1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
|
// Copyright Yahoo. Licensed under the terms of the Apache 2.0 license. See LICENSE in the project root.
package ai.vespa.rankingexpression.importer.operations;
import ai.vespa.rankingexpression.importer.DimensionRenamer;
import ai.vespa.rankingexpression.importer.OrderedTensorType;
import com.yahoo.searchlib.rankingexpression.Reference;
import com.yahoo.searchlib.rankingexpression.evaluation.DoubleValue;
import com.yahoo.searchlib.rankingexpression.rule.ConstantNode;
import com.yahoo.searchlib.rankingexpression.rule.EmbracedNode;
import com.yahoo.searchlib.rankingexpression.rule.ExpressionNode;
import com.yahoo.tensor.TensorType;
import com.yahoo.tensor.functions.Join;
import com.yahoo.tensor.functions.Reduce;
import com.yahoo.tensor.functions.ScalarFunctions;
import com.yahoo.tensor.functions.Slice;
import com.yahoo.tensor.functions.TensorFunction;
import java.util.ArrayList;
import java.util.List;
import java.util.Optional;
import static com.yahoo.searchlib.rankingexpression.rule.TensorFunctionNode.wrapScalar;
public class MatMul extends IntermediateOperation {
public MatMul(String modelName, String nodeName, List<IntermediateOperation> inputs) {
super(modelName, nodeName, inputs);
}
@Override
protected OrderedTensorType lazyGetType() {
if ( ! allInputTypesPresent(2)) return null;
OrderedTensorType typeA = inputs.get(0).type().get();
OrderedTensorType typeB = inputs.get(1).type().get();
if (typeA.type().rank() < 1 || typeB.type().rank() < 1)
throw new IllegalArgumentException("Tensors in matmul must have rank of at least 1");
OrderedTensorType.Builder typeBuilder = new OrderedTensorType.Builder(resultValueType());
OrderedTensorType largestRankType = typeA.rank() >= typeB.rank() ? typeA : typeB;
OrderedTensorType smallestRankType = typeA.rank() >= typeB.rank() ? typeB : typeA;
for (int i = 0; i < largestRankType.rank() - 2; ++i) {
TensorType.Dimension dim = largestRankType.dimensions().get(i);
// broadcasting
int j = smallestRankType.rank() - largestRankType.rank() + i;
if (j >= 0 && smallestRankType.dimensions().get(j).size().get() > dim.size().get()) {
dim = smallestRankType.dimensions().get(j);
}
typeBuilder.add(dim);
}
if (typeA.rank() >= 2) {
typeBuilder.add(typeA.dimensions().get(typeA.rank() - 2));
}
if (typeB.rank() >= 2) {
typeBuilder.add(typeB.dimensions().get(typeB.rank() - 1));
}
return typeBuilder.build();
}
@Override
protected TensorFunction<Reference> lazyGetFunction() {
if ( ! allInputTypesPresent(2)) return null;
if ( ! allInputFunctionsPresent(2)) return null;
OrderedTensorType typeA = inputs.get(0).type().get();
OrderedTensorType typeB = inputs.get(1).type().get();
TensorFunction<Reference> functionA = handleBroadcasting(inputs.get(0).function().get(), typeA, typeB);
TensorFunction<Reference> functionB = handleBroadcasting(inputs.get(1).function().get(), typeB, typeA);
return new com.yahoo.tensor.functions.Reduce<Reference>(
new Join<Reference>(functionA, functionB, ScalarFunctions.multiply()),
Reduce.Aggregator.sum,
typeA.dimensions().get(typeA.rank() - 1).name());
}
private TensorFunction<Reference> handleBroadcasting(TensorFunction<Reference> tensorFunction, OrderedTensorType typeA, OrderedTensorType typeB) {
List<Slice.DimensionValue<Reference>> slices = new ArrayList<>();
for (int i = 0; i < typeA.rank() - 2; ++i) {
long dimSizeA = typeA.dimensions().get(i).size().get();
String dimNameA = typeA.dimensionNames().get(i);
int j = typeB.rank() - typeA.rank() + i;
