@@ -59,6 +67,8 @@ rankprofile[].fef.property[].name "vespa.rank.firstphase"
 rankprofile[].fef.property[].value "rankingExpression(firstphase)"
 rankprofile[].fef.property[].name "rankingExpression(firstphase).rankingScript"
 rankprofile[].fef.property[].value "reduce(rankingExpression(joinedtensors), sum)"
+rankprofile[].fef.property[].name "vespa.type.attribute.f7"
+rankprofile[].fef.property[].value "tensor<int8>(p{},x[5])"
 rankprofile[].fef.property[].name "vespa.type.attribute.f2"
 rankprofile[].fef.property[].value "tensor<float>(x[2],y[1])"
 rankprofile[].fef.property[].name "vespa.type.attribute.f3"
@@ -72,6 +82,8 @@ rankprofile[].fef.property[].name "vespa.rank.firstphase"
 rankprofile[].fef.property[].value "rankingExpression(firstphase)"
 rankprofile[].fef.property[].name "rankingExpression(firstphase).rankingScript"
 rankprofile[].fef.property[].value "reduce(attribute(f5), sum)"
+rankprofile[].fef.property[].name "vespa.type.attribute.f7"
+rankprofile[].fef.property[].value "tensor<int8>(p{},x[5])"
 rankprofile[].fef.property[].name "vespa.type.attribute.f2"
 rankprofile[].fef.property[].value "tensor<float>(x[2],y[1])"
 rankprofile[].fef.property[].name "vespa.type.attribute.f3"
@@ -85,6 +97,8 @@ rankprofile[].fef.property[].name "vespa.rank.firstphase"
 rankprofile[].fef.property[].value "rankingExpression(firstphase)"
 rankprofile[].fef.property[].name "rankingExpression(firstphase).rankingScript"
 rankprofile[].fef.property[].value "reduce(tensor<float>(d0[1],x[2]):{{d0:0,x:0}:(attribute(f6)),{d0:0,x:1}:(reduce(attribute(f5), sum))}, sum)"
+rankprofile[].fef.property[].name "vespa.type.attribute.f7"
+rankprofile[].fef.property[].value "tensor<int8>(p{},x[5])"
 rankprofile[].fef.property[].name "vespa.type.attribute.f2"
 rankprofile[].fef.property[].value "tensor<float>(x[2],y[1])"
 rankprofile[].fef.property[].name "vespa.type.attribute.f3"
@@ -102,6 +116,8 @@ rankprofile[].fef.property[].name "vespa.rank.firstphase"
 rankprofile[].fef.property[].value "rankingExpression(firstphase)"
 rankprofile[].fef.property[].name "rankingExpression(firstphase).rankingScript"
 rankprofile[].fef.property[].value "reduce(tensor<float>(d0[1],x[2]):{{d0:0,x:0}:(attribute(f6)),{d0:0,x:1}:(reduce(rankingExpression(joinedtensors), sum))}, sum)"
+rankprofile[].fef.property[].name "vespa.type.attribute.f7"
+rankprofile[].fef.property[].value "tensor<int8>(p{},x[5])"
 rankprofile[].fef.property[].name "vespa.type.attribute.f2"
 rankprofile[].fef.property[].value "tensor<float>(x[2],y[1])"
 rankprofile[].fef.property[].name "vespa.type.attribute.f3"
@@ -123,6 +139,8 @@ rankprofile[].fef.property[].name "vespa.rank.firstphase"
 rankprofile[].fef.property[].value "rankingExpression(firstphase)"
 rankprofile[].fef.property[].name "rankingExpression(firstphase).rankingScript"
 rankprofile[].fef.property[].value "reduce(rankingExpression(reshaped) * rankingExpression(literal), sum)"
+rankprofile[].fef.property[].name "vespa.type.attribute.f7"
+rankprofile[].fef.property[].value "tensor<int8>(p{},x[5])"
 rankprofile[].fef.property[].name "vespa.type.attribute.f2"
 rankprofile[].fef.property[].value "tensor<float>(x[2],y[1])"
 rankprofile[].fef.property[].name "vespa.type.attribute.f3"
@@ -138,6 +156,8 @@ rankprofile[].fef.property[].name "vespa.rank.firstphase"
 rankprofile[].fef.property[].value "rankingExpression(firstphase)"
