1 /* Copyright 2015 The TensorFlow Authors. All Rights Reserved.
2
3 Licensed under the Apache License, Version 2.0 (the "License");
4 you may not use this file except in compliance with the License.
5 You may obtain a copy of the License at
6
7 http://www.apache.org/licenses/LICENSE-2.0
8
9 Unless required by applicable law or agreed to in writing, software
10 distributed under the License is distributed on an "AS IS" BASIS,
11 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 See the License for the specific language governing permissions and
13 limitations under the License.
14 ==============================================================================*/
15
16 #ifndef TENSORFLOW_CORE_KERNELS_GPU_UTILS_H_
17 #define TENSORFLOW_CORE_KERNELS_GPU_UTILS_H_
18
19 #if GOOGLE_CUDA || TENSORFLOW_USE_ROCM
20
21 #include <unordered_map>
22
23 #include "absl/types/span.h"
24 #include "tensorflow/core/framework/tensor.h"
25 #include "tensorflow/core/lib/core/status.h"
26 #include "tensorflow/core/lib/strings/str_util.h"
27 #include "tensorflow/core/lib/strings/strcat.h"
28 #include "tensorflow/core/lib/strings/stringprintf.h"
29 #include "tensorflow/core/platform/logging.h"
30 #include "tensorflow/core/platform/stream_executor.h"
31
32 namespace stream_executor {
33 class RedzoneAllocator;
34 } // namespace stream_executor
35
36 namespace tensorflow {
37
38 class NodeDef;
39 class AutotuneResult;
40
41 // Return whether the redzone check is disabled.
42 //
43 // Controlled by the TF_DISABLE_RZ_CHECK environment variable.
44 bool RedzoneCheckDisabled();
45
46 // Return an allocated buffer with redzones the size of `buffer`. Does
47 // *not* copy the contents of the `buffer` into the newly allocated buffer:
48 // assumes that buffer is a pure out-parameter.
49 //
50 // Returns `buffer` if RedzoneCheckDisabled() is true.
51 //
52 // On error, return `buffer`, and log an error message (once).
53 se::DeviceMemoryBase WrapRedzoneBestEffort(se::RedzoneAllocator* rz_allocator,
54 se::DeviceMemoryBase buffer);
55
56 // Check the passed allocator for redzone violations.
57 // If violations have occurred, mark the corresponding autotune result
58 // as a failure.
59 void CheckRedzones(const se::RedzoneAllocator& rz_allocator,
60 tensorflow::AutotuneResult* autotune_result);
61
62 template <typename T>
AsDeviceMemory(const T * cuda_memory,uint64 size)63 inline se::DeviceMemory<T> AsDeviceMemory(const T* cuda_memory, uint64 size) {
64 se::DeviceMemoryBase wrapped(const_cast<T*>(cuda_memory), size * sizeof(T));
65 se::DeviceMemory<T> typed(wrapped);
66 return typed;
67 }
68
69 // A helper class that looks up the best autotuned config from parameters.
70 // Due to the noisy nature of autotune, especially with multiple devices, it
71 // only accepts a config if its margin exceeds a threshold.
72 // For the same shape configs, if a new best config matches the previous best,
73 // they get promoted; otherwise, the winner gets demoted. This process stops
74 // when the winner's score exceeds the threshold.
75 // In a bad case when two configs are very close to each other and flips
76 // back and forth randomly, the expected number of experiments before autotune
77 // settles is O(threshold ^ 2). So we recommend that number of warmup runs
78 // for any benchmarks.
79 template <typename Parameters, typename Config>
80 class AutoTuneMap {
81 public:
Find(const Parameters & params,Config * config)82 bool Find(const Parameters& params, Config* config) const {
83 mutex_lock lock(mu_);
84 auto iter = params_config_map_.find(params);
85 if (iter == params_config_map_.end() ||
86 (iter->second.score < min_score_threshold_ &&
87 iter->second.count <= max_autotune_count_)) {
88 return false;
89 }
90 *config = iter->second.config;
91 return true;
92 }
Insert(const Parameters & params,const Config & config)93 void Insert(const Parameters& params, const Config& config) {
94 mutex_lock lock(mu_);
95 auto iter = params_config_map_.find(params);
96 int new_score = 0;
97 if (iter == params_config_map_.end()) {
98 // Create a new entry if params is new.
99 VLOG(1) << GetActionSummary("creates", params, config);
100 params_config_map_.insert(
101 std::make_pair(params, ValueType{config, 1, 1}));
102 new_score = 1;
103 } else if (iter->second.score < min_score_threshold_ &&
104 iter->second.count <= max_autotune_count_) {
105 DCHECK_GT(iter->second.score, 0);
106 if (iter->second.config != config) {
107 // If it is different from the current winner, demotes the winner.
108 VLOG(1) << GetActionSummary("demotes", params, config);
109 new_score = --iter->second.score;
110 ++iter->second.count;
111 if (new_score <= 0) {
112 VLOG(1) << GetActionSummary("erases", params, config);
113 params_config_map_.erase(iter);
114 }
115 } else {
116 // If it is the same as the current winner, promotes the winner.
117 VLOG(1) << GetActionSummary("promotes", params, config);
118 new_score = ++iter->second.score;
119 ++iter->second.count;
120 }
121 }
122 if (new_score >= min_score_threshold_) {
123 VLOG(1) << GetActionSummary("accepts", params, config);
124 } else if (autotune_global_count_ >= max_autotune_global_count_) {
125 // The autotuning exceeds the max iteration threshold and we accept the
126 // the winner if it exists in the map, otherwise we accept the current
127 // winner.
