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
|
// Copyright Vespa.ai. Licensed under the terms of the Apache 2.0 license. See LICENSE in the project root.
package com.yahoo.tensor;
import com.yahoo.tensor.evaluation.MapEvaluationContext;
import com.yahoo.tensor.evaluation.Name;
import com.yahoo.tensor.evaluation.VariableTensor;
import com.yahoo.tensor.functions.ConstantTensor;
import com.yahoo.tensor.functions.Join;
import com.yahoo.tensor.functions.Reduce;
import com.yahoo.tensor.functions.TensorFunction;
import java.util.*;
import java.util.stream.Collectors;
/**
* Microbenchmark of tensor operations.
*
* @author bratseth
*/
public class TensorFunctionBenchmark {
private final static Random random = new Random();
public double benchmark(int iterations, List<Tensor> modelVectors, TensorType.Dimension.Type dimensionType,
boolean extraSpace) {
Tensor queryVector = vectors(1, 300, dimensionType).get(0);
if (extraSpace) {
queryVector = queryVector.multiply(unitVector("j"));
modelVectors = modelVectors.stream().map(t -> t.multiply(unitVector("k"))).toList();
}
dotProduct(queryVector, modelVectors, Math.max(iterations/10, 10)); // warmup
System.gc();
long startTime = System.currentTimeMillis();
dotProduct(queryVector, modelVectors, iterations);
long totalTime = System.currentTimeMillis() - startTime;
return (double)totalTime / (double)iterations;
}
private Tensor unitVector(String dimension) {
return Tensor.Builder.of(new TensorType.Builder().indexed(dimension, 1).build())
.cell().label(dimension, 0).value(1).build();
}
private double dotProduct(Tensor tensor, List<Tensor> tensors, int iterations) {
double result = 0;
for (int i = 0 ; i < iterations; i++)
result = dotProduct(tensor, tensors);
return result;
}
private double dotProduct(Tensor tensor, List<Tensor> tensors) {
double largest = Double.MIN_VALUE;
TensorFunction<Name> dotProductFunction = new Reduce<>(new Join<>(new ConstantTensor<>(tensor),
new VariableTensor<>("argument"), (a, b) -> a * b),
Reduce.Aggregator.sum).toPrimitive();
MapEvaluationContext<Name> context = new MapEvaluationContext<>();
for (Tensor tensorElement : tensors) { // tensors.size() = 1 for larger tensor
context.put("argument", tensorElement);
double dotProduct = dotProductFunction.evaluate(context).asDouble();
if (dotProduct > largest) {
largest = dotProduct;
}
}
return largest;
}
private static List<Tensor> vectors(int vectorCount, int vectorSize, TensorType.Dimension.Type dimensionType) {
List<Tensor> tensors = new ArrayList<>();
TensorType type = vectorType(new TensorType.Builder(), "x", dimensionType, vectorSize);
for (int i = 0; i < vectorCount; i++) {
Tensor.Builder builder = Tensor.Builder.of(type);
for (int j = 0; j < vectorSize; j++) {
builder.cell().label("x", String.valueOf(j)).value(random.nextDouble());
}
tensors.add(builder.build());
}
return tensors;
}
private static List<Tensor> matrix(int vectorCount, int vectorSize, TensorType.Dimension.Type dimensionType) {
TensorType.Builder typeBuilder = new TensorType.Builder();
typeBuilder.dimension("i", dimensionType == TensorType.Dimension.Type.indexedBound ? TensorType.Dimension.Type.indexedUnbound : dimensionType);
vectorType(typeBuilder, "x", dimensionType, vectorSize);
Tensor.Builder builder = Tensor.Builder.of(typeBuilder.build());
for (int i = 0; i < vectorCount; i++) {
for (int j = 0; j < vectorSize; j++) {
builder.cell()
.label("i", String.valueOf(i))
.label("x", String.valueOf(j))
.value(random.nextDouble());
}
}
return Collections.singletonList(builder.build());
}
private static TensorType vectorType(TensorType.Builder builder, String name, TensorType.Dimension.Type type, int size) {
switch (type) {
case mapped: builder.mapped(name); break;
case indexedUnbound: builder.indexed(name); break;
case indexedBound: builder.indexed(name, size); break;
default: throw new IllegalArgumentException("Dimension type " + type + " not supported");
}
return builder.build();
}
public static void main(String[] args) {
double time = 0;
// ---------------- 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:
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);
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):
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);
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);
// ---------------- 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
*/
}
}
|
