xref: /freebsd/contrib/llvm-project/llvm/lib/Analysis/TrainingLogger.cpp (revision a8089ea5aee578e08acab2438e82fc9a9ae50ed8)
1 //===- TrainingLogger.cpp - mlgo feature/reward logging -------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file implements logging infrastructure for extracting features and
10 // rewards for mlgo policy training.
11 //
12 //===----------------------------------------------------------------------===//
13 #include "llvm/Analysis/TensorSpec.h"
14 #include "llvm/Config/config.h"
15 
16 #include "llvm/ADT/Twine.h"
17 #include "llvm/Analysis/Utils/TrainingLogger.h"
18 #include "llvm/Support/CommandLine.h"
19 #include "llvm/Support/Debug.h"
20 #include "llvm/Support/JSON.h"
21 #include "llvm/Support/MemoryBuffer.h"
22 #include "llvm/Support/Path.h"
23 #include "llvm/Support/raw_ostream.h"
24 
25 #include <cassert>
26 #include <numeric>
27 
28 using namespace llvm;
29 
30 void Logger::writeHeader(std::optional<TensorSpec> AdviceSpec) {
31   json::OStream JOS(*OS);
32   JOS.object([&]() {
33     JOS.attributeArray("features", [&]() {
34       for (const auto &TS : FeatureSpecs)
35         TS.toJSON(JOS);
36     });
37     if (IncludeReward) {
38       JOS.attributeBegin("score");
39       RewardSpec.toJSON(JOS);
40       JOS.attributeEnd();
41     }
42     if (AdviceSpec.has_value()) {
43       JOS.attributeBegin("advice");
44       AdviceSpec->toJSON(JOS);
45       JOS.attributeEnd();
46     }
47   });
48   *OS << "\n";
49 }
50 
51 void Logger::switchContext(StringRef Name) {
52   CurrentContext = Name.str();
53   json::OStream JOS(*OS);
54   JOS.object([&]() { JOS.attribute("context", Name); });
55   *OS << "\n";
56 }
57 
58 void Logger::startObservation() {
59   auto I = ObservationIDs.insert({CurrentContext, 0});
60   size_t NewObservationID = I.second ? 0 : ++I.first->second;
61   json::OStream JOS(*OS);
62   JOS.object([&]() {
63     JOS.attribute("observation", static_cast<int64_t>(NewObservationID));
64   });
65   *OS << "\n";
66 }
67 
68 void Logger::endObservation() { *OS << "\n"; }
69 
70 void Logger::logRewardImpl(const char *RawData) {
71   assert(IncludeReward);
72   json::OStream JOS(*OS);
73   JOS.object([&]() {
74     JOS.attribute("outcome", static_cast<int64_t>(
75                                  ObservationIDs.find(CurrentContext)->second));
76   });
77   *OS << "\n";
78   writeTensor(RewardSpec, RawData);
79   *OS << "\n";
80 }
81 
82 Logger::Logger(std::unique_ptr<raw_ostream> OS,
83                const std::vector<TensorSpec> &FeatureSpecs,
84                const TensorSpec &RewardSpec, bool IncludeReward,
85                std::optional<TensorSpec> AdviceSpec)
86     : OS(std::move(OS)), FeatureSpecs(FeatureSpecs), RewardSpec(RewardSpec),
87       IncludeReward(IncludeReward) {
88   writeHeader(AdviceSpec);
89 }
90