 rankprofile[].fef.property[].name "rankingExpression(firstphase).rankingScript"
 rankprofile[].fef.property[].value "reduce(tensor(d0[1])((attribute(f3){x:(rankingExpression(functionNotLabel))})), sum)"
+rankprofile[].fef.property[].name "vespa.type.attribute.f7"
+rankprofile[].fef.property[].value "tensor<int8>(p{},x[5])"
 rankprofile[].fef.property[].name "vespa.type.attribute.f2"
 rankprofile[].fef.property[].value "tensor<float>(x[2],y[1])"
 rankprofile[].fef.property[].name "vespa.type.attribute.f3"
@@ -153,6 +173,8 @@ rankprofile[].fef.property[].name "vespa.rank.firstphase"
 rankprofile[].fef.property[].value "rankingExpression(firstphase)"
 rankprofile[].fef.property[].name "rankingExpression(firstphase).rankingScript"
 rankprofile[].fef.property[].value "reduce(tensor(shadow[1])((attribute(f3){x:(shadow + rankingExpression(shadow))})), sum)"
+rankprofile[].fef.property[].name "vespa.type.attribute.f7"
+rankprofile[].fef.property[].value "tensor<int8>(p{},x[5])"
 rankprofile[].fef.property[].name "vespa.type.attribute.f2"
 rankprofile[].fef.property[].value "tensor<float>(x[2],y[1])"
 rankprofile[].fef.property[].name "vespa.type.attribute.f3"
@@ -161,3 +183,26 @@ rankprofile[].fef.property[].name "vespa.type.attribute.f4"
 rankprofile[].fef.property[].value "tensor(x[10],y[10])"
 rankprofile[].fef.property[].name "vespa.type.attribute.f5"
 rankprofile[].fef.property[].value "tensor<float>(x[10])"
+rankprofile[].name "with-unpack"
+rankprofile[].fef.property[].name "rankingExpression(myunpack).rankingScript"
+rankprofile[].fef.property[].value "map_subspaces(attribute(f7), f(denseSubspaceInput)(tensor<float>(x[40])(bit(denseSubspaceInput{x:(x/8)}, 7-(x % 8)))))"
+rankprofile[].fef.property[].name "rankingExpression(myunpack).type"
+rankprofile[].fef.property[].value "tensor<float>(p{},x[40])"
+rankprofile[].fef.property[].name "vespa.rank.firstphase"
+rankprofile[].fef.property[].value "rankingExpression(firstphase)"
+rankprofile[].fef.property[].name "rankingExpression(firstphase).rankingScript"
+rankprofile[].fef.property[].value "reduce(query(para) * rankingExpression(myunpack) * query(qvec), sum)"
+rankprofile[].fef.property[].name "vespa.type.attribute.f7"
+rankprofile[].fef.property[].value "tensor<int8>(p{},x[5])"
+rankprofile[].fef.property[].name "vespa.type.attribute.f2"
+rankprofile[].fef.property[].value "tensor<float>(x[2],y[1])"
+rankprofile[].fef.property[].name "vespa.type.attribute.f3"
+rankprofile[].fef.property[].value "tensor(x{})"
+rankprofile[].fef.property[].name "vespa.type.attribute.f4"
+rankprofile[].fef.property[].value "tensor(x[10],y[10])"
+rankprofile[].fef.property[].name "vespa.type.attribute.f5"
+rankprofile[].fef.property[].value "tensor<float>(x[10])"
+rankprofile[].fef.property[].name "vespa.type.query.para"
+rankprofile[].fef.property[].value "tensor<float>(p{})"
+rankprofile[].fef.property[].name "vespa.type.query.qvec"
+rankprofile[].fef.property[].value "tensor<float>(x[40])"
diff --git a/config-model/src/test/derived/tensor/tensor.sd b/config-model/src/test/derived/tensor/tensor.sd
index f2fc57a2018..81230e5c54c 100644
--- a/config-model/src/test/derived/tensor/tensor.sd
+++ b/config-model/src/test/derived/tensor/tensor.sd
@@ -20,6 +20,9 @@ schema tensor {
     field f6 type float {
       indexing: attribute
     }
+    field f7 type tensor<int8>(p{},x[5]) {
