[16285] | 1 | using System;
|
---|
| 2 | using System.Collections.Generic;
|
---|
| 3 | using System.Linq;
|
---|
| 4 |
|
---|
| 5 | using HeuristicLab.Common;
|
---|
| 6 | using HeuristicLab.Core;
|
---|
| 7 | using HeuristicLab.Data;
|
---|
| 8 | using HeuristicLab.Encodings.SymbolicExpressionTreeEncoding;
|
---|
| 9 | using HeuristicLab.Parameters;
|
---|
| 10 | using HeuristicLab.Persistence.Default.CompositeSerializers.Storable;
|
---|
| 11 |
|
---|
| 12 | using static HeuristicLab.Problems.DataAnalysis.Symbolic.BatchOperations;
|
---|
| 13 |
|
---|
| 14 | namespace HeuristicLab.Problems.DataAnalysis.Symbolic {
|
---|
| 15 | [Item("SymbolicDataAnalysisExpressionTreeBatchInterpreter", "An interpreter that uses batching and vectorization techniques to achieve faster performance.")]
|
---|
| 16 | [StorableClass]
|
---|
| 17 | public class SymbolicDataAnalysisExpressionTreeBatchInterpreter : ParameterizedNamedItem, ISymbolicDataAnalysisExpressionTreeInterpreter {
|
---|
| 18 | private const string EvaluatedSolutionsParameterName = "EvaluatedSolutions";
|
---|
| 19 |
|
---|
| 20 | #region parameters
|
---|
| 21 | public IFixedValueParameter<IntValue> EvaluatedSolutionsParameter {
|
---|
| 22 | get { return (IFixedValueParameter<IntValue>)Parameters[EvaluatedSolutionsParameterName]; }
|
---|
| 23 | }
|
---|
| 24 | #endregion
|
---|
| 25 |
|
---|
| 26 | #region properties
|
---|
| 27 | public int EvaluatedSolutions {
|
---|
| 28 | get { return EvaluatedSolutionsParameter.Value.Value; }
|
---|
| 29 | set { EvaluatedSolutionsParameter.Value.Value = value; }
|
---|
| 30 | }
|
---|
| 31 | #endregion
|
---|
| 32 |
|
---|
| 33 | public void ClearState() { }
|
---|
| 34 |
|
---|
| 35 | public SymbolicDataAnalysisExpressionTreeBatchInterpreter() {
|
---|
| 36 | Parameters.Add(new FixedValueParameter<IntValue>(EvaluatedSolutionsParameterName, "A counter for the total number of solutions the interpreter has evaluated", new IntValue(0)));
|
---|
| 37 | }
|
---|
| 38 |
|
---|
| 39 |
|
---|
| 40 | [StorableConstructor]
|
---|
| 41 | protected SymbolicDataAnalysisExpressionTreeBatchInterpreter(bool deserializing) : base(deserializing) { }
|
---|
| 42 | protected SymbolicDataAnalysisExpressionTreeBatchInterpreter(SymbolicDataAnalysisExpressionTreeBatchInterpreter original, Cloner cloner) : base(original, cloner) {
|
---|
| 43 | }
|
---|
| 44 | public override IDeepCloneable Clone(Cloner cloner) {
|
---|
| 45 | return new SymbolicDataAnalysisExpressionTreeBatchInterpreter(this, cloner);
|
---|
| 46 | }
|
---|
| 47 |
|
---|
| 48 | private void LoadData(BatchInstruction instr, int[] rows, int rowIndex, int batchSize) {
|
---|
| 49 | for (int i = 0; i < batchSize; ++i) {
|
---|
| 50 | var row = rows[rowIndex] + i;
|
---|
| 51 | instr.buf[i] = instr.weight * instr.data[row];
|
---|
| 52 | }
|
---|
| 53 | }
|
---|
| 54 |
|
---|
| 55 | private void Evaluate(BatchInstruction[] code, int[] rows, int rowIndex, int batchSize) {
|
---|
| 56 | for (int i = code.Length - 1; i >= 0; --i) {
|
---|
| 57 | var instr = code[i];
|
---|
| 58 | var c = instr.childIndex;
|
---|
| 59 | var n = instr.narg;
|
---|
| 60 |
|
---|
| 61 | switch (instr.opcode) {
|
---|
| 62 | case OpCodes.Variable: {
|
---|
