1 /*
2 * Copyright (C) 2021 The Android Open Source Project
3 *
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
7 *
8 * http://www.apache.org/licenses/LICENSE-2.0
9 *
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
15 */
16
17 #include "ResilientExecution.h"
18
19 #include "InvalidBurst.h"
20 #include "ResilientBurst.h"
21
22 #include <android-base/logging.h>
23 #include <android-base/thread_annotations.h>
24 #include <nnapi/IExecution.h>
25 #include <nnapi/Result.h>
26 #include <nnapi/TypeUtils.h>
27 #include <nnapi/Types.h>
28
29 #include <functional>
30 #include <memory>
31 #include <mutex>
32 #include <sstream>
33 #include <utility>
34 #include <vector>
35
36 namespace android::hardware::neuralnetworks::utils {
37 namespace {
38
39 template <typename FnType>
protect(const ResilientExecution & resilientExecution,const FnType & fn)40 auto protect(const ResilientExecution& resilientExecution, const FnType& fn)
41 -> decltype(fn(*resilientExecution.getExecution())) {
42 auto execution = resilientExecution.getExecution();
43 auto result = fn(*execution);
44
45 // Immediately return if prepared model is not dead.
46 if (result.has_value() || result.error().code != nn::ErrorStatus::DEAD_OBJECT) {
47 return result;
48 }
49
50 // Attempt recovery and return if it fails.
51 auto maybeExecution = resilientExecution.recover(execution.get());
52 if (!maybeExecution.has_value()) {
53 const auto& [message, code] = maybeExecution.error();
54 std::ostringstream oss;
55 oss << ", and failed to recover dead prepared model with error " << code << ": " << message;
56 result.error().message += oss.str();
57 return result;
58 }
59 execution = std::move(maybeExecution).value();
60
61 return fn(*execution);
62 }
63
64 } // namespace
65
create(Factory makeExecution)66 nn::GeneralResult<std::shared_ptr<const ResilientExecution>> ResilientExecution::create(
67 Factory makeExecution) {
68 if (makeExecution == nullptr) {
69 return NN_ERROR(nn::ErrorStatus::INVALID_ARGUMENT)
70 << "utils::ResilientExecution::create must have non-empty makeExecution";
71 }
72 auto execution = NN_TRY(makeExecution());
73 CHECK(execution != nullptr);
74 return std::make_shared<ResilientExecution>(PrivateConstructorTag{}, std::move(makeExecution),
75 std::move(execution));
76 }
77
ResilientExecution(PrivateConstructorTag,Factory makeExecution,nn::SharedExecution execution)78 ResilientExecution::ResilientExecution(PrivateConstructorTag /*tag*/, Factory makeExecution,
79 nn::SharedExecution execution)
80 : kMakeExecution(std::move(makeExecution)), mExecution(std::move(execution)) {
81 CHECK(kMakeExecution != nullptr);
82 CHECK(mExecution != nullptr);
83 }
84
getExecution() const85 nn::SharedExecution ResilientExecution::getExecution() const {
86 std::lock_guard guard(mMutex);
87 return mExecution;
88 }
89
recover(const nn::IExecution * failingExecution) const90 nn::GeneralResult<nn::SharedExecution> ResilientExecution::recover(
91 const nn::IExecution* failingExecution) const {
92 std::lock_guard guard(mMutex);
93
94 // Another caller updated the failing prepared model.
95 if (mExecution.get() != failingExecution) {
96 return mExecution;
97 }
98
99 mExecution = NN_TRY(kMakeExecution());
100 return mExecution;
101 }
102
103 nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>>
compute(const nn::OptionalTimePoint & deadline) const104 ResilientExecution::compute(const nn::OptionalTimePoint& deadline) const {
105 const auto fn = [&deadline](const nn::IExecution& execution) {
106 return execution.compute(deadline);
107 };
108 return protect(*this, fn);
109 }
110
111 nn::GeneralResult<std::pair<nn::SyncFence, nn::ExecuteFencedInfoCallback>>
computeFenced(const std::vector<nn::SyncFence> & waitFor,const nn::OptionalTimePoint & deadline,const nn::OptionalDuration & timeoutDurationAfterFence) const112 ResilientExecution::computeFenced(const std::vector<nn::SyncFence>& waitFor,
113 const nn::OptionalTimePoint& deadline,
114 const nn::OptionalDuration& timeoutDurationAfterFence) const {
115 const auto fn = [&waitFor, &deadline,
116 &timeoutDurationAfterFence](const nn::IExecution& execution) {
117 return execution.computeFenced(waitFor, deadline, timeoutDurationAfterFence);
118 };
119 return protect(*this, fn);
120 }
121
isValidInternal() const122 bool ResilientExecution::isValidInternal() const {
123 return true;
124 }
125
126 } // namespace android::hardware::neuralnetworks::utils
127