+      indexing: attribute
+    }
   }
   
   rank-profile profile1 {
@@ -119,4 +122,17 @@ schema tensor {
 
   }
 
+  rank-profile with-unpack {
+    inputs {
+      query(para) tensor<float>(p{})
+      query(qvec) tensor<float>(x[40])
+    }
+    function myunpack() {
+      expression: unpack_bits_from_int8(attribute(f7))
+    }
+    first-phase {
+      expression: sum(query(para)*myunpack*query(qvec))
+    }
+  }
+
 }
diff --git a/searchlib/abi-spec.json b/searchlib/abi-spec.json
index ca475a77b6c..4f0a99a117d 100644
--- a/searchlib/abi-spec.json
+++ b/searchlib/abi-spec.json
@@ -953,6 +953,8 @@
       "public final com.yahoo.searchlib.rankingexpression.rule.TensorFunctionNode tensorArgmax()",
       "public final com.yahoo.searchlib.rankingexpression.rule.TensorFunctionNode tensorArgmin()",
       "public final com.yahoo.searchlib.rankingexpression.rule.TensorFunctionNode tensorCellCast()",
+      "public final com.yahoo.searchlib.rankingexpression.rule.ExpressionNode tensorMacro()",
+      "public final com.yahoo.searchlib.rankingexpression.rule.ExpressionNode tensorUnpackBitsFromInt8()",
       "public final com.yahoo.searchlib.rankingexpression.rule.LambdaFunctionNode lambdaFunction()",
       "public final com.yahoo.tensor.functions.Reduce$Aggregator tensorReduceAggregator()",
       "public final com.yahoo.tensor.TensorType tensorType(java.util.List)",
@@ -1086,6 +1088,7 @@
       "public static final int HAMMING",
       "public static final int MAP",
       "public static final int MAP_SUBSPACES",
+      "public static final int UNPACK_BITS_FROM_INT8",
       "public static final int REDUCE",
       "public static final int JOIN",
       "public static final int MERGE",
@@ -1707,5 +1710,22 @@
       "public int hashCode()"
     ],
     "fields" : [ ]
+  },
+  "com.yahoo.searchlib.rankingexpression.rule.UnpackBitsFromInt8" : {
+    "superClass" : "com.yahoo.searchlib.rankingexpression.rule.CompositeNode",
+    "interfaces" : [ ],
+    "attributes" : [
+      "public"
+    ],
+    "methods" : [
+      "public void <init>(com.yahoo.searchlib.rankingexpression.rule.ExpressionNode, com.yahoo.tensor.TensorType$Value, java.lang.String)",
+      "public java.util.List children()",
+      "public java.lang.StringBuilder toString(java.lang.StringBuilder, com.yahoo.searchlib.rankingexpression.rule.SerializationContext, java.util.Deque, com.yahoo.searchlib.rankingexpression.rule.CompositeNode)",
+      "public com.yahoo.searchlib.rankingexpression.evaluation.Value evaluate(com.yahoo.searchlib.rankingexpression.evaluation.Context)",
+      "public com.yahoo.tensor.TensorType type(com.yahoo.tensor.evaluation.TypeContext)",
+      "public com.yahoo.searchlib.rankingexpression.rule.CompositeNode setChildren(java.util.List)",
+      "public int hashCode()"
+    ],
+    "fields" : [ ]
   }
 }
\ No newline at end of file
diff --git a/searchlib/src/main/java/com/yahoo/searchlib/rankingexpression/rule/UnpackBitsFromInt8.java b/searchlib/src/main/java/com/yahoo/searchlib/rankingexpression/rule/UnpackBitsFromInt8.java
new file mode 100644
index 00000000000..84203da4a7e
--- /dev/null
+++ b/searchlib/src/main/java/com/yahoo/searchlib/rankingexpression/rule/UnpackBitsFromInt8.java
@@ -0,0 +1,183 @@
+// Copyright Vespa.ai. Licensed under the terms of the Apache 2.0 license. See LICENSE in the project root.