| 63 | LoadData(instr, rows, rowIndex, batchSize);
|
---|
| 64 | break;
|
---|
| 65 | }
|
---|
[16293] | 66 |
|
---|
[16285] | 67 | case OpCodes.Add: {
|
---|
| 68 | Load(instr.buf, code[c].buf);
|
---|
| 69 | for (int j = 1; j < n; ++j) {
|
---|
| 70 | Add(instr.buf, code[c + j].buf);
|
---|
| 71 | }
|
---|
| 72 | break;
|
---|
| 73 | }
|
---|
| 74 |
|
---|
| 75 | case OpCodes.Sub: {
|
---|
| 76 | if (n == 1) {
|
---|
| 77 | Neg(instr.buf, code[c].buf);
|
---|
| 78 | } else {
|
---|
| 79 | Load(instr.buf, code[c].buf);
|
---|
| 80 | for (int j = 1; j < n; ++j) {
|
---|
| 81 | Sub(instr.buf, code[c + j].buf);
|
---|
| 82 | }
|
---|
| 83 | }
|
---|
[16293] | 84 | break;
|
---|
[16285] | 85 | }
|
---|
| 86 |
|
---|
| 87 | case OpCodes.Mul: {
|
---|
| 88 | Load(instr.buf, code[c].buf);
|
---|
| 89 | for (int j = 1; j < n; ++j) {
|
---|
| 90 | Mul(instr.buf, code[c + j].buf);
|
---|
| 91 | }
|
---|
| 92 | break;
|
---|
| 93 | }
|
---|
| 94 |
|
---|
| 95 | case OpCodes.Div: {
|
---|
| 96 | if (n == 1) {
|
---|
| 97 | Inv(instr.buf, code[c].buf);
|
---|
| 98 | } else {
|
---|
| 99 | Load(instr.buf, code[c].buf);
|
---|
| 100 | for (int j = 1; j < n; ++j) {
|
---|
| 101 | Div(instr.buf, code[c + j].buf);
|
---|
| 102 | }
|
---|
| 103 | }
|
---|
[16293] | 104 | break;
|
---|
[16285] | 105 | }
|
---|
| 106 |
|
---|
[16293] | 107 | case OpCodes.Square: {
|
---|
| 108 | Square(instr.buf, code[c].buf);
|
---|
| 109 | break;
|
---|
| 110 | }
|
---|
| 111 |
|
---|
| 112 | case OpCodes.Root: {
|
---|
| 113 | Root(instr.buf, code[c].buf);
|
---|
| 114 | break;
|
---|
| 115 | }
|
---|
| 116 |
|
---|
| 117 | case OpCodes.SquareRoot: {
|
---|
| 118 | Sqrt(instr.buf, code[c].buf);
|
---|
| 119 | break;
|
---|
| 120 | }
|
---|
| 121 |
|
---|
| 122 | case OpCodes.Power: {
|
---|
| 123 | Pow(instr.buf, code[c].buf);
|
---|
| 124 | break;
|
---|
| 125 | }
|
---|
| 126 |
|
---|
[16285] | 127 | case OpCodes.Exp: {
|
---|
| 128 | Exp(instr.buf, code[c].buf);
|
---|
| 129 | break;
|
---|
| 130 | }
|
---|
| 131 |
|
---|
| 132 | case OpCodes.Log: {
|
---|
| 133 | Log(instr.buf, code[c].buf);
|
---|
| 134 | break;
|
---|
| 135 | }
|
---|
[16293] | 136 |
|
---|
| 137 | case OpCodes.Sin: {
|
---|
| 138 | Sin(instr.buf, code[c].buf);
|
---|
| 139 | break;
|
---|
| 140 | }
|
---|
| 141 |
|
---|
| 142 | case OpCodes.Cos: {
|
---|
| 143 | Cos(instr.buf, code[c].buf);
|
---|
| 144 | break;
|
---|
| 145 | }
|
---|
| 146 |
|
---|
| 147 | case OpCodes.Tan: {
|
---|
| 148 | Tan(instr.buf, code[c].buf);
|
---|
| 149 | break;
|
---|
| 150 | }
|
---|
[16285] | 151 | }
|
---|
| 152 | }
|
---|
| 153 | }
|
---|
| 154 |
|
---|
[16296] | 155 | [ThreadStatic]
|
---|
| 156 | private Dictionary<string, double[]> cachedData;
|
---|
| 157 |
|
---|
| 158 | private void InitCache(IDataset dataset) {
|
---|
| 159 | cachedData = new Dictionary<string, double[]>();
|
---|
| 160 | foreach (var v in dataset.DoubleVariables) {
|
---|
[16301] | 161 | cachedData[v] = dataset.GetDoubleValues(v).ToArray();
|
---|
[16296] | 162 | }
|
---|
| 163 | }
|
---|
| 164 |
|
---|
| 165 | public void InitializeState() {
|
---|
| 166 | cachedData = null;
|
---|
| 167 | EvaluatedSolutions = 0;
|
---|
| 168 | }
|
---|
| 169 |
|
---|