128 auto winner = params_config_map_.find(params);
129 if (winner == params_config_map_.end()) {
130 VLOG(1) << GetActionSummary("creates", params, config);
131 for (int i = 0; i < min_score_threshold_; ++i) {
132 VLOG(1) << GetActionSummary("promotes", params, config);
133 }
134 params_config_map_.insert(
135 std::make_pair(params, ValueType{config, min_score_threshold_, 1}));
136 } else {
137 int promotes_times = min_score_threshold_ - winner->second.score;
138 for (int i = 0; i < promotes_times; ++i) {
139 VLOG(1) << GetActionSummary("promotes", params, config);
140 }
141 winner->second.score = min_score_threshold_;
142 }
143 VLOG(1) << GetActionSummary("accepts", params, config);
144 }
145 autotune_global_count_++;
146 }
147
148 private:
AutoTuneMap(const std::string & name)149 AutoTuneMap(const std::string& name) : name_(name) {
150 min_score_threshold_ = 1;
151 int min_warmup_iterations = 10;
152 const char* threshold_str = getenv("TF_AUTOTUNE_THRESHOLD");
153 if (threshold_str != nullptr) {
154 VLOG(1) << "TF_AUTOTUNE_THRESHOLD = " << threshold_str;
155 strings::safe_strto32(threshold_str, &min_score_threshold_);
156 }
157 const char* min_warmup_iteration_str =
158 getenv("TF_AUTOTUNE_MIN_WARMUP_ITERATIONS");
159 if (min_warmup_iteration_str != nullptr) {
160 strings::safe_strto32(min_warmup_iteration_str, &min_warmup_iterations);
161 }
162 min_score_threshold_ = std::max(min_score_threshold_, 1);
163 max_autotune_count_ = std::max(
164 5 * min_score_threshold_ * min_score_threshold_, min_warmup_iterations);
165 max_autotune_global_count_ = 2 * max_autotune_count_;
166 autotune_global_count_ = 0;
167 }
168
169 template <class Group, class Params, class Cfg>
170 friend class AutoTuneSingleton;
171
172 struct Hasher {
operatorHasher173 std::size_t operator()(const Parameters& parameter) const {
174 return parameter.hash();
175 }
176 };
177
GetActionSummary(StringPiece action,const Parameters & params,const Config & config)178 std::string GetActionSummary(StringPiece action, const Parameters& params,
179 const Config& config) {
180 return strings::Printf("autotune_map %s %s: %s -> (%s)", name_.c_str(),
181 string(action).c_str(), params.ToString().c_str(),
182 config.ToString().c_str());
183 }
184
185 mutable mutex mu_;
186 struct ValueType {
187 Config config;
188 int32 score;
189 int32 count;
190 };
191 std::unordered_map<Parameters, ValueType, Hasher> params_config_map_
192 TF_GUARDED_BY(mu_);
193 std::string name_;
194 int32 min_score_threshold_;
195 int32 max_autotune_count_;
196 int32 max_autotune_global_count_;
197 int32 autotune_global_count_;
198
199 TF_DISALLOW_COPY_AND_ASSIGN(AutoTuneMap);
200 };
201
202 // A Singleton helper that manages the global autotune results by groups.
203 // The caller specified arbitrary Group type that can distinguish between
204 // different autotune results, even if their Parameters and Configs are the
205 // same.
206 template <class Group, typename Parameters, typename Config>
207 class AutoTuneSingleton {
208 public:
209 typedef AutoTuneMap<Parameters, Config> AutoTuneType;
GetInstance()210 static AutoTuneType* GetInstance() {
211 static AutoTuneType* instance = new AutoTuneType(Group::name());
212 return instance;
213 }
214 };
215
216 // Logs convolution results to customized back-storage.
217 void LogConvAutotuneResults(se::dnn::ConvolutionKind kind,
218 se::dnn::DataType element_type,
219 se::DeviceMemoryBase input_buffer,
220 se::DeviceMemoryBase filter_buffer,
221 se::DeviceMemoryBase output_buffer,
222 const se::dnn::BatchDescriptor& input_desc,
223 const se::dnn::FilterDescriptor& filter_desc,
224 const se::dnn::BatchDescriptor& output_desc,
225 const se::dnn::ConvolutionDescriptor& conv_desc,
226 se::StreamExecutor* stream_exec,
227 absl::Span<const AutotuneResult> results);
228
229 // Logs fused convolution results to customized back-storage.
230 void LogFusedConvForwardAutotuneResults(
231 se::dnn::DataType element_type, se::DeviceMemoryBase input_buffer,
232 se::DeviceMemoryBase filter_buffer, se::DeviceMemoryBase output_buffer,
233 se::DeviceMemoryBase bias_buffer, se::DeviceMemoryBase side_input_buffer,
234 const se::dnn::BatchDescriptor& input_desc,
235 const se::dnn::FilterDescriptor& filter_desc,
236 const se::dnn::BatchDescriptor& output_desc,
237 const se::dnn::ConvolutionDescriptor& conv_desc, double conv_scale,
238 double side_value_scale, se::dnn::ActivationMode activation_mode,
239 se::StreamExecutor* stream_exec, absl::Span<const AutotuneResult> results);
240
241 // Returns the best algorithms for the config, one is the fastest, the other is
242 // other is fastest with 0 scratch space. Unsuccessful autotuning results are
243 // allowed and ignored.
244 Status BestCudnnConvAlgorithm(absl::Span<const AutotuneResult> results,
245 se::dnn::AlgorithmConfig* algo);
246
247 } // namespace tensorflow
248
249 #endif // GOOGLE_CUDA || TENSORFLOW_USE_ROCM
250
251 #endif // TENSORFLOW_CORE_KERNELS_GPU_UTILS_H_
252