+package com.yahoo.searchlib.rankingexpression.rule;
+
+import com.yahoo.api.annotations.Beta;
+import com.yahoo.searchlib.rankingexpression.Reference;
+import com.yahoo.searchlib.rankingexpression.evaluation.Context;
+import com.yahoo.searchlib.rankingexpression.evaluation.TensorValue;
+import com.yahoo.searchlib.rankingexpression.evaluation.Value;
+import com.yahoo.tensor.Tensor;
+import com.yahoo.tensor.TensorAddress;
+import com.yahoo.tensor.TensorType;
+import com.yahoo.tensor.evaluation.TypeContext;
+
+import java.util.ArrayList;
+import java.util.Collections;
+import java.util.Deque;
+import java.util.List;
+import java.util.Optional;
+import java.util.Objects;
+
+/**
+ * Macro that expands to the appropriate map_subspaces magic incantation
+ *
+ * @author arnej
+ */
+@Beta
+public class UnpackBitsFromInt8 extends CompositeNode {
+
+    private static String operationName = "unpack_bits_from_int8";
+    private enum EndianNess {
+        BIG_ENDIAN("big"), LITTLE_ENDIAN("little");
+
+        private final String id;
+        EndianNess(String id) { this.id = id; }
+        public String toString() { return id; }
+        public static EndianNess fromId(String id) {
+            for (EndianNess value : values()) {
+                if (value.id.equals(id)) {
+                    return value;
+                }
+            }
+            throw new IllegalArgumentException("EndianNess must be either 'big' or 'little', but was '" + id + "'");
+        }
+    };
+
+    final ExpressionNode input;
+    final TensorType.Value targetCellType;
+    final EndianNess endian;
+
+    public UnpackBitsFromInt8(ExpressionNode input, TensorType.Value targetCellType, String endianNess) {
+        this.input = input;
+        this.targetCellType = targetCellType;
+        this.endian = EndianNess.fromId(endianNess);
+    }
+
+    @Override
+    public List<ExpressionNode> children() {
+        return Collections.singletonList(input);
+    }
+
+    private static record Meta(TensorType outputType, TensorType outputDenseType, String unpackDimension) {}
+
+    @Override
+    public StringBuilder toString(StringBuilder string, SerializationContext context, Deque<String> path, CompositeNode parent) {
+        var optTC = context.typeContext();
+        if (optTC.isPresent()) {
+            TensorType inputType = input.type(optTC.get());
+            var meta = analyze(inputType);
+            string.append("map_subspaces").append("(");
+            input.toString(string, context, path, this);
+            string.append(", f(denseSubspaceInput)(");
+            string.append(meta.outputDenseType()).append("("); // generate
+            string.append("bit(denseSubspaceInput{");
+            for (var dim : meta.outputDenseType().dimensions()) {
+                String dName = dim.name();
+                boolean last = dName.equals(meta.unpackDimension);
+                string.append(dName);
+                string.append(":(");
+                string.append(dName);
+                if (last) {
+                    string.append("/8");
+                }
+                string.append(")");
+                if (! last) {
+                    string.append(", ");
+                }
+            }
+            if (endian.equals(EndianNess.BIG_ENDIAN)) {
+                string.append("}, 7-(");
+            } else {
+                string.append("}, (");
+            }
+            string.append(meta.unpackDimension);
+            string.append(" % 8)");
+            string.append("))))"); // bit, generate, f, map_subspaces
+        } else {
+            string.append(operationName);