if (j >= 0) {
long dimSizeB = typeB.dimensions().get(j).size().get();
if (dimSizeB > dimSizeA && dimSizeA == 1) {
ExpressionNode dimensionExpression = new EmbracedNode(new ConstantNode(DoubleValue.zero));
slices.add(new Slice.DimensionValue<>(Optional.of(dimNameA), wrapScalar(dimensionExpression)));
}
}
}
return slices.size() == 0 ? tensorFunction : new Slice<>(tensorFunction, slices);
}
@Override
public void addDimensionNameConstraints(DimensionRenamer renamer) {
if ( ! allInputTypesPresent(2)) return;
OrderedTensorType typeA = inputs.get(0).type().get();
OrderedTensorType typeB = inputs.get(1).type().get();
String lastDimA = typeA.dimensions().get(typeA.rank()-1).name();
String lastDimB = typeB.dimensions().get(typeB.rank()-1).name();
String secondLastDimA = typeA.dimensions().get(Math.max(0,typeA.rank()-2)).name();
String secondLastDimB = typeB.dimensions().get(Math.max(0,typeB.rank()-2)).name();
// The last dimension of A should have the same name as the second-to-last dimension of B
renamer.addConstraint(lastDimA, secondLastDimB, DimensionRenamer.Constraint.equal(false), this);
// The second-to-last dimension of a should have a different name than the last dimension of b
if (typeA.rank() >= 2 && typeB.rank() >= 2) {
renamer.addConstraint(secondLastDimA, lastDimB, DimensionRenamer.Constraint.lessThan(false), this);
}
// For efficiency, the dimensions to join over should be innermost - soft constraint
if (typeA.rank() >= 2) {
renamer.addConstraint(secondLastDimA, lastDimA, DimensionRenamer.Constraint.lessThan(true), this);
}
if (typeB.rank() >= 2) {
renamer.addConstraint(secondLastDimB, lastDimB, DimensionRenamer.Constraint.greaterThan(true), this);
}
// Handle different cases when at least one of the tensors have rank > 2
for (int i = 0; i < typeA.rank() - 2; ++i) {
String iDim = typeA.dimensionNames().get(i);
// a1 < a2 < a3 < a4
for (int j = i+1; j < typeA.rank(); ++j) {
String jDim = typeA.dimensionNames().get(j);
renamer.addConstraint(iDim, jDim, DimensionRenamer.Constraint.lessThan(false), this);
}
// not equal to last 2 dimensions in B
for (int j = typeB.rank()-2; j < typeB.rank(); ++j) {
if (j < 0) continue;
String jDim = typeB.dimensionNames().get(j);
renamer.addConstraint(iDim, jDim, DimensionRenamer.Constraint.notEqual(false), this);
}
// equal to matching dimension in tensor B
int j = typeB.rank() - typeA.rank() + i;
if (j >= 0) {
String jDim = typeB.dimensionNames().get(j);
renamer.addConstraint(iDim, jDim, DimensionRenamer.Constraint.equal(false), this);
}
}
for (int i = 0; i < typeB.rank() - 2; ++i) {
String iDim = typeB.dimensionNames().get(i);
// b1 < b2 < b3 < b4
for (int j = i+1; j < typeB.rank(); ++j) {
String jDim = typeB.dimensionNames().get(j);
renamer.addConstraint(iDim, jDim, DimensionRenamer.Constraint.lessThan(false), this);
}
// not equal to last 2 dimensions in A
for (int j = typeA.rank()-2; j < typeA.rank(); ++j) {
if (j < 0) continue;
String jDim = typeA.dimensionNames().get(j);
renamer.addConstraint(iDim, jDim, DimensionRenamer.Constraint.notEqual(false), this);
}
// equal to matching dimension in tensor A
int j = typeA.rank() - typeB.rank() + i;
if (j >= 0) {
String jDim = typeA.dimensionNames().get(j);
renamer.addConstraint(iDim, jDim, DimensionRenamer.Constraint.equal(false), this);
}
}
}
@Override
public MatMul withInputs(List<IntermediateOperation> inputs) {
return new MatMul(modelName(), name(), inputs);
}
@Override
public String operationName() { return "MatMul"; }
}
|
