xref: /aosp_15_r20/external/tensorflow/tensorflow/core/grappler/optimizers/auto_mixed_precision.h (revision b6fb3261f9314811a0f4371741dbb8839866f948)
1 /* Copyright 2017 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_GRAPPLER_OPTIMIZERS_AUTO_MIXED_PRECISION_H_
17 #define TENSORFLOW_CORE_GRAPPLER_OPTIMIZERS_AUTO_MIXED_PRECISION_H_
18 
19 #include "tensorflow/core/grappler/optimizers/graph_optimizer.h"
20 #include "tensorflow/core/platform/cpu_info.h"
21 #include "tensorflow/core/protobuf/rewriter_config.pb.h"
22 
23 namespace tensorflow {
24 namespace grappler {
25 
26 // CUDA: convert to float16 on GPU
27 // BF16: convert to bfloat16 on CPU
28 // CPU: emulate float16 on CPU without changing operator kernel
29 enum class AutoMixedPrecisionMode { CUDA, BF16, CPU };
30 
31 // Convert data types to float16 or bfloat16 where appropriate to improve
32 // performance on GPUs or CPUs.
33 class AutoMixedPrecision : public GraphOptimizer {
34  public:
35   // If 'mode' is CUDA, converts nodes to float16 on Nvidia GPUs. If BF16,
36   // converts nodes to bfloat16 on CPUs in order to take advantage of oneDNN
37   // performance improvements with bfloat16.
38   explicit AutoMixedPrecision(
39       AutoMixedPrecisionMode mode = AutoMixedPrecisionMode::CUDA)
mode_(mode)40       : mode_(mode) {}
41 
~AutoMixedPrecision()42   ~AutoMixedPrecision() override {}
43 
name()44   string name() const override {
45     switch (mode_) {
46       case AutoMixedPrecisionMode::CUDA:
47         return "auto_mixed_precision";
48       case AutoMixedPrecisionMode::BF16:
49         return "auto_mixed_precision_onednn_bfloat16";
50       case AutoMixedPrecisionMode::CPU:
51         return "auto_mixed_precision_cpu";
52       default:
53         LOG(FATAL) << "Invalid value for AutoMixedPrecisionMode: "  // Crash Ok
54                    << static_cast<int>(mode_);
55     }
56   };
57 
UsesFunctionLibrary()58   bool UsesFunctionLibrary() const override { return false; }
59 
60   Status Optimize(Cluster* cluster, const GrapplerItem& item,
61                   GraphDef* output) override;
62 
63  private:
64   const AutoMixedPrecisionMode mode_;
65 };
66 
67 }  // end namespace grappler
68 }  // end namespace tensorflow
69 
70 #endif  // TENSORFLOW_CORE_GRAPPLER_OPTIMIZERS_AUTO_MIXED_PRECISION_H_
71