+            string.append("(");
+            input.toString(string, context, path, this);
+            string.append(",");
+            string.append(targetCellType);
+            string.append(",");
+            string.append(endian);
+            string.append(")");
+        }
+        return string;
+    }
+
+    @Override
+    public Value evaluate(Context context) {
+        Tensor inputTensor = input.evaluate(context).asTensor();
+        TensorType inputType = inputTensor.type();
+        var meta = analyze(inputType);
+        var builder = Tensor.Builder.of(meta.outputType());
+        for (var iter = inputTensor.cellIterator(); iter.hasNext(); ) {
+            var cell = iter.next();
+            var oldAddr = cell.getKey();
+            for (int bitIdx = 0; bitIdx < 8; bitIdx++) {
+                var addrBuilder = new TensorAddress.Builder(meta.outputType());
+                for (int i = 0; i < inputType.dimensions().size(); i++) {
+                    var dim = inputType.dimensions().get(i);
+                    if (dim.name().equals(meta.unpackDimension())) {
+                        long newIdx = oldAddr.numericLabel(i) * 8 + bitIdx;
+                        addrBuilder.add(dim.name(), String.valueOf(newIdx));
+                    } else {
+                        addrBuilder.add(dim.name(), oldAddr.label(i));
+                    }
+                }
+                var newAddr = addrBuilder.build();
+                int oldValue = (int)(cell.getValue().doubleValue());
+                if (endian.equals(EndianNess.BIG_ENDIAN)) {
+                    float newCellValue = 1 & (oldValue >>> (7 - bitIdx));
+                    builder.cell(newAddr, newCellValue);
+                } else {
+                    float newCellValue = 1 & (oldValue >>> bitIdx);
+                    builder.cell(newAddr, newCellValue);
+                }
+            }
+        }
+        return new TensorValue(builder.build());
+    }
+
+    private Meta analyze(TensorType inputType) {
+        TensorType inputDenseType = inputType.indexedSubtype();
+        if (inputDenseType.rank() == 0) {
+            throw new IllegalArgumentException("bad " + operationName + "; input must have indexed dimension, but type was: " + inputType);
+        }
+        var lastDim = inputDenseType.dimensions().get(inputDenseType.rank() - 1);
+        if (lastDim.size().isEmpty()) {
+            throw new IllegalArgumentException("bad " + operationName + "; last indexed dimension must be bound, but type was: " + inputType);
+        }
+        List<TensorType.Dimension> outputDims = new ArrayList<>();
+        var ttBuilder = new TensorType.Builder(targetCellType);
+        for (var dim : inputType.dimensions()) {
+            if (dim.name().equals(lastDim.name())) {
+                long sz = dim.size().get();
+                ttBuilder.indexed(dim.name(), sz * 8);
+            } else {
+                ttBuilder.set(dim);
+            }
+        }
+        TensorType outputType = ttBuilder.build();
+        return new Meta(outputType, outputType.indexedSubtype(), lastDim.name());
+    }
+
+    @Override
+    public TensorType type(TypeContext<Reference> context) {
+        TensorType inputType = input.type(context);
+        var meta = analyze(inputType);
+        return meta.outputType();
+    }
+
+    @Override
+    public CompositeNode setChildren(List<ExpressionNode> newChildren) {
+        if (newChildren.size() != 1)
+            throw new IllegalArgumentException("Expected 1 child but got " + newChildren.size());
+        return new UnpackBitsFromInt8(newChildren.get(0), targetCellType, endian.toString());