[16285] | 170 | private double[] GetValues(ISymbolicExpressionTree tree, IDataset dataset, int[] rows) {
|
---|
| 171 | var code = Compile(tree, dataset, OpCodes.MapSymbolToOpCode);
|
---|
| 172 | var remainingRows = rows.Length % BATCHSIZE;
|
---|
| 173 | var roundedTotal = rows.Length - remainingRows;
|
---|
| 174 |
|
---|
| 175 | var result = new double[rows.Length];
|
---|
| 176 |
|
---|
| 177 | for (int rowIndex = 0; rowIndex < roundedTotal; rowIndex += BATCHSIZE) {
|
---|
| 178 | Evaluate(code, rows, rowIndex, BATCHSIZE);
|
---|
| 179 | Array.Copy(code[0].buf, 0, result, rowIndex, BATCHSIZE);
|
---|
| 180 | }
|
---|
| 181 |
|
---|
| 182 | if (remainingRows > 0) {
|
---|
| 183 | Evaluate(code, rows, roundedTotal, remainingRows);
|
---|
| 184 | Array.Copy(code[0].buf, 0, result, roundedTotal, remainingRows);
|
---|
| 185 | }
|
---|
| 186 |
|
---|
| 187 | return result;
|
---|
| 188 | }
|
---|
| 189 |
|
---|
[16293] | 190 | public IEnumerable<double> GetSymbolicExpressionTreeValues(ISymbolicExpressionTree tree, IDataset dataset, int[] rows) {
|
---|
[16296] | 191 | if (cachedData == null) {
|
---|
| 192 | InitCache(dataset);
|
---|
| 193 | }
|
---|
[16293] | 194 | return GetValues(tree, dataset, rows);
|
---|
| 195 | }
|
---|
| 196 |
|
---|
[16285] | 197 | public IEnumerable<double> GetSymbolicExpressionTreeValues(ISymbolicExpressionTree tree, IDataset dataset, IEnumerable<int> rows) {
|
---|
[16296] | 198 | return GetSymbolicExpressionTreeValues(tree, dataset, rows.ToArray());
|
---|
[16285] | 199 | }
|
---|
| 200 |
|
---|
| 201 | private BatchInstruction[] Compile(ISymbolicExpressionTree tree, IDataset dataset, Func<ISymbolicExpressionTreeNode, byte> opCodeMapper) {
|
---|
| 202 | var root = tree.Root.GetSubtree(0).GetSubtree(0);
|
---|
| 203 | var code = new BatchInstruction[root.GetLength()];
|
---|
| 204 | if (root.SubtreeCount > ushort.MaxValue) throw new ArgumentException("Number of subtrees is too big (>65.535)");
|
---|
| 205 | int c = 1, i = 0;
|
---|
| 206 | foreach (var node in root.IterateNodesBreadth()) {
|
---|
[16296] | 207 | if (node.SubtreeCount > ushort.MaxValue) throw new ArgumentException("Number of subtrees is too big (>65.535)");
|
---|
| 208 | code[i] = new BatchInstruction {
|
---|
| 209 | opcode = opCodeMapper(node),
|
---|
| 210 | narg = (ushort)node.SubtreeCount,
|
---|
| 211 | buf = new double[BATCHSIZE],
|
---|
| 212 | childIndex = c
|
---|
| 213 | };
|
---|
[16285] | 214 | if (node is VariableTreeNode variable) {
|
---|
| 215 | code[i].weight = variable.Weight;
|
---|
[16296] | 216 | if (cachedData.ContainsKey(variable.VariableName)) {
|
---|
| 217 | code[i].data = cachedData[variable.VariableName];
|
---|
| 218 | } else {
|
---|
| 219 | code[i].data = dataset.GetReadOnlyDoubleValues(variable.VariableName).ToArray();
|
---|
| 220 | cachedData[variable.VariableName] = code[i].data;
|
---|
| 221 | }
|
---|
[16285] | 222 | } else if (node is ConstantTreeNode constant) {
|
---|
| 223 | code[i].value = constant.Value;
|
---|
[16287] | 224 | for (int j = 0; j < BATCHSIZE; ++j)
|
---|
| 225 | code[i].buf[j] = code[i].value;
|
---|
[16285] | 226 | }
|
---|
| 227 | c += node.SubtreeCount;
|
---|
| 228 | ++i;
|
---|
| 229 | }
|
---|
| 230 | return code;
|
---|
| 231 | }
|
---|
| 232 | }
|
---|
| 233 | }
|
---|