+    }
+
+    @Override
+    public int hashCode() { return Objects.hash(operationName, input, targetCellType); }
+
+}
diff --git a/searchlib/src/main/javacc/RankingExpressionParser.jj b/searchlib/src/main/javacc/RankingExpressionParser.jj
index 6cd01151dc1..1da8a5ece89 100755
--- a/searchlib/src/main/javacc/RankingExpressionParser.jj
+++ b/searchlib/src/main/javacc/RankingExpressionParser.jj
@@ -129,6 +129,7 @@ TOKEN :
 
     <MAP: "map"> |
     <MAP_SUBSPACES: "map_subspaces"> |
+    <UNPACK_BITS_FROM_INT8: "unpack_bits_from_int8"> |
     <REDUCE: "reduce"> |
     <JOIN: "join"> |
     <MERGE: "merge"> |
@@ -344,7 +345,7 @@ ExpressionNode function() :
     ExpressionNode function;
 }
 {
-    ( LOOKAHEAD(2) function = scalarOrTensorFunction() | function = tensorFunction() )
+    ( LOOKAHEAD(2) function = scalarOrTensorFunction() | function = tensorFunction() | function = tensorMacro() )
     { return function; }
 }
 
@@ -669,6 +670,32 @@ TensorFunctionNode tensorCellCast() :
     { return new TensorFunctionNode(new CellCast(TensorFunctionNode.wrap(tensor), TensorType.Value.fromId(valueType)));}
 }
 
+ExpressionNode tensorMacro() :
+{
+    ExpressionNode tensorExpression;
+}
+{
+    (
+        tensorExpression = tensorUnpackBitsFromInt8()
+    )
+    { return tensorExpression; }
+}
+
+ExpressionNode tensorUnpackBitsFromInt8() :
+{
+    ExpressionNode tensor;
+    String targetCellType = "float";
+    String endianNess = "big";
+}
+{
+    <UNPACK_BITS_FROM_INT8> <LBRACE> tensor = expression() (
+      <COMMA> targetCellType = identifier() (
+        <COMMA> endianNess = identifier() )? )? <RBRACE>
+    {
+        return new UnpackBitsFromInt8(tensor, TensorType.Value.fromId(targetCellType), endianNess);
+    }
+}
+
 LambdaFunctionNode lambdaFunction() :
 {
     List<String> variables;
diff --git a/searchlib/src/test/java/com/yahoo/searchlib/rankingexpression/evaluation/EvaluationTestCase.java b/searchlib/src/test/java/com/yahoo/searchlib/rankingexpression/evaluation/EvaluationTestCase.java
index 637e9be5fc3..6ac734e8771 100644
--- a/searchlib/src/test/java/com/yahoo/searchlib/rankingexpression/evaluation/EvaluationTestCase.java
+++ b/searchlib/src/test/java/com/yahoo/searchlib/rankingexpression/evaluation/EvaluationTestCase.java
@@ -177,6 +177,28 @@ public class EvaluationTestCase {
     }
 
     @Test
+    public void testUnpack() {
+        EvaluationTester tester = new EvaluationTester();
+        tester.assertEvaluates("tensor<float>(a{},x[16]):{foo:[" +
+                               "0,0,0,0, 0,0,0,0," +
+                               "1,1,1,1, 1,1,1,1" +
+                               "],bar:[" +
+                               "0,0,0,0, 0,0,0,1," +
+                               "1,1,1,1, 1,0,0,0]}",
+                               "unpack_bits_from_int8(tensor0, float, big)",
+                               "tensor<int8>(a{},x[2]):{foo:[0,255],bar:[1,-8]}");
+
+        tester.assertEvaluates("tensor<int8>(a{},x[16]):{foo:[" +
+                               "0,0,0,0, 0,0,0,0," +
+                               "1,1,1,1, 1,1,1,1" +
+                               "],bar:[" +
+                               "1,0,0,0, 0,0,0,0," +
+                               "0,0,0,1, 1,1,1,1]}",
+                               "unpack_bits_from_int8(tensor0, int8, little)",
+                               "tensor<int8>(a{},x[2]):{foo:[0,255],bar:[1,-8]}");
+    }
+
+    @Test
     public void testMapSubspaces() {
         EvaluationTester tester = new EvaluationTester();
         tester.assertEvaluates("tensor<float>(a{},x[2]):{foo:[2,3],bar:[7